Failure mode identification method and device of new energy ship power system, computer equipment, storage medium and computer program product

Abstract

The application provides a new energy ship power system failure mode identification method, device, computer equipment, storage medium and computer program product. The method comprises the following steps: acquiring real-time multi-modal data of a new energy ship power system, and extracting multi-modal fusion features based on the real-time multi-modal data; based on the multi-modal fusion features, searching in a pre-constructed failure mode knowledge graph to obtain each knowledge graph associated result associated with the real-time state of the new energy ship power system; splicing the multi-modal fusion features and the graph features of each knowledge graph associated result to obtain a feature splicing result; inputting the feature splicing result into a large language model to obtain a failure mode identification result for the new energy ship power system. The method can accurately and reliably identify the failure mode of the new energy ship power system.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer equipment, storage medium and computer program product for failure mode identification of a new energy ship power system. Background Technology

[0002] As new energy ship power systems evolve towards high power density and multi-energy source integration, their failure modes exhibit characteristics of multi-mode coupling and cross-domain transmission.

[0003] Currently, failure mode identification of new energy ship propulsion systems mainly relies on data-driven models such as general-purpose large language models, which utilize their powerful pattern recognition and reasoning capabilities to analyze real-time data. However, when faced with complex new energy ship propulsion systems, the reasoning process of these general-purpose large language models is like a "black box," which may produce misjudgments that violate physical mechanisms or common sense in the field, making it difficult to guarantee the reliability and credibility of the results.

[0004] Therefore, existing technologies cannot accurately and reliably identify the failure modes of new energy ship power systems. Summary of the Invention

[0005] Based on this, the purpose of this application is to at least solve one of the above-mentioned technical defects, in particular the technical defect that the prior art cannot accurately and reliably identify the failure modes of new energy ship power systems. This application provides a method, device, computer equipment, computer-readable storage medium and computer program product for identifying failure modes of new energy ship power systems that can accurately and reliably identify failure modes of new energy ship power systems.

[0006] Firstly, this application provides a failure mode identification method for a new energy ship power system, the method comprising:

[0007] Acquire real-time multimodal data of the power system of new energy ships, and extract multimodal fusion features based on the real-time multimodal data;

[0008] Based on multimodal fusion features, the system searches in a pre-constructed failure mode knowledge graph to obtain the association results of various knowledge graphs related to the real-time status of the new energy ship power system.

[0009] The multimodal fusion features are concatenated with the graph features of the association results of each knowledge graph to obtain the feature concatenation result;

[0010] The feature concatenation results are input into a large language model to obtain failure mode identification results for new energy ship power systems.

[0011] In an exemplary embodiment, the feature concatenation result is input into a large language model to obtain failure mode identification results for new energy ship power systems, including:

[0012] The feature concatenation results are input into a pre-built Bayesian network model to obtain the probability distribution results of the new energy ship power system in each candidate failure mode; the Bayesian network model is constructed based on the causal relationship between historical failure data and failure mode knowledge graph;

[0013] The probability distribution results and feature concatenation results are concatenated and then input into the large language model to obtain the failure mode identification results for new energy ship power systems.

[0014] In one exemplary embodiment, multimodal fusion features are extracted based on real-time multimodal data, including:

[0015] Real-time multimodal data is preprocessed to obtain processed multimodal data;

[0016] Obtain weight information for different modalities, and then perform weighted processing on the processed multimodal data based on the weight information for different modalities to obtain multimodal fusion features.

[0017] In one exemplary embodiment, based on multimodal fusion features, a search is performed in a pre-constructed failure mode knowledge graph to obtain association results of various knowledge graphs related to the real-time state of the new energy ship power system, including:

[0018] Determine the similarity between the multimodal fusion features and each candidate feature in the pre-constructed failure mode knowledge graph;

[0019] Based on the similarity between the multimodal fusion features and each candidate feature, the association results of each knowledge graph are selected from the association results of each candidate knowledge graph.

[0020] In one exemplary embodiment, the failure mode knowledge graph is constructed through the following steps:

[0021] Acquire multimodal datasets of new energy ship power systems throughout their entire lifecycle, and extract multimodal fusion feature sets based on these datasets;

[0022] Based on the multimodal fusion feature set, multiple failure modes are extracted, and simulation test data for each failure mode are obtained.

[0023] A large language model is used to infer the simulation test data of each failure mode, determine the failure mechanism of each failure mode, and determine the correlation strength information between each failure mode and the corresponding failure mechanism.

[0024] A failure mode knowledge graph is constructed based on each failure mode, the failure mechanism of each failure mode, and the correlation strength between each failure mode and the corresponding failure mechanism.

[0025] In one exemplary embodiment, multiple failure modes are extracted based on a multimodal fusion feature set, including:

[0026] Based on the multimodal fusion feature set, a multi-dimensional collaborative analysis of the new energy ship power system is conducted from the dimensions of the entire life cycle stage, component level, and operation scenario type to identify various potential failure modes.

[0027] A confidence model is used to evaluate each potential failure mode, determine the confidence level of each potential failure mode, and identify potential failure modes with a confidence level higher than a preset confidence threshold as failure modes.

[0028] Secondly, this application provides a failure mode identification device for a new energy ship power system, the device comprising:

[0029] The acquisition module is used to acquire real-time multimodal data of the new energy ship power system and extract multimodal fusion features based on the real-time multimodal data;

[0030] The retrieval module is used to search in a pre-built failure mode knowledge graph based on multimodal fusion features to obtain the association results of various knowledge graphs related to the real-time status of the new energy ship power system.

[0031] The splicing module is used to splice the multimodal fusion features with the graph features of the association results of each knowledge graph to obtain the feature splicing result;

[0032] The recognition module is used to input the feature splicing results into the large language model to obtain the failure mode recognition results for the new energy ship power system.

[0033] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0034] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0035] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0036] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0037] The failure mode identification method, apparatus, computer equipment, storage medium, and computer program product for new energy ship power systems provided in this application acquire real-time multimodal data of the new energy ship power system and extract multimodal fusion features based on the real-time multimodal data; based on the multimodal fusion features, a search is performed in a pre-constructed failure mode knowledge graph to obtain the association results of each knowledge graph associated with the real-time state of the new energy ship power system; the multimodal fusion features are concatenated with the graph features of each knowledge graph association result to obtain the feature concatenation result; the feature concatenation result is input into a large language model to obtain the failure mode identification result for the new energy ship power system; thus, multimodal fusion ensures... To enhance the comprehensiveness of state perception, knowledge graph retrieval introduces domain prior knowledge, providing professional context for identification. The large language model enables deep semantic understanding and comprehensive judgment of complex features and knowledge associations. Based on the constructed analysis process consisting of multimodal fusion features, knowledge graph retrieval, feature concatenation, and large language model reasoning, it combines data-driven state perception with knowledge-driven logical reasoning. This solves the problem that traditional methods cannot accurately and reliably identify failure modes due to the complex structure, hidden failure modes, and strong coupling of new energy ship power systems. It improves the accuracy of failure mode identification and the interpretability and credibility of the conclusions, providing core technical support for the safe operation and intelligent maintenance of ships. Attached Figure Description

[0038] 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.

