A Ship Trajectory Prediction System and Method Based on Large Models

Abstract

The present invention discloses a ship trajectory prediction system and method based on a large model, which relates to the field of marine intelligent navigation technology. It includes collecting multi-modal data through multiple channels; respectively extracting features from the multi-modal data and outputting fused environmental perception features; constructing an anomaly detection model based on the fused environmental perception features, inputting the fused environmental perception feature vector into the anomaly detection model, identifying and filtering out abnormal data through a causal inference algorithm, and outputting pure features; constructing a meta-learning model by using a meta-learning mechanism and the pure features, inputting the pure features into the meta-learning model, and outputting a personalized prediction model through a sub-model that quickly adapts to the specific ship behavior pattern; combining the personalized prediction model with real-time data, predicting the ship trajectory through a large model, and outputting a predicted trajectory; establishing a safety decision-making assistance system, comparing the predicted trajectory with the navigation rules, and when the prediction result exceeds the safety threshold, triggering an alarm and providing safety suggestions to the ship.

Description

Technical Field

[0001] The present invention relates to the technical field of marine intelligent navigation, and particularly to a ship trajectory prediction system and method based on a large model. Background Art

[0002] In marine transportation and military operations, ship trajectory prediction is crucial for ensuring navigation safety, avoiding collision accidents, planning efficient routes, and tactical deployment. In recent years, with the rapid development of artificial intelligence technology, especially the application of deep learning and big data analysis, ship trajectory prediction technology has been significantly improved. Traditional trajectory prediction methods mainly rely on physical models and statistical analysis. Although these methods can reflect the basic laws of ship movement, in a complex and changing marine environment, their prediction accuracy is often limited. Physical models are limited by simplified assumptions about environmental factors, while statistical methods are difficult to capture the impact of non-linear and sudden events.

[0003] In recent years, ship trajectory prediction methods based on machine learning have gradually become a research hotspot. Such methods utilize multi-source information of historical trajectory data, environmental data, and ship characteristics, and can more accurately predict the future course and speed of ships by training deep neural network models. However, there are still some key problems in the current prediction models. On the one hand, the generalization ability of the models is limited, and the prediction effect for new environments or specific ship behaviors is not good. Especially when dealing with abnormal data and extreme conditions, the robustness of the models needs to be enhanced. On the other hand, most methods lack the ability to quickly adapt to real-time data, resulting in a decline in prediction timeliness and accuracy. In addition, there are also limitations in safety decision-making assistance in the existing technology. It is difficult to evaluate the safety of the predicted trajectory in real time, and there is a lack of effective alarm and suggestion mechanisms, which may pose safety hazards in complex sea conditions. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a ship trajectory prediction system and method based on a large model to solve the problems that the generalization ability of the model is limited, the prediction effect for new environments or specific ship behaviors is not good, especially when dealing with abnormal data and extreme conditions, the robustness of the model needs to be enhanced. On the other hand, most methods lack the ability to quickly adapt to real-time data, resulting in a decline in prediction timeliness and accuracy. In addition, there are also limitations in safety decision-making assistance in the existing technology. It is difficult to evaluate the safety of the predicted trajectory in real time, and there is a lack of effective alarm and suggestion mechanisms, which may pose safety hazards in complex sea conditions.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a large model-based ship trajectory prediction method, which includes:

[0008] Collect multi-modal data through multiple channels;

[0009] Extract features from the multi-modal data respectively, and output fused environmental perception features;

[0010] Construct an anomaly detection model based on the fused environmental perception features, input the fused environmental perception feature vector into the anomaly detection model, identify and filter out abnormal data through a causal inference algorithm, and output pure features;

[0011] Construct a meta-learning model using the meta-learning mechanism and pure features, input the pure features into the meta-learning model, and output a personalized prediction model through a sub-model that quickly adapts to specific ship behavior patterns;

[0012] Combine the personalized prediction model with real-time data, and perform ship trajectory prediction through a large model to output a predicted trajectory;

[0013] Establish a safety decision-making assistance system, compare the predicted trajectory with navigation rules, and when the prediction result exceeds the safety threshold, trigger an alarm and provide safety suggestions to the ship.

