Intelligence intention understanding and report generation method and system based on multi-agent

Abstract

The application discloses a multi-agent-based intelligence intention understanding and report generation method and system, relates to the technical field of big data processing, and comprises the following steps: acquiring a marine intelligence text, extracting an event timestamp, a dynamic reference target and a relative spatial description parameter of a target to be positioned; searching for an instantaneous motion state vector of the reference target at a corresponding moment in an AIS historical database; converting the relative position into an absolute geodetic azimuth, calculating the absolute geographic coordinates of the target to be positioned by using a geodetic subject direct calculation algorithm; projecting the absolute geographic coordinates into a multi-dimensional marine GIS layer, and performing spatial topology analysis; according to the spatial topology analysis result, reasoning the intention of the target to be positioned by using a large language model, and generating an intelligence report. The application solves the positioning failure problem caused by the dynamic change of the reference system by utilizing multi-agent cooperation, realizes the conversion from fuzzy relative semantics to accurate absolute coordinates, and improves the depth and accuracy of marine intelligence intention mining.

Description

Technical Field

[0001] This application relates to the field of big data processing technology, and in particular to a method and system for intelligence intent understanding and report generation based on multi-agent intelligence. Background Technology

[0002] With the exponential growth of marine data, situational awareness using open-source intelligence has become a core industry requirement. Currently, large language models and multi-agent collaborative architectures have made significant progress in automated information processing, effectively improving the efficiency of complex processes by assigning different agents to tasks such as information extraction, logical reasoning, and content generation. However, general multi-agent architectures still have certain technical limitations when processing unstructured text in the marine vertical domain.

[0003] Maritime texts widely employ relative reference frames based on moving entities, such as a reference frame located 15 nautical miles to port on a ship's course. Their spatial attributes exhibit high temporal sensitivity and dynamic change characteristics. Existing multi-agent systems are mostly limited to purely semantic-level interactions, lacking a deep coupling mechanism between natural language and spatiotemporal data from the physical world. This means that the agent swarm cannot combine the specific timestamp of the text release to trace back the actual latitude and longitude of the reference point in the physical world at that time. Consequently, the computer cannot understand the true geographical scope described in the text, and therefore cannot accurately determine whether the event has a substantial intentional impact on a specific shipping route or port. The resulting reports often contain numerous location errors or semantically ambiguous factual illusions, reducing the credibility of intelligence reports.

[0004] Therefore, there is an urgent need for a technical solution that can understand dynamic spatiotemporal reference relationships and automatically convert relative semantics into absolute geographic coordinates. Summary of the Invention

[0005] To address the aforementioned shortcomings, this application provides a method and system for intelligence intent understanding and report generation based on multi-agent intelligence, which solves the technical problem that existing technologies cannot accurately perform spatial positioning and intent inference when processing unstructured maritime text containing moving reference objects.

[0006] Firstly, this application provides a method for understanding intelligence intent and generating reports based on multi-agent intelligence, including: The system acquires marine intelligence text, identifies and extracts event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located; the relative spatial description parameters include relative azimuth description and relative distance values ​​based on the dynamic reference target's body coordinate system. Based on the dynamic reference target and the event timestamp, the instantaneous motion state vector of the target at the corresponding moment is retrieved from the AIS historical database; the instantaneous motion state vector includes reference latitude and longitude coordinates and reference heading angle; Based on the reference heading angle, the relative orientation description is converted into the absolute geodetic azimuth angle in the true north coordinate system; using the reference latitude and longitude coordinates as the starting point, the absolute geographic coordinates of the target to be located are calculated by using the geodetic forward calculation algorithm and combining the absolute geodetic azimuth angle with the relative distance value. The absolute geographic coordinates are projected onto a multidimensional marine GIS layer, and spatial topology analysis is performed to determine whether the target to be located falls into a preset sensitive area or whether it has spatiotemporal intersection with a predetermined route. Based on the spatial topology analysis results, the intent of the target to be located is inferred using a large language model, and an intelligence report is generated.

[0007] Optionally, the identification and extraction of event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located includes: The marine intelligence text is input into a pre-trained entity extraction model to identify time entities, ship name entities, and bearing distance descriptors. The time entity is preprocessed and converted into a unified timestamp format as an event timestamp; The vessel name entity is matched with preset vessel file data to determine the unique identifier of the dynamic reference target; Syntactic dependency analysis is performed on the directional distance descriptors to separate the directional words representing direction and the numerical values ​​representing distance, which are used as relative directional descriptions and relative distance values, respectively.

[0008] Optionally, the relative orientation description includes a fuzzy description based on clock face orientation or a precise description based on hull angle; In response to the relative orientation description being a fuzzy description based on clock face orientation, a mapping relationship between clock face dots and angles is established, and the extracted clock face dots are converted into relative angle values ​​relative to the dynamic reference target ship's bow baseline. In response to the relative bearing description being a precise description based on the hull angle, the port and starboard attributes in the bearing term are identified, and a sign transformation is performed on the extracted angle value based on the port and starboard attributes to obtain the relative angle value relative to the bow baseline of the dynamic reference target ship.

[0009] Optionally, the step of retrieving the instantaneous motion state vector of the target at the corresponding moment in the AIS historical database based on the dynamic reference target and the event timestamp includes: Based on the unique identifier of the dynamic reference target, query the AIS historical database for a time window including the event timestamp, and obtain the first AIS data point at the start time of the time window and the second AIS data point at the end time of the time window. Extract the latitude and longitude coordinates and heading angles of the first and second AIS data points; Using an interpolation algorithm, the reference latitude and longitude coordinates and reference heading angle of the dynamic reference target at the event timestamp are calculated based on the time proportion of the event timestamp within the time window, and combined to form an instantaneous motion state vector.

[0010] Optionally, the step of converting the relative bearing description into an absolute geodetic azimuth in the true north coordinate system based on the reference heading angle includes: Obtain the relative angle value corresponding to the relative orientation description; The reference heading angle and the relative angle value are summed and calculated. The summation result is normalized based on the period of the circular angle to obtain the absolute geodetic azimuth relative to true north.

[0011] Optionally, the step of using the geodetic forward calculation algorithm, combining the absolute geodetic azimuth and the relative distance value, to calculate the absolute geographic coordinates of the target to be located includes: Initialize the parameters of the reference ellipsoid model, including the semi-major axis and flattening of the ellipsoid; Using the reference latitude and longitude coordinates as the starting point of the geodetic line, the absolute geodetic azimuth angle as the initial geodetic azimuth angle, and the relative distance value as the length of the geodetic line; Based on the parameters of the reference ellipsoid model, the geodetic equations are solved using an iterative approximation method or an ellipsoidal series expansion method to obtain the latitude and longitude of the geodetic endpoint, which are then used as the absolute geographic coordinates of the target to be located.

[0012] Optionally, the step of performing spatial topology analysis to determine whether the target to be located falls within a preset sensitive area includes: Obtain vector boundary data of the preset sensitive area; Calculate the geometric inclusion relationship between the absolute geographic coordinates and the vector boundary data to generate a region attribution identifier; the region attribution identifier is used to characterize whether the target to be located is located inside, on the boundary, or outside the sensitive area.

[0013] Optionally, the step of performing spatial topology analysis to determine whether the target to be located has spatiotemporal intersection with the predetermined flight path includes: Based on the center trajectory line of the predetermined route, a preset safety distance threshold is extended to both sides to generate a route buffer zone; Calculate the spatial overlap between the absolute geographic coordinates and the route buffer, and verify whether the event timestamp is within the planned travel time window of the predetermined route; In response to the absolute geographic coordinates being located within the route buffer and the event timestamp being within the planned travel time window of the predetermined route, a spatiotemporal intersection is determined to exist, and a spatiotemporal intersection confirmation identifier is generated. In response to the absolute geographic coordinates being outside the route buffer or the event timestamp being outside the planned travel time window of the predetermined route, a spatiotemporal intersection negation flag is generated.