[0039] Figure 1 This is a flowchart illustrating a failure mode identification method for a new energy ship power system in one embodiment.

[0040] Figure 2 This is a flowchart illustrating a failure mode identification method for a new energy ship power system in another embodiment;

[0041] Figure 3 This is a structural block diagram of a failure mode identification device for a new energy ship power system in one embodiment;

[0042] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0043] 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.

[0044] In one exemplary embodiment, such as Figure 1 As shown, a failure mode identification method for a new energy ship power system is provided. Taking the application of this method to a server as an example, the method includes the following steps S102 to S108. Wherein:

[0045] Step S102: Obtain real-time multimodal data of the new energy ship power system, and extract multimodal fusion features based on the real-time multimodal data.

[0046] Among them, the new energy ship power system refers to the ship propulsion and energy management system that uses non-traditional fossil fuels such as lithium batteries, hydrogen fuel cells, solar energy, and wind energy as energy sources. It includes subsystems such as energy storage, power conversion, propulsion motors, and intelligent control.

[0047] Among them, real-time multimodal data refers to multi-source heterogeneous data that is collected in real time by various data acquisition devices during the operation of a ship and reflects the real-time status of the power system of a new energy ship. It can be sensor data, unstructured text data such as fault / maintenance logs, structured text data such as standard texts and expert experience, labeled data such as scene parameter (inland sea / coastal) data, and visual data such as thermal imaging maps.

[0048] Optionally, the server continuously collects real-time multimodal data from the sensor network and monitoring platform deployed on the new energy ship power system. Subsequently, the various types of data are cleaned, aligned, and feature extracted, and then fused into a unified multimodal fusion feature.

[0049] Step S104: Based on the multimodal fusion features, search the pre-constructed failure mode knowledge graph to obtain the association results of each knowledge graph related to the real-time status of the new energy ship power system.

[0050] Among them, the Failure Mode Knowledge Graph is a professional knowledge base for new energy ship power systems organized in a graphical structure. It is based on entity-relationship-attribute triple modeling. The core entities of the graph include "power components, failure modes, failure mechanisms, scenario types, inducing factors, and prevention and control measures". The relationships include "component-failure mode, failure mode-failure mechanism, scenario-inducing factor, inducing factor-failure mode, and failure mode-prevention and control measures". The attributes include failure probability, severity, scenario adaptability, and prevention and control effectiveness.

[0051] Among them, the knowledge graph association result is a set of subgraphs or nodes that match the current multimodal fusion features obtained by retrieving the knowledge graph, and contains information such as relevant failure modes and causes of failure.

[0052] Optionally, the server uses the multimodal fusion features as query vectors to perform an approximate nearest neighbor search or graph traversal in the vector index of the failure mode knowledge graph to obtain the most relevant knowledge graph association results and their corresponding graph features.

[0053] Step S106: The multimodal fusion features are concatenated with the graph features of the association results of each knowledge graph to obtain the feature concatenation result.

[0054] Among them, graph features are structured features extracted from the association results of knowledge graphs, such as node embedding vectors and subgraph embeddings. The feature concatenation result is an enhanced feature vector formed by concatenating multimodal fusion features and graph features along the feature dimension.

[0055] Optionally, the server concatenates the multimodal fusion features with the graph features of the association results of each knowledge graph to obtain the feature concatenation result.

[0056] Step S108: Input the feature splicing result into the large language model to obtain the failure mode identification result for the new energy ship power system.

[0057] Among them, the large language model is an artificial intelligence model that has been pre-trained on massive amounts of text and has powerful semantic understanding and generation capabilities. In this method, it acts as the inference engine and generates diagnostic conclusions in natural language description based on the enhanced features of the input.

[0058] Optionally, the server inputs the feature concatenation results into the large language model. Based on its internalized engineering knowledge and logical reasoning ability, the large language model outputs structured failure mode identification results, such as specific failure mode names, possible causes, risk levels, and recommended measures.

[0059] In practical applications, retrieval-enhanced generation (RAG) technology is introduced to achieve deep integration of knowledge graphs and large language models. The knowledge graph retrieval model is represented as follows: In the formula, For the top-K knowledge graph association results retrieved ( Values ​​range from 3 to 5, with 3 being the preferred value. Input query features (such as the feature vector of "battery pack heating") into the large language model, dimensionality ); For knowledge graph entity / relation features (encoded via the TransE algorithm, dimensionality) ); The cosine similarity function is used. value range The higher the similarity, the stronger the correlation. The RAG fusion reasoning model is represented as... In the formula, The fault diagnosis results output by the large language model; For feature splicing operations; The result concatenation function converts the Top-K knowledge graph results into feature vectors and then concatenates them. Through this formula, the large language model can combine the retrieved knowledge graph relationships (such as "battery overcharge → thermal runaway") with the input query features to generate accurate inference results.

[0060] The aforementioned failure mode identification method for new energy ship propulsion systems acquires real-time multimodal data of the new energy ship propulsion system and extracts multimodal fusion features based on this data. Then, based on these multimodal fusion features, a search is performed in a pre-constructed failure mode knowledge graph to obtain association results from various knowledge graphs related to the real-time state of the new energy ship propulsion system. The multimodal fusion features are then concatenated with the graph features of the association results from each knowledge graph to obtain a feature concatenation result. Finally, this feature concatenation result is input into a large language model to obtain the failure mode identification result for the new energy ship propulsion system. Thus, multimodal fusion ensures comprehensive state perception, and the knowledge graph... Spectral retrieval introduces domain prior knowledge, providing professional context for identification, while the large language model enables deep semantic understanding and comprehensive judgment of complex features and knowledge associations. Based on the constructed analysis process of multimodal fusion features, knowledge graph retrieval, feature concatenation, and large language model reasoning, it combines data-driven state perception with knowledge-driven logical reasoning. This solves the problem that traditional methods cannot accurately and reliably identify failure modes due to the complex structure, hidden failure modes, and strong coupling of new energy ship power systems. It improves the accuracy of failure mode identification and the interpretability and credibility of the conclusions, providing core technical support for the safe operation and intelligent maintenance of ships.

