A ship energy efficiency abnormal cross-modal traceability attribution method and system based on a structural causal model

Abstract

This invention discloses a cross-modal attribution method and system for ship energy efficiency anomalies based on a structural causal model, relating to the field of intelligent ship energy efficiency monitoring and anomaly causal attribution technology. The method includes: collecting multimodal ship energy efficiency monitoring data and mapping it to a multimodal operating condition heterogeneous graph; performing posterior completion of missing nodes based on structural equation constraints and historical samples of the same operating conditions; performing online Bayesian structural updates using speed, ballast draft, and sea state as trigger conditions to obtain an adaptive directed acyclic graph; applying causal intervention to the main engine load rate deviation and propeller propulsion efficiency deviation on the adaptive directed acyclic graph, calculating the causal contribution rate of each variable to the overall ship energy efficiency deviation, and generating cross-modal attribution results. This invention can improve the accuracy, interpretability, and targeted operation and maintenance handling of ship energy efficiency anomaly root cause localization.

Description

Technical Field

[0001] This invention relates to the field of intelligent energy efficiency monitoring and anomaly causal attribution technology for ships, and in particular to a cross-modal attribution method and system for ship energy efficiency anomalies based on a structural causal model. Background Technology

[0002] With the increasing demand for green shipping, intelligent operation and maintenance, and energy efficiency compliance management, ship energy consumption prediction, power system anomaly monitoring, and operational status diagnosis technologies are gradually evolving from traditional experience-based threshold judgments to intelligent analysis methods based on multi-source sensor data, machine learning models, and multimodal information fusion. Existing technologies typically predict and identify ship energy consumption levels or equipment anomalies by collecting data such as main engine speed, fuel flow, weather conditions, speed, and control parameters, which can improve the automation level of ship operation monitoring to some extent. However, ship energy efficiency anomalies are often not caused by a single variable, but rather by the coupling of multiple factors such as main engine load, propeller propulsion efficiency, sea state disturbances, ballast draft, and engine room equipment status. Existing methods mostly focus on "predictive results" or "anomaly alarms," ​​lacking sufficient understanding of the causal dependencies between cross-modal data, anomaly propagation paths, and hierarchical explanations of root variables. This makes it difficult to meet the needs for interpretable tracing, operation and maintenance priority ranking, and analysis of the contribution of overall ship fuel consumption deviations in ship energy efficiency anomaly scenarios.

[0003] CN118364945A discloses a ship energy consumption prediction method based on a lightweight model. This method acquires historical navigation data, including meteorological, operational, and energy consumption data, constructs a training sample set, and trains a lightweight convolutional neural network model to obtain the energy consumption prediction results for the target ship. This method primarily addresses the problem of low training efficiency in energy consumption prediction models, improving the overall efficiency of ship energy consumption prediction. However, its core remains based on numerical energy consumption prediction using historical samples, without structured modeling of the causal relationships between the main engine, propeller, navigation conditions, and visual status. When deviations occur between predicted and actual energy consumption, this method struggles to determine whether the deviation originates from abnormal main engine load, decreased propulsion efficiency, sea state changes, or abnormal engine room equipment status. Furthermore, it cannot provide causal propagation paths at the component, system, or ship-wide levels.

[0004] CN120597152A discloses a method for monitoring anomalies in ship propulsion systems. This method integrates multi-source, multi-modal monitoring status information of the ship propulsion system and combines 3D models, simulation models, signal processing and transformation techniques, and CLIP large-scale models to achieve anomaly monitoring and alarm push notifications for ship propulsion systems under conditions of limited anomaly data. This method improves the intelligence level of propulsion system anomaly monitoring and can generate anomaly descriptions based on image and text binding. However, its focus is on anomaly state identification and alarm information push notifications. It has not yet established a structural causal model for ship energy efficiency anomalies, lacks a processing mechanism for adaptive updating of the causal structure under changes in operating conditions such as speed, ballast draft, and sea state, and does not intervene to calculate the causal contribution rate of anomaly variables to the overall ship energy efficiency deviation. Therefore, it is difficult to support cross-modal attribution for fuel consumption deviations.

[0005] CN121936608A discloses a method, apparatus, and computer device for tracing the root causes of anomalies using a multi-causal model integration approach. This method trains at least three causal inference models employing different prediction functions on multivariate time-series data under abnormal conditions in industrial equipment. It then determines a unified root cause score vector based on multiple root cause score vectors to identify root cause variables. While this method improves the efficiency and accuracy of identifying the root causes of anomalies in industrial equipment, its application is limited to general industrial equipment. It does not incorporate cross-modal information features such as navigation GIS trajectories, engine room CCTV images, sensor time-series data, and navigation operating condition indicators in ship energy efficiency anomalies. Furthermore, it does not construct a hierarchical attribution logic around the physical correlation between main engine load rate deviation, propeller propulsion efficiency deviation, and overall ship energy efficiency deviation values. Therefore, it is difficult to directly apply to the explanation of multi-level propagation paths and the generation of maintenance and handling priorities for ship energy efficiency anomalies.

[0006] Therefore, existing technologies for ship energy consumption prediction, power system anomaly monitoring, and industrial equipment root cause tracing still suffer from problems such as insufficient expression of causal relationships in cross-modal data, difficulty in adaptively updating the causal structure under changing operating conditions, lack of interpretable and quantifiable contribution of anomalies, and difficulty in directly serving operation and maintenance. This invention provides a cross-modal attribution method and system for ship energy efficiency anomalies based on a structural causal model. It maps multimodal performance monitoring data into a multimodal heterogeneous graph of operating conditions, uses structural equation constraints and historical samples of the same operating conditions to complete the posterior completion of missing nodes, and performs online Bayesian structure updates based on navigation operating condition indicators. On an adaptive directed acyclic graph, the do operator is used to calculate the causal contribution rate of variables to the overall ship energy efficiency deviation, thereby effectively solving the problems of difficulty in cross-modal tracing of ship energy efficiency anomalies, difficulty in hierarchical interpretation, and difficulty in quantifying root cause contributions. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0008] In view of the problems of existing ship energy consumption prediction, power system anomaly monitoring and industrial equipment root cause tracing technologies, such as insufficient expression of causal relationships in cross-modal data, difficulty in adaptively updating the causal structure under changing operating conditions, lack of interpretable quantification of the contribution of abnormal root causes, and difficulty in directly supporting operation and maintenance decision-making with tracing results, this invention is proposed.

[0009] Therefore, the problem to be solved by this invention is how to construct a causal structure that can adaptively update with changes in navigation conditions based on multimodal performance monitoring data such as sensor time-series data, navigation GIS trajectory data, and engine room CCTV equipment status image data under complex ship navigation conditions, and further explain, trace and quantify the impact of variables such as main engine load, propeller propulsion efficiency and fuel consumption on the abnormal energy efficiency of the whole ship.

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model, comprising, Collect multimodal performance monitoring data of ships, construct an initial directed acyclic graph, and map the multimodal performance monitoring data into the temporal nodes, spatial edge weights and visual feature vectors of the initial directed acyclic graph to obtain a multimodal operating condition heterogeneous graph; Using the directed dependencies of the initial directed acyclic graph as structural equation constraints and historical samples of the same working condition as priors, posterior inference is performed on the missing nodes in the multimodal working condition heterogeneous graph, and the inferred expected values ​​are backfilled into the corresponding nodes to obtain the complete multimodal working condition heterogeneous graph. Using navigation condition indicators consisting of speed, ballast draft, and sea state as trigger conditions, online Bayesian structure update is performed on the completed multimodal heterogeneous diagram to obtain an adaptive directed acyclic graph. Based on the conditional probability weights of the adaptive directed acyclic graph, the observation deviations of the main engine speed and propeller torque in the sensor time series data are converted into the deviations of the main engine load rate and propeller propulsion efficiency. The overall ship energy efficiency deviation value is defined by the ratio of fuel flow deviation to rated fuel consumption rate. The do operator intervention is applied to the deviations of the main engine load rate and propeller propulsion efficiency on the adaptive directed acyclic graph. The causal contribution rate of each variable to the overall ship energy efficiency deviation value is calculated. The causal propagation path is decomposed at three levels of granularity: component level, system level, and ship level to obtain a hierarchical attribution result set. The hierarchical attribution results are aggregated into causal propagation paths at each level and transformed into structured attribution reports. The operation and maintenance priorities are determined from high to low based on the causal contribution rate, and cross-modal tracing attribution results are output.

