Ship anti-collision monitoring method and system based on large model self-optimization

Through the ship collision prevention monitoring method based on large model self-optimization, and the self-supervised pre-trained ship position information prediction model, the problem of insufficient accuracy and real-time performance of ship collision prevention algorithms in the existing technology is solved, and more efficient collision avoidance is achieved Development of decision-making and navigation automation technology.

CN118940800APending Publication Date: 2024-11-12CHENGDU SPACEON IND CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410957444.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing ship collision prevention algorithms have insufficient accuracy and real-time performance in improving maritime traffic safety, and rely on labor costs.

Method used

The ship collision prevention monitoring method based on large model self-optimization is adopted. The big model is predicted by self-supervised pre-trained ship position information, and self-optimized with actual data, which improves prediction accuracy and real-timeness and achieves more accurate collision avoidance decisions.

Benefits of technology

It significantly improves the accuracy and refinement of ship position prediction, reduces the work intensity of the driver, improves the accuracy of decision-making, and reduces labor costs, promoting the development of navigation automation technology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940800A_ABST
    Figure CN118940800A_ABST
Patent Text Reader

Abstract

The invention discloses a ship anti-collision monitoring method and system based on large model self-optimization, and belongs to the technical field of ship anti-collision monitoring, and the method comprises the steps: collecting the historical data of a ship automatic recognition system; interpolating the missing record according to a fixed time sequence; constructing a large ship position prediction model, and training the constructed prediction model; inputting to-be-predicted ship feature data into the trained prediction model to obtain position information of the ship at the next moment; distance calculation is carried out on the position information of all ships at the next moment; performing anti-collision early warning on the to-be-predicted ship according to the distance threshold value; acquiring AIS real-time data; and inputting into a model self-optimization system, and updating the ship position prediction large model. According to the invention, an automatic early warning system for ship collision is realized, and the prediction model is optimized, so that the prediction model has self-updating and optimization capabilities, and powerful technical support can be provided for future intelligent ship management.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of ship anti-collision monitoring, and more specifically, to a ship anti-collision monitoring method and system based on large model self-optimization. Background Art

[0002] With the growth of global trade and economic activities, the number of ships in the world is increasing. However, the increase in the number of ships has also increased the risk of maritime accidents, making it increasingly important to strengthen ship safety supervision and improve collision avoidance technology. Ship collision avoidance algorithm is one of the key technologies to ensure maritime traffic safety. It can help crew members or autonomous driving systems to identify potential collision risks in a timely manner and take appropriate collision avoidance measures. At present, collision avoidance algorithms are mainly rule-based collision avoidance and automatic collision avoidance. Rule-based collision avoidance is to formulate reasonable avoidance strategies by judging parameters such as the relative position, speed and heading between ships. Automatic collision avoidance uses data collected by equipment such as AIS and radar to assess the risk of collision and guide ship drivers to take collision avoidance measures based on this assessment. Although the existing ship collision avoidance scheme has played a positive role in improving maritime traffic safety, it still needs to be improved to improve its accuracy and real-time performance. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a ship collision avoidance monitoring method and system based on large model self-optimization, which predicts a large model of ship position information through self-supervision and pre-training, and achieves collision avoidance by calculating the distance of the predicted output ship's future position, and further self-optimizes the prediction model through actual data, thereby improving the accuracy and real-time performance, and solving the accuracy and labor cost problems of existing ship collision avoidance technology.

[0004] The object of the present invention is achieved through the following solutions:

[0005] A ship anti-collision monitoring method based on large model self-optimization includes the following steps:

[0006] S1, collect ship data;

[0007] S2, data preprocessing;

[0008] S3, building a large model for predicting ship position information;

[0009] S4, training a large model for predicting ship position information;

[0010] S5, model prediction, predicts the position of the ship at the next moment based on the trained ship position information prediction model;

[0011] S6, calculating the distance between the ships outputted by the prediction;

[0012] S7, based on the threshold information, a warning is issued to the ship that is about to collide at the next moment.

[0013] Furthermore, after step S7, the method further includes:

[0014] Step S8, construct a self-optimizing model system and optimize it regularly.

