Multi-agent task routing and research and judgment method for abnormal situation at sea

Abstract

The application discloses a kind of multi-agent task routing and abnormal situation of sea research and judgment method, comprising: receiving control task request and semantic analysis generation structured input package;Task routing activation agent based on input package;By first agent, execute multidimensional representation abnormal quantification analysis and output abnormal determination result;By second agent, execute deep qualitative inference and output qualitative conclusion;Compare and judge to both;In response to the direction inconsistency triggers second agent supplementary data re-reasoning;In response to the judgment through, execute compliance review and assemble output research and judgment report.The application can realize the multi-dimensional accurate identification of abnormal situation of sea, depth logic reasoning and compliance report automatic generation.

Description

Technical Field

[0001] This invention relates to the fields of public security informatization and artificial intelligence technology, and in particular to a method for multi-agent task routing and judgment of abnormal situations at sea. Background Technology

[0002] In the fields of port shoreline management and coastal defense monitoring, the identification and assessment of illegal activities such as illegal fishing are core tasks. Current technologies primarily rely on single-rule triggers or manual cross-system queries for anomaly assessment, which presents the following prominent problems: First, in automated assessment processes, anomaly detection based on superficial data and in-depth qualitative analysis based on business logic are often disconnected. When the conclusions of the two are inconsistent, there is a lack of effective feedback and correction mechanisms, leading to high false alarm rates or lost leads. Second, traditional single-dimensional identification models struggle to penetrate collaborative networks when facing new countermeasures such as AIS signal spoofing, segmented shutdowns, and multiple vessels providing cover for each other. Finally, the process from anomaly detection to the formation of a compliant evidence chain is cumbersome and lacks fully automated closed-loop processing capabilities. Therefore, there is an urgent need for an intelligent assessment method that integrates multi-dimensional perception and deep reasoning, possesses conflict self-correction capabilities, and supports collaborative group analysis. Summary of the Invention

[0003] The purpose of this invention is to provide a method for multi-agent task routing and assessment of abnormal maritime situations. This invention can achieve multi-dimensional and accurate identification of abnormal maritime situations, deep logical reasoning, and automatic generation of compliance reports.

[0004] The technical solution of this invention: a method for multi-agent task routing and judgment of abnormal maritime situations, comprising the following steps: Step S1: Receive and perform semantic parsing of port shoreline control task requests, extract task feature information to generate structured input packets; Step S2: Based on the structured input packet, perform task routing, determine and activate the agents participating in the processing; Step S3: The first intelligent agent performs multi-dimensional quantitative analysis of the appearance anomalies of the target object based on the structured input package, and outputs the anomaly judgment result. Step S4: The second intelligent agent performs deep qualitative reasoning based on the anomaly determination result and outputs a qualitative conclusion; Step S5: Compare and judge the anomaly determination result with the qualitative conclusion; Step S6: In response to the inconsistency between the direction of the anomaly determination result and the qualitative conclusion, the second agent is triggered to supplement the query of related data and then re-execute the deep qualitative reasoning to update the qualitative conclusion; Step S7: In response to the comparison judge's approval, perform a compliance review on the updated or consistent qualitative conclusion, and assemble and output an analysis report.

[0005] The above method, specifically includes the following: The first intelligent agent sequentially performs anomaly identification on a single ship target from four dimensions: statistical anomaly, temporal behavior recognition, tactical form matching, and identity and behavior consistency. Specifically, the statistical anomaly is used to identify unknown abnormal behaviors lacking prior labels; the temporal behavior recognition is used to identify abnormal behavior patterns with time-dependent characteristics; the tactical morphology matching is used to identify morphology matching anomalies between real-time trajectories and preset tactical templates; and the identity and behavior consistency is used to detect conflicting relationships between target identity attributes and their actual behaviors.

[0006] The aforementioned method, in which the first agent performs anomaly identification on a single ship target from the dimension of statistical anomaly, employs the Isolation Forest algorithm to perform unsupervised mining of the behavioral features of the single ship, outputting anomaly alienation scores, including: Using the standard deviation of airspeed, rate of change of heading, and variance of acceleration as inputs, a randomized isolated tree is constructed in memory to calculate the average path length for each target; The abnormal alienation score of the Isolation Forest algorithm is defined as follows: ; In the formula: This is a sample of single-ship behavioral characteristics; The total number of samples; This represents the path length of the sample in the isolated tree. For the sample Average path length in the forest; This is a path length normalization correction term; Scoring for sample anomalies; The path length normalization correction term is: ; In the formula: It is a harmonic function; When the abnormal alienation score exceeds the threshold, it is judged as an abnormal sample: ; In the formula: This is the initially set threshold for anomaly detection.

[0007] The aforementioned method, in which the first intelligent agent performs anomaly recognition on a single ship target from the perspective of temporal behavior recognition, employs a long short-term memory network to model multi-dimensional behavior sequences. Let the input behavior feature sequence be... The calculations for each gate are as follows: ; ; ; ; ; ; The final output behavior category probability is: ; In the formula: Output for the forget gate. For input gate output, Candidate memory states, This is a unit memory state. For output gate output, Let be the hidden state vector at the current moment. This is the hidden state from the previous moment. It is the Sigmoid activation function. It is the hyperbolic tangent function. This is Hadamard's element-wise multiplication; Both represent model parameter matrices. Both represent bias terms; For a moment The probability distribution of behavioral categories It is a normalized exponential function.

