Coal and ore ship identification and monitoring method based on AIS and multi-source data fusion

Abstract

This application belongs to the field of intelligent shipping technology, specifically relating to a method for identifying and monitoring coal and ore vessels based on AIS and multi-source data fusion. The method includes extracting and fusing feature vectors from multiple heterogeneous data sources containing AIS data of the target vessel to form a feature vector containing static attribute features, dynamic behavior features, and contextual semantic features; inputting the feature vector into a progressive fusion model for processing; firstly, inputting the static attribute features into a first classification model to obtain a prior probability distribution of the vessel type; then, inputting the dynamic behavior features into a second classification model to obtain a likelihood probability distribution based on the dynamic behavior; finally, fusing the prior probability distribution, the likelihood probability distribution, and the contextual likelihood obtained based on the contextual semantic features to obtain the final posterior probability distribution; and outputting the identification result that the cargo transported by the target vessel is coal or ore based on the posterior probability distribution.

Description

Technical Field

[0001] This application belongs to the field of intelligent shipping technology, specifically relating to a method for identifying and monitoring coal and ore ships based on AIS and multi-source data fusion. Background Technology

[0002] Accurate identification and dynamic monitoring of vessels transporting bulk commodities such as coal and ore are of great value in intelligent shipping, supply chain management, and the protection of national strategic resources. The Automatic Identification System (AIS), as the primary source of vessel activity data, provides the data foundation for achieving this goal. However, while the AIS system was designed with navigational safety at its core, and its standard data messages include dynamic information such as position, speed, and draft, it conspicuously lacks the crucial field of "cargo type." This inherent "information gap" constitutes the core technical bottleneck for automatically identifying vessels transporting specific cargo types from massive amounts of AIS data.

[0003] To overcome this bottleneck, the industry has proposed various technical solutions. Initially, methods relied on static information (such as registered vessel types) in commercial vessel databases for rule matching. However, these methods, due to severe data lag and coarse granularity in categories like "bulk carrier," failed to meet the demands for real-time and accurate identification. Subsequently, research shifted to utilizing the dynamic data from AIS itself. One approach employed a simple threshold method, such as judging solely by whether the port of call was a specialized bulk carrier port. While this method offered real-time performance, its shortcomings were obvious: the feature utilization was too singular and one-sided, unable to address the diverse cargo types transported by general bulk carriers, resulting in a very high misclassification rate. More advanced solutions attempted to introduce machine learning models, but their feature engineering was largely limited to macroscopic geometric features such as trajectory and speed, failing to delve into the key information contained within the data that directly reflects the vessel's operational status, leading to limited feature discriminative power. Furthermore, the decision-making process of such data-driven models is usually a "black box," with opaque inference logic, difficulty in tracing the impact of evidence from different sources on the final conclusion, and a lack of effective mechanisms for incorporating domain knowledge for constrained reasoning. As a result, the credibility and reliability of their decision results in complex scenarios are difficult to guarantee.

[0004] Therefore, existing technologies fail to effectively integrate AIS with data from other sources to compensate for information gaps, lack the ability to deeply mine key features, and lack a transparent and interpretable decision-making framework. These shortcomings systematically limit the accuracy and reliability of identification and surveillance. Summary of the Invention

[0005] This application proposes a coal ore vessel identification and monitoring method based on AIS and multi-source data fusion. It aims to systematically solve the defects in the existing technology, such as information loss and insufficient inference basis due to single data source, limited identification accuracy due to superficial feature utilization, and poor model credibility and lack of dynamic monitoring capability due to opaque decision-making mechanism.

[0006] The first aspect of this application provides a method for identifying and monitoring coal ore carriers based on AIS and multi-source data fusion, including: From multiple heterogeneous data sources containing AIS data of the target vessel, feature vectors are extracted and fused to form a feature vector containing static attribute features, dynamic behavior features and contextual semantic features. The dynamic behavior features are at least based on the time-series changes in the vessel's draft, and the contextual semantic features include at least features related to port professionalism and route matching. The feature vectors are input into a progressive fusion model for processing. First, the static attribute features are input into a first classification model to obtain the prior probability distribution of the ship type, so as to establish the classification benchmark for the identification task. Then, the dynamic behavior features are input into a second classification model to obtain the likelihood probability distribution based on dynamic behavior, so as to correct the classification benchmark using real-time behavioral evidence. Finally, the prior probability distribution, the likelihood probability distribution, and the contextual likelihood obtained based on the contextual semantic features are fused to obtain the final posterior probability distribution, so as to form the final identification decision based on comprehensive heterogeneous evidence. Based on the posterior probability distribution, the output will identify whether the cargo being transported by the target vessel is coal or ore.

[0007] In some embodiments of this application, the plurality of heterogeneous data sources include the AIS data of the target vessel, a vessel static database, and an external knowledge base.

[0008] In some embodiments of this application, the method further includes: Based on the time-series changes in the ship's draft, the cargo capacity of the target ship is estimated.

