A ship fault early warning method and system

Abstract

The application discloses a ship fault early warning method and system, the method comprises the following steps: acquiring multi-modal data of a ship system, performing semantic alignment and data preprocessing on the multi-modal data to obtain standardized multi-modal data, performing feature enhancement on the standardized multi-modal data to obtain multi-dimensional enhanced feature data, fusing the enhanced feature data in each dimension to obtain fusion features, inputting the fusion features into a pre-trained multi-level response state machine model, outputting a fault response level, and performing early warning corresponding to the fault response level. As can be seen, through multi-modal data semantic alignment and preprocessing, data heterogeneity and noise interference are eliminated, data availability is improved, weak fault signal features are strengthened through feature enhancement and multi-dimensional fusion, early fault recognition capability is significantly improved, and the pre-trained multi-level response state machine model is combined to quickly output an early warning level. The early warning level can realize accurate fault grading, early warning, effectively reduce early warning lag and false alarm risk.

Description

Technical Field

[0001] This application relates to the field of ship operation and maintenance safety technology, and more specifically, to a ship fault early warning method and system. Background Technology

[0002] With the advancement of dual-carbon goals and the green transformation of the shipbuilding industry, new energy ships such as pure batteries and fuel cells are developing rapidly. The operating conditions of their core components, such as lithium batteries and fuel cell stacks, are complex and affected by environmental factors such as temperature, humidity, salt spray, and dynamic loads. Faults are characterized by weak early signs, multimodal coupling, and easy noise suppression. Accurate and early fault warning has become a key technical prerequisite for ensuring the safe operation and large-scale promotion of new energy ships.

[0003] In existing early warning systems for new energy ships, there is a semantic gap in multi-source heterogeneous data, insufficient ability to extract weak fault signals, and difficulty in effectively capturing early warning features of serious faults such as thermal runaway. This results in delayed warnings, low accuracy, and an inability to meet the needs of early warning.

[0004] How to achieve data semantic alignment and extract early weak signals in the early warning of faults in new energy ships, so as to improve the accuracy of early warning, is a technical issue that needs attention. Summary of the Invention

[0005] In view of the above problems, this application provides a ship fault early warning method and system to achieve data semantic alignment and extract early weak signals in the fault early warning of new energy ships, so as to improve the accuracy of early warning.

[0006] To achieve the above objectives, the following specific solutions are proposed:

[0007] A method for early warning of ship malfunctions, comprising:

[0008] Acquire multimodal data of the ship system;

[0009] The multimodal data is semantically aligned and preprocessed to obtain standardized multimodal data;

[0010] The standardized multimodal data is enhanced to obtain multidimensional enhanced feature data;

[0011] By fusing the enhanced feature data from each dimension, a fused feature is obtained;

[0012] The fused features are input into a pre-trained multi-level response state machine model, which outputs a fault response level and issues an early warning corresponding to the fault response level.

[0013] Optionally, the multimodal data is semantically aligned and preprocessed to obtain standardized multimodal data, including:

[0014] The multimodal data is semantically aligned using a pre-trained cross-ship semantic alignment model to obtain semantically aligned multimodal data. The loss function for training the cross-ship semantic alignment model is as follows:

[0015] ;

[0016] in, Let the loss function be the conditional random field. For real semantic tags, For predicting labels, For the association loss of the pre-built knowledge graph, This refers to the set of nodes in the knowledge graph. This is the set of edges associated with the nodes of the knowledge graph. This is the correlation loss coefficient;

[0017] The semantically aligned multimodal data is denoised using a wavelet denoising model to obtain wavelet-denoised multimodal data. The wavelet denoising model is as follows:

[0018] ;

[0019] in, This is a db4 wavelet transform. It is a soft threshold function. This is the inverse wavelet transform. For the semantically aligned multimodal data, For the wavelet-denoised multimodal data;

[0020] The wavelet-denoised multimodal data is normalized using a normalization model to obtain normalized multimodal data. The normalization model is as follows:

[0021] ;

[0022] in, The maximum value in the wavelet-denoised multimodal data. This is the minimum value in the wavelet-denoised multimodal data;

[0023] The normalized multimodal data is then augmented and purged to obtain standardized multimodal data.

[0024] Optionally, feature enhancement is performed on the standardized multimodal data to obtain multidimensional enhanced feature data, including:

[0025] The standardized multimodal data is enhanced using a pre-trained contrastive learning feature enhancement model to obtain multidimensional enhanced feature data. The loss function for training the contrastive learning feature enhancement model is:

[0026] ;

[0027] in, and These are enhanced view features of different dimensions for the same sample. The cosine similarity function is used. This refers to the temperature parameter.

[0028] Optionally, the method further includes:

[0029] The temporal feature data of the multidimensional enhanced feature data is enhanced by using a variational mode decomposition model to obtain the time-series enhanced multidimensional enhanced feature data. The variational mode decomposition model is as follows:

[0030] ;

[0031] in, For the set of modal components, Let be the center frequency of the k-th modal component. For the first-order partial derivative with respect to time t, For the Dirac function, This indicates a convolution operation.

[0032] Optionally, the fusion of the enhanced feature data from each dimension to obtain the fused feature includes:

[0033] Using a pre-built adaptive allocation model for modal feature weights, the fusion weights for each dimension of the enhanced feature data are determined. The adaptive allocation model for modal feature weights is as follows:

[0034] ;

[0035] in, This represents the feature corresponding to index k. As a time series feature, For spatial features, For semantic features, Features The corresponding attention weights Features of fusion;

[0036] Based on the fusion weights of the enhanced feature data in each dimension, the enhanced feature data in each dimension are fused to obtain fused features.

[0037] Optionally, the optimization process for the model parameters of the multi-level response state machine model is as follows:

[0038] ;

[0039] in, In response to the trajectory, For the trajectory probability distribution, For the reward function, These are the model parameters before adjustment of the multi-level response state machine model. The adjusted model parameters are those for the multi-level response state machine model.

[0040] Optionally, the method further includes:

[0041] The contribution of each feature of the multimodal data to the fault response level is calculated using the SHAP value calculation model. The SHAP value calculation model is as follows:

[0042] ;

[0043] in, Let be the SHAP value of the i-th feature. For feature set, For feature subset, This is the prediction function.

[0044] Optionally, the method further includes:

[0045] Based on the multimodal data, the ship system is optimized for prevention and control using a ship multi-objective prevention and control optimization model, which is as follows:

[0046] ;

[0047] in, For safety losses, For safety loss weighting coefficient, For economic costs, This is the economic cost weighting coefficient;

[0048] The ship multi-target defense optimization model is enhanced through adversarial training, and the adversarial training loss function of the ship multi-target defense optimization model is:

[0049] ;

[0050] in, To counteract the total training loss, To predict losses, For the original sample, For adversarial examples, The distance loss between the original sample and the adversarial sample is... This is the distance loss weighting coefficient.