[0014] Optionally, the step of inferring the intent of the target to be located using a large language model based on the spatial topology analysis results and generating an intelligence report includes: Construct inference prompt text, which includes the absolute geographic coordinates of the target to be located, the event timestamp, and the spatial topology analysis results; The inference prompt text is input into the large language model, which combines the marine domain knowledge base to analyze the behavioral semantics of the target to be located in the spatiotemporal context and outputs a natural language description text representing the navigation intention or operation purpose of the target to be located. According to the preset intelligence report structure, the event timestamp, the absolute geographic coordinates, the spatial topology analysis results, and the natural language description text are filled into the corresponding fields to generate an intelligence report.

[0015] Secondly, this application provides a multi-agent-based intelligence intent understanding and report generation system, including: The semantic parsing intelligent agent module is used to acquire marine intelligence text, identify and extract event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located; the relative spatial description parameters include relative orientation description and relative distance value based on the dynamic reference target's ontological coordinate system; The spatiotemporal backtracking intelligent agent module is used to retrieve the instantaneous motion state vector of the target at the corresponding moment in the AIS historical database based on the dynamic reference target and the event timestamp; the instantaneous motion state vector includes reference latitude and longitude coordinates and reference heading angle; The target localization agent module is used to convert the relative orientation description into an absolute geodetic azimuth angle in the true north coordinate system based on the reference heading angle; and to calculate the absolute geographic coordinates of the target to be located by using the geodetic forward calculation algorithm, combined with the absolute geodetic azimuth angle and the relative distance value, starting from the reference latitude and longitude coordinates. The intent reasoning intelligent agent module is used to project the absolute geographic coordinates onto a multi-dimensional marine GIS layer, perform spatial topology analysis, determine whether the target to be located falls into a preset sensitive area or whether it has spatiotemporal intersection with a predetermined route; based on the spatial topology analysis results, it uses a large language model to infer the intent of the target to be located and generates an intelligence report.

[0016] Compared with the prior art, the beneficial effects of the present invention are: A semantic and physical dual-space alignment mechanism was introduced. Through multi-agent collaboration, the instantaneous heading and position of the reference target were traced back in the historical AIS database using timestamps. Combined with the geodetic forward calculation algorithm, the relative description dependent on the ontology attitude was accurately converted into latitude and longitude in the true north coordinate system. This effectively solved the positioning failure problem caused by the dynamic change of the reference system in marine scenarios and realized the conversion from fuzzy relative semantics to precise absolute coordinates.

[0017] After obtaining the absolute geographic coordinates, this application further projects them into a marine geographic information system for spatial topology analysis. The system can automatically detect whether the target falls into a sensitive sea area or whether it constitutes a spatiotemporal approach to a high-value shipping route, thereby improving the tactical depth and accuracy of intent mining. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the multi-agent-based intelligence intent understanding and report generation method provided in this application embodiment; Figure 2 A flowchart for text semantic parsing of marine intelligence is provided as an embodiment of this application; Figure 3 A flowchart for generating an instantaneous motion state vector of a dynamic reference target is provided in an embodiment of this application; Figure 4 A schematic diagram of a multi-agent intelligence intent understanding and report generation system provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0020] See Figure 1 The flowchart of the multi-agent intelligence intent understanding and report generation method provided in the embodiments of this application includes steps S101 to S104, wherein: S101, acquire marine intelligence text, identify and extract event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located; the relative spatial description parameters include relative orientation description and relative distance values ​​based on the dynamic reference target's body coordinate system; S102, based on the dynamic reference target and the event timestamp, retrieve the instantaneous motion state vector of the target at the corresponding moment from the AIS historical database; the instantaneous motion state vector includes reference latitude and longitude coordinates and reference heading angle; S103, Based on the reference heading angle, the relative orientation description is converted into the absolute geodetic azimuth angle in the true north coordinate system; Taking the reference latitude and longitude coordinates as the starting point, the absolute geographic coordinates of the target to be located are calculated using the geodetic forward calculation algorithm, combined with the absolute geodetic azimuth angle and the relative distance value. S104, the absolute geographic coordinates are projected onto a multidimensional marine GIS layer, spatial topology analysis is performed to determine whether the target to be located falls into a preset sensitive area or whether it has a spatiotemporal intersection with a predetermined route; based on the spatial topology analysis results, the intent of the target to be located is inferred using a large language model, and an intelligence report is generated.

[0021] Regarding the above S101: Existing marine intelligence data sources, such as social media, public news, and radio voice transcripts, are mostly in unstructured natural language. The key spatiotemporal information contained therein exhibits significant ambiguity, dispersion, and non-standardization. Specifically, the formats for describing event times vary, such as 14:00 on October 20, 2025, late May, and yesterday afternoon, lacking a unified benchmark. Spatial location information often relies on moving reference objects, such as ships, for relative descriptions rather than directly providing latitude and longitude coordinates, making it difficult for computers to directly use physical models for calculations.

[0022] Specifically, traditional general natural language processing (NLP) models struggle to handle specific semantics in the marine vertical domain, such as unstructured relative descriptions that rely on the attitude of moving reference objects, like "a ship at 45 degrees to port" or "10 o'clock." They cannot directly obtain absolute coordinates that can be used for GIS positioning, resulting in low automation in intelligence processing and lagging spatiotemporal perception.

[0023] In this embodiment, step S101 constructs a semantic parsing agent for the marine vertical domain, using natural language processing technology to map unstructured text into a machine-readable structured parameter sequence. Specifically, this application employs entity recognition technology based on deep neural networks. Through pre-training and fine-tuning, it utilizes the general language representation capabilities obtained from training on large-scale corpora, combined with marine-specific corpora for transfer learning, to achieve the capture of specific entities.

[0024] See Figure 2 The flowchart provided in this application embodiment for text semantic parsing of marine intelligence includes steps S201-S204, wherein: S201, Input the marine intelligence text into a pre-trained entity extraction model to identify time entities, ship name entities, and bearing distance descriptors; S202, preprocess the time entity and convert it into a unified timestamp format as an event timestamp; S203, Match the vessel name entity with the preset vessel file data to determine the unique identifier of the dynamic reference target; S204, perform syntactic dependency analysis on the directional distance descriptor to separate the directional words representing direction and the numerical values ​​representing distance, which are respectively used as relative directional description and relative distance value.

[0025] In practical implementation, the first step is to construct and preprocess a dedicated dataset for the marine field. This implementation method involves writing a targeted web crawler program to collect data from the official website of the Maritime Safety Administration and professional shipping news websites. It connects to the electronic navigation log system via API interfaces and utilizes automatic speech recognition equipment to access the ship's Very High Frequency (VHF) communication system to collect marine intelligence text. This marine intelligence text refers to unstructured text data including information on marine navigation events, ship operation information, and marine area dynamics. Sources include maritime reports, publicly available maritime information announcements, ship VHF communication records, and marine monitoring logs, providing raw intelligence materials for analysis.

[0026] This application collected over 10,000 pieces of marine-specific text data, including historical maritime accident investigation reports, publicly available ship logs, maritime search and rescue exercise reports, and transcribed ship VHF radio communication records. The collected corpus was cleaned, irrelevant symbols were removed, and a BIO (Begin, Inside, Outside) annotation system was established for annotation. The following entity labels were defined: the label B / I-TIME is used to annotate the event timestamp, i.e., the specific time when the observation or event occurred; the label B / I-REF is used to annotate dynamic reference targets; and the labels B / I-BEARING and B / I-DIST are used to annotate relative bearing descriptions and relative distance values, respectively. After annotation, the dataset was divided into training, validation, and test sets in an 8:1:1 ratio as model input.

