A large model-based ship intelligence research and judgment analysis method and platform

Abstract

The application provides a ship intelligence analysis method and platform based on a large model, and belongs to the technical field of large model analysis, and comprises the following steps: acquiring AIS data, radar images, ship communication messages and other multi-modal historical data and processing the multi-modal historical data in a mode, combining ship maneuvering characteristics and sea area attributes, generating a supplementary sample set through time-space correlation analysis; constructing a single modal feature set according to a mode matching feature extraction mode, outputting an analysis index and a fusion confidence after analyzing a feature distribution vector; integrating an index set, a sample set and a ship field knowledge graph, optimizing a preset large model hierarchical structure, and obtaining a ship enhanced model through multi-stage training, so that structured intelligence can be output by inputting current multi-modal data. The application improves sample integrity and model adaptability, and guarantees analysis accuracy and real-time performance.

Description

Technical Field

[0001] This invention relates to the field of large model analysis technology, and in particular to a method and platform for ship intelligence analysis based on large models. Background Technology

[0002] With the rapid development of the shipping industry, the navigation range of ships is constantly expanding, and the traffic flow in sea areas is continuously increasing, which places higher demands on the comprehensiveness, accuracy, and real-time nature of ship intelligence analysis. Ship intelligence data exhibits multimodal characteristics, mainly including AIS data, radar images, and ship communication messages. These data are scattered across different acquisition devices and systems. How to effectively integrate and deeply analyze multimodal data has become a core technical challenge for ship intelligence analysis.

[0003] In existing technologies, ship intelligence analysis largely relies on single-modal data processing, lacking systematic integration of multimodal data. This leads to biased analysis results that fail to comprehensively reflect the ship's navigation status and the marine environment. Furthermore, due to factors such as data acquisition intervals, equipment malfunctions, and environmental interference during ship navigation, historical sample data often suffers from temporal or spatial gaps, resulting in incomplete model training data and impacting the model's generalization ability and analysis accuracy. In addition, existing technologies do not fully utilize the distributed computing power advantages of industrial cloud computing, nor do they effectively leverage the resource integration capabilities of industrial cloud platforms and the real-time data transmission capabilities of industrial internet platforms. This results in low multimodal data processing efficiency, failing to meet the needs of real-time intelligence analysis.

[0004] In terms of feature extraction and model building, existing methods have insufficient matching between feature extraction methods and modality types, lack specific optimizations for the ship domain, and do not incorporate ship domain knowledge graphs and specific constraints. They also have poor adaptability to ship maneuvering characteristics and sea area attributes, resulting in weak targeting of intelligence analysis indicators and an inability to effectively support accurate analysis.

[0005] These issues collectively result in insufficient accuracy, real-time performance, and reliability of existing ship intelligence analysis methods, making it difficult to meet the high requirements of maritime management, ship operation safety, and other scenarios.

[0006] Therefore, this invention proposes a ship intelligence analysis method and platform based on a large model. Summary of the Invention

[0007] This invention provides a ship intelligence analysis method and platform based on a large model to solve the aforementioned technical problems.

[0008] This invention provides a method for ship intelligence analysis based on a large model, comprising: Step 1: Acquire historical multimodal ship intelligence data and perform submodal processing to obtain an initial sample set for each mode. At the same time, combined with ship maneuvering characteristic constraints and sea area attributes, perform fusion analysis on all initial sample sets according to temporal and spatial correlations to obtain a supplementary sample set. Step 2: Obtain the feature extraction method matching each modality type from the type-extraction lookup table, and extract features from the corresponding modality data according to the feature extraction method to construct a single modality feature set. Analyze the feature distribution vector of the single modality feature set and input it into the distribution network to output the intelligence judgment index and fusion confidence of the corresponding single modality feature set. The modality types include: AIS data modality, radar image modality, and ship communication message modality. Step 3: Perform dynamic analysis on the intelligence assessment indicators under all modal types with all initial sample sets and all supplementary sample sets according to the fusion confidence level, to obtain the first indicator set based on all initial sample sets and the second indicator set based on all supplementary sample sets; Step 4: Based on the knowledge graph of the ship domain, the initial sample set and the supplementary sample set, the first indicator set and the second indicator set, construct and train the preset large model to obtain the ship augmented model, obtain the current multimodal intelligence data after modal processing and input it into the ship augmented model for intelligence analysis and judgment, and output ship intelligence content.

[0009] Preferably, a supplementary sample set is obtained, including: The temporal and spatial labels of each sample in the initial sample set of the same modality type are extracted respectively. Based on the temporal coherence of each temporal label and the characteristics of the ship's navigation stage, the historical time quantity is dynamically determined and historical decoding is performed to obtain the historical potential variable set. Based on the spatial coherence of each spatial label and the constraints of the sea area functional attributes, the future time quantity is dynamically determined and future decoding is performed to obtain the future potential variable set. The temporal labels include: sample collection time, identification of continuous segments of ship track time sequence, and temporal correlation degree of track status. The spatial labels include: ship latitude and longitude, sea area functional attribute level, and spatial neighborhood distance between adjacent ships. A temporal baseline is constructed for each initial sample set, and a first supplementary location point based on the historical latent variable set is determined based on the temporal baseline. Simultaneously, a spatial baseline is constructed for each initial sample set, and a second supplementary location point based on the spatial baseline is determined based on the future latent variable set. Retrieve the spatial labels immediately adjacent to each supplementary location point and analyze the spatial characteristics of the corresponding supplementary location point. At the same time, retrieve the temporal labels immediately adjacent to each supplementary location point and analyze the temporal characteristics of the corresponding supplementary location point. The spatial features are subjected to temporal analysis and the temporal features are subjected to spatial analysis to obtain a single supplementary sample of the corresponding supplementary location point; Extract the temporal-spatial feature vectors of all single supplementary samples under each initial sample set, and perform residual analysis on the temporal-spatial feature vectors of all initial sample sets to obtain retained supplementary samples. Supplementary sample sets are then constructed based on all retained supplementary samples.

[0010] Preferably, obtaining a single supplementary sample at the corresponding supplementary location point includes: Based on the preset time baseline of ship intelligence, historical time series fragments matching the same type of ship in the same sea area and the corresponding spatial features are retrieved. The spatial features and the historical time series fragments are mapped and analyzed to determine the temporal dimension information corresponding to the spatial features and obtain the temporal spatial features. Based on the spatial baseline preset by the ship intelligence, the set of neighboring spatial attributes that are compatible with the corresponding temporal features within the same time interval is retrieved. The temporal features are matched and analyzed with the set of neighboring spatial attributes to determine the spatial dimension information corresponding to the temporal features and obtain the spatialized temporal features. The spatial features, temporal features, temporally sequenced spatial features, and spatially temporally sequenced features of the same supplementary location point are fused across dimensions, and the content that meets the residual threshold is selected by combining the weight of ship interaction relationship, and used as a single supplementary sample for the corresponding supplementary location point.

[0011] Preferably, residual analysis is performed on the temporal-spatial feature vectors of all initial sample sets to obtain retained supplementary samples, including: Extract the temporal-spatial feature vectors of each initial sample set to determine the global and local variables of the residual analysis; Based on the global variables, the initial global residual between the temporal-spatial feature vector and the spatiotemporal reference feature vector is calculated. The initial global residual is then summed with the dynamic disturbance zone value in the ship domain to obtain the adjusted global residual sequence. Based on the local variables, local scene adaptation modal decomposition is performed on the temporal-spatial feature vector to obtain local modal components; Calculate the initial local residuals of each local modal component and the navigation time-space feature vector in the local variables, and calibrate the initial local residuals by combining the contextual correlation features of the local modal components to obtain the calibrated local residual sequence; The fusion weights are determined based on the scene matching degree between the global and local variables, and the fusion residuals of the global residual sequence and the local residual sequence are calculated according to the fusion weights. The sum of squared residuals of all fusion residuals is then calculated. Samples with residual sum of squares lower than the ship scenario adaptive threshold are selected as retained supplementary samples.

[0012] Preferably, analyzing the feature distribution vector of the single-modality feature set includes: Extract the ship-specific feature dimensions of the modality type corresponding to the single modality feature set; Based on the unique feature dimension, the basic statistical distribution vector and the domain association distribution vector based on the single modality feature set are analyzed; The adaptation degree of the basic statistical distribution vector, the domain-related distribution vector and the ship intelligence preset distribution benchmark of the corresponding modality type is calculated to obtain the distribution deviation vector.

[0013] Preferably, a pre-defined large model is constructed and trained to obtain an enhanced ship model, including: Optimize the hierarchical structure of the preset large model; Multi-stage progressive training is performed on the hierarchical structure, and ship domain constraints are introduced during the training process to obtain an enhanced ship model.

[0014] Preferably, the hierarchical structure includes: a knowledge graph fusion layer, a multi-granularity temporal feature parsing layer, an adaptive weighted feature adaptation layer, a cross-modal feature fusion interaction layer, a confidence dynamic calibration layer, and an intelligence analysis and reasoning layer, wherein the rule constraint feature vector output by the knowledge graph fusion layer dynamically adjusts the granularity division threshold of the multi-granularity temporal feature parsing layer.