[0008] The aforementioned method, in which the first intelligent agent performs anomaly identification on a single ship target from the tactical form matching dimension, employs a dynamic time warping method for non-uniform length sequence alignment. Let the trajectory sequence to be detected be... The template trajectory sequence is The dynamic time-warped cumulative distance is recursively calculated as follows: ; Its normalized similarity is expressed as: ; In the formula: The normalized similarity score. The global minimum normalized distance between the two trajectories is given. To normalize the path length, This is the distance-normalized scale parameter; They are the first in the sequence The and the first A trajectory point, The local distance between two points. To reach the grid point The minimum cumulative distance; The optimal template matching result is defined as: ; In the formula: This is the optimal matching template number.

[0009] The aforementioned method, in which the first agent performs anomaly identification on a single ship target from the dimension of identity and behavior consistency, uses Mahalanobis distance to statistically test the consistency between identity and behavior, defining the target behavior feature vector as: ; In the formula: For the goal at all times behavioral feature vectors, Indicates average speed. For the maximum permissible speed, Indicates the duration of stay. This indicates the percentage of activities in sensitive areas. Indicates the number of times the device has been stopped and started; For identity categories The standard behavior distribution, whose squared Mahalanobis distance is defined as: ; In the formula: Indicates the target behavior relative to the identity category The squared Mahalanobis distance, For identity categories The behavior is the mean vector. For identity categories The behavioral covariance matrix; when the distance exceeds a threshold, it is determined to be a conflict between identity and behavior: ; In the formula: This is the preset distance threshold parameter.

[0010] In the aforementioned method, the first intelligent agent also performs collaborative analysis on the multi-ship group, the collaborative analysis including two levels: long-term collaborative identification and short-term synchronization identification; The long-term collaborative identification includes: within a continuously preset long time window, calculating the similarity between any two ships in three aspects: overall trajectory outline, speed habits, and synchronization of action sequence, and then constructing a global ship similarity matrix to discover hidden gang structures; The short-term synchronization identification includes: based on the identification of a suspected gang structure in the long-term collaborative identification, within a preset short time window, the degree of synchronization of multiple ships experiencing the same abnormal event is counted, and when the degree of synchronization exceeds a preset synchronization threshold, it is determined to be an organized collaborative violation.

[0011] In the aforementioned method, the similarity between any two ships in the long-term collaborative identification is defined as: ; In the formula: These are the weighting coefficients for spatial distance, speed difference, and heading difference, respectively, and they satisfy... ; The Frescher distance between the two trajectories within the time window. , and These are the scale parameters for spatial distance, speed difference, and heading difference, respectively; It measures the difference in velocity magnitude or average velocity. This measures the deviation of the heading angle or the average heading difference. Further construct the spatiotemporal similarity matrix: ; In the formula: This indicates the size at a specific time step or state. The matrix, Indicates a ship With ships Spatiotemporal comprehensive similarity, This indicates the number of ships included in the calculation within the window.

[0012] The aforementioned method, wherein the step of performing deep qualitative reasoning based on the anomaly determination result by a second intelligent agent and outputting a qualitative conclusion, includes: Trajectory tracing: Starting from the anomaly determination result, examine each of the marked abnormal evidence in the multi-dimensional quantitative analysis of apparent anomalies, determine whether the trajectory anomaly has a subjective intent to resist, and output the trajectory evidence score and the corresponding supporting fact description. Uncover Relationships: Call the ship relationship graph interface to verify the relationship between the target ship and other ships, personnel, and locations, and output the relationship score and a description of the relationship facts; Conclusion: By combining trajectory evidence, correlation evidence, and similarity to historical cases, a qualitative conclusion is formed and a similarity score to historical cases is output; The credibility is calculated based on the trajectory evidence score, the correlation score, and the historical case similarity score: ; In the formula: To assess the credibility of the conclusions or the overall evaluation value, These are the weight coefficients for trajectory features, correlation features, and historical features, respectively, and their sum is 1. This refers to the similarity score or trajectory evidence score calculated based on the motion trajectory. The similarity or association score is calculated based on spatiotemporal relationships. This refers to the similarity score calculated based on historical behavioral patterns or historical case similarity scores.

[0013] The aforementioned method, specifically the compliance review of the updated or consistent qualitative conclusions, includes performing the following three reviews sequentially: Isolated evidence test: Determine whether the conclusion is supported by evidence from a single source. If so, mark the evidence as isolated evidence and indicate that it cannot be used to make a judgment on its own. Language review: Detects and assesses whether the text contains speculative expressions, and automatically replaces them with standard law enforcement language if found; Legal provision matching: Based on the type of illegal behavior indicated by the assessment conclusion, the applicable legal provisions are retrieved from the legal knowledge base and attached to the assessment report.

[0014] Compared to existing technologies, this invention decouples maritime anomaly situation assessment into two independent yet interconnected stages: quantification of apparent anomalies and deep qualitative reasoning, by constructing a collaborative architecture between a first and a second intelligent agent. The first agent focuses on objectively quantifying anomaly characteristics from multiple dimensions, including statistics, time series, morphology, and identity consistency, avoiding interference from subjective inferences in the perception of raw data. The second agent, based on this, performs deep logical reasoning by combining archives and case studies. When the two conclusions are inconsistent, the system does not simply discard one and discard the other, but triggers a feedback adjudication mechanism where the second agent supplements the query with related data and re-reasons. This mechanism leverages the complementarity between apparent data and deep knowledge, eliminating cognitive biases from a single perspective through iterative verification, and significantly improving the accuracy and robustness of assessment conclusions in complex adversarial scenarios.