[0011] Secondly, embodiments of the present invention provide a cross-modal attribution system for ship energy efficiency anomalies based on a structural causal model, comprising: The multimodal graph construction module is used to collect multimodal performance monitoring data of ships, construct an initial directed acyclic graph, and map the multimodal performance monitoring data into the temporal nodes, spatial edge weights and visual feature vectors of the initial directed acyclic graph to obtain a multimodal operating condition heterogeneous graph. The node completion module is used to perform posterior inference on missing nodes in the multimodal heterogeneous graph, with the directed dependencies of the initial directed acyclic graph as structural equation constraints and historical samples of the same working condition as priors. The inferred expected values ​​are then used to fill in the corresponding nodes to obtain a completed multimodal heterogeneous graph. The structure update module is used to perform online Bayesian structure update on the completed multimodal heterogeneous graph, using navigation condition indicators consisting of speed, ballast draft, and sea state as trigger conditions, to obtain an adaptive directed acyclic graph. The causal attribution module, based on the conditional probability weights of an adaptive directed acyclic graph, converts the observation deviations of the main engine speed and propeller torque in the sensor time series data into the deviations of the main engine load rate and propeller propulsion efficiency. It defines the overall ship energy efficiency deviation value as the ratio of fuel flow deviation to rated fuel consumption rate. It applies the do operator intervention to the main engine load rate deviation and propeller propulsion efficiency deviation on the adaptive directed acyclic graph, calculates the causal contribution rate of each variable to the overall ship energy efficiency deviation value, and decomposes the causal propagation path at three levels of granularity: component level, system level, and ship level to obtain a hierarchical attribution result set. The results output module is used to transform the causal propagation paths of each layer in the hierarchical attribution results set into a structured attribution report, determine the priority of operation and maintenance handling according to the causal contribution rate from high to low, and output cross-modal tracing attribution results.

[0012] Thirdly, embodiments of the present invention provide 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 any step of the above-described cross-modal attribution method for ship energy efficiency anomalies based on structural causal models.

[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the above-described cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model.

[0014] Compared with existing technologies, the advantages of this invention are as follows: By collecting sensor time-series data, GIS trajectory data, and engine room CCTV image data, and mapping them to time-series nodes, spatial edge weights, and visual feature vectors, a multimodal heterogeneous graph of operating conditions is formed. This allows dispersed power, navigation, and equipment status information to be expressed under a unified causal graph structure, avoiding biased anomaly judgments caused by relying solely on single sensor data. By using the directed dependencies of the initial directed acyclic graph as structural equation constraints, and combining historical samples of the same operating conditions to perform posterior extrapolation and backfilling of missing nodes, the attribution bias caused by sensor packet loss, image occlusion, or data asynchrony under complex ship operating conditions is reduced, improving the data integrity and stability of subsequent causal calculations. By triggering online Bayesian structure updates using navigation operating condition indicators composed of speed, ballast draft, and sea state level, the causal graph can adaptively adjust its topology according to changes in navigation operating conditions. The structure and conditional probability weights prevent causal mismatches in fixed models when sea state, load, or speed changes. By converting observed deviations such as main engine speed, propeller torque, and fuel flow into main engine load rate deviation, propeller propulsion efficiency deviation, and overall ship energy efficiency deviation values ​​on an adaptive directed acyclic graph, and applying do operator intervention to key variables, it can distinguish between correlation anomalies and true causal anomalies. It decomposes causal propagation paths at the component, system, and ship levels, elevating the discovery of energy efficiency anomalies to locating the source and propagation chain of anomalies. By transforming hierarchical attribution results into structured attribution reports and determining maintenance priorities based on causal contribution rates, it can provide interpretable results with correlated modal evidence, fuel consumption deviation contribution, and response order for crew members or shore-based maintenance platforms, thereby improving the accuracy, interpretability, and maintenance efficiency of ship energy efficiency anomaly diagnosis. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 The flowchart shows a cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model. Figure 2 This is a structural diagram of a cross-modal attribution system for ship energy efficiency anomalies based on a structural causal model. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] As mentioned in the background section, while existing technologies can predict ship energy consumption, identify power system anomalies, or perform root cause analysis on anomalies in general industrial equipment, most methods remain at the level of anomaly detection, energy consumption prediction, or root cause variable scoring. They fail to fully integrate navigation conditions such as ship speed, ballast draft, and sea state to dynamically model the causal dependencies between different modal data. Furthermore, existing methods lack mechanisms for applying causal intervention to anomalous variables and calculating their contribution to overall ship energy efficiency deviations. This results in outputs that fail to clearly indicate which component the anomaly originates from, through which system path it propagates, and ultimately, how much fuel consumption deviation it causes. To address these issues, this invention provides a cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model.