[0014] As a preferred solution of the large model-based ship trajectory prediction method of the present invention, wherein: the specific steps of collecting multi-modal data through multiple channels are as follows:

[0015] Collect unstructured text description data, historical trajectory data, and environmental condition data from multiple sources;

[0016] Collect radar images and sonar signals through sensors;

[0017] Perform preprocessing of denoising and missing value filling on the collected data.

[0018] As a preferred solution of the large model-based ship trajectory prediction method of the present invention, wherein: the specific steps of extracting features from the multi-modal data respectively and outputting fused environmental perception features are as follows:

[0019] Use the BERT model to convert the unstructured text description into a numerical vector to obtain a text feature vector, and the expression is:

[0020] T = BERT(wo);

[0021] Wherein, T represents the text feature obtained after being processed by the BERT model, and wo represents the input text sequence;

[0022] The sliding window technique is used to extract the rates of change of the speed, direction, and position of the track, obtaining a track feature vector. The size of the sliding window is set to W, and the step size of the sliding window is s. The expression is as follows:

[0023]

[0024] Among them, H represents the track feature, i represents the starting time point of the window, represents the summation operation starting from time point i and ending at time point i + W - 1, v t represents the speed vector at time t, θ t represents the direction angle at time t, x t represents the position coordinate at time t, x t+s -x t represents the displacement vector of the position between time t and time t + s;

[0025] The discrete Fourier transform is used to perform spectral analysis on the environmental condition data. The meteorological data at time t is set as m t , and the ocean current data at time t is set as o t . The expression for extracting the environmental condition feature is:

[0026] E = (DFT(m), DFT(o));

[0027] Among them, E represents the environmental condition feature, and DFT represents the discrete Fourier transform;

[0028] The convolutional neural network CNN is used to extract features from the sensor data. The two-dimensional matrix of the sensor data is set as so, and the expression is:

[0029] S = CNN(so) = ReLU(K * so);

[0030] Among them, S represents the sensor feature, K represents the CNN filter kernel, ReLU represents the activation function, and * represents the convolution operation;

[0031] The extracted text features, track features, environmental condition features, and sensor features are fused by the weighted average method. The expression is:

[0032] Z = α T T + α H H + α E E + α S S;

[0033] α T + α H + α E + α S = 1;

[0034] Among them, Z represents the fused environmental perception feature, and α T , α H , α E and α S respectively represent the weights of text features, track features, environmental condition features, and sensor features.

[0035] As a preferred solution of the ship track prediction method based on a large model according to the present invention, wherein: an anomaly detection model is constructed based on the fused environmental perception feature, the fused environmental perception feature vector is input into the anomaly detection model, and causal inference algorithms are used to identify and filter out abnormal data, and pure features are output. The specific steps are as follows:

[0036] Use a multivariate Gaussian distribution, information entropy, and combine the latent causal variable model LCM in causal inference to construct an anomaly detection model. The expression of the anomaly detection model is:

[0037]

[0038] Among them, A(Z) represents the output of the anomaly detection model, represents the integral symbol, μ represents the mean vector, Σ represents the covariance matrix, N(Z; μ, Σ) represents the multivariate Gaussian distribution function, ι(Z) represents the information filtering function, exp(-ι(Z)) represents the exponential function, du is the differential element, representing the infinitesimal change of the integral variable μ;

[0039] The expression of the information filtering function ι(Z) is:

[0040]

[0041] Among them, p i represents the probability of the occurrence of the i-th feature, n represents the number of features in Z j , log represents the natural logarithm function. When p i is smaller, the absolute value of log is larger, λ represents the regularization parameter, represents the sum of all terms from 1 to m, m represents the number of features in j Z represents the actual observed value of the j-th feature in Z, represents the expected value of the j-th feature, and q represents the exponential operation.

