Method for constructing fault cause and effect relationship of new energy ship

Abstract

The new energy ship fault causal relationship construction method provided in the application comprises the following steps: when a fault occurs, multi-modal heterogeneous data is fused and a core feature set is obtained through screening; an environment-invariant feature matrix is obtained through invariant risk minimization learning; the causal edge strength between each variable feature is evaluated based on mutual information to obtain a causal edge strength matrix; a core fault causal skeleton is mined by combining the causal edge strength threshold and conditional independence test; based on the skeleton and the hierarchical node system, an initial hierarchical fault causal graph is constructed, the initial causal edge weight is determined by fusing the causal edge strength matrix, and the target hierarchical fault causal graph is obtained by dynamically updating the meta-learning model; after the target hierarchical fault causal graph is corrected and optimized by counterfactual samples, the reasoning path is optimized, then the implicit causal relationship is mined by combining the structured reasoning and the large language model, and the fault causal relationship of the new energy ship is obtained. Thus, stable and self-adaptive fault causal relationship mining is realized under complex dynamic working conditions.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis and predictive maintenance technology for marine new energy systems, and in particular to a method for constructing causal relationships for faults in new energy ships. Background Technology

[0002] With the green and intelligent transformation of the shipbuilding industry, new energy ships have become the mainstream direction of shipping development. Their multi-subsystem coupled architecture, consisting of power systems, propulsion systems, and energy storage systems, has become the core unit of ship operation. This architecture is complex, with variable operating conditions, and faults exhibit typical characteristics such as multiple causes and effects, significant differences in cross-environmental manifestations, dynamic evolution of fault chains, and implicit causal relationships. Accurately identifying fault causal relationships and constructing fault cause-effect graphs adapted to dynamic operating conditions are the core prerequisites for achieving accurate fault diagnosis and predictive maintenance of complex faults in new energy ships.

[0003] In current technologies for fault diagnosis and causal relationship mining of new energy ships, traditional static cause-effect graph models and data-driven models struggle to adapt to the dynamic changes in the same fault characterization parameters as ships operate under different sea areas, climates, and load conditions. This results in a high misdiagnosis rate in cross-environment fault diagnosis and poor stability of diagnostic results. Furthermore, existing cause-effect graph models are mostly statically constructed, failing to adapt to real-time changes in system operating states and the dynamic evolution of fault chains, leading to problems such as diagnostic lag and difficulty in tracing multi-cause, multi-effect fault chains.

[0004] In summary, it is currently difficult to achieve stable and adaptive fault causality mining under complex dynamic operating conditions. Summary of the Invention

[0005] The purpose of this application is to at least address one of the aforementioned technical deficiencies, particularly the current difficulty in achieving stable and adaptive fault causality mining under complex dynamic operating conditions.

[0006] Firstly, this application provides a method for constructing causal relationships for failures in new energy ships, the method comprising:

[0007] When a new energy vessel malfunctions, the core feature set is obtained by integrating the multimodal heterogeneous data of the new energy vessel and then filtering it.

[0008] The core feature set is subjected to invariant risk minimization learning to obtain the environment invariant feature matrix. Based on mutual information, the causal edge strength between the features of each variable in the environment invariant feature matrix is ​​evaluated to obtain the causal edge strength matrix. Based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, the core fault causal skeleton is mined.

[0009] Based on the core fault causal skeleton and the preset hierarchical node system, an initial hierarchical fault causal graph is constructed. The causal edge strength matrix is ​​integrated to determine the initial causal edge weights. Based on the meta-learning model and the initial causal edge weights, the initial hierarchical fault causal graph is dynamically updated to obtain the target hierarchical fault causal graph.

[0010] By using counterfactual samples and correcting the causal edge weights of the hierarchical fault causal graph, the reasoning path is optimized. Based on the optimized hierarchical fault causal graph, the causal relationships of faults in new energy ships are obtained through structured reasoning and the implicit causal relationships mined by the large language model.

[0011] In one embodiment, the step of fusing multimodal heterogeneous data from new energy ships and then filtering it to obtain a core feature set includes:

[0012] Preprocessing of multimodal heterogeneous data from new energy ships yields sensor features, text features, and environmental features;

[0013] The obtained sensor features, text features, and environmental features are mapped to a unified feature space and fused through a cross-modal self-attention mechanism to obtain fused features;

[0014] A large language model fine-tuned for the shipbuilding domain is used to evaluate the fault correlation of each feature channel in the fused features, and the importance score of each feature channel is obtained.

[0015] Based on the importance scores of each feature channel, feature channels that are higher than a preset threshold are selected to form the core feature set.

[0016] In one embodiment, the step of performing invariant risk minimization learning on the core feature set to obtain the environment-invariant feature matrix includes:

[0017] The environment-invariant feature matrix is ​​obtained using the following formula:

[0018]

[0019] in, Represents the environment-invariant feature matrix. This represents a feature extraction network. For feature extraction network parameters, Represents the core feature set, This represents the activation function. This represents the weight matrix of the feature extraction network. The bias vector represents the weighted evidence; the feature extraction network is trained using an optimization objective of invariant risk minimization.

[0020]

[0021] in, This represents a subset of environments, obtained by dividing the ship's operating environment based on the core feature set. Represents a single subset of the environment. , These represent the feature data and fault labels of a single environmental subset, respectively. The loss function represents a single subset of environments. Indicates the parameters of the prediction head network. This represents the gradient of the loss function with respect to the parameters of the feature extraction network. This represents the gradient penalty coefficient.

[0022] In one embodiment, the step of evaluating the causal edge strength matrix based on mutual information to assess the causal edge strength between the variable features in the environment-invariant feature matrix includes:

[0023] The causal edge strength matrix is ​​obtained using the following formula:

[0024]

[0025]

[0026] in, Representing variable characteristics arrive The causal edge strength, i.e., the causal edge strength matrix, , The first element in the environment-invariant characteristic matrix The and the first Individual variable characteristics, Indicates the fault label variable. Indicates control Under unchanged conditions, and Conditional mutual information, used for quantization and The degree of direct correlation, Control in the display Under unchanged conditions, and Conditional mutual information, used for quantization and The degree of spurious association, This represents the preset spurious association correction coefficient, used to adjust the strength of spurious association removal. express , , The joint probability density function, Indicates in Given the conditions, and The joint conditional probability density, , They represent in Given the conditions, and The conditional probability density.

[0027] In one embodiment, the steps of mining the core fault causal skeleton based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, include:

[0028] Using the features of each variable in the environment-invariant feature matrix as nodes, and the pairs of variable features in the causal edge strength matrix whose strength values ​​are greater than the causal edge strength threshold as candidate causal edges, an initial undirected graph is constructed.

[0029] For each edge connecting two nodes in the initial undirected graph, a candidate condition set is selected from the remaining nodes. The conditional independence test is used to determine whether the two nodes are conditionally independent under the conditions of the candidate condition set. If there is at least one candidate condition set that makes the two nodes conditionally independent, the edge is determined to be indirectly related and is removed from the initial undirected graph. If after traversing all candidate condition sets, the two nodes do not satisfy the conditional independence, the edge is retained as a valid causal edge.

[0030] By determining the direction of valid causal edges and identifying the causal orientation between variable features, a directed core fault causal skeleton is obtained.

[0031] In one embodiment, the step of fusing the causal edge strength matrix to determine the initial causal edge weights includes:

[0032] The initial causal edge weights are determined using the following formula:

[0033]

[0034] in, Represents nodes in the initial hierarchical fault cause graph To the node The initial causal edge weights, Represents the nodes obtained from the causal edge strength matrix. To the node The causal edge strength, This represents the set of all causal edges in the initial hierarchical fault causality graph. This represents the maximum causal edge strength among all causal edges in the initial hierarchical fault causal graph. Represents a node To the node Experience rating of causal relationships.

[0035] In one embodiment, the step of dynamically updating the initial hierarchical fault causality graph based on the meta-learning model and initial causal edge weights to obtain the target hierarchical fault causality graph includes:

[0036] By pre-training a meta-learning model using an update task set containing various environmental changes and fault types, general update parameters applicable to different update scenarios can be obtained.

[0037] When the current scenario of a new energy ship is detected to be a new environment or a new fault, a small number of current samples are obtained, and the initial causal edge weights of the initial hierarchical fault causal graph are fine-tuned based on the general update parameters to obtain the updated causal edge weights. Based on a small number of current samples, the causal edge strength under the current scenario is recalculated to obtain the updated causal edge strength matrix.

[0038] In the initial hierarchical fault cause graph, invalid causal edges with weights lower than a preset removal threshold are removed based on the updated causal edge weights, and new causal edges with causal edge strengths higher than a preset addition threshold are added based on the updated causal edge strength matrix, so as to obtain the target hierarchical fault cause graph.

[0039] In one embodiment, the steps of removing invalid causal edges with weights lower than a preset removal threshold based on the updated causal edge weights, and adding new causal edges with strengths higher than a preset addition threshold based on the updated causal edge strength matrix, include:

[0040] The following formula is used to remove invalid causal changes and add new causal edges:

[0041]

[0042] in, This represents the set of causal edges in the updated hierarchical fault cause-effect graph. This represents the set of causal edges before the update. Let represent any two nodes in the initial hierarchical fault cause-effect graph. Represents a node To the node Updated causal edge weights This indicates the preset rejection threshold. Represents a node To the node The updated causal edge strength, i.e., the updated causal edge strength matrix. This indicates a preset threshold for adding values.

[0043] In one embodiment, the step of correcting the causal edge weights of a target hierarchical fault causal graph using counterfactual samples includes:

[0044] Using a large language model, based on real fault samples and a hierarchical fault cause-effect graph, counterfactual samples that conform to causal logic are generated.

[0045] The counterfactual samples are input into the target hierarchical fault cause-effect graph for inference to obtain the inference loss corresponding to the counterfactual samples, and the real fault samples are input into the target hierarchical fault cause-effect graph for inference to obtain the inference loss corresponding to the real samples.

[0046] Based on the difference between the inference loss of counterfactual samples and the inference loss of true samples, the weight correction of each causal edge in the hierarchical fault causal graph of the target is calculated.

[0047] Based on the weight correction amount, the weights of the causal edges of each causal edge in the hierarchical fault causal graph of the target are corrected to obtain the corrected hierarchical fault causal graph of the target.

[0048] In one embodiment, the step of inference path optimization includes:

[0049] The modified hierarchical fault cause graph of the target is transformed into a Markov decision process model, in which the nodes in the modified hierarchical fault cause graph of the target are used as the state space, the causal edges between nodes are used as the action space, and the weights of the causal edges in the modified hierarchical fault cause graph of the target are used as the state transition probabilities.

[0050] Based on the Markov decision process model, a reinforcement learning algorithm is used for iterative training to update the Q-values ​​of each state-action pair until convergence to obtain the optimal Q-value table. The reinforcement learning algorithm is trained based on a preset reward function, which is positively correlated with inference accuracy and negatively correlated with inference path length.

[0051] Based on the optimal Q-value table, the causal edge with the largest Q-value is selected as the corresponding optimal action at each node to form the optimal reasoning path from the fault feature node to the fault root cause node.