[0020] Reference Figures 1-2 , Figure 1 This is a flowchart illustrating a cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model, according to an embodiment of the present invention. Figure 1 As shown, a cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model includes: S1: Collect multimodal performance monitoring data of ships, construct an initial directed acyclic graph, and map the multimodal performance monitoring data into the temporal nodes, spatial edge weights and visual feature vectors of the initial directed acyclic graph to obtain a multimodal operating condition heterogeneous graph; S1.1: Collect multimodal performance monitoring data of the ship, including sensor time-series data, navigation GIS trajectory data and engine room CCTV equipment status image data; sensor time-series data includes main engine speed, propeller torque, fuel flow and compartment temperature; Specifically, sensor time-series data is continuously collected by physical sensors deployed in the main engine, shafting, and cabins at a fixed sampling period; navigation GIS trajectory data is collected by the shipborne GNSS positioning device according to the navigation update rate, recording the latitude and longitude coordinates, ground speed, and heading angle at each sampling time; engine room CCTV equipment status image data is continuously collected by fixed-point camera units in the engine room at a frame rate, with each frame image having an additional collection timestamp. Furthermore, time-stamp alignment is performed on the sensor time-series data, navigation GIS trajectory data, and cabin CCTV equipment status image data. The sampling period of the sensor time-series data is used as the reference time granularity. The positioning points of the navigation GIS trajectory data are interpolated to the reference time granularity. Keyframes are extracted from the cabin CCTV equipment status image data according to the reference time granularity. After time-stamp alignment, the three data streams correspond one-to-one on the time axis, forming a synchronous multimodal performance monitoring data sequence. It should be noted that the reference time granularity is determined based on the sampling period of each modality data. The reference time granularity is not less than the sampling period of the sensor with the lowest sampling rate in the sensor timing data. When the sampling period of some sensor channels is less than the reference time granularity, the multiple sampled values ​​of these sensor channels within the reference time granularity are aggregated according to the mean, maximum, minimum or first-order difference mean to avoid introducing false interpolation that exceeds the sensor resolution into the low sampling rate data during the time alignment process. S1.2: The navigation condition indicators are composed of speed, ballast draft, and sea state level, and the observations corresponding to the navigation condition indicators in the synchronous multimodal performance monitoring data sequence are extracted: the ground speed component of the navigation GIS trajectory data is used as the observed value of speed, the reading of the pre-loaded draft sensor is used as the observed value of ballast draft, and the meaningful wave height output by the shipborne meteorological instrument is mapped to the discrete value of sea state level according to the predetermined classification rule. Furthermore, based on the range of values ​​for navigation condition indicators, the multidimensional observation space of navigation condition indicators is divided into several condition intervals. The division method is as follows: the observed values ​​of speed are segmented at equal intervals according to a preset step size; the observed values ​​of ballast draft are discretized into three levels: full load, half load, and ballast; and the discrete values ​​of sea state level are directly used as the segment boundaries. The Cartesian product of each dimension segment constitutes a set of condition intervals, and each condition interval corresponds to a typical navigation condition. S1.3: Using navigation condition indicators as conditional variables, based on prior causal knowledge of ship power transmission, predefine the directed dependencies between variables and construct an initial directed acyclic graph; Specifically, the initial set of nodes in the directed acyclic graph includes operating condition nodes, sensor timing nodes, navigation status nodes, and equipment status nodes. Among them, operating condition nodes include speed nodes, ballast draft nodes, and sea state level nodes; sensor timing nodes include main engine speed nodes, propeller torque nodes, fuel flow nodes, and compartment temperature nodes; navigation status nodes include segment status nodes and ground speed nodes; and equipment status nodes include the equipment nodes corresponding to each monitored device in the engine room CCTV equipment status image data. Furthermore, the set of directed dependencies in the initial directed acyclic graph is determined based on prior causal knowledge of ship power transmission, including: the main engine speed node points to the propeller torque node, the propeller torque node points to the fuel flow node, the fuel flow node points to the ground speed node, the compartment temperature node points to the corresponding equipment node, and the speed node, ballast draft node, and sea state level node point to the main engine speed node, propeller torque node, and fuel flow node, respectively, to characterize the modulation effect of navigation conditions on the propulsion chain energy efficiency state; the above set of directed dependencies does not constitute a directed loop in the graph structure and satisfies the topological constraints of the directed acyclic graph; Furthermore, the edge types in the initial directed acyclic graph include causal dependency edges and spatial segment edges; causal dependency edges are used to store the conditional probability weights between variables and participate in structural equation constraints, Bayesian structure updates, and do operator intervention calculations; spatial segment edges are used to store the spatial edge weights formed by the geographical distance between adjacent segments, the mean ground speed, and the rate of change of heading angle, and do not participate in the conditional probability weight updates of causal dependency edges. S1.4: For the sensor time series data in the synchronous multimodal performance monitoring data sequence, perform sliding window feature extraction on the time series samples of main engine speed, propeller torque, fuel flow and cabin temperature in a preset time window, and map the samples of each sensor channel in the time window into statistical feature vectors, which are used as the node features of the corresponding time series nodes in the initial directed acyclic graph. Preferably, the statistical feature vector includes the mean, standard deviation, maximum value, minimum value, and first-order difference mean of the observations of each sensor channel within the time window. The above five statistical dimensions are concatenated to form the node feature vector of the time series node. z-score normalization is performed on the node feature vector of the time series node: using the mean and standard deviation of the statistical features of historical samples under the same operating conditions within the time window as normalization parameters, the node feature vector of the current time series node is normalized dimension by dimension. The normalization parameters are stored separately according to the operating condition interval to which the navigation operating condition index belongs, so as to eliminate the influence of the sensor dimension difference on the feature vector amplitude under different operating condition intervals. S1.5: For the navigation GIS trajectory data in the synchronous multimodal performance monitoring data sequence, the continuous positioning point sequence is divided into adjacent segments according to the reference time granularity. The geographical distance, average ground speed and heading angle change rate of each adjacent segment are calculated as three-dimensional spatial attribute vectors. The scalar value of the three-dimensional spatial attribute vector after weighted summation is used as the spatial edge weight of the corresponding spatial segment edge in the multimodal heterogeneous diagram. Specifically, the geographical distance is calculated from the latitude and longitude coordinates of two adjacent positioning points using the Haversine formula. The average ground speed is the average of the observed ground speed values ​​at corresponding times for two adjacent positioning points. The rate of change of the heading angle is the ratio of the difference in the heading angles of two adjacent positioning points to the reference time granularity. The weighting coefficients of the weighted summation are preset based on the degree of influence of each spatial attribute on the ship's propulsion load, with the average ground speed having the largest weighting coefficient and the rate of change of the heading angle having the smallest weighting coefficient. Furthermore, maximum-minimum normalization is performed on the spatial edge weights, using the maximum and minimum values ​​of the spatial edge weights of each adjacent flight segment in the historical samples under the same working conditions as the normalization boundary, and mapping the spatial edge weights to the interval [0,1], so that the spatial edge weights are comparable under different working condition intervals. S1.6: For each key frame of the cabin CCTV equipment status image data in the synchronous multimodal performance monitoring data sequence, a pre-trained convolutional neural network is used to perform feature extraction on the key frame. The output vector of the global average pooling layer of the pre-trained convolutional neural network is taken as the visual feature vector of the key frame, and the visual feature vector is written into the node feature domain of the corresponding equipment node in the initial directed acyclic graph. Furthermore, the pre-trained convolutional neural network uses a database of cabin equipment status sample images as its training set and equipment operating status category labels as its supervision signals. After supervised training, the output vector of the global average pooling layer of the pre-trained convolutional neural network has the ability to express the discriminative features of the equipment status. The dimension of the visual feature vector is consistent with the output dimension of the global average pooling layer of the pre-trained convolutional neural network. L2 norm normalization is performed on the visual feature vector, dividing the visual feature vector by its L2 norm so that the normalized visual feature vector lies on the unit hypersphere. This eliminates the influence of differences in imaging conditions such as image brightness and shooting distance on the amplitude of the visual feature vector. The visual feature vectors of different equipment nodes in the initial directed acyclic graph are comparable in the sense of cosine similarity after normalization. S1.7: Integrate the node feature vectors, spatial edge weights, and visual feature vectors of the time-series nodes according to the topological structure of the initial directed acyclic graph to obtain a multimodal heterogeneous graph. Specifically, the node set of the multimodal heterogeneous graph is the same as the node set of the initial directed acyclic graph; for each node belonging to the sensor channel in the multimodal heterogeneous graph, the node feature vector of the corresponding time-series node is used as the node feature; for each node belonging to the equipment monitoring in the multimodal heterogeneous graph, the corresponding visual feature vector is used as the node feature; the set of directed dependencies in the multimodal heterogeneous graph is the same as the set of directed dependencies in the initial directed acyclic graph, and each directed dependency is assigned a corresponding spatial edge weight as an edge attribute. The spatial edge weight of the directed dependency between flight segment nodes is obtained from S1.5, and the edge weight of the other directed dependencies is initialized to 1; Furthermore, the current observation values ​​of navigation condition indicators are added to the multimodal heterogeneous map as map-level attributes to record the operating condition intervals corresponding to the multimodal heterogeneous map. It should be noted that the dimension of the node feature vector in the multimodal heterogeneous graph varies depending on the node type. The dimension of the node feature vector of the sensor channel node is determined by the statistical feature dimension in S1.4, and the dimension of the node feature vector of the equipment monitoring node is determined by the pre-trained convolutional neural network structure in S1.6. The multimodal heterogeneous graph is stored in a heterogeneous graph data structure, and the node type field and edge type field record the modal origin of each node and each directed dependency relationship, respectively. S2: Using the directed dependencies of the initial directed acyclic graph as structural equation constraints and historical samples of the same working condition as priors, perform posterior inference on the missing nodes in the multimodal working condition heterogeneous graph, and backfill the corresponding nodes with the inferred expected values ​​to obtain