[0027] Secondly, a marine intelligence entity recognition model is constructed and trained. To accurately capture domain features while retaining generalization ability, this application can adopt a deep learning-based sequence labeling architecture. As a preferred embodiment, a BERT-BiLSTM-CRF composite architecture is used for illustration. However, those skilled in the art will understand that the backbone network of the model can also be replaced with similar pre-trained models such as RoBERTa, ALBERT, or ERNIE, and the decoding layer can also adopt other structures such as Softmax.

[0028] Specifically, the first layer of the model is an embedding layer based on BERT (Bidirectional Encoder Representations from Transformers). This layer first preprocesses the input natural language text: specifically, the original text is input into the WordPiece tokenizer自带 by BERT, which splits words into subword units based on a preset vocabulary, adds a <[BOS_never_used_51bce0c785ca2f68081bfa7d91973934]> token at the beginning of the sentence and a [SEP] token at the end of the sentence, and then converts the split Token sequence into a machine-readable character index sequence according to the vocabulary mapping relationship.

[0029] Wherein, the preset vocabulary refers to the vocabulary file carried when the pre-trained model is released, such as vocab.txt, which includes about 21128 commonly used Chinese characters, characters and special symbols. The specific vocabulary mapping process is as follows: the tokenizer splits the input marine intelligence text into character or subword units, automatically adds the classification identifier at the beginning of the sentence and the [SEP] separation identifier at the end of the sentence. Then, traverse the split Token sequence and find the unique integer index corresponding to each Token in the preset vocabulary. For example, the Chinese character "海" (sea) in the text corresponds to an index of 3146 in the vocabulary, so "海" is converted to 3146. Finally, the text is converted into a Tensor that can be calculated by the model and input to the BERT layer.

[0030] The data input to the model is the character index sequence after tokenization processing , wherein N0 is the maximum length of the input sequence supported by the model for a single inference, that is, the maximum number of characters or Tokens input to the model after tokenization processing, which can be set to 128 in this embodiment. The BERT layer loads a general Chinese pre-trained model, such as the weight parameters of BERT-base-Chinese, and uses its internal 12-layer Transformer encoder structure to capture the deep semantic dependencies between characters, and outputs the corresponding word vector sequence , where each represents the context semantic vector obtained after BERT encoding of the n-th character or Token in the input sequence, with a value range of 1≤n≤N0. Since the hidden layer configuration of the BERT-base model is fixed at 768, each is a 768-dimensional vector.

[0031] The BERT-base-Chinese pre-trained weights, having been trained on a large-scale general corpus, possess basic word embedding representation capabilities and syntactic feature extraction capabilities. Using the aforementioned marine-specific dataset, the model undergoes full parameter fine-tuning, and the model parameters are updated via backpropagation. This allows the general language representation capabilities to be transferred and adapted to entity recognition tasks in the marine intelligence domain, thereby addressing the problem of training difficulties caused by the scarcity of domain-labeled samples.

[0032] The second layer is a Bi-directional Long-Short Term Memory (Bi-LSTM) feature fusion layer, consisting of two LSTM units: a forward LSTM and a backward LSTM. The hidden layer dimension is set to 256, and the dropout ratio is set to 0.3. The forward LSTM reads the vector sequence output by BERT from left to right, while the backward LSTM reads from right to left. Through bidirectional recursive computation, long-range dependencies in the text sequence are captured. Finally, the forward and backward hidden states are concatenated to output a hidden state vector sequence containing rich contextual information. At this point, the feature vector dimension of each time step is 512.

[0033] The output of the Bi-LSTM layer is fed into a fully connected layer for dimensionality reduction mapping, mapping the 512-dimensional features to the label space dimension. That is, the number of tag types, to obtain the emission score matrix for each tag at each time step. , dimension ,in The length of the text sequence, and the elements in the matrix. Indicates at time The model predicts the corresponding character based on context information. The nonnormalized probability score of each label.

[0034] Finally, the Conditional Random Field (CRF) decoding layer is connected, and the CRF layer receives the transmit-receive fraction matrix. And a learnable label transition matrix is ​​introduced. This constrains the legality of the predicted label sequence. Based on the BIO labeling system defined in this application embodiment, it includes four entity types: TIME, REF, BEARING, and DIST. The label set includes {O, B-TIME, I-TIME, B-REF, I-REF, B-BEARING, I-BEARING, B-DIST, I-DIST} and start and stop markers START and STOP. The total number of labels... Approximately 11.

[0035] Therefore, the label transition matrix It is an 11×11 square matrix used to model the adjacency dependencies between tags. Matrix elements... This represents the probability score weight for transitioning from label i at the previous time step to label j at the current time step. For example, in the BIO annotation system, the label "B-REF" (beginning of reference target) can only be followed by "I-REF" (inside the reference target) or "O" (non-entity), but not "I-TIME" (inside time). Therefore, the weight for transitioning from label "B-REF" to "I-REF"... It would be a large positive value; however, directly transferring from the label "B-REF" to "I-TIME" is logically illegal, as the entity type changes abruptly and there is no start label. During training, the model will automatically learn... This is a very small value. In this way, the CRF layer can effectively suppress the generation of illegal paths during decoding, ensuring that the output tag sequence conforms to BIO syntax constraints.

[0036] In its implementation, the CRF layer utilizes the Viterbi algorithm for global decoding. The Viterbi algorithm is a dynamic programming-based optimal path search algorithm that finds the global optimum by recursively calculating the maximum cumulative score at each time step for each state. Specifically, the input data is the emission score matrix output by the Bi-LSTM. and the transition matrix of the CRF layer Define variables Indicates at time That is, the first of the sequence The character is predicted to be a tag. The maximum cumulative score when all previous paths were optimal; and define variables. Used to record the label index of the previous time step that makes the maximum value true. For each time step in the sequence... From the second character to the... Each character is used to calculate the score for the current state using the following recursive formula: ; The above formula represents the current time. Selected label The maximum score is equal to the maximum cumulative score of all possible labels i in the previous time step plus the score from the previous time step. Transferred to In the transfer score The maximum value, plus the value determined by Bi-LSTM at the current time. launch fraction .

[0037] Through the above recursive calculations, the sequence is processed until the last time step. Select The maximum value in the record is taken as the highest score for the entire path. Using the record index, the algorithm backtracks from back to front to determine the best label for each time step, ultimately outputting a path of length [length missing]. Global optimal label sequence .

[0038] During model training, the constructed training set is input into the model. During training, the loss function can be set to the negative log-likelihood function, the AdamW optimizer can be used, and a hierarchical learning rate strategy can be implemented. The BERT layer uses a low learning rate of 2e-5 for fine-tuning to preserve general semantics, while the BiLSTM and CRF layers use a higher learning rate of 1e-3 to quickly adapt to the task. The batch size is set to 16, and the training epochs are set to 100. During training, if the F1 score on the validation set does not improve within 10 consecutive epochs, an early stopping mechanism is triggered, and the optimal parameter file is saved.

[0039] During the model deployment and application phase, real-time acquired marine intelligence text is input into the trained model, which outputs the optimal label sequence. Based on the boundaries of the BIO labels, entity text fragments are segmented from the original text. These entity text fragments include time entity strings representing time, ship name entity strings representing reference targets, and bearing distance strings representing positional relationships.

[0040] The aforementioned entity text fragments are then post-processed to generate relative spatial description parameters. These parameters refer to the geometric positional relationship data of the target to be located relative to a dynamic reference target. The post-processing involves performing syntactic dependency analysis, numerical extraction, and logical operations on the fragments extracted by the model, tailored to specific entity types. Specifically, for the extracted time entity strings, regular expressions are used to extract the year, month, date, and time numbers. If the suffix "Z" or the keyword "UTC" is detected, it is processed directly according to International Standard Time (USST). If no time zone identifier is found, it is calibrated by default according to the time zone of the report's origin. Finally, a time library function is called to convert it to the ISO 8601 standard UTC timestamp format as the event timestamp. .