[0052] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0053] The method for constructing causal relationships for new energy ship failures provided in this application firstly obtains a core feature set by fusing and filtering multimodal heterogeneous data, and then performs invariant risk minimization learning on the core feature set to obtain an environment-invariant feature matrix, thus removing the influence of environmental disturbances on failure features. Based on mutual information, the causal edge strength between the variable features in the environment-invariant feature matrix is ​​evaluated to obtain a causal edge strength matrix. A dual screening process, combining a causal edge strength threshold and a conditional independence test, is then performed to uncover a stable and reliable core failure causal framework, solving the problem of unstable causal relationship mining under complex operating conditions. Secondly, an initial hierarchical failure causal graph is constructed based on the core failure causal framework and a pre-defined hierarchical node system. The initial causal edge weights are determined by fusing the causal edge strength matrix, and the initial hierarchical failure causal graph is dynamically updated based on a meta-learning model and the initial causal edge weights to obtain a target hierarchical failure causal graph. This enables the causal graph to adaptively update to new environments or new failures, solving the problem that static causal graphs are difficult to update dynamically. Finally, after correcting the causal edge weights of the hierarchical fault causal graph using counterfactual samples, the inference path is optimized. Based on the optimized hierarchical fault causal graph, the causal relationships of faults in new energy ships are obtained through structured reasoning combined with implicit causal relationships mined by a large language model. This enhances the robustness of the causal graph to unknown scenarios and improves the accuracy and efficiency of fault reasoning. Therefore, this application achieves stable and adaptive fault causal relationship mining under complex dynamic conditions, significantly improving the generalization ability, dynamic adaptability, and diagnostic accuracy of fault diagnosis for new energy ships. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating the method for constructing causal relationships for new energy ship failures provided in this application embodiment;

[0056] Figure 2 A schematic diagram of the structure of the new energy ship fault causation construction device provided in the embodiments of this application;

[0057] Figure 3 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] This application provides a method for constructing causal relationships for failures in new energy vessels. The following embodiments illustrate this method using a computer device as an example. It is understood that the computer device can be any device with data processing capabilities, including but not limited to a single server and server clusters. Figure 1 As shown, the method may include the following steps:

[0060] S101: When a new energy vessel malfunctions, the core feature set is obtained by integrating the multimodal heterogeneous data of the new energy vessel and then filtering it.

[0061] In this embodiment, a new energy ship refers to a ship that uses clean energy technologies such as electric propulsion, hybrid power, or fuel cells as its power source, and its operating status is monitored in real time by a multi-source heterogeneous sensor and data acquisition system; a fault refers to an abnormality or failure event that occurs in the key components of the new energy ship's power system, propulsion system, or energy storage system during navigation or berthing; multimodal heterogeneous data refers to various types of data from different sampling frequencies, different physical meanings, and different data structures, including but not limited to electrical parameters, mechanical vibration, temperature, pressure, and ship operating conditions; the core feature set refers to the set of key features extracted from the original multimodal heterogeneous data through data processing and feature selection methods that can effectively characterize the fault mode and has the lowest redundancy.

[0062] When a new energy vessel experiences a malfunction, the data acquisition process is first triggered, obtaining real-time multimodal heterogeneous data from sensor networks, controller area network buses, and navigation data recorders deployed in various key systems of the vessel. This data covers the voltage, current, and temperature sequences of the power battery, vibration signals and speed of the propulsion motor, power parameters of the power converter, ship position information provided by the global positioning system, and wind, wave, and current data collected by marine environmental sensors.

[0063] Because the sampling frequencies, data formats, and physical dimensions of the various data points differ, data fusion processing is required. Specifically, timestamp alignment technology is used to synchronize asynchronous data to a unified time base, and normalization methods are used to transform data with different dimensions to the same scale, thus forming a structured fused dataset. Next, feature extraction algorithms are used to extract time-domain statistical features, frequency-domain energy features, and time-frequency-domain wavelet packet features from the aligned data to fully reflect the fault's performance across different dimensions. To reduce data dimensionality and focus on fault-related information, feature importance assessment based on random forests or feature selection based on mutual information are further employed to select features with the highest correlation to the current fault mode and minimal redundancy, forming a core feature set. For example, in a fault scenario where the propeller experiences a sudden load increase due to biofouling, the above processing may allow the selection of specific harmonic components of the propulsion motor current, shaft torsional vibration frequency, and ship speed reduction rate as core features from the original data.

[0064] It's important to note that when new energy ships experience malfunctions, fusing multimodal heterogeneous data and then filtering it to obtain a core feature set is crucial. This is because the ship's monitoring system involves different components such as power batteries, propulsion motors, and power converters. The raw data collected by various sensors differ significantly in sampling frequency, physical dimensions, and data structure. Without fusion processing, this heterogeneous data would be difficult to use within the same analytical framework. Furthermore, the raw data contains a large amount of normal operating condition information and environmental noise unrelated to the malfunction. Directly inputting all of this into subsequent analysis would not only waste computational resources and reduce inference efficiency but also obscure key clues truly representing the malfunction due to redundant information. By fusing multimodal data, data from different sources can be unified in terms of time reference and physical scale, making them comparable and computable. Further filtering of the fused data allows for the extraction of the core features most representative of the malfunction from the complex information, effectively removing redundancy and noise. Data fusion solves the consistency problem of multi-source heterogeneous data, providing standardized input for subsequent analysis. Feature filtering achieves data dimensionality reduction and information purification, enabling the core feature set to retain the key characteristics of the fault in the most compact way. This lays a high-quality data foundation for the subsequent accurate mining of fault causal relationships and significantly improves the accuracy and robustness of fault diagnosis.

[0065] S102: Perform invariant risk minimization learning on the core feature set to obtain the environment invariant feature matrix. Based on mutual information, evaluate the causal edge strength between the features of each variable in the environment invariant feature matrix to obtain the causal edge strength matrix. Based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, mine the core fault causal skeleton.

[0066] In this embodiment, invariant risk minimization learning is a method aimed at learning feature representations that maintain stable predictive relationships in different environments or operating conditions from data under different environments or operating conditions. Its core objective is to remove the influence of environmental factors on features and extract common features that are strongly correlated with the fault itself and are insensitive to environmental changes. The environmental invariant feature matrix refers to the data representation in matrix form obtained after invariant risk minimization learning, which consists of fault features that maintain a stable distribution in different operating environments. Mutual information is an index in information theory used to measure the degree of interdependence between two random variables; the larger the value, the stronger the correlation between the two variables. Causal edge strength is calculated based on mutual information and is used to quantify variable features. The numerical values ​​representing the degree of causal relationship between variables; the causal edge strength matrix is ​​a matrix composed of the strength values ​​of causal edges between each pair of features of all variables; the causal edge strength threshold is a pre-set numerical limit for screening strong causal relationships, and only causal edges higher than this threshold are initially considered as potential valid causal associations; the conditional independence test is a statistical hypothesis testing method used to determine whether two variables are independent of each other given a third variable, and is an important means of eliminating spurious correlations in causal discovery; the core fault causal skeleton refers to the simplest directed graph structure that reflects the propagation path and causal dependence of faults in new energy ships after environmental invariant feature extraction and dual causal screening, and is the basic skeleton for constructing the final fault causal graph.

[0067] After a new energy vessel experiences a malfunction and its core feature set is obtained, this core feature set is used as input to perform invariant risk minimization learning. Specifically, the collected core feature set is first divided into multiple subsets based on different sea areas, climate conditions, or load conditions encountered during the vessel's voyage. Each subset represents a specific environment or operating condition. Then, a feature extraction model is constructed. The goal of this model is to simultaneously minimize the prediction loss across all environmental subsets, and constraints are introduced to minimize the distribution differences of the extracted features across different environments, thus forcing the model to learn feature representations insensitive to environmental changes. Through optimization, the model outputs a transformation matrix, which maps the original core feature set to a new feature space, resulting in an environment-invariant feature matrix. Each column of this matrix corresponds to an environment-invariant feature, which has been freed from the interference of environmental factors such as sea state and climate, and better reflects the intrinsic attributes of the malfunction itself.

[0068] Next, a preliminary assessment of causal relationships is performed based on the environment-invariant feature matrix. The mutual information value between any two variable features in this matrix is ​​calculated; the magnitude of the mutual information value reflects the statistical dependence between the two features, including linear and nonlinear associations. All pairwise mutual information values ​​are arranged into a square matrix, yielding the causal edge strength matrix. The diagonal elements of this matrix are 0, and the off-diagonal elements represent the causal edge strength between the corresponding two features. Subsequently, a causal edge strength threshold is set; this threshold can be determined based on the statistical distribution of historical fault data or through cross-validation. Elements in the causal edge strength matrix below this threshold are set to 0, while elements above the threshold are retained as candidate causal edges, thus initially screening feature pairs that may have causal relationships.

[0069] To further eliminate spurious correlations and confirm genuine causal relationships, conditional independence tests are performed on candidate causal edges. For each pair of feature variables with candidate causal edges, other variables are progressively introduced as a condition set, and the conditional independence of these two variables is tested. If a condition set exists that makes them conditionally independent, the original candidate edge is considered a spurious correlation and is removed from the candidate set. By traversing all candidate edges and performing conditional independence tests, the remaining causal edges constitute a concise and reliable network structure, namely the core fault causal skeleton. This skeleton describes the causal transmission paths between key features of new energy ships in the current fault scenario in the form of a directed graph, providing a solid topological foundation for the subsequent construction of a hierarchical fault causal graph.

[0070] By performing invariant risk minimization learning on the core feature set, the perturbations caused by external factors such as navigation environment, climate conditions, and load changes on fault features can be fundamentally removed, thereby obtaining an environment-invariant feature matrix that truly reflects the essence of the fault, significantly improving the stability and generalization ability of feature representation. Based on this, the strength of causal edges between variables is evaluated using mutual information, and a causal edge strength matrix is ​​constructed, achieving a comprehensive quantitative measurement of potential causal relationships and capturing complex linear and nonlinear dependencies. Furthermore, a dual screening process is performed combining a causal edge strength threshold and a conditional independence test. The threshold quickly eliminates weak associations to reduce computational complexity, while the conditional independence test rigorously removes spurious correlations caused by confounding variables, ensuring the high accuracy and reliability of the discovered core fault causal framework. This process lays a solid foundation for subsequently constructing fault causal graphs that can adapt to complex dynamic operating conditions, effectively avoiding misdiagnosis caused by environmental changes or spurious associations, thereby improving the stability and accuracy of fault diagnosis for new energy vessels.

[0071] S103: Based on the core fault causal skeleton and the preset hierarchical node system, an initial hierarchical fault causal graph is constructed, the causal edge strength matrix is ​​integrated, the initial causal edge weights are determined, and the initial hierarchical fault causal graph is dynamically updated based on the meta-learning model and the initial causal edge weights to obtain the target hierarchical fault causal graph.

[0072] In this embodiment, the preset hierarchical node system refers to a multi-level node classification framework predefined according to the system composition structure and fault propagation law of new energy ships. This framework organizes the ship's power system, electric system, energy storage system, and various subsystem components according to a hierarchical relationship from the whole to the part, providing a structured basis for the placement of nodes in the fault causal graph. The initial hierarchical fault causal graph refers to the preliminary causal graph formed by arranging each causal node in a hierarchical manner according to the preset hierarchical node system based on the structure of the core fault causal skeleton. The initial causal edge weight refers to assigning the values ​​of the corresponding positions in the causal edge strength matrix to each causal edge in the initial hierarchical fault causal graph as the initial quantification value of the causal relationship strength of that edge. The meta-learning model is a machine learning model with the ability to learn how to learn. It can quickly adapt to new tasks after being trained on multiple related tasks. Here, it is used to capture the evolution law of the fault causal graph under different operating conditions. Dynamic update refers to the process of adjusting the structure and weight of the initial hierarchical fault causal graph using the meta-learning model based on the real-time operating data of new energy ships. The target hierarchical fault causal graph refers to the final causal graph that can be adapted to the current ship operating conditions after being dynamically updated by the meta-learning model.

[0073] After obtaining the core fault causal framework, the pre-defined hierarchical node system is first retrieved from the storage unit. This system is pre-designed based on the system engineering decomposition structure of new energy ships, dividing nodes into four levels: system layer, subsystem layer, equipment layer, and component layer. The system layer includes the power system, electrical system, and energy storage system, with each system layer further subdivided into corresponding subsystems and specific equipment. All feature variable nodes in the core fault causal framework are matched one by one to their corresponding positions in the pre-defined hierarchical node system according to their physical system and functional level. For example, feature nodes related to the voltage of a single power battery cell are assigned to the component layer of the energy storage system, and feature nodes related to the vibration of the propulsion motor are assigned to the equipment layer of the power system. After completing the hierarchical assignment of nodes, directed edges are added between nodes at each level according to the causal relationship determined in the core fault causal framework, thereby constructing an initial hierarchical fault causal graph with a clear hierarchical structure. This graph not only preserves the transmission relationship of the original causal framework but also makes the fault propagation path more intuitive through hierarchical organization, facilitating subsequent reasoning and analysis.