the complete multimodal working condition heterogeneous graph. S2.1: Perform missing feature vector detection on the node feature vectors of each node in the heterogeneous graph of multimodal operating conditions; Furthermore, missing node detection includes determining whether the node feature domain is null or below the effective signal threshold on a node-by-node basis; if the node feature domain is null or below the effective signal threshold, the node is marked as a missing node; otherwise, the node is marked as an observable node. Specifically, the effective signal thresholds are set according to the modal type of the node: for time-series nodes belonging to sensor channels, the lower limit of the range of the sensor channel and the sensor self-test fault code are used as the effective signal thresholds; for visual feature vector nodes belonging to equipment monitoring, the lower limit of the L2 norm of the output vector of the pre-trained convolutional neural network is used as the effective signal threshold; for spatial edge weights belonging to flight segment positioning, the distance between positioning points in the flight GIS trajectory data exceeding the maximum credible jump distance is used as the effective signal threshold. Furthermore, all missing nodes are classified according to their modal origin: missing nodes belonging to sensor time-series data channels are assigned to the sensor modal missing set, and missing nodes belonging to cabin CCTV equipment status image data channels are assigned to the visual modal missing set; the union of the sensor modal missing set and the visual modal missing set constitutes the missing node set, which is processed by S2.2 to S2.5 according to modal type. It should be noted that the navigation GIS trajectory data is continuous at the reference time granularity after timestamp interpolation. Its missing data is manifested as abnormal spatial edge weights rather than missing node features. Therefore, the missing data of navigation GIS trajectory data is replaced by the detection of empty values ​​in node feature domains by detecting the validity of spatial edge weights. Spatial edge weights that are detected as abnormal are also included in the set of missing nodes. In S2.5, the weighted interpolation of adjacent valid spatial edge weights is used as the posterior expectation value. S2.2: Using the directed dependencies of the initial directed acyclic graph as constraints for the structural equation, construct a structural equation for each missing node in the heterogeneous graph of the multimodal working conditions, with the set of its parent nodes as the independent variable. Furthermore, for any missing node in the heterogeneous graph of multimodal operating conditions, the set of parent nodes of the missing node is located in the initial directed acyclic graph. The node feature vector of the missing node is represented as a linear combination of the node feature vectors of its parent node set plus a residual term. The linear combination coefficients and the residual term constitute the structural equation parameters of the missing node. The structural equation parameters are initially estimated using the ordinary least squares regression results of the observed values ​​of the missing node and its parent node set in historical samples of the same operating conditions. These parameters are then used as prior parameters in the posterior extrapolation of S2.5 for updating. The form of the structural equation is as follows: ; in, Let be the variable to be calculated for the i-th missing node. Let be the vector of observable variables corresponding to the set of parent nodes of the i-th missing node in the initial directed acyclic graph. This is a mapping function parameterized based on the initial conditional probability weights of the corresponding directed edges in the initial directed acyclic graph. For the residual terms of the structural equations; Specifically, for cases where there are multiple levels of missing nodes in the set of missing nodes (i.e., the set of parent nodes of a missing node also contains missing nodes), the posterior inference order is determined according to the topological sort of the initial directed acyclic graph: posterior inference is performed first on missing nodes whose parent nodes are all observable nodes, and the expected value is temporarily written into the node. Then, posterior inference is performed on missing nodes whose parent nodes contain the aforementioned inferred nodes. This process is advanced layer by layer according to the topological sort until all nodes in the set of missing nodes have undergone posterior inference. Preferably, the distribution of the residual terms in the structural equation is assumed to be normally distributed, the mean of the residual terms is set to zero, and the variance is estimated by the sample variance of the structural equation regression residuals of the missing node in the historical samples of the same working conditions. The posterior inference under the normal distribution assumption is a closed-form solution, which does not require numerical integration and is suitable for online scenarios with limited onboard computing resources. S2.3: Using the current observed value of the navigation condition index as an index, retrieve historical samples belonging to the same operating condition interval from the historical sample database, and record the retrieved historical sample set as the historical same operating condition sample set; Furthermore, the attribution of the operating condition interval is determined based on the boundary of the operating condition interval: each dimension component of the current observation value of the navigation operating condition index is compared with the boundary of the operating condition interval to determine the operating condition interval number to which it belongs. All samples in the historical sample library whose operating condition interval numbers are consistent with the above operating condition interval numbers are retrieved as the historical same operating condition sample set; each sample in the historical same operating condition sample set records the complete node feature vector observation value of all nodes of the multimodal operating condition heterogeneous graph at that moment. Furthermore, for the historical observation sequence of each missing node in the historical working condition sample set, the mean vector and covariance matrix are calculated, and the multivariate normal distribution parameterized by the mean vector and covariance matrix is ​​used as the prior distribution of the missing node. The mean vector and covariance matrix of the prior distribution are automatically updated as the historical working condition sample set is expanded. New historical samples update the mean vector and covariance matrix incrementally, without having to retrace the entire historical sample database. It should be noted that if the number of historical samples retrieved from the historical sample set of the same working conditions is less than the preset minimum sample size threshold, the global statistic of the full historical sample database is used as the backoff value of the prior distribution parameter. At the same time, the diagonal elements in the covariance matrix of the prior distribution are increased to reflect the decrease in prior confidence. The preset minimum sample size threshold is determined based on the minimum sample size requirement for the invertibility of the covariance matrix of the multivariate normal distribution. It is usually required that the sample size is at least 5 to 10 times the dimension of the node feature vector to ensure that the estimation of the covariance matrix in the prior distribution is non-singular and stable. S2.4: For all observable nodes in the heterogeneous graph of multimodal operating conditions, under the constraints of the structural equation, calculate the residuals of the observed values ​​of the node eigenvectors of each observable node relative to the predicted values ​​of its structural equation, and construct the likelihood function of the missing nodes using the residuals. Specifically, for any observable node in the heterogeneous graph of multimodal operating conditions, the observed values ​​of the node eigenvectors of its parent node set are substituted into the structural equation of this observable node to obtain the predicted value of the structural equation of this observable node; the difference between the observed value of the node eigenvectors of this observable node and the predicted value of the structural equation is defined as the structural equation residual of this observable node; the joint distribution of the structural equation residuals of all observable nodes is taken, and the probability density value of the joint residuals under the assumption of the normal distribution of the structural equation residuals is used as the likelihood function with respect to the parameters of the missing nodes; Furthermore, the scope of the likelihood function construction is selectively limited: only observable nodes that have a direct directed dependency relationship (i.e., parent-child relationship) with the missing node in the directed dependency relationship of the initial directed acyclic graph are selected to participate in the likelihood function construction, and observable nodes that do not have a direct directed dependency relationship with the missing node are not included in the likelihood function; this selective limitation ensures that the likelihood function only reflects the observation information that is causally related to the missing node, avoiding the introduction of noise into the posterior inference by irrelevant modal observations; S2.5: Perform posterior inference for each missing node in the set of missing nodes layer by layer according to the determined topological sort, wherein the posterior inference is based on Bayes' theorem: ; in, For the prior distribution, For the structural equation residual term The distribution is a likelihood function in parameterized form. represents the posterior distribution of the missing nodes; It should be noted that the residual term The distribution is obtained by fitting the sample distribution of the residual sequence obtained by substituting each observable node in the observable node set into the structural equation. When Gaussian distribution parameterization is used, the analytical solution of the posterior distribution is a Gaussian distribution. The posterior mean and posterior variance are calculated by the weighted harmonic sum of the prior distribution parameters and the likelihood function parameters. Furthermore, posterior inference: multiply the prior distribution by the likelihood function and normalize to obtain the posterior distribution of each missing node; under the combination of normal prior distribution and normal likelihood function, the posterior distribution is also normal, and its posterior mean vector and posterior covariance matrix both have closed-form solutions: the posterior precision matrix is ​​the sum of the prior precision matrix and the likelihood precision matrix, and the posterior mean vector is the product of the inverse of the posterior precision matrix and the sum of the prior precision matrix and the likelihood precision matrix multiplied by their respective mean vectors; the above closed-form calculation only involves matrix addition and inversion operations, and the computational complexity increases with the cube of the dimension of the missing node feature vector. For scenarios where the node feature vector dimension does not exceed 100 dimensions, it can be completed in a single data frame time. Specifically, a confidence assessment is performed on the inference results of the posterior distribution: the trace of the posterior covariance matrix is ​​used as the posterior uncertainty index. When the posterior uncertainty index exceeds the preset uncertainty upper limit, the posterior inference result of the missing node is marked as a low confidence state, and a low confidence mark is added to the node in the completed multimodal heterogeneous diagram. S2.6: Take the mean vector of the posterior distribution of each missing node as the posterior expectation value, write the posterior expectation value into the node feature domain of the corresponding node in the multimodal working condition heterogeneous graph, and obtain the completed multimodal working condition heterogeneous graph. Furthermore, the graph structure of the completed multimodal heterogeneous graph is completely consistent with the directed dependency set, spatial edge weight set, and graph-level attributes of the multimodal heterogeneous graph. Only the node feature domain of the missing nodes that were originally null values ​​is replaced with the corresponding posterior expected value; the node feature vector of the original observable nodes remains unchanged. Furthermore, a node source labeling domain is added to complete the heterogeneous graph of the multimodal operating conditions: the node source labeling domain of each node is assigned a value of either the observed value or the posterior inferred value to distinguish between the original observable nodes and the nodes completed by the posterior inference; the node source labeling domain is used to apply the structural update weight discount to the posterior inferred value nodes in the Bayesian structure update in S3, and to add a confidence interval to the causal contribution rate calculation results of the posterior inferred value nodes in the do operator application in S4.