[0015] Preferably, optimizing the hierarchical structure of the preset large model includes: The knowledge graph fusion layer is configured to perform vectorized embedding extraction of entity-relationships on the knowledge graph of the ship domain, to obtain the knowledge graph entity embedding vector group and entity relation constraint matrix, and to transform the intelligence judgment rules of the ship domain into rule constraint feature vectors. The multi-granularity temporal feature parsing layer is configured to extract temporal correlation feature vectors at different granularities from the multimodal data in the initial sample set and the supplementary sample set based on the granularity partitioning threshold, thereby obtaining a multi-granularity modal temporal feature set. The adaptive weighted feature adaptation layer is configured to encode the weights of the judgment indicators in the multi-granularity modal temporal feature set, the first indicator set, and the second indicator set, respectively. Modal adaptive weight coefficients are introduced, and a three-level fully connected mapping network and a sigmoid activation function are used to generate the weight allocation of each feature. At the same time, an asymmetric adjustment parameter is introduced into the high-dimensional modal feature branch to generate weighted features, which are then mapped to the feature space that matches the knowledge graph entity embedding vector group to obtain adaptive sample-indicator adaptation features. The cross-modal feature fusion interaction layer is configured as a composite feature processing module consisting of a Swin-Transformer feature mapping unit and a ResNet feature extraction unit. The adaptive sample-index matching features are normalized in dimension to construct a cross-modal fusion feature map. Furthermore, it is combined with knowledge graph entity embedding vector groups, entity relationship constraint matrices, and rule constraint feature vectors for deep interactive fusion to output a global fusion feature vector. The confidence dynamic calibration layer is configured to call the fused confidence sequence to perform temporal-spatial joint dynamic adjustment on the weights of each modal component in the global fused feature vector, and generate the calibrated feature vector. The intelligence analysis and reasoning layer is configured as a deep fully connected inference network with a ship domain activation function. It performs classification reasoning and regression operations on the calibrated feature vectors in the ship intelligence dimension, and outputs structured ship intelligence analysis results.

[0016] This invention provides a platform for executing any of the aforementioned ship intelligence analysis methods based on large models.

[0017] Compared with the prior art, the beneficial effects of this application are as follows: By integrating three core modalities of data—AIS data, radar imagery, and ship communication messages—and combining the distributed computing power of industrial cloud computing with the resource integration capabilities of industrial cloud platforms, this solution achieves end-to-end processing of multimodal ship intelligence data. By supplementing the sample set to improve training data, precise feature extraction is achieved based on a type-extraction lookup table, and the large model structure is optimized using a ship domain knowledge graph. The resulting enhanced ship model possesses high adaptability and high analytical accuracy. This solution effectively addresses the problems of incomplete data, insufficient targeted feature extraction, and poor model adaptability in existing technologies. Relying on the real-time data transmission capabilities of the industrial internet platform, it ensures the real-time nature and reliability of intelligence analysis, providing strong support for maritime management and safe ship operations.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of a ship intelligence analysis method based on a large model, as described in an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] This invention provides a method for ship intelligence analysis based on a large model, such as... Figure 1 As shown, it includes: Step 1: Acquire historical multimodal ship intelligence data and perform submodal processing to obtain an initial sample set for each mode. At the same time, combined with ship maneuvering characteristic constraints and sea area attributes, perform fusion analysis on all initial sample sets according to temporal and spatial correlations to obtain a supplementary sample set. Step 2: Obtain the feature extraction method matching each modality type from the type-extraction lookup table, and extract features from the corresponding modality data according to the feature extraction method to construct a single modality feature set. Analyze the feature distribution vector of the single modality feature set and input it into the distribution network to output the intelligence judgment index and fusion confidence of the corresponding single modality feature set. The modality types include: AIS data modality, radar image modality, and ship communication message modality. Step 3: Perform dynamic analysis on the intelligence assessment indicators under all modal types with all initial sample sets and all supplementary sample sets according to the fusion confidence level, to obtain the first indicator set based on all initial sample sets and the second indicator set based on all supplementary sample sets; Step 4: Based on the knowledge graph of the ship domain, the initial sample set and the supplementary sample set, the first indicator set and the second indicator set, construct and train the preset large model to obtain the ship augmented model, obtain the current multimodal intelligence data after modal processing and input it into the ship augmented model for intelligence analysis and judgment, and output ship intelligence content.

[0023] In this embodiment, historical multimodal ship intelligence data refers to historical data generated by ships during past voyages, containing various data types and reflecting the ship's navigation status, location information, and interaction information. It is collected through an industrial internet platform by accessing data collection nodes such as ship AIS receiving equipment, radar monitoring equipment, and communication terminals distributed in various sea areas. Among them, AIS data includes continuous data such as ship name, MMSI code, speed, heading, latitude and longitude; radar images are two-dimensional grayscale radar images of the ship, containing information such as the ship's outline, distance from the radar station, and relative motion trajectory; ship communication messages include VHF communication text records between the ship and the port dispatch center and other ships.

[0024] In this embodiment, for AIS data mode, the modality processing includes data format standardization, data time sequence alignment, and invalid data removal; for radar image mode, the modality processing includes image grayscale conversion, removal of salt-and-pepper noise using a median filtering algorithm, extraction of ship target areas using a threshold segmentation algorithm, and removal of marine background interference; for ship communication message mode, the modality processing includes text segmentation, stop word removal, and key information extraction.

[0025] In this embodiment, the maneuverability parameters of different types of ships, such as cargo ships, fishing boats, dangerous goods ships, and passenger ships, are collected through channels such as ship design manuals, ship operation logs, and ship inspection reports. A ship maneuverability constraint database is established and stored on an industrial cloud platform. For example, the maneuverability constraints of a cargo ship include a maximum speed of 25 knots, which will not be exceeded in actual navigation; a maximum turning angular velocity of 0.5 degrees / second, which must be within the range of its turning angular velocity; and a minimum braking distance of 500 meters at a speed of 10 knots, which must be no less than the travel distance during emergency braking.

[0026] In this embodiment, the sea area attributes include: Marine functional attributes: such as the core area of ​​the port with a water depth of 15-20 meters, which allows large ships to berth and operate; the water depth of the channel area with a water depth of 12-15 meters, where ships must navigate along fixed routes and the speed is limited to 10-15 knots; and the water depth of the open sea navigation area is greater than 50 meters, where there is no strict speed limit but ships must avoid windy and wave areas, etc. Marine geographical attributes: such as latitude and longitude range of 29°-31° north latitude and 120°-122° east longitude, water depth of 10-30 meters, and flat seabed topography; Marine environmental attributes: such as wind and wave level 2-4, current speed 0.5-1 knot and visibility 10-15 nautical miles; The sea area has the following traffic characteristics: a ship density of 5-10 ships per square nautical mile, an east-west route, and traffic control rules requiring ships to keep to the right.

[0027] In this embodiment, the temporal correlation includes the order of sample collection time, the temporal continuity of track data, and the temporal synchronization of different modal data, etc. The spatial correlation includes the spatial proximity of ship positions, the spatial overlap of track routes, and the spatial correspondence between ships and marine features such as ports, waterways, islands, etc.

[0028] In this embodiment, the type-extraction lookup table refers to a pre-established standardized table that records the mapping relationship between different modality types and corresponding feature extraction methods, as shown in Table 1, which contains part of the lookup table: Table 1 Type-Extraction Reference Table

[0029] In this embodiment, the basic statistical distribution vector is extracted using the mean and variance calculation method, and the domain association distribution vector is calculated using cosine similarity. By calculating the fit between the first two and the preset ship domain feature distribution benchmark, the distribution deviation vector is obtained. The three are combined to form the feature distribution vector. For example, the basic statistical distribution vector of the single feature set of AIS data modality is [14.2, 3.1, 0.8], which corresponds to the mean speed, standard deviation of heading, and latitude and longitude distribution concentration, respectively; the domain association distribution vector is [0.92, 0.88, 0.75], which corresponds to the matching degree with cargo ship domain features, the matching degree with waterway navigation features, and the matching degree with normal navigation status, respectively; the distribution deviation vector is [0.05, 0.03, 0.08], which corresponds to the deviation value from the preset cargo ship AIS data feature distribution benchmark, respectively.

[0030] In this embodiment, the distributed network serves as a mapping model from feature distribution vectors to evaluation indicators and fusion confidence scores, with the following specific parameters: Network structure: Input layer → Hidden layer 1 → Hidden layer 2 → Hidden layer 3 → Output layer (two branches).

[0031] Input layer: Dimension = 20 (corresponding to the dimension of the feature distribution vector, including 8 dimensions of basic statistical distribution vector, 8 dimensions of domain association distribution vector, and 4 dimensions of distribution deviation vector).

[0032] Hidden layer parameters: Hidden layer 1: Number of neurons = 64, activation function = ReLU, dropout rate = 0.1, weight initialization method = He normal distribution. Hidden layer 2: Number of neurons = 32, activation function = ReLU, dropout rate = 0.1, regularization method = L2. Hidden layer 3: Number of neurons = 16, activation function = ReLU, dropout rate = 0.05.

[0033] Output Layer: Branch 1 (Intelligence Analysis Indicators): Number of neurons = 5, activation function = Sigmoid, output range [0,1], corresponding to track anomaly degree, navigation compliance degree, sea area adaptability, ship identification accuracy, and interaction risk degree. Branch 2 (Fusion Confidence): Number of neurons = 1, activation function = Sigmoid, output range [0,1], confidence ≥ 0.6 is considered a valid indicator.

[0034] Training parameters: epoch=50, batchsize=32, learning rate=3e-4, optimizer=Adam (β1=0.9, β2=0.999), loss function=cross-entropy loss+MSE loss (weight ratio=0.7:0.3).

[0035] For example, when the feature distribution vector of a certain AIS data modality is input [14.2,3.1,0.8,0.92,0.88,0.75,0.05,0.03,0.08,...] (20 dimensions in total), the output intelligence analysis index is [0.12,0.95,0.89,0.98,0.03], and the fusion confidence is 0.92. Among them, 0.12 indicates that the track anomaly is slightly deviated, 0.95 indicates that the navigation compliance is basically compliant, 0.89 indicates that the sea area adaptability is moderately adapted, 0.98 indicates that the ship identification accuracy is high, and 0.03 indicates that the interaction risk is extremely low.

[0036] In this embodiment, the fusion confidence score is used as the weight to weight the indicators of the same core dimension, and unreliable indicators with a fusion confidence score below the threshold of 0.6 are removed. The integrated effective indicators are associated and stored according to the classification of the initial sample set to form the first indicator set. The second indicator set is constructed using the same algorithm and process as the first indicator set, which will not be described in detail here.