[0051] Optionally, the method further includes:

[0052] For any intelligent model, when incremental data exceeding a preset amount is obtained, incremental learning is performed on the intelligent model using the incremental data to update the intelligent model. The incremental learning process is as follows:

[0053] ;

[0054] in, Let be the model parameters of the intelligent model at time t. The updated model parameters for the intelligent model. The parameters are trained based on the incremental data. The learning coefficient is the incremental learning factor.

[0055] A ship fault early warning system, comprising:

[0056] The ship system data acquisition module is used to acquire multimodal data of the ship system;

[0057] The data alignment and preprocessing module is used to perform semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data;

[0058] The feature enhancement module is used to enhance the features of the standardized multimodal data to obtain multidimensional enhanced feature data;

[0059] The feature fusion module is used to fuse the enhanced feature data from each dimension to obtain fused features;

[0060] The fault response level early warning module is used to input the fused features into a pre-trained multi-level response state machine model, output the fault response level, and issue an early warning corresponding to the fault response level.

[0061] Optionally, the data alignment preprocessing module includes:

[0062] The semantic alignment module is used to perform semantic alignment on the multimodal data using a pre-trained cross-ship semantic alignment model to obtain semantically aligned multimodal data. The loss function for training the cross-ship semantic alignment model is:

[0063] ;

[0064] in, Let the loss function be the conditional random field. For real semantic tags, For predicting labels, For the association loss of the pre-built knowledge graph, This refers to the set of nodes in the knowledge graph. This is the set of edges associated with the nodes of the knowledge graph. This is the correlation loss coefficient;

[0065] The wavelet denoising module is used to denoise the semantically aligned multimodal data using a wavelet denoising model to obtain wavelet-denoised multimodal data. The wavelet denoising model is as follows:

[0066] ;

[0067] in, This is a db4 wavelet transform. It is a soft threshold function. This is the inverse wavelet transform. For the semantically aligned multimodal data, For the wavelet-denoised multimodal data;

[0068] The normalization module is used to normalize the wavelet-denoised multimodal data using a normalization model to obtain normalized multimodal data. The normalization model is as follows:

[0069] ;

[0070] in, The maximum value in the wavelet-denoised multimodal data. This is the minimum value in the wavelet-denoised multimodal data;

[0071] The data completion and removal module is used to complete and remove data from the normalized multimodal data to obtain standardized multimodal data.

[0072] Optionally, the feature enhancement module includes:

[0073] The contrastive learning feature enhancement module is used to enhance the features of each dimension of the standardized multimodal data using a pre-trained contrastive learning feature enhancement model, resulting in multidimensional enhanced feature data. The loss function for training the contrastive learning feature enhancement model is:

[0074] ;

[0075] in, and These are enhanced view features of different dimensions for the same sample. The cosine similarity function is used. This refers to the temperature parameter.

[0076] Optionally, the system may also include:

[0077] The variational mode decomposition module is used to enhance the temporal feature data of the multidimensional enhanced feature data through a variational mode decomposition model, thereby obtaining the enhanced multidimensional enhanced feature data. The variational mode decomposition model is as follows:

[0078] ;

[0079] in, For the set of modal components, Let be the center frequency of the k-th modal component. For the first-order partial derivative with respect to time t, For the Dirac function, This indicates a convolution operation.

[0080] Optionally, the feature fusion module includes:

[0081] The fusion weight determination module is used to determine the fusion weight of each dimension of the enhanced feature data using a pre-built modal feature weight adaptive allocation model. The modal feature weight adaptive allocation model is as follows:

[0082] ;

[0083] in, This represents the feature corresponding to index k. As a time series feature, For spatial features, For semantic features, Features The corresponding attention weights Features of fusion;

[0084] The fusion module is used to fuse the enhanced feature data of each dimension based on the fusion weights of each dimension to obtain fused features.

[0085] Optionally, the optimization process for the model parameters of the multi-level response state machine model is as follows:

[0086] ;

[0087] in, In response to the trajectory, For the trajectory probability distribution, For the reward function, These are the model parameters before adjustment of the multi-level response state machine model. The adjusted model parameters are those for the multi-level response state machine model.

[0088] Optionally, the system may also include:

[0089] The SHAP value calculation module is used to calculate the contribution of the multimodal data using the SHAP value calculation model, thereby obtaining the contribution of each feature of the multimodal data to the fault response level. The SHAP value calculation model is as follows:

[0090] ;

[0091] in, Let be the SHAP value of the i-th feature. For feature set, For feature subset, This is the prediction function.

[0092] Optionally, the system may also include:

[0093] The prevention and control optimization module is used to optimize the ship system's prevention and control based on the multimodal data using a ship multi-objective prevention and control optimization model. The ship multi-objective prevention and control optimization model is as follows:

[0094] ;

[0095] in, For safety losses, For safety loss weighting coefficient, For economic costs, This is the economic cost weighting coefficient;

[0096] The ship multi-target defense optimization model is enhanced through adversarial training, and the adversarial training loss function of the ship multi-target defense optimization model is:

[0097] ;

[0098] in, To counteract the total training loss, To predict losses, For the original sample, For adversarial examples, The distance loss between the original sample and the adversarial sample is... This is the distance loss weighting coefficient.

[0099] Optionally, the system may also include:

[0100] An incremental learning module is used to perform incremental learning on any intelligent model when more than a preset amount of incremental data is obtained, in order to update the intelligent model. The incremental learning process is as follows:

[0101] ;

[0102] in, Let be the model parameters of the intelligent model at time t. The updated model parameters for the intelligent model. The parameters are trained based on the incremental data. The learning coefficient is the incremental learning factor.

[0103] By employing the aforementioned technical solution, this application acquires multimodal data from a ship system, performs semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data, enhances the standardized multimodal data to obtain multidimensional enhanced feature data, fuses the enhanced feature data from each dimension to obtain fused features, inputs the fused features into a pre-trained multi-level response state machine model, outputs a fault response level, and provides an early warning corresponding to the fault response level. Thus, by semantically aligning and preprocessing multimodal data, data heterogeneity and noise interference are eliminated, improving data usability. Feature enhancement and multi-dimensional fusion strengthen weak fault signal characteristics, significantly improving early fault identification capabilities. Combined with the pre-trained multi-level response state machine model to quickly output early warning levels, accurate fault classification and early warning can be achieved, effectively reducing the risk of delayed warnings and false alarms. Attached Figure Description

[0104] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0105] Figure 1 This is a schematic diagram of a process for implementing ship fault early warning, provided in an embodiment of this application.

[0106] Figure 2 This is a schematic diagram of a process for implementing multimodal data preprocessing of ships, provided in an embodiment of this application.

[0107] Figure 3 This is a schematic diagram of a system structure for realizing ship fault early warning, provided as an embodiment of this application. Detailed Implementation

[0108] 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.