[0021] It should be noted that the posterior expectation backfilling operation does not modify the directed dependencies of the initial directed acyclic graph. The directions of the directed dependencies in the completed multimodal heterogeneous graph are completely consistent with those of the initial directed acyclic graph. S3: Using navigation condition indicators consisting of speed, ballast draft, and sea state as trigger conditions, perform online Bayesian structure update on the completed multimodal heterogeneous diagram to obtain an adaptive directed acyclic graph. S3.1: Quantitatively detect the degree of deviation between the current observed distribution and the historical baseline distribution of the navigation condition indicators in the completed multimodal heterogeneous diagram. Determine whether to trigger online Bayesian structure update based on whether the degree of deviation exceeds the preset drift threshold.

[0022] Specifically, the current observation distribution is constructed using the current sliding window samples of each node (speed, ballast draft, and sea state) in the multimodal heterogeneous diagram. The historical baseline distribution is constructed using the kernel density estimation results of the corresponding variables in the historical normal navigation data of the same route. The KL divergence between the current observation distribution and the historical baseline distribution is calculated, and the specific formula is as follows: ; Furthermore, when the KL divergence exceeds the preset drift threshold, it is determined that the navigation condition index has shifted in distribution, triggering an online Bayesian structure update; when the KL divergence does not exceed the preset drift threshold, the current topology and conditional probability weights of the initial directed acyclic graph are retained, and the current initial directed acyclic graph is directly used as the adaptive directed acyclic graph and output to step S4. It should be noted that the preset drift threshold is determined based on the 95th percentile of the KL divergence distribution corresponding to the normal operating condition switching event in the historical navigation data, in order to distinguish between normal operating condition fluctuations and significant operating condition offsets; the window length of the current sliding window is consistent with the window length of the sliding window used for sensor time series data feature extraction, so as to complete the time alignment benchmark of the time series features of each node in the multimodal operating condition heterogeneous graph as a common index. S3.2: Under the condition that the distribution of navigation condition indicators is offset, the node feature vector of the observable node set in the complete multimodal heterogeneous graph is used as the observation data. Based on the Bayesian scoring criterion, causal edge addition and deletion operations are performed on the topology of the initial directed acyclic graph to obtain the updated topology directed acyclic graph. Furthermore, the Bayesian scoring criterion uses the Bayesian Information Criterion (BIC) for scoring, with the specific formula as follows: ; in, Candidate directed acyclic graph topology, To complete the observation dataset composed of the node feature vectors of the observable node set in the heterogeneous graph of multimodal operating conditions, For maximum likelihood parameter estimation under topology G, Let G be the number of free parameters of the topology G, and N be the number of observed samples. Specifically, delete each directed edge in the initial directed acyclic graph and add each candidate directed edge that does not currently exist in the initial directed acyclic graph. Calculate the BIC score increment of the candidate topology after each addition or deletion operation, and retain the addition or deletion operations with positive BIC score increments to obtain the updated topological directed acyclic graph. Furthermore, the range of candidate directed edges is limited by the constraint of the maximum number of parent nodes among nodes in the initial directed acyclic graph. The maximum number of parent nodes is configured according to the physical reachable causal order of each variable in the prior causal knowledge of ship power transmission, so as to avoid overfitting the current working condition sample and causing the updated topological directed acyclic graph to deviate from the physical causal mechanism. Furthermore, after each insertion or deletion operation, a directed acyclic graph (DAG) topology sorting verification is performed on the candidate topologies. If an insertion or deletion operation introduces a cycle, the operation is discarded. The cycle detection uses a depth-first search algorithm with a detection complexity of O(log n). , where V and E are the number of nodes and the number of directed edges in the candidate topology, respectively; Preferably, the causal edge addition and deletion operations adopt an incremental update variant of the Greedy Equivalence Search (GES) algorithm, which uses the initial directed acyclic graph as the search starting point instead of re-searching from an empty graph. The incremental update variant only searches within the current topological neighborhood of the initial directed acyclic graph to reduce online computational overhead. In an optional embodiment, the Bayesian scoring criterion is replaced by the Bayesian Dirichlet Equivalent Uniform (BDeu) score, which has better structural consistency than the BIC score when the proportion of discrete variable nodes is high. It is suitable for situations where discrete navigation condition index nodes such as sea state level dominate the completion of multimodal heterogeneous maps. S3.3: Using the topological structure of the updated topological directed acyclic graph as a fixed framework, perform Bayesian parameter updates on the conditional probability weights of each retained directed edge in the updated topological directed acyclic graph to obtain an adaptive directed acyclic graph. Specifically, for each directed edge preserved in the updated topological directed acyclic graph The posterior conditional probability weights are calculated using the initial conditional probability weights of the corresponding directed edges in the initial directed acyclic graph as the prior mean, and the observation data within the current sliding window of the corresponding node in the completed multimodal heterogeneous graph as the likelihood data source. The conjugate prior update rule is used to calculate the posterior conditional probability weights, and the specific formula is as follows: ; in, For directed edges The conditional probability weight parameters, To complete the observation data of the corresponding node in the current sliding window of the heterogeneous graph of multimodal working conditions, the posterior mean is used as the update conditional probability weight of the corresponding directed edge in the adaptive directed acyclic graph. Furthermore, for newly added directed edges, the prior mean of their conditional probability weights is initialized with the sample correlation coefficient of the corresponding variable pairs in historical samples of the same working conditions; for deleted directed edges, their corresponding conditional probability weights are removed from the adaptive directed acyclic graph. Furthermore, the update of the posterior conditional probability weights adopts a recursive Bayesian filtering method. When new observation data enters the current sliding window, the posterior conditional probability weights of the previous time step are used as the new priors. The new observation samples are gradually absorbed, and the forgetting factor of the recursive update is adaptively adjusted according to the distribution offset rate of the navigation condition index: the larger the KL divergence, the smaller the forgetting factor, and the faster the historical prior weights decay. It should be noted that if the KL divergence does not exceed the preset drift threshold, the adaptive directed acyclic graph is directly composed of the topology and conditional probability weights of the current initial directed acyclic graph. The conditional probability weights of each directed edge in the adaptive directed acyclic graph serve as the basis for calculating the host load rate deviation and propeller propulsion efficiency deviation, applying the do operator intervention, and calculating the causal contribution rate. Preferably, the update conditional probability weights of each directed edge in the adaptive directed acyclic graph are stored in the form of directed edge confidence intervals. The confidence intervals are determined by the posterior variance of the posterior distribution. Directed edges with larger posterior variances are assigned lower causal propagation confidence in the S4 do operator intervention calculation to reflect the high uncertainty of parameter estimation in the early stage of the working condition switching. In an optional embodiment, when the number of observed samples in the current sliding window of the completed multimodal heterogeneous graph is lower than the preset minimum sample size threshold, the Bayesian parameter update in S3.3 is only performed on the condition modulation causal edges that are directly connected to the navigation condition index in the updated topological directed acyclic graph. The conditional probability weights of the other directed edges retain the initial conditional probability weights in the initial directed acyclic graph. The full graph parameter update is performed after the number of observed samples accumulates to the preset minimum sample size threshold, so as to avoid excessive fluctuations in parameter estimation under small sample conditions. S4: Based on the conditional probability weights of the adaptive directed acyclic graph, the observation deviations of the main engine speed and propeller torque in the sensor time series data are converted into the deviations of the main engine load rate and propeller propulsion efficiency. The overall ship energy efficiency deviation value is defined by the ratio of fuel flow deviation to rated fuel consumption rate. The do operator intervention is applied to the deviations of the main engine load rate and propeller propulsion efficiency on the adaptive directed acyclic graph. The causal contribution rate of each variable to the overall ship energy efficiency deviation value is calculated. The causal propagation path is decomposed at three levels of granularity: component level, system level, and ship level to obtain a hierarchical attribution result set. S4.1: Using the mean value of the node feature vectors of each node in the sensor time series data of the completed multimodal operating condition heterogeneous diagram under the historical same operating condition sample as the benchmark value, calculate the difference between the mean value of the feature vectors of each node in the current sliding window of the sensor time series data of the completed multimodal operating condition heterogeneous diagram and the benchmark value, and obtain the observation deviation of the main engine speed, propeller torque and fuel flow.