[0009] In some embodiments of this application, the dynamic behavioral characteristics are obtained at least based on the temporal changes in the ship's draft, including: A penalized likelihood change point detection model was used to segment the draft depth sequence to identify loading or unloading intervals. The penalized likelihood change point detection is achieved by solving a minimization problem, the objective function of which is the sum of the fitting cost and the penalty term for model complexity. Specifically, for a given sequence of segmentation points τ, the fitting cost is... The objective function is the sum of squared residuals of the draft depth data within each data segment relative to the linear regression model, expressed as: , Among them, fitting cost The expression is: Where K is the number of split points, the index is the data segment index, β is the complexity penalty parameter, and the draft data in the data segment is fitted with a linear model. Based on the identified loading and / or unloading intervals, the total change in draft within the interval is calculated as a dynamic behavior feature. The total change in draft is obtained by calculating the difference between the draft value at the end of the interval and the draft value at the beginning of the interval.

[0010] In some embodiments of this application, the dynamic behavior characteristics further include speed-draft joint distribution characteristics; The combined speed-draft distribution characteristics were obtained through the following methods: Normalize the draft depth sequence d(t) and the speed sequence v(t) respectively to obtain the normalized draft depth sequence. and normalized speed sequence ; calculate and 2x2 covariance matrix The eigenvalues ​​and eigenvector directions of the matrix are extracted as the combined speed-draft distribution features.

[0011] In some embodiments of this application, both the first classification model and the second classification model are gradient boosting decision tree models.

[0012] In some embodiments of this application, the contextual semantic features include at least features related to port specialization and route matching, including: The contextual semantic features are determined based on port specialization scores and route matching scores; wherein, The port specialization score is a quantitative value of port specialization for a specific cargo type, calculated using a logistic regression model. ,in, This represents the port's expertise score for cargo type c, where σ is the sigmoid function. The weight vector is obtained by training for cargo type c. For port feature vectors; The route matching score is a quantitative value of the similarity between the ship's trajectory and a typical route, which is obtained by calculating the trajectory similarity: , in, This represents the route matching score, where γ is the scaling parameter. This is the dynamic time-normalized distance between the ship's trajectory and the typical route.

[0013] In some embodiments of this application, the static attribute feature vector includes the following features: ship type code, gross tonnage, deadweight tonnage, ship length, ship width, approximate waterline area (which is the product of ship length and ship width), and a binary indicator variable representing whether a self-unloading system is equipped.

[0014] In some embodiments of this application, estimating the cargo capacity of the target vessel based on the time-series changes in the vessel's draft includes: The cargo capacity of the target vessel is estimated using the trapezoidal numerical integration method, wherein the estimation formula is: , in, This is an estimated value for the cargo weight. For the density of water, It is the acceleration due to gravity; This is obtained by interpolation from the ship's hull form data table based on the draft. The square factor at time step, The time interval between adjacent AIS messages; for The approximate displacement volume at time L, where L is the ship's length and B is the ship's beam. for The draft at any given time, This refers to the draft depth under light load.

[0015] In some embodiments of this application, the step of fusing the prior probability distribution, the likelihood probability distribution, and the contextual likelihood obtained based on the contextual semantic features to obtain the final posterior probability distribution includes: A hierarchical Bayesian model is used to fuse the prior probability distribution, the likelihood probability distribution, and the contextual likelihood into a multi-source evidence fusion to obtain the final posterior probability distribution. The fusion formula for the hierarchical Bayesian model is as follows: , in, This indicates that a ship belongs to a category given all the characteristics. The posterior probability; This represents the static attribute features obtained from the first classification model. The prior probability; This represents a given category obtained by the second classification model. Dynamic behavioral characteristics appear under certain conditions The likelihood probability; Represents a given category Contextual semantic features appear under certain conditions The likelihood probability; This indicates the scope of all candidate cargo categories, including coal, ore, and others. The summation operation.

[0016] In summary, the coal ore carrier identification and monitoring method based on AIS and multi-source data fusion proposed in the embodiments of this application firstly extracts and fuses static, dynamic, and contextual semantic features from multi-source heterogeneous data such as AIS data, ship static databases, and external knowledge bases to construct a feature vector. This systematically integrates multi-dimensional information such as the ship's inherent attributes, real-time behavior, and navigation environment semantics, fundamentally compensating for the lack of cargo information in AIS data and solving the deficiency of insufficient inference basis due to a single data source. Furthermore, it refines the method by deeply mining the temporal changes in the ship's draft and employing a penalized likelihood change point detection model. The refined modeling, by focusing feature extraction on key physical quantities that directly characterize the cargo loading status, achieves a leap from macroscopic trajectory description to microscopic loading and unloading behavior perception, thus overcoming the problem of limited recognition accuracy caused by superficial feature utilization. Finally, by designing a three-level progressive probability fusion model of "prior-likelihood-posterior" and integrating a cargo volume estimation module, this architecture constructs a transparent and interpretable Bayesian inference framework, and expands the recognition output from static labels to monitoring signals covering dynamic processes and quantitative indicators. This eliminates the model unreliability problem caused by the "black box" decision-making mechanism, and achieves a leap from static recognition to dynamic process monitoring capabilities. Attached Figure Description