[0015] Furthermore, this invention integrates single-ship multidimensional orthogonal recognition with multi-ship spatiotemporal collaborative analysis technology. At the single-ship level, it comprehensively utilizes algorithms such as Isolation Forest, LSTM, dynamic time warping, and Mahalanobis distance to achieve full-coverage recognition of unknown anomalies, temporal patterns, tactical forms, and identity conflicts. At the multi-ship level, it effectively judges collaborative violations through long-term trajectory similarity matrix and short-term event synchronization analysis. Combined with the automatic isolated evidence verification, terminology standardization, and legal provision matching functions of the compliance review agent, it achieves full-process automation from natural language command input to the output of a judgment report conforming to public security document standards, significantly reducing the cost of manual cross-system operations and improving the response speed and law enforcement standardization level of maritime defense control. Attached Figure Description

[0016] Figure 1 This is a flowchart of the processing of the present invention.

[0017] Figure 2 This is a flowchart of the first intelligent agent of the present invention.

[0018] Figure 3This is a flowchart of the operation of the second intelligent agent of the present invention.

[0019] Figure 4 This is a flowchart illustrating the compliance review process for this invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0021] Example 1: This example provides a method for multi-agent task routing and assessment of abnormal maritime situations. This method decouples the assessment of abnormal maritime situations into two stages: superficial anomaly quantification and deep qualitative reasoning. When the conclusions of the two stages are inconsistent, a feedback reassessment mechanism is triggered, ultimately outputting an assessment report that conforms to law enforcement standards. Specifically, as follows... Figure 1 As shown, the method includes the following steps: Step S1: Receive and perform semantic parsing of port shoreline control task requests, extract task feature information to generate structured input packets.

[0022] In this embodiment, the system receives natural language commands input by police officers, such as "Check the high-powered boat with the prefix 'Zhepuyu' that was lingering slowly near Dongjiao last night." The semantic parsing module first classifies the intent, identifying the command as a "single-vehicle trajectory check" type with a confidence level of 0.92. Then, it extracts key entity information, locking the target vessel as one with the prefix "Zhepuyu" and the characteristic "high-powered." Simultaneously, it performs spatiotemporal normalization, estimating the colloquial "last night" to specific start and end times (e.g., 00:00 to 04:00 on October 25, 2025), and mapping the relative location "near Dongjiao" to center coordinates (122.34°E, 29.87°N) and a preset search radius (e.g., 3 nautical miles). The parsing results are encapsulated into a structured input package containing task type, target identifier, time window, and spatial range, serving as a unified data foundation for subsequent agent scheduling. It should be understood that although this example uses single-ship trajectory verification, in other embodiments, semantic parsing can also support various intentions such as group collaborative verification or resume file verification, as long as the corresponding structured elements can be extracted from natural language.

[0023] Step S2: Based on the structured input packet, perform task routing, determine and activate the intelligent agents participating in the processing.

[0024] The task routing engine automatically matches and activates the corresponding agent combination based on the task type field in the structured input package. In this embodiment, since the task type is "single ship trajectory verification," the routing engine activates the first agent (i.e., the situational awareness agent) to perform anomaly identification on the target ship. If the task type is "group collaborative verification," then both the first agent and the correlation analysis module are activated simultaneously; if the credibility output of the subsequent second agent (i.e., the judgment and analysis agent) is higher than a preset threshold (e.g., 0.70), compliance review will also be forcibly activated. This dynamic routing mechanism avoids the resource waste caused by the constant operation of all agents and achieves on-demand scheduling.

[0025] Step S3: The first intelligent agent performs multi-dimensional quantitative analysis of the appearance anomalies of the target object based on the structured input package, and outputs the anomaly judgment result.

[0026] The first intelligent agent focuses on discovering anomalies from the perspective of objective data, without relying on subjective business rules. Its analysis process covers four dimensions in sequence to ensure comprehensive coverage of various adversarial techniques.

[0027] Specifically, such as Figure 2 As shown, the first intelligent agent sequentially performs anomaly identification on a single ship target from four dimensions: statistical anomaly, temporal behavior recognition, tactical form matching, and identity and behavior consistency.

[0028] In terms of statistical anomaly, the method for anomaly identification of a single ship target is to use the isolated forest algorithm to perform unsupervised mining of the behavioral features of the single ship and output an anomaly alienation score.

[0029] This dimension aims to identify unknown anomalous behaviors lacking prior labels, such as AIS data falsification, multi-vessel relay cover, deliberate detours to avoid checkpoints, and long-term lurking in radar blind spots. The algorithm uses the standard deviation of speed, rate of change of heading, and variance of acceleration as input features, constructs a randomized isolated tree in memory, and calculates the average path length for each target.

[0030] The abnormal alienation score of the Isolation Forest algorithm is defined as follows: ; The path length normalization correction term is: ; In the formula: It is a harmonic function; When the abnormal alienation score exceeds the threshold, it is judged as an abnormal sample: ; In the formula: This is the initially set threshold for anomaly detection.