[0074] For example, the hierarchical cause-effect graph structure model is as follows:

[0075]

[0076] In the formula, For hierarchical fault cause-effect graphs; For a causal graph, the set of nodes. This is a set of environmental layer nodes (containing environmental variables such as sea area, temperature and humidity, salt spray, and load, with a total of 6 nodes). It is a set of system-level nodes (including the power system, power system, energy storage system and various subsystem units, with a total of 9 nodes). This is a set of fault layer nodes (containing various typical faults and fault development stages, with 15 nodes). It is a set of feature layer nodes (containing core multimodal feature variables that are strongly correlated with faults, with 248 nodes). It is a set of causal edges, which only includes causal edges between different levels or between adjacent faulty nodes at the same level, and does not include edges between unrelated nodes across levels to avoid redundancy; For the set of causal edge weights, For nodes To the node The causal edge weights range from [0,1]. A larger weight indicates a higher node weight. For nodes The stronger the causal influence, the better. This addresses the technical problem of traditional causal graphs having messy nodes and no hierarchical division, leading to unclear causal relationships and difficulties in tracing fault chains.

[0077] Next, the edge weights of the initial hierarchical fault cause-effect graph are initialized. The previously calculated causal edge strength matrix is ​​read, which stores the pairwise causal edge strength values ​​between all feature variables. For each directed edge in the initial hierarchical fault cause-effect graph, the corresponding elements of its starting and target nodes in the causal edge strength matrix are found, and the value of these elements is assigned as the initial causal edge weight for that edge. In this way, each causal relationship in the initial hierarchical fault cause-effect graph has a quantified strength index, which reflects the tightness of the causal association under the statistical significance of historical data.

[0078] To adapt the fault cause-effect graph to the dynamic operating conditions encountered by new energy vessels in actual operation, a meta-learning model is further introduced to dynamically update the initial cause-effect graph. Specifically, data from multiple fault cases occurring during different historical voyages, in different sea areas, and under different load conditions are first collected, and each case data is organized into an independent update task. Each task includes operational data for a period before and after the fault occurrence, as well as the corresponding initial hierarchical fault cause-effect graph. A graph neural network is constructed as the basic architecture of the meta-learning model, which can receive graph structure data and output the updated graph structure and node features. A model-independent meta-learning algorithm is used to train the graph neural network, with the training objective of learning how to quickly adjust the cause-effect graph to adapt to the data distribution of new tasks across multiple tasks. After training, the meta-learning model is deployed to the vessel's online monitoring system. When the vessel enters a new navigation condition or experiences a new fault, the real-time collected operational data stream and the initial hierarchical fault cause-effect graph are updated. Figure 1 The same meta-learning model is used as the input. Based on the data characteristics of the current operating conditions, the meta-learning model fine-tunes the edge weights in the initial causal graph, increasing the weights of causal edges strongly correlated with the current operating conditions and decreasing the weights of weakly correlated or irrelevant edges. It even adjusts some connection structures of the graph when necessary. After dynamic updates by the meta-learning model, a causal graph that accurately reflects the current ship operating environment and fault state is output—a target hierarchical fault causal graph. This graph possesses adaptive capabilities to new environments, ensuring the timeliness and accuracy of causal relationships.

[0079] It should be noted that constructing an initial hierarchical fault causal graph based on the core fault causal framework and a pre-defined hierarchical node system embeds the originally abstract causal framework into the physical hierarchical structure of the ship system. This allows the fault propagation path to correspond to the actual components of the ship, enhancing the interpretability and engineering applicability of the causal graph. The initial causal edge weights are determined by integrating the causal edge strength matrix, assigning each causal relationship a quantitative credibility based on historical data. This gives the causal graph not only structural information but also statistically significant strength measurements, laying the foundation for subsequent quantitative analysis. Dynamically updating the initial hierarchical fault causal graph based on a meta-learning model and initial causal edge weights enables the graph to learn evolutionary patterns from historical fault experience. It can adaptively adjust causal relationships and weights according to changes in the ship's real-time operating conditions, effectively overcoming the limitation of static causal graphs in adapting to dynamic environments. This ensures that the obtained target hierarchical fault causal graph accurately reflects the actual fault propagation mechanism under the current operating conditions, significantly improving the real-time performance and adaptability of fault diagnosis.

[0080] S104: Using counterfactual samples, after correcting the causal edge weights of the hierarchical fault causal graph of the target, the reasoning path is optimized. Based on the optimized hierarchical fault causal graph of the target, the causal relationship of the fault of new energy ships is obtained through structured reasoning and combined with the implicit causal relationship mined by the large language model.

[0081] In this embodiment, counterfactual samples refer to virtual samples generated through hypothetical intervention based on actual operational data of new energy ships. The core idea is to change the value of a key variable while keeping other conditions constant to observe its impact on the outcome, thereby simulating the evolution path of the fault under different hypothetical scenarios. Correcting the causal edge weights of the hierarchical fault causal graph involves recalibrating the strength of each causal relationship in the current causal graph using counterfactual samples, making the weights more reflective of the direction and magnitude of causal effects in the real world. Inference path optimization involves using a search algorithm, based on the corrected causal edge weights, to select the most probable and concise fault propagation path from numerous possible causal chains, eliminating redundant connections and highlighting the core fault chain. Structured inference refers to, based on the optimized hierarchical fault causal graph, following the graph's topology... The structure involves a logical reasoning process that traces backward from the fault symptom node to the root cause node, or forward from the root cause node to the consequence node; the large language model refers to a large neural network model pre-trained on massive text corpora, possessing powerful natural language understanding and knowledge association capabilities, and able to extract implicit fault association information from unstructured texts such as technical documents, maintenance records, and expert experience; implicit causal relationships refer to fault associations that are not directly detected by sensors or explicitly reflected in data, but actually exist in the physical logic of the system or operational experience, such as chain reactions caused by design defects, material aging patterns, or improper operation; the fault causal relationship of new energy ships refers to the final output that can completely describe the complete causal chain from the root cause to the final manifestation in the current fault event, including explicit data-driven causality and implicit knowledge-driven causality.

[0082] After obtaining the hierarchical fault causal graph, a counterfactual sample set is first constructed. Specifically, typical fault samples of the same type as the current fault are selected from the historical operation database of new energy ships. For key characteristic variables in the samples, such as the voltage of a single battery cell or the current of the propulsion motor, intervention-based reconstruction is performed using generative models such as generative adversarial networks or variational autoencoders. Intervention operations include artificially setting the value of a variable to a normal range, reversing the trend of a variable, or shifting the occurrence time of a variable, thereby generating a series of hypothetical virtual fault samples. These counterfactual samples simulate how the fault would evolve if a certain condition changed; for example, if the battery cell voltage did not suddenly drop at the time of the fault, would the motor still stop? All generated counterfactual samples are compiled into a dataset for subsequent causal weight correction.

[0083] Next, the weights of causal edges in the hierarchical fault causal graph are corrected using a counterfactual sample set. Both actual observation samples and counterfactual samples are simultaneously input into the causal graph model, and the average causal effect of each causal edge under different sample sets is calculated. Specifically, for a causal edge in the graph pointing from node A to node B, the strength of the causal effect is calculated by comparing the changes in the value of node B under the conditions of intervention A and no intervention A. If the counterfactual samples show that the change in B after intervention A is much smaller than the correlation between the two in actual observations, it indicates that the original weights may have overestimated the causal effect and need to be adjusted downwards. By traversing all causal edges and calibrating based on counterfactual inferences, a more robust set of causal edge weights that more closely reflects real physical laws is obtained, thus completing the correction of the hierarchical fault causal graph.

[0084] After weight correction, the causal graph is optimized for inference paths. Using either a shortest path algorithm or a path search algorithm based on causal strength weighting, the optimal causal chain from the final manifestation node of the fault to the potential root cause node is found in the corrected causal graph. During optimization, redundant edges with excessively low weights or those deemed false by conditional independence tests are removed, retaining the core path with the strongest causal effect and the simplest propagation path. For example, in a fault scenario involving a ship's electric propulsion system, the optimization algorithm might select the simplest and most effective path from multiple possible causal chains—from an abnormal inverter control signal to motor torque fluctuations and then to a decrease in propulsion power—while discarding other weakly correlated side paths, making the causal graph more focused on the core propagation mechanism of the fault.

[0085] Subsequently, structured reasoning is performed based on the optimized hierarchical fault causal graph. Following the graph's hierarchical structure, starting from the top-level fault symptom node, the process traverses downstream or upstream along directed edges. During backward reasoning, starting from the detected abnormal symptom, all possible preceding nodes are searched along the reverse direction of the causal edges. The confidence level of each candidate root cause node is calculated by combining the causal weights of each node with real-time monitoring data. During forward reasoning, starting from the suspected root cause node, the potential subsequent fault chains it might trigger are deduced to predict the evolution trend of the fault. The entire structured reasoning process fully utilizes the topological and edge weight information of the causal graph, achieving logically clear and interpretable fault tracing and deduction.

[0086] While performing structured reasoning, a large language model is invoked to uncover implicit causal relationships. Textual materials such as design specifications for new energy ships, historical maintenance work orders, fault case libraries, and expert experience documents are input into the large language model. Specialized prompts guide the model to extract implicit fault-related knowledge. For example, the prompting model summarizes the potential correlation between battery pack consistency differences and thermal runaway from maintenance records, or identifies the physical logic between cooling system faults and inverter overheating from design documents. The large language model outputs structured implicit causal relationship triples, including cause nodes, result nodes, and relationship descriptions. These implicit causal relationships are then integrated with the optimized hierarchical fault causal graph. Implicit relationships between existing nodes in the graph are added as new edges or supplementary weights to the graph structure; new nodes not covered in the graph are incorporated into the appropriate level of the causal graph after expert confirmation.

[0087] Finally, the explicit results of structured reasoning are integrated with the implicit causal relationships mined by the large language model to form a complete causal relationship for new energy ship failures. This causal relationship is presented in the form of a visualized causal graph, which not only includes the explicit failure propagation paths mined from real-time data but also labels the implicit associations extracted from textual knowledge, with a corrected causal weight attached to each edge. Simultaneously, a detailed failure reasoning report is output, tracing back from the final failure phenomenon to the root cause step by step, and using implicit knowledge mined by the large language model as supplementary explanations at key nodes, thus providing ship maintenance personnel with comprehensive, accurate, and interpretable failure diagnosis conclusions.

[0088] By using counterfactual samples to correct the causal edge weights of the hierarchical fault causal graph, the evolutionary behavior of faults under assumed conditions can be simulated, effectively eliminating spurious associations caused by data randomness. This makes the estimation of causal weights more robust and closer to physical reality, significantly enhancing the robustness of the causal graph to unknown scenarios and extreme conditions. Based on the weight correction, inference path optimization is performed. By selecting the simplest and most effective causal path, the interference of redundant information on fault tracing is eliminated, making subsequent structured inference more focused and efficient. Structured inference based on the optimized causal graph fully utilizes the graph's hierarchical topology and quantified weights, achieving accurate tracing and deduction consistent with causal logic. Simultaneously, by combining implicit causal relationships mined from large language models, the limitations of purely data-driven methods are overcome. Expert experience, design logic, and operational knowledge accumulated in textual materials are integrated into the fault analysis process. This ensures that the final causal relationships of new energy ship faults not only include explicit data associations but also encompass implicit physical logic and experiential wisdom, thereby significantly improving the comprehensiveness, accuracy, and interpretability of fault diagnosis.

[0089] Furthermore, addressing the issue of insufficient interpretability of complex fault reasoning results, an interpretability framework is constructed based on causal reasoning theory and large-scale model generation capabilities. The core objective of this framework is to transform the fault reasoning process into explicit causal logic, thereby enhancing maintenance personnel's trust in diagnostic results and their willingness to apply them, ultimately forming a complete closed loop from construction to reasoning to interpretation. Within this framework, Do-Calculus intervention calculus is first used to logically analyze the reasoning process of the causal graph. By quantifying the impact of intervention operations on fault nodes, the causal logical relationships between the root cause, intermediate fault nodes, and fault characteristics are clarified, ensuring that the interpretability quantification score of the causal logic is no less than 0.90. Subsequently, the reasoning results of the causal graph, including the root cause, fault chain, and causal edge strength, are input into a large-scale model fine-tuned for the marine domain. The large-scale model then generates a natural language fault diagnosis and interpretation report. This report details the root cause, fault development process, fault impact range, and preliminary maintenance recommendations, thus achieving a natural language presentation of the reasoning logic. At the same time, a ship-shore cloud collaborative cause-effect graph visualization platform was developed. This platform can realize the dynamic display of hierarchical fault cause-effect graphs, real-time tracing of fault chains, and visualization of the reasoning process. It also supports maintenance personnel to manually correct the cause-effect graphs and input expert experience, thereby realizing human-machine collaborative cause-effect graph optimization.