[0109] The proposed solution can be implemented based on a terminal with data processing capabilities, such as a computer, cloud, or server.

[0110] Next, combined Figure 1 The ship fault early warning method of this application may include the following steps:

[0111] Step S110: Obtain multimodal data of the ship system.

[0112] Specifically, a three-tiered collaborative data acquisition architecture involving ship, shore, and cloud can be constructed. High-precision sensors deployed on core ship components can collect real-time operational data such as voltage, current, temperature, vibration, and impedance from critical equipment like lithium batteries, fuel cell stacks, and propulsion motors. Simultaneously, unstructured data on equipment field distribution and structural damage can be acquired using infrared thermal imagers and ultrasonic testing equipment. Furthermore, external environmental data such as temperature, humidity, salt spray concentration, and sea state can be collected, along with textual data including historical fault cases, equipment maintenance records, and classification society specifications. This forms a comprehensive dataset covering operational status, equipment health, external environment, and historical experience. All data is synchronously recorded with a unified timestamp, ensuring accurate alignment of multi-source, multi-type, and multi-format data in both time and space. This provides comprehensive, accurate, and high-quality data support for subsequent semantic alignment, data preprocessing, feature extraction, and fault early warning.

[0113] Step S120: Perform semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data.

[0114] Specifically, cross-domain semantic alignment can be achieved based on the BERT-CRF joint embedding model and dynamic knowledge graph, unifying parameter definitions and semantic mappings for data from different ship types, equipment, and sources, thus resolving inconsistencies in the representation of spatiotemporally heterogeneous data. Subsequently, integrated data preprocessing is performed, using db4 wavelet transform to denoise the original signal, filtering out sea state interference and sensor noise while retaining weak fault signal characteristics. Min-Max normalization unifies the data magnitude, eliminating dimensional differences, and the 3σ criterion is used to remove outliers, while the KNN algorithm completes missing values, ensuring data integrity and reliability. This transforms heterogeneous, noisy, and non-standard data into standardized multimodal data that is spatiotemporally consistent, semantically unified, and of reliable quality.

[0115] Understandably, semantic alignment and data preprocessing can effectively bridge the semantic gap between multi-source data and improve data compatibility across ship types and equipment. Wavelet denoising can significantly suppress environmental and sensor noise, maximize the preservation of weak features of early faults, and improve the identifiability of weak signals. Normalization and outlier handling unify data scale, purify sample quality, and reduce model training difficulty and computational errors. Overall, this can improve data availability and consistency, enhance the accuracy of subsequent feature extraction and fault identification, reduce false alarms and false negatives, and achieve earlier and more accurate fault warnings, providing stable and reliable data support for the safe operation of new energy vessels.

[0116] Step S130: Perform feature enhancement on the standardized multimodal data to obtain multidimensional enhanced feature data.

[0117] Specifically, multi-dimensional feature enhancement can be performed on standardized data of temporal, spatial, and semantic types based on contrastive learning and variational mode decomposition techniques. A SimCLR+MoCo dual-channel feature extractor is used to construct a feature space alignment and temporal enhancement mechanism, amplifying the similarity of features of similar fault types and reducing the confusion of features of dissimilar types. Variational mode decomposition decomposes the original signal into multiple intrinsic mode components, accurately extracting and reconstructing weak fault signal components, highlighting early fault precursor features. Simultaneously, a temporal-spatial-semantic multi-dimensional attention mechanism can be constructed, adaptively allocating feature weights according to fault type and impact, strengthening the expression of key fault features. Through multi-level enhancement processing, redundant information interference is effectively suppressed, fault-sensitive features are highlighted, and ultimately, multi-dimensional enhanced feature data with high signal-to-noise ratio, strong discriminative power, and high representativeness is generated, providing high-quality input for subsequent feature fusion and fault early warning.

[0118] Understandably, feature enhancement can significantly amplify weak fault signals, preventing fault features from being easily drowned out by noise and greatly improving the ability to detect early faults. Contrastive learning can improve feature discriminative power and reduce the probability of false faults, while variational mode decomposition can accurately uncover hidden precursors, enabling early fault warnings. Multidimensional attention mechanisms can focus on key features, reducing redundant computation and improving model efficiency. Overall, this effectively improves feature quality and fault identification, making it easier to capture early features of high-risk faults such as thermal runaway during the warning process, significantly improving warning accuracy and reducing false negatives.

[0119] Step S140: Fuse the enhanced feature data of each dimension to obtain the fused features.

[0120] Specifically, enhanced temporal, spatial, and semantic features can be uniformly aggregated. Based on a multi-dimensional attention mechanism, adaptive weights are assigned to each dimension, dynamically adjusting the weight allocation according to the degree of fault sensitivity and the difference in operating condition impact, highlighting high-contribution features and weakening redundant information. Through weighted summation and feature concatenation, deep coupling of multi-dimensional features is achieved, realizing the complementarity and enhancement of information from different modalities. Furthermore, PCA can be used to reduce the dimensionality of the fused high-dimensional features, retaining core feature components with a cumulative contribution rate of no less than 90%. While ensuring the integrity of fault discrimination information, this simplifies feature dimensions and reduces computational overhead, thereby integrating scattered, multi-dimensional enhanced features into a fused feature with comprehensive representation, high signal-to-noise ratio, and strong discriminative power.

[0121] Understandably, multi-dimensional feature fusion can fully integrate temporal, spatial, and semantic information, compensating for the incomplete representation of single-modal features and significantly improving the completeness and recognizability of fault features. Adaptive weight allocation can focus on key fault information, strengthening the ability to capture weak and coupled faults. PCA dimensionality reduction simplifies calculations while retaining core information, improving early warning response speed. Overall, it effectively improves feature representation capabilities and model discrimination accuracy, reduces missed detections and false positives, and makes early fault features easier to identify, providing a reliable guarantee for achieving rapid, accurate, and stable fault early warning for new energy ships.

[0122] Step S150: Input the fused features into the pre-trained multi-level response state machine model, output the fault response level, and issue an early warning corresponding to the fault response level.

[0123] Specifically, the fused features are input into a pre-trained multi-level response state machine model. Based on the learned and optimized parameters and decision logic, the model rapidly analyzes the input fused features and determines the fault level. According to the fault severity, development trend, and risk level information carried in the features, a reinforcement learning-optimized state transition mechanism automatically matches and outputs the corresponding levels for Level 1 minor faults, Level 2 local faults, and Level 3 severe faults. After determining the level, the model synchronously triggers the corresponding early warning mechanism, executing pre-defined strategies such as tiered alarms, parameter locking, and data uploading. Information such as fault location, fault type, risk level, and recommended handling methods are pushed to the ship-shore control terminal in real time. The entire discrimination and early warning process is executed in a closed loop within a unified state machine framework, ensuring stable early warning output, consistent response logic, and reliable action execution, providing clear and accurate level criteria and instruction guidance for subsequent emergency response.