[0017] The features and advantages of this application will become clearer with reference to the accompanying drawings, which are illustrative and should not be construed as limiting the application in any way. In the drawings: Figure 1 This is the overall algorithm framework of a coal ore ship identification and monitoring method based on AIS and multi-source data fusion, as shown in some embodiments of this application. Figure 2 This is a flowchart illustrating a coal ore vessel identification and monitoring method based on AIS and multi-source data fusion, according to some embodiments of this application. Figure 3 This is a flowchart of a quantitative modeling and feature engineering algorithm for a multi-source feature system, as shown in some embodiments of this application. Figure 4 This is a flowchart of a three-level progressive probability fusion decision algorithm according to some embodiments of this application; Figure 5 It is a multi-source feature fingerprint map shown according to some embodiments of this application; Figure 6This is a schematic diagram of ship AIS timing data and loading event detection results according to some embodiments of this application; Figure 7 This is a schematic diagram illustrating the combined speed-draft distribution characteristics according to some embodiments of this application; Figure 8 This is a schematic diagram illustrating load inference and DWT verification based on some embodiments of this application; Figure 9 This is a schematic diagram of a three-level progressive probability fusion decision-making process according to some embodiments of this application; Figure 10 This is a schematic diagram of an identification report according to some embodiments of this application. Detailed Implementation

[0018] In the following detailed description, numerous specific details of this application are illustrated by example to provide a thorough understanding of the relevant disclosure. However, it will be apparent to those skilled in the art that this application can be practiced without these details. It should be understood that the terms “system,” “apparatus,” “unit,” and / or “module” used in this application are one way of distinguishing different parts, elements, sections, or components at different levels in a sequential arrangement. However, these terms may be replaced with other expressions if other expressions can achieve the same purpose.

[0019] It should be understood that when a device, unit, or module is referred to as being "on," "connected to," or "coupled to" another device, unit, or module, it may be directly connected to or coupled to, or communicate with, other devices, units, or modules, or there may be intermediate devices, units, or modules present, unless the context explicitly indicates otherwise. For example, the term "and / or" as used herein includes any one and all combinations of one or more of the relevant listed items.

[0020] The terminology used in this application is for the purpose of describing specific embodiments only and is not intended to limit the scope of this application. As shown in the specification and claims of this application, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate that explicitly identified features, integrals, steps, operations, elements, and / or components are included, and such expressions do not constitute an exclusive list, and other features, integrals, steps, operations, elements, and / or components may also be included.

[0021] Referring to the following description and accompanying drawings, these and other features and characteristics, operating methods, functions of related structural elements, combinations of parts, and economics of manufacture of this application can be better understood, wherein the description and drawings form part of the specification. However, it is clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. It is understood that the drawings are not drawn to scale.

[0022] Various structural diagrams are used in this application to illustrate various variations of the embodiments according to this application. It should be understood that the preceding or following structures are not intended to limit this application. The scope of protection of this application is determined by the claims.

[0023] As described in the background section, existing technologies for automatically identifying cargo carriers from massive amounts of AIS data suffer from systemic defects, including a failure to effectively coordinate multi-source data to compensate for information gaps, a lack of ability to deeply mine key features, and a lack of a transparent and interpretable decision-making framework.

[0024] To address this technical deficiency, this application proposes a coal ore vessel identification and monitoring method based on AIS and multi-source data fusion. The aim is to systematically solve the core problems of existing technologies, such as single data source, superficial feature utilization, and opaque decision-making mechanism, by constructing a hierarchical feature extraction and probability fusion framework.

[0025] The core design concept of this application lies in constructing a three-tiered probabilistic reasoning system that progresses through "data-features-decision." The algorithmic design framework of this system is as follows: Figure 1 As shown. Specifically, the system takes multiple heterogeneous data sources, including the target vessel's AIS data, as input, and uses a feature extraction engine (corresponding to...) Figure 1 In the S1 stage, a multi-level feature vector containing static attributes, dynamic behaviors, and contextual semantics is quantified and constructed to compensate for the lack of information in AIS and form a comprehensive description. Then, the feature vector is input into the core three-level progressive fusion model (corresponding to...). Figure 1 In the S2 stage, the model first generates prior probabilities for the recognition task based on static features, then uses dynamic behavioral features to correct their likelihood, and finally integrates contextual semantic evidence to obtain the posterior probability distribution through Bayes' theorem, thus achieving an interpretable final decision. Furthermore, the system integrates a cargo volume estimation module in parallel to achieve dynamic monitoring of the transportation process. Through this architecture, this method simultaneously achieves high-precision recognition, transparent decision-making process, and full-cycle dynamic monitoring capabilities.

[0026] Figure 2 This is a flowchart illustrating a coal ore vessel identification and monitoring method based on AIS and multi-source data fusion, according to some embodiments of this application. Figure 2As shown, the method mainly includes two core processes: first, extracting and fusing features from multi-source heterogeneous data to construct a multi-level feature vector; second, processing the feature vector through a three-level progressive fusion model to complete the recognition decision. The following is combined with... Figure 3 (Quantitative modeling and feature engineering of multi-source feature systems) and Figure 4 (Three-level progressive probability fusion decision-making) provides a detailed explanation of each step.

[0027] like Figure 2 As shown, the method specifically includes the following steps: S210, extract and fuse from multiple heterogeneous data sources containing AIS data of the target vessel to form a feature vector containing static attribute features, dynamic behavior features and contextual semantic features, wherein the dynamic behavior features are at least based on the time-series changes in the vessel's draft, and the contextual semantic features include at least features related to port professionalism and route matching.

[0028] like Figure 3 As shown, this step extracts multi-level features from multiple heterogeneous data sources based on AIS data containing the target vessel. Specifically: First, it is necessary to access and integrate data from multiple sources: AIS data stream interface: It receives AIS messages from shore-based base stations or satellites in real time or near real time, and parses the dynamic information (MMSI, timestamp, latitude and longitude, speed to ground SOG, heading to ground COG, heading to ground HDG, draft, etc.) and static information (ship name, call sign, IMO number, ship type, length, beam, etc.).