[0023] Specifically, the matching conditions for historical samples under the same operating conditions are consistent with the matching conditions for prior data sources, namely, the speed deviation does not exceed 0.5 knots, the ballast draft deviation does not exceed 0.2 meters, and the sea state is the same; the main engine speed observation deviation... Propeller torque observation deviation Deviation from fuel flow observation They are defined as follows: ; in, , , These are respectively used to complete the unnormalized mean values ​​of the main engine speed, propeller torque, and fuel flow rate nodes in the heterogeneous diagram of multimodal operating conditions within the current sliding window. , , These are the baseline values ​​for the main engine speed, propeller torque, and fuel flow rate in historical samples under the same operating conditions. It should be noted that the current window length of the sliding window is consistent with the window length of the sliding window, so as to use the time alignment benchmark of the temporal features of each node in the multimodal heterogeneous graph as a common index; S4.2: Based on the conditional probability weights of the causal edges of the propulsion chain and the causal edges of the energy input in the adaptive directed acyclic graph, the observation deviation of the main engine speed and the observation deviation of the propeller torque are converted into the deviation of the main engine load rate and the deviation of the propeller propulsion efficiency, and the overall ship energy efficiency deviation value is defined by the ratio of the observation deviation of the fuel flow rate to the rated fuel consumption rate. Furthermore, host load rate deviation Deviation observed from main engine speed Based on the conditional probability weights of energy input causal edges in an adaptive directed acyclic graph The conversion yields the following formula: ; Propeller propulsion efficiency deviation Deviation observed from propeller torque Based on the conditional probability weights of causal edges in the advancing chain of an adaptive directed acyclic graph The conversion yields the following formula: ; Overall ship energy efficiency deviation Observation deviation based on fuel flow rate With rated fuel consumption rate Definition of ratio: ; in, For adaptive directed edges in a directed acyclic graph The conditional probability weights attached to the (fuel flow node pointing to the main engine speed node) For adaptive directed edges in a directed acyclic graph The conditional probability weights attached to (the main engine speed node pointing to the propeller torque node), and the rated fuel consumption rate. The nominal fuel consumption rate under rated operating conditions of the main engine is taken from the ship's design specifications. Furthermore, conditional probability weights and The mean of the posterior conditional probability weights of the corresponding directed edges in the adaptive directed acyclic graph after Bayesian parameter updates is taken. When the confidence interval width of the posterior conditional probability weights exceeds the preset confidence width threshold, the lower bound of the confidence interval is used for the conversion to reflect the conservative correction of the bias conversion result by the uncertainty of parameter estimation. The preset confidence width threshold is determined based on the maximum relative error limit allowed by the parameter estimation project. It is usually taken as the upper limit of the confidence interval width corresponding to 1.96 times the standard deviation of the posterior conditional probability weights, which corresponds to the relative error not exceeding the preset percentage at a 95% confidence level. Preferably, rated fuel consumption rate The ballast draft and speed are segmented and corrected according to the current navigation conditions. The segmented interpolation table is composed of the nominal values ​​of the rated fuel consumption rate under different combinations of draft and speed in the ship design specifications, so as to eliminate the influence of the difference in operating conditions on the definition of the overall ship energy efficiency deviation value. S4.3: On the adaptive directed acyclic graph, the do operator intervention is applied sequentially to the main engine load rate deviation and the propeller propulsion efficiency deviation. By truncating all the incoming edges of the intervention nodes in the adaptive directed acyclic graph, the intervention subgraph is constructed. The causal effect of the intervention variables on the overall ship energy efficiency deviation value is calculated on the intervention subgraph to obtain the causal contribution rate of each intervention variable. Specifically, apply the do operator intervention to nodes with skewed host load rates. At that time, truncate all directed edges pointing to the host load rate deviation node in the adaptive directed acyclic graph, and fix the host load rate deviation node value. After intervention, the causal propagation path along the adaptive directed acyclic graph is forward to the whole ship energy efficiency deviation value node on the subgraph, and the expected change in the whole ship energy efficiency deviation value before and after intervention is calculated. Apply the do operator to the propeller propulsion efficiency deviation node. At the same time, all incoming edges of the propeller propulsion efficiency deviation node are truncated in the same way and the corresponding expected change is calculated; Furthermore, the causal contribution rate between the main engine load rate deviation node and the propeller propulsion efficiency deviation node. and They are defined as follows: ; in, The intervention setting value for host load rate deviation, Intervention setpoint for propeller propulsion efficiency deviation, causal contribution rate and The sum does not exceed 1, and the difference is partly attributed to the joint effect of the other node variables in the adaptive directed acyclic graph; Furthermore, the calculation of the expected change adopts Monte Carlo forward sampling of the posterior conditional probability weights of each directed edge in the adaptive directed acyclic graph. The number of samplings is adaptively configured according to the confidence interval width of the posterior conditional probability weights: the wider the confidence interval, the more samplings are performed, in order to control the variance of the causal contribution rate estimation. Preferably, when there is a direct directed edge between the host load rate deviation node and the propeller propulsion efficiency deviation node in the adaptive directed acyclic graph, the path-specific effect (PSE) decomposition is used instead of the simple do operator intervention. The direct causal effect between the two nodes and the indirect causal effect through the intermediate node are calculated separately and then summed to avoid double counting of the intervention effect. S4.4: Based on the node affiliation relationship of the adaptive directed acyclic graph, the causal contribution rate and its corresponding causal propagation path are decomposed into three levels of granularity: component level, system level, and ship level, to obtain a hierarchical attribution result set; Specifically, the node affiliation relationships of the adaptive directed acyclic graph are pre-configured according to the following three levels of granularity: component-level nodes correspond to single mechanical components such as main engine, propeller, fuel pump, and heat exchanger; system-level nodes correspond to functional subsystems such as propulsion system, fuel system, and cooling system, which are composed of multiple component-level nodes; and ship-wide nodes correspond to the ship-wide energy efficiency deviation value nodes. Furthermore, the three-level granularity decomposition steps of the causal propagation path are as follows: Component level: Extract all directed paths in the subgraph after intervention, starting from the main engine load rate deviation node or the propeller propulsion efficiency deviation node, passing through each component level node to the whole ship energy efficiency deviation value node, and use the product of the conditional probability weights of each component level node on the path as the component level causal propagation coefficient of that component. System level: Summing the component-level causal propagation coefficients of all component-level nodes belonging to the same system yields the system-level causal propagation coefficients of each functional subsystem; Ship-wide level: The weighted summation of all system-level causal propagation coefficients to the ship-wide energy efficiency deviation value node yields the ship-wide causal convergence coefficient; Furthermore, the hierarchical attribution result set is stored in a directed tree structure. The root node is the whole ship energy efficiency deviation value node, the first-level child nodes are the system-level nodes, the second-level child nodes are the component-level nodes, and the storage fields of each node include node identifier, granularity level, causal propagation coefficient, corresponding directed path node sequence and causal contribution rate component. Specifically, for component-level nodes whose component-level causal propagation coefficients in the hierarchical attribution result set are lower than the preset minimum contribution threshold, they are merged into the residual terms of their respective system-level nodes and are not separately included in the component-level output of the hierarchical attribution result set. This is to compress the size of the hierarchical attribution result set and highlight the main component. The preset minimum contribution threshold is determined based on the readability of the hierarchical attribution result set and the limited operational resources. It is usually taken as 1% to 5% of the normalized causal propagation coefficient as the cutoff value. Preferably, the causal propagation coefficients of each granular level in the hierarchical attribution result set are stored in percentage form after normalization. The normalization benchmark is the current energy efficiency deviation value of the whole ship, and the normalized causal propagation coefficient is the causal contribution rate component of each node at each granular level to the energy efficiency deviation value of the whole ship. In an optional embodiment, the three-level granular configuration of node affiliation supports dynamic adjustment based on ship type: for ship types equipped with shaft generators, a new shaft generator system node is added at the system level, and causal edges from the shaft generator speed node to the overall ship energy efficiency deviation value node are added in the adaptive directed acyclic graph; for ship types equipped with waste heat recovery systems, a waste heat boiler node is added at the component level to cover the offsetting effect of waste heat utilization on fuel flow deviation. The dynamically adjusted node affiliation is configured according to the ship design specifications during the system initialization phase and does not change with subsequent online Bayesian structure updates. S5: Convert the hierarchical attribution results into a structured attribution report, determine the priority of operation and maintenance handling according to the causal contribution rate from high to low, and output cross-modal tracing attribution results; S5.1: Based on the normalized causal propagation coefficient of the component-level nodes in the hierarchical attribution result set, sort all component-level nodes in the hierarchical attribution result set from high to low according to the normalized causal propagation coefficient, determine the operation and maintenance priority of each component-level node, and extract the directed path node sequence corresponding to each component-level node after sorting as the causal propagation path. Furthermore, the priority of operation and maintenance is represented by a positive integer sequence number. Sequence number 1 corresponds to the component-level node with the highest normalized causal propagation coefficient, and the sequence numbers increase in descending order according to the normalized causal propagation coefficient. When the difference in the normalized causal propagation coefficients of two component-level nodes is lower than the preset parallel threshold, the two nodes are determined to have the same priority. The parallel nodes are sorted by the number of hops of the directed path from the component-level node to the whole ship energy efficiency deviation value node in the adaptive directed acyclic graph, from least to most, with the fewer hops, the higher the priority. It should be noted that the preset parallel threshold is configured based on the overall distribution variance of the normalized causal propagation coefficient of the hierarchical attribution results set. When the variance is small and the normalized causal propagation coefficients of each component-level node are relatively close, the preset parallel threshold is appropriately narrowed to improve the priority differentiation. The operation and maintenance handling priority and the causal propagation path together constitute the index basis for the extraction of associated modal evidence and the calculation of the contribution of the whole ship fuel consumption deviation. Each subsequent sub-step organizes the output fields of each component-level node with the operation and maintenance handling priority number as the primary key. S5.2: Based on the node identifiers in the causal propagation path, extract the associated modal evidence corresponding to each node in the causal propagation path from the completed multimodal heterogeneous diagram and the cabin CCTV equipment status image data; Furthermore, the extraction of relevant modal evidence is performed in three categories based on modal type: For the sensor time series data modes, extract the node feature vector time series corresponding to each node of the causal propagation path in the heterogeneous diagram of the multimodal operating conditions, extract the time series segment within the current sliding window, and use the time series change curves of the host speed observation deviation, propeller torque observation deviation, and fuel flow observation deviation as evidence of sensor time series data mode association. For the navigation GIS trajectory data modalities, extract the spatial edge weight time series of the directed edges corresponding to each node of the causal propagation path in the multimodal heterogeneous graph to complete the causal propagation path. Use the sequence of segment distance, speed to ground and rate of change of heading to ground within the current sliding window as evidence of the association between navigation GIS trajectory data modalities. For the cabin CCTV equipment status image data modality, the visual feature vectors stored in the node visual attribute fields of the corresponding devices of each component-level node in the causal propagation path are used as indexes. The corresponding time frame with the highest cosine similarity of the visual feature vectors is retrieved from the cabin CCTV equipment status image data, and the corresponding time frame is extracted as evidence of the association between the cabin CCTV equipment status image data modality. Specifically, the retrieval scope of the corresponding time frame is limited to the set of cabin CCTV equipment status image data frames within the current sliding window time range. When multiple camera devices correspond to the same component-level node in the causal propagation path, the frames collected by the camera device covering the maximum field of view of that component are given priority as the retrieval source. The three types of associated modal evidence are organized into an associated modal evidence set with the maintenance and handling priority number as the primary key and the modal type identifier as the secondary index. The storage fields of each entry in the associated modal evidence set include the maintenance and handling priority number, modal type identifier, evidence data fragment, and evidence collection time timestamp. Furthermore, the corresponding time frames of the modal association evidence of the cabin CCTV equipment status image data are processed by image annotation, and the image areas of the equipment corresponding to the component-level nodes are marked with rectangular boxes. The color of the annotation boxes is graded according to the priority number of the operation and maintenance, with the smaller the number, the darker the color, so as to intuitively distinguish the visual evidence of different priority components in the cross-modal tracing and attribution results. Preferably, when a component-level node in the causal propagation path has a missing cabin CCTV equipment status image data frame in the current sliding window and has been backfilled, the corresponding time frame is replaced by the historical nearest neighbor frame corresponding to the expected value of the posterior distribution under the same working condition, and a missing mark is added to the evidence data fragment field of the corresponding entry in the associated modality evidence set to inform the subsequent processing steps that the frame is a historical replacement frame rather than the current actual frame. The backfilling process is step S2. S5.3: Calculate the contribution of each component-level node to the overall fuel consumption deviation by multiplying the overall energy efficiency deviation value with the normalized causal propagation coefficient of each component-level node in the hierarchical attribution result set. Furthermore, the contribution of the kth component-level node to the overall ship fuel consumption deviation. The specific formula is as follows: ; in, This represents the overall ship energy efficiency deviation value. This represents the normalized causal propagation coefficient of the k-th component-level node in the hierarchical attribution result set. Rated fuel consumption rate, This represents the current duration of the sliding window. It should be noted that when the rated fuel consumption rate When the unit is tons / hour, Hourly; when rated fuel consumption rate When the unit is kg / hour, first set the rated fuel consumption rate. Converted to tons per hour before being included in the calculation of the overall ship fuel consumption deviation contribution, so as to ensure the overall ship fuel consumption deviation contribution. The unit is uniformly ton; Specifically, the contribution of overall ship fuel consumption deviation Further converted into the contribution of equivalent carbon dioxide emission deviation : ; in, The carbon emission conversion factor for the corresponding fuel type in Annex VI of the International Maritime Organization's MARPOL Convention, and the equivalent CO2 emission deviation contribution. As a supplementary field to the contribution of the overall ship fuel consumption deviation, it is stored in the cross-modal attribution results to support the needs of ship carbon emission compliance verification. It should be noted that when the hierarchical attribution results are concentrated at the component level nodes that are merged into the system-level residual items, the contribution of the ship's fuel consumption deviation corresponding to the residual item is calculated by the product of the normalized causal propagation coefficient of the system-level residual node and the ship's energy efficiency deviation value, and is included in the cross-modal attribution results with the system-level residual identifier, without being broken down to the specific component level. S5.4: Organize the operation and maintenance handling priority, causal propagation path, related modal evidence and the contribution of the whole ship fuel consumption deviation according to the preset structured attribution report format, and output the cross-modal tracing attribution results; Furthermore, the cross-modal attribution results are organized sequentially from smallest to largest, using the operation and maintenance handling priority number as the primary key. Each record contains the following fields: Root Cause Component Field: Stores the component-level node identifier and its system-level node identifier corresponding to the operation and maintenance handling priority number; Causal propagation path field: Stores the sequence of directed path nodes, arranged in the topological sorting order of the adaptive directed acyclic graph in the form of a list of node identifiers; Associated Modal Evidence Field: Stores all entries in the associated modal evidence set that correspond to the operation and maintenance handling priority number, including the time frames corresponding to the sensor time series data modal association evidence, the navigation GIS trajectory data modal association evidence, and the cabin CCTV equipment status image data modal association evidence; The "Overall Fuel Consumption Deviation Contribution" field stores the overall fuel consumption deviation contribution. Contribution to deviation from equivalent carbon dioxide emissions ; Operation and maintenance handling priority field: Stores the operation and maintenance handling priority number; Furthermore, the cross-modal attribution results are serialized and stored in JSON format, with the Unix timestamp of the current sliding window termination time of the complete multimodal heterogeneous diagram as the file name index. The data is persisted to the shipboard data storage unit and synchronized to the shore-based ship energy efficiency management platform via the ship-to-shore communication link. Specifically, the structured attribution report adds a summary field to the JSON format of the cross-modal attribution results. The summary field includes the current sliding window period identifier, the ship-wide causal convergence coefficient, the root cause component identifier corresponding to priority number 1, and the total contribution of the ship-wide fuel consumption deviation. The summary field is used for lightweight priority transmission under the condition of limited ship-to-shore communication bandwidth. After receiving the summary field, the shore-based ship energy efficiency management platform can retrieve the complete cross-modal attribution results as needed. Preferably, the output period of the cross-modal attribution result is consistent with the sliding step size of the current sliding window. The sliding step size is configured as an integer multiple of the sampling period of the sensor time series data to complete the latest timestamp of the time series node of the heterogeneous graph of multimodal working conditions and drive the current sliding window to roll and update. After each roll, the complete execution link from step S4.1 to step S5.4 is triggered, and the updated cross-modal attribution result is output. In an optional embodiment, the engine room CCTV equipment status image data modal association evidence corresponding to the modal evidence field in the cross-modal attribution results is stored as a thumbnail in the Base64 encoded field of a JSON file, and the full-resolution frame is stored in the shipborne image storage unit and mounted to the JSON file by a URI index. The shore-based ship energy efficiency management platform selects to pull the thumbnail or the full-resolution frame according to the network conditions to balance the ship-shore communication bandwidth usage and the readability of the image evidence.