[0031] In this example, the trajectory points of the target vessel "Zhepuyu 00128" overlapped with the radar blind zone boundary for more than 30 minutes, resulting in an anomaly alienation score of 0.82, which is higher than the threshold of 0.70. This successfully identified the concealed behavior of "lurking in the radar blind zone." It should be understood that although this embodiment preferably uses the Isolation Forest algorithm due to its robustness to high-dimensional sparse data and the lack of labeled samples, in other implementations, it can be replaced with unsupervised anomaly detection algorithms such as One-Class SVM or Autoencoder, as long as anomaly quantification based on feature space alienation can be achieved.

[0032] In the temporal behavior recognition dimension, the method for performing anomaly recognition on a single ship target is to use a Long Short-Term Memory (LSTM) network to model multi-dimensional behavior sequences.

[0033] This dimension is used to identify anomalous behavioral patterns with time-dependent characteristics, such as reconnaissance, emergency escape, camouflage operations, and covert landings. Let the input behavioral feature sequence be... The calculations for each gate are as follows: ; ; ; ; ; ; The final output behavior category probability is: ; In the formula: Output for the forget gate. For input gate output, Candidate memory states, This is a unit memory state. For output gate output, Let be the hidden state vector at the current moment. This is the hidden state from the previous moment. It is the Sigmoid activation function. It is the hyperbolic tangent function. This is Hadamard's element-wise multiplication; Both represent model parameter matrices. Both represent bias terms; For a moment The probability distribution of behavioral categories It is a normalized exponential function.

[0034] In this case, the LSTM model analyzed the ship's continuous 5-minute trajectory sequence and output a probability of 0.88 for the "reconnaissance mode," effectively filtering out random fluctuations caused by wind and waves and identifying temporal actions with a subjective intention to evade. Although this embodiment uses LSTM, other temporal models such as GRU or Transformer can also be used in other embodiments. The advantage of LSTM is that its gating mechanism can more stably maintain key state information when dealing with long-range dependent and noisy sequences such as sea trajectories.

[0035] In the tactical form matching dimension, the method for anomaly identification of single-ship targets is to use dynamic time warping to perform non-equal length sequence alignment.

[0036] This dimension is used to identify anomalies in the morphological matching between real-time trajectories and preset tactical templates, such as four patterns: zigzag, S-shaped evasion, shore-hugging concealed navigation, paperclip-like turning back to await resupply, and fixed-point turning. This is used to address the misalignment between real-time trajectories and templates on the time axis. Let the sequence of trajectories to be detected be... The template trajectory sequence is The cumulative distance for Dynamic Time Warping (DTW) is recursively calculated as follows: ; Its normalized similarity is expressed as: ; In the formula: The normalized similarity score. The global minimum normalized distance between the two trajectories is given. To normalize the path length, This is the distance-normalized scale parameter; They are the first in the sequence The and the first A trajectory point, The local distance between two points. To reach the grid point The minimum cumulative distance; The optimal template matching result is defined as: ; In the formula: This is the optimal matching template number.

[0037] In this case, the target vessel's trajectory matches the "paperclip-like turnaround waiting for loading" template. =0.93, which is higher than the threshold of 0.90, accurately identifying its tactical maneuver of circling back and forth near the port while waiting for loading. Although this embodiment uses DTW, other embodiments may also use resilient measurement methods such as LCSS (Longest Common Subsequence) or EDR (Edit Distance Real Number Sequence). The advantage of DTW in maritime scenarios is that it has better tolerance for inconsistent trajectory sampling rates and local deformations.

[0038] In terms of identity and behavior consistency, the method for anomaly identification of single-ship targets is to use Mahalanobis distance to statistically test the consistency between identity and behavior.

[0039] This dimension is used to detect conflicts between a target's identity attributes and its actual behavior, such as large-scale aircraft camouflage or providing supplies to the mothership. The target behavior feature vector is defined as follows: ; In the formula: For the goal at all times behavioral feature vectors, Indicates average speed. For the maximum permissible speed, Indicates the duration of stay. This indicates the percentage of activities in sensitive areas. Indicates the number of times the device has been stopped and started; For identity categories The standard behavior distribution, whose squared Mahalanobis distance is defined as: ; In the formula: Indicates the target behavior relative to the identity category The squared Mahalanobis distance, For identity categories The behavior is the mean vector. For identity categories The behavioral covariance matrix; when the distance exceeds a threshold, it is determined to be a conflict between identity and behavior: ; In the formula: This is the preset distance threshold parameter.

[0040] In this case, the vessel was registered as a fishing boat but had a maximum speed of 18 knots (far exceeding the upper limit for similar fishing boats) and stayed in a non-open port for 92 minutes. The calculated Mahalanobis distance squared was 38.7, which is much higher than the threshold of 9.49, thus determining that its "identity and behavior do not match". The advantage of Mahalanobis distance over Euclidean distance is that it considers the correlation and dimensional differences between features, making it more suitable for conflict testing of multidimensional heterogeneous behavioral features.

[0041] In addition, the first intelligent agent also performs collaborative analysis on multi-ship groups, which includes two levels: long-term collaborative identification and short-term synchronization identification.

[0042] The long-term collaborative identification includes: within a continuously preset long time window (such as 7 days), calculating the similarity between any two ships in three aspects: overall trajectory outline, speed habits, and synchronization of action sequence, and then constructing a global ship similarity matrix to discover hidden gang structures.

[0043] The short-term synchronization identification includes: based on the identification of a suspected gang structure in the long-term collaborative identification, within a preset short time window (such as 1 minute), the synchronization degree of multiple ships experiencing the same abnormal event is counted. When the synchronization degree exceeds a preset synchronization threshold, it is determined to be an organized collaborative violation, such as collective shutdown or scattering and escaping.