[0029] Ship static database: Connect to commercial or official ship databases (such as HIS Markit, Lloyd's Register database) and obtain detailed design parameters of the ship through MMSI or IMO numbers, including gross tonnage (GT), deadweight tonnage (DWT), design draft, and whether it is equipped with a self-unloading system.

[0030] External knowledge bases, including: Port Professional Knowledge Base: Records berth information, historical throughput (categorized by cargo type), maximum water depth of the terminal, and whether it is a specialized coal / ore terminal for major ports worldwide.

[0031] Typical route database: Typical route path sequence for storing bulk commodities (such as from Newcastle, Australia to coastal ports in China, and from Tubarão, Brazil to coastal ports in China).

[0032] Historical identification record database: Used to store and update the historical identification results and confidence levels for each vessel.

[0033] Then, multi-level feature extraction is performed. In the feature extraction stage, this application fuses multi-source information from AIS data, static databases, and external knowledge bases to construct feature vectors containing static, dynamic, and contextual semantic features. The distribution and association patterns of these features can be analyzed using multi-source feature fingerprints (…). Figure 5 The diagram presents the overall structure of the feature system in a visual form, providing a foundation for subsequent classification.

[0034] To extract effective information distinguishing cargo types from raw data, this application constructs a three-layer heterogeneous feature system and quantitatively describes it using a mathematical model. Specifically, it includes: Static attribute feature vector: Static features originate from the inherent properties of the ship and constitute the basic prior space for identification. The feature vector is defined as: in: : This is the ship type coding function. The input is the IMO ship type code, and the output is a high-dimensional vector that has been embedded or one-hot encoded to represent the functional design category of the ship.

[0035] These are gross tonnage, deadweight tonnage, length, and beam, respectively, and are continuous variables.

[0036] It approximates the waterline surface area and is linearly related to the ship's cargo carrying capacity.

[0037] This is a binary indicator variable, marking whether the ship is equipped with a self-unloading system (common in some coal transport ships).

[0038] This feature vector is used to construct the prior probability distribution for a classification task.

[0039] Dynamic behavioral feature vector: Dynamic features are key to mining AIS time-series information, especially draft data, in this application. For a time series of a cruise... Define the observation data: Draft sequence: Ground speed sequence: Position sequence: (a) Loading event detection model based on optimized segmentation To accurately extract the loading / unloading phase, this application employs a penalized likelihood change point detection model. Assuming that the draft follows a noisy linear increase during the stable loading phase, we divide the sequence into segments, with the data within each segment following a linear regression model.

[0040] The model is defined as follows: Let the first... The data index range for each segment is: Its internal data satisfies in For a given sequence of split points Its negative log-likelihood (i.e., cost) is: in For the first Number of data points These are the least-squares fitting parameters for this segment. To control model complexity, a penalty term is introduced. The optimal segmentation is obtained by minimizing the following objective function: in This represents the complexity penalty coefficient. This optimization problem can be solved efficiently using the PELT algorithm. The identified loading interval... Used for calculation: Changes in load draft: Average load rate: Load stability index: Standard deviation of the fitted residuals within the loading range .

[0041] By using the penalized likelihood change point detection model, this application can accurately segment the draft depth sequence and identify the loading or unloading interval. Figure 6 This is a visualization of the process, which clearly shows the trend of draft change over time and the detected change points, helping to intuitively understand the dynamic characteristics of the loading event.

[0042] (b) Extraction of behavioral pattern features during the voyage phase Based on draft, speed, and location, the voyage was divided into stages such as "port berthing," "coastal maneuvering," and "ocean voyage," and stage-specific statistics were extracted: Full-load steady-state characteristics: defining the full-load factor Statistics on high load ratio in .

[0043] Speed-draft joint distribution matrix: Calculate normalized draft With normalized speed covariance matrix And extract its feature values And the direction of the eigenvectors, used to quantify behavioral patterns. Ore carriers, due to their high cargo density and high stability requirements when fully loaded, often exhibit… and A stronger negative correlation.

[0044] The joint distribution characteristics of speed and draft are quantified by extracting eigenvalues ​​from the covariance matrix to quantify behavioral patterns. For example, the speed and draft of an ore carrier at full load are negatively correlated. A schematic diagram of this distribution is shown below. Figure 7 As shown, it displays the statistical relationship between normalized speed and draft in the form of scatter plots or contour lines, which helps to distinguish the behavior of ships carrying different types of cargo.

[0045] Finally, the dynamic behavior feature vector is: in The mean and standard deviation of the speed.

[0046] Contextual semantic feature vector: Contextual features place ships within a macro-logistics network, providing geographical and historical semantic information. Definition: The port professionalism score was calculated using a logistic regression model. For ports... and cargo type ,have .in For the sigmoid function, For the port feature vector, This is the learned weight vector.

[0047] Route matching degree calculation of ship trajectory Similarity to a preset typical route (such as the "Western Australia-China" iron ore route). Distance based on dynamic time warping is used. but in This is the scale parameter.

[0048] Historical consistency The probability of a ship's history being classified into a certain type is calculated using the exponential smoothing method. .

[0049] deadweight tonnage calibration deviation Cargo volume based on Level 2 estimation The standardized deviation from the ship's registered DWT is used to detect anomalies or verify classification.