[0015] Furthermore, in step S1, the collecting of ship data specifically includes collecting historical data of the ship automatic identification system AIS, and an adaptive sampling rate method is adopted in the data collection process to adjust the data sending frequency according to the actual dynamics of the ship and the surrounding environment.

[0016] Further, in step S2, the data preprocessing specifically includes abnormal data processing and interpolation processing.

[0017] Further, in step S3, the constructing of the ship position information prediction model includes: constructing the ship position information prediction model based on the decoder model, specifically including an input module, a decoding module, and an output module;

[0018] The input module is the time series data {x1, x2, x3, …, xt} of the ship's position information, where xt represents the vector of the position information of a certain ship at time t, and xt = [longitude, latitude, course over the ground, speed over the ground, heading, turning rate, heading status]; the vector of the ship's position information is used as the input of the Decoder module;

[0019] The decoding module contains six Decoder Blocks. The output of the previous Block is the input of the next Block. Each Block contains a Masked Self-Attention unit and an FNN unit. The Masked Self-Attention unit is used to receive data from the input module, and the FNN unit is used to receive the output of the Masked Self-Attention unit.

[0020] In the output module, its content remains consistent with the input module, but it moves forward one step in time.

[0021] Furthermore, in step S4, the training of the large model for predicting ship position information specifically includes the following sub-steps:

[0022] First, the preprocessed data are grouped according to the MMSI code of the ship, and the trajectory of a single ship is divided into segments by setting a sliding window;

[0023] Then, a loss function is constructed to calculate the loss generated by the training samples for gradient update;

[0024] Finally, set the training hyperparameters and input trajectory segments for training until the loss function value becomes stable and then stop training.

[0025] Furthermore, in step S5, the position of the ship at the next moment is predicted based on the trained large model for predicting the ship's position information, specifically including: according to the trained large model for predicting the ship's position information, the trajectory fragment of the ship to be predicted is input into the model to predict and save the position of the ship at the next moment.

[0026] Furthermore, in step S6, the distance between the ships output by calculating the prediction specifically includes: for the position information of the ships at the next moment saved in step S5, only the longitude and latitude therein are obtained and the distance between the two ships is calculated, and the distance is calculated by the Haversine formula.

[0027] Further, in step S8, the self-optimizing model system is constructed and optimized regularly, specifically including: constructing the self-optimizing model system, first receiving the real-time data of the ship automatic identification system AIS and storing it in the data cache area; then initializing the Lora layer and fine-tuning it; then outputting the loss Loss1 and loss Loss2 of the LLM model and the fine-tuned Lora layer respectively; finally, comparing the sizes of the two losses, if Loss2 is less than Loss1, it means that the fine-tuned model has higher accuracy, then outputting the gradient ΔW, and updating the LLM model with the gradient, if Loss2 is greater than Loss1, it means that the original LLM model has higher accuracy, then reinitializing the Lora layer and waiting for the next fine-tuning; wherein the LLM model is the large model for predicting ship position information constructed in step S3.

[0028] A ship anti-collision monitoring system based on large model self-optimization, comprising:

[0029] Data acquisition module, used to collect ship data;

[0030] Preprocessing module, used for data preprocessing;

[0031] Prediction model building module, used to build a large model for predicting ship position information;

[0032] Training module, used to train the large model for predicting ship position information;

[0033] The prediction module is used for model prediction and predicts the position of the ship at the next moment based on the trained ship position information prediction model;

[0034] A distance calculation module is used to calculate the distance between ships of the predicted output;

[0035] The early warning module is used to issue an early warning to a ship that is about to collide at the next moment according to the threshold information; and the system is used to execute the ship anti-collision monitoring method based on large model self-optimization as described in any of the above items.

[0036] The beneficial effects of the present invention include:

[0037] (1) The method of the present invention cleverly uses large model technology to extract rich information about ship navigation and can learn more features and patterns to predict the ship's position. Compared with direct prediction by traditional models, this method significantly improves the accuracy and detail of ship position prediction.