[0090] In the above embodiments, firstly, a core feature set is obtained by fusing and filtering multimodal heterogeneous data, and then invariant risk minimization learning is performed on the core feature set to obtain an environment-invariant feature matrix, thus removing the influence of environmental disturbances on fault features. Based on mutual information, the strength of causal edges between the variable features in the environment-invariant feature matrix is ​​evaluated to obtain a causal edge strength matrix. A dual screening process is then performed using a causal edge strength threshold and conditional independence test to uncover a stable and reliable core fault causal framework, solving the problem of unstable causal relationship mining under complex operating conditions. Secondly, an initial hierarchical fault causal graph is constructed based on the core fault causal framework and a pre-defined hierarchical node system. The initial causal edge weights are determined by fusing the causal edge strength matrix, and the initial hierarchical fault causal graph is dynamically updated based on the meta-learning model and the initial causal edge weights to obtain a target hierarchical fault causal graph. This enables the causal graph to adaptively update to new environments or new faults, solving the problem of static causal graphs being difficult to update dynamically. Finally, after correcting the causal edge weights of the hierarchical fault causal graph using counterfactual samples, the inference path is optimized. Based on the optimized hierarchical fault causal graph, the causal relationships of faults in new energy ships are obtained through structured reasoning combined with implicit causal relationships mined by a large language model. This enhances the robustness of the causal graph to unknown scenarios and improves the accuracy and efficiency of fault reasoning. Therefore, this application achieves stable and adaptive fault causal relationship mining under complex dynamic conditions, significantly improving the generalization ability, dynamic adaptability, and diagnostic accuracy of fault diagnosis for new energy ships.

[0091] In one embodiment, the step of fusing multimodal heterogeneous data from new energy ships and then filtering it to obtain a core feature set includes:

[0092] Preprocessing of multimodal heterogeneous data from new energy ships yields sensor features, text features, and environmental features;

[0093] The obtained sensor features, text features, and environmental features are mapped to a unified feature space and fused through a cross-modal self-attention mechanism to obtain fused features;

[0094] A large language model fine-tuned for the shipbuilding domain is used to evaluate the fault correlation of each feature channel in the fused features, and the importance score of each feature channel is obtained.

[0095] Based on the importance scores of each feature channel, feature channels that are higher than a preset threshold are selected to form the core feature set.

[0096] Among them, sensor features refer to numerical time-series or statistical quantities such as voltage, current, temperature, and vibration extracted from raw sensor readings of equipment such as power batteries and propulsion motors; text features refer to semantic features extracted from unstructured text such as ship operation logs and maintenance records; and environmental features refer to external environmental parameters such as sea area, climate, and load obtained through positioning systems and environmental sensors. A unified feature space maps these three types of heterogeneous features to an aligned space with the same dimension through mathematical transformation, enabling cross-modal processing. The cross-modal self-attention mechanism is a neural network structure that dynamically calculates the degree of mutual attention between features of different modalities to capture deep-level interaction relationships, forming a fusion feature that integrates the three types of information. The large language model, fine-tuned for the shipbuilding domain, is a model with fault diagnosis understanding capabilities, optimized using specialized corpora for new energy ships based on a general model. It is used to assess the fault relevance of each independent dimension in the fusion features, outputting a quantified importance score, which reflects the contribution of the corresponding feature to the current diagnostic task. The preset threshold is a numerical limit set based on historical data distribution or expert experience, used to distinguish useful features from redundant or irrelevant features.

[0097] To address the heterogeneity of multimodal data related to faults in new energy ships, this study breaks down the barriers between cross-modal data formats by data acquisition, preprocessing, cross-modal embedding, and attention feature screening. It integrates feature information from time-series sensor data, textual data, and environmental data to screen out core features strongly correlated with the causal relationship of the faults, providing high-quality data support for subsequent causal relationship mining.

[0098] First, a three-level multimodal data acquisition system based on ship, shore, and cloud was designed to collect time-series sensor data such as voltage, current, temperature, and vibration from the power system, propulsion system, and energy storage system of new energy ships, with a sampling frequency of 1kHz. Simultaneously, textual fault records, maintenance logs, expert experience, and environmental data including sea area, temperature, humidity, salt spray, and load were collected and incorporated into ship industry standard data to construct a multimodal database. The preprocessing process is divided into two categories: For time-series sensor data, outliers are removed based on the 3σ criterion, with a removal rate not exceeding 3%. The KNN algorithm (k=5) is used to fill in missing values, and Min-Max normalization is used to unify the data magnitude to eliminate the influence of dimensions. For textual data, natural language processing techniques such as word segmentation, stop word removal, and entity recognition are used for preprocessing. A pre-trained word vector model, Word2Vec, is used to transform text features into low-dimensional dense vectors. For environmental data, classification, encoding, and normalization are performed to align features with other data.

[0099] Subsequently, a feature fusion model based on a cross-modal self-attention mechanism is constructed to map temporal sensor features, text features, and environmental features to a unified feature space, achieving deep fusion of multimodal features and addressing the problem of one-sided causal relationship mining caused by heterogeneous multimodal data formats and insufficient fusion. The fusion model learns the correlation between cross-modal features through the mutual information maximization criterion. The cross-modal feature mapping model is as follows:

[0100]

[0101]

[0102]

[0103] In the formula, The preprocessed time-series sensor feature matrix (dimension 1) ), The preprocessed text feature matrix (dimension 1) ), The preprocessed environmental feature matrix (dimension 1) ); , , These are the linear mapping weight matrices for the three types of features, with dimensions respectively. , , It is used to map features of different dimensions to a unified dimension. ; , , These are the linear mapping bias vectors for the three types of features, each with dimension [missing information]. , used to correct the feature offset after mapping; , , These are the mapped sensor feature, text feature, and environmental feature matrices, respectively, each with a dimension of [missing information]. N is the sample size. , , These represent the dimensions of sensor features, text features, and environmental features before mapping, respectively. To unify feature mapping dimensions. To address the technical problem of insufficient cross-modal feature correlation mining caused by inconsistent dimensions of multimodal data and the inability to directly fuse them.

[0104] Cross-modal self-attention weight calculation model:

[0105]

[0106]

[0107]

[0108]

[0109] In the formula, , , These are the query matrix, key matrix, and value matrix, each with a dimension of [missing information]. ; , , These are the query, key, and value weight matrices for the attention mechanism, each with a dimension of [missing information]. ; This is a scaling factor used to mitigate the vanishing gradient problem during attention weight calculation; This is the softmax activation function, used to normalize the attention weights to the [0,1] interval; The fusion feature matrix is ​​a cross-modal self-attention weighted matrix with dimension . This addresses the technical problem of neglecting the differences in importance between different modal features during multimodal feature fusion, resulting in weak targeted fused features.

[0110] Fusion loss function model:

[0111]

[0112] In the formula, Mutual information is used to quantify the degree of correlation between three modal features. Its value ranges from [0, +∞). The greater the mutual information, the stronger the correlation between the three modal features. , , The first The sensor features, text features, and environmental feature vectors of each sample, each with a dimension of [dimensionality missing]. N is the sample size; The fusion loss function value ranges from (-∞, 0]. The smaller the loss function, the more comprehensive the multimodal core information contained in the fused features. This addresses the technical problem that the lack of quantitative evaluation standards for the fusion effect during multimodal feature fusion, which leads to the fused features failing to fully retain the core information related to each modality and fault.

[0113] Finally, a finely tuned large model from the shipbuilding domain is introduced as a feature evaluator. This model is finely tuned based on a pre-trained LLaMA or BERT model to score the importance of the fused multimodal features. It also uses a channel attention mechanism to select core features that are strongly correlated with the causal relationship of the fault. This solves the technical problems of redundant information, high computational complexity, and weak targeting of fault-related features in the fused features. It reduces the feature dimension by more than 50% while maintaining the discriminativeness of fault features at no less than 95%.

[0114] Large model feature importance scoring model:

[0115]

[0116] In the formula, For the first multimodal feature after fusion Each feature channel ( ), dimension ; This is a fully connected layer for a finely tuned LLaMA large model in the shipbuilding domain, used to evaluate the importance of individual feature channels, with an output dimension of 1×1; The sigmoid activation function is used to normalize the feature importance scores to the [0,1] interval; For the large model to the first The importance score for each feature channel is assigned; the closer the value is to 1, the stronger the causal relationship between the feature channel and the fault; the closer the value is to 0, the more redundant the feature channel is. This addresses the technical problem that general large models lack knowledge of the ship domain, making it impossible to accurately assess the correlation between features and ship faults.

[0117] Channel attention feature weight correction model:

[0118]

[0119] In the formula, For the revised first The attention weights of each feature channel range from [0,1]. The larger the weight, the more important the feature channel is in the fault causal relationship mining. For the large model to the first Score the importance of each feature channel; The denominator is an exponential function used to amplify the differences in importance between different feature channels; This is a normalization term used to ensure that the sum of the attention weights for all feature channels is 1. It addresses the technical problem that relying solely on large model scoring may lead to biased feature importance assessments, resulting in the omission of core fault features.

[0120] Core Feature Filtering Model:

[0121]

[0122] In the formula, The core feature set after filtering has dimensions of 1. ( The core feature channel number, ); τ is the feature selection threshold, with a value of 0.7, which can be adaptively adjusted according to the accuracy requirements of ship fault diagnosis; For the first multimodal feature after fusion One feature channel; For the revised first Attention weights for each feature channel. This addresses the technical problem of high feature dimensionality and computational complexity leading to low efficiency in subsequent causal relationship mining.

[0123] It should be noted that preprocessing the multimodal heterogeneous data of new energy ships to obtain sensor features, text features, and environmental features is to transform the raw data into a structured representation that can be processed uniformly. Mapping the three types of features to a unified feature space and fusing them through a cross-modal self-attention mechanism is to integrate multi-source information to obtain more comprehensive fused features. Using a large language model fine-tuned for the shipbuilding domain to assess the fault correlation of each feature channel is to accurately identify key features related to faults. Selecting feature channels with importance scores above a preset threshold to form the core feature set is to reduce data dimensionality while retaining high-value information. Through multimodal fusion and professional screening, a clean core feature set rich in fault information is obtained, laying a high-quality data foundation for subsequent accurate mining of fault causal relationships.

[0124] In one embodiment, the step of performing invariant risk minimization learning on the core feature set to obtain the environment-invariant feature matrix includes:

[0125] The environment-invariant feature matrix is ​​obtained using the following formula:

[0126]

[0127] in, Represents the environment-invariant feature matrix. This represents a feature extraction network. For feature extraction network parameters, Represents the core feature set, This represents the activation function. This represents the weight matrix of the feature extraction network. The bias vector represents the weighted evidence; the feature extraction network is trained using an optimization objective of invariant risk minimization.

[0128]

[0129] in, This represents a subset of environments, obtained by dividing the ship's operating environment based on the core feature set. Represents a single subset of the environment. , These represent the feature data and fault labels of a single environmental subset, respectively. The loss function represents a single subset of environments. Indicates the parameters of the prediction head network. This represents the gradient of the loss function with respect to the parameters of the feature extraction network. This represents the gradient penalty coefficient.

[0130] Specifically, by dividing the core feature set into multiple environmental subsets according to different operating conditions of the ship, and introducing an optimization objective based on an invariant risk minimization algorithm for training, a stable fault feature representation across environments can be effectively learned. This optimization objective, while minimizing the prediction loss of each environmental subset, adds a penalty constraint on the gradient of the feature extraction network parameters, forcing the model to make the predicted performance of features extracted in different environments tend to be consistent with fault labels, thereby eliminating the perturbation of fault features by environmental factors such as sea state, climate, and load. Based on this, the core feature set is mapped through a feature extraction network containing a nonlinear activation function to obtain an environment-invariant feature matrix.