[0061] In an exemplary embodiment, the feature concatenation result is input into a large language model to obtain the failure mode identification result for the new energy ship power system. This includes: inputting the feature concatenation result into a pre-built Bayesian network model to obtain the probability distribution result of the new energy ship power system in each candidate failure mode; the Bayesian network model is constructed based on the causal relationship between historical fault data and the failure mode knowledge graph; and the probability distribution result and the feature concatenation result are concatenated and then input into the large language model to obtain the failure mode identification result for the new energy ship power system.

[0062] Among them, the Bayesian network model is a model that uses failure modes, failure mechanisms, inducing factors, and impact consequences as network nodes, and quantifies the node association strength with prior probability and conditional probability.

[0063] Historical fault data is a collection of case data recorded in the past, including abnormal system states, finally confirmed failure modes, and repair processes.

[0064] The probability distribution result is the posterior probability of each candidate failure mode occurring, calculated by the Bayesian network model under the given current observed features (feature concatenation results), forming a probability vector.

[0065] Optionally, before calling the large language model for final inference, the server first inputs the feature concatenation result into a pre-trained Bayesian network model. The structure and parameters of this model are learned based on historical failure data and incorporate causal relationships from the failure mode knowledge graph as prior constraints. The model performs probabilistic inference, outputting the probability value of the system currently in each known failure mode, i.e., the probability distribution result. Subsequently, this probability distribution vector is concatenated again with the original feature concatenation result to form the final input feature that integrates deterministic features and uncertain probability assessment. Finally, this enhanced feature is input into the large language model, guiding it to focus on high-probability failure modes during text generation and incorporating probability information into the output.

[0066] In practical applications, by integrating the system theory accident model (STAMP), Bayesian networks, and the temporal modeling capabilities of large language models, a three-dimensional failure propagation mechanism of "system hierarchy + probabilistic causality + temporal dynamics" can be constructed. Relying on the Transformer architecture of the large language model, the dynamic failure propagation laws of the entire chain from "component to unit to whole ship" can be revealed. Specifically, this includes:

[0067] First, based on the STAMP theory, the power system of new energy ships is divided into a four-level structure: "ship-wide safety target layer, power unit control layer, core component execution layer, and environmental scenario influence layer," clearly defining the safety constraints, control relationships, and feedback paths at each level. For example, the control layer controls the load distribution of the execution layer's battery packs and fuel cell stacks through the power management system, while the environmental layer influences the operating status of the execution layer's components through parameters such as salt spray and water flow. The failure propagation interfaces between each level are identified (e.g., component failure → abnormal control signal → unit failure → ship-wide safety risk), constructing a system-level failure propagation framework to overcome the limitations of traditional single-component analysis.

[0068] Secondly, probabilistic causal modeling is performed by integrating Bayesian networks and large language models. Based on the failure mode and mechanism analysis results, and combined with the correlation relationships output by the large language model, a Bayesian network failure causal model for the core dynamic unit is constructed. Failure modes, failure mechanisms, inducing factors, and impact consequences are treated as network nodes, and the node correlation strength is quantified using prior probability and conditional probability. The Bayesian network inference formula (modified version of the large language model) is expressed as follows: In the formula, Normalization coefficient ( ); Nodes in failure mode; Evidence nodes (sensor data, visual features); Knowledge graph association results retrieved from a large language model; The similarity between the association result and the "failure mode-evidence" path (values ​​from 0 to 1) is used to correct the prior probability and conditional probability, thereby improving the accuracy of inference. Prior probabilities of failure modes (based on historical data statistics, such as battery thermal runaway) ); The conditional probability of evidence occurring under a given failure mode (based on experimental data and expert experience, such as a sudden temperature rise during thermal runaway). ).

[0069] The formula for updating the auxiliary conditional probability of a large language model is expressed as follows: In the formula, , They are respectively time, The conditional probability at time t; The update factor (values ​​from 0.01 to 0.05, preferably 0.03); The conditional probability is dynamically optimized for the "failure mode-evidence" association probability output by the large language model.

[0070] Finally, the STAMP system hierarchy framework, Bayesian network model, and temporal features of the large language model are integrated. By constraining the association of Bayesian network nodes through system hierarchy, quantifying the transmission strength of system hierarchy through probabilistic causality, and capturing dynamic transmission patterns through the temporal features of the large language model, a failure mode transmission mechanism that is "clearly hierarchical, probabilistically controllable, and temporally dynamic" is formed. For key failure modes (such as battery thermal runaway), a transmission path diagram is drawn to clarify the entire chain logic of "cause → mechanism → failure mode → cross-component transmission → overall ship impact." Combined with Chain-of-Thought (CoT) reasoning of the large language model, interpretable transmission steps are generated.

[0071] In this embodiment, the probability distribution provided by the Bayesian network provides a scientific confidence measure for the failure mode identification results, so that the output of the large language model is no longer a single judgment, but a set of possible options with probability ranking. Under the guidance of this probability, the large language model makes reasoning, and its conclusions are more rigorous and reliable, which is more in line with the needs of decision-making based on risk assessment in engineering practice. Thus, on the basis of accurate identification, the reliability of the results is further enhanced.

[0072] In one exemplary embodiment, extracting multimodal fusion features based on real-time multimodal data includes: preprocessing the real-time multimodal data to obtain processed multimodal data; obtaining weight information for different modalities; and weighting the processed multimodal data based on the weight information for different modalities to obtain multimodal fusion features.

[0073] Preprocessing includes data cleaning (removing outliers and filling in missing values), format standardization, time alignment, and enhancement, all aimed at improving data quality and consistency.

[0074] The processed multimodal data is preprocessed data that is clean, well-organized, and spatiotemporally aligned.

[0075] The weight information is the importance coefficient assigned to the data of each modality, reflecting the relative contribution or credibility of the modality in the current system state assessment. The weights can be fixed (based on prior knowledge) or dynamically calculated (based on data quality or the current scenario).

[0076] Optionally, the server performs parallel preprocessing pipeline operations on the received real-time multimodal data. For example, it uses moving average filtering on time-series data, deblurs and crops image data, and segments and vectorizes text data. In the feature extraction and feature fusion stages, weight information is determined according to predefined rules or real-time calculation. Then, when fusing the feature vectors extracted from each modality, a weighted summation or weighted concatenation method is used, with the modality features with higher weights dominating the final multimodal fusion features.

[0077] In this embodiment, the dynamic weight allocation mechanism can automatically enhance reliable and information-rich modalities and suppress contaminated or secondary modalities, making the generated multimodal fusion features more robust and accurate in reflecting the core abnormal state of the system. This ensures that subsequent steps can obtain high-quality input, thereby improving the accuracy and robustness of the entire recognition process from the source.