[0044] Specifically, in the long-term collaborative identification, the similarity between any two ships is defined as: ; In the formula: These are the weighting coefficients for spatial distance, speed difference, and heading difference, respectively, and the sum of the three is 1. The Frescher distance between the two trajectories within the time window. , and These are the scale parameters for spatial distance, speed difference, and heading difference, respectively; It measures the difference in velocity magnitude or average velocity. This measures the deviation of the heading angle or the average heading difference. Further construct the spatiotemporal similarity matrix: ; In the formula: This indicates the size at a specific time step or state. The matrix, Indicates a ship With ships Spatiotemporal comprehensive similarity, This indicates the number of ships included in the calculation within the window.

[0045] This matrix can automatically detect hidden gang structures such as unified command and escort. It should be understood that although this embodiment uses the Gaussian kernel function to calculate similarity, other embodiments may also use cosine similarity or Pearson correlation coefficient. The advantage of the Gaussian kernel is that its nonlinear mapping of distance is more in line with the physical attenuation characteristics of maritime ship interactions.

[0046] Step S4: The second intelligent agent performs deep qualitative reasoning based on the anomaly determination result and outputs a qualitative conclusion.

[0047] The second intelligent agent (i.e., the analytical agent) is responsible for interpreting anomalies and determining illegality. Its reasoning process includes: tracing patterns, identifying correlations, and drawing conclusions. For example... Figure 3 As shown, specifically, trajectory investigation starts with the anomaly determination result, examines each marked abnormal evidence in the multi-dimensional quantitative analysis of apparent anomalies, determines whether the trajectory anomaly has a subjective intent to resist, and outputs a trajectory evidence score. And the corresponding supporting factual descriptions; correlation analysis involves calling the ship relationship graph interface to verify the relationship between the target ship and other ships, personnel, and locations, and outputting a correlation score. The conclusion is drawn by comprehensively considering trajectory evidence, related evidence, and similarity to historical cases, forming a qualitative conclusion, and outputting a similarity score to historical cases. .

[0048] In this embodiment, the quantitative results of the first intelligent agent in four dimensions—statistical anomaly, temporal behavior, tactical form, and identity consistency—are reviewed item by item to determine whether anomalies in each dimension indicate subjective adversarial intent. The trajectory evidence score is calculated by weighting the subjective intent and is defined as follows: ; in, The values ​​are, in order: the isolated forest anomaly alienation score of the statistical anomaly dimension, the LSTM anomaly category probability of the temporal behavior dimension, the DTW normalized matching degree of the tactical form dimension, and the identity deviation degree after Mahalanobis distance squared normalization of the identity consistency dimension, all of which take values ​​in [0,1]. The coefficients for subjective confrontational intent are in four dimensions, ranging from [0,1]. A value close to 1 is taken when a certain dimension abnormally exhibits clear characteristics of evading law enforcement, and a value close to 0 is taken when it can be explained by objective factors such as sea conditions, equipment failure, or normal operations.

[0049] In this case, the first agent's four-dimensional output is: Isolation Forest Anomaly Alienation Score. =0.82, probability of timing-based behavior reconnaissance mode =0.88, Tactical Form Paperclip-like Turnaround Matching Degree =0.93, the squared Mahalanobis distance between identity and behavior conflict is 38.7 (far higher than the threshold of 9.49, normalized identity deviation). (≈1.00); all four dimensions exhibit clear subjective avoidance and confrontation characteristics, with no explanation possible from objective factors, therefore the four intention coefficients are... All values ​​are 1.0, and the trajectory evidence score is calculated.

[0050] ; The system uses a ship relationship graph API to investigate the relationships between the target ship and three other types of objects: other ships, personnel, and locations. For each type of object, the highest associated risk level is assigned, and then a weighted sum is calculated using a fixed weighting to obtain the associated relationship score, defined as follows: ; in, The values ​​are the highest associated risk levels of the target with other vessels, personnel, and locations, respectively, and are all set to [0,1]. The more direct the relationship (such as synchronous abnormal accompaniment, the same shadow shipowner, joint involvement) and the higher the risk of the associated object itself (such as prior convictions, being under surveillance, red code), the larger the value will be. , , The three categories of objects have fixed weights, and the sum of the three weights is 1.

[0051] In this case, through investigation of the vessel relationship map, it was found that the related vessel "Zhepuyu 00219" exhibited synchronous anomalies on the same day. =0.90; Regarding the connections with personnel, the shipowner and crew members have prior convictions for smuggling. =0.72; In the correlation with location, the target activity points to the sensitive Aokou hideout. =0.55; by weight =0.5、 =0.3、 =0.2, the correlation score is calculated as follows: ; The conclusion is that the trajectory evidence and related evidence of the target are encoded into a case feature vector in a fixed order. Then, in a historical typical case database, the most similar typical illegal case is retrieved using cosine similarity. The product of the case similarity and the degree of confirmation of this typical case is used as the historical case similarity score, defined as: ; In the formula: The cosine similarity between the feature vector of the target case and the feature vector of the matched typical illegal case is taken in [0,1]. express The maximum value of c; c is the confirmation degree of the typical case, which takes the value [0,1]. Cases that have been investigated and determined take the value of 1, while cases that are under investigation or have not been concluded take the lower value.