[0037] In this embodiment, the Neo4j knowledge graph construction tool is used to construct a knowledge graph for the ship domain. For example, the entities in the ship domain knowledge graph include ship entities, sea area entities, equipment entities, and rule entities; the relationships between entities include ship-equipment-equipment, ship-navigation-sea area, equipment-collection-data, and rule-constraint-ship behavior. The judgment rules include that the speed of cargo ships in the core area of ​​the port shall not exceed 10 knots, and dangerous goods ships must be detected by radar in the designated waterway and an alarm must be triggered if the ship's position deviates from the route by more than 5 nautical miles.

[0038] In this embodiment, the ship-enhanced model refers to a deep learning model adapted to ship intelligence analysis scenarios, obtained by optimizing the hierarchical structure of a pre-set large model and combining it with a ship domain knowledge graph, an initial sample set, a supplementary sample set, a first indicator set, and a second indicator set for multi-stage training.

[0039] In this embodiment, current multimodal intelligence data refers to data collected within the current moment, such as within 5 minutes, that has the same modality type as historical multimodal ship intelligence data.

[0040] In this embodiment, three sets of comparative experiments were designed. The experimental data were based on 100,000 pieces of ship multimodal data (including 30,000 AIS data, 30,000 radar images, and 40,000 ship communication messages), covering scenarios such as port core areas, waterway areas, and open sea navigation areas. The comparison objects were a single-modal model, a multimodal model without knowledge graph fusion, and a traditional sample interpolation supplementation model. The experimental indicators included judgment accuracy, real-time performance, and sample utilization rate, as shown in Table 2.

[0041] The beneficial effects of the above technical solution are as follows: By integrating three core modal data types—AIS data, radar images, and ship communication messages—and combining the distributed computing power of industrial cloud computing with the resource integration capabilities of the industrial cloud platform, the entire process of multimodal ship intelligence data processing is achieved. By supplementing the sample set to improve training data, achieving accurate feature extraction based on a type-extraction lookup table, and optimizing the large model structure with the help of a ship domain knowledge graph, the final enhanced ship model possesses high adaptability and high judgment accuracy. This solution effectively solves the problems of incomplete data, insufficient targeted feature extraction, and poor model adaptability in existing technologies. Relying on the real-time data transmission capabilities of the industrial internet platform, it ensures the real-time nature and reliability of intelligence analysis, providing strong support for maritime management and safe ship operation.

[0042] This invention provides a ship intelligence analysis method based on a large model, which obtains a supplementary sample set, including: The temporal and spatial labels of each sample in the initial sample set of the same modality type are extracted respectively. Based on the temporal coherence of each temporal label and the characteristics of the ship's navigation stage, the historical time quantity is dynamically determined and historical decoding is performed to obtain the historical potential variable set. Based on the spatial coherence of each spatial label and the constraints of the sea area functional attributes, the future time quantity is dynamically determined and future decoding is performed to obtain the future potential variable set. The temporal labels include: sample collection time, identification of continuous segments of ship track time sequence, and temporal correlation degree of track status. The spatial labels include: ship latitude and longitude, sea area functional attribute level, and spatial neighborhood distance between adjacent ships. A temporal baseline is constructed for each initial sample set, and a first supplementary location point based on the historical latent variable set is determined based on the temporal baseline. Simultaneously, a spatial baseline is constructed for each initial sample set, and a second supplementary location point based on the spatial baseline is determined based on the future latent variable set. Retrieve the spatial labels immediately adjacent to each supplementary location point and analyze the spatial characteristics of the corresponding supplementary location point. At the same time, retrieve the temporal labels immediately adjacent to each supplementary location point and analyze the temporal characteristics of the corresponding supplementary location point. The spatial features are subjected to temporal analysis and the temporal features are subjected to spatial analysis to obtain a single supplementary sample of the corresponding supplementary location point; Extract the temporal-spatial feature vectors of all single supplementary samples under each initial sample set, and perform residual analysis on the temporal-spatial feature vectors of all initial sample sets to obtain retained supplementary samples. Supplementary sample sets are then constructed based on all retained supplementary samples.

[0043] In this embodiment, ; in, It is the minimum value; For historical moments; Measured for future moments; This is the quantification value of the temporal coherence of the current sample's time series label; This is a correction factor for the ship's navigation phase, and , The temporal correlation degree of the current sample's track state; For the time-series consistency of similar vessels across the entire sea area; The historical time reference for the same type of ships; This is the quantification value of the spatial coherence of the current sample spatial label; For the sea area attribute adaptation coefficient, and , This is the normalized value of the spatial neighborhood distance between the current sample and its adjacent ships; For spatial reference consistency across the entire sea area; For future time-based spatial reference quantities of the same sea area type; This represents the spatial constraint threshold corresponding to the current functional attribute level of the sea area. This is the floor symbol.

[0044] In this embodiment, the functional attribute levels of the sea area are: Level 1 is the port core area, Level 2 is the waterway area, Level 3 is the open sea navigation area, and Level 4 is the fishery operation area.

[0045] In this embodiment, historical decoding refers to the process of parsing and extracting features from historical sample data corresponding to historical time points, mining the potential information contained in the historical samples, and forming a set of historical latent variables. The decoding objects include the time series labels, spatial labels, and original data features of the historical samples. The set of historical latent variables refers to the set of variables obtained through historical decoding that reflects the core characteristics and potential patterns of the historical samples. Each latent variable corresponds to a key feature dimension of the historical sample. The variable dimensions are dynamically adjusted according to the modal type and ship type. It should be noted that the principle of future decoding is similar to that of future latent variable sets, and will not be elaborated here. For example: AIS data modality: 15 potential variables, including mean rate of change of speed, course stability, distance of position movement, sea area attribute fit, frequency of interaction between adjacent vessels, number of track inflection points, speed fluctuation coefficient, smoothness of course change, concentration of latitude and longitude distribution, fit of navigation phase, equipment working status, signal transmission stability, data integrity, temporal correlation strength, and spatial correlation strength.

[0046] Radar image modalities: 18 latent variables, including ship contour change features, position positioning accuracy trend, image texture similarity, target discrimination change, grayscale mean stability, edge detection accuracy, background noise suppression effect, target tracking continuity, distance measurement accuracy, relative velocity calculation error, image acquisition angle adaptability, sea area illumination influence coefficient, fog penetration capability, ship size recognition accuracy, track prediction deviation, neighboring target interference degree, image frame correlation degree, and feature extraction repetition rate.

[0047] Ship communication message modality: 12 potential variables, including semantic consistency trend, frequency of key information occurrence, stability of interaction objects, message length fluctuation coefficient, keyword matching degree, syntactic regularity, timestamp accuracy, information integrity, encrypted transmission security, communication signal strength, multi-ship interaction conflict rate, and rule compliance.

[0048] In this embodiment, the functional attribute constraints of the sea area refer to the constraints on ship navigation behavior and spatial characteristics based on the functional attribute level of the sea area where the sample is located. The constraint requirements are different for different functional attribute levels of the sea area, mainly reflected in speed limits, navigation area range, and ship density limits. For example, the constraints for the port core area (sea area functional attribute level 1) are: speed ≤ 10 knots, navigation area limited to within 1 nautical mile around the designated berth, and ship density ≤ 8 ships / square nautical mile.

[0049] In this embodiment, the time baseline is usually established with the collection time of the first sample in the initial sample set as the starting point, and an isochronous time axis is established according to the sample collection time interval. It includes three core parameters: the time axis origin, the time interval, and the time range. The first supplementary position point is the time position corresponding to the missing historical collection moment in the initial sample set, and each first supplementary position point corresponds to a missing historical moment.

[0050] In this embodiment, the spatial baseline is usually established with the average latitude and longitude of all samples in the initial sample set as the origin. Combined with the spatial range of the sea area functional attribute level, a two-dimensional spatial coordinate system is established, which includes three core parameters: spatial origin, coordinate unit, and spatial boundary. The second supplementary position point is the coordinate point corresponding to the missing spatial position in the initial sample set, and each second supplementary position point corresponds to a missing spatial coordinate.

[0051] In this embodiment, the spatial characteristics of the supplementary location points refer to the attribute characteristics of each supplementary location point in the spatial dimension. For example, the spatial characteristics of the first supplementary location point are: latitude and longitude 30.08°N, longitude 120.15°E, sea area functional attribute level 2, predicted spatial neighborhood distance of adjacent ships 2.5 nautical miles, and spatial correlation strength 0.91 (indicating the degree of spatial correlation with surrounding samples); the spatial characteristics of the second supplementary location point are: latitude and longitude 30.125°N, longitude 120.227°E, sea area functional attribute level 2, predicted spatial neighborhood distance of adjacent ships 3.1 nautical miles, and spatial correlation strength 0.88.

[0052] The temporal features of supplementary location points refer to the attribute features of each supplementary location point in the time dimension. For example, the temporal features of the first supplementary location point are: acquisition time T02, track temporal continuous segment identifier TRACK_T02, track status temporal correlation prediction value 0.93, and temporal correlation strength 0.89, which represent the degree of temporal correlation with surrounding samples; the temporal features of the second supplementary location point are: acquisition time prediction value T03, track temporal continuous segment identifier TRACK_T03, track status temporal correlation prediction value 0.90, and temporal correlation strength 0.86.

[0053] In this embodiment, a time-series analysis model is constructed based on a knowledge graph and historical sample database in the shipbuilding field. The spatial features and corresponding time ranges of the supplementary location points are input, and the temporal dimension patterns corresponding to the spatial features are mined through the ARIMA model. The time-series spatial feature parameters are output to achieve time-series analysis of spatial features. Similarly, a spatial analysis model is constructed based on a marine attribute database and a spatial association rule base. The time-series features and corresponding spatial ranges of the supplementary location points are input, and the spatial dimension patterns corresponding to the time-series features are mined through the K-means algorithm. The spatialized time-series feature parameters are output to achieve spatial analysis of time-series features.