[0050] In a preferred embodiment of this application, the specific process of multi-dimensional feature extraction is as follows: First, perform static attribute feature vector ( )extract: For the target vessel, the system extracts the original fields from its static data records and performs the following calculations: 1) Feature encoding of the IMO ship type code T. One implementation is to use one-hot encoding, mapping subcategories under bulk carriers (such as Bulk Carrier, OBO Carrier, etc.) to sparse binary vectors. Another, more preferred implementation is to use an embedding layer, learning a low-dimensional dense vector through training. .

[0051] 2) Calculate the approximate value of the ship's waterline area. .

[0052] 3) Based on the ship design data or equipment list Set the indicator variable for the self-unloading system; if equipped, it should be 1, otherwise 0.

[0053] 4) Concatenate all the above feature values ​​in a predetermined order to form a static feature vector, and then standardize it (such as Z-score standardization) to eliminate the influence of dimensions.

[0054] Then perform dynamic behavior feature vector ( )extract: This step processes AIS time-series data for a single voyage or a single time window.

[0055] 1) Data preprocessing and alignment: Sort the original AIS dynamic messages by timestamp and perform simple smoothing filtering (such as moving average) on key fields (such as draft) to suppress transient noise.

[0056] 2) Loading event detection (implementing a penalized likelihood change point model): Input: timestamp sequence and the corresponding draft depth sequence .

[0057] Implementation: The PELT algorithm is used. The cost function is set as the negative log-likelihood of the linear model fit, and the complexity penalty parameter is... This was determined through cross-validation.

[0058] Output: Identify all significant points of change, with the intervals of sustained and rapid draft increase marked as loading intervals. .

[0059] Calculation: Based on the loading range, calculate three core characteristics: total draft change. Average loading rate Stability index of the loading process (standard deviation of the fitted residuals) ).

[0060] 3) Extraction of behavioral pattern features during the voyage phase: Full-load steady-state characteristics: Traverse the entire time window and calculate the full-load factor at each moment. .statistics The proportion of time, to obtain .

[0061] Speed-draft combined distribution characteristics: For draft and speed Normalize them separately to obtain and .

[0062] calculate and covariance matrix .

[0063] Perform eigenvalue decomposition on Σ and extract the largest eigenvalue. and eigenvalue ratio These characteristics reflect the strength of the coupling between speed and draft changes; for example, ore carriers tend to maintain lower speeds when fully loaded, exhibiting a stronger negative correlation. Larger).

[0064] Speed ​​statistics: Directly calculate the mean of the speed series and standard deviation .

[0065] 4) Feature vector assembly: The features calculated above are assembled... , , , , , , , Features are combined to form a dynamic behavioral feature vector. And standardize it.

[0066] Then perform contextual semantic feature vector ( )extract 1) Port professionalism score ( ): Query the knowledge base to obtain the feature vectors of the voyage's origin and destination ports. .

[0067] For each candidate cargo type (coal, ore), a score is calculated using a pre-trained logistic regression model: The weight vector It was obtained through training on historical port-cargo correlation data.

[0068] 2) Route matching degree ( ): The sequence of the ship's trajectory positions during the voyage Compared with typical coal / ore routes in the knowledge base Compare them.

[0069] Using dynamic time warping ( The algorithm calculates the minimum cumulative distance between the trajectory and the flight path. .

[0070] Calculate the matching degree: ,in To adjust the parameters and control the rate at which the matching degree decays.

[0071] 3) Historical consistency ( ): Read the probability distribution of the type of the ship that was last identified from the historical database.

[0072] Update using exponential smoothing: .in As a smoothing factor, This is an indicator function.

[0073] 4) Cadre weight verification deviation ( ): The second-level model completes the cargo loading. After estimation, calculate its relative deviation from the ship's registered deadweight tonnage: .

[0074] 5) Combine the above five feature values ​​into a contextual semantic feature vector. .

[0075] S220, the feature vector is input into the progressive fusion model for processing. First, the static attribute features are input into the first classification model to obtain the prior probability distribution of the ship type, so as to establish the classification benchmark for the identification task. Then, the dynamic behavior features are input into the second classification model to obtain the likelihood probability distribution based on dynamic behavior, so as to correct the classification benchmark using real-time behavioral evidence. Finally, the prior probability distribution, the likelihood probability distribution, and the contextual likelihood obtained based on the contextual semantic features are fused to obtain the final posterior probability distribution, so as to form the final identification decision based on comprehensive heterogeneous evidence.

[0076] like Figure 4 As shown, this application first constructs a three-level progressive fusion recognition model, and based on this model, obtains the posterior probability distribution used for the final recognition decision. Specifically: Level 1: Static prior generation based on gradient boosting decision tree This layer uses the GBDT model to process static features. GBDT is a forward additive ensemble model whose prediction function is a weighted sum of multiple weak learners (decision trees): , in, For the first A decision tree, Its parameters (such as split node and leaf node values). The shrinkage rate is used to prevent overfitting. Each new tree is built to fit the residuals along the negative gradient direction of the current model. For multi-class classification tasks, multi-class log loss (cross-entropy) is commonly used as the loss function, along with a gradient boosting strategy. This stage outputs an initial class probability distribution: in It corresponds to the category The GBDT output value. This probability serves as prior knowledge for subsequent fusion.

