A ship navigation risk alarm information intelligent identification method and system

Abstract

This invention relates to the field of port safety management technology, and more particularly to an intelligent identification method and system for ship navigation risk alarm information. In extracting entities and relationships, this invention iteratively updates the initial node feature matrix using a directional entity-aware graph attention network to obtain a final high-order node feature matrix. This final high-order node feature matrix is ​​then input into a classifier for entity recognition and relationship classification. During the iterative update phase, the output is a completely new final high-order node feature matrix. Each row of this matrix is ​​a deep feature rich in global contextual information and directional entity awareness, tailored for entity recognition and relationship classification. These deep features are directly input into the subsequent classifier to complete the final extraction task, thereby significantly improving the extraction accuracy and robustness for complex alarm statements.

Description

Technical Field

[0001] This invention relates to the field of port safety management technology, and in particular to a method and system for intelligent identification of ship navigation risk alarm information. Background Technology

[0002] The navigation environment for ships is becoming increasingly complex, facing various risks such as collisions, groundings, fires, equipment failures, and severe weather. Modern ships are equipped with advanced navigation systems, engine room monitoring systems, and weather instruments, which trigger numerous alarm messages when potential hazards are detected.

[0003] In the field of ship intelligent alarms, the use of knowledge graphs to achieve intelligent identification of alarm information is a relatively in-depth research branch of technology. When constructing a ship alarm information knowledge graph, most traditional entity and relation extraction methods adopt the approach of first identifying entities and then classifying entity pairs. However, traditional approaches are prone to error propagation, as small errors in the previous entity identification step can be directly transmitted to the relation classification stage. At the same time, traditional approaches view the relationships between entity pairs in isolation, ignoring the inherent syntactic structure and logical framework of the alarm statement itself. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent identification method and system for ship navigation risk alarm information, which solves the problems existing in the prior art.

[0005] This invention provides an intelligent identification method for ship navigation risk warning information, the method comprising the following steps:

[0006] S1: Collect multi-source alarm information;

[0007] S2: Perform data preprocessing on the multi-source alarm information;

[0008] S3: Perform data processing on the pre-processed multi-source alarm information to extract entities and relationships;

[0009] Specifically, S3 is:

[0010] S3.1: The preprocessed multi-source alarm information undergoes preliminary encoding to obtain the initial node feature matrix;

[0011] S3.2: Construct a fully connected graph structure;

[0012] S3.3: The graph neural network model based on the graph attention mechanism iteratively updates the initial node feature matrix using a directional entity perception graph attention network to obtain the final high-order node feature matrix;

[0013] S3.4: Input the final high-order node feature matrix into the classifier to perform entity recognition and relationship classification;

[0014] S4: Construct a knowledge graph of ship navigation risks based on the entities and relationships described;

[0015] S5: Intelligent identification of ship navigation risk alarm information based on the ship navigation risk knowledge graph.

[0016] Preferably, in S3.3, the iteration process starts from the first... From the beginning of layer L to the end of layer L, the input to the graph neural network model based on the graph attention mechanism is the first layer. The node feature matrix of all nodes in the first layer; for the first layer, the input of the graph neural network model based on the graph attention mechanism is the initial node feature matrix. ;

[0017] First, the input to the graph neural network model based on the graph attention mechanism is linearly transformed, and then the importance of node j to node i is calculated as the basic attention score. The calculation formula is:

[0018] ;

[0019] In the formula, This is a vector concatenation operation. These are learnable weight vectors used to map the concatenated high-dimensional vector into a scalar attention score. For transpose operation, For activation function, The i-th node after linear transformation Layer node feature vectors For the j-th node after linear transformation Layer node feature vectors;

[0020] Calculate and add directional entity-aware bias terms ;

[0021] The directional entity perception bias term The expression is:

[0022] ;

[0023] In the formula, λ is the coefficient, and sigmoid is the normalization function. Let be the relative distance between the words of node i and node j. Let i be the entity vector of node i. Let be the entity vector of node j;

[0024] The attention scores of nodes i and j are the sum of the base attention scores and the directional entity perception bias term, expressed as:

[0025] ;

[0026] For node i, the attention scores of all its neighboring nodes j are normalized using softmax to obtain the final attention weights. Then feature aggregation is performed;

[0027] The attention score calculation and feature aggregation steps are performed in parallel on all nodes in the fully connected graph to obtain the first... Node feature matrix of all nodes in the layer The iterative process is executed, and finally, after the Lth layer of iterations, the final high-order node feature matrix of each node is output. The expression is:

[0028] ;

[0029] In the formula, Let be the node feature matrix of the i-th node iterated up to the L-th layer.