[0078] In an exemplary embodiment, based on multimodal fusion features, a search is performed in a pre-built failure mode knowledge graph to obtain the association results of each knowledge graph associated with the real-time state of the new energy ship power system. This includes: determining the similarity between the multimodal fusion features and each candidate feature in the pre-built failure mode knowledge graph; and selecting each knowledge graph association result from the candidate knowledge graph association results based on the similarity between the multimodal fusion features and each candidate feature.

[0079] Candidate features refer to the pre-computed feature vector representations of each entity or subgraph in the failure mode knowledge graph, which are usually obtained through graph embedding techniques.

[0080] Similarity is a measure of how close two feature vectors are to each other; cosine similarity is commonly used.

[0081] Among them, the candidate knowledge graph association result is all entities or subgraphs that can be retrieved in the knowledge graph before similarity calculation.

[0082] Among them, filtering refers to determining the final set of associations based on similarity scores by setting a threshold or selecting the Top-K most similar results.

[0083] Optionally, during the offline phase, the server has already generated candidate feature vectors for all nodes in the failure mode knowledge graph using a graph embedding algorithm and established a vector index. During online retrieval, the server uses the multimodal fusion features extracted online as query vectors, performs a fast approximate nearest neighbor search in the vector index, calculates its similarity with all candidate features, and then filters according to a preset strategy (e.g., similarity greater than 0.8, or returning the top 5 results with the highest similarity) to obtain the set of most matching node IDs. Based on these node IDs, the server extracts complete node attributes and their associated edges from the knowledge graph to form the final knowledge graph association result.

[0084] In this embodiment, a retrieval mechanism based on vector similarity is used to efficiently and accurately locate the most relevant professional knowledge fragments to the current system state from a vast failure mode knowledge graph. This solves the problems of low recall and poor accuracy of traditional keyword or rule-based matching in complex engineering knowledge graphs. Vectorized retrieval can capture deep semantic similarity. Even if the real-time features and the records in the knowledge graph are not exactly the same in specific values, they can be associated as long as the patterns are similar. This ensures that even when faced with unprecedented compound faults or slight symptoms, relevant historical experience and theoretical basis can be found, which greatly improves the reliability and coverage of failure mode recognition when facing unknown variations.

[0085] In an exemplary embodiment, the failure mode knowledge graph is constructed through the following steps: obtaining a multimodal dataset of the new energy ship power system throughout its entire life cycle, and extracting a multimodal fusion feature set based on the multimodal dataset; extracting multiple failure modes based on the multimodal fusion feature set, and obtaining simulation test data for each failure mode; using a large language model to infer the simulation test data of each failure mode, determining the failure mechanism of each failure mode, and determining the correlation strength information between each failure mode and the corresponding failure mechanism; and constructing a failure mode knowledge graph based on each failure mode, the failure mechanism of each failure mode, and the correlation strength information between each failure mode and the corresponding failure mechanism.

[0086] Among them, the full life cycle multimodal dataset covers all relevant multimodal data generated in each stage of the new energy ship power system from design simulation, manufacturing testing, operation monitoring to maintenance and decommissioning.

[0087] Among them, the multimodal fusion feature set is a set of fusion features extracted in batches from the above full-cycle data, which is used for systematic pattern mining.

[0088] Among them, failure modes are the specific manifestations of system function loss or performance degradation identified by analyzing feature sets.

[0089] Among them, the simulation test data is a data packet that records the occurrence process of each identified failure mode in detail, obtained by conducting simulation tests specifically for in-depth research on each identified failure mode, and contains complete multimodal records.

[0090] Among them, the failure mechanism is the underlying physical, chemical or logical cause that leads to the occurrence of the failure mode, such as material fatigue, control logic error, electromagnetic interference, etc.

[0091] Among them, the correlation strength information quantifies the confidence level or influence of the causal relationship between failure mode and failure mechanism, and is usually expressed as probability or weight.

[0092] Optionally, the server first collects and integrates various types of data throughout the system's entire lifecycle to form a multimodal dataset. A batch processing workflow is then used to extract a multimodal fusion feature set. Clustering, anomaly detection algorithms, and expert review are employed to mine and confirm multiple failure modes from the feature set. For each mode, corresponding simulation test data (potentially from fault reproduction experiments, high-fidelity simulations, or real accident reports) is collected. Subsequently, the description of each failure mode and its simulation test data summary are constructed as prompt words and submitted to a large language model. This model is then tasked with analyzing and inferring the failure mechanism and assessing the correlation strength between the mechanism and the mode. After expert verification, the output of the large language model is stored as a triple (failure mode, correlation strength, failure mechanism) in a graph database, gradually constructing a complete failure mode knowledge graph.

[0093] In this embodiment, data mining is performed using a multimodal dataset covering the entire lifecycle, and a knowledge graph is constructed by leveraging the deep reasoning capabilities of a large language model, ensuring the completeness, depth, and accuracy of the knowledge graph content.

[0094] In practical applications, the construction of a failure mode knowledge graph involves operations such as data standardization, feature fusion, data alignment, data augmentation, and data cleaning.

[0095] When standardizing multimodal data, sensor data adopts a time-series database format, unstructured text such as failure modes / maintenance logs adopts JSON format, standard text and expert experience adopt structured text format, scene parameters (inland waterway / coastal) adopt a labeled format, and visual data such as thermal imaging adopts a standardized pixel format.

[0096] When performing feature fusion on multimodal data, a multimodal feature fusion model is adopted. ,in Feature fusion processing is performed to obtain multimodal fusion features, thus obtaining a multimodal fusion feature set.

[0097] When performing data augmentation on multimodal fusion feature sets, a few-shot data augmentation model is adopted. To achieve data augmentation. In the formula, For the enhanced fault dataset; Label data for a small number of samples (quantity) =5 to 20, with 10 core fault cases selected (out of 5 to 20). For large language models (such as LLaMA-7B); The input is the template function to be suggested. Each sample is labeled with a fault type, and the output is a structured prompt text. Similar fault cases are generated through context learning using a large language model to solve the problem of insufficient data for rare failure modes.

[0098] When performing data alignment and cleaning, for multimodal fusion feature sets ,use After alignment, enhancement, and cleaning, the following is obtained: In the formula, This is a multi-source data alignment function that achieves accurate correspondence of heterogeneous data through semantic matching, temporal calibration, and component number association. This is a data cleaning function that removes outliers, missing values, and redundant data to ensure data accuracy.

[0099] In this application, the failure mode knowledge graph is dynamically updated, and the dynamic update model is represented as follows: In the formula, for The state of the knowledge graph at any given moment; For knowledge update actions (adding new entities, correcting relationships, supplementing attributes); The reward function for knowledge consistency verification (values ​​from 0 to 1); To update the step size (values ​​from 0.03 to 0.07, preferably 0.05); The similarity between the output of the large language model and expert feedback is used to constrain the rationality of update actions and ensure that the updated knowledge is consistent with the actual situation in the domain.