[0124] Understandably, employing a pre-trained multi-level response state machine model can significantly improve the speed and stability of fault level determination, avoiding delays and errors caused by manual intervention. Hierarchical output of fault levels enables refined risk management, preventing resource waste or insufficient response caused by simple warnings. The model directly outputs matching warning actions, shortening the response chain, improving real-time warnings, and achieving early detection, early alerts, and early intervention. Simultaneously, standardized decision-making logic effectively reduces the probability of false alarms and missed alarms, improving warning reliability and providing efficient, accurate, and controllable intelligent early warning protection for the safe operation of core components of new energy ships.

[0125] The optimization process for the model parameters of the multi-level response state machine model is as follows:

[0126] ;

[0127] in, In response to the trajectory, For the trajectory probability distribution, For the reward function, It can be designed based on response time and loss control parameters. These are the model parameters before adjustment of the multi-level response state machine model. These are the adjusted model parameters for the multi-level response state machine model.

[0128] Once the fault response level (e.g., Level 3 response) is determined, the corresponding fault warning can be identified.

[0129] For example, when only local parameter fluctuations occur, and the core parameters (temperature, voltage fluctuations, etc.) are analyzed by the Sobol index and the sensitivity index is ≥0.2, reaching the first-level warning threshold, and there is no risk of component performance degradation or systemic failure, the first-level fault response is output, and the first-level automatic isolation response process is triggered. The local branches and non-core modules associated with the fault are automatically electrically and logically isolated to block the fault propagation path, while maintaining the normal operation of the main system. The entire process response time is controlled within 500ms.

[0130] When component performance degradation exceeds the range of minor parameter fluctuations, posing a risk of partial functional failure, and the impact of the fault cannot be eliminated by isolating a single branch, a secondary fault response is output, and a secondary load transfer response process is triggered. The load of the faulty component / module is automatically scheduled to redundant units or healthy modules of the same type to ensure that the core functions of the system are not interrupted. The offline isolation and alarm of the faulty unit are completed simultaneously, and the response time of the entire process is controlled within 500ms.

[0131] In the event of serious malfunctions such as thermal runaway or fuel leakage that could lead to systemic safety risks, or significant potential hazards such as equipment damage or navigational safety accidents, a Level 3 fault response will be output, triggering a Level 3 emergency shutdown response procedure. This will involve executing the highest priority safety operations, such as emergency shutdown of the power system, linkage of safety protection devices, and emergency interlocking, prioritizing the safety of personnel and the core of the hull. The entire response time will be controlled within 500ms.

[0132] The state transition logic of the three-level response can be globally optimized through reinforcement learning. The reward function is designed based on the response time and loss control quantity, and the state machine parameters are iteratively optimized to achieve accurate and rapid switching of response states under different fault levels, ensuring closed-loop matching of early warning, decision-making and execution.

[0133] Building upon this foundation, digital twin emergency simulations can be conducted: constructing multiphysics simulation models of ship propulsion systems to simulate fault propagation processes and emergency response effects, and optimizing evacuation routes and resource allocation schemes. Digital twin emergency simulation models can encompass multi-field coupling of electrical, thermal, and mechanical fields, quantifying the impact of environmental factors (temperature, humidity, salt spray) on fault evolution, thereby reducing errors in digital twin emergency simulations.

[0134] The ship fault early warning method provided in this embodiment acquires multimodal data of the ship system, performs semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data, performs feature enhancement on the standardized multimodal data to obtain multidimensional enhanced feature data, fuses the enhanced feature data of each dimension to obtain fused features, inputs the fused features into a pre-trained multi-level response state machine model, outputs the fault response level, and issues an early warning corresponding to the fault response level. Thus, by semantically aligning and preprocessing the multimodal data, data heterogeneity and noise interference are eliminated, improving data usability. Feature enhancement and multi-dimensional fusion strengthen weak fault signal characteristics, significantly improving early fault identification capabilities. Combined with the pre-trained multi-level response state machine model to quickly output early warning levels, accurate fault classification and early warning can be achieved, effectively reducing the risk of early warning lag and false negatives.

[0135] In some embodiments of this application, the process of performing semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data is described, with reference to... Figure 2 This process may specifically include:

[0136] Step S210: Semantically align the multimodal data using a pre-trained cross-ship semantic alignment model to obtain semantically aligned multimodal data.

[0137] The loss function for training the cross-ship semantic alignment model is:

[0138] ;

[0139] in, The conditional random field loss function is used to quantify the semantic label prediction bias. For real semantic tags, For predicting labels, The association loss of the pre-constructed knowledge graph is used to constrain the rationality of semantic node associations. It is a set of nodes in the knowledge graph (fault type, equipment parameters, operating condition characteristics). For the set of edges associated with nodes in a knowledge graph, For example, the correlation loss coefficient, A value of 0.3-0.5 is used to balance the accuracy of semantic prediction and graph association. The cross-ship semantic alignment model can improve the accuracy of semantic transformation of cross-ship data while reducing the alignment error of spatiotemporally heterogeneous data.

[0140] Understandably, unifying the parameter definitions and semantic mapping relationships of data from different ship types, equipment, and sources significantly improves the alignment accuracy and consistency of spatiotemporally heterogeneous data. Cross-ship type semantic alignment models can automatically complete data semantic transformation across devices and scenarios, avoiding errors and inefficiencies caused by manual annotation and format adaptation, ensuring data semantic uniformity and standardization. After semantic alignment, multimodal data can be efficiently compatible and reused, significantly improving the accuracy and stability of subsequent feature extraction and fault identification, while reducing model deployment and adaptation costs and enhancing the overall solution's engineering feasibility.

[0141] Step S220: Use a wavelet denoising model to denoise the semantically aligned multimodal data to obtain wavelet denoised multimodal data.

[0142] The wavelet denoising model is as follows:

[0143] ;

[0144] in, This is a db4 wavelet transform. It is a soft threshold function. Inverse wavelet transform is used to filter out sea state interference and sensor noise, while retaining the characteristics of weak fault signals. For semantically aligned multimodal data, For wavelet-denoised multimodal data.

[0145] Understandably, employing wavelet denoising models to denoise semantically aligned multimodal data can accurately filter out redundant signals such as sea state interference, sensor noise, and electromagnetic interference generated during ship operation, improving data purity while preserving the original fault characteristics to the maximum extent. Wavelet transform possesses the advantage of time-frequency localization, adapting to the non-stationary and nonlinear characteristics of ship time-series signals, specifically preserving weak early fault features, and preventing weak signals from being submerged by noise. Wavelet denoising causes less damage to fault features and has a more stable denoising effect, significantly improving the data signal-to-noise ratio. The denoised data can effectively reduce the probability of model misjudgment and missed judgment, improving the sensitivity and reliability of fault early warning, and providing crucial support for accurate early fault identification.