[0052] in, ; In the formula: A is the target case feature vector, specifically A = ( , , , ), the case feature vector of the typical case matched by B.

[0053] In this case, the four-dimensional case feature vector of the target vessel "Zhepuyu 00128" is A=(0.82,0.88,0.93,1.00), which are respectively: isolated forest abnormal alienation score 0.82, reconnaissance probability 0.88, paperclip reversal and reloading matching degree 0.93, identity deviation normalized value 1.00 (Mathial distance squared 38.7 is much higher than the threshold 9.49, so it is normalized to 1.00); historical typical cases of "sinking on the beach" are pre-stored with the same four features, and its representative vector is B=(0.30,0.30,0.95,0.92).

[0054] Therefore ≈0.91; This typical case has been investigated and determined, with a certainty of c=1.0.

[0055] The similarity score of historical cases was calculated as follows: .

[0056] Finally, the credibility is calculated based on the trajectory evidence score, the correlation score, and the historical case similarity score: ; In the formula: To assess the credibility of the conclusions or the overall evaluation value, These are the weight coefficients for trajectory features, correlation features, and historical features, respectively, and the sum of the three is 1. This refers to the similarity score or trajectory evidence score calculated based on the motion trajectory. The similarity or association score is calculated based on spatiotemporal relationships. This refers to the similarity score calculated based on historical behavioral patterns or historical case similarity scores.

[0057] In this case, the second agent retrieved the ship's history and historical case database, and found that the ship's trajectory evidence highly matched the sinking case. =0.91), and the related vessel "Zhepuyu 00219" also showed a similar anomaly on the same day ( =0.78), combined with historical case similarity ( =0.91), by weight =0.4、 =0.3、 =0.3 to calculate the confidence level =0.87, classified as "high risk of beach sinking".

[0058] Step S5: Compare and judge the anomaly determination result with the qualitative conclusion.

[0059] This step is the core of resolving conflicts between surface-level data and deep qualitative reasoning. When the anomaly detection direction of the first agent aligns with the qualitative conclusion of the second agent, the qualitative conclusion is directly adopted. If the anomaly detection result is inconsistent with the qualitative conclusion, the second agent is triggered to supplement the query of related data and then re-execute the deep qualitative reasoning to update the qualitative conclusion. If inconsistencies remain after re-scoring, the specific reasons for the inconsistency are written into the report and marked as awaiting manual review.

[0060] The associated data includes supplementary queries of ship history records, personnel criminal records, or historical case databases.

[0061] For example, if the first agent determines something to be "highly abnormal" but the second agent initially classifies it as "low risk" due to a lack of relevant evidence, the system will force the second agent to expand its search scope, querying deeper data such as the crew member's criminal record and past involvement in cases, and then re-evaluate the situation, rather than simply discarding it. This feedback reassessment mechanism utilizes the complementarity between surface-level and deep-level knowledge, significantly reducing the risk of false positives and false negatives.

[0062] Step S6: In response to the comparison judge's approval, perform a compliance review on the updated or consistent qualitative conclusion, and assemble and output an analysis report.

[0063] like Figure 4 As shown, the compliance review includes the following three reviews performed sequentially: Isolated Evidence Verification: determining whether the assessment conclusion is supported by evidence from a single source; if so, marking the evidence as isolated evidence and indicating that it cannot be used as the sole basis for a verdict; Terminology Verification: checking whether the assessment text contains speculative expressions (such as "should" or "perhaps"); if so, automatically replacing them with standard law enforcement terminology; and Legal Provision Matching: based on the type of illegal behavior indicated by the assessment conclusion, retrieving applicable legal provisions from the legal knowledge base and attaching them to the assessment report, as shown in Table 1 below. Table 1

[0064] The final analysis report automatically includes target vessel information, anomaly type, credibility score, complete evidence chain (including anomaly alienation score, temporal pattern probability, tactical matching degree, identity deviation degree, and historical similar cases), applicable laws, and handling recommendations. This report is directly pushed to the police officer's mobile terminal, achieving a fully automated closed loop from natural language commands to compliant official documents.

[0065] For cases where the conclusions are inconsistent but consistent after verification, for example, in a certain task, the first agent performed a multi-dimensional quantitative analysis of the anomalies of the target ship "Zhe Dai Yu 03567": its trajectory points overlapped with the radar blind zone boundary for more than 28 minutes, and the calculated anomaly alienation score was 0.86, which was higher than the threshold of 0.70, and the "radar blind zone lurking" behavior was identified; the LSTM model output a probability of 0.81 for "stop-and-go" tracking of its continuous 5-minute trajectory sequence, and the first agent judged the direction as "highly abnormal" based on this. In the first round of in-depth qualitative reasoning, the second agent, during the trajectory investigation stage, considered that the sea area was at the edge of the fishing moratorium and that the swells on that day exceeded 2 meters, which could objectively explain some of the turning back and low-speed wandering, and judged that the subjective intention of confrontation was insufficient, outputting a trajectory evidence score of 0.42; in the correlation investigation stage, no relevant case was found within the three-layer relationship penetration range, outputting a correlation score of 0.20; in the conclusion stage, no highly matching cases were found after searching the historical case database, outputting a historical case similarity score of 0.30; calculated with weights of 0.4, 0.3, and 0.3, the credibility was 0.31, and the initial classification was "low risk". At this time, the direction of the first agent's "highly abnormal" was inconsistent with the direction of the second agent's "low risk", triggering the second agent to supplement the query of relevant data and reason again: after retrieving the ship's resume file, crew members' criminal records and historical case records, it was found that the registered owner of the ship had another ship that was involved in the case, and the target ship had been closely accompanied by the monitored mother ship "Zhepuyu 00219" three times at night within the past 30 days. Based on this, the correlation score jumped from 0.20 to 0.79. The conclusion phase matched the "mother ship transshipment" case, and the historical case similarity score rose to 0.83. The trajectory tracking phase was adjusted to 0.71 because the accompanying behavior confirmed the evasion intention. The credibility was recalculated to 0.78, and the qualitative assessment was reversed from "low risk" to "transshipment suspected". The direction of the analysis is consistent with the direction of the first intelligent agent's anomaly judgment. The comparison and judgment passed, and the case entered the compliance review and was assembled and output as an analysis report.