[0038] (2) In the prediction of the ship's position, the model in the method of the present invention not only incorporates the longitude and latitude of the ship, but also incorporates information such as the ship's heading relative to the ground and speed, thereby realizing anti-collision warning from multiple dimensions, which helps to improve the accuracy of position prediction.

[0039] (3) As the ship's navigation track increases, the self-optimizing model system in the method of the present invention can regularly update the model according to the new position, thereby optimizing the prediction model of the ship's position, thereby improving the generalization of the model.

[0040] (4) The anti-collision scheme of the present invention adopts advanced large model technology and realizes the liberalization of the model, providing more accurate collision avoidance auxiliary decision-making, reducing the workload of the driver and improving the accuracy of decision-making. It also effectively promotes the development of navigation automation technology and provides a reference and application platform for technological innovation in other related fields.

[0041] (5) The solution of the present invention can not only optimize the ship's route planning, reduce unnecessary stops and detours, thereby saving time and fuel costs, and improving overall navigation efficiency. This is particularly important for global trade, as maritime transportation is a key link in the global supply chain. At the same time, preventing ship collisions is not only to protect people and cargo, but also to prevent the leakage of oil and other harmful substances caused by collisions, which is crucial to protecting the marine environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 is a flowchart of the steps of the method according to the embodiment of the present invention;

[0044] Figure 2 It is the basic structure of the large model for predicting the ship position information in the embodiment of the present invention;

[0045] Figure 3 This is the structural diagram of the Masked Self-Attention unit;

[0046] Figure 4 It is the structural diagram of the FNN unit;

[0047] Figure 5 It is the loss function value change graph;

[0048] Figure 6 For collision warning example;

[0049] Figure 7 This is a diagram of the model optimization system architecture of an embodiment of the present invention;

[0050] Figure 8 Comparison of loss values ​​before and after optimization. DETAILED DESCRIPTION

[0051] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0052] In order to reduce the reliance on human factors and better assess the risk of collision. The present invention proposes a ship position information prediction model and application scheme based on large-scale pre-training in the technical concept, and realizes collision avoidance through position calculation, and can automatically optimize the prediction model, thereby improving the prediction accuracy and reducing labor costs, and better adapting to complex marine environments and situations. In a preferred embodiment, the specific implementation process is as follows:

[0053] Specifically, the present invention provides a ship anti-collision monitoring method based on large model self-optimization, such as Figure 1 As shown, the following steps are included:

[0054] Step 1: Through the shore-based receiving station, detailed monitoring and data collection work is carried out on ships sailing in specific sea areas. The collected data covers the static information, dynamic information and navigation status information of the ship. In the data collection process, an adaptive sampling rate method is used to adjust the data transmission frequency according to the actual dynamics of the ship and the surrounding environment. Reducing the sampling rate in areas with dense ships can reduce the receiving and processing pressure of the AIS base station and avoid network congestion. In critical operations or high-risk areas (such as port entrances, narrow waterways, etc.), increasing the sampling rate can provide more accurate ship location and dynamic information and improve data quality. This method can effectively achieve network load balancing and data quality optimization.

[0055] Step 2: Perform data preprocessing, including abnormal data processing and interpolation processing.

[0056] Data exception processing includes: 1) Deleting records with empty MMSI code or MMSI code length not equal to 9. 2) Deleting records with empty latitude, longitude, and heading. 3) Deleting records with longitude values ​​not in the range of -180 to +180 degrees or latitude values ​​not in the range of -90 to +90 degrees.

[0057] Data interpolation process: First, construct the time series of AIS data by 10 seconds; then, perform linear interpolation on the empty points in the time series. Then, delete the sample points that are not in the time series. The formula for linear interpolation is as follows:

[0058]

[0059] Among them, x t represents the estimated value of field x at time t, represents the true value of x at time t2, Represents the true value at time t1.