[0054] In this embodiment, a single supplementary sample refers to a single complete sample that supplements the missing data in the initial sample set by integrating the spatial features, temporal features, temporally sequenced spatial features, and spatially sequenced temporal features of each supplementary location point through cross-dimensional fusion and filtering. For example, the single supplementary sample (AIS data modality) of the first supplementary location point includes: original data (speed 11.5 knots, heading 92 degrees, latitude and longitude 30.08°N 120.15°E), temporal tags (collection time T02, track temporal continuous segment identifier TRACK_T02, track status temporal correlation degree 0.93), and spatial tags (sea area functional attribute level 2, adjacent ship spatial neighborhood distance 2.5 nautical miles), which is completely consistent with the AIS sample structure of the initial sample set.

[0055] In this embodiment, the temporal-spatial feature vector refers to the comprehensive feature extracted from a single supplementary sample that simultaneously contains core information of both the temporal and spatial dimensions. It is a high-level summary of the single supplementary sample data, and its core features include temporal-spatial correlation strength, spatiotemporal feature change rate, and spatiotemporal constraint fit. For example, the temporal-spatial feature vector of a single supplementary sample (radar image modality) is as follows: temporal-spatial correlation strength 0.92 (indicating the degree of correlation between the temporal and spatial features of the sample), spatiotemporal feature change rate 0.05 (indicating the magnitude of spatiotemporal feature change between the sample and its neighboring samples), and spatiotemporal constraint fit 0.89 (indicating the degree to which the sample conforms to the constraints of ship maneuvering characteristics and sea area attributes).

[0056] The beneficial effects of the above technical solution are as follows: By extracting the temporal and spatial labels of the samples and dynamically determining historical / future time values ​​based on features such as temporal and spatial coherence, a set of latent variables is obtained through decoding. This allows for the identification of supplementary location points and the generation of single supplementary samples. Finally, residual analysis is used to filter and retain these supplementary samples, constructing a supplementary sample set. This effectively compensates for the temporal and spatial deficiencies in the initial sample set, improving the completeness and quality of the sample data. Compared to the simple interpolation method used in existing technologies for supplementing samples, the supplementary samples generated by this solution better conform to ship navigation patterns and sea area attribute constraints, providing a high-quality data foundation for subsequent model training and further improving the accuracy of ship intelligence analysis.

[0057] This invention provides a ship intelligence analysis method based on a large model, which obtains a single supplementary sample of corresponding supplementary location points, including: Based on the preset time baseline of ship intelligence, historical time series fragments matching the same type of ship in the same sea area and the corresponding spatial features are retrieved. The spatial features and the historical time series fragments are mapped and analyzed to determine the temporal dimension information corresponding to the spatial features and obtain the temporal spatial features. Based on the spatial baseline preset by the ship intelligence, the set of neighboring spatial attributes that are compatible with the corresponding temporal features within the same time interval is retrieved. The temporal features are matched and analyzed with the set of neighboring spatial attributes to determine the spatial dimension information corresponding to the temporal features and obtain the spatialized temporal features. The spatial features, temporal features, temporally sequenced spatial features, and spatially temporally sequenced features of the same supplementary location point are fused across dimensions, and the content that meets the residual threshold is selected by combining the weight of ship interaction relationship, and used as a single supplementary sample for the corresponding supplementary location point.

[0058] In this embodiment, the preset time baseline for ship intelligence refers to a standard time reference system pre-set based on the business needs of ship intelligence analysis and historical data statistics, applicable to specific sea areas and ship types. Core parameters include the time baseline origin, time interval, and time window size, set by ship industry experts in conjunction with actual application scenarios. For example, the preset time baseline for cargo ships in the port core area (sea area functional attribute level 1) is: time baseline origin 00:00:00, time interval 1 minute, and time window size 1 hour.

[0059] Vessels of the same type in the same sea area refer to vessels with the same sea area functional attribute level and the same vessel type. For example, if the current supplementary location point is located in the waterway area (sea area functional attribute level 2) and the corresponding vessel type is a container ship, then vessels of the same type in the same sea area are all container ships that have sailed in that waterway area.

[0060] In this embodiment, the historical time series fragment set refers to a continuous time segment data set in the historical navigation data of the same type of ship in the same sea area that matches the spatial characteristics of the current supplementary location point, and includes the original data, time series label, and spatial label of the corresponding historical moment.

[0061] In this embodiment, mapping analysis refers to the correlation analysis between the spatial features of the current supplementary location point and the historical time series fragment set, to explore the correspondence between the spatial features and the temporal dimension information in the historical time series fragments. This includes three steps: feature matching, pattern extraction, and information mapping. For example, mapping analysis is performed on the spatial features of the current supplementary location point (channel area, 30.10°N, 120.20°E, distance to adjacent ships 2.8 nautical miles) and the historical time series fragment set. The analysis reveals that the temporal dimension information corresponding to this spatial feature in the historical time series fragments is: speed 12-14 knots, heading change rate 0.1-0.2 degrees / second, and track status temporal correlation degree 0.9-0.95. Through mapping analysis, this temporal dimension information is assigned to the current spatial features, resulting in temporally sequenced spatial features and temporal dimension information. Information refers to information related to spatial characteristics that reflects the motion status and patterns of a ship in the time dimension. The core includes quantifiable time-series parameters such as speed range, rate of change of course, track continuity, and frequency of interaction with adjacent ships. For example, the time-series information corresponding to the spatial characteristics of the current supplementary location point is: speed range 13-15 knots, rate of change of course 0.08-0.15 degrees / second, track continuity 0.92-0.96, and frequency of interaction with adjacent ships 0.2 times / minute (i.e., one interaction with adjacent ships every 5 minutes).

[0062] In this embodiment, a preset spatial baseline library is constructed based on the marine attribute database and ship navigation time statistics. Spatial baseline parameters are stored according to time interval and ship type. For example, the preset spatial baseline for passenger ships in the channel area (level 2) during the daytime (06:00-18:00) is: the origin of the spatial baseline is the midpoint of the channel centerline (30.00°N, 120.00°E), the coordinate unit is nautical miles, and the spatial window size is 10 nautical miles × 10 nautical miles.

[0063] In this embodiment, the time interval is usually determined by taking the acquisition time of the supplementary location point as the center and using a 1-minute window size, and then retrieving the neighborhood spatial attribute set that matches the time sequence features within the time interval.

[0064] In this embodiment, the boundary coordinates of the neighborhood range are determined based on the coordinates of the supplementary location point and the spatial window size of the preset spatial baseline. All spatial attribute data of the neighborhood range within the same time interval are retrieved from the historical database of marine attributes, and classified and organized according to attribute type to form a neighborhood spatial attribute set. For example, if the same time interval of the supplementary location point is T04 and the preset spatial window size is 10 nautical miles × 10 nautical miles, its neighborhood spatial attribute set is as follows: the distribution of marine functional attributes in the neighborhood is that the waterway area accounts for 80% and the open sea navigation area accounts for 20%; the water depth distribution is 12-15 meters; the ship density distribution is 3-5 ships / square nautical mile; the traffic control rules are that the ship should sail on the right and the speed limit is 10-15 knots.

[0065] In this embodiment, a matching analysis model is constructed. The temporal features of the supplementary location point and the set of neighborhood spatial attributes are input. The matching relationship between the temporal features and spatial information in the neighborhood spatial attributes is extracted by machine learning algorithm. Based on the matching relationship, the spatial dimension information corresponding to the current temporal feature is predicted, realizing the matching analysis from temporal features to spatial information. For example, the temporal features of the current supplementary location point (speed 14 knots, heading 95 degrees, track status temporal correlation degree 0.94) are matched with the set of neighborhood spatial attributes. It is found that the neighborhood spatial attributes corresponding to the temporal feature are: water depth 13-14 meters, ship density 4 ships / square nautical mile, and located in a straight section of the channel (heading adjustment is restricted). By matching analysis, these spatial dimension information are assigned to the current temporal feature to obtain the spatialized temporal feature. The spatialized temporal feature formed after fusion is: speed 14 knots - heading 95 degrees - track status temporal correlation degree 0.94 - water depth 13-14 meters - ship density 4 ships / square nautical mile - straight section of the channel.

[0066] At this time, spatial features (channel area, 30.10°N, 120.20°E), temporal features (speed 14 knots, course 95 degrees), temporalized spatial features (speed 13-15 knots - channel area), and spatialized temporal features (water depth 12-14 m - speed 14 knots) are subjected to cross-dimensional fusion, and the extracted comprehensive feature vector is [channel area, 30.10°N, 120.20°E, speed 14 knots, course 95 degrees, water depth 13 m, ship density 3.5 ships per square nautical mile], which comprehensively covers the core information of four dimensions.

[0067] In this embodiment, the ship interaction relationship weight is used for screening single supplementary samples, and the calculation steps are as follows: Input parameters: ship type (Type), sea area functional attribute level (Level), navigation stage (Stage), distance between adjacent ships (D).

[0068] Ship interaction relationship weight W0=w1×f(Type)+w2×g(Level)+w3×h(Stage)+w4×k(D).

[0069] w1=0.4, w2=0.3, w3=0.2, w4=0.1: parameter weights.

[0070] f(Type): ship type coefficient: cargo ship=1.0, fishing vessel=1.2, dangerous goods ship=1.5, passenger ship=1.1.

[0071] g(Level): sea area level coefficient: level 1=1.4, level 2=1.1, level 3=0.7, level 4=0.9.

[0072] h(Stage): navigation stage coefficient: departure=1.2, cruising=1.0, berthing=0.8.

[0073] k(D): distance coefficient: D≤2 nautical miles=1.0, 2<D≤5 nautical miles=0.8, D>5 nautical miles=0.5.