[0146] Step S230: Normalize the wavelet denoising multimodal data using a normalization model to obtain normalized multimodal data.

[0147] The normalization model is as follows:

[0148] ;

[0149] in, The maximum value in wavelet-denoised multimodal data. It represents the minimum value in wavelet-denoised multimodal data.

[0150] Understandably, normalizing multimodal data after wavelet denoising using a normalization model can unify the data magnitudes of different physical quantities such as voltage, temperature, vibration, and impedance, eliminating computational biases caused by differences in dimensions and significant numerical amplitudes, and ensuring that all features have equal weight in the model. Normalization can significantly improve the training stability and convergence speed of subsequent feature enhancement, fusion, and early warning models, preventing features with excessively large values ​​from dominating model learning and causing key fault features with small amplitudes to be ignored. Simultaneously, it can reduce the impact of data distribution differences on model accuracy and improve the algorithm's generalization ability under different operating conditions and equipment.

[0151] Step S240: Perform data completion and data removal on the normalized multimodal data to obtain standardized multimodal data.

[0152] Specifically, for missing and outlier values ​​caused by signal interruption, packet loss, or sudden environmental changes during ship sensor data acquisition, the 3σ criterion can be used to detect and remove anomalies in the normalized data. This identifies and removes abnormal sampling points that exceed reasonable fluctuation ranges, preventing interference from affecting subsequent feature extraction and model judgment. Building on this, the K-nearest neighbor algorithm can be used to intelligently complete missing items in the data sequence. Missing values ​​are inferred from valid data at adjacent times and under similar operating conditions, ensuring the continuity and integrity of the data sequence. Through the coordinated processing of anomaly removal and missing data completion, the dataset is further purified and data gaps are repaired. The resulting standardized multimodal data possesses complete, continuous, reliable, and anomaly-free characteristics, stably supporting subsequent feature enhancement, fusion, and fault early warning processes, providing a qualified data foundation for high-precision intelligent early warning.

[0153] In some embodiments of this application, the process of performing feature enhancement on the standardized multimodal data to obtain multidimensional enhanced feature data in step S130 is described. This process may include:

[0154] By using a pre-trained contrastive learning feature enhancement model, the features of each dimension of standardized multimodal data are enhanced to obtain multidimensional enhanced feature data.

[0155] The loss function for training the contrastive learning feature enhancement model is:

[0156] ;

[0157] in, and These are enhanced view features of different dimensions for the same sample. The cosine similarity function is used. For temperature parameters, such as The value is 0.1.

[0158] Specifically, the contrastive learning feature enhancement model can use the SimCLR+MoCo dual-channel feature extractor to construct a feature space alignment and temporal enhancement mechanism, thereby strengthening the fault feature discrimination through contrastive learning and effectively improving the fault identification rate.

[0159] Furthermore, for weak fault precursors in time-series data, variational mode decomposition (VMD) can be used to decompose the signal into multiple intrinsic mode components, extract the components corresponding to the fault features, and reconstruct them. Specifically, the time-series feature data of the multidimensional enhanced feature data can be enhanced using a variational mode decomposition model to obtain multidimensional enhanced feature data after time-series feature enhancement.

[0160] The variational mode decomposition model is as follows:

[0161] ;

[0162] in, For the set of modal components, The first-order partial derivative with respect to time t is used to constrain the bandwidth of the components, ensuring a smooth decomposition result. Let be the center frequency of the k-th modal component. For the Dirac function, Representing the convolution operation, the variational mode decomposition model can effectively capture the weak precursor features of sudden failures such as thermal runaway.

[0163] In some embodiments of this application, the process of fusing the enhanced feature data of each dimension to obtain the fused feature is described, which may include:

[0164] S1. Using a pre-built modal feature weight adaptive allocation model, determine the fusion weight of each dimension of enhanced feature data.

[0165] The modal feature weight adaptive allocation model is as follows:

[0166] ;

[0167] in, This represents the feature corresponding to index k. The time series features are voltage / temperature time series statistics, encoded using Time2Vec, with a dimension of 768. Spatial features (equipment field distribution parameters, encoded via GCN, dimension 768). The semantic features are (knowledge graph mapping features, extracted via BERT-base, with a dimension of 768). Features The corresponding attention weights (adaptively assigned based on the impact of the fault, such as thermal runaway faults) =0.4、 =0.3), This is a feature of fusion.

[0168] S2. Based on the fusion weights of the enhanced feature data of each dimension, fuse the enhanced feature data of each dimension to obtain the fused feature.

[0169] Specifically, based on different fault types and operating scenarios, an adaptive algorithm can dynamically allocate the fusion weights of features across various dimensions, highlighting key features that contribute significantly to fault identification while minimizing redundant and interfering information. The enhanced features of each dimension are then weighted, superimposed, and concatenated into vectors to achieve deep coupling and complementary advantages of multimodal information. This fully integrates the fault characterization capabilities of features from different dimensions, ultimately resulting in a fusion feature that is comprehensive, highly discriminative, and has a high signal-to-noise ratio.

[0170] In some embodiments of this application, the process of dynamic risk prevention and interpretability optimization of the ship system mentioned in the above embodiments is described. Specifically, in terms of interpretability optimization, the contribution of multimodal data can be calculated using the SHAP value calculation model to obtain the contribution of each feature of the multimodal data to the fault response level.

[0171] The SHAP value calculation model is as follows:

[0172] ;

[0173] in, Let be the SHAP value of the i-th feature. For feature set, For feature subset, This is the prediction function.

[0174] Specifically, the XAI interpretable framework can be used to quantify the impact of each feature on the early warning results, visualize the fault propagation path and decision logic, and make the decision logic transparent and more interpretable by integrating SHAP value analysis and causal reasoning decision tree.

[0175] In terms of dynamic risk prevention and control, multi-modal data can be used to optimize the prevention and control of ship systems using a multi-objective ship prevention and control optimization model.

[0176] The optimization model for multi-target ship defense is as follows:

[0177] ;

[0178] in, For safety losses (including failure rate and degree of loss). For safety loss weighting coefficient, Economic costs (including operation and maintenance costs and downtime losses). This is the economic cost weighting coefficient. and Dynamic adjustments based on operating conditions, prioritizing safety in scenarios. It can take the value 0.7. The possible value is 0.3.

[0179] Specifically, the ship multi-target defense optimization model is strengthened through adversarial training. The adversarial training loss function of the ship multi-target defense optimization model is:

[0180] ;

[0181] in, To counteract the total training loss, To predict losses, For the original sample, For adversarial examples, The distance loss between the original sample and the adversarial sample. This is the distance loss weighting coefficient.