[0066] If the conclusions are inconsistent and remain inconsistent after review, they are transferred to manual review. For example, the first agent performs tactical pattern matching on the target vessel "Zhepuyu 07712". Its trajectory matches the template of "paperclip-like turning back and forth waiting for loading" with a degree of 0.92, which is higher than the threshold of 0.90. It identifies the tactical action of turning back and forth near the port. Based on this, the first agent judges the direction as "highly abnormal". In the first round of reasoning, the second agent, after reviewing the trajectory item by item, finds that the turning pattern is highly consistent with the "grid-based survey line operation" of a certain marine surveying unit. Moreover, the operation time and area are consistent with the reported surveying task. It is judged that it does not have subjective intention to confront, and the trajectory evidence score is 0.28. In the link mining step, only the normal cooperative relationship between it and the marine monitoring agency is found within the three-layer relationship penetration range, and the link relationship score is 0.18. In the conclusion step, no comparable cases are found after searching the historical case database, and the historical case similarity score is 0.22. The credibility is calculated by weighting 0.4, 0.3 and 0.3, and the confidence is 0.24, which is classified as "low risk". At this point, the "highly anomaly" direction of the first intelligent agent is inconsistent with the "low-risk" direction of the second intelligent agent. The system triggers the second intelligent agent to supplement the query of the operation report records and the crew's criminal record, and then re-evaluates the case. The correlation score and the similarity score of historical cases still do not increase significantly, and the direction remains "low-risk," continuing to be inconsistent with the "highly anomaly" direction of the first intelligent agent. The system does not force a conclusion, but writes the specific reasons for the inconsistency (although the apparent trajectory hits the tactical template, it has legitimate operation report support, lacks correlation and historical evidence) into the analysis report, marking it "pending manual review," and the police will make the final decision based on the on-site verification and report information.

[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for multi-agent task routing and assessment of abnormal maritime situations, characterized in that, The steps include the following: Step S1: Receive and perform semantic parsing of port shoreline control task requests, extract task feature information to generate structured input packets; Step S2: Based on the structured input packet, perform task routing, determine and activate the intelligent agents participating in the processing; Step S3: The first intelligent agent performs multi-dimensional quantitative analysis of the appearance anomalies of the target object based on the structured input package, and outputs the anomaly judgment result. Step S4: The second intelligent agent performs deep qualitative reasoning based on the anomaly determination result and outputs a qualitative conclusion; Step S5: Compare and judge the anomaly determination result with the qualitative conclusion; Step S6: In response to the inconsistency between the direction of the anomaly determination result and the qualitative conclusion, the second agent is triggered to supplement the query of related data and then re-execute the deep qualitative reasoning to update the qualitative conclusion; Step S7: In response to the comparison judge's approval, perform a compliance review on the updated or consistent qualitative conclusion, and assemble and output an analysis report.

2. The method according to claim 1, characterized in that, The output anomaly determination results specifically include: The first intelligent agent sequentially performs anomaly identification on a single ship target from four dimensions: statistical anomaly, temporal behavior recognition, tactical form matching, and identity and behavior consistency. Specifically, the statistical anomaly is used to identify unknown abnormal behaviors lacking prior labels; the temporal behavior recognition is used to identify abnormal behavior patterns with time-dependent characteristics; the tactical morphology matching is used to identify morphology matching anomalies between real-time trajectories and preset tactical templates; and the identity and behavior consistency is used to detect conflicting relationships between target identity attributes and their actual behaviors.

3. The method according to claim 2, characterized in that, The first intelligent agent performs anomaly identification on single-ship targets from the dimension of statistical anomaly by using the Isolation Forest algorithm to perform unsupervised mining of single-ship behavioral features and outputting anomaly alienation scores, including: Using the standard deviation of airspeed, rate of change of heading, and variance of acceleration as inputs, a randomized isolated tree is constructed in memory to calculate the average path length for each target; The abnormal alienation score of the Isolation Forest algorithm is defined as follows: ； In the formula: This is a sample of single-ship behavioral characteristics. The total number of samples; This represents the path length of the sample in the isolated tree. For the sample Average path length in a forest; This is a path length normalization correction term; Scoring for sample anomalies; The path length normalization correction term is: ； In the formula: It is a harmonic function; When the abnormal alienation score exceeds the threshold, it is judged as an abnormal sample: ； In the formula: This is the initially set threshold for anomaly detection.