[0030] Preferably, in step S3.1, the preprocessed multi-source alarm information is converted into a word vector S, and the expression of the word vector S is: In the formula, Let be the i-th word in the preprocessed multi-source alarm information, and n be the total number of words in the preprocessed multi-source alarm information;

[0031] Each word in the word vector S is encoded using the BERT-Marine-Domain model to obtain a context vector. Simultaneously, positional encoding is added to preserve the sequence order information of each word. At this point, the word vector S is represented as the initial node feature matrix. Its mathematical expression is: .

[0032] Preferably, in step S3.2, each word in the word vector S is... Defined as a node in the fully connected graph structure, thus forming a node set V, the mathematical expression of which is: In the formula, For the i-th node, corresponding to the word ; at any two distinct nodes and An undirected edge is established between them, forming a set of connecting edges E, thus constituting the fully connected graph structure. .

[0033] Preferably, in step S3.4, the classifier includes an entity recognition head and a relationship classification head. The entity recognition head is used to predict the entity label corresponding to each word, and the relationship classification head is used to predict the relationship type of any pair of entities.

[0034] Preferably, in step S1, the multi-source alarm information originates from the ship navigation and safety system, engine room and power monitoring system, safety and fire monitoring system, and meteorological and platform system.

[0035] Preferably, the ship navigation and safety system includes an automatic identification system, radar, electronic chart display and information system, global positioning system, depth sounder, and log; the engine room and power monitoring system includes a centralized engine room monitoring system, main engine remote control system, generator control system, boiler control system, and pump control system; the safety and fire monitoring system includes a fire detection and alarm system, a general alarm system, a carbon dioxide release alarm system, and an emergency shutdown system; and the meteorological and platform system includes an anemometer and compass.

[0036] Preferably, in step S2, the data preprocessing operation includes data cleaning, data deduplication, and timestamp alignment.

[0037] Preferably, in step S4, a graph node is created for each uniquely identified entity and classified according to its type. According to the extracted relationship, labeled edges are created between the corresponding nodes. The values ​​and real-time data extracted from the preprocessed multi-source alarm information are added as attributes to the attribute table of the corresponding node, thereby realizing the construction of a knowledge graph of ship navigation risks.

[0038] According to another aspect of the present invention, a ship navigation risk alarm information intelligent identification system is provided, the system employing the above-described ship navigation risk alarm information intelligent identification method, the system comprising:

[0039] The data acquisition module is used to collect alarm information from multiple sources.

[0040] The data preprocessing module is used to perform data preprocessing operations on the multi-source alarm information;

[0041] The entity and relationship extraction module is used to process the pre-processed multi-source alarm information and extract entities and relationships.

[0042] The knowledge graph construction module is used to construct a knowledge graph of ship navigation risks based on the entities and relationships.

[0043] The intelligent identification module is used to intelligently identify ship navigation risk alarm information based on the ship navigation risk knowledge graph.

[0044] The embodiments of the present invention have the following technical effects:

[0045] This invention employs an attention-based graph neural network model to extract entities and relationships from multi-source alarm information. The pre-processed multi-source alarm information undergoes initial encoding to obtain an initial node feature matrix. A fully connected graph structure is then constructed. The initial node feature matrix is ​​iteratively updated using a directional entity-aware graph attention network to obtain a final high-order node feature matrix. This final high-order node feature matrix is ​​input into a classifier for entity recognition and relationship classification. During the iterative update phase, the output is a completely new final high-order node feature matrix. Each row of this matrix represents a deep feature rich in global contextual information and directional entity awareness, tailored for entity recognition and relationship classification. These deep features are directly input into the subsequent classifier to complete the final extraction task, thereby significantly improving the extraction accuracy and robustness for complex alarm statements.