[0100] In an exemplary embodiment, multiple failure modes are extracted based on a multimodal fusion feature set, including: based on the multimodal fusion feature set, multi-dimensional collaborative analysis of the new energy ship power system is performed from the dimensions of the entire life cycle stage, component level, and operation scenario type of the new energy ship power system to identify each potential failure mode; a confidence model is used to evaluate each potential failure mode to determine the confidence level of each potential failure mode, and potential failure modes with confidence levels higher than a preset confidence threshold are identified as failure modes.

[0101] The full life cycle stage dimension refers to different time stages such as design, manufacturing, operation, maintenance, and decommissioning.

[0102] Among them, the component level dimension refers to different composition levels from single unit to module to unit.

[0103] Among them, the operational scenario type dimension refers to different environments such as inland rivers or coastal areas.

[0104] Among them, multi-dimensional collaborative analysis refers to integrating the feature information of the above three dimensions, performing cross-comparison and correlation mining, in order to discover those failure modes that only appear under specific combinations of dimensions.

[0105] The confidence model can be expressed as follows: In the formula, For the first Extraction confidence level of failure mode (value from 0 to 1, threshold set to 0.7, if higher than the threshold, it is judged as a valid failure mode); For the first Failure modes (such as battery thermal runaway, membrane electrode failure); For multilayer perceptron (input dimension) The output dimension is the number of failure mode categories. , ); The function transforms the output into a probability distribution, enabling confidence assessment of failure modes.

[0106] The pre-set confidence threshold is the minimum confidence level standard for determining whether a potential failure mode is valid.

[0107] Optionally, the server performs multi-dimensional analysis on the multimodal fusion feature set. Through pattern mining under this multi-dimensional constraint, a series of potential failure modes are identified. For each potential failure mode, a confidence model is invoked for evaluation. This model may consider factors such as the frequency of the pattern in the data, its stability on different subsets of data, and its degree of conformity with known physical constraints, to calculate a comprehensive confidence score. Finally, the confidence score is compared with a preset confidence threshold, and only those failure modes with high confidence scores are retained and formally confirmed as failure modes that can be used for knowledge graph construction.

[0108] In practical applications, a fusion approach of "large language model-driven + data-driven + expert knowledge + experimental verification" can be adopted. Relying on the entity extraction and logical reasoning capabilities of the large language model, the failure modes of the core power unit can be comprehensively identified, the fault mechanism can be revealed, and a foundation can be provided for the construction of the transmission mechanism and causal reasoning.

[0109] First, based on a multimodal fusion feature dataset, combined with Fault Tree Analysis (FTA), Failure Mode and Effects Analysis (FMEA) methods, and large language model entity extraction technology, potential failure modes can be identified from three dimensions: "full life cycle stage (design, manufacturing, operation, maintenance, decommissioning), component level (unit-module-apartment), and scenario type (inland waterway / coastal)". The confidence model for failure mode entity extraction is as follows: The system employs statistical analysis (failure frequency, severity, and probability of occurrence) and confidence level to screen core failure modes, including battery pack thermal runaway, SEI membrane decomposition, overcharge / over-discharge short circuits, fuel cell stack membrane electrode failure, water / gas shortage, catalyst poisoning, propulsion motor winding short circuits, bearing wear, overload burnout, power management system signal anomalies, and module failures. An expert review mechanism is then introduced to verify and supplement the identification results, forming a list of core failure modes for new energy ship power systems, clarifying the triggering conditions, impact range, and related components for each failure mode. A fusion approach of "large language model-driven + data-driven + expert knowledge + experimental verification" is adopted, leveraging the entity extraction and logical reasoning capabilities of the large language model to comprehensively identify core power unit failure modes, revealing failure mechanisms and providing a foundation for constructing transmission mechanisms and causal reasoning.

[0110] Then, combining theories from multiple disciplines such as electrochemistry, thermodynamics, and mechanical engineering, the core mechanisms of each failure mode are revealed through experimental testing, numerical simulation, and large-scale linguistic model mechanistic reasoning: For battery failures, the electrochemical and thermodynamic coupling mechanism is analyzed through cycle life testing, thermal runaway simulation experiments, and the "SEI membrane decomposition → electrolyte combustion" path inferred by the large-scale linguistic model; for fuel cell failures, the chemical mechanisms of membrane electrode proton conductivity decay and catalyst activity reduction are revealed through polarization curve testing, electron microscopy analysis, and the infusion of domain knowledge into the large-scale linguistic model; for propulsion motor and power management system failures, the mechanical and electrical coupling mechanisms of winding insulation aging, mechanical vibration wear, and signal interference are elucidated through mechanical wear testing and circuit simulation. A "failure mode-failure mechanism-influence path" correlation table is established to quantify the correlation strength between mechanisms and failure modes, providing mechanistic support for the subsequent construction of causal reasoning models.

[0111] In this embodiment, a strategy combining multi-dimensional collaborative analysis and confidence model evaluation is adopted to achieve high accuracy and high reliability in failure mode mining, ensuring that the failure mode set ultimately used for constructing knowledge graphs and online identification are all real modes that have been rigorously verified, have statistical significance and engineering significance.

[0112] This application also provides a method for constructing and optimizing a scenario-based dynamic causal reasoning model. Targeting the dynamic characteristics of dynamic systems, scenario-specific needs, and real-time requirements, it integrates large language model temporal modeling, dynamic Bayesian networks (DBN), and counterfactual reasoning capabilities to construct a scenario-based dynamic causal reasoning model. This solves the "causal lag" problem of traditional static models, achieving failure mode temporal propagation, scenario-based accurate matching, and real-time reasoning. Specifically, this method includes:

[0113] Step 1: Construction of a scenario-based failure cause library and adaptation to a large language model.

[0114] Specifically, based on the differentiated characteristics of inland waterways and coastal areas, scenario-based failure causes are categorized and analyzed. Combined with the scenario recognition capabilities of large language models, a scenario-based failure cause database is constructed. The scenario feature matching formula is expressed as follows: In the formula, The matching score between the i-th scene and the input scene features (values ​​range from 0 to 1, with a threshold of 0.6; if the score is higher than the threshold, it is considered the target scene). For scenario i (inland river / coastal); The input scene feature vector (including water flow velocity, salt spray concentration, etc., dimension 768×1); Let j be the j-th feature component of the input scene; This represents the j-th standard feature component of the i-th scene type; Feature weights (salt spray concentration in coastal scenes) Inland river scene with shallow beach distribution Through data statistics, expert evaluation, and large language model quantification, the probability of occurrence and the ability to trigger failure modes of each factor in different scenarios are determined, achieving a precise mapping between "scenario and factor" and providing support for scenario adaptation of subsequent inference models.