[0041] For the extracted vessel name entities, an edit distance algorithm is used to perform fuzzy string matching with a pre-set vessel archive database. The similarity between the extracted name and the standard name in the database is calculated. The record with the highest similarity exceeding a preset similarity threshold is selected, and the corresponding Maritime Mobile Service Identifier (MMSI) code is extracted from this record to determine the unique identity of the dynamic reference target. The preset similarity threshold can be determined based on the matching accuracy of the validation set, with a preferred value of 0.80 to 0.95.

[0042] For location and distance descriptions, a proximity-based syntactic dependency analysis is performed. That is, when multiple numerical values ​​or multiple directional words exist in the text, regular expressions are used to separate the directional words representing direction from the numerical values ​​representing distance, based on the modification relationships between words. Specifically, for relative location descriptions, the extracted numerical values ​​are labeled as angle values, and the words preceding and following the numerical values ​​are identified as directional words, such as "port," "starboard," and "o'clock." For relative distance values, the extracted numerical values ​​are labeled as distance modulus, and adjacent quantifiers, such as "nautical miles" and "kilometers," are identified as units.

[0043] Subsequently, the parameters are quantified and normalized. For relative orientation descriptions, based on the varying precision of the text content, they are categorized into fuzzy descriptions based on clock face orientation and precise descriptions based on hull angle. If the relative orientation description is a fuzzy description based on clock face orientation, such as "10 o'clock direction," a mapping relationship between clock face numbers and angles is established, and the extracted clock face numbers are... (Values ​​range from 1 to 12) are converted to relative angle values ​​relative to the bow baseline of the dynamic reference target. The calculation formula is: ; The relative bearing description is a precise description based on the hull angle, such as "45 degrees to port." It identifies the port, starboard, or left and right attributes in the bearing term and extracts the angle value. Based on the port and starboard attributes, the extracted angle values ​​are sign-transformed. If it is port, the relative angle is calculated according to the navigational convention of port being negative and starboard being positive, using the following formula: ; If the keyword includes starboard or starboard, the calculation formula is: ; For relative distance values, normalization is calculated; specifically, the values ​​in the distance segments are extracted. Using the keyword "unit," this method unifies different units of measurement into the standard unit, the meter. Let the unit conversion factor be... If the unit is nautical miles, It is 1852; if it is a chain, It is 185.2; if it is kilometers, It is 1000. Relative distance. The calculation formula is: ; For example, suppose we receive an input text that reads, "According to VHF reports, at 14:30 on November 15, 2025, the 'Yuanyang 9' observed an unidentified floating object at a distance of approximately 5 nautical miles, 60 degrees to its port side." First, the model performs inference. The text is segmented into words and mapped to ID sequences before being input into the BERT-BiLSTM-CRF model. The model outputs the corresponding label sequences and performs entity segmentation based on the label boundaries. It identifies the B-TIME to I-TIME label sequences, segmenting the text into the fragment: "November 15, 2025, 14:30"; identifies the B-REF to I-REF label sequences, segmenting the text into the fragment: "Yuanyang 9"; identifies the B-BARING and other label sequences, segmenting the text into the fragment: "60 degrees to port side"; and identifies the B-DIST and other label sequences, segmenting the text into the fragment: "5 nautical miles". Next, data processing was performed. For the time entity "November 15, 2025, 14:30Z", the numbers were extracted and Z was identified as a UTC identifier, converting it to a standard timestamp 2025-11-15, 14:30:00. For the reference target's ship name entity, "Yuanyang 9", a fuzzy match was successfully found by comparing it with the database, and its MMSI code X was extracted. For the bearing, the keywords "port" and the value "60" were extracted from "port 60 degrees". According to the port negative, starboard positive rule, the formula was applied. The relative angle is calculated. The value is -60°; for distance, the numerical value "5" and the unit "nautical mile" are extracted from "5 nautical miles", and the unit conversion factor is identified. The value is 1852, applying the formula. The relative distance was calculated. Meters. The final structured feature vector is generated {Text: 2025-11-15 14:30:00, ID: X, θ} rel -60°, D rel The result quantifies the spatial geometric relationship between the target to be located and the reference vessel.

[0044] Regarding S102 above: In practical maritime intelligence processing scenarios, there is a technical challenge in the spatiotemporal alignment of data. Because AIS data transmission mechanisms are typically based on discrete time intervals—for example, a position report every 2 to 3 minutes depending on ship speed—while the event timestamps extracted from intelligence text through semantic parsing are often continuous and arbitrary time points, the two rarely coincide precisely in the time domain. If the historical AIS record closest to the event timestamp is directly used as reference data, for high-speed ships, a time difference of a few minutes can lead to position errors of several nautical miles or even more, accompanied by changes in heading angle. This cumulative deviation caused by the asynchronous pose of the dynamic intelligence reference system itself will be passed as a system error to subsequent coordinate transformation stages, affecting the positioning accuracy of intelligence targets. Therefore, step S102 constructs a continuous motion trajectory model through temporal interpolation, reconstructing the continuous state at any given moment using discrete observation data, thus achieving spatiotemporal alignment of asynchronous data.

[0045] See Figure 3 The flowchart provided in this application embodiment for generating an instantaneous motion state vector of a dynamic reference target includes steps S301 to S303, wherein: S301, based on the unique identifier of the dynamic reference target, query the AIS historical database for a time window including the event timestamp, and obtain the first AIS data point at the start time of the time window and the second AIS data point at the end time of the time window; S302, extract the latitude and longitude coordinates and heading angles of the first and second AIS data points; S303, using an interpolation algorithm, based on the time proportion of the event timestamp in the time window, calculate the reference latitude and longitude coordinates and reference heading angle of the dynamic reference target at the event timestamp, and combine them to form an instantaneous motion state vector.

[0046] In practical implementation, the spatiotemporal backtracking agent, based on the unique identifier (MMSI code) of the dynamic reference target determined in S101, accesses the AIS historical database through a database interface. This database stores massive amounts of historical ship trajectory data, and the target is quickly located using an index. Subsequently, using the extracted event timestamp as the retrieval key, a bidirectional scan is performed on the target's historical trajectory sequence to query and lock onto a time window covering that timestamp. The time window refers to the time period in the AIS historical database sequence consisting of two consecutive data frames with adjacent time spans that include the event timestamp, used to define the boundary conditions for interpolation calculations.

[0047] Specifically, first, obtain the first AIS data point at the start of the time window, denoted as... And the second AIS data point at the termination time, denoted as ,make sure Next, the corresponding latitude and longitude coordinates are extracted from these two data points. , and heading angle , .

[0048] To obtain the precise state at the moment of the event, an interpolation algorithm based on time proportion is executed. This application clarifies the specific construction process and parameter settings of the interpolation algorithm. The input data of the interpolation algorithm are the timestamps, latitude and longitude, and heading angles of the two boundary moments mentioned above. Specifically, firstly, the normalized time proportion factor of the event timestamp within the time window is calculated. The calculation formula is as follows: ; Based on this factor, for latitude and longitude coordinates, since the distance between adjacent AIS points is usually short, linear interpolation can be used to approximate the target's position. Reference latitude and longitude for the time The calculation formula is as follows: ; ; For reference heading angle The calculation of the heading difference cannot be performed by direct linear addition due to the 360-degree periodicity of the angles; minimum angle difference interpolation is required. First, the heading difference is calculated. .like If the angle is greater than 180°, then correct it. ;like If less than -180°, then correct. The revised formula for calculating the heading angle is: ; The modulo operation is used to calculate the instantaneous motion state vector. The instantaneous motion state vector refers to the set of precise motion attributes of the dynamic reference target in the physical world at a specific millisecond time. It includes at least reference latitude and longitude coordinates to determine the starting point of geodesy and reference heading angle to determine the rotation reference in the true north coordinate system.