[0046] Furthermore, unlike the traditional sequential pipeline of entity recognition followed by relation classification, this invention employs an end-to-end joint modeling approach, unifying entity recognition and relation classification within a single graph attention network for simultaneous learning and optimization. This model iteratively updates deep node features that simultaneously contain global context and directional entity perception information, directly inputting them into parallel entity recognition and relation classification heads. This mechanism ensures that both tasks share the same high-quality feature representation, enabling entity information to enhance the accuracy of relation judgments, while relational context also assists in the precise definition of entities. This fundamentally avoids the performance degradation in relation classification caused by the accumulation and propagation of entity recognition errors in traditional solutions. Attached Figure Description

[0047] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a flowchart of an intelligent identification method for ship navigation risk alarm information provided in an embodiment of the present invention;

[0049] Figure 2 This is a flowchart provided by an embodiment of the present invention for processing preprocessed multi-source alarm information and extracting entities and relationships. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0051] Example 1, Figure 1 A flowchart of a method for intelligent identification of ship navigation risk alarm information is shown, such as... Figure 1 As shown, a method for intelligent identification of ship navigation risk alarm information includes the following steps:

[0052] S1: Collect multi-source alarm information;

[0053] In this step, the multi-source alarm information originates from the ship's navigation and safety system, engine room and power monitoring system, safety and fire monitoring system, and meteorological and platform system. The ship's navigation and safety system includes Automatic Identification System (AIS), radar (ARPA), Electronic Chart Display and Information System (ECDIS), Global Positioning System (GPS), depth sounder, log, etc. The engine room and power monitoring system includes Centralized Engine Room Monitoring System (ECS / AMS), Main Engine Remote Control System (MCR), generator control system, boiler control system, pump control system, etc. The safety and fire monitoring system includes Fire Detection and Alarm System (FAS), General Alarm System (GS), Carbon Dioxide Release Alarm System, Emergency Shutdown System (ESD), etc. The meteorological and platform system includes anemometers, compasses, etc.

[0054] To achieve unified collection of the multi-source alarm information, this embodiment adopts a distributed, loosely coupled collection agent architecture. Specifically, a central coordinator is set up and deployed on the ship's server. It is responsible for managing the status, configuration, and logs of all collection agents and provides a unified data receiving API. Lightweight dedicated collection agents are deployed for different types of devices or systems. The collection agents are responsible for establishing connections with the data sources, reading data, and sending it to the central coordinator.

[0055] S2: Perform data preprocessing on the multi-source alarm information;

[0056] The data preprocessing operations include data cleaning, data deduplication, and timestamp alignment. Data cleaning specifically involves filtering for garbled characters and illegal characters, as well as handling null values ​​and missing fields. Data deduplication involves assigning a unique, incrementing ID to each alarm message and comparing the ID of newly acquired alarm messages with the IDs of previously processed alarm messages to eliminate duplicate alarm messages. Format unification and timestamp alignment specifically involve using GPS time as the unified time reference for the ship. At the entry point for collecting multi-source alarm information, the clocks of all acquisition agents and the central server are strictly synchronized with the GPS signal through the NTP client service.

[0057] S3: Perform data processing on the pre-processed multi-source alarm information to extract entities and relationships;

[0058] In the field of intelligent ship alarms, traditional entity and relation extraction methods mostly adopt a scheme of first identifying entities and then classifying entity pairs. However, traditional schemes are prone to error propagation; small errors in the previous entity identification step can be directly passed to the relation classification stage. At the same time, traditional schemes view the relationships between entity pairs in isolation, ignoring the inherent syntactic structure and logical framework of the alarm statement itself. For example, when analyzing the alarm "Generator No.1 tripped due to LO pressure low at 1230," it would naturally be parsed as an event framework of "[Generator No.1] - [tripped due to] - [LO pressure low]." But traditional schemes may only capture the broad and vague relationship of "[Generator No.1] - [related to] - [pressure]," failing to precisely extract the causal logic "tripped due to."

[0059] This embodiment proposes a graph neural network model based on graph attention mechanism and calculates attention scores based on content and directionality entity perception. It performs end-to-end joint modeling and optimization of entity extraction and relation classification tasks on a unified graph structure, thereby significantly improving the extraction accuracy and robustness of complex alarm statements.

[0060] Specifically, such as Figure 2 As shown, S3 specifically includes:

[0061] S3.1: The preprocessed multi-source alarm information is initially encoded to obtain the initial node feature matrix.

[0062] The preprocessed multi-source alarm information is converted into a word vector S, and the expression of the word vector S is: In the formula, Let be the i-th word in the preprocessed multi-source alarm information, and n be the total number of words in the processed multi-source alarm information.

[0063] Each word in the word vector S is encoded using the BERT-Marine-Domain model to obtain a context vector containing rich semantic information. Simultaneously, positional encoding is added to preserve the sequence order information of each word. At this point, the word vector S is represented as the initial node feature matrix. Its mathematical expression is:

[0064] .