[0042] As a preferred solution of the ship track prediction method based on a large model according to the present invention, wherein: a meta-learning model is constructed using the meta-learning mechanism and pure features, the pure features are input into the meta-learning model, and a sub-model that quickly adapts to the specific ship behavior pattern is used to output a personalized prediction model. The specific steps are as follows:

[0043] The meta - learning structure is set to consist of a shared base model and sub - models for a series of tasks;

[0044] Set each ship as k, and then construct a sub - model to quickly adapt to the specific behavior pattern of ship k through historical ship track data;

[0045] Based on ship k and the base model P, construct an adaptability metric function, and the expression is:

[0046]

[0047] Among them, R k (P, Z k ) represents the reliability metric of ship k under the pure feature Z k and the prediction P. P represents the prediction result of the base model, Z k represents the set of pure features of ship k, T represents the total number of time steps, d k (t) represents the deviation at time t, α k represents a constant, β k represents another constant, μ d represents the mean of the deviation, represents the variance of the deviation, N represents the normal distribution function, represents calculating the sum of the probability densities of the normal distribution within the entire deviation space;

[0048] Use the adaptability metric function as the objective function R k , and optimize the sub - model;

[0049] When the sub - model is optimized, that is, when the personalized prediction model is optimized, the output vector is R k (t).

[0050] As a preferred solution of the ship track prediction method based on the large model described in the present invention, among them: The combination of the personalized prediction model and real - time data, and the ship track prediction through the large model, and the output of the predicted track, the specific steps are as follows:

[0051] Set the real - time data input vector at time t as X(t);

[0052] Based on the real - time data X(t) and R k (t), and at the same time combined with the set of pure features Z of ship k k Construct a composite function G, and the expression is:

[0053]

[0054] Among them, Y k (t) represents the predicted track of ship k at time t, P k(t) represents the output vector of the personalized prediction model of ship k at time t, w i is used to adjust the importance of each sensor data in the comprehensive prediction, ReLU represents the activation function, K represents the convolution kernel, B represents the matrix for adjusting the output of the personalized prediction model, C represents the matrix for adjusting the pure features, μ k represents the mean value of the real-time data, represents the variance of the real-time data.

[0055] As a preferred solution of the ship trajectory prediction method based on the large model described in the present invention, wherein: the safety decision-making assistance system is established, the predicted trajectory is compared with the navigation rules, and when the prediction result exceeds the safety threshold, an alarm is triggered and safety suggestions are provided to the ship. The specific steps are as follows:

[0056] Set the set of navigation rules as R;

[0057] Construct a safety evaluation function based on the predicted trajectory of the ship, and the expression is:

[0058]

[0059] wherein, S(t) represents the output value of the safety evaluation function at time t, R i represents the i-th navigation rule, Y k (t) i represents the value of the i-th dimension of the k-th predicted trajectory at time t, n represents the number of navigation rules, γ represents the exponent, τ represents the preset safety threshold, and min represents the minimum value;

[0060] Set an alarm trigger function A(t) based on the safety evaluation function and the safety threshold, and the expression is:

[0061]

[0062] This formula means that when the value of S(t) is greater than the preset safety threshold τ, A(t) returns 1, indicating that an alarm needs to be triggered; otherwise it returns 0, indicating that no alarm is triggered;

[0063] Design a safety suggestion function to provide safety suggestions SG k (t), and the expression is:

[0064]

[0065] wherein, v c represents the correction speed vector based on the deviation of the predicted trajectory from the navigation rules, a a represents the adjustment acceleration vector based on the deviation of the predicted trajectory from the navigation rules, α k and β k respectively represent the adjustment coefficients of the correction speed and the adjustment acceleration.