[0049] For example, a piece of intelligence can be used to extract the event timestamp. The timeframe is 14:30:30 on October 20, 2025, with the reference target being "Yuanyang". A query for "Yuanyang" data in the AIS database revealed the two most recent records as follows: The time is 14:30:00, location ,course ; The time is 14:31:00, location ,course At this point, the time window length is 60 seconds, and the event occurs in the middle of the window. Substituting these values ​​into the formula, the time percentage factor is calculated. For latitude and longitude, the calculation yields: ; .

[0050] For the heading angle, calculate the original difference. Since -340° is less than -180°, a correction is performed. That is, rotate 20 degrees clockwise. Substitute into the formula to calculate: ; 0° represents true north. Finally, the system generates the instantaneous motion state vector for that moment as follows: Instead of directly using the 350-degree heading at 14:30:00.

[0051] Thus, by introducing a time window-based interpolation reconstruction mechanism and an angle periodic correction algorithm, this application enables the system to overcome the limitations of discrete sampling of AIS data, achieving accurate "sub-data frame level" state backtracking of dynamic reference targets at any intelligence event moment. This effectively eliminates reference origin drift and baseline deviation caused by data asynchrony, and in particular solves the problem of heading calculation error when crossing the true north direction (0 / 360 degrees). It significantly improves the geometric calculation accuracy of subsequent conversion of relative descriptions into absolute geographic coordinates, thereby optimizing the reliability of marine intelligence positioning.

[0052] Regarding the above S103: After obtaining the precise instantaneous motion state of a dynamic reference target at the moment of an event, the next step is to convert the relative geometric description based on the reference target's perspective into universal absolute coordinates in geospatial space. Traditional methods often assume the Earth is a plane or a regular sphere and directly use plane trigonometric functions for coordinate calculation. However, in long-range maritime intelligence scenarios, the projection distortion caused by the Earth's curvature leads to significant positioning errors, and the essential influence of the reference frame rotation on the azimuth angle is ignored.

[0053] Therefore, step S103 utilizes the principles of geodesy to construct a spatial geometric transformation model based on a reference ellipsoid, corrects the relative orientation to true north by rotating the coordinate system, and uses the geodetic forward calculation algorithm to accurately solve for the coordinates of the geodesic endpoint on the ellipsoid surface.

[0054] In practical implementation, the target localization agent uses the relative angle values ​​from the structured feature vector output in step S101. The value has already undergone quantization based on hull angle or clock face orientation in step S101, and has been converted into a relative distance value in standard metric units. Simultaneously, the instantaneous motion state vector output in step S102, i.e., the reference latitude and longitude, is received. With reference heading angle .

[0055] After the data enters the processing flow, the absolute geodetic azimuth is calculated first, using the reference heading angle. As a rotation reference, for relative angle values Perform a spatial reference transformation. Among these, Defined as the clockwise angle between the reference target ship's bow line and true north.

[0056] Specifically, the relative angles are superimposed onto true north, and a modulo operation is used to handle cases where the superimposed angle exceeds 360 degrees or is less than 0 degrees. For example, 350° + 20° = 370° is converted to 10° to obtain the initial geodetic azimuth angle from the reference point along the tangent to the ellipsoidal surface pointing towards the target to be located. The formula is: ; Among them, the initial geodetic azimuth angle Defined in the interval [0, 360).

[0057] Subsequently, using the Earth-themed forward calculation algorithm, combined with the above... Relative distance and reference point coordinates Solve the target coordinates To eliminate positioning errors caused by the Earth's curvature, the parameters of the reference ellipsoid model are first initialized, preferably using the WGS-84 standard. The semi-major axis *a* is set to 6,378,137 meters, the flattening *f* to 1 / 298.257223563, and the semi-minor axis *b* to *a(1-f)*. Simultaneously, the relative distance values ​​extracted in step S101 are... Convert to standard metric units, such as multiplying nautical miles by 1852 to convert to meters, and record this as the geodetic length. The solution is obtained using the Vincenty iterative algorithm or the Bessel ellipsoidal series expansion method. Through iterative approximation, the geodesic differential equations along the ellipsoidal surface between two points are solved, outputting the absolute geographic coordinates of the target. .

[0058] As an optional implementation, the Vincenty iterative algorithm is used for solution. By iteratively calculating the geodesic differential equations on the ellipsoid, millimeter-level positioning accuracy is achieved. Specifically, the normalized latitude of the reference point is first calculated. Its tangent value Based on this, we can conclude and Next, calculate the angular distance from the starting point to the equator. and the great circle azimuth at the equator parameters And thus obtain and intermediate variables : ; ; To perform iterative calculations, two fixed coefficients are pre-calculated. and , and The specific calculation formula is as follows: ; ; Based on this, angular distance is set. The initial value is Then it enters an iterative loop. In each iteration, the system calculates the auxiliary variables sequentially. and angular distance correction The specific calculation formula is as follows: ; After calculating the correction amount, use the formula Update angular distance This iterative process continues until the results of the two consecutive calculations are... When the absolute value of the difference is less than a preset convergence threshold, the loop is considered converged and the loop exits. The convergence threshold is preferably set to 10. -12 The convergence threshold, measured in radians, is determined based on the accuracy of the WGS-84 ellipsoid model. The corresponding geometric distance on the Earth's surface is approximately 0.006 millimeters. Although this is much higher than the physical measurement accuracy of AIS data, it ensures that the iterative process converges mathematically, eliminating the impact of computational truncation errors on long-distance positioning.

[0059] After iterative convergence, the system utilizes the finally determined... And auxiliary parameters, the latitude of the target to be located is calculated using the following formula: ; Simultaneously, calculate auxiliary parameters of longitude. and coefficients The longitude difference is calculated using the following formula. Add it to the starting longitude to get the final longitude. .

[0060] As an optional implementation, the Bessel ellipsoidal series expansion method is used for solution. This method is suitable for embedded environments where real-time computation requirements are not high or complex iterations are not supported. During the solution process, the first eccentricity is first calculated based on the ellipsoidal parameters. , Then, the first auxiliary function at the latitude of the reference point is calculated. This is used to characterize the local curvature features on the ellipsoid. Further calculations are made of the meridional radius at the reference point. and the radius of curvature of the y-o circle .

[0061] Utilizing the length of the earth wire and initial geodetic azimuth Latitude difference can be calculated directly using a Bessel series truncated to the third order to expand a polynomial. and longitude difference Among them, the latitude difference is calculated using the formula... Perform a quadratic polynomial calculation, and use the formula for longitude difference. The system performs calculations. After converting the calculated exact radian difference into degrees, it directly superimposes it onto the latitude and longitude of the reference point, thus outputting the absolute coordinates. and .

[0062] For example, suppose step S102 backtracks to find the reference point is located at (20°N, 110°E), and the reference heading is due north at 0°; step S101 extracts the information as "90 degrees to starboard, distance 100 nautical miles". First, parameter initialization and quantization are performed, multiplying the relative distance of 100 nautical miles by a conversion factor of 1852 to convert it into geodetic length. The distance is 185,200 meters; the reference heading of 0° is compared with the relative angle. By superimposing 90° angles, the initial geodetic azimuth angle is obtained. The angle is 90°. Initialize the WGS-84 reference ellipsoid parameters: a = 6378137 m, flattening f = 1 / 298.257223563, and minor axis b = a × (1-f). The calculated value of b is 6356752.3142 m.

[0063] The data then enters the geodetic forward calculation process. Specifically, if the Vincenty iterative algorithm is used for calculation, the latitude of the reference point is first... 20°, substitute into the formula The normalized latitude tangent of the reference point latitude is calculated. Approximately 0.3627, i.e., naturalized latitude. It is approximately 19.9383°, and further... Approximately 0.3410, It is approximately 0.9401.