[0065] S3.2: Construct a fully connected graph structure;

[0066] In this step, each word in the word vector S is... Defined as a node in the fully connected graph structure, thus forming a node set V, the mathematical expression of which is: In the formula, For the i-th node, corresponding to the word ; at any two distinct nodes and An undirected edge is established between them, forming a set of connecting edges E, thus constituting the fully connected graph structure. .

[0067] S3.3: Iteratively update the initial node feature matrix using a directional entity-aware graph attention network to obtain the final high-order node feature matrix;

[0068] This step is performed iteratively, using superscript. This indicates that the current number is the [number]. Layer, assuming the initial node feature matrix has a total of L layers, the iteration process starts from the The layers begin up to the Lth layer. In the Lth layer... In a graph neural network model based on graph attention mechanisms, the input to the previous layer (the first layer) is... The node feature matrix of all nodes in the first layer (layer 1). For the first layer, the input of the graph neural network model based on the graph attention mechanism is the initial node feature matrix. .

[0069] To enhance the expressive power of the graph neural network model based on graph attention mechanism, the input of the model is first linearly transformed, and then the importance of node j to node i is calculated as the basic attention score. The calculation formula is as follows:

[0070] ;

[0071] In the formula, This is a vector concatenation operation. These are learnable weight vectors used to map the concatenated high-dimensional vector into a scalar attention score. For activation function, For transpose operation, The i-th node after linear transformation Layer node feature vectors For the j-th node after linear transformation The feature vectors of the nodes in the layer.

[0072] Calculate and add directional entity-aware bias terms ;

[0073] Wherein, the directional entity perception bias term The directional entity perception bias term is used to guide the graph neural network model based on the graph attention mechanism to focus on entity structures. The expression is:

[0074] ;

[0075] In the formula, λ is a coefficient used to control the overall strength of the bias term, and sigmoid is the normalization function. Let be the relative distance between words in nodes i and j. When I was very young, The value will be relatively large, which will affect the directional entity perception bias term. When the relative distance between words in node i and node j increases, As the value decreases, the bias effect weakens. The above formula encourages the model to prioritize combinations of neighboring words. Let i be the entity vector of node i. Let be the entity vector of node j.

[0076] The attention scores of nodes i and j are the sum of the base attention scores and the directional entity perception bias term, expressed as:

[0077] ;

[0078] Therefore, the attention scores of nodes i and j are fused with the attention scores of content and directional entity perception.

[0079] In this embodiment, an undirected fully connected graph structure is first established. This fully connected graph structure does not impose any hard constraints on which words may have relationships, thus ensuring that any potential inter-entity relationship, regardless of how far apart they are in a sentence, is physically connected at the graph topology level. At the same time, by adding a directional entity-aware bias term, the directed attention mechanism allows the fully connected graph structure to focus its main efforts on neighboring nodes that are more relevant to the current node in terms of entity structure during each aggregation, thereby achieving an optimal balance between structural flexibility and semantic targeting.

[0080] For node i, the attention scores of all its neighboring nodes j are normalized using softmax to obtain the final attention weights. Then feature aggregation is performed.

[0081] The attention score calculation and feature aggregation steps are performed in parallel on all nodes in the fully connected graph to obtain the first... Node feature matrix of all nodes in the layer ; The above process is from the first The process is repeated up to the Lth layer, with each layer refining the representation from the previous layer at a deeper level. Finally, after the Lth layer iteration, the final high-order node feature matrix for each node is output. The expression is:

[0082] ;

[0083] In the formula, Let be the node feature matrix of the i-th node iterated up to the L-th layer.

[0084] The output of this step is a completely new final high-order node feature matrix. Each row of this matrix, i.e., each... These are no longer the original contextual semantic vectors; they are deep features rich in global contextual information and directional entity perception, tailored for entity recognition and relationship classification. These deep features are directly input into the subsequent classifier to complete the final extraction task, thereby significantly improving the extraction accuracy and robustness of complex alarm statements.

[0085] S3.4: Input the final high-order node feature matrix into the classifier to perform entity recognition and relationship classification;

[0086] The classifier includes an entity recognition head and a relationship classification head. The entity recognition head is used to predict the entity label corresponding to each word, and the relationship classification head is used to predict the relationship type of any pair of entities.

[0087] The entity recognition head is a shallow classification network used to map the representation of each node to its corresponding entity label; the relationship classification head is another independent classification network used to process paired entity node representations and predict the relationship between them; the entity recognition head and the relationship classification head are trained simultaneously, and the losses during the training process are weighted and merged into a total loss function, which is then used to jointly optimize the parameters of the entire model through backpropagation.