[0074] In this embodiment, the residual threshold is a decision threshold used for screening cross-dimensional fusion results, which is set based on historical fusion data statistics in the ship field and expert experience, and quantifies the deviation degree between the fusion result and ship navigation rules and sea area attribute constraints. If the residual of the fusion result is lower than the threshold, the fusion result is determined to be valid and can be used as a single supplementary sample. For example, the residual threshold of the AIS data modality is 0.05, the residual threshold of the radar image modality is 0.08, and the residual threshold of the ship communication message modality is 0.06.

[0075] The beneficial effects of the above technical solution are: it fully utilizes historical data resources and domain knowledge bases to achieve deep integration of features across different dimensions, generating a single supplementary sample that possesses both temporal and spatial completeness and rationality. Compared to existing technologies that generate supplementary samples based on a single dimension, this solution effectively improves the quality of supplementary samples and their adaptability to actual navigation scenarios, providing high-quality single-sample support for the construction of supplementary sample sets and further enhancing the effectiveness of model training data.

[0076] This invention provides a ship intelligence analysis method based on a large model, which performs residual analysis on the temporal-spatial feature vectors of all initial sample sets to obtain retained supplementary samples, including: Extract the temporal-spatial feature vectors of each initial sample set to determine the global and local variables of the residual analysis; Based on the global variables, the initial global residual between the temporal-spatial feature vector and the spatiotemporal reference feature vector is calculated. The initial global residual is then summed with the dynamic disturbance zone value in the ship domain to obtain the adjusted global residual sequence. Based on the local variables, local scene adaptation modal decomposition is performed on the temporal-spatial feature vector to obtain local modal components; Calculate the initial local residuals of each local modal component and the navigation time-space feature vector in the local variables, and calibrate the initial local residuals by combining the contextual correlation features of the local modal components to obtain the calibrated local residual sequence; The fusion weights are determined based on the scene matching degree between the global and local variables, and the fusion residuals of the global residual sequence and the local residual sequence are calculated according to the fusion weights. The sum of squared residuals of all fusion residuals is then calculated. Samples with residual sum of squares lower than the ship scenario adaptive threshold are selected as retained supplementary samples.

[0077] In this embodiment, the method for determining the adaptive threshold for the ship scene is as follows: Based on the classification of ship type and sea area attributes, a threshold database was established, covering 6 types of ships, including cargo ships, fishing boats, and dangerous goods ships, and 4 types of sea areas, including port core areas and waterway areas.

[0078] Calculate the mean of the sum of squared residuals of historical valid samples for each scenario. and standard deviation Threshold= +0.5 .

[0079] In this embodiment, the initial global residual The determination method is as follows: ,in, This is the temporal-spatial feature vector of the current initial sample set; This refers to the spatiotemporal reference feature vector of ship navigation across the entire sea area in the global variables; The global energy benchmark for the spatiotemporal constraints of ships across the entire sea area in the global variables; This is the preset reference value for the global energy reference.

[0080] In this embodiment, For container ships, the following settings are used: [15 knots (baseline speed), 90° (baseline heading), 25°-35° North latitude and 115°-125° East longitude (baseline latitude and longitude range), 0.1 degrees / second (baseline speed change rate), 4 ships / square nautical mile (baseline ship density)].

[0081] The value range is [80, 120]. The value is 100.

[0082] In this embodiment, the global residual sequence is defined as: Its element calculation formula is: ,in, For the initial global residual The i1th component; This refers to randomly selected values ​​for dynamic disturbance zones in the shipbuilding field, and the range of dynamic disturbance zones is [value missing]. k0 is the ship domain coefficient; is the sequence value of the i1th global residual.

[0083] In this embodiment, the value of k0 is obtained by matching the ship type-navigation area-coefficient lookup table. For example, for a cargo ship in coastal waters, k0 = 0.1; for a cargo ship in offshore waters, k0 = 0.15; for a fishing vessel in coastal waters, k0 = 0.2, etc. In this embodiment, the initial local residual The calculation formula is: ,in, The eigenvectors of the local modal components; This refers to the current ship type's navigation time-space feature vector in the local variables; Let be the current ship type fit coefficient in the local variables, and ; In this embodiment, the calibrated local residual sequence is: Its element calculation formula is: ,in, For the initial local residual The i2th component; This is a calibration coefficient for the marine industry, with a value of 0.2. The context-related feature quantization value for local modal components; is the sequence value of the i2th component.

[0084] In this embodiment, global variables refer to the set of variables that reflect the common laws of navigation of all types of ships in the entire sea area during the residual analysis process, while local variables refer to the set of variables that reflect the individual laws of a specific ship, a specific sea area, and a specific navigation stage corresponding to the current initial sample set during the residual analysis process.

[0085] In this embodiment, the spatiotemporal reference feature refers to the reference feature vector contained in the global variables, which reflects the common spatiotemporal patterns of navigation of the same type of ships in the entire sea area. It is a reference benchmark for calculating the initial global residual. The core features include parameters such as reference speed, reference heading, reference latitude and longitude distribution range, and reference speed change rate, which are obtained from historical data of the entire sea area. For example, the spatiotemporal reference feature vector of container ships in the entire sea area is [reference speed 15 knots, reference heading 90 degrees, reference latitude and longitude distribution range 25°-35°N, 115°-125°E, reference speed change rate 0.1 degrees / second, reference ship density 4 ships / square nautical mile].

[0086] In this embodiment, the dynamic disturbance zone value in the ship domain refers to a numerical range reflecting the impact of various dynamic disturbance factors within the ship domain, such as wind, waves, currents, and temporary equipment failures, on ship navigation. It is used to adjust the initial global residual, making it more closely reflect the actual navigation scenario. The range of the disturbance zone value is set by ship domain experts in conjunction with historical disturbance data. For example, the dynamic disturbance zone value range for a cargo ship navigating in near-shore waters is [-3, 3], indicating a relatively small disturbance impact; while the dynamic disturbance zone value range for a fishing vessel navigating in the open ocean is [-6, 6], indicating a larger disturbance impact and significant influence from wind and waves.

[0087] In this embodiment, the Empirical Mode Decomposition (EMD) algorithm with local scene adaptation is used to decompose the system into four local modal components: speed modal component, heading modal component, position modal component, and interaction modal component. Each local modal component refers to a local feature component that constitutes the temporal-spatial feature vector, obtained through modal decomposition with local scene adaptation. Each component focuses on a specific dimension of the temporal-spatial feature vector and is adapted to the current local scene, forming the basis for calculating the initial local residuals. For example, the speed modal component is [14.2, 14.3, 14.4, 14.5]; the heading modal component is [91, 92, 93, 94]; and the position modal component is [30.04°N, 30.05°N, 30.06°N, 30.07°N].

[0088] In this embodiment, the navigation time-space feature vector refers to the feature vector contained in the local variables that reflects the spatiotemporal characteristics of a specific ship in a local scenario. It serves as a reference benchmark for calculating the initial local residual. The core features include the ship's personalized speed, heading, position, and interaction parameters, which are obtained from the analysis of the current initial sample set. For example, the navigation time-space feature vector of the current cargo ship during the cruise phase in the waterway area is [personalized speed 14.3 knots, personalized heading 92.5 degrees, personalized position 30.05°N 120.10°E, personalized speed change rate 0.08 degrees / second].

[0089] In this embodiment, contextual correlation features refer to the features corresponding to each local modal component that reflect the correlation between that component and the preceding and following local modal components. The core features include parameters such as similarity between components, consistency of change trends, and correlation strength, which are used to calibrate the initial local residuals and improve the accuracy of residual calculation. For example, the contextual correlation features of the speed modal component [14.2, 14.3, 14.4, 14.5] are: similarity to the preceding component (position modal component) 0.85, consistency of change trend 0.9, and correlation strength 0.8; similarity to the following component (heading modal component) 0.75, consistency of change trend 0.85, and correlation strength 0.7.

[0090] In this embodiment, the scenario matching degree refers to the degree of matching between the common scenario reflected by the global variable and the individual scenario reflected by the local variable, quantified as a value between 0 and 1. For example, if the common scenario reflected by the global variable is cargo ship - open sea navigation area - cruising phase, and the individual scenario reflected by the local variable is container ship - open sea navigation area - cruising phase, the scenario matching degree is 0.95 (high matching); if the individual scenario is fishing boat - port core area - departure phase, the scenario matching degree is 0.3 (low matching).

[0091] In this embodiment, the fusion weight refers to the weight coefficients used to weight the fusion of the global residual sequence and the local residual sequence, determined based on the scene matching degree. It includes the global fusion weight and the local fusion weight, the sum of which is 1. The weight reflects the importance of the corresponding residual sequence in the fusion process. Specifically, a fusion weight calculation model is constructed, the scene matching degree is input, and the fusion weight is calculated through a linear mapping algorithm (e.g., when the scene matching degree is ≥0.8, the global weight = 0.6 and the local weight = 0.4; when the scene matching degree is 0.5 ≤ the scene matching degree <0.8, the global weight = 0.4 and the local weight = 0.6; when the scene matching degree <0.5, the global weight = 0.2 and the local weight = 0.8), and stored in the fusion weight library.

[0092] In this embodiment, the fusion residual refers to the residual obtained by weighted summation of the global residual sequence and the local residual sequence according to the fusion weight. It comprehensively reflects the overall deviation of a single supplementary sample from common and individual patterns and is the basis for calculating the sum of squared residuals. For example, if the global residual sequence is [0.05, 0.04, 0.03, 0.03], the local residual sequence is [0.0084, 0, 0.00816, 0.0156], the global fusion weight is 0.6, and the local fusion weight is 0.4, the calculated fusion residual is [0.03336, 0.024, 0.020464, 0.02424].

[0093] In this embodiment, the residual sum of squares refers to the value obtained by summing the squares of all fused residual components. It is a comprehensive indicator that quantifies the degree of deviation between a single supplementary sample and the overall ship navigation pattern. The smaller the value, the more the sample conforms to the navigation pattern and the more suitable it is as a retained supplementary sample.