[0024] In summary, this invention collects sensor time-series data, GIS trajectory data, and engine room CCTV image data, mapping them to time-series nodes, spatial edge weights, and visual feature vectors to form a multimodal heterogeneous graph of operating conditions. This allows dispersed power, navigation, and equipment status information to be expressed under a unified causal graph structure, avoiding biased anomaly judgments caused by relying solely on single sensor data. By using the directed dependencies of the initial directed acyclic graph as structural equation constraints, and combining historical samples of the same operating conditions to perform posterior extrapolation and backfilling of missing nodes, it reduces attribution bias caused by sensor packet loss, image occlusion, or data asynchrony under complex ship operating conditions, improving the data integrity and stability of subsequent causal calculations. Furthermore, by triggering online Bayesian structure updates using navigation condition indicators composed of speed, ballast draft, and sea state, the causal graph can adaptively adjust its topology and conditional probability as navigation conditions change. The system employs rate weighting to avoid causal mismatches in fixed models when sea state, load, or speed changes. By converting observed deviations such as main engine speed, propeller torque, and fuel flow into main engine load rate deviation, propeller propulsion efficiency deviation, and overall ship energy efficiency deviation values ​​on an adaptive directed acyclic graph, and applying the do operator intervention to key variables, it can distinguish between correlation anomalies and true causal anomalies. It decomposes causal propagation paths at the component, system, and ship levels, elevating the discovery of energy efficiency anomalies to locating the source and propagation chain of anomalies. By transforming hierarchical attribution results into structured attribution reports and determining maintenance priorities based on causal contribution rates, it can provide interpretable results with correlated modal evidence, fuel consumption deviation contribution, and response order for crew members or shore-based maintenance platforms, thereby improving the accuracy, interpretability, and maintenance efficiency of ship energy efficiency anomaly diagnosis.

[0025] Based on the teachings of the above embodiments, other aspects of the present invention also propose a cross-modal attribution system for ship energy efficiency anomalies based on a structural causal model, such as... Figure 2 As shown, it includes: The multimodal graph construction module is used to collect multimodal performance monitoring data of ships, construct an initial directed acyclic graph, and map the multimodal performance monitoring data into the temporal nodes, spatial edge weights and visual feature vectors of the initial directed acyclic graph to obtain a multimodal operating condition heterogeneous graph. The node completion module is used to perform posterior inference on missing nodes in the multimodal heterogeneous graph, with the directed dependencies of the initial directed acyclic graph as structural equation constraints and historical samples of the same working condition as priors. The inferred expected values ​​are then used to fill in the corresponding nodes to obtain a completed multimodal heterogeneous graph. The structure update module is used to perform online Bayesian structure update on the completed multimodal heterogeneous graph, using navigation condition indicators consisting of speed, ballast draft, and sea state as trigger conditions, to obtain an adaptive directed acyclic graph. The causal attribution module, based on the conditional probability weights of an adaptive directed acyclic graph, converts the observation deviations of the main engine speed and propeller torque in the sensor time series data into the deviations of the main engine load rate and propeller propulsion efficiency. It defines the overall ship energy efficiency deviation value as the ratio of fuel flow deviation to rated fuel consumption rate. It applies the do operator intervention to the main engine load rate deviation and propeller propulsion efficiency deviation on the adaptive directed acyclic graph, calculates the causal contribution rate of each variable to the overall ship energy efficiency deviation value, and decomposes the causal propagation path at three levels of granularity: component level, system level, and ship level to obtain a hierarchical attribution result set. The results output module is used to transform the causal propagation paths of each layer in the hierarchical attribution results set into a structured attribution report, determine the priority of operation and maintenance handling according to the causal contribution rate from high to low, and output cross-modal tracing attribution results.

[0026] This embodiment also provides a computer device applicable to the cross-modal attribution method for ship energy efficiency anomalies based on structural causal models, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the cross-modal attribution method for ship energy efficiency anomalies based on structural causal models proposed in the above embodiment.

[0027] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0028] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as proposed in the above embodiments.