[0064] Next, based on the initial geodetic azimuth... Equal to 90°, substitute The sine value of the great circle azimuth at the equator was calculated. Approximately 0.9401, the intermediate variable is obtained by combining the ellipsoidal eccentricity calculation. Approximately equal to 0.00078, and the spherical angular distance is set. The initial value is approximately 0.03 radians, and the iteration loop begins.

[0065] In each iteration, update sequentially. Approximately 0.0291 Approximately 0.9996 and auxiliary variables Approximately -0.9996, dynamically updated using a higher-order small-quantity formula. It can iteratively converge to 0.03 radians, and the angular distance correction amount... It converges to approximately 1.5 × 10⁻⁶. -7 The process continues until the difference between two consecutive calculations is less than a preset convergence threshold, such as 10 radians. -12 Stop when the iteration converges. After convergence, the final determined value will be... And auxiliary parameters are substituted into the arctangent formula to calculate the latitude of the target to be located. It is approximately equal to 19.96783°, and the longitude difference is calculated using the longitude difference formula. It is approximately equal to 1.77452°. When superimposed on the starting longitude, the final output coordinates are (19.96783°N, 111.77452°E).

[0066] Optionally, if the Bessel ellipsoid series expansion method is used for solution, the latitude difference... Approximately equal to -0.03218° and the difference in longitude. Approximately 1.77450°, this is superimposed on the reference point coordinates, resulting in the calculated output coordinates (19.96782°N, 111.77450°E). Compared to the traditional planar approximation calculation result (20°N, 111.77°E), the algorithm in this application corrects for a latitude deviation of approximately 3.5 kilometers, thereby ensuring high-precision mapping of intelligence targets in the GIS system and providing a reliable geometric basis for the subsequent sensitive area determination of S104.

[0067] Regarding S104 above: After obtaining the absolute geographic coordinates of the target to be located using the geodetic forward calculation algorithm in step S103, the problem of locating the target in physical space is solved. However, the numerical coordinates only represent a geometric point to the operator and lack semantic association with the surrounding marine environment.

[0068] Existing intelligence processing systems typically only display locations as dots on a map, failing to automatically determine whether a location involves specific maritime management rules or navigation risks. This data silo phenomenon means that subsequent intelligence compilation still heavily relies on manual verification against nautical charts, which is not only inefficient but also prone to overlooking dynamic correlations in the spatiotemporal dimension. In this application's embodiment, absolute geographic coordinates are projected onto a multi-dimensional marine GIS layer. Through the cascading of computational geometry algorithms and large language model inference, an automated analysis link is constructed from physical coordinates to tactical intent.

[0069] In specific implementation, the intent reasoning agent first receives the absolute geographic coordinates output in step S103 and the event timestamp extracted in step S101, and initializes and loads the preset multidimensional marine GIS database. This database pre-stores layer data including various functional zones, mainly including preset sensitive area layers and predetermined route layers.

[0070] First, spatial topology analysis is performed on the pre-defined sensitive areas to determine whether the target has entered the controlled sea area. Addressing the problem in existing technologies where the use of rectangular bounding boxes for coarse screening leads to a high false alarm rate for irregular sea areas, this application employs a vector boundary geometric inclusion algorithm. Specifically, vector boundary data of the pre-defined sensitive areas is first obtained. This vector boundary data refers to a polygonal geometric description composed of a series of ordered closed latitude and longitude coordinate points, used to define the geographical scope of sensitive areas such as core areas for marine ecological protection, offshore wind farm operation areas, submarine optical cable laying corridors, or fishing ban areas.

[0071] In the data processing flow, the absolute geographic coordinates of the target to be located are used as test points, and the geometric inclusion relationships of the remaining vector boundary data are calculated using a ray casting algorithm. Specifically, a virtual ray is emitted from the test point along an arbitrary fixed direction, and the total number of intersections between this ray and each boundary segment of the sensitive area polygon is calculated. If the total number of intersections is odd, according to topological principles, the test point is determined to be inside the polygon; if the total number of intersections is even, it is determined to be outside; if the test point is located on a boundary segment, it is determined to be on the boundary. Based on this calculation result, an area attribution identifier is generated to characterize whether the target to be located is inside, on, or outside the sensitive area. In this way, this application can effectively distinguish between illegal targets located inside irregular boundaries and legitimate navigation targets located outside.

[0072] Secondly, considering the dynamic nature of the marine environment, analyzing only static areas is insufficient to comprehensively assess risks. Therefore, a spatiotemporal intersection analysis is performed on the planned route. To address the technical problem of traditional methods that only compare spatial distances while ignoring the asynchronous nature of the time dimension, leading to the misclassification of unrelated vessels that passed through the location at a historical time as interfering with the current route, this application's embodiments construct a spatiotemporal dual verification mechanism.

[0073] First, the center trajectory data of the predetermined flight path is read, which consists of a series of ordered geodetic coordinate points. Composition, among which, This represents the number of discrete geodetic coordinate points that constitute the center trajectory line of the predetermined flight path. Indicates the first A trajectory point, Using the center trajectory line of the predetermined route as a reference, a preset safety distance threshold is extended to both sides along the normal direction of the trajectory line. A route buffer zone is constructed. The route buffer zone is a strip-shaped polygonal geometric region distributed along the route direction, used to define the spatial range that may pose a close-range threat or interference to navigation safety.

[0074] Specifically, for each segment on the trajectory line Calculate its normal vector , take the two endpoints of the line segment along and- Directional translation distance This yields four vertices, forming a rectangular buffer sub-unit. The union of all sub-units is then processed by a polygon merging operation to form the final strip-shaped polygonal region.

[0075] Wherein, the safe distance threshold The settings can be dynamically configured based on the type of sensitive area and maritime safety regulations. For example, for a submarine fiber optic cable / pipeline protection zone, a setting can be configured... The maximum distance is 500 meters; for the main shipping lanes of merchant ships, a maximum distance can be set. It is 2 nautical miles.

[0076] Next, it is determined whether the absolute geographic coordinates lie within the flight path buffer zone. This can be done using the ray casting method or by calculating the shortest spherical distance from the point to the line segment. As an optional implementation, this embodiment uses the Haversine distance formula to calculate the shortest spherical distance from the target point to the center trajectory line of the flight path. Specifically, set the absolute geographic coordinates of the target to be located. for The coordinates of the closest projection point or node to the target on the flight path are: ,in, Represents latitude (unit: radians). Represents longitude (unit: radians). The average radius of the Earth (taken as the standard value of 6,371,000 meters).

[0077] First, calculate the latitude difference between the two points. Difference of longitude Substitute into the semi-versus formula to calculate the intermediate variables. : ; Then calculate the spherical distance. : ; If the calculated spherical distance Less than or equal to the safe distance threshold If the spatial verification is successful, meaning the coordinates fall within the buffer polygon, then the temporal dimension verification will be initiated.

[0078] One method is to use the vector dot product projection method to calculate the coordinates of the projection point or node closest to the target on the flight path. Let the absolute geographic coordinates of the target to be located be point . Traverse all adjacent waypoints in the flight path. and The line segment vector formed For each line segment, construct a path from the waypoints. Point to target point vector ; Calculate the projection scale factor : ; like If less than or equal to 0, then the nearest point is ;like If the value is greater than or equal to 1, then the nearest point is... ;like If the value is greater than 0 and less than 1, then the nearest point is the projection point of the perpendicular on the line segment, and its coordinates can be obtained by linear interpolation, as shown in the formula: ; After traversing all line segments, select the point with the smallest geometric distance from the point as the final nearest projection point, denoted as . .