[0088] Wherein, the loss function of the entity recognition head Using standard cross-entropy loss, for all n nodes, the cross-entropy between the predicted distribution and the true label of each node is calculated, and then averaged; the specific formula is as follows:

[0089] ;

[0090] In the formula, S is the set of entities. Let i be the actual entity label of node i.

[0091] The loss function of the relation classification head The cross-entropy loss is also used, but it is calculated only on all valid positive entity pairs. To avoid excessive negative samples diluting the loss, a small number of negative samples are typically sampled within a sentence; the expression for the loss function of the relation classification head is:

[0092] ;

[0093] In the formula, P is the set of all valid positive entity pairs in a training batch. For entity pairs The true relationship label.

[0094] The loss function of the entity recognition head and the loss function of the relation classification head are weighted and summed to obtain the total loss function. The expression is:

[0095] ;

[0096] In the formula, γ is an adjustable hyperparameter, and the selection of γ directly affects the balance of the model's performance on the two tasks. If γ is too large, the model will focus more on learning entity recognition, which may impair the performance of relation classification, and vice versa. In this step, γ=0.5.

[0097] S4: Construct a knowledge graph of ship navigation risks based on the entities and relationships described;

[0098] A graph node is created for each uniquely identified entity and classified according to its type. Based on the extracted relationships, labeled edges are created between the corresponding nodes. The values ​​extracted from the preprocessed multi-source alarm information and real-time data are added as attributes to the attribute table of the corresponding node, thereby realizing the construction of a knowledge graph of ship navigation risks.

[0099] S5: Intelligent identification of ship navigation risk alarm information based on the ship navigation risk knowledge graph;

[0100] By leveraging the association query capabilities of the aforementioned ship navigation risk knowledge graph, the root cause of alarms can be traced, enabling intelligent identification of ship navigation risk alarm information. For example, upon detecting a "generator trip" alarm, the path in the ship navigation risk knowledge graph can be queried to find its upstream association, such as a "high fuel filter differential pressure" alarm, thereby pinpointing the root cause.

[0101] Example 2: The present invention also provides an intelligent identification system for ship navigation risk alarm information. The system employs an intelligent identification method for ship navigation risk alarm information as described in Example 1. The system includes:

[0102] The data acquisition module is used to collect alarm information from multiple sources.

[0103] The data preprocessing module is used to perform data preprocessing operations on the multi-source alarm information;

[0104] The entity and relationship extraction module is used to process the pre-processed multi-source alarm information and extract entities and relationships.

[0105] The knowledge graph construction module is used to construct a knowledge graph of ship navigation risks based on the entities and relationships.

[0106] The intelligent identification module is used to intelligently identify ship navigation risk alarm information based on the ship navigation risk knowledge graph.

[0107] Example 3: The present invention also provides an electronic device, including one or more processors and a memory.

[0108] A processor can be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and can control other components in an electronic device to perform desired functions.

[0109] The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may execute the program instructions to implement the intelligent identification method for ship navigation risk alarm information described in any embodiment of this application above, and / or other desired functions. Various contents such as initial external parameters and thresholds may also be stored in the computer-readable storage medium.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A ship navigation risk alarm information intelligent identification method, characterized in that, The method includes the following steps: S1: Collect multi-source alarm information; S2: Perform data preprocessing on the multi-source alarm information; S3: Perform data processing on the pre-processed multi-source alarm information to extract entities and relationships; Specifically, S3 is: S3.1: The preprocessed multi-source alarm information undergoes preliminary encoding to obtain an initial node feature matrix; S3.2: A fully connected graph structure is constructed; S3.3: A graph neural network model based on a graph attention mechanism iteratively updates the initial node feature matrix using a directional entity perception graph attention network to obtain the final high-order node feature matrix; In S3.3, the iteration process starts from the first... From the beginning of layer L to the end of layer L, the input to the graph neural network model based on the graph attention mechanism is the first layer. The node feature matrix of all nodes in the first layer; for the first layer, the input of the graph neural network model based on the graph attention mechanism is the initial node feature matrix. ; First, the input to the graph neural network model based on the graph attention mechanism is linearly transformed, and then the importance of node j to node i is calculated as the basic attention score. The calculation formula is: ； In the formula, This is a vector concatenation operation. These are learnable weight vectors used to map the concatenated high-dimensional vector into a scalar attention score. For transpose operation, For activation function, The i-th node after linear transformation Layer node feature vectors For the j-th node after linear transformation Layer node feature vectors; Calculate and add directional entity-aware bias terms ; The directional entity perception bias term The expression is: ； In the formula, λ is the coefficient, and sigmoid is the normalization function. Let be the relative distance between the words of node i and node j. Let i be the entity vector of node i. Let be the entity vector of node j; The attention scores of nodes i and j are the sum of the base attention scores and the directional entity perception bias term, expressed as: ； For node i, the attention scores of all its neighboring nodes j are normalized using softmax to obtain the final attention weights. Then feature aggregation is performed; The attention score calculation and feature aggregation steps are performed in parallel on all nodes in the fully connected graph to obtain the first... Node feature matrix of all nodes in the layer The iterative process is executed, and finally, after the Lth layer of iterations, the final high-order node feature matrix of each node is output. The expression is: ； In the formula, Let be the node feature matrix of the i-th node iterated up to the L-th layer; S3.4: Input the final high-order node feature matrix into the classifier to perform entity recognition and relationship classification; S4: Construct a knowledge graph of ship navigation risks based on the entities and relationships described; S5: Intelligent identification of ship navigation risk alarm information based on the ship navigation risk knowledge graph.