[0029] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0030] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model, characterized in that, include: Collect multimodal performance monitoring data of ships, construct an initial directed acyclic graph, and map the multimodal performance monitoring data to the temporal nodes, spatial edge weights and visual feature vectors of the initial directed acyclic graph to obtain a multimodal operating condition heterogeneous graph; Using the directed dependencies of the initial directed acyclic graph as structural equation constraints and historical samples of the same working condition as priors, posterior inference is performed on the missing nodes in the multimodal working condition heterogeneous graph, and the inferred expected values ​​are backfilled into the corresponding nodes to obtain a complete multimodal working condition heterogeneous graph. Using navigation condition indicators consisting of speed, ballast draft, and sea state as trigger conditions, online Bayesian structure update is performed on the completed multimodal heterogeneous diagram to obtain an adaptive directed acyclic graph. Based on the conditional probability weights of the adaptive directed acyclic graph, the observation deviations of the main engine speed and propeller torque in the sensor time series data are converted into the deviations of the main engine load rate and propeller propulsion efficiency. The overall ship energy efficiency deviation value is defined by the ratio of fuel flow deviation to rated fuel consumption rate. The do operator intervention is applied to the deviations of the main engine load rate and propeller propulsion efficiency on the adaptive directed acyclic graph. The causal contribution rate of each variable to the overall ship energy efficiency deviation value is calculated. The causal propagation path is decomposed at three levels of granularity: component level, system level, and ship level to obtain a hierarchical attribution result set. The hierarchical attribution results are converted into structured attribution reports, and the operation and maintenance priorities are determined according to the causal contribution rate from high to low. Cross-modal tracing attribution results are then output.

2. The cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in claim 1, characterized in that, The cross-modal attribution results include: Using the normalized causal propagation coefficient of the component-level nodes in the hierarchical attribution result set as the sorting basis, all component-level nodes in the hierarchical attribution result set are sorted from high to low according to the normalized causal propagation coefficient, the operation and maintenance priority of each component-level node is determined, and the directed path node sequence corresponding to each component-level node after sorting is extracted as the causal propagation path. Based on the node identifiers in the causal propagation path, the associated modal evidence corresponding to each node in the causal propagation path is extracted from the completed multimodal heterogeneous diagram and the cabin CCTV equipment status image data. The contribution of each component-level node to the overall ship fuel consumption deviation is calculated by multiplying the overall ship energy efficiency deviation value with the normalized causal propagation coefficient of each component-level node in the hierarchical attribution result set. The operation and maintenance priority, the causal propagation path, the related modal evidence, and the contribution of the overall ship fuel consumption deviation are organized according to a preset structured attribution report format, and the cross-modal tracing attribution results are output.

3. The cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in claim 2, characterized in that, The hierarchical attribution result set includes: Using the mean value of the node feature vectors of each node in the sensor time series data of the completed multimodal working condition heterogeneous diagram under the historical same working condition sample as the benchmark value, the difference between the mean value of the node feature vectors of each node in the current sliding window of the sensor time series data of the completed multimodal working condition heterogeneous diagram and the benchmark value is calculated to obtain the host speed observation deviation, propeller torque observation deviation and fuel flow observation deviation. Based on the conditional probability weights of the causal edges of the propulsion chain and the causal edges of the energy input in the adaptive directed acyclic graph, the observed deviation of the main engine speed and the observed deviation of the propeller torque are converted into the deviation of the main engine load rate and the deviation of the propeller propulsion efficiency, and the overall ship energy efficiency deviation value is defined by the ratio of the observed deviation of the fuel flow rate to the rated fuel consumption rate. On the adaptive directed acyclic graph, the do operator intervention is applied sequentially to the host load rate deviation and the propeller propulsion efficiency deviation. By truncating all incoming edges of the intervention nodes in the adaptive directed acyclic graph, an intervention subgraph is constructed. The causal effect of the intervention variables on the overall ship energy efficiency deviation value is calculated on the intervention subgraph to obtain the causal contribution rate of each intervention variable. Based on the node affiliation relationship of the adaptive directed acyclic graph, the causal contribution rate and its corresponding causal propagation path are decomposed into three levels of granularity: component level, system level, and ship level, to obtain a hierarchical attribution result set.

4. The cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in claim 3, characterized in that, The adaptive directed acyclic graph includes: The degree of deviation between the current observed distribution and the historical baseline distribution of the navigation condition index in the completed multimodal heterogeneous diagram is quantitatively detected. Whether the degree of deviation exceeds a preset drift threshold is used as the criterion to determine whether to trigger online Bayesian structure update. Under the condition that the distribution of the navigation condition index is offset, the node feature vector of the observable node set in the completed multimodal heterogeneous graph is used as the observation data. Based on the Bayesian scoring criterion, causal edge addition and deletion operations are performed on the topology of the initial directed acyclic graph to obtain the updated topology directed acyclic graph. Using the topological structure of the updated topological directed acyclic graph as a fixed framework, Bayesian parameter updates are performed on the conditional probability weights of each retained directed edge in the updated topological directed acyclic graph to obtain an adaptive directed acyclic graph.

5. The cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in claim 4, characterized in that, Determining whether to trigger an online Bayesian structure update includes: The current observation distribution is constructed using the current sliding window samples of each node of speed, ballast draft, and sea state level in the completed multimodal heterogeneous diagram. The historical baseline distribution is constructed using the kernel density estimation results of the corresponding variables in the historical normal navigation data of the same route. The KL divergence between the current observation distribution and the historical baseline distribution is calculated. When the KL divergence exceeds the preset drift threshold, it is determined that the navigation condition index has shifted in distribution, triggering an online Bayesian structure update; when the KL divergence does not exceed the preset drift threshold, the current topology and conditional probability weights of the initial directed acyclic graph are retained, and the current initial directed acyclic graph is directly used as the adaptive directed acyclic graph.

6. The cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in claim 5, characterized in that, The completion of the multimodal heterogeneous diagram includes: Perform missing feature vector detection on the node feature vectors of each node in the multimodal heterogeneous graph; Using the directed dependencies of the initial directed acyclic graph as constraints for the structural equations, a structural equation is constructed for each missing node in the multimodal heterogeneous graph, with its set of parent nodes as the independent variable. Using the current observed value of the navigation condition index as an index, retrieve historical samples belonging to the same operating condition interval from the historical sample database, and record the retrieved historical sample set as the historical same operating condition sample set. For all observable nodes in the multimodal heterogeneous graph, under the constraints of the structural equation, the residual between the observed value of the node feature vector of each observable node and its predicted value of the structural equation is calculated, and the likelihood function of the missing node is constructed using the residual. Perform posterior inference on each missing node in the set of missing nodes layer by layer according to the determined topological sort; The mean vector of the posterior distribution of each missing node is taken as the posterior expectation value. The posterior expectation value is written into the node feature domain of the corresponding node in the multimodal heterogeneous graph to obtain the completed multimodal heterogeneous graph.

7. The cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in claim 6, characterized in that, The missing node detection includes determining whether the node feature field is empty or below the effective signal threshold on a node-by-node basis; if the node feature field is empty or below the effective signal threshold, the node is marked as a missing node, otherwise the node is marked as an observable node.

8. A cross-modal attribution system for ship energy efficiency anomalies based on a structural causal model, comprising the cross-modal attribution method for ship energy efficiency anomalies based on a structural causal model as described in any one of claims 1 to 7, characterized in that, include: The multimodal graph construction module is used to collect multimodal performance monitoring data of ships, construct an initial directed acyclic graph, and map the multimodal performance monitoring data into the temporal nodes, spatial edge weights and visual feature vectors of the initial directed acyclic graph to obtain a multimodal operating condition heterogeneous graph. The node completion module is used to perform posterior inference on the missing nodes in the multimodal heterogeneous graph, with the directed dependency relationship of the initial directed acyclic graph as the structural equation constraint and the historical samples of the same working condition as the prior, and backfill the corresponding nodes with the inferred expected value to obtain the completed multimodal heterogeneous graph. The structure update module is used to perform online Bayesian structure update on the completed multimodal heterogeneous graph, using navigation condition indicators consisting of speed, ballast draft, and sea state as trigger conditions, to obtain an adaptive directed acyclic graph. The causal attribution module, based on the conditional probability weights of the adaptive directed acyclic graph, converts the observed deviations of the main engine speed and propeller torque in the sensor time series data into the deviations of the main engine load rate and propeller propulsion efficiency. It defines the overall ship energy efficiency deviation value as the ratio of fuel flow deviation to rated fuel consumption rate. It applies the do operator intervention to the deviations of the main engine load rate and propeller propulsion efficiency on the adaptive directed acyclic graph, calculates the causal contribution rate of each variable to the overall ship energy efficiency deviation value, and decomposes the causal propagation path at three levels of granularity: component level, system level, and overall ship level to obtain a hierarchical attribution result set. The results output module is used to transform the causal propagation paths of each layer in the hierarchical attribution results set into a structured attribution report, determine the operation and maintenance handling priority according to the causal contribution rate from high to low, and output cross-modal tracing attribution results.

9. 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 cross-modal attribution method for ship energy efficiency anomalies based on structural causal models as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the cross-modal attribution method for ship energy efficiency anomalies based on structural causal models as described in any one of claims 1 to 7.

Images