[0060] Table 1 Partial data example table

[0061]

[0062]

[0063] Step 3: The ship position information prediction model consists of three modules: input module, decoding module, and output module. The model architecture is shown in the figure below: Figure 2 As shown. The input module is the time series data of the ship's position information {x1, x2, x3, …, xt}, where xt represents the vector of the position information of a certain ship at time t, and xt = [longitude, latitude, course over the ground, speed over the ground, heading, turning rate, heading state]. The vector consisting of the ship's position information is used as the input of the decoder module. When processing tasks related to ship navigation, monitoring or autonomous driving, it has the following advantages: 1) Comprehensiveness: The input vector contains multiple key navigation parameters of the ship, providing comprehensive contextual information for the model. 2) Context-awareness: The model can understand the current state and environment of the ship, including heading, speed and steering behavior, so as to make more reasonable decisions. 3) Enhanced generalization ability: Input vectors containing a variety of information help the model better generalize to different marine environments and operating conditions.

[0064] The decoding module contains 6 Decoder Blocks, and the output of the previous Block is the input of the next Block. Each Block contains a Masked Self-Attention unit and an FNN unit. The Masked Self-Attention unit receives data from the input module, and the FNN unit receives the output of the Masked Self-Attention unit. Using the vector composed of the ship's position information for the decoder's mask self-attention calculation can bring the following advantages: 1) Dynamic information filtering: It can dynamically determine which information is important and which can be temporarily ignored based on the current state of the ship (such as the heading state). 2) Contextual focus: The model can focus on the information that is most critical to the current decision, such as turning rate and heading when turning. 3) Reduce interference: In ship navigation, some information may be noise or interference in certain circumstances. Mask self-attention can help the model ignore these interferences.

[0065] The structure of the Masked Self-Attention unit is as follows Figure 3 As shown. Among them, X = {x1, x2, x3, ..., xt} is the data output by the input module, H = {h1, h2, h3, ..., ht} is the hidden vector output by the Masked Self-Attention unit, and the dotted line part indicates the calculation method of Attention. The specific calculation formula is as follows:

[0066]

[0067] Q=W Q X

[0068] K=W K X

[0069] V=W V X

[0070] Among them, W Q , W K , W V are three different parameter matrices.

[0071] The structure of the FNN unit is as follows Figure 4 As shown in the figure, the input layer receives the output of the Masked Self-Attention unit, and the hidden layer and output layer undergo a nonlinear transformation. The number of neurons in the hidden layer and output layer is 128 and 7 respectively. The formula for nonlinear transformation is as follows:

[0072] y=sigmoid(WX+b)

[0073] Among them, sigmoid is a nonlinear activation function.

[0074] In the output module, its content is consistent with the input module, one step forward in time. This self-supervised learning method has the following advantages: 1) The consistency of the input and output modules provides a symmetrical learning framework, which may help the model to more effectively capture and utilize the intrinsic structure of the data. 2) Self-supervised learning can use a large amount of unlabeled data to learn representations through prediction tasks (such as masked language models), which usually saves more annotation costs than supervised learning.

[0075] Step 4: Train the large model for predicting ship position information. First, group the preprocessed data by the ship's MMSI code, and divide the trajectory of a single ship into segments by setting a sliding window. Then, construct a loss function to calculate the loss generated by the training sample for gradient update. Finally, set the training hyperparameters, such as learning rate, batch size, etc., input the trajectory segment for training, and stop training until the loss function value tends to be stable. The formula of the loss function is as follows:

[0076]

[0077] Among them, w i represents the loss weight of the i-th feature, with a value range of [0,1], y i represents the true value of the i-th feature, Represents the predicted value of the i-th feature. Figure 5 This is the change graph of the loss function value.

[0078] Step 5: Based on the trained ship position information prediction model, the trajectory fragment of the ship to be predicted is input into the model, and the position of the ship at the next moment is predicted and saved.

[0079] Step six, for the next moment position information of the ship saved in step five, only the longitude and latitude are obtained and the distance between the two ships is calculated. The distance is realized by the Haversing formula. This calculation method has the following advantages over the general plane distance. 1) Accuracy: The Haversing formula takes into account the curvature of the earth, so it is more accurate than the simple Euclidean distance (straight-line distance) when calculating across larger distances. 2) Applicability: It is suitable for calculating the distance between any two points on the earth, especially when the two points are not on the same longitude or latitude. 3) Versatility: The Haversing formula can be applied to various fields, such as navigation, aviation, geographic information system (GIS), map making, etc.