[0094] The beneficial effects of the above technical solution are: by distinguishing between global and local variables in residual analysis, taking into account the common laws of ship navigation in the entire sea area, and fitting the individual scenarios of specific ships and sea areas, it effectively eliminates invalid samples that deviate from the constraints of ship maneuvering characteristics and sea area attributes, significantly improves the reliability, adaptability and data quality of the supplementary sample set, solves the problem of data messiness caused by the lack of targeted screening in traditional sample supplementation, and provides high-quality sample support for the efficient training of subsequent ship augmentation models.

[0095] This invention provides a ship intelligence analysis method based on a large model, which analyzes the feature distribution vector of the single modality feature set, including: Extract the ship-specific feature dimensions of the modality type corresponding to the single modality feature set; Based on the unique feature dimension, the basic statistical distribution vector and the domain association distribution vector based on the single modality feature set are analyzed; The adaptation degree of the basic statistical distribution vector, the domain-related distribution vector and the ship intelligence preset distribution benchmark of the corresponding modality type is calculated to obtain the distribution deviation vector.

[0096] In this embodiment, for example, the ship-specific feature dimensions of the AIS data modality include four core dimensions: speed stability dimension, which quantifies the degree of speed fluctuation; course continuity dimension, which describes the smoothness of course changes; track deviation dimension, which measures the deviation between the track and the planned route; and sea area adaptability dimension, which assesses the degree of matching between the ship's navigation and sea area attributes.

[0097] In this embodiment, the basic statistical distribution vector refers to the vector data calculated by statistical analysis methods on the unique feature dimension of a single modal feature set in the ship domain. For example, the basic statistical distribution vector of the single feature set of AIS data modality is: [mean speed, variance of speed, median of heading, variance of heading, mean of track deviation, variance of track deviation, mean of sea area fit].

[0098] In this embodiment, the domain association distribution vector refers to the vector data formed by analyzing the degree of association between a single modal feature set and other core elements of the ship domain, such as ship type, navigation stage, sea area attributes, and maritime rules, on the dimension of ship domain-specific features. For example, the domain association distribution vector of AIS data modality is: [matching degree with cargo ship type 0.95, adaptability with cruise stage 0.93, fit with waterway area 0.91, conformity with maritime navigation rules 0.94].

[0099] In this embodiment, the ship intelligence preset distribution benchmark refers to the ideal reference benchmark vector of feature distribution that is preset for each modality type based on a large amount of historical valid data, ship industry standards, and expert experience.

[0100] In this embodiment, a weighted Euclidean distance algorithm is used to construct the fit calculation model. The actual distribution vector (basic statistical distribution vector or domain-related distribution vector) and a preset distribution benchmark are input. Weights are assigned according to the importance of each feature dimension. For example, in the AIS data modality, the weights for speed stability, heading continuity, track deviation, and sea area fit are 0.2 and 0.3 respectively. The fit value is automatically calculated. For instance, the basic statistical distribution vector of the AIS data modality is [14.5]. The preset distribution baseline is [15, 1.0, 90, 4.0, 0.5, 0.2, 0.95]. The basic statistical fit is calculated to be 0.91 using the fit calculation formula. At this time, if the basic statistical fit of the AIS data modality is 0.91 and the domain association fit is 0.93, the corresponding distribution deviation vector is [basic statistical deviation 0.09, domain association deviation 0.07], where the deviation value = 1 - fit.

[0101] The beneficial effects of the above technical solution are as follows: By extracting the unique feature dimensions of the shipbuilding domain, analyzing the basic statistical distribution vector and the domain-related distribution vector, and finally calculating the distribution deviation vector, a complete feature distribution vector of a single modality feature set is constructed. This ensures the domain-specificity and accuracy of feature distribution analysis and effectively avoids the problem of core information loss caused by general feature analysis. Compared with existing feature distribution analysis methods that lack domain adaptation, this solution can accurately capture the distribution patterns and deviations of ship modal data, providing a solid foundation for the distribution network to output reliable intelligence judgment indicators and fusion confidence, thereby improving the accuracy of ship intelligence judgment.

[0102] This invention provides a ship intelligence analysis method based on a large model, which constructs and trains a pre-set large model to obtain an enhanced ship model, including: Optimize the hierarchical structure of the preset large model; Multi-stage progressive training is performed on the hierarchical structure, and ship domain constraints are introduced during the training process to obtain an enhanced ship model.

[0103] Preferably, the hierarchical structure includes: a knowledge graph fusion layer, a multi-granularity temporal feature parsing layer, an adaptive weighted feature adaptation layer, a cross-modal feature fusion interaction layer, a confidence dynamic calibration layer, and an intelligence analysis and reasoning layer, wherein the rule constraint feature vector output by the knowledge graph fusion layer dynamically adjusts the granularity division threshold of the multi-granularity temporal feature parsing layer.

[0104] The granularity partitioning threshold is dynamically adjusted based on the rule-constrained feature vector, as follows: Let the rule-constrained feature vector output by the knowledge graph fusion layer be... The target granularity partitioning threshold for the multi-granularity temporal feature parsing layer is: ,and For fine-grained, medium-grained, and coarse-grained particles, respectively, the dynamic adjustment formula is: ,in, These are respectively the strength of safety-related rules, the strength of operational rules, the correlation of maritime risks, and the vessel type suitability coefficient; To set a pre-defined particle size threshold based on ship type, , For particle size adaptation coefficient, and ; For correction factors during the navigation phase; For sea area attribute adaptation coefficients; For rule priority weights, ; It is the minimum value; Let be the sensitivity coefficient for the rate of change of the rule constraints, and ; The time rate of change of the feature vector constrained by the rule.

[0105] Preferably, optimizing the hierarchical structure of the preset large model includes: The knowledge graph fusion layer is configured to perform vectorized embedding extraction of entity-relationships on the knowledge graph of the ship domain, to obtain the knowledge graph entity embedding vector group and entity relation constraint matrix, and to transform the intelligence judgment rules of the ship domain into rule constraint feature vectors. The multi-granularity temporal feature parsing layer is configured to extract temporal correlation feature vectors at different granularities from the multimodal data in the initial sample set and the supplementary sample set based on the granularity partitioning threshold, thereby obtaining a multi-granularity modal temporal feature set. The adaptive weighted feature adaptation layer is configured to encode the weights of the judgment indicators in the multi-granularity modal temporal feature set, the first indicator set, and the second indicator set, respectively. Modal adaptive weight coefficients are introduced, and a three-level fully connected mapping network and a sigmoid activation function are used to generate the weight allocation of each feature. At the same time, an asymmetric adjustment parameter is introduced into the high-dimensional modal feature branch to generate weighted features, which are then mapped to the feature space that matches the knowledge graph entity embedding vector group to obtain adaptive sample-indicator adaptation features. The cross-modal feature fusion interaction layer is configured as a composite feature processing module consisting of a Swin-Transformer feature mapping unit and a ResNet feature extraction unit. The adaptive sample-index matching features are normalized in dimension to construct a cross-modal fusion feature map. Furthermore, it is combined with knowledge graph entity embedding vector groups, entity relationship constraint matrices, and rule constraint feature vectors for deep interactive fusion to output a global fusion feature vector. The confidence dynamic calibration layer is configured to call the fused confidence sequence to perform temporal-spatial joint dynamic adjustment on the weights of each modal component in the global fused feature vector, and generate the calibrated feature vector. The intelligence analysis and reasoning layer is configured to use a deep fully connected reasoning network with a ship domain activation function. The calibrated feature vectors are subjected to classification reasoning and regression operations in the ship intelligence dimension, and the structured ship intelligence analysis results are output.

[0106] In this embodiment, ; in, It is a first-level fully connected mapping network; It is a second-level fully connected mapping network; It is a third-level fully connected mapping network; Low-dimensional modal features; High-dimensional modal features; The sigmoid activation function is used; α is the modality-adaptive weighting coefficient. This is element-wise multiplication; These are weighted features.

[0107] In this embodiment, ,in, The feature output of the Swin-Transformer feature mapping unit; The feature output of the ResNet feature extraction unit; This is a function for cross-modal attention interaction mechanisms; For feature concatenation and fusion operations; I represents adaptive sample-index matching features; This is a cross-modal fusion feature map.

[0108] In this embodiment, the multi-stage progressive training of the ship enhancement model is divided into three stages: Pre-training phase: A large-scale unlabeled dataset of 500,000 samples (initial and supplementary sets) is used. The training objective is feature reconstruction with a reconstruction error ≤ 0.05.

[0109] Semi-supervised training phase: 20,000 labeled data points and labeled intelligence analysis indicators are used for training, combined with a knowledge graph in the shipbuilding field. The training objective is to achieve a prediction accuracy of ≥85% for the analysis indicators.

[0110] Fine-tuning phase: 5,000 high-quality labeled core samples were used, including 2,000 complex navigation scenarios and 3,000 high-risk interaction scenarios. Constraints in the shipping field, such as speed limits for dangerous goods vessels, were introduced. The training objective was to achieve an accuracy rate of ≥92%.

[0111] In this embodiment, the entity cargo ship AIS equipment waterway area in the ship domain knowledge graph, and the entity relationship cargo ship-equipped-AIS equipment cargo ship-navigating-waterway area, are transformed into vectors [0.9,0.1,0.6,0.4], AIS equipment into [0.8,0.2,0.5,0.3], and waterway area into [0.7,0.3,0.4,0.5] through the TransE embedding algorithm; the entity relationship equipped into vectors [0.1,0.9,0.2,0.3], and navigating into vectors [0.2,0.1,0.9,0.4], and then an entity relationship constraint matrix is ​​constructed to quantify the strength of the relationship between entities.