[0182] Understandably, adversarial training enhances the robustness and generalization ability of the ship multi-target control optimization model under complex sea conditions and extreme operating conditions, effectively mitigating prediction biases caused by data noise and environmental interference. Adversarial training generates near-realistic disturbance samples, driving the model to learn more robust feature patterns and avoiding overfitting and insufficient scenario adaptability. It also strengthens the model's ability to identify and control abnormal operating conditions and edge failure scenarios, resulting in a more stable trade-off between safety and economy. After adversarial enhancement, the model maintains stable output under diverse operating conditions, including inland waterways and coastal areas, making control decisions more reliable and early warnings more accurate, significantly improving the overall solution's engineering implementation and long-term operational capabilities.

[0183] In some embodiments of this application, considering that the real-time collected data can be used to enhance the training of the model during the fault early warning process of the ship system mentioned in the above embodiments, for any intelligent model, when incremental data greater than the preset data amount is obtained, incremental learning can be performed on the intelligent model through the incremental data to update the intelligent model.

[0184] The incremental learning process is as follows:

[0185] ;

[0186] in, Let be the model parameters of the intelligent model at time t. The updated model parameters are for the intelligent model. These are parameters trained based on incremental data. The learning coefficient is the incremental learning factor.

[0187] Understandably, incremental learning for updating intelligent models allows for lightweight updates using new data after acquiring sufficient incremental data, significantly shortening the model iteration cycle and reducing computational resources and time costs. Incremental learning retains the model's existing mature knowledge while continuously absorbing new operating conditions and fault samples, enabling the model to adapt to complex and ever-changing navigation environments such as inland waterways and coastal areas, and continuously improving the accuracy and generalization ability of fault warnings. This dynamic update mechanism can quickly respond to changes in the distribution of ship operation data, effectively avoiding model aging and accuracy decay, maintaining the accuracy and reliability of the early warning system in the long term, and providing stable and sustainable intelligent support for the safe operation and maintenance of new energy ships throughout their entire life cycle.

[0188] The following embodiments will provide a detailed introduction to the ship fault early warning, response and risk prevention scheme of this application through specific examples.

[0189] In the process of multimodal data perception and cross-domain semantic alignment, model initialization is performed: a BERT-Base model (12-layer Transformer) and a CRF layer are jointly constructed, pre-trained based on 50,000 equipment parameter descriptions and 30,000 fault case texts, with an initial λ=0.4 (balanced). and (Weights), number of semantic tags: 86 categories, number of tags adapted to the dimensions of the transition matrix.

[0190] Iterative solution: Set batch size to 32, learning rate to 0.001, and Adam optimizer (β1=0.9, β2=0.999). Iterate for 500 rounds, validating semantic similarity every 50 rounds. Stop training when the semantic similarity on the validation set is ≥95%. L_align converged to a minimum value of 0.032, satisfying the requirement of semantic conversion accuracy ≥ 92%.

[0191] Spatiotemporal alignment error control: Based on the model solution results, the alignment error of inland waterway vessels is controlled within 0.03s (timestamp synchronization every 100ms), and the error of coastal vessels is controlled within 0.05s through GPS timing + linear interpolation correction, which meets the semantic node association rationality of L_align constraint.

[0192] Wavelet denoising model: For inland waterway conditions, τ = 0.018 * max|W( The decomposition layer is 3 layers. For coastal conditions, τ = 0.022 * max|W( The decomposition process involves 4 layers, and the db4 wavelet transform is implemented using MATLAB. After denoising, the signal-to-noise ratio is ≥35dB, ensuring... Retain weak fault signals.

[0193] Normalization model: Statistical extreme values ​​are calculated according to equipment type, with lithium battery voltage ranging from 2.75V to 4.2V and fuel cell voltage ranging from 0.6V to 1.0V. The values ​​are then substituted into the formula to solve the problem, mapping the data to the [0,1] interval to eliminate the influence of dimensions.

[0194] Data completion and elimination: Outliers were eliminated based on the 3σ criterion (2.8% outlier rate in inland waterways and 3.2% in coastal areas). Missing values ​​were filled in using the KNN algorithm (k=5) to ensure data integrity and provide high-quality data for model input.

[0195] Data Acquisition and Alignment: Based on the model input requirements, real ship time-series data (1kHz sampling frequency), micro-experimental data, and case data were collected. Multi-source data alignment was achieved through timestamp calibration, and a multimodal dataset of 500,000+ samples was constructed, which was divided into training set, validation set, and test set in a 7:2:1 ratio.

[0196] In the process of multimodal feature fusion and early fault extraction, a contrastive learning model is used to solve the problem: (1) SimCLR model: set τ=0.1, feature dimension 128, generate two sample augmentation views, batch size 64, learning rate 0.001, momentum 0.9, SGD optimizer iterates for 300 rounds, when the similarity of similar fault features is ≥85% and the dissimilarity is ≤30%, The training stopped when the convergence reached 0.12. (2) MoCo model: queue size 65536, momentum parameter 0.999, ResNet-50 encoder, initial learning rate 0.03 (decayed by 1 / 10 every 100 rounds), 300 rounds of iteration, fused with SimCLR model to form 256-dimensional feature vector, improving feature discrimination.

[0197] VMD decomposition model solution: For lithium battery temperature and fuel cell impedance time series data, the decomposition layer is set to 4 layers, the penalty factor α=2000, and the maximum number of iterations is 500. Substitute into the VMD formula to solve, screen the fault component with an energy ratio ≥20% to reconstruct the signal. After reconstruction, the correlation coefficient with the original signal is greatly improved, and weak precursor features are captured.

[0198] Solving the attention fusion model: (1) Weight calibration: determined by the analytic hierarchy process. Such as lithium battery thermal runaway ( ); fuel cell membrane rupture ( (2) Feature fusion and dimensionality reduction: Substitute The formula weightedly fuses temporal, spatial, and semantic features, and uses PCA to reduce dimensionality (covariance matrix Σ=1 / N×). T × The 16-dimensional core features with a cumulative contribution rate ≥91.5% and a sum of importance of core features ≥70% were selected. On the test set, the accuracy of weak signal extraction for lithium battery thermal runaway and the accuracy of feature extraction for fuel cell membrane rupture were improved.

[0199] In the process of graded early warning and intelligent emergency response, the following steps were taken: (1) Discrimination network training: CNN+Transformer architecture, input 16-dimensional core features, batch size 32, learning rate 0.001, Adam optimizer, 200 iterations, convergence loss ≤0.02, and test set discrimination accuracy ≥90%. (2) Threshold calibration: Sobol exponential sensitivity analysis (core parameter sensitivity ≥0.2), combined with case calibration thresholds: Level 1 (lithium battery voltage fluctuation 3%-5%, fuel cell membrane impedance 10%-15%), Level 2 (5%-15%, 15%-30%), Level 3 (≥15%, ≥30%), threshold error ≤±0.5%.