[0080] The calculation code is as follows:

[0081]

[0082] Step 7: Set the distance threshold to 30 meters. When the distance is less than 30 meters, output the ship's MMSI code for warning. Figure 6 As shown in the figure, the ship 667001322 and the ship 311001296 were traveling from northeast to southwest and were about to collide near 121.67 east longitude. After the collision warning was issued by the model, the two ships maintained a certain distance in the subsequent journey, ensuring navigation safety.

[0083] The accuracy of ship collision recognition through this model has been significantly improved in recall rate compared to the method of establishing a straight line model to identify collisions. The straight line model is to establish a straight line model through the ground heading and longitude and latitude coordinates, and predict the location of the ship collision through the intersection of the straight lines, so as to achieve the purpose of anti-collision recognition.

[0084] Table 2 Comparison of indicators between the decoder large model and the linear model

[0085] Accuracy Recall Decoder large model 85.64% 78.77% Straight Line Model 80.33% 69.60%

[0086] Step 8: Build a self-optimizing model system. First, receive AIS real-time data and store it in the data cache. Then initialize the Lora layer and perform fine-tuning. Then output the losses Loss1 and Loss2 of the LLM model and the fine-tuned Lora layer respectively. Finally, compare the two losses. If Loss2 is less than Loss1, it means that the fine-tuned model has higher accuracy. Then output the gradient ΔW and update the LLM model with the gradient. If Loss2 is greater than Loss1, it means that the original LLM model has higher accuracy. Then reinitialize the Lora layer and wait for the next fine-tuning. The system architecture is as follows: Figure 7 As shown in the figure, LLM is the trained ship position information prediction model, and W is the parameter of the model. The parameter update formula is as follows:

[0087]

[0088] In the above formula, W ′ represents the updated parameters of the LLM model, α is the scaling factor of the gradient, r is the low-rank coefficient, and B and A are the parameter matrices of the Lora layer. Figure 8 It is the comparison of loss values ​​before and after optimization.

[0089] It should be noted that within the scope of protection defined in the claims of the present invention, the following embodiments can be combined and / or expanded or replaced in any logical way from the above specific implementation methods, such as disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0090] In other embodiments, including but not limited to the following embodiments:

[0091] Example 1

[0092] A ship anti-collision monitoring method based on large model self-optimization includes the following steps:

[0093] S1, collect ship data;

[0094] S2, data preprocessing;

[0095] S3, building a large model for predicting ship position information;

[0096] S4, training a large model for predicting ship position information;

[0097] S5, model prediction, predicts the position of the ship at the next moment based on the trained ship position information prediction model;

[0098] S6, calculating the distance between the ships outputted by the prediction;

[0099] S7, based on the threshold information, a warning is issued to the ship that is about to collide at the next moment.

[0100] Example 2

[0101] On the basis of Embodiment 1, after step S7, the method further includes:

[0102] Step S8, construct a self-optimizing model system and optimize it regularly.

[0103] Example 3

[0104] On the basis of Example 1, in step S1, the collecting of ship data specifically includes collecting historical data of the ship automatic identification system AIS, and an adaptive sampling rate method is used in the data collection process to adjust the data sending frequency according to the actual dynamics of the ship and the surrounding environment.

[0105] Example 4

[0106] On the basis of Example 1, in step S2, the data preprocessing specifically includes abnormal data processing and interpolation processing.

[0107] Example 5

[0108] On the basis of Example 1, in step S3, the construction of the ship position information prediction model includes: constructing the ship position information prediction model based on the decoder model, specifically including an input module, a decoding module, and an output module;

[0109] The input module is the time series data {x1, x2, x3, …, xt} of the ship's position information, where xt represents the vector of the position information of a certain ship at time t, and xt = [longitude, latitude, course over the ground, speed over the ground, heading, turning rate, heading status]; the vector of the ship's position information is used as the input of the Decoder module;

[0110] The decoding module contains six Decoder Blocks. The output of the previous Block is the input of the next Block. Each Block contains a Masked Self-Attention unit and an FNN unit. The Masked Self-Attention unit is used to receive data from the input module, and the FNN unit is used to receive the output of the Masked Self-Attention unit.