[0112] In this embodiment, the rule constraint feature vector refers to the transformation of intelligence assessment rules in the shipping field into low-dimensional vector data that the model can recognize, ensuring that the model output conforms to the domain rules. For example, the intelligence assessment rule in the shipping field that dangerous goods vessels must not exceed 10 knots in the core area of ​​a port is transformed into a rule constraint feature vector [0.95 (rule importance), 0.8 (port core area fit), 0.7 (dangerous goods vessel fit), 0.9 (speed limit sensitivity)] through rule parsing and feature extraction; the rule that ship communication messages must contain MMSI code, time, and location information is transformed into a rule constraint feature vector [0.98 (rule importance), 0.92 (communication specification fit), 0.88 (information integrity sensitivity), 0.95 (compliance requirements)].

[0113] In this embodiment, the multi-granularity modal temporal feature set refers to the feature set formed by the multi-granularity temporal feature parsing layer after extracting temporal features from the multi-modal data in the initial sample set and supplementary sample set according to the fine-grained, medium-grained, and coarse-grained division thresholds. For example, the multi-granularity modal temporal feature set of AIS data modality includes fine-grained features: instantaneous airspeed values ​​at 0.1-second intervals [14.5, 14.6, 14.4, ...], and minute changes in heading [90, 90]. Medium-grained features: average speed at 1-second intervals [14.5, 14.4, 14.6, ...], rate of change of heading [0.1, 0.2, 0.1, ...], coarse-grained features: coordinates of track segments at 10-second intervals [(30.0°N, 120.0°E), (30.01°N, 120.02°E), ...], distance of position movement [0.2, 0.3, 0.25, ...].

[0114] In this embodiment, the granularity benchmark threshold of the multi-granularity temporal feature parsing layer The parameters are set according to ship type and particle size class, as shown in Table 3: Table 3 Particle size reference thresholds

[0115]

[0116] In this embodiment, feature encoding refers to the digital processing of features and indicators in the multi-granularity modal temporal feature set, the first indicator set, and the second indicator set, transforming unstructured or semi-structured feature information into structured vector data to ensure that the features can be processed by the fully connected network of the model. This is a preliminary step for feature adaptation. For example, the track anomaly degree of 0.12 (numerical feature) in the first indicator set is directly encoded as [0.12]; the navigation compliance degree of 0.95 is encoded as [0.95]; the high speed stability of AIS data in the multi-granularity temporal feature set (textual feature) is converted into [1] through label encoding, with high corresponding to 1, medium corresponding to 0.5, low corresponding to 0, etc.; good heading continuity is encoded as [1].

[0117] In this embodiment, a modal adaptive weight calculation model is constructed. The fusion confidence, data integrity, and relevance score of each modality are input, and the adaptive weight coefficient of each modality is calculated using a weighted summation algorithm. For example, in a certain navigation scenario, the fusion confidence of AIS data is 0.92 (high reliability), the fusion confidence of radar image data is 0.95 (highest reliability), and the fusion confidence of ship communication message data is 0.89 (relatively high reliability). Then, the modal adaptive weight coefficients Q(AIS) = 0.32, Q(radar) = 0.35, and Q(message) = 0.33.

[0118] In this embodiment, the three-layer fully connected mapping network has 10 neurons in the input layer (corresponding to 10 encoded features), 20 neurons in the first hidden layer, 10 neurons in the second hidden layer, and 10 neurons in the output layer (corresponding to the weight values ​​of the 10 features). The input encoded feature vector [0.12, 0.95, 0.88, ...] is processed by the linear transformation of the network and the sigmoid activation function to output the weight distribution of each feature [0.2, 0.8, 0.3, ...].

[0119] In this embodiment, the Swin-Transformer feature mapping unit has the following parameters: window size = 7, number of layers = 12, number of heads = 12, hidden layer dimension = 768, MLP ratio = 4, dropout rate = 0.1, and LayerNormε = 1e-5.

[0120] ResNet feature extraction unit: ResNet50 is used, number of convolutional layers = 49, convolutional kernel size is 7×7 (first layer) and 3×3 (other layers), max pooling window = 3×3, stride = 2, dropout rate = 0.1, final feature dimension = 2048.

[0121] Composite feature processing collaborative logic: The output features of Swin-Transformer and ResNet are concatenated according to weights, with the weights being Swin-Transformer (0.6) and ResNet (0.4). After concatenation, the dimensions are reduced to 512 dimensions through 1×1 convolution.

[0122] In this embodiment, the cross-modal fusion feature map refers to the graphical structure data constructed by normalizing the dimensions of different modal features according to their correlation relationships. Nodes represent feature vectors of each modality, and edges represent the correlation strength between different modal features. For example, the cross-modal fusion feature map contains three nodes: AIS feature node (32-dimensional vector), radar feature node (32-dimensional vector), and message feature node (32-dimensional vector). The edge weight between the AIS feature node and the radar feature node is 0.9 (calculated based on location information correlation), the edge weight between the AIS feature node and the message feature node is 0.85 (calculated based on navigation status correlation), and the edge weight between the radar feature node and the message feature node is 0.8 (calculated based on target identity correlation), forming a complete cross-modal fusion feature map.

[0123] Among them, dimensionality normalization: AIS data modality (32 dimensions): keeping the dimension unchanged, the feature values ​​are normalized, with a range of [0,1]; radar image modality (64 dimensions): using the PCA algorithm to reduce the dimension, resulting in a 32-dimensional modality; ship communication message modality (128 dimensions): using the PCA algorithm to reduce the dimension, resulting in a 32-dimensional modality.

[0124] Based on the entity relationship constraint matrix of the knowledge graph, the edge weights are calculated using cosine similarity: , Let be the edge weight between node s1 and node s2; Let be the feature vectors of nodes s1 and s2.

[0125] In this embodiment, the node feature vector of the cross-modal fusion feature map is [AIS: 0.85, radar: 0.91, message: 0.89], and the edge weight is [AIS-radar: 0.9, AIS-message: 0.85, radar-message: 0.8]. Combining the knowledge graph entity embedding vector group [0.8, 0.2, 0.5, 0.3], entity relationship constraint matrix [[1, 0, 0], [0, 1, 0], [0, 0, 1]], and rule constraint feature vector [0.95, 0.8, 0.7, 0.9], the attention weight of each node is calculated through the GAT algorithm, and feature aggregation and interaction are performed. Finally, the global fusion feature vector [0.88, 0.92, 0.89, 0.93] is output.

[0126] In this embodiment, the temporal-spatial joint adjustment formula is as follows: ,in, The weights of each modal component after calibration, The initial weights are set as follows: AIS component weight is 0.25, radar component weight is 0.25, message component weight is 0.25, and correlation component weight is 0.25. The values ​​of 0.4, 0.3, and 0.3 are pre-set; C is the mean of the fusion confidence series; T is the temporal correlation coefficient; and S is the spatial correlation coefficient. For example, =[0.25,0.25,0.25,0.25], C=0.92, T=0.93, S=0.96, then =[0.26,0.27,0.24,0.25].

[0127] In this embodiment, the activation function for compliance is: The output range is 0 to 1, and the gradient in the interval from 0.8 to 1 is ≤0.1, ensuring high compliance and stable results.

[0128] For the speed prediction activation function: The output range is [0,25].

[0129] For level activation functions: Output range The normalized map is mapped to (0,5), where 0 represents low risk, 1 represents relatively low risk, 2 represents medium risk, 3 represents relatively high risk, 4 represents high risk, and 5 represents extremely high risk.

[0130] In this embodiment, a deep fully connected inference network is constructed based on the TensorFlow framework. The number of hidden layers, the number of neurons, the ship domain activation function, the optimizer (Adam), and the learning rate (0.0001) are set. The deep fully connected inference network contains 3 hidden layers (each with 64, 32, and 16 neurons respectively) and 1 output layer. The input is a calibrated feature vector [0.88, 0.92, 0.89, 0.93]. After processing through linear transformation of each hidden layer and the ship domain activation function, the output classification results are: navigation status: normal and regression results: predicted speed: 14.3 knots, predicted heading: 93 degrees.

[0131] In this embodiment, a pre-set structured result template is used, specifying the fields included in the result, such as ship identification, speed, heading, and navigation status, as well as the field order and data format. The output of the intelligence analysis and reasoning layer is filled according to the template to form a structured ship intelligence analysis result. For example, the structured analysis result is: Ship identification: MMSI412345678; Ship type: cargo ship; Current speed: 14.5 knots; Current heading: 92 degrees; Current location: 30.05°N, 120.10°E; Navigation status: Normal; Predicted speed in the next 10 minutes: 14.3 knots; Predicted heading in the next 10 minutes: 93 degrees; Navigation compliance: 95%; Sea area adaptability: 92%; Interaction risk level: Low; Key reminder: Please maintain the current heading and pay attention to maintaining a safe distance from the fishing vessel 3 nautical miles ahead.

[0132] The beneficial effects of the above technical solution are as follows: By refining the hierarchical structure of the pre-set large model, clarifying the core functions, processing algorithms, and input-output relationships of each level, it achieves deep fusion of knowledge graphs in the shipbuilding field with multimodal data, accurate extraction of multi-granular temporal features, adaptive feature weight allocation, efficient interaction of cross-modal features, dynamic confidence calibration, and structured intelligence reasoning. Compared with the simple model structure optimization in existing technologies, the hierarchical design of this solution is more in line with the business logic of ship intelligence analysis. The functions of each level are complementary and synergistically efficient, significantly improving the model's ability to process shipbuilding data, feature fusion accuracy, and analysis and reasoning accuracy, ensuring that the output ship intelligence is complete, reliable, and practical.

[0133] This invention provides a platform for executing any of the aforementioned ship intelligence analysis methods based on large models.