[0200] Solving the multi-level response state machine model (three-level response state machine model): (1) Setting the reward function: =0.6×(1-response time / 500ms)+0.4×(1-failure loss rate), optimized using the policy gradient algorithm. 500 iterations Convergence, average response time 480ms (≤500ms). (2) Response process triggering: based on The output state transition logic includes level 1 automatic isolation, level 2 load transfer, and level 3 emergency stop, linked to audible and visual alarms and ship-to-shore data transmission (delay ≤ 100ms).

[0201] Based on this, a digital twin model is solved: COMSOL is used to construct an electro-thermal-mechanical multiphysics model, with the thermal conductivity of the lithium battery k=2.0 (inland waterway) / 1.9 (coastal) W / (m•K) and the membrane corrosion rate of the fuel cell 0.02μm / day. This simulates fault propagation, effectively reduces simulation errors, optimizes emergency response plans, and improves the consistency between simulation and actual ship operating conditions.

[0202] In the process of dynamic risk prevention and interpretability optimization, the SHAP value model is solved: using the Python shap library, the calculation is performed on 1000 to 2000 test samples. Visualize the impact of key factors (lithium batteries: temperature, SEI membrane thickness; fuel cells: salt spray concentration, membrane impedance) to improve consistency with Sobol index analysis.

[0203] Solving the multi-objective ship defense optimization model: (1) Setting the objective function: For coastal and offshore areas, α=0.7 and β=0.3; for inland and nearshore areas, α=0.6 and β=0.4. (2) Genetic algorithm solution: population size 100, iterations 200, crossover probability 0.8, mutation probability 0.05, after optimization This reduces safety losses and economic costs.

[0204] Adversarial training: Adversarial examples are generated using the FGSM algorithm, and the meta-learning fine-tuning step size is 0.005, with 50 iterations. After fine-tuning, the model's accuracy on adversarial examples is improved, thereby enhancing its generalization ability.

[0205] Therefore, based on Output an interpretable policy, and Strengthen and reduce the failure loss rate, and improve the understanding of operation and maintenance personnel.

[0206] Finally, during the model closed-loop update process, incremental learning is performed: incremental learning coefficients. The parameter can be set to 0.8, and the update coefficient to 0.05. 50 sets of experimental data and 120 sets of real-ship data are collected weekly. The top-level parameters are updated using the Finetune incremental method, with each update taking ≤2 hours and requiring no retraining. Through a closed-loop iterative approach, the model parameters and knowledge graph are updated weekly, and hyperparameters are adjusted monthly, accumulating 24 iterations over 6 months. Through continuous optimization, the prediction error (MAE) has been significantly reduced.

[0207] Performance verification: Real-world testing on 8 vessels showed that the key indicators—early warning accuracy rate of 92.3%, response time of 470ms, false alarm rate of 2.1%, false alarm rate of 1.6%, and interpretability score of 0.87—all met the preset requirements. The incidence of core failures was reduced by 23%, and maintenance costs were reduced by 18%.

[0208] The system for realizing ship fault early warning provided in the embodiments of this application is described below. The system for realizing ship fault early warning described below can be referred to in correspondence with the method for realizing ship fault early warning described above.

[0209] See Figure 3 , Figure 3 This is a schematic diagram of a system module structure for realizing ship fault early warning, as disclosed in an embodiment of this application.

[0210] like Figure 3 As shown, the system may include:

[0211] Ship system data acquisition module 11 is used to acquire multimodal data of the ship system;

[0212] The data alignment and preprocessing module 12 is used to perform semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data;

[0213] Feature enhancement module 13 is used to enhance the features of the standardized multimodal data to obtain multidimensional enhanced feature data;

[0214] Feature fusion module 14 is used to fuse the enhanced feature data of each dimension to obtain fused features;

[0215] The fault response level early warning module 15 is used to input the fused features into a pre-trained multi-level response state machine model, output the fault response level, and issue an early warning corresponding to the fault response level.

[0216] Optionally, the data alignment preprocessing module includes:

[0217] The semantic alignment module is used to perform semantic alignment on the multimodal data using a pre-trained cross-ship semantic alignment model to obtain semantically aligned multimodal data. The loss function for training the cross-ship semantic alignment model is:

[0218] ;

[0219] in, Let the loss function be the conditional random field. For real semantic tags, For predicting labels, For the association loss of the pre-built knowledge graph, This refers to the set of nodes in the knowledge graph. This is the set of edges associated with the nodes of the knowledge graph. This is the correlation loss coefficient;

[0220] The wavelet denoising module is used to denoise the semantically aligned multimodal data using a wavelet denoising model to obtain wavelet-denoised multimodal data. The wavelet denoising model is as follows:

[0221] ;

[0222] in, This is a db4 wavelet transform. It is a soft threshold function. This is the inverse wavelet transform. For the semantically aligned multimodal data, For the wavelet-denoised multimodal data;

[0223] The normalization module is used to normalize the wavelet-denoised multimodal data using a normalization model to obtain normalized multimodal data. The normalization model is as follows:

[0224] ;

[0225] in, The maximum value in the wavelet-denoised multimodal data. This is the minimum value in the wavelet-denoised multimodal data;

[0226] The data completion and removal module is used to complete and remove data from the normalized multimodal data to obtain standardized multimodal data.

[0227] Optionally, the feature enhancement module includes:

[0228] The contrastive learning feature enhancement module is used to enhance the features of each dimension of the standardized multimodal data using a pre-trained contrastive learning feature enhancement model, resulting in multidimensional enhanced feature data. The loss function for training the contrastive learning feature enhancement model is:

[0229] ;

[0230] in, and These are enhanced view features of different dimensions for the same sample. The cosine similarity function is used. This refers to the temperature parameter.

[0231] Optionally, the system may also include:

[0232] The variational mode decomposition module is used to enhance the temporal feature data of the multidimensional enhanced feature data through a variational mode decomposition model, thereby obtaining the enhanced multidimensional enhanced feature data. The variational mode decomposition model is as follows:

[0233] ;

[0234] in, For the set of modal components, Let be the center frequency of the k-th modal component. For the first-order partial derivative with respect to time t, For the Dirac function, This indicates a convolution operation.

[0235] Optionally, the feature fusion module includes:

[0236] The fusion weight determination module is used to determine the fusion weight of each dimension of the enhanced feature data using a pre-built modal feature weight adaptive allocation model. The modal feature weight adaptive allocation model is as follows:

[0237] ;

[0238] in, This represents the feature corresponding to index k. As a time series feature, For spatial features, For semantic features, Features The corresponding attention weights Features of fusion;

[0239] The fusion module is used to fuse the enhanced feature data of each dimension based on the fusion weights of each dimension to obtain fused features.

[0240] Optionally, the optimization process for the model parameters of the multi-level response state machine model is as follows:

[0241] ;

[0242] in, In response to the trajectory, For the trajectory probability distribution, For the reward function, These are the model parameters before adjustment of the multi-level response state machine model. The adjusted model parameters are those for the multi-level response state machine model.