[0111] In the output module, its content remains consistent with the input module, but it moves forward one step in time.

[0112] Example 6

[0113] On the basis of Example 1, in step S4, the training of the large model for predicting ship position information specifically includes the following sub-steps:

[0114] First, the preprocessed data are grouped according to the MMSI code of the ship, and the trajectory of a single ship is divided into segments by setting a sliding window;

[0115] Then, a loss function is constructed to calculate the loss generated by the training samples for gradient update;

[0116] Finally, set the training hyperparameters and input trajectory segments for training until the loss function value becomes stable and then stop training.

[0117] Example 7

[0118] On the basis of Example 6, in step S5, the position of the ship at the next moment is predicted based on the trained ship position information prediction model, specifically including: according to the trained ship position information prediction model, the trajectory fragment of the ship to be predicted is input into the model to predict and save the position of the ship at the next moment.

[0119] Example 8

[0120] Based on Example 7, in step S6, the distance between the ships output by calculating the predicted output specifically includes: for the position information of the ships at the next moment saved in step S5, only the longitude and latitude therein are obtained and the distance between each pair of ships is calculated, and the distance is calculated by the Haversine formula.

[0121] Example 9

[0122] On the basis of Example 2, in step S8, the self-optimizing model system is constructed and optimized regularly, specifically including: constructing the self-optimizing model system, first receiving the real-time data of the ship automatic identification system AIS and storing it in the data cache area; then initializing the Lora layer and fine-tuning it; then outputting the loss Loss1 and loss Loss2 of the LLM model and the fine-tuned Lora layer respectively; finally, comparing the sizes of the two losses, if Loss2 is less than Loss1, it means that the fine-tuned model has higher accuracy, then outputting the gradient ΔW, and updating the LLM model with the gradient, if Loss2 is greater than Loss1, it means that the original LLM model has higher accuracy, then reinitializing the Lora layer and waiting for the next fine-tuning; wherein the LLM model is the large model for predicting ship position information constructed in step S3.

[0123] Example 10

[0124] A ship anti-collision monitoring system based on large model self-optimization, comprising:

[0125] Data acquisition module, used to collect ship data;

[0126] Preprocessing module, used for data preprocessing;

[0127] Prediction model building module, used to build a large model for predicting ship position information;

[0128] Training module, used to train the large model for predicting ship position information;

[0129] The prediction module is used for model prediction and predicts the position of the ship at the next moment based on the trained ship position information prediction model;

[0130] A distance calculation module is used to calculate the distance between ships of the predicted output;

[0131] The early warning module is used to issue an early warning to a ship that is about to collide at the next moment according to the threshold information; and the system is used to execute the ship anti-collision monitoring method based on large model self-optimization described in any one of Examples 1 to 9.

[0132] The units involved in the embodiments of the present invention may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, limit the units themselves.

[0133] According to one aspect of an embodiment of the present invention, a computer program product or a computer program is provided, the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above various optional implementations.

[0134] As another aspect, an embodiment of the present invention further provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.

Claims

1. A ship anti-collision monitoring method based on large model self-optimization, characterized in that: The following steps are involved: S1, collect ship data; S2, data preprocessing; S3, building a large model for predicting ship position information; S4, training a large model for predicting ship position information; S5, model prediction, predicts the position of the ship at the next moment based on the trained ship position information prediction model; S6, calculating the distance between the ships outputted by the prediction; S7, based on the threshold information, a warning is issued to the ship that is about to collide at the next moment.

2. The ship anti-collision monitoring method based on large model self-optimization according to claim 1 is characterized in that: After step S7, the method further includes: S8, build a self-optimizing model system and optimize it regularly.

3. The ship anti-collision monitoring method based on large model self-optimization according to claim 1 is characterized in that: In step S1, the collecting of ship data specifically includes collecting historical data of the ship automatic identification system AIS, and in the data collection process, an adaptive sampling rate method is adopted to adjust the data transmission frequency according to the actual dynamics of the ship and the surrounding environment.