[0134] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for ship intelligence analysis based on a large model, characterized in that, include: Step 1: Acquire historical multimodal ship intelligence data and perform submodal processing to obtain an initial sample set for each mode. At the same time, combined with ship maneuvering characteristic constraints and sea area attributes, perform fusion analysis on all initial sample sets according to temporal and spatial correlations to obtain a supplementary sample set. Step 2: Obtain the feature extraction method matching each modality type from the type-extraction lookup table, and extract features from the corresponding modality data according to the feature extraction method to construct a single modality feature set. Analyze the feature distribution vector of the single modality feature set and input it into the distribution network to output the intelligence judgment index and fusion confidence of the corresponding single modality feature set. The modality types include: AIS data modality, radar image modality, and ship communication message modality. Step 3: Perform dynamic analysis on the intelligence assessment indicators under all modal types with all initial sample sets and all supplementary sample sets according to the fusion confidence level, to obtain the first indicator set based on all initial sample sets and the second indicator set based on all supplementary sample sets; Step 4: Based on the knowledge graph of the ship domain, the initial sample set and the supplementary sample set, the first indicator set and the second indicator set, the preset large model is constructed and trained to obtain the ship augmented model. The current multimodal intelligence data after submodal processing is obtained and input into the ship augmented model for intelligence analysis and judgment, and the ship intelligence content is output. The AIS data includes the ship's name, MMSI code, speed, heading, and continuous latitude and longitude data. The radar image is a two-dimensional grayscale monitoring image that includes the ship's outline, ranging information, and relative motion trajectory; Ship communication messages include VHF communication text records between ships and port dispatch centers, as well as other ships; The analysis of the feature distribution vector of the single modality feature set includes: Extract the ship-specific feature dimensions of the modality type corresponding to the single modality feature set; Based on the unique feature dimension, the basic statistical distribution vector and the domain association distribution vector based on the single modality feature set are analyzed; The adaptation degree of the basic statistical distribution vector and the domain-related distribution vector with the preset distribution benchmark of ship intelligence of the corresponding modality type is calculated to obtain the distribution deviation vector. Among them, the basic statistical distribution vector refers to the vector data obtained by statistical analysis methods on the unique feature dimension of a single modal feature set in the shipbuilding field. Domain-related distribution vectors refer to vector data formed by analyzing the degree of correlation between a single modal feature set and other core elements of the shipping domain, including ship type, navigation stage, sea area attributes, and maritime rules, on the dimension of features unique to the shipping domain. The feature distribution vector is composed of a combination of a basic statistical distribution vector, a domain-related distribution vector, and a distribution deviation vector. The distribution network is a two-branch mapping neural network. The input is the feature distribution vector, and the output branch 1 is the intelligence analysis index and branch 2 is the fusion confidence score. Intelligence analysis indicators include track anomaly, navigation compliance, sea area adaptability, vessel identification accuracy, and interaction risk. Among them, the fusion confidence score is used as the weight to weight the indicators of the same core dimension, and unreliable indicators with a fusion confidence score lower than a preset threshold are removed. The integrated effective indicators are associated and stored according to the classification of the initial sample set to form the first indicator set. The construction logic of the second indicator set is completely consistent with that of the first indicator set, that is, the effective indicators corresponding to the supplementary sample set are integrated into the second indicator set. This involves constructing and training a pre-defined large model to obtain an enhanced ship model, including: The hierarchical structure of the preset large model is optimized, wherein the hierarchical structure includes: a knowledge graph fusion layer, a multi-granularity temporal feature parsing layer, an adaptive weighted feature adaptation layer, a cross-modal feature fusion interaction layer, a confidence dynamic calibration layer, and an intelligence analysis and reasoning layer, wherein the rule constraint feature vector output by the knowledge graph fusion layer dynamically adjusts the granularity division threshold of the multi-granularity temporal feature parsing layer; Multi-stage progressive training is performed on the hierarchical structure, and ship domain constraints are introduced during the training process to obtain an enhanced ship model. Optimizing the hierarchical structure of the preset large model includes: The knowledge graph fusion layer is configured to perform vectorized embedding extraction of entity-relationships on the knowledge graph of the ship domain, to obtain the knowledge graph entity embedding vector group and entity relation constraint matrix, and to transform the intelligence judgment rules of the ship domain into rule constraint feature vectors. The multi-granularity temporal feature parsing layer is configured to extract temporal correlation feature vectors at different granularities from the multimodal data in the initial sample set and the supplementary sample set based on the granularity partitioning threshold, thereby obtaining a multi-granularity modal temporal feature set. The adaptive weighted feature adaptation layer is configured to encode the weights of the judgment indicators in the multi-granularity modal temporal feature set, the first indicator set, and the second indicator set, respectively. Modal adaptive weight coefficients are introduced, and a three-level fully connected mapping network and a sigmoid activation function are used to generate the weight allocation of each feature. At the same time, an asymmetric adjustment parameter is introduced into the high-dimensional modal feature branch to generate weighted features, which are then mapped to the feature space that matches the knowledge graph entity embedding vector group to obtain adaptive sample-indicator adaptation features. The cross-modal feature fusion interaction layer is configured as a composite feature processing module consisting of a Swin-Transformer feature mapping unit and a ResNet feature extraction unit. The adaptive sample-index matching features are normalized in dimension to construct a cross-modal fusion feature map. Furthermore, it is combined with knowledge graph entity embedding vector groups, entity relationship constraint matrices, and rule constraint feature vectors for deep interactive fusion to output a global fusion feature vector. The confidence dynamic calibration layer is configured to call the fused confidence sequence to perform temporal-spatial joint dynamic adjustment on the weights of each modal component in the global fused feature vector, and generate the calibrated feature vector. The intelligence analysis and reasoning layer is configured to use a deep fully connected reasoning network with a ship domain activation function to perform classification reasoning and regression operations on the calibrated feature vectors in the ship intelligence dimension, and output structured ship intelligence analysis results. The modality adaptive weighting coefficient is a weighting coefficient calculated by weighting the fusion confidence of each modality, data integrity, and relevance to the judgment target. The ship domain activation function includes: Activation function for compliance: The output range is 0 to 1, and the gradient in the interval from 0.8 to 1 is ≤0.1, ensuring the stability of the results with high compliance. For the speed prediction activation function: The output range is [0,25]. For level activation functions: Output range The normalized map is mapped to (0,5), where 0 represents low risk, 1 represents relatively low risk, 2 represents medium risk, 3 represents relatively high risk, 4 represents high risk, and 5 represents extremely high risk.

2. The ship intelligence analysis method based on a large model according to claim 1, characterized in that, The supplementary sample set includes: The temporal and spatial labels of each sample in the initial sample set of the same modality type are extracted respectively. Based on the temporal coherence of each temporal label and the characteristics of the ship's navigation stage, the historical time quantity is dynamically determined and historical decoding is performed to obtain the historical potential variable set. Based on the spatial coherence of each spatial label and the constraints of the sea area functional attributes, the future time quantity is dynamically determined and future decoding is performed to obtain the future potential variable set. The temporal labels include: sample collection time, identification of continuous segments of ship track time sequence, and temporal correlation degree of track status. The spatial labels include: ship latitude and longitude, sea area functional attribute level, and spatial neighborhood distance between adjacent ships. A temporal baseline is constructed for each initial sample set, and a first supplementary location point based on the historical latent variable set is determined based on the temporal baseline. Simultaneously, a spatial baseline is constructed for each initial sample set, and a second supplementary location point based on the spatial baseline is determined based on the future latent variable set. Retrieve the spatial labels immediately adjacent to each supplementary location point and analyze the spatial characteristics of the corresponding supplementary location point. At the same time, retrieve the temporal labels immediately adjacent to each supplementary location point and analyze the temporal characteristics of the corresponding supplementary location point. The spatial features are subjected to temporal analysis and the temporal features are subjected to spatial analysis to obtain a single supplementary sample of the corresponding supplementary location point; Extract the temporal-spatial feature vectors of all single supplementary samples under each initial sample set, and perform residual analysis on the temporal-spatial feature vectors of all initial sample sets to obtain retained supplementary samples. Supplementary sample sets are then constructed based on all retained supplementary samples.

3. The ship intelligence analysis method based on a large model according to claim 2, characterized in that, A single supplementary sample of the corresponding supplementary location points is obtained, including: Based on the preset time baseline of ship intelligence, historical time series fragments matching the same type of ship in the same sea area and the corresponding spatial features are retrieved. The spatial features and the historical time series fragments are mapped and analyzed to determine the temporal dimension information corresponding to the spatial features and obtain the temporal spatial features. Based on the spatial baseline preset by the ship intelligence, the set of neighboring spatial attributes that are compatible with the corresponding temporal features within the same time interval is retrieved. The temporal features are matched and analyzed with the set of neighboring spatial attributes to determine the spatial dimension information corresponding to the temporal features and obtain the spatialized temporal features. The spatial features, temporal features, temporally sequenced spatial features, and spatially temporally sequenced features of the same supplementary location point are fused across dimensions, and the content that meets the residual threshold is selected by combining the weight of ship interaction relationship, and used as a single supplementary sample for the corresponding supplementary location point.

4. The ship intelligence analysis method based on a large model according to claim 2, characterized in that, Residual analysis was performed on the temporal-spatial feature vectors of all initial sample sets to obtain retained supplementary samples, including: Extract the temporal-spatial feature vectors of each initial sample set to determine the global and local variables of the residual analysis; Based on the global variables, the initial global residual between the temporal-spatial feature vector and the spatiotemporal reference feature vector is calculated. The initial global residual is then summed with the dynamic disturbance zone value in the ship domain to obtain the adjusted global residual sequence. Based on the local variables, local scene adaptation modal decomposition is performed on the temporal-spatial feature vector to obtain local modal components; Calculate the initial local residuals of each local modal component and the navigation time-space feature vector in the local variables, and calibrate the initial local residuals by combining the contextual correlation features of the local modal components to obtain the calibrated local residual sequence; The fusion weights are determined based on the scene matching degree between the global and local variables, and the fusion residuals of the global residual sequence and the local residual sequence are calculated according to the fusion weights. The sum of squared residuals of all fusion residuals is then calculated. Samples with residual sum of squares lower than the ship scenario adaptive threshold are selected as retained supplementary samples.

5. A ship intelligence analysis system based on a large model, characterized in that, Used to perform the ship intelligence analysis method based on a large model as described in any one of claims 1-4.

Images