[0243] Optionally, the system may also include:

[0244] The SHAP value calculation module is used to calculate the contribution of the multimodal data using the SHAP value calculation model, thereby obtaining the contribution of each feature of the multimodal data to the fault response level. The SHAP value calculation model is as follows:

[0245] ;

[0246] in, Let be the SHAP value of the i-th feature. For feature set, For feature subset, This is the prediction function.

[0247] Optionally, the system may also include:

[0248] The prevention and control optimization module is used to optimize the ship system's prevention and control based on the multimodal data using a ship multi-objective prevention and control optimization model. The ship multi-objective prevention and control optimization model is as follows:

[0249] ;

[0250] in, For safety losses, For safety loss weighting coefficient, For economic costs, This is the economic cost weighting coefficient;

[0251] The ship multi-target defense optimization model is enhanced through adversarial training, and the adversarial training loss function of the ship multi-target defense optimization model is:

[0252] ;

[0253] in, To counteract the total training loss, To predict losses, For the original sample, For adversarial examples, The distance loss between the original sample and the adversarial sample is... This is the distance loss weighting coefficient.

[0254] Optionally, the system may also include:

[0255] An incremental learning module is used to perform incremental learning on any intelligent model when more than a preset amount of incremental data is obtained, in order to update the intelligent model. The incremental learning process is as follows:

[0256] ;

[0257] in, Let be the model parameters of the intelligent model at time t. The updated model parameters for the intelligent model. The parameters are trained based on the incremental data. The learning coefficient is the incremental learning factor.

[0258] 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.

[0259] 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.

[0260] 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 early warning of ship malfunctions, characterized in that, include: Acquire multimodal data of the ship system; The multimodal data is semantically aligned and preprocessed to obtain standardized multimodal data; The standardized multimodal data is enhanced to obtain multidimensional enhanced feature data; By fusing the enhanced feature data from each dimension, a fused feature is obtained; The fused features are input into a pre-trained multi-level response state machine model, which outputs a fault response level and issues an early warning corresponding to the fault response level.

2. The method according to claim 1, characterized in that, The multimodal data undergoes semantic alignment and data preprocessing to obtain standardized multimodal data, including: The multimodal data is semantically aligned using a pre-trained cross-ship semantic alignment model to obtain semantically aligned multimodal data. The loss function for training the cross-ship semantic alignment model is as follows: ； in, Let the loss function be the conditional random field. For real semantic tags, For predicting labels, For the association loss of the pre-built knowledge graph, This refers to the set of nodes in the knowledge graph. This is the set of edges associated with the nodes of the knowledge graph. This is the correlation loss coefficient; The semantically aligned multimodal data is denoised using a wavelet denoising model to obtain wavelet-denoised multimodal data. The wavelet denoising model is as follows: ； in, This is a db4 wavelet transform. It is a soft thresholding function. This is the inverse wavelet transform. For the semantically aligned multimodal data, For the wavelet-denoised multimodal data; The wavelet-denoised multimodal data is normalized using a normalization model to obtain normalized multimodal data. The normalization model is as follows: ； in, The maximum value in the wavelet-denoised multimodal data. This is the minimum value in the wavelet-denoised multimodal data; The normalized multimodal data is then augmented and purged to obtain standardized multimodal data.

3. The method according to claim 1, characterized in that, The standardized multimodal data is subjected to feature enhancement to obtain multidimensional enhanced feature data, including: The standardized multimodal data is enhanced using a pre-trained contrastive learning feature enhancement model to obtain multidimensional enhanced feature data. The loss function for training the contrastive learning feature enhancement model is: ； in, and These are enhanced view features of different dimensions for the same sample. The cosine similarity function is used. This refers to the temperature parameter.

4. The method according to claim 3, characterized in that, Also includes: The temporal feature data of the multidimensional enhanced feature data is enhanced by using a variational mode decomposition model to obtain the time-series enhanced multidimensional enhanced feature data. The variational mode decomposition model is as follows: ； in, For the set of modal components, Let be the center frequency of the k-th modal component. For the first-order partial derivative with respect to time t, For the Dirac function, This indicates a convolution operation.

5. The method according to claim 1, characterized in that, The fusion of the enhanced feature data from each dimension yields fused features, including: Using a pre-built adaptive allocation model for modal feature weights, the fusion weights for each dimension of the enhanced feature data are determined. The adaptive allocation model for modal feature weights is as follows: ； in, This represents the feature corresponding to index k. As a time series feature, For spatial features, For semantic features, Features The corresponding attention weights Features of fusion; Based on the fusion weights of the enhanced feature data in each dimension, the enhanced feature data in each dimension are fused to obtain fused features.

6. The method according to claim 1, characterized in that, The optimization process for the model parameters of the multi-level response state machine model is as follows: ； in, In response to the trajectory, For the trajectory probability distribution, For the reward function, These are the model parameters before adjustment of the multi-level response state machine model. The adjusted model parameters are those for the multi-level response state machine model.

7. The method according to claim 1, characterized in that, Also includes: The contribution of each feature of the multimodal data to the fault response level is calculated using the SHAP value calculation model. The SHAP value calculation model is as follows: ； in, Let be the SHAP value of the i-th feature. For feature set, For feature subset, This is the prediction function.

8. The method according to claim 1, characterized in that, Also includes: Based on the multimodal data, the ship system is optimized for prevention and control using a ship multi-objective prevention and control optimization model, which is as follows: ； in, For safety losses, For safety loss weighting coefficient, For economic costs, This is the economic cost weighting coefficient; The ship multi-target defense optimization model is enhanced through adversarial training, and the adversarial training loss function of the ship multi-target defense optimization model is: ； in, To counteract the total training loss, To predict losses, For the original sample, For adversarial examples, The distance loss between the original sample and the adversarial sample is the loss function. This is the distance loss weighting coefficient.

9. The method according to any one of claims 1-8, characterized in that, Also includes: For any intelligent model, when incremental data exceeding a preset amount is obtained, incremental learning is performed on the intelligent model using the incremental data to update the intelligent model. The incremental learning process is as follows: ； in, Let be the model parameters of the intelligent model at time t. The updated model parameters for the intelligent model. The parameters are trained based on the incremental data. The learning coefficient is the incremental learning factor.

10. A ship fault early warning system, characterized in that, include: The ship system data acquisition module is used to acquire multimodal data of the ship system; The data alignment and preprocessing module is used to perform semantic alignment and data preprocessing on the multimodal data to obtain standardized multimodal data; The feature enhancement module is used to enhance the features of the standardized multimodal data to obtain multidimensional enhanced feature data; The feature fusion module is used to fuse the enhanced feature data from each dimension to obtain fused features; The fault response level early warning module is used to input the fused features into a pre-trained multi-level response state machine model, output the fault response level, and issue an early warning corresponding to the fault response level.

Images