4. The method according to claim 2, characterized in that, The first intelligent agent performs anomaly recognition on a single ship target from the perspective of temporal behavior recognition by using a long short-term memory network to model multi-dimensional behavior sequences. Let the input behavior feature sequence be... The calculations for each gate are as follows: ； ； ； ； ； ； The final output behavior category probability is: ； In the formula: Output for the forget gate. For input gate output, Candidate memory states, This is a unit memory state. For output gate output, Let be the hidden state vector at the current moment. This is the hidden state from the previous moment. It is the Sigmoid activation function. It is the hyperbolic tangent function. This is Hadamard's element-wise multiplication; Both represent model parameter matrices. Both represent bias terms; For a moment The probability distribution of behavioral categories It is a normalized exponential function.

5. The method according to claim 2, characterized in that, The first intelligent agent performs anomaly identification on single-ship targets from the tactical form matching dimension by using dynamic time warping to compare non-uniform length sequences. Let the trajectory sequence to be detected be... The template trajectory sequence is The dynamic time-warped cumulative distance is recursively calculated as follows: ； Its normalized similarity is expressed as: ； In the formula: The normalized similarity score. The global minimum normalized distance between the two trajectories is given. To normalize the path length, This is the distance-normalized scale parameter; They are respectively the first in the sequence The and the first A trajectory point, The local distance between two points. To reach the grid point The minimum cumulative distance; The optimal template matching result is defined as: ； In the formula: This is the optimal matching template number.

6. The method according to claim 2, characterized in that, The first intelligent agent performs anomaly identification on single-ship targets from the dimension of identity and behavior consistency by using Mahalanobis distance to statistically test the consistency between identity and behavior, and defining the target behavior feature vector as: ； In the formula: For the goal at all times behavioral feature vectors, Indicates average speed. For the maximum permissible speed, Indicates the duration of stay. This indicates the percentage of activities in sensitive areas. Indicates the number of times the device has been stopped and started; For identity categories The standard behavior distribution, whose squared Mahalanobis distance is defined as: ； In the formula: Indicates the target behavior relative to the identity category The squared Mahalanobis distance, For identity categories The behavior mean vector, For identity categories The behavioral covariance matrix; when the distance exceeds a threshold, it is determined to be a conflict between identity and behavior: ； In the formula: This is the preset distance threshold parameter.

7. The method according to claim 1, characterized in that, The first intelligent agent also performs collaborative analysis on multi-ship groups, which includes two levels: long-term collaborative identification and short-term synchronization identification. The long-term collaborative identification includes: within a continuously preset long time window, calculating the similarity between any two ships in three aspects: overall trajectory outline, speed habits, and synchronization of action sequence, and then constructing a global ship similarity matrix to discover hidden gang structures; The short-term synchronization identification includes: based on the identification of a suspected gang structure in the long-term collaborative identification, within a preset short time window, the degree of synchronization of multiple ships experiencing the same abnormal event is counted, and when the degree of synchronization exceeds a preset synchronization threshold, it is determined to be an organized collaborative violation.

8. The method according to claim 7, characterized in that, In the long-term collaborative identification, the similarity between any two ships is defined as: ； In the formula: These are the weighting coefficients for spatial distance, speed difference, and heading difference, respectively, and they satisfy... ; Let the Frescher distance between the two trajectories be within the time window. , and These are the scale parameters for spatial distance, speed difference, and heading difference, respectively; This measures the difference in velocity magnitude or average velocity. This measures the deviation of the heading angle or the average heading difference. Further construct the spatiotemporal similarity matrix: ； In the formula: This indicates the size at a specific time step or state. The matrix, Indicates a ship With ships Spatiotemporal comprehensive similarity, This indicates the number of ships participating in the calculation within the window.

9. The method according to claim 1, characterized in that, The step of performing deep qualitative reasoning based on the anomaly determination result by the second intelligent agent and outputting a qualitative conclusion includes: Trajectory tracing: Starting from the anomaly determination result, examine each of the marked abnormal evidence in the multi-dimensional quantitative analysis of apparent anomalies, determine whether the trajectory anomaly has a subjective intent to resist, and output the trajectory evidence score and the corresponding supporting fact description. Uncover Relationships: Call the ship relationship graph interface to verify the relationship between the target ship and other ships, personnel, and locations, and output the relationship score and a description of the relationship facts; Conclusion: By combining trajectory evidence, correlation evidence, and similarity to historical cases, a qualitative conclusion is formed and a similarity score to historical cases is output; The credibility is calculated based on the trajectory evidence score, the correlation score, and the historical case similarity score: ； In the formula: To assess the credibility of the conclusions or the overall evaluation value, These are the weight coefficients for trajectory features, correlation features, and historical features, respectively, and their sum is 1; This refers to the similarity score or trajectory evidence score calculated based on the motion trajectory. The similarity or association score is calculated based on spatiotemporal relationships. This refers to the similarity score calculated based on historical behavioral patterns or historical case similarity.

10. The method according to claim 1, characterized in that, Perform a compliance review on the updated or consistent qualitative conclusions, including the following three reviews in sequence: Isolated evidence test: Determine whether the conclusion is supported by evidence from a single source. If so, mark the evidence as isolated evidence and indicate that it cannot be used to make a judgment on its own. Language review: Detects and assesses whether the text contains speculative expressions, and automatically replaces them with standard law enforcement language if found; Legal provision matching: Based on the type of illegal behavior indicated by the assessment conclusion, the applicable legal provisions are retrieved from the legal knowledge base and attached to the assessment report.

Images