[0115] Step 2: Construct a model that integrates large language model, Transformer, and DBN for temporal reasoning.

[0116] Specifically, based on a static Bayesian network, a large language model Transformer architecture and a DBN model are introduced to construct a fusion temporal inference model to capture the temporal propagation patterns of failure modes. The Transformer temporal feature extraction formula is expressed as follows: In the formula, For the extraction of long-range temporal features (dimensions) , For timing step size, =60 to 120, preferably 90); For the first The temporal feature vector at each moment; For multi-head attention mechanism (number of heads) ); For layer normalization operation; To mitigate the vanishing gradient problem, residual connections are used to capture the long-range time-series relationship of "temperature change → internal resistance change → voltage fluctuation". The state transition formula of the dynamic Bayesian network (corrected for the large language model) is expressed as follows: In the formula, for The posterior probability of the failure mode state at any given time; The state transition probability matrix (dimension) , (Number of failure mode categories) The Sigmoid activation function (values ​​from 0 to 1); For a multilayer perceptron, temporal features are... This is transformed into state correction coefficients, enabling dynamic adjustment of state transition probabilities and resolving the causal lag problem. The counterfactual reasoning formula is expressed as follows: In the formula, Counterfactual intervention (e.g., "no battery overcharge occurred") (as intervention variables) The intervention coefficient (values ​​from 0.2 to 0.5, preferably 0.3); Intervention variables Failure Mode The correlation (calculated using a large language model, with values ​​ranging from 0 to 1) can be used to assess the impact of interventions on failure modes, such as "if overcharging is avoided, the probability of thermal runaway decreases from 70% to 20%".

[0117] Step 3: Scenario-based model adaptation and lightweight optimization of large language models.

[0118] Specifically, a scenario-based failure cause library is integrated with the fusion inference model, and model parameters (prior probabilities, state transition probabilities) are adjusted for different scenarios to construct a scenario-specific dynamic causal inference model. Simultaneously, lightweight large language model quantization technology is employed to meet real-time requirements. The model quantization formula (INT8 quantization) is expressed as follows: In the formula, The quantized weights (8-bit integers); Original floating-point weights (32 bits); For quantization step size ( ); , These are the maximum and minimum weight values, respectively. Zero offset ( This reduces the weights of the large language model from 32 bits to 8 bits, improving inference speed by more than 4 times. The cloud-edge collaborative inference latency constraint formula is expressed as follows: In the formula, For total inference delay (requirement) (to meet the requirements for thermal runaway early warning) For edge device inference latency (lightweight model), ); For cloud-based large language model inference latency ( ); For indicator functions ( hour ,otherwise ); Risk level (calculated by the model, with a value from 0 to 1); A risk threshold (value 0.6) is set to dynamically allocate "low-risk edge reasoning and high-risk cloud-based deep reasoning".

[0119] This application also provides a knowledge graph visualization and closed-loop optimization method. Based on knowledge graph technology, the interpretability of large language models, and reinforcement learning, it visualizes the failure propagation path and, by combining inference results, operational feedback, and large language model output, constructs a "data-knowledge-model-feedback" closed-loop optimization mechanism to continuously improve the method's accuracy, practicality, and real-time performance. The method specifically includes:

[0120] Step 1: Visualization of failure propagation path (driven by large language model CoT).

[0121] Specifically, based on the constructed knowledge graph and the CoT reasoning model, combined with the failure mode propagation mechanism and reasoning results, a visual interface is developed. The CoT reasoning path generation model is represented as follows: ,in In the formula, This is the generated failure propagation path; This is a function to prompt the thought process; For the first One reasoning step; For the first Step failure mode / state; For the first Step probability (values ​​from 0 to 1); For the first The knowledge graph nodes are linked step by step; Number of reasoning steps ( Values ​​range from 3 to 5, with 4 being preferred, to achieve step-by-step path generation of "cause → mechanism → failure mode → consequence". The visualized correlation strength model is represented as follows: In the formula, For the first The association strength of each path edge (values ​​from 0 to 1); For path probability; Failure mode confidence; Similarity between the path and expert experience; adjust edge thickness and color (intensity) based on intensity values. The thick red border is 0.5 to 0.8 mm, while the middle yellow border is 0.5 to 0.8 mm. (Using a thin blue border) to achieve visual differentiation. The visualization interface implements three main functions: visual matching of failure modes and scenario-based causes, graphical display of failure propagation paths, and visual labeling of risk levels, providing intuitive support for operation and maintenance decisions.

[0122] Step 2: Construction and operation of closed-loop optimization mechanism (reinforcement learning feedback).

[0123] Specifically, a closed-loop optimization mechanism is constructed by combining dynamic causal reasoning results, operational feedback data, new fault cases, updated specifications, and the output of a large language model. The reinforcement learning feedback reward function is represented as follows: In the formula, The total reward value (ranging from 0 to 1); ( ), ( ), ( ) represents the weighting coefficient; Rewards for reasoning accuracy ( ); The interpretability reward is based on the operations and maintenance personnel's approval rating of the CoT path, with a value ranging from 0 to 1. For real-time rewards ( , (For maximum allowable delay). Model parameter update model representation is as follows: In the formula, These are the model parameters (including Bayesian network probabilities and fine-tuning parameters for the large language model). The learning rate (ranging from 0.001 to 0.01, preferably 0.005); The gradient of the reward function with respect to the parameters is used to maximize the reward value through the gradient ascent algorithm, thereby optimizing the parameters.

[0124] In practical applications, a closed-loop iteration cycle is set (it is recommended to do this every 1 to 2 months). Through continuous iteration, the knowledge base, transmission mechanism, reasoning model, and lightweight model are dynamically optimized to adapt to the needs of technological upgrades and scenario expansion of new energy ship power systems.

[0125] This application, through the deep integration of multiple theories, technologies, and large language models, constructs a failure mode propagation mechanism and a dynamic causal reasoning method, which has the following significant advantages compared to existing technologies:

[0126] First, the comprehensiveness of failure mode identification and propagation mechanism is significantly improved. By integrating STAMP theory, Bayesian networks, and large language models, and relying on multimodal fusion and few-sample learning capabilities, it achieves three-dimensional analysis of "system hierarchy + probabilistic causality + temporal dynamics," covering the entire life cycle of core power units, multi-component collaboration, and failure modes in inland / coastal waterways. The failure mode identification coverage is ≥95%, and the completeness of propagation path characterization is improved by more than 60% compared with traditional methods. It can effectively identify rare failure modes, making up for the limitations of single theory, static analysis, and scarce samples.