[0131] Therefore, the gradient penalty constraint enables the model to focus on learning the stable correlation between the fault itself and the label, eliminating the interference of environmental diversity on causal relationship mining and significantly improving the generalization ability of feature representation under different working conditions; the nonlinear feature extraction network can capture complex fault feature patterns, ensuring the integrity of information of environmentally invariant features; the finally obtained environmentally invariant feature matrix provides a clean and robust data foundation for subsequent mining of stable and reliable core fault causal skeletons, fundamentally solving the technical problems of unstable causal relationship mining and high cross-environment misdiagnosis rate caused by environmental changes.

[0132] In one embodiment, the step of evaluating the causal edge strength matrix based on mutual information to assess the causal edge strength between the variable features in the environment-invariant feature matrix includes:

[0133] The causal edge strength matrix is ​​obtained using the following formula:

[0134]

[0135]

[0136] in, Representing variable characteristics arrive The causal edge strength, i.e., the causal edge strength matrix, , The first element in the environment-invariant characteristic matrix The and the first Individual variable characteristics, Indicates the fault label variable. Indicates control Under unchanged conditions, and Conditional mutual information, used for quantization and The degree of direct correlation, Control in the display Under unchanged conditions, and Conditional mutual information, used for quantization and The degree of spurious association, This represents the preset spurious association correction coefficient, used to adjust the strength of spurious association removal. express , , The joint probability density function, Indicates in Given the conditions, and The joint conditional probability density, , They represent in Given the conditions, and The conditional probability density.

[0137] By introducing a causal edge strength evaluation function based on mutual information and conditional mutual information, this study aims to accurately quantify the degree of direct causal association between features of various variables in an environment-invariant feature matrix, while effectively eliminating spurious associations caused by confounding variables or data noise. Specifically, the first term, conditional mutual information, measures the strength of the direct association between a feature and the fault label while controlling for other variables, reflecting the true contribution of the feature to the fault. The second term, conditional mutual information, captures the residual association between two features after controlling for the fault label; this association usually originates from common causes or accidental correlations and is the main source of spurious associations. By introducing a correction coefficient to weighted subtract the second term, the final causal edge strength more purely reflects the direct causal effect between features.

[0138] Therefore, the dual measurement mechanism based on conditional mutual information enables the evaluation of causal edge strength to not only focus on the direct correlation between features and faults, but also effectively eliminate spurious correlations caused by data confusion, significantly improving the accuracy of causal relationship mining. The introduction of the correction coefficient provides adjustable flexibility for the elimination intensity, avoiding the risk of losing effective causal edges due to excessive elimination. The final causal edge strength matrix provides a precise quantitative basis for the subsequent construction of the core fault causal skeleton, fundamentally solving the technical problem of one-sided and distorted causal mining caused by false correlations after multimodal feature fusion.

[0139] In one embodiment, the steps of mining the core fault causal skeleton based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, include:

[0140] Using the features of each variable in the environment-invariant feature matrix as nodes, and the pairs of variable features in the causal edge strength matrix whose strength values ​​are greater than the causal edge strength threshold as candidate causal edges, an initial undirected graph is constructed.

[0141] For each edge connecting two nodes in the initial undirected graph, a candidate condition set is selected from the remaining nodes. The conditional independence test is used to determine whether the two nodes are conditionally independent under the conditions of the candidate condition set. If there is at least one candidate condition set that makes the two nodes conditionally independent, the edge is determined to be indirectly related and is removed from the initial undirected graph. If after traversing all candidate condition sets, the two nodes do not satisfy the conditional independence, the edge is retained as a valid causal edge.

[0142] By determining the direction of valid causal edges and identifying the causal orientation between variable features, a directed core fault causal skeleton is obtained.

[0143] Here, the initial undirected graph refers to a graph structure constructed using the variable features in the environmental invariant feature matrix as nodes and the variable feature pairs in the causal edge strength matrix whose strength values ​​exceed a certain threshold as edges, without considering causal orientation for the time being; the candidate condition set refers to the combination of variable features selected from the remaining nodes in the initial undirected graph except for the two nodes to be tested; indirect association refers to the association between two variables that do not have a direct causal effect but are transmitted through other variables; effective causal edge refers to the edge that is retained after the conditional independence test and is considered to have a direct causal association; direction determination refers to the process of determining the causal orientation in the effective causal edge, that is, clarifying which variable is the cause and which variable is the effect.

[0144] After obtaining the environment-invariant feature matrix and the causal edge strength matrix, each variable feature in the environment-invariant feature matrix is ​​first used as a node. All off-diagonal elements in the causal edge strength matrix are traversed, and variable feature pairs with values ​​greater than or equal to a preset causal edge strength threshold are considered to have potential causal relationships and added as undirected edges, thus constructing an initial undirected graph. Furthermore, to eliminate potential indirect relationships, for each undirected edge connecting two nodes in the initial undirected graph, different subsets are selected from the remaining nodes as candidate condition sets for conditional independence testing. If any candidate condition set makes the two nodes conditionally independent, the edge is determined to be a false association and removed; if conditional independence is not satisfied after traversing all candidate condition sets, the edge is retained as a valid causal edge. Furthermore, the direction of the remaining valid undirected edges is determined. Based on the temporal relationship between variable features and the asymmetric calculation based on mutual information, the causal direction of each edge is determined, and finally the core fault causal skeleton in the form of a directed acyclic graph is obtained. This skeleton clearly reveals the core causal transmission relationship between each fault feature with a clear direction, laying a precise topological foundation for the subsequent construction of a hierarchical fault causal graph.

[0145] In one example, the core features obtained after environment invariance learning and feature selection are used as input, and an improved PC algorithm is used for causal discovery. This algorithm adds a causal edge strength threshold constraint to the traditional PC algorithm, which can effectively mine the causal relationship between variables and remove redundant causal edges. It solves the technical problem that the traditional PC algorithm is prone to retaining redundant edges or missing effective edges due to not considering the strength of causal edges, resulting in inaccurate causal skeleton structure. Finally, the core causal skeleton of new energy ship failure is obtained, and the structural accuracy of the causal skeleton is not less than 90%.

[0146] Improved PC algorithm causal edge selection model:

[0147]

[0148] In the formula, The core fault causal framework is composed of a set of effective causal edges. For variables and The causal edge between them; For variables arrive The causal edge strength; This is the threshold for causal edge strength, used to filter valid causal edges; Indicates in the control condition set ( To remove , If other variables (excluding subsets) remain unchanged, and Conditional independence; "Not valid" means and There is a direct causal relationship, rather than an indirect relationship through other variables. This addresses the technical problem of traditional PC algorithms that only select causal edges based on conditional independence without considering the strength of causal edges, resulting in redundant and invalid edges in the causal skeleton and low structural accuracy.

[0149] Conditional independence test model:

[0150]

[0151] In the formula, This is the conditional independence test statistic; the larger the value, the stronger the conditional independence test statistic. and In control If conditions remain unchanged, the stronger the correlation, the less independent the relationship. The number of samples; For condition set The number of variables; for , In control Empirical joint probability density under the invariant condition; , They are respectively , In control Empirical conditional probability density under the condition of invariance; It is a function of the natural logarithm. When ( significance level Degrees of freedom are When the chi-square distribution critical value is reached, the determination is made. Not valid, that is and A direct causal relationship exists. This addresses the technical problem of inaccurate causal edge screening due to the lack of quantitative standards in conditional independence tests.

[0152] It should be noted that using the features of each variable in the environment-invariant feature matrix as nodes and constructing an initial undirected graph based on features greater than a threshold in the causal edge strength matrix is ​​to quickly filter weak associations based on quantized strength, thereby reducing subsequent computational complexity. Furthermore, performing a conditional independence test on each edge to eliminate indirect associations transmitted by other variables ensures that the remaining edges represent genuine direct causal effects. Finally, determining the direction of valid causal edges to clarify the causal orientation is to enable the causal graph to reveal the directionality of fault propagation.

[0153] By using threshold filtering, conditional independence testing, and direction determination, false and indirect associations were effectively eliminated, and the causal flow was clarified. The resulting core fault causal framework has high accuracy, reliability, and interpretability, laying a precise topological foundation for the subsequent construction of a hierarchical fault causal graph.

[0154] In one embodiment, the step of fusing the causal edge strength matrix to determine the initial causal edge weights includes:

[0155] The initial causal edge weights are determined using the following formula:

[0156]

[0157] in, Represents nodes in the initial hierarchical fault cause graph To the node The initial causal edge weights, Represents the nodes obtained from the causal edge strength matrix. To the node The causal edge strength, This represents the set of all causal edges in the initial hierarchical fault causality graph. This represents the maximum causal edge strength among all causal edges in the initial hierarchical fault causal graph. Represents a node To the node Experience rating of causal relationships.

[0158] Specifically, the initial weights of each causal edge in the hierarchical fault cause-effect graph are determined through a weighted fusion approach. This design balances the strength of causal relationships driven by data with the empirical judgment of expert knowledge. The first part of the formula divides the corresponding strength value in the causal edge strength matrix by the maximum value of all causal edge strengths, normalizing the data-driven information and enabling comparison and fusion of causal strengths of different dimensions on a unified scale. The second part incorporates empirical scores from shipbuilding experts on causal relationships, integrating long-accumulated industry knowledge into the weight calculation. The two parts are linearly combined with weight coefficients of 0.7 and 0.3, maintaining the dominance of data-driven approaches to ensure objectivity while incorporating expert experience to correct for potential biases in purely data-driven methods.

[0159] It should be noted that maximum value normalization eliminates the impact of differences in causal intensity dimensions on weight fusion, making the weights of different edges comparable. The introduction of empirical scoring compensates for sample bias or statistical randomness that may be caused by relying on single data, improving the robustness and scenario adaptability of the weights. The reasonable allocation of weight coefficients achieves the complementary advantages of data objectivity and empirical rationality. The final initial causal edge weights not only accurately reflect the causal effect strength contained in historical data, but also fit the physical logic and operation and maintenance experience in actual engineering scenarios, providing a reliable foundation for subsequent dynamic updates based on meta-learning models.

[0160] In one embodiment, the step of dynamically updating the initial hierarchical fault causality graph based on the meta-learning model and initial causal edge weights to obtain the target hierarchical fault causality graph includes:

[0161] By pre-training a meta-learning model using an update task set containing various environmental changes and fault types, general update parameters applicable to different update scenarios can be obtained.

[0162] When the current scenario of a new energy ship is detected to be a new environment or a new fault, a small number of current samples are obtained, and the initial causal edge weights of the initial hierarchical fault causal graph are fine-tuned based on the general update parameters to obtain the updated causal edge weights. Based on a small number of current samples, the causal edge strength under the current scenario is recalculated to obtain the updated causal edge strength matrix.

[0163] In the initial hierarchical fault cause graph, invalid causal edges with weights lower than a preset removal threshold are removed based on the updated causal edge weights, and new causal edges with causal edge strengths higher than a preset addition threshold are added based on the updated causal edge strength matrix, so as to obtain the target hierarchical fault cause graph.

[0164] The update task set refers to a pre-collected set of historical operational data covering different environments and fault types, with each task corresponding to a specific combination of environment and fault. The general update parameters are model initialization parameters obtained after training the meta-learning model on the update task set, applicable to different update scenarios. The current scenario refers to the real-time operating condition of the new energy vessel, including newly emerging environments or faults. A small number of current samples are a small amount of real-time operational data collected from the current scenario for rapid adaptation. The preset removal threshold is a set lower limit for weights, used to determine whether causal edges are effective in the current scenario; the preset addition threshold is a set upper limit for the strength of causal edges, used to identify newly emerging strong causal relationships. Invalid causal edges refer to causal edges whose weights are lower than the removal threshold after the update, indicating that their causal effect has weakened or disappeared; new causal edges refer to causal edges whose strength is higher than the addition threshold after the update but do not exist in the original graph, indicating the emergence of new causal relationships.