[0077] Level 2: Likelihood estimation and transport volume estimation based on dynamic characteristics This layer is the core of the classification, using a deeper GBDT model. Processing dynamic features This model focuses on learning the conditional probability relationship between dynamic behavior patterns and cargo types, and its output corresponds to the logarithm of the likelihood function: Right now .

[0078] Simultaneously, this level integrates a cargo capacity estimation module. After determining that loading is complete, the weight of the cargo is estimated using the trapezoidal numerical integration method. in, For approximate drainage volume, The block coefficient is estimated based on draft and ship shape. This represents gravitational acceleration. The ratio of this estimate to the ship's DWT can be used as a feature feedback to the classification model.

[0079] The cargo capacity estimation employs the trapezoidal numerical integration method, validated in deadweight tonnage (DWT) to ensure the reliability of the results. A visualization of this inference and validation process is provided below. Figure 8 As shown, it includes the draft depth variation curve, the estimated cargo load calculation steps, and DWT deviation analysis, which intuitively demonstrates the dynamic monitoring capability.

[0080] Level 3: Probabilistic fusion decision-making based on Bayesian theory This level is the decision-making hub of this application, employing a hierarchical Bayesian model for multi-source evidence fusion. This application assumes that, given the true category, static, dynamic, and contextual features are independent of each other (an extension of the Naive Bayes assumption). This assumption is valid in engineering practice and greatly simplifies computation.

[0081] The joint posterior probability after fusion is: in: Provided by the first-level GBDT model.

[0082] Likelihood estimates provided by the second-level GBDT model.

[0083] It is the context likelihood, which can be computed by a standalone classifier (such as logistic regression) or a rule-based deterministic function. For example, it can be defined as: in Adjustable parameters The final decision is given by the maximum a posteriori probability criterion: The posterior probability This also serves as the confidence score output for this identification.

[0084] The three-level progressive fusion model achieves the final decision through prior generation, likelihood correction, and Bayesian fusion. A flowchart illustrating the entire decision-making process is shown below. Figure 9 As shown, it progressively demonstrates the probabilistic fusion path from static features to dynamic evidence, and then combines it with contextual semantics, thereby enhancing the interpretability and transparency of the model.

[0085] In a preferred embodiment of this application, the specific process of the three-level progressive probability fusion decision is as follows: First-level model training and prior generation: 1) Training data preparation: Collect a large amount of ship voyage data with confirmed cargo types. Each sample contains its static feature vector and real cargo type label (coal, ore, others).

[0086] 2) Model Training: Train a multi-class classification model using Gradient Boosting Decision Tree Library (XGBoost). Set an appropriate number of trees ( ), depth, learning rate ( Hyperparameters such as cross-entropy are optimized using multi-class cross-entropy as the loss function.

[0087] 3) Online inference: For new ships, extract... Its input is the trained model The model output corresponds to the raw scores of the three categories. These scores are converted into probability distributions using the Softmax function, serving as prior knowledge.

[0088] Second-level model training, likelihood estimation, and transport estimation: 1) Model training: Use the same labeled dataset as the first level, but the features are dynamic behavioral feature vectors. Train another, deeper model. .

[0089] 2) Likelihood estimation: During online inference, the target ship's... Input Model Similarly, the output score is converted into conditional probability using the Softmax function. That is, the likelihood of dynamic evidence.

[0090] 3) Parallel calculation of cargo capacity estimation: After loading is detected as complete (draft stabilizes at a high level), the draft-time series of that stable segment is obtained.

[0091] Estimate the block coefficient at different drafts based on the ship's lines diagram data or empirical formulas. .

[0092] Applying the trapezoidal numerical integration formula, based on and seawater density Calculate the cargo weight .

[0093] This estimate is used to calculate the characteristics. And can be output in the final result.

[0094] S3.3, Third-Level Bayesian Fusion and Decision Making 1) Context likelihood calculation: A lightweight method is used for calculation. For example, it can be defined .parameter It can be adjusted based on the validation set.

[0095] 2) Bayesian fusion: This integrates the prior information from the first-level output. Likelihood of the second-level output Contextual similarity Substitute into the Bayesian fusion formula to calculate the complete posterior probability. .

[0096] 3) Final decision and output: The category with the highest posterior probability is selected as the final identification result: .

[0097] The maximum posterior probability value is also used as the confidence level for this identification.

[0098] The system outputs structured results, including: vessel MMSI, identified cargo type (coal / ore / other), confidence level, current estimated loading status (e.g., "fully loaded"), and estimated cargo volume. Timestamps, etc.

[0099] Update the identification results (type and probability) for this vessel. In historical records.

[0100] The system ultimately outputs a structured identification report, including key information such as cargo type, confidence level, and cargo load. An example report is shown below. Figure 10 As shown, it demonstrates the output formats used in practical applications, such as ship MMSI, identification results, and monitoring indicators, making the results of the entire identification process clear and easy for users to understand and use.

[0101] S230, based on the posterior probability distribution, output the identification result that the cargo transported by the target vessel is coal or ore.

[0102] One embodiment of this application provides the entire process of identifying a Capesize bulk carrier en route from Port Hedland, Australia to Caofeidian Port, China.

[0103] 1) Obtain its static data: Vessel type is "Bulk Carrier", DWT is 205,000 tons, L=300 meters, B=50 meters, no self-unloading system. Construction. .