[0079] In the time dimension verification stage, the predetermined route data is first obtained through a pre-set navigation mission database. The navigation mission database establishes a mapping relationship between spatial waypoints and the time dimension, and stores the estimated arrival time of each waypoint on the route in a specific voyage mission.

[0080] Next, based on the spatial location of the target, the specific flight segment corresponding to the target is determined. Optionally, this embodiment uses the minimum projection distance method to retrieve the nearest preceding waypoint to the target from the flight route data. and subsequent waypoints ,in, This is the index of the waypoints along the route. Indicates the first One waypoint, Indicates the number of adjacent elements. There are waypoints, by to The line segment formed represents the first There are several candidate flight segments. Specifically, the process iterates through all line segments formed by adjacent waypoints along the flight path, calculates the vertical projection distance from the target point to each line segment, and selects the line segment with the smallest projection distance as the target-related flight segment. The starting endpoint of this related flight segment is the preceding waypoint, and the ending endpoint is the following waypoint.

[0081] After determining the associated flight segments, the estimated arrival times of the preceding waypoints are extracted from the database. and estimated arrival time of subsequent waypoints Using this as a reference time range, and taking into account navigation errors, a relaxation process is applied to the reference time range to construct a planned passage time window. The calculation formula is: ; ; in, The preset time tolerance can be set based on the statistical standard deviation of historical navigation data. Specifically, it can be achieved by collecting actual navigation data of similar vessels passing through this segment over the past year and calculating the set of deviations between the actual arrival time and the expected arrival time. And use the standard deviation formula to calculate the statistical standard deviation of this set of deviations. The formula is: ; in, This represents the total number of historical navigation data samples collected. The arithmetic mean of all deviation values ​​can be set in this embodiment. .

[0082] Then, verify the event timestamps in the intelligence text. Is it within the planned travel window? Within the specified time frame, a spatiotemporal intersection is determined to exist only when both spatial overlap and temporal synchronization conditions are simultaneously met—that is, the absolute geographic coordinates are within the flight path buffer zone and the event timestamp is within the planned travel time window of the predetermined flight path. A spatiotemporal intersection confirmation flag is then generated to indicate the intersection status. Conversely, if either condition is not met, a spatiotemporal intersection negation flag is generated. This effectively filters out invalid targets that are spatially close but temporally misaligned, accurately identifying immediate targets with substantial interaction risks.

[0083] After obtaining the structured topology analysis results, to address the issues of traditional rule engines generating reports that are rigid, lack logical coherence, and are difficult to explain the target's behavioral motivations, a large language model is used for intent reasoning and report generation. The data obtained in the preceding steps is assembled into a reasoning prompt text, which includes the absolute geographic coordinates of the target to be located, the original event timestamp, the spatial topology analysis results calculated above, and a spatiotemporal intersection confirmation identifier.

[0084] Subsequently, the inference prompt text is input into a large language model that has been fine-tuned with marine domain knowledge. Specifically, a marine intelligence instruction fine-tuning dataset is first constructed. Historical accident investigation reports are crawled from the official website of the Maritime Safety Administration, abnormal trajectory data is extracted from the public AIS database, and a marine domain knowledge base is introduced as a source of rule samples. Unstructured documents such as navigation prohibition regulations for various sea areas are digitized and processed in blocks. During the data processing stage, regular expressions are used to clean HTML tags and garbled characters, and unstructured text is converted into a unified JSON format. The dataset is organized according to the instruction, input, and output structure. For example, the instruction is "Analyze the ship's tactical intentions," the input is "Time: 02:00, Coordinates: Within the fiber optic cable area, Topology status: Spatiotemporal intersection confirmed," and the output is "The target is lingering at low speed in the sensitive fiber optic cable area at night and is within the scheduled maintenance window, which is determined to be illegal reconnaissance or sabotage operations." The dataset contains 50,000 samples, which are randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0085] The model can be a decoder-only model using the Transformer architecture, including 32 Transformer decoding layers with a hidden layer dimension of 4096. The input information prompt text is first segmented by a tokenizer and mapped to a TokenID sequence; then it enters the Embedding layer to be converted into a word vector matrix. The data flows sequentially through each Transformer decoding layer. In each layer, it is first normalized using Root Mean Square Normalization (RMSNorm) to stabilize the gradient; then it enters the multi-head self-attention module, where the input vector is linearly projected to generate Query(Q), Key(K), and Value(V) matrices. To capture temporal and positional relationships, Rotated Position Encoding (RoPE) is applied to Q and K, injecting positional information into the vector through a rotation transformation in the complex domain. Next, the attention score is calculated using the formula: ; in, The feature dimension represents each attention head.

[0086] This process captures the long-distance dependency and causal logic weights between spatiotemporal intersection identifiers and intended conclusions. The output of the attention layer, after residual connections and a second RMSnorm, enters the feedforward neural network SwiGLU. This network employs a gated linear unit structure, enhancing the model's semantic representation capability through nonlinear transformations.

[0087] During model fine-tuning training, LoRA technique was used for efficient parameter tuning. The weights of the fully connected layers of the base model were frozen during training, and low-rank matrices were introduced and trained only in the Query and Value projection layers, with a rank of 8 and a scaling factor of 16. The training optimizer used was AdamW, with a learning rate of 2×10⁻⁶. -4 The batch size was set to 16. The cross-entropy loss function was used. After 100 epochs of training, training was stopped when the validation set loss value converged to below 0.2, and the LoRA adapter weights were saved, thus obtaining the final model with marine semantic understanding capabilities.

[0088] In the application phase, the absolute geographic coordinates of the target to be located, the event timestamp, and the previously generated spatiotemporal intersection confirmation identifier are assembled into a reasoning prompt text input model. Next, system-level instructions are added to the prompt text, directing the model to invoke its internalized knowledge parameters, such as: "As a marine intelligence expert, please analyze the target's intent using your knowledge of the marine domain." After receiving this prompt text, the model combines the regulatory knowledge learned during training with the input spatiotemporal state to output a natural language description representing the target's navigation intent or operational purpose.

[0089] Finally, the event timestamp, absolute geographic coordinates, detailed spatial topology analysis conclusions, and natural language description text generated by the large model are automatically filled into the corresponding fields of the preset intelligence report structure template to generate a complete intelligence report.

[0090] For example, the preset intelligence report structure template can be a data structure built based on the JSON Schema standard, including four core parts: metadata header, target status, topological conclusions, and intent understanding. Specifically, the event timestamp extracted in step S101 is written into the Timestamp field; the absolute geographic coordinates calculated in step S103 are written into the Location field; the spatiotemporal intersection identifier generated in step S104 is filled into the corresponding key-value pair under the Metrics module; and the natural language description text output by the large language model is written into the Intent field. After completion, it is converted into a standard JSON format file as the final generated complete intelligence report.

[0091] In this way, by combining geometric topology calculation with large model semantic reasoning, this application not only achieves accurate and automated identification of targets intruding into sensitive areas and approaching shipping routes, but also interprets the tactical intentions of targets from the data, thereby improving the depth and timeliness of marine intelligence analysis.