[0127] Second, it achieves collaborative optimization of dynamic reasoning, scenario adaptation, and real-time performance. Through a fusion model of large language model, Transformer, and DBN, it effectively captures the temporal propagation patterns of failure modes, solving the "causal lag" problem, achieving a temporal reasoning accuracy of ≥92%. A scenario-based cause library and dedicated model design enable precise matching of "scenario-cause-failure mode," with scenario-based risk assessment error ≤8%. Lightweight large language model design and cloud-edge collaboration technology reduce inference latency to ≤1s, meeting real-time early warning requirements and adapting to the differentiated navigation and real-time operation and maintenance needs of inland waterways and coastal areas.

[0128] Third, decision support capabilities and visualization levels have been significantly improved. The combination of large language model CoT inference and knowledge graph visualization enables a step-by-step and intuitive presentation of failure propagation paths and risk levels, reducing fault tracing time by more than 70% compared to traditional methods. Operation and maintenance personnel can quickly locate the root cause of the fault and key prevention and control points. The "data-knowledge-model-feedback" closed-loop mechanism, combined with reinforcement learning, ensures that the method continuously adapts to technological upgrades. After three months of cumulative operation, the inference accuracy has further improved by 5-7%.

[0129] Fourth, it exhibits outstanding domain adaptability and generalization capabilities. Through LoRA fine-tuning and domain knowledge injection, the general-purpose language model is transformed into a model specifically for new energy ships, increasing the domain fault identification accuracy to over 94%. Its few-sample learning capability allows for rapid adaptation to new power components (such as new fuel cells), with cross-component generalization accuracy decreasing by ≤5%, eliminating the need for large-scale retraining and reducing adaptation costs. It provides precise theoretical support for the inherently safe design and fault prevention strategy formulation of new energy ship power systems, reducing the failure rate of core power units by ≥15%, significantly improving the navigation safety of new energy ships, and promoting the green and intelligent transformation of the shipbuilding industry.

[0130] In another embodiment, such as Figure 2 As shown, a failure mode identification method for a new energy ship power system is provided. Taking the application of this method to a server as an example, the method includes the following steps:

[0131] Step S202: Obtain real-time multimodal data of the new energy ship power system, and extract multimodal fusion features based on the real-time multimodal data.

[0132] Step S204: Based on the multimodal fusion features, search the pre-constructed failure mode knowledge graph to obtain the association results of each knowledge graph related to the real-time status of the new energy ship power system.

[0133] Step S206: The multimodal fusion features are concatenated with the graph features of the association results of each knowledge graph to obtain the feature concatenation result.

[0134] Step S208: Input the feature splicing result into the pre-built Bayesian network model to obtain the probability distribution results of the new energy ship power system in each candidate failure mode; the Bayesian network model is constructed based on the causal relationship between historical fault data and failure mode knowledge graph.

[0135] Step S210: The probability distribution result and the feature concatenation result are concatenated and then input into the large language model to obtain the failure mode identification result for the new energy ship power system.

[0136] It should be noted that the specific limitations of the above steps can be found in the specific limitations of the failure mode identification method for a new energy ship power system described above.

[0137] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0138] The following describes the failure mode identification device for a new energy ship power system provided in the embodiments of this application. The failure mode identification device for a new energy ship power system has the same inventive concept as the failure mode identification method for a new energy ship power system described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more failure mode identification device embodiments for a new energy ship power system provided below can be referred to the limitations of the failure mode identification method for a new energy ship power system described above. The failure mode identification device for a new energy ship power system described below and the failure mode identification method for a new energy ship power system described above can be referred to each other, and will not be repeated here.

[0139] In one exemplary embodiment, Figure 3 This application provides a schematic diagram of the structure of a failure mode identification device for a new energy ship power system, as shown in the embodiments of this application. Figure 3 As shown, the failure mode identification device for the new energy ship power system includes: an acquisition module 302, a retrieval module 304, a splicing module 306, and an identification module 308, wherein:

[0140] The acquisition module 302 is used to acquire real-time multimodal data of the new energy ship power system and extract multimodal fusion features based on the real-time multimodal data.

[0141] The retrieval module 304 is used to retrieve the results of each knowledge graph associated with the real-time status of the new energy ship power system by searching in the pre-constructed failure mode knowledge graph based on multimodal fusion features.

[0142] The splicing module 306 is used to splice the multimodal fusion features with the graph features of the association results of each knowledge graph to obtain the feature splicing result;

[0143] The recognition module 308 is used to input the feature splicing results into the large language model to obtain the failure mode recognition results for the new energy ship power system.

[0144] In an exemplary embodiment, the identification module 308 is specifically used to input the feature splicing result into a pre-built Bayesian network model to obtain the probability distribution result of the new energy ship power system in each candidate failure mode; the Bayesian network model is constructed based on the causal relationship in historical fault data and failure mode knowledge graph; the probability distribution result and the feature splicing result are spliced ​​together and then input into the large language model to obtain the failure mode identification result for the new energy ship power system.

[0145] In an exemplary embodiment, the acquisition module 302 is specifically used to preprocess real-time multimodal data to obtain processed multimodal data; acquire weight information for different modalities; and perform weighted processing on the processed multimodal data based on the weight information for different modalities to obtain multimodal fusion features.

[0146] In an exemplary embodiment, the retrieval module 304 is specifically used to determine the similarity between the multimodal fusion feature and each candidate feature in the pre-constructed failure mode knowledge graph; and to filter out each knowledge graph association result from the association results of each candidate knowledge graph based on the similarity between the multimodal fusion feature and each candidate feature.

[0147] In an exemplary embodiment, the apparatus further includes: a construction module, configured to acquire a multimodal dataset of the new energy ship power system throughout its entire lifecycle, and extract a multimodal fusion feature set based on the multimodal dataset; extract multiple failure modes based on the multimodal fusion feature set, and acquire simulated test data for each failure mode; use a large language model to infer the simulated test data for each failure mode, determine the failure mechanism of each failure mode, and determine the correlation strength information between each failure mode and the corresponding failure mechanism; and construct a failure mode knowledge graph based on each failure mode, the failure mechanism of each failure mode, and the correlation strength information between each failure mode and the corresponding failure mechanism.

[0148] In an exemplary embodiment, the construction module is further configured to perform multi-dimensional collaborative analysis of the new energy ship power system based on a multi-modal fusion feature set, from the dimensions of the entire life cycle stage, component level, and operation scenario type of the new energy ship power system, to identify each potential failure mode; to evaluate each potential failure mode using a confidence model, to determine the confidence level of each potential failure mode, and to identify potential failure modes with a confidence level higher than a preset confidence threshold as failure modes.