4. The ship anti-collision monitoring method based on large model self-optimization according to claim 1 is characterized in that: In step S2, the data preprocessing specifically includes abnormal data processing and interpolation processing.

5. The ship anti-collision monitoring method based on large model self-optimization according to claim 1 is characterized in that: In step S3, the construction of the ship position information prediction model includes: constructing the ship position information prediction model based on the Decoder model, specifically including an input module, a decoding module, and an output module; The input module is the time series data {x1, x2, x3, …, xt} of the ship's position information, where xt represents the vector of the position information of a certain ship at time t, and xt = [longitude, latitude, course over the ground, speed over the ground, heading, turning rate, heading status]; the vector of the ship's position information is used as the input of the Decoder module; The decoding module contains six Decoder Blocks. The output of the previous Block is the input of the next Block. Each Block contains a Masked Self-Attention unit and an FNN unit. The Masked Self-Attention unit is used to receive data from the input module, and the FNN unit is used to receive the output of the Masked Self-Attention unit. In the output module, its content remains consistent with the input module, but it moves forward one step in time.

6. The ship anti-collision monitoring method based on large model self-optimization according to claim 1 is characterized in that: In step S4, the training of the ship position information prediction model specifically includes the following sub-steps: First, the preprocessed data are grouped according to the MMSI code of the ship, and the trajectory of a single ship is divided into segments by setting a sliding window; Then, a loss function is constructed to calculate the loss generated by the training samples for gradient update; Finally, set the training hyperparameters and input trajectory segments for training until the loss function value becomes stable and then stop training.

7. The ship anti-collision monitoring method based on large model self-optimization according to claim 6 is characterized in that: In step S5, the position of the ship at the next moment is predicted based on the trained large model for predicting the ship's position information, specifically including: according to the trained large model for predicting the ship's position information, the trajectory fragment of the ship to be predicted is input into the model to predict and save the position of the ship at the next moment.

8. The ship anti-collision monitoring method based on large model self-optimization according to claim 7 is characterized in that: In step S6, the distance between the ships is calculated and predicted, specifically including: for the position information of the ships at the next moment saved in step S5, only the longitude and latitude are obtained and the distance between the two ships is calculated, and the distance is calculated by the Haversine formula.

9. The ship anti-collision monitoring method based on large model self-optimization according to claim 2 is characterized in that: In step S8, the self-optimizing model system is constructed and optimized regularly, specifically including: constructing the self-optimizing model system, first receiving the real-time data of the ship automatic identification system AIS and storing it in the data cache area; then initializing the Lora layer and fine-tuning it; then outputting the loss Loss1 and loss Loss2 of the LLM model and the fine-tuned Lora layer respectively; finally, comparing the sizes of the two losses, if Loss2 is less than Loss1, it means that the fine-tuned model has higher accuracy, then outputting the gradient ΔW, and updating the LLM model with the gradient, if Loss2 is greater than Loss1, it means that the original LLM model has higher accuracy, then reinitializing the Lora layer and waiting for the next fine-tuning; wherein the LLM model is the large model for predicting ship position information constructed in step S3.

10. A ship anti-collision monitoring system based on large model self-optimization, characterized in that: include: Data acquisition module, used to collect ship data; Preprocessing module, used for data preprocessing; Prediction model building module, used to build a large model for predicting ship position information; Training module, used to train the large model for predicting ship position information; The prediction module is used for model prediction and predicts the position of the ship at the next moment based on the trained ship position information prediction model; A distance calculation module is used to calculate the distance between ships of the predicted output; The early warning module is used to issue an early warning to a ship that is about to collide at the next moment according to threshold information; and the system is used to execute the ship anti-collision monitoring method based on large model self-optimization as described in any one of claims 1 to 9.

Citation Information

Cited By

  • Ship operation efficiency intelligent control system and method based on large model optimization

    CN119916692A

  • Multi-feature trajectory prediction and cross-defense early warning method for edge and coast defense control objects

    CN121011112A