[0165] Fault cases covering different environmental conditions, such as tropical high-temperature sea areas, polar low-temperature sea areas, and inland waterway light-load conditions, as well as different fault types, such as thermal runaway of power batteries, wear of propulsion motor bearings, and overcurrent of power converters, are extracted from the historical operation database of new energy ships in advance. These cases are divided into multiple independent update tasks according to the combination of environment and fault. Each task contains corresponding input features and fault labels. A graph neural network based on a model-independent meta-learning algorithm is constructed as a meta-learning model. This model receives a hierarchical fault causal graph and the current sample and outputs the updated causal edge weights. During training, a small number of inner loop updates are performed on each task, and the model is optimized through outer loop optimization to achieve the best overall performance across multiple tasks. After training, the initialization parameters of the model are saved as general update parameters. When a new environment or new fault is detected in the current scenario of a new energy vessel, a small amount of real-time operational data is collected from the current scenario. After data preprocessing and feature extraction, a small number of current samples are obtained. The initial hierarchical fault causality graph and these small number of current samples are input into a meta-learning model loaded with general update parameters for fine-tuning, quickly obtaining the updated causal edge weights. At the same time, based on the small number of current samples, the mutual information or conditional mutual information between variable features is recalculated to obtain the updated causal edge strength matrix. Finally, a preset removal threshold is read to traverse each causal edge in the graph, and invalid causal edges with updated weights lower than the threshold are removed. A preset addition threshold is read to traverse the updated causal edge strength matrix, and variable feature pairs with strengths higher than the threshold and not present in the initial graph are added as new causal edges and assigned initial weights. After invalid edge removal and new edge addition, a target hierarchical fault causality graph that accurately reflects the real fault propagation mechanism in the current new scenario is obtained.

[0166] In one example, a model-agnostic meta-learning algorithm (MAML) is used to construct an incremental update framework for causal graphs. This framework treats causal graph update tasks under different environments and operating states as meta-tasks, learning general meta-update parameters. This addresses the technical problem of traditional static causal graphs being unable to adapt to changes in system state, new environments, or new faults, resulting in delayed updates and the need for extensive retraining. When new environments, faults, or operating states arise, rapid incremental updates of the causal graph can be achieved with only a small number of samples.

[0167] Meta-learning and meta-update parameter training model:

[0168]

[0169] In the formula, The meta-update parameters are universal parameters applicable to all causal graph update tasks, with dimensions of [missing information]. ( (This represents the total number of parameters in the causal graph, including the weights of all causal edges). These are the initial parameters for meta-learning; The task set for updating the cause-effect graph includes three types of update tasks: new environment adaptation, new fault identification, and changes in operating state, totaling 10 categories. For a single update task; For the task The mean squared error loss function is used to quantify the inference error of the updated causal graph, and its value range is [0,+∞). To adapt to the task The output of the causal graph reasoning model, i.e., the fault diagnosis result, has a dimension of ( For the task (number of samples); For the task The true fault labels, in dimensions of ; To adapt to the task The fine-tuning parameters are updated by the meta-update parameters. After fine-tuning, the updated formula is: ,in The learning rate for the inner loop is used to control the fine-tuning amplitude. This addresses the technical challenges of requiring a large number of samples, long training periods, and delayed updates when new environments or faults occur, by using meta-learning to achieve rapid updates with a small number of samples.

[0170] Causal edge weight incremental update model:

[0171]

[0172] In the formula, For the node after incremental update To the node Causal edge weights; The weights of the causal edges before the update; The outer loop learning rate is used to control the magnitude of weight updates and prevent weight oscillations caused by excessively rapid updates. For the loss function of task T, fine-tuning parameters The gradient; Output the weights of the causal edges before the update for the causal graph inference model. The partial derivatives are used to quantify the impact of weight changes on the inference results. This addresses the technical problem of decreased inference accuracy in causal graphs due to a lack of quantitative basis for updating causal edge weights.

[0173] It's important to note that pre-training the meta-learning model with multiple update tasks to obtain general update parameters enables the model to quickly adapt to new scenarios based on historical experience. When a new environment or fault is detected, the causal edge weights are fine-tuned based on the general update parameters and a small number of current samples, while the causal edge strength matrix is ​​recalculated. This is to achieve low-cost model transfer to the new scenario. Furthermore, invalid edges are removed based on the updated weights, and new edges are added based on the updated strength matrix to ensure that the causal graph adapts to the new scenario in terms of topology. Thus, rapid updates with a small number of samples are achieved through meta-learning pre-training, and the combination of weight fine-tuning and structural updates ensures that the target causal graph accurately adapts to the current scenario, effectively solving the technical problems of lagging updates and large sample requirements of traditional static models.

[0174] In one embodiment, the steps of removing invalid causal edges with weights lower than a preset removal threshold based on the updated causal edge weights, and adding new causal edges with strengths higher than a preset addition threshold based on the updated causal edge strength matrix, include:

[0175] The following formula is used to remove invalid causal changes and add new causal edges:

[0176]

[0177] in, This represents the set of causal edges in the updated hierarchical fault cause-effect graph. This represents the set of causal edges before the update. Let represent any two nodes in the initial hierarchical fault cause-effect graph. Represents a node To the node Updated causal edge weights This indicates the preset rejection threshold. Represents a node To the node The updated causal edge strength, i.e., the updated causal edge strength matrix. This indicates a preset threshold for adding values.

[0178] Specifically, this formula updates the structure of the hierarchical fault causal graph through set operations. Its core lies in simultaneously removing invalid causal edges and adding new causal edges. The removal operation compares the updated causal edge weights with a preset removal threshold, removing edges with excessively low weights from the original edge set. This signifies that associations whose causal effects have significantly weakened or disappeared in the current new scenario are promptly cleaned up. The addition operation compares the updated causal edge strength with a preset addition threshold, adding edges with sufficiently high strength that are not present in the original graph to a new edge set. This indicates that newly emerging strong causal associations in the current new scenario are successfully captured and incorporated into the graph structure.

[0179] Thus, the dynamic evolution of the causal graph topology was achieved through a dual threshold mechanism. The elimination operation removed outdated or invalid causal relationships, avoiding the interference of redundant information on subsequent reasoning. The addition operation incorporated key causal paths in the new scenario, ensuring that the causal graph fully covers the current fault mechanism. The resulting hierarchical fault causal graph is structurally highly matched to the actual scenario, effectively solving the technical problem of distortion of the original causal graph structure due to changes in environment or fault, and significantly improving the timeliness and accuracy of the causal graph.

[0180] In one embodiment, the step of correcting the causal edge weights of a target hierarchical fault causal graph using counterfactual samples includes:

[0181] Using a large language model, based on real fault samples and a hierarchical fault cause-effect graph, counterfactual samples that conform to causal logic are generated.

[0182] The counterfactual samples are input into the target hierarchical fault cause-effect graph for inference to obtain the inference loss corresponding to the counterfactual samples, and the real fault samples are input into the target hierarchical fault cause-effect graph for inference to obtain the inference loss corresponding to the real samples.

[0183] Based on the difference between the inference loss of counterfactual samples and the inference loss of true samples, the weight correction of each causal edge in the hierarchical fault causal graph of the target is calculated.

[0184] Based on the weight correction amount, the weights of the causal edges of each causal edge in the hierarchical fault causal graph of the target are corrected to obtain the corrected hierarchical fault causal graph of the target.

[0185] Among them, inference loss refers to the quantitative difference between the inference result obtained after structured inference by inputting the sample into the hierarchical fault causal graph and the real fault label; real fault sample refers to sample data with real fault labels collected from the actual operation of new energy ships, which are formed after preprocessing and feature extraction; weight correction amount refers to the value used to adjust the weight of causal edge, which is calculated based on the difference between the inference loss of counterfactual samples and the inference loss of real samples.

[0186] Specifically, firstly, real fault samples are selected from the historical fault database of new energy ships, and then compared with the target hierarchical fault causality. Figure 1 A large language model, fine-tuned for the marine domain, is used as input, and prompts are designed to guide the model in counterfactual reasoning. For example, different values ​​for key variables are assumed to predict the failure evolution, thereby generating counterfactual samples that conform to causal logic. Counterfactual samples and real failure samples are input into the target hierarchical failure causal graph for structured reasoning, obtaining their respective reasoning results and comparing them with real failure labels. The reasoning loss of the counterfactual samples and the reasoning loss of the real samples are calculated. By comparing the difference in reasoning losses between the two types of samples, for each causal edge, the difference between the gradient of the counterfactual loss relative to the edge weight and the gradient of the real loss relative to the edge weight is calculated as the correction direction, combined with a preset learning rate to obtain the weight correction amount. Finally, each directed edge in the graph is traversed, and its current weight is updated based on the weight correction amount, resulting in the corrected target hierarchical failure causal graph. After calibration with counterfactual samples, the weights of this graph are more consistent with actual physical laws and cross-scenario robustness requirements, providing a more reliable foundation for subsequent failure diagnosis.

[0187] For example, by leveraging the implicit relationship mining capabilities of large models, counterfactual fault samples are generated based on existing fault samples to supplement the fault sample set, thus solving the technical problems of scarce rare fault samples, poor robustness of causal graphs, and low accuracy of rare fault inference. At the same time, the counterfactual samples are input into the causal graph for inference verification, the causal edge weights are corrected, the robustness of the causal graph is improved, the correlation coefficient between the generated samples and the real samples is ≥0.92, and the causal inference accuracy of rare faults is ≥85%.

[0188] Large-scale counterfactual sample generation model:

[0189]

[0190] In the formula, The generated counterfactual fault sample matrix has dimensions of ( (Number of counterfactual samples); A fine-tuned LLaMA model for the shipbuilding field, used to generate counterfactual samples based on real samples; This is a matrix of real fault samples, with dimensions... for( (the actual number of samples); This is a counterfactual intervention parameter, with a value range of [0.1, 0.3], used to control the degree of difference between the counterfactual sample and the true sample. If the sample size is too small, the counterfactual sample lacks diversity; if the sample size is too large, it deviates too much from the true characteristics of the fault. This is a hierarchical fault cause-effect graph used to constrain the causal logic of large models generating counterfactual samples, ensuring that the generated samples conform to the fault causal relationship and avoiding the generation of invalid samples that do not conform to actual fault patterns. It addresses the technical problem of insufficient training, poor robustness, and low accuracy of rare fault inference caused by the scarcity of rare fault samples.

[0191] Counterfactual sample validity test model:

[0192]

[0193] In the formula, The correlation coefficient between the counterfactual sample and the true sample ranges from [-1, 1]. When a counterfactual sample is deemed valid, it shall be retained. The covariance between the counterfactual sample and the true sample is used to quantify the degree of linear correlation between the two. , These are the variances of the counterfactual samples and the true samples, used to quantify their dispersion. This addresses the technical problem of the lack of quantitative verification standards for the validity of generated counterfactual samples, which leads to invalid samples interfering with the robustness of the causal graph.

[0194] Causal edge weight correction model (based on counterfactual sample verification):

[0195]

[0196] In the formula, The causal edge weights are corrected after verification using counterfactual samples; The weights of the causal edges are updated after the meta-learning increment; γ=0.05 is a correction coefficient used to control the correction magnitude. After inputting the counterfactual sample into the causal graph, the nodes To the node Inference loss; After inputting the causal graph with real samples, the nodes To the node Inference loss; The inference loss bias rate quantifies the degree of deviation between the inference results of counterfactual samples and true samples. The smaller the bias rate, the more accurate the causal edge weights are, and the greater the correction magnitude. This addresses the technical problem of insufficient validation of causal edge weights after incremental updates, which may lead to biases and insufficient robustness of causal graphs.