2. The intelligent identification method for ship navigation risk alarm information according to claim 1, characterized in that, In step S3.1, the preprocessed multi-source alarm information is converted into a word vector S, and the expression of the word vector S is: In the formula, Let be the i-th word in the preprocessed multi-source alarm information, and n be the total number of words in the preprocessed multi-source alarm information; Each word in the word vector S is encoded using the BERT-Marine-Domain model to obtain a context vector. Simultaneously, positional encoding is added to preserve the sequence order information of each word. At this point, the word vector S is represented as the initial node feature matrix. Its mathematical expression is: .

3. The intelligent identification method for ship navigation risk alarm information according to claim 2, characterized in that, In step S3.2, each word in the word vector S... Defined as a node in the fully connected graph structure, thus forming a node set V, the mathematical expression of which is: In the formula, For the i-th node, corresponding to the word ; at any two distinct nodes and An undirected edge is established between them, forming a set of connecting edges E, thus constituting the fully connected graph structure. .

4. The intelligent identification method for ship navigation risk alarm information according to claim 1, characterized in that, In step S3.4, the classifier includes an entity recognition head and a relationship classification head. The entity recognition head is used to predict the entity label corresponding to each word, and the relationship classification head is used to predict the relationship type of any pair of entities.

5. The intelligent identification method for ship navigation risk alarm information according to claim 1, characterized in that, In S1, the multi-source alarm information comes from the ship navigation and safety system, engine room and power monitoring system, safety and fire monitoring system, and meteorological and platform system.

6. The intelligent identification method for ship navigation risk alarm information according to claim 5, characterized in that, The ship navigation and safety system includes an automatic identification system, radar, electronic chart display and information system, global positioning system, depth sounder, and log; the engine room and power monitoring system includes a centralized engine room monitoring system, main engine remote control system, generator control system, boiler control system, and pump control system; the safety and fire monitoring system includes a fire detection and alarm system, a general alarm system, a carbon dioxide release alarm system, and an emergency shutdown system; the meteorological and platform system includes an anemometer and compass.

7. The intelligent identification method for ship navigation risk alarm information according to claim 1, characterized in that, In step S2, the data preprocessing operation includes data cleaning, data deduplication, and timestamp alignment.

8. The intelligent identification method for ship navigation risk alarm information according to claim 1, characterized in that, In step S4, a graph node is created for each uniquely identified entity and classified according to its type. Based on the extracted relationships, labeled edges are created between the corresponding nodes. The values ​​extracted from the preprocessed multi-source alarm information and real-time data are added as attributes to the attribute table of the corresponding node, thereby realizing the construction of a knowledge graph of ship navigation risks.

9. A smart identification system for ship navigation risk alarm information, characterized in that, The system employs a ship navigation risk alarm information intelligent identification method according to any one of claims 1-8, and the system includes: The data acquisition module is used to collect alarm information from multiple sources. The data preprocessing module is used to perform data preprocessing operations on the multi-source alarm information; The entity and relationship extraction module is used to process the pre-processed multi-source alarm information and extract entities and relationships. The knowledge graph construction module is used to construct a knowledge graph of ship navigation risks based on the entities and relationships. The intelligent identification module is used to intelligently identify ship navigation risk alarm information based on the ship navigation risk knowledge graph.

Images