[0104] 2) Processing its voyage AIS data, the PELT algorithm revealed a continuous draft rise of approximately 36 hours in Port Hedland, and calculations were performed to obtain... rice, meters per hour. Reached during the voyage. 92%. Eigenvalue ratio of the speed-draft covariance matrix. High. Construction .

[0105] 3) Query Context: Port Hedland (origin) has a ore expertise score of 0.98, and Caofeidian Port (destination) has a ore expertise score of 0.90. Its trajectory matches the "Western Australia-China" iron ore route with a score of 0.85. This ship has been identified as an ore carrier in eight out of ten historical voyages. The value is 0.80. (Build) .

[0106] 4) The first-level model outputs prior probabilities: , , .

[0107] 5) The second-level model is based on Output likelihood: , At the same time, it was estimated that tons, calculation normal.

[0108] The third level calculates the context likelihood (assuming parameters). ),For example: After Bayesian fusion, the posterior probability is calculated: , .

[0109] Final output: The ship is identified as an ore carrier with a confidence level of 0.92, currently fully loaded, and an estimated cargo volume of 198,000 tons. Update the ship's historical records.

[0110] This application: 1) Achieved high-precision recognition and strong generalization ability: By employing the Gradient Boosting Decision Tree (GBDT) model to process the static and dynamic features of ships, it can automatically learn and capture the complex nonlinear interaction relationships between multi-source heterogeneous features. Combined with a multi-level probability fusion mechanism based on Bayes' theorem, it effectively integrates prior knowledge, real-time behavioral patterns, and contextual evidence, significantly improving the classification accuracy of coal and ore transport ships in complex and ever-changing maritime transport environments. It also exhibits good generalization performance for unknown samples or unseen behavioral patterns, overcoming the problems of low recognition rate and poor adaptability caused by traditional methods relying on single features or simple rules.

[0111] 2) It provides an interpretable and transparent decision-making process: The three-level progressive identification framework of this application is essentially an implementation of a probabilistic graphical model. The final classification decision can be explicitly decomposed into three contributions: the prior probability generated by static features, the likelihood probability estimated by dynamic behavioral features, and the evidence probability provided by contextual features. The weight and contribution of each level of evidence can be quantified and output as probability values, making the entire identification process highly interpretable and traceable. This greatly facilitates system administrators in reviewing results, assessing credibility, and diagnosing errors, enhancing the credibility and acceptability of the system in practical applications.

[0112] 3) Breaking through static type determination, this application achieves dynamic process perception and quantitative monitoring: This application innovatively conducts in-depth mining of AIS time-series data, especially draft data. Through a loading event detection model based on optimized segmentation (such as the PELT algorithm), it can accurately identify the loading and unloading operation status of ships. Furthermore, combined with ship physical parameters, it achieves real-time judgment of cargo loading status and approximate estimation of transport volume. This allows the method to leap from a single, static identification of ship type to dynamic perception and quantitative monitoring of the entire lifecycle of ship transportation activities, providing unprecedented fine-grained data support for logistics analysis, capacity scheduling, and trade insights.

[0113] 4) Flexible and scalable modular system architecture: The three-level recognition model (static filtering, dynamic classification, and context fusion) constructed in this application is functionally independent and coupled through a clear probabilistic interface. This modular design enables the system to have high flexibility, allowing independent upgrades, replacements, or optimizations of the model or algorithm at any stage (e.g., introducing a more advanced temporal neural network model in dynamic behavior analysis, or integrating a richer knowledge graph in context analysis) without reconstructing the entire system. This ensures that this technical solution can continuously absorb new algorithmic results and data sources, possessing long-term vitality and broad adaptability.

[0114] In summary, the coal ore carrier identification and monitoring method based on AIS and multi-source data fusion proposed in the embodiments of this application firstly extracts and fuses static, dynamic, and contextual semantic features from multi-source heterogeneous data such as AIS data, ship static databases, and external knowledge bases to construct a feature vector. This systematically integrates multi-dimensional information such as the ship's inherent attributes, real-time behavior, and navigation environment semantics, fundamentally compensating for the lack of cargo information in AIS data and solving the deficiency of insufficient inference basis due to a single data source. Furthermore, it refines the method by deeply mining the temporal changes in the ship's draft and employing a penalized likelihood change point detection model. The refined modeling, by focusing feature extraction on key physical quantities that directly characterize the cargo loading status, achieves a leap from macroscopic trajectory description to microscopic loading and unloading behavior perception, thus overcoming the problem of limited recognition accuracy caused by superficial feature utilization. Finally, by designing a three-level progressive probability fusion model of "prior-likelihood-posterior" and integrating a cargo volume estimation module, this architecture constructs a transparent and interpretable Bayesian inference framework, and expands the recognition output from static labels to monitoring signals covering dynamic processes and quantitative indicators. This eliminates the model unreliability problem caused by the "black box" decision-making mechanism, and achieves a leap from static recognition to dynamic process monitoring capabilities.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding descriptions in the foregoing device embodiments, and will not be repeated here.

[0116] Although the subject matter described herein is provided in the general context of execution on a computer system in conjunction with an operating system and applications, those skilled in the art will recognize that other implementations can also be executed in conjunction with other types of program modules. Generally, program modules include routines, programs, components, data structures, and other types of structures that perform specific tasks or implement specific abstract data types. Those skilled in the art will understand that the subject matter described herein can be practiced using other computer system configurations, including handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframes, etc., and can also be used in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules may reside on both local and remote memory storage devices.