[0197] It should be noted that generating counterfactual samples based on real fault samples and a hierarchical fault causal graph using a large language model is intended to obtain causal verification information beyond the observed data by simulating the fault evolution path under the intervention of key variables. Inputting the counterfactual samples and real fault samples into the causal graph for inference and comparing the difference in inference loss is to quantify the rationality of the current causal graph's response under hypothetical intervention, thereby identifying edges where the weight settings do not match the actual causal effect. Calculating the weight correction amount for each causal edge based on the loss difference and updating the weights accordingly ensures that the corrected causal graph not only performs well on observed data but also maintains inference accuracy in counterfactual scenarios. The beneficial effects of this approach are that counterfactual verification effectively calibrates the causal edge weights, eliminates biases caused by data randomness, significantly improves the robustness and generalization ability of the causal graph weights, and ultimately results in a corrected hierarchical fault causal graph that more realistically and reliably reflects the causal mechanism of new energy ship faults.

[0198] In one embodiment, the step of inference path optimization includes:

[0199] The modified hierarchical fault cause graph of the target is transformed into a Markov decision process model, in which the nodes in the modified hierarchical fault cause graph of the target are used as the state space, the causal edges between nodes are used as the action space, and the weights of the causal edges in the modified hierarchical fault cause graph of the target are used as the state transition probabilities.

[0200] Based on the Markov decision process model, a reinforcement learning algorithm is used for iterative training to update the Q-values ​​of each state-action pair until convergence to obtain the optimal Q-value table. The reinforcement learning algorithm is trained based on a preset reward function, which is positively correlated with inference accuracy and negatively correlated with inference path length.

[0201] Based on the optimal Q-value table, the causal edge with the largest Q-value is selected as the corresponding optimal action at each node to form the optimal reasoning path from the fault feature node to the fault root cause node.

[0202] The Markov Decision Process (MDP) model is a sequential decision-making framework used to describe how an agent changes its state and obtains rewards through actions. The state space here corresponds to all nodes in the modified hierarchical fault causal graph. The action space corresponds to all directed causal edges originating from the current node in the graph. The state transition probability uses the weights of the causal edges as the probability of transitioning to the next state after performing an action. Reinforcement learning algorithms are methods by which an agent learns optimal decision-making strategies through trial and error with its environment. The Q-value is an estimate of the expected cumulative reward obtained by taking a certain action in a given state. The optimal Q-value table is a table containing the optimal Q-values ​​for each state-action pair, obtained after iterative convergence of reinforcement learning. The optimal inference path is a sequence of paths starting from the fault feature node and progressing step-by-step along the edge with the maximum Q-value at each node until reaching the root cause node of the fault.

[0203] Specifically, after revising the hierarchical fault causal graph, it is transformed into a Markov decision process model for inference path optimization. Each node in the graph is defined as a state, forming the state space; each directed causal edge is defined as an action that can be taken in that state, forming the action space; and the weight of the causal edge is defined as the state transition probability, ensuring a one-to-one correspondence between the state, action, and transition probability and the structure and weight of the causal graph. Based on this model, reinforcement learning algorithms such as Q-learning are used for iterative training. A Q-value table is initialized, and a reward function that is positively correlated with inference accuracy and negatively correlated with inference path length is set. By continuously selecting actions from the starting node, obtaining rewards, and updating Q-values ​​until convergence, an optimal Q-value table containing the optimal estimate for each state-action pair is obtained. Finally, based on the optimal Q-value table, the causal edge with the largest Q-value at each node is selected as the optimal action, thus forming the optimal inference path from any fault feature node to the fault root cause node. These paths minimize path length while ensuring inference accuracy, significantly improving the efficiency and interpretability of fault diagnosis.

[0204] For example, the fault reasoning process of the causal graph is transformed into a Markov decision process (MDP). The reasoning path of "fault feature - fault node - fault root cause" is used as the action space, and the reasoning accuracy and reasoning efficiency are used as reward functions. The Q-learning algorithm is used to optimize the reasoning path, which solves the technical problems of fixed reasoning path, low efficiency of tracing fault chains with multiple causes and effects and long fault root cause location time in traditional causal graphs. The reasoning weight of the causal graph is dynamically adjusted to achieve rapid tracing of fault chains with multiple causes and effects. The optimized reasoning path length is shortened by more than 40%, and the fault root cause location time is ≤10s.

[0205] Markov Decision Process (MDP) Modeling:

[0206]

[0207] In the formula, The state space is the set of all nodes in the causal graph. Each state Each corresponds to a cause-effect graph node (feature node, fault node, system node, or environment node). The action space is the set of all causal edges in the causal graph. Every action Corresponding to a causal edge , indicating from state Transition to state ; Let be the state transition probability function. Indicates from state Execution After transitioning to state The probability is equal to the causal edge weight. ; The reward function quantifies the benefit of taking an action; a higher reward indicates that the action is more helpful in quickly locating the root cause of the failure. γ=0.95 is a discount factor used to balance current rewards with long-term future rewards; a value closer to 1 indicates a greater emphasis on long-term future rewards. This addresses the technical problems of lacking systematic modeling in the fault reasoning process and the absence of clear objectives for reasoning path optimization.

[0208] Reward function model:

[0209]

[0210] In the formula, From state Execution The reward value is in the range of [0,1]; 0.7 and 0.3 are the weighting coefficients for reasoning accuracy and reasoning efficiency, respectively, with priority given to ensuring reasoning accuracy; From state Execution The inference accuracy, i.e. the degree of matching between the inference result and the actual root cause of the fault, has a value range of [0,1]. From state Execution The length of the resulting reasoning path is determined by the length of the reasoning path. The shorter the reasoning path, the larger the value of this item, and the higher the corresponding reward value, thereby optimizing reasoning efficiency.

[0211] It should be noted that transforming the modified hierarchical fault causal graph into a Markov decision process model, with nodes as the state space, causal edges as the action space, and causal edge weights as state transition probabilities, is to formalize the fault reasoning problem into a sequential decision problem, providing a mathematical foundation for applying reinforcement learning to solve for the optimal path. Based on this model, reinforcement learning algorithms are used for iterative training, and a reward function positively correlated with inference accuracy and negatively correlated with inference path length is designed to guide the algorithm to automatically balance diagnostic accuracy and inference efficiency during the exploration process, thereby learning an accurate yet concise inference strategy. According to the converged optimal Q-value table, the causal edge with the largest Q-value at each node is selected as the optimal action, ultimately forming the optimal inference path from the fault feature node to the fault root cause node. The beneficial effect of this implementation is that global optimal selection of the inference path is achieved through autonomous optimization by reinforcement learning, minimizing path length while ensuring diagnostic accuracy, effectively improving the efficiency and interpretability of fault diagnosis.

[0212] To facilitate understanding of the scheme in this application, specific examples are provided below.

[0213] To address the pain points in fault diagnosis of new energy ships, such as heterogeneous multimodal data, poor environmental adaptability, static causal reasoning, and insufficient interpretability, this embodiment proposes a complete technical solution covering data fusion, causal mining, dynamic updating, reasoning optimization, and visualization. The entire implementation process consists of five main steps and a total of 190 days of practical operation.

[0214] In the first step of multimodal heterogeneous data fusion and feature enhancement, a raw multimodal dataset covering three operating conditions—inland waterways, coastal areas, and offshore areas—was first constructed. For the collected sensor time-series data, outliers were removed using the 3σ criterion, with the outlier rate controlled between 2.2% and 2.9%. The KNN algorithm was used to fill in missing values, ensuring data integrity of at least 99.7%. Min-Max normalization was applied to map all data to the 0-1 range to eliminate the influence of unit dimensions. For textual fault records and maintenance logs, preprocessing was performed using natural language processing techniques such as word segmentation, stop word removal, and entity recognition. The Word2Vec model was then used to transform them into 256-dimensional text feature vectors. Environmental data underwent classification, encoding, and normalization. After standardized preprocessing, sensor features, text features, and environmental features were mapped to a unified 512-dimensional feature space through a linear layer, constructing a cross-modal self-attention fusion model. This model achieves deep fusion of multimodal information by calculating attention weights between features of different modalities. A loss function is designed with maximizing mutual information as the optimization objective. After 300 rounds of training, the fused multimodal features are obtained. Subsequently, the fused features are input into a large-scale LLaMA model fine-tuned with shipbuilding fault data. The large model scores the causal correlation of each feature dimension, and uses a channel attention mechanism to weight and correct the scores, ultimately selecting core features with an importance score of at least 0.7. The original fused features totaled 512 dimensions; after screening, 248 core features were obtained, reducing the dimensionality by 51.6% while maintaining the fault feature discriminative power at 96.3%. This effectively eliminated redundant information and provided high-quality input for subsequent causal mining.

[0215] The second step, cross-environment invariant causal skeleton mining, aims to eliminate the interference of different navigation environments on causal relationship mining. The dataset is further subdivided into six environmental subsets according to three operating conditions: inland waterway, coastal, and open sea. A feature extraction network consisting of a three-layer MLP and a prediction head consisting of a single fully connected layer are constructed. An invariant risk minimization algorithm is introduced, incorporating gradient penalty constraints into the optimization objective. This forces the prediction head to maintain consistent gradients to the feature extraction network under different environments, thereby learning environment-independent feature representations. After 200 rounds of iterative training, an environment-invariant feature matrix is ​​obtained. Based on this, a causal edge strength evaluation function is designed based on mutual information and conditional mutual information. By calculating the degree of correlation between features and fault labels under the control of other variables and subtracting spurious correlations between features, the quantified causal edge strength is obtained, and 0.6 is set as the threshold for determining potentially valid causal edges. Further, an improved PC algorithm was used to discover the causal skeleton. An initial undirected graph was constructed using environmental invariant features as nodes and potential effective causal edges. Indirectly related edges were gradually eliminated through conditional independence tests. At the same time, the causal edge strength threshold constraint was combined to ensure that the retained edges had sufficient strength. Finally, a core fault causal skeleton containing 128 effective causal edges was discovered, with a structural accuracy of 92.5%.

[0216] The third step, hierarchical dynamic causal graph construction and meta-learning incremental update, transforms the core causal framework into a structured hierarchical fault causal graph. This graph comprises four levels: environment, system, fault, and feature. The environment layer includes six nodes (e.g., ocean area, temperature and humidity); the system layer covers nine nodes (e.g., electricity, power, and energy storage); the fault layer includes 15 nodes (three typical faults and sub-faults); and the feature layer consists of 248 selected core feature nodes. Edges in the core causal framework are mapped to nodes at the corresponding levels, and initial causal edge weights are determined by fusing causal edge strength with expert experience in a 7:3 ratio. To enable rapid adaptation of the causal graph to new environments and faults, a model-independent meta-learning algorithm is used to construct an incremental update framework. Ten types of update tasks covering new environments, faults, and operating states are pre-collected as a meta-task set. After 150 rounds of meta-training, general update parameters are obtained. When a new scenario appears, only eight sets of fine-tuning samples are needed to quickly adapt to the general update parameters, adjusting causal edge weights, adding new edges, and removing invalid edges. Experiments show that the cross-environment update accuracy of cause-effect graphs decreases by only 4.1%, and the update accuracy under new faults can reach 91.8%, effectively solving the diagnostic lag problem of static cause-effect graphs.

[0217] The fourth step, large-scale collaborative causal reasoning and reinforcement learning path optimization, aims to improve reasoning efficiency and the accuracy of complex fault diagnosis. First, a finely tuned LLaMA model, adapted for the shipbuilding domain, is used to generate counterfactual samples. By setting counterfactual reasoning prompts, the fault evolution path under key variable intervention is simulated, generating 2000 sets of counterfactual samples. After removing invalid samples with a correlation coefficient lower than 0.92 with the real samples, the valid samples are input into the causal graph for reasoning verification. Based on the reasoning results, the causal edge weights are adjusted, enhancing the robustness of the causal graph to rare faults, improving the rare fault reasoning accuracy to over 85%. Subsequently, the adjusted causal graph is transformed into a Markov decision process model. Using nodes in the graph as the state space, causal edges as the action space, and causal edge weights as state transition probabilities, a reward function is designed to balance reasoning accuracy and efficiency, with an accuracy weight of 0.7 and an efficiency weight of 0.3. A Q-learning algorithm is used for 500 rounds of iterative training to obtain an optimal Q-value table. Then, based on the optimal Q-value, the optimal action is selected at each node, forming the optimal reasoning path from the fault feature node to the fault root cause node. The optimized reasoning path length was shortened by 42.3%, and the root cause location time was reduced to 8.5 seconds. In complex fault reasoning, fault features are input into the causal graph for explicit reasoning to obtain preliminary results. The preliminary results and features are then input into a large model to mine implicit causal relationships. The final diagnostic conclusion is obtained by fusing the two. The diagnostic accuracy of three typical complex faults reached 95.8%, which is 6.2 percentage points higher than that of single causal graph reasoning.