[0149] In one exemplary embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the failure mode identification method for any of the new energy ship power systems described above.

[0150] In one exemplary embodiment, this application also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the failure mode identification method for any of the new energy ship power systems described in the above embodiments.

[0151] In one exemplary embodiment, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the failure mode identification method for any of the new energy ship power systems described in the above embodiments.

[0152] Indicatively, such as Figure 4 As shown, Figure 4 This is a schematic diagram of the internal structure of a computer device 400 provided in an embodiment of this application. The computer device 400 can be provided as a server. (Refer to...) Figure 4 The computer device 400 includes a processing component 402, which further includes one or more processors, and memory resources represented by memory 401 for storing instructions executable by the processing component 402, such as application programs. The application programs stored in memory 401 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 402 is configured to execute instructions to perform the failure mode identification method for the new energy ship propulsion system of any of the above embodiments.

[0153] The computer device 400 may also include a power supply component 403 configured to perform power management of the computer device 400, a wired or wireless network interface 404 configured to connect the computer device 400 to a network, and an input / output (I / O) interface 405. The computer device 400 may operate on an operating system stored in memory 401, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0154] Those skilled in the art will understand that Figure 4 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.

[0155] 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. Without further limitations, 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.

[0156] 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.

[0157] 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 failure mode identification method for a new energy ship power system, characterized in that, The method includes: Acquire a multimodal dataset of the new energy ship power system throughout its entire life cycle, and extract a multimodal fusion feature set based on the multimodal dataset; Based on the multimodal fusion feature set, multiple failure modes are extracted, and simulation test data for each failure mode are obtained. A large language model is used to infer the simulation test data of each failure mode to determine the failure mechanism of each failure mode and to determine the correlation strength information between each failure mode and the corresponding failure mechanism. A failure mode knowledge graph is constructed based on each failure mode, the failure mechanism of each failure mode, and the correlation strength information between each failure mode and the corresponding failure mechanism. Acquire real-time multimodal data of the new energy ship power system, and extract multimodal fusion features based on the real-time multimodal data; Based on the multimodal fusion features, a search is performed in the failure mode knowledge graph to obtain the association results of each knowledge graph associated with the real-time status of the new energy ship power system. The multimodal fusion features are concatenated with the graph features of the association results of each knowledge graph to obtain the feature concatenation result; Based on the system theory accident model, a system-level failure propagation framework for the new energy ship power system is constructed. Under the constraints of the system-level failure propagation framework, a Bayesian network model is constructed based on historical failure data and the causal relationships in the failure mode knowledge graph. The feature concatenation result is input into the Bayesian network model to obtain the probability distribution of the new energy ship power system in each candidate failure mode; the Bayesian network model is represented as follows: In the formula, The normalization coefficient is... , For failure mode nodes, As evidence nodes, The association results of the knowledge graph. The knowledge graph association results and failure paths similarity, This represents the prior probability of the failure mode. Let M be the conditional probability of evidence node E occurring under failure mode node M; the Bayesian network model uses a large language model to assist in updating the conditional probability, and the update method of the conditional probability is expressed as follows: In the formula, , They are respectively time, The conditional probability at time 1. To update the coefficients, The failure mode node output by the large language model With the evidence node The probability of association; The probability distribution result and the feature concatenation result are concatenated and then input into the large language model to obtain the failure mode identification result for the new energy ship power system.

2. The method according to claim 1, characterized in that, The extraction of multimodal fusion features based on the real-time multimodal data includes: The real-time multimodal data is preprocessed to obtain processed multimodal data; Obtain weight information for different modalities, and perform weighted processing on the processed multimodal data based on the weight information for different modalities to obtain multimodal fusion features.

3. The method according to claim 1, characterized in that, Based on the multimodal fusion features, a search is performed in the failure mode knowledge graph to obtain the association results of each knowledge graph related to the real-time state of the new energy ship power system, including: Determine the similarity between the multimodal fusion features and each candidate feature in the failure mode knowledge graph; Based on the similarity between the multimodal fusion features and each of the candidate features, the association results of each knowledge graph are selected from the association results of each candidate knowledge graph.

4. The method according to claim 1, characterized in that, Based on the multimodal fusion feature set, multiple failure modes are extracted, including: Based on the multimodal fusion feature set, the new energy ship power system is analyzed in a multi-dimensional collaborative manner from the dimensions of the entire life cycle stage, component level, and operation scenario type to identify potential failure modes. A confidence model is used to evaluate each of the potential failure modes, determine the confidence level of each potential failure mode, and identify the potential failure modes with a confidence level higher than a preset confidence threshold as failure modes.

5. A failure mode identification device for a new energy ship power system, characterized in that, The device includes: A construction module is used to acquire a multimodal dataset of the new energy ship power system throughout its entire life cycle, and extract a multimodal fusion feature set based on the multimodal dataset; based on the multimodal fusion feature set, extract multiple failure modes, and acquire simulated test data for each failure mode; use a large language model to infer the simulated test data of each failure mode to determine the failure mechanism of each failure mode, and determine the correlation strength information between each failure mode and the corresponding failure mechanism; based on each failure mode, the failure mechanism of each failure mode, and the correlation strength information between each failure mode and the corresponding failure mechanism, construct a failure mode knowledge graph. The acquisition module is used to acquire real-time multimodal data of the new energy ship power system and extract multimodal fusion features based on the real-time multimodal data. The retrieval module is used to search the failure mode knowledge graph based on the multimodal fusion features to obtain the association results of each knowledge graph associated with the real-time status of the new energy ship power system. The splicing module is used to splice the multimodal fusion features with the graph features of the association results of each knowledge graph to obtain the feature splicing result; based on the system theory accident model, a system-level failure transmission framework for the new energy ship power system is constructed; under the constraints of the system-level failure transmission framework, a Bayesian network model is constructed based on historical fault data and the causal relationship in the failure mode knowledge graph. The identification module is used to input the feature concatenation result into the Bayesian network model to obtain the probability distribution results of the new energy ship power system in each candidate failure mode; the Bayesian network model is represented as follows: In the formula, The normalization coefficient is... , For failure mode nodes, As evidence nodes, The association results of the knowledge graph. The knowledge graph association results and failure paths similarity, This represents the prior probability of the failure mode. Let M be the conditional probability of evidence node E occurring under failure mode node M; the Bayesian network model uses a large language model to assist in updating the conditional probability, and the update method of the conditional probability is expressed as follows: In the formula, , They are respectively time, The conditional probability at time 1. To update the coefficients, The failure mode node output by the large language model With the evidence node The correlation probability is calculated; the probability distribution result and the feature concatenation result are concatenated and then input into the large language model to obtain the failure mode identification result for the new energy ship power system.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images