[0117] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0118] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of this application and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of this application should be included within the protection scope of this application. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A coal ore ship identification monitoring method based on AIS and multi-source data fusion, characterized in that, include: From multiple heterogeneous data sources containing AIS data of the target vessel, feature vectors are extracted and fused to form a feature vector containing static attribute features, dynamic behavior features and contextual semantic features. The dynamic behavior features are at least based on the time-series changes in the vessel's draft, and the contextual semantic features include at least features related to port professionalism and route matching. The feature vectors are input into a progressive fusion model for processing. First, the static attribute features are input into a first classification model to obtain the prior probability distribution of the ship type, so as to establish the classification benchmark for the identification task. Then, the dynamic behavior features are input into a second classification model to obtain the likelihood probability distribution based on dynamic behavior, so as to correct the classification benchmark using real-time behavioral evidence. Finally, the prior probability distribution, the likelihood probability distribution, and the contextual likelihood obtained based on the contextual semantic features are fused to obtain the final posterior probability distribution, so as to form the final identification decision based on comprehensive heterogeneous evidence. Based on the posterior probability distribution, the output will identify whether the cargo being transported by the target vessel is coal or ore.

2. The method according to claim 1, characterized in that: The multiple heterogeneous data sources include the target vessel's AIS data, a vessel static database, and an external knowledge base.

3. The method of claim 1, wherein, The method further includes: Based on the time-series changes in the ship's draft, the cargo capacity of the target ship is estimated.

4. The method of claim 1 or 3, wherein, The dynamic behavioral characteristics are derived at least from the temporal changes in the ship's draft, including: A penalized likelihood change point detection model was used to segment the draft depth sequence to identify loading or unloading intervals. The penalized likelihood change point detection is achieved by solving a minimization problem, the objective function of which is the sum of the fitting cost and the penalty term for model complexity. Specifically, for a given sequence of segmentation points τ, the fitting cost is... The objective function is the sum of squared residuals of the draft depth data within each data segment relative to the linear regression model, expressed as: ， Among them, fitting cost The expression is: Where K is the number of split points, the index is the data segment index, β is the complexity penalty parameter, and the draft data in the data segment is fitted with a linear model. Based on the identified loading and / or unloading intervals, the total change in draft within the interval is calculated as a dynamic behavior feature. The total change in draft is obtained by calculating the difference between the draft value at the end of the interval and the draft value at the beginning of the interval.

5. The method according to claim 4, characterized in that: The dynamic behavioral characteristics also include the combined speed-draft distribution characteristics; The combined speed-draft distribution characteristics were obtained through the following methods: The draft sequence d(t) and the speed sequence v(t) are normalized respectively to obtain the normalized draft sequence and the normalized speed sequence ; calculate and 2x2 covariance matrix The eigenvalues ​​and eigenvector directions of the matrix are extracted as the combined speed-draft distribution features.

6. The method according to claim 1, characterized in that: Both the first classification model and the second classification model are gradient boosting decision tree models.

7. The method of claim 1, wherein, The contextual semantic features include at least the features related to port specialization and route matching, including: The contextual semantic features are determined based on port specialization scores and route matching scores; wherein, The port specialization score is a quantitative value of the degree of specialization of a port for a specific cargo type, which is calculated by a logistic regression model: wherein, denotes the specialization score of a port for cargo type c, σ is a sigmoid function, is a weight vector trained for cargo type c, is a port feature vector; The route matching score is a quantitative value of the similarity between the ship trajectory and the typical route, which is obtained by calculating the trajectory similarity: , wherein, denotes the route matching score, and γ is a scale parameter, is the dynamic time warping distance between the ship trajectory and the typical route.

8. The method according to claim 1, characterized in that: The static attribute feature vector includes the following features: ship type code, gross tonnage, deadweight tonnage, ship length, ship width, approximate waterline area (which is the product of ship length and ship width), and a binary indicator variable representing whether a self-unloading system is equipped.

9. The method of claim 3, wherein, The estimation of the target vessel's cargo capacity based on the time-series changes in the vessel's draft includes: The cargo capacity of the target vessel is estimated using the trapezoidal numerical integration method, wherein the estimation formula is: , in, This is an estimated value for the cargo weight. For the density of water, It is the acceleration due to gravity; This is obtained by interpolation from the ship's hull form data table based on the draft. The square factor at time step, The time interval between adjacent AIS messages; for The approximate displacement volume at time L, where L is the ship's length and B is the ship's beam. for The draft at any given time, This refers to the draft depth under light load.

10. The method of claim 1, wherein, The process of fusing the prior probability distribution, the likelihood probability distribution, and the contextual likelihood obtained based on the contextual semantic features to obtain the final posterior probability distribution includes: A hierarchical Bayesian model is used to fuse the prior probability distribution, the likelihood probability distribution, and the contextual likelihood into a multi-source evidence fusion to obtain the final posterior probability distribution. The fusion formula for the hierarchical Bayesian model is as follows: ， wherein, represents the posterior probability that the ship belongs to the class under given all characteristic conditions; representing a prior probability based on static attribute features derived by the first classification model; representing a likelihood probability of a given class being obtained by the second classification model dynamic behavior features under the condition of the given class representing a given class contextual semantic features under the condition that the likelihood probability; represents a summation operation over all candidate cargo classes including coal, ore, and others .

Images