[0218] The fifth step, enhancing interpretability and visualization, focuses on improving the understandability of diagnostic results and human-machine collaboration capabilities. Based on Do-Calculus, the reasoning process of the causal graph is analyzed through intervention, quantifying the changes in the probability of failure after intervention at each feature node, clarifying the contribution of each node to the root cause of the failure, and forming explicit causal logical relationships. The causal logic analysis results, including the root cause of the failure, contribution, and failure chain, are input into a finely tuned large model. Natural language generation prompts are set, and the model automatically generates a standardized natural language explanation report containing the root cause of the failure, the failure development process, the scope of the failure's impact, and preliminary maintenance recommendations. Simultaneously, a web-based ship-shore cloud collaborative visualization platform is developed to achieve dynamic display of hierarchical causal graphs, real-time tracing of failure chains, and visual replay of the reasoning process. It also supports manual correction of the causal graph by maintenance personnel and the input of expert experience, forming a human-machine collaborative causal graph optimization mechanism.

[0219] Finally, twelve new energy vessels—four inland waterway pure battery ships, four coastal fuel cell ships, and four offshore hybrid ships—were selected for real-world testing and verification. Test results showed a 38.2% reduction in cross-environmental fault misdiagnosis rate, a 92.5% accuracy rate for causal framework structure, a 95.8% accuracy rate for complex fault diagnosis, a 42.3% reduction in reasoning path length, a fault root cause location time of 8.5 seconds, a causal logic interpretability score of 0.91, and a 4.1% decrease in the accuracy rate of causal graph cross-environment updates. All core indicators met the preset requirements. In actual ship operation, the ship failure rate decreased by 31%, and maintenance costs decreased by 25.6%, fully verifying the engineering practicality and effectiveness of this technical solution.

[0220] The following describes the apparatus for constructing causal relationships for new energy ship failures provided in the embodiments of this application. The apparatus for constructing causal relationships for new energy ship failures described below can be referred to in correspondence with the method for constructing causal relationships for new energy ship failures described above. Figure 2 As shown, this application provides a device for constructing causal relationships for failures in new energy ships. The device includes:

[0221] The core feature set determination module 201 is used to obtain the core feature set by fusing multimodal heterogeneous data of new energy ships when a new energy ship malfunctions.

[0222] The core fault causal skeleton mining module 202 is used to perform invariant risk minimization learning on the core feature set to obtain the environment invariant feature matrix. Based on mutual information, it evaluates the causal edge strength between the variable features in the environment invariant feature matrix to obtain the causal edge strength matrix. Based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, the core fault causal skeleton is mined.

[0223] The target hierarchical fault cause-effect graph determination module 203 is used to construct an initial hierarchical fault cause-effect graph based on the core fault cause-effect skeleton and the preset hierarchical node system, fuse the causal edge strength matrix, determine the initial causal edge weights, and dynamically update the initial hierarchical fault cause-effect graph based on the meta-learning model and the initial causal edge weights to obtain the target hierarchical fault cause-effect graph.

[0224] The fault causality determination module 204 is used to optimize the reasoning path after correcting the causal edge weights of the target hierarchical fault causality graph using counterfactual samples. Based on the optimized target hierarchical fault causality graph, the fault causality of new energy ships is obtained through structured reasoning and the implicit causal relationships mined by the large language model.

[0225] In one embodiment, this application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the new energy ship fault causation construction method as described in any of the above embodiments.

[0226] In one embodiment, this application also provides a computer device storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the new energy ship fault causality construction method as described in any of the above embodiments.

[0227] Indicatively, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 3 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions executable by the processing component 302, such as application programs. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the new energy ship fault causality construction method of any of the above embodiments.

[0228] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0229] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0230] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the related listed items.

[0231] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0232] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing causal relationships for failures in new energy ships, characterized in that, The method includes: When a new energy vessel malfunctions, the multimodal heterogeneous data of the new energy vessel is fused and filtered to obtain a core feature set; The core feature set is subjected to invariant risk minimization learning to obtain an environment-invariant feature matrix. Based on mutual information, the causal edge strength between the features of each variable in the environment-invariant feature matrix is ​​evaluated to obtain a causal edge strength matrix. Based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, the core fault causal skeleton is mined. Based on the core fault causal framework and the preset hierarchical node system, an initial hierarchical fault causal graph is constructed. The causal edge strength matrix is ​​fused to determine the initial causal edge weights. Based on the meta-learning model and the initial causal edge weights, the initial hierarchical fault causal graph is dynamically updated to obtain the target hierarchical fault causal graph. Using a large language model, counterfactual samples conforming to causal logic are generated based on real fault samples and the target hierarchical fault causal graph. These counterfactual samples are then input into the target hierarchical fault causal graph for inference, yielding the inference loss corresponding to each counterfactual sample. Similarly, real fault samples are input into the target hierarchical fault causal graph for inference, yielding the inference loss corresponding to each real fault sample. Based on the difference between the inference loss of the counterfactual samples and the inference loss of the real fault samples, the weight correction amount for each causal edge in the target hierarchical fault causal graph is calculated. Based on the weight correction amount, the causal edge weights of each causal edge in the target hierarchical fault causal graph are corrected, resulting in a corrected target hierarchical fault causal graph. The modified hierarchical fault causal graph is transformed into a Markov decision process model. Nodes in the modified hierarchical fault causal graph serve as the state space, causal edges between nodes serve as the action space, and the weights of the causal edges in the modified hierarchical fault causal graph serve as state transition probabilities. Based on the Markov decision process model, a reinforcement learning algorithm is used for iterative training, updating the Q-values ​​of each state-action pair until convergence yields the optimal Q-value table. The reinforcement learning algorithm is trained based on a preset reward function, which is positively correlated with inference accuracy and negatively correlated with inference path length. According to the optimal Q-value table, the causal edge with the largest Q-value at each node is selected as the corresponding optimal action to form the optimal inference path from the fault feature node to the fault root cause node. Based on the optimized hierarchical fault cause-effect graph, the fault cause-effect relationship of the new energy ship is obtained through structured reasoning and the implicit causal relationship mined by the large language model.

2. The method for constructing causal relationships for new energy ship failures according to claim 1, characterized in that, The step of filtering the multimodal heterogeneous data of the new energy ships to obtain the core feature set includes: The multimodal heterogeneous data of the new energy vessel are preprocessed to obtain sensor features, text features, and environmental features. The obtained sensor features, text features, and environmental features are mapped to a unified feature space and fused through a cross-modal self-attention mechanism to obtain fused features; A large language model fine-tuned for the shipbuilding domain is used to evaluate the fault correlation of each feature channel in the fused features, and the importance score of each feature channel is obtained. Based on the importance scores of each feature channel, feature channels with scores higher than a preset threshold are selected to form the core feature set.

3. The method for constructing causal relationships for new energy ship failures according to claim 1, characterized in that, The step of performing invariant risk minimization learning on the core feature set to obtain the environment-invariant feature matrix includes: The environment-invariant feature matrix is ​​obtained using the following formula: in, This represents the environment-invariant feature matrix. This represents a feature extraction network. For feature extraction network parameters, This represents the core feature set. This represents the activation function. This represents the weight matrix of the feature extraction network. This represents the bias vector for weighted evidence collection; the feature extraction network is trained using an optimization objective of invariant risk minimization learning. in, This represents a subset of environments, obtained by dividing the ship's operating environment based on the corresponding ship operating conditions of the core feature set. Represents a single subset of the environment. , These represent the feature data and fault labels of a single environmental subset, respectively. The loss function represents a single subset of environments. Indicates the parameters of the prediction head network. This represents the gradient of the loss function with respect to the parameters of the feature extraction network. This represents the gradient penalty coefficient.

4. The method for constructing causal relationships for new energy ship failures according to claim 1, characterized in that, The step of evaluating the causal edge strength between the variable features in the environment-invariant feature matrix based on mutual information to obtain the causal edge strength matrix includes: The causal edge strength matrix is ​​obtained using the following formula: in, Representing variable characteristics arrive The causal edge strength, i.e., the causal edge strength matrix, , The first element in the environment-invariant feature matrix The and the first Individual variable characteristics, Indicates the fault label variable. Indicates control Under unchanged conditions, and Conditional mutual information, used for quantization and The degree of direct correlation, Control in the display Under unchanged conditions, and Conditional mutual information, used for quantization and The degree of spurious association, This represents the preset spurious association correction coefficient, used to adjust the strength of spurious association removal. express , , The joint probability density function, Indicates in Given the conditions, and The joint conditional probability density, , They represent in Given the conditions, and The conditional probability density.

5. The method for constructing causal relationships for new energy ship failures according to claim 1, characterized in that, The step of mining the core fault causal framework based on the causal edge strength matrix, combined with the causal edge strength threshold and conditional independence test, includes: Using each variable feature in the environment-invariant feature matrix as a node, and using the variable feature pairs in the causal edge strength matrix whose strength value is greater than the causal edge strength threshold as candidate causal edges, an initial undirected graph is constructed. For each edge in the initial undirected graph that connects two nodes, a candidate condition set is selected from the remaining nodes excluding the two nodes. The conditional independence test is used to determine whether the two nodes are conditionally independent under the conditions of the candidate condition set. If there is at least one candidate condition set that makes the two nodes conditionally independent, the edge is determined to be an indirect association and is removed from the initial undirected graph. If after traversing all the candidate condition sets, the two nodes do not satisfy the conditional independence, the edge is retained as a valid causal edge. The direction of the effective causal edges is determined to identify the causal orientation between variable features, thus obtaining a directed core fault causal skeleton.

6. The method for constructing causal relationships for new energy ship failures according to claim 1, characterized in that, The step of fusing the causal edge strength matrix to determine the initial causal edge weights includes: The initial causal edge weights are determined using the following formula: in, Represents the nodes in the initial hierarchical fault cause graph To the node The initial causal edge weights, Represents the nodes obtained from the causal edge strength matrix. To the node The causal edge strength, This represents the set of all causal edges in the initial hierarchical fault causality graph. This represents the maximum causal edge strength of all causal edges in the initial hierarchical fault causal graph. Represents a node To the node Experience rating of causal relationships.

7. The method for constructing causal relationships for new energy ship failures according to claim 1, characterized in that, The step of dynamically updating the initial hierarchical fault causality graph based on the meta-learning model and the initial causal edge weights to obtain the target hierarchical fault causality graph includes: By pre-training a meta-learning model using an update task set containing various environmental changes and fault types, general update parameters applicable to different update scenarios can be obtained. When the current scenario of the new energy ship is detected to be a new environment or a new fault, a small number of current samples are obtained, and the initial causal edge weights of the initial hierarchical fault causal graph are fine-tuned based on the general update parameters to obtain the updated causal edge weights. Based on a small number of current samples, the causal edge strength under the current scenario is recalculated to obtain the updated causal edge strength matrix. In the initial hierarchical fault causal graph, invalid causal edges with weights lower than a preset removal threshold are removed according to the updated causal edge weights, and new causal edges with causal edge strengths higher than a preset addition threshold are added according to the updated causal edge strength matrix, so as to obtain the target hierarchical fault causal graph.

8. The method for constructing causal relationships for new energy ship failures according to claim 7, characterized in that, The steps of removing invalid causal edges with weights lower than a preset removal threshold based on the updated causal edge weights, and adding new causal edges with strengths higher than a preset addition threshold based on the updated causal edge strength matrix, include: The following formula is used to remove invalid causal changes and add new causal edges: in, This represents the set of causal edges in the updated hierarchical fault causality graph. This represents the set of causal edges before the update. Represent any two nodes in the initial hierarchical fault cause-effect graph. Represents a node To the node Updated causal edge weights This indicates the preset rejection threshold. Represents a node To the node The updated causal edge strength, i.e., the updated causal edge strength matrix. This indicates the preset addition threshold.

Images