[0092] See Figure 4 The diagram illustrates a multi-agent intelligence intent understanding and report generation system provided in this application embodiment, including: The semantic parsing intelligent agent module 10 is used to acquire marine intelligence text, identify and extract event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located; the relative spatial description parameters include relative orientation description and relative distance value based on the dynamic reference target body coordinate system; The spatiotemporal backtracking intelligent agent module 20 is used to retrieve the instantaneous motion state vector of the target at the corresponding moment in the AIS historical database based on the dynamic reference target and the event timestamp; the instantaneous motion state vector includes reference latitude and longitude coordinates and reference heading angle; The target localization intelligent agent module 30 is used to convert the relative orientation description into the absolute geodetic azimuth angle in the true north coordinate system based on the reference heading angle; and to calculate the absolute geographic coordinates of the target to be located by using the geodetic forward calculation algorithm, combined with the absolute geodetic azimuth angle and the relative distance value, starting from the reference latitude and longitude coordinates. The intent reasoning intelligent agent module 40 is used to project the absolute geographic coordinates onto a multi-dimensional marine GIS layer, perform spatial topology analysis, determine whether the target to be located falls into a preset sensitive area or whether it has a spatiotemporal intersection with a predetermined route; based on the spatial topology analysis results, it uses a large language model to reason about the intent of the target to be located and generates an intelligence report.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for understanding intelligence intent and generating reports based on multi-agent intelligence, characterized in that, include: Acquire marine intelligence text, identify and extract event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located; The relative spatial description parameters include relative orientation description and relative distance values ​​based on the dynamic reference target body coordinate system; Based on the dynamic reference target and the event timestamp, the instantaneous motion state vector of the target at the corresponding moment is retrieved from the AIS historical database; the instantaneous motion state vector includes reference latitude and longitude coordinates and reference heading angle; Based on the reference heading angle, the relative azimuth description is converted into the absolute geodetic azimuth angle in the true north coordinate system; Starting from the reference latitude and longitude coordinates, the absolute geographic coordinates of the target to be located are calculated using the geodetic forward calculation algorithm, combined with the absolute geodetic azimuth and the relative distance value. The absolute geographic coordinates are projected onto a multidimensional marine GIS layer, and spatial topology analysis is performed to determine whether the target to be located falls into a preset sensitive area or whether it has a spatiotemporal intersection with a predetermined route. Based on the spatial topology analysis results, the intent of the target to be located is inferred using a large language model, and an intelligence report is generated.

2. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 1, characterized in that, The process of identifying and extracting event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located includes: The marine intelligence text is input into a pre-trained entity extraction model to identify time entities, ship name entities, and bearing distance descriptors. The time entity is preprocessed and converted into a unified timestamp format as an event timestamp; The vessel name entity is matched with preset vessel file data to determine the unique identifier of the dynamic reference target; Syntactic dependency analysis is performed on the directional distance descriptors to separate the directional words representing direction and the numerical values ​​representing distance, which are used as relative directional descriptions and relative distance values, respectively.

3. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 2, characterized in that, The relative orientation description includes a fuzzy description based on clock face orientation or a precise description based on hull angle; In response to the relative orientation description being a fuzzy description based on clock face orientation, a mapping relationship between clock face dots and angles is established, and the extracted clock face dots are converted into relative angle values ​​relative to the dynamic reference target ship's bow baseline. In response to the relative bearing description being a precise description based on the hull angle, the port and starboard attributes in the bearing term are identified, and a sign transformation is performed on the extracted angle value based on the port and starboard attributes to obtain the relative angle value relative to the bow baseline of the dynamic reference target ship.

4. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 2, characterized in that, The step of retrieving the instantaneous motion state vector of the target at the corresponding moment in the AIS historical database based on the dynamic reference target and the event timestamp includes: Based on the unique identifier of the dynamic reference target, query the AIS historical database for a time window including the event timestamp, and obtain the first AIS data point at the start time of the time window and the second AIS data point at the end time of the time window. Extract the latitude and longitude coordinates and heading angles of the first and second AIS data points; Using an interpolation algorithm, the reference latitude and longitude coordinates and reference heading angle of the dynamic reference target at the event timestamp are calculated based on the time proportion of the event timestamp within the time window, and combined to form an instantaneous motion state vector.

5. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 3, characterized in that, The process of converting the relative bearing description into an absolute geodetic azimuth angle in the true north coordinate system based on the reference heading angle includes: Obtain the relative angle value corresponding to the relative orientation description; The reference heading angle and the relative angle value are summed and calculated. The summation result is normalized based on the period of the circular angle to obtain the absolute geodetic azimuth relative to true north.

6. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 1, characterized in that, The step of using the geodetic forward calculation algorithm, combining the absolute geodetic azimuth and the relative distance value, to calculate the absolute geographic coordinates of the target to be located includes: Initialize the parameters of the reference ellipsoid model, including the semi-major axis and flattening of the ellipsoid; Using the reference latitude and longitude coordinates as the starting point of the geodetic line, the absolute geodetic azimuth angle as the initial geodetic azimuth angle, and the relative distance value as the length of the geodetic line; Based on the parameters of the reference ellipsoid model, the geodetic equations are solved using an iterative approximation method or an ellipsoidal series expansion method to obtain the latitude and longitude of the geodetic endpoint, which are then used as the absolute geographic coordinates of the target to be located.

7. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 1, characterized in that, The execution of spatial topology analysis to determine whether the target to be located falls within a preset sensitive area includes: Obtain vector boundary data of the preset sensitive area; Calculate the geometric inclusion relationship between the absolute geographic coordinates and the vector boundary data to generate a region attribution identifier; the region attribution identifier is used to characterize whether the target to be located is located inside, on the boundary, or outside the sensitive area.

8. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 1, characterized in that, The execution of spatial topology analysis to determine whether the target to be located has spatiotemporal intersection with the predetermined flight path includes: Based on the center trajectory line of the predetermined route, a preset safety distance threshold is extended to both sides to generate a route buffer zone; Calculate the spatial overlap between the absolute geographic coordinates and the route buffer, and verify whether the event timestamp is within the planned travel time window of the predetermined route; In response to the absolute geographic coordinates being located within the route buffer and the event timestamp being within the planned travel time window of the predetermined route, a spatiotemporal intersection is determined to exist, and a spatiotemporal intersection confirmation identifier is generated. In response to the absolute geographic coordinates being outside the route buffer or the event timestamp being outside the planned travel time window of the predetermined route, a spatiotemporal intersection negation flag is generated.

9. The method for understanding intelligence intent and generating reports based on multi-agent intelligence according to claim 1, characterized in that, The process of inferring the intent of the target to be located using a large language model based on spatial topology analysis results and generating an intelligence report includes: Construct inference prompt text, which includes the absolute geographic coordinates of the target to be located, the event timestamp, and the spatial topology analysis results; The inference prompt text is input into the large language model, which combines the marine domain knowledge base to analyze the behavioral semantics of the target to be located in the spatiotemporal context and outputs a natural language description text representing the navigation intention or operation purpose of the target to be located. According to the preset intelligence report structure, the event timestamp, the absolute geographic coordinates, the spatial topology analysis results, and the natural language description text are filled into the corresponding fields to generate an intelligence report.

10. A multi-agent intelligence intent understanding and report generation system, used to implement the multi-agent intelligence intent understanding and report generation method according to any one of claims 1-9, characterized in that, include: The semantic parsing intelligent agent module is used to acquire marine intelligence text, identify and extract event timestamps, dynamic reference targets, and relative spatial description parameters of the target to be located; the relative spatial description parameters include relative orientation description and relative distance value based on the dynamic reference target's ontological coordinate system; The spatiotemporal backtracking intelligent agent module is used to retrieve the instantaneous motion state vector of the target at the corresponding moment in the AIS historical database based on the dynamic reference target and the event timestamp; the instantaneous motion state vector includes reference latitude and longitude coordinates and reference heading angle; The target localization agent module is used to convert the relative orientation description into an absolute geodetic azimuth angle in the true north coordinate system based on the reference heading angle; and to calculate the absolute geographic coordinates of the target to be located by using the geodetic forward calculation algorithm, combined with the absolute geodetic azimuth angle and the relative distance value, starting from the reference latitude and longitude coordinates. The intent reasoning agent module is used to project the absolute geographic coordinates onto a multi-dimensional marine GIS layer, perform spatial topology analysis, and determine whether the target to be located falls into a preset sensitive area or whether it has a spatiotemporal intersection with a predetermined route. Based on the spatial topology analysis results, the intent of the target to be located is inferred using a large language model, and an intelligence report is generated.

Images