[0066] In a second aspect, the present invention provides a large model-based ship trajectory prediction system, including,

[0067] A collection module, responsible for collecting multimodal data from various sources, including unstructured text descriptions, historical trajectories, environmental conditions, radar images, and sonar signals, and performing preprocessing such as denoising and missing value filling on this data;

[0068] A feature module, responsible for extracting features from the preprocessed data, converting text data into numerical vectors, extracting the speed, direction, and position change rate of the trajectory, performing spectral analysis on environmental conditions, processing sensor data through a convolutional neural network, and finally fusing all features through weighted averaging to form environmental perception features;

[0069] A detection module, responsible for constructing an anomaly detection model, identifying and filtering out abnormal data, constructing a meta-learning model using the meta-learning mechanism and pure features, quickly adapting to the behavior patterns of specific ships through sub-models, and outputting a personalized prediction model;

[0070] A prediction module, responsible for combining the personalized prediction model with real-time data, integrating sensor data, environmental conditions, navigation data, and communication data using a composite function, performing ship trajectory prediction through a large model, and outputting the predicted trajectory and its confidence level;

[0071] A decision-making module, responsible for establishing a safety decision-making assistance system, comparing the predicted trajectory with navigation rules, calculating the safety level through a safety assessment function, triggering an alarm when the prediction result exceeds the safety threshold, and generating safety suggestions for correcting speed and adjusting acceleration based on the deviation degree to help the ship avoid risks.

[0072] In a third aspect, an embodiment of the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the large model-based ship trajectory prediction method described in the first aspect of the present invention is implemented.

[0073] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of the large model-based ship trajectory prediction method described in the first aspect of the present invention is implemented.

[0074] The beneficial effects of the present invention are as follows: The present invention collects multi-modal data through multiple channels, fuses environmental perception information using advanced feature extraction techniques, constructs an anomaly detection model to enhance data quality, and utilizes a meta-learning mechanism to quickly adapt to specific ship behaviors, ultimately achieving high-precision and real-time ship track prediction and safety decision-making assistance. The present invention focuses on solving the accuracy and safety problems of ship track prediction in complex marine environments, aiming to improve the intelligent level and safety of ship navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0076] Figure 1 It is a flowchart of the ship track prediction method based on a large model in Embodiment 1.

[0077] Figure 2 It is a flowchart for judging alarm triggering in the ship track prediction method based on a large model in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0078] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0079] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0080] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or selectively exclusive embodiment with other embodiments.

[0081] Embodiment 1, referring to Figure 1 and Figure 2 , is the first embodiment of the present invention. This embodiment provides a ship track prediction method based on a large model, including the following steps:

[0082] S1 Collect multi-modal data through multiple channels;

[0083] Collect unstructured text description data, historical track data, and environmental condition data from multiple sources;

[0084] Collect radar images and sonar signals through sensors;

[0085] Perform preprocessing on the collected data, including denoising and filling missing values;

[0086] Unstructured text description data, including navigation logs and chart annotations;

[0087] Historical track data contains GPS records;

[0088] Environmental condition data, including meteorological data and ocean current data.

[0089] S2 Extract features from multi-modal data respectively and output fused environmental perception features;

[0090] Use the BERT model to convert unstructured text descriptions into numerical vectors to obtain text feature vectors. The expression is:

[0091] T = BERT(wo);

[0092] Where T represents the text features obtained after being processed by the BERT model, wo represents the input text sequence, and the value range is from 1 to n, where n represents the number of words;

[0093] Use the sliding window technique to extract the change rates of the speed, direction, and position of the track to obtain the track feature vector. Set the sliding window size to W and the step size of the sliding window to s. The expression is:

[0094]

[0095] Where H represents the track features, i represents the starting time point of the window, represents the summation operation starting from time point i and ending at time point i + W - 1, v t represents the speed vector at time t, θ t represents the direction angle at time t, x t represents the position coordinate at time t, x t+s -x t represents the displacement vector of the position between time t and time t + s;

[0096] Use the discrete Fourier transform to perform spectral analysis on the environmental condition data. Set the meteorological data at time t to m t , and the ocean current data at time t to o t , and the expression for extracting environmental condition features is:

[0097] E = (DFT(m), DFT(o));

[0098] Among them, E represents environmental condition features, and DFT represents discrete Fourier transform;

[0099] A convolutional neural network CNN is used to extract features from sensor data. The two-dimensional matrix of sensor data is set as so, and the expression is:

[0100] S = CNN(so) = ReLU(K * so);

[0101] Among them, S represents sensor features, K represents the CNN filter kernel, ReLU represents the activation function, and * represents the convolution operation;

[0102] The extracted text features, track features, environmental condition features, and sensor features are fused by weighted averaging. The expression is:

[0103] Z = α T T + α H H + α E E + α S S;

[0104] α T + α H + α E + α S = 1;

[0105] Among them, Z represents the fused environmental perception features, and αT, αH, αE, and αS represent the weights of text features, track features, environmental condition features, and sensor features respectively. The sum of the weights of the four features is 1.

[0106] Based on the fused environmental perception features, S3 constructs an anomaly detection model. The fused environmental perception feature vector is input into the anomaly detection model, and abnormal data is identified and filtered through a causal inference algorithm, and pure features are output;

[0107] A multivariate Gaussian distribution, information entropy, and the latent causal variable model LCM in causal inference are used to construct an anomaly detection model. The expression of the anomaly detection model is:

[0108]

[0109] Among them, A(Z) represents the output of the anomaly detection model, represents the integral symbol, μ represents the mean vector, Σ represents the covariance matrix, N(Z; μ, Σ) represents the multivariate Gaussian distribution function, which is used to describe the distribution of normal data, ι(Z) represents the information filtering function, which is used to evaluate the information entropy and feature deviation degree of the feature vector Z, exp(-ι(Z)) represents the exponential function, and du is the differential element, which represents the infinitesimal change of the integral variable μ;

[0110] The expression of the information filtering function ι(Z) is as follows:

[0111]

[0112] Among them, p i represents the probability of the occurrence of the i-th feature, n represents the number of features in Z j log represents the natural logarithm function. When p i is smaller, the absolute value of log is larger, indicating that the possibility of the occurrence of this feature is lower and the amount of information carried is larger. λ represents the regularization parameter, which is used to balance the weights between the information entropy and the feature deviation degree. represents the summation of all terms from 1 to m, and m represents the number of features in j Z represents the actual observed value of the j-th feature in Z. represents the expected value of the j-th feature, and q represents the exponential operation, which is used to adjust the calculation method of the deviation degree.

[0113] S4 uses the meta-learning mechanism and pure features to build a meta-learning model, inputs the pure features into the meta-learning model, and outputs a personalized prediction model through a sub-model that quickly adapts to the specific ship behavior pattern;

[0114] It is set that the meta-learning structure consists of a shared basic model and sub-models for a series of tasks. The basic model can make a preliminary prediction for any ship track, while the sub-model is a personalized model for a specific ship and is used for fine-tuning the prediction;

[0115] Set each ship as k, and then build a sub-model to quickly adapt to the specific behavior pattern of ship k through the historical ship track data;

[0116] Based on the ship and the basic model P, build an adaptability metric function, and the expression is:

[0117]

[0118] Among them, R k (P, Z k ) represents the reliability metric of ship k under the pure feature Z k and the prediction P. P represents the prediction result of the basic model, and Z k represents the set of pure features of ship k, T represents the total number of time steps, and d k (t) represents the deviation at time t, and α k represents a constant, which is the ideal deviation of the model prediction, and β k represents another constant, which controls the non-linear degree of the influence of the deviation on the reliability metric. A larger β kmeans that the deviation has a greater penalty on the reliability metric, μ d represents the mean of the deviation, represents the variance of the deviation, N represents the normal distribution function, represents the sum of the probability densities of the normal distribution calculated over the entire deviation space;

[0119] Use the adaptability metric function as the objective function R k , and optimize the sub-model with the goal of minimizing R k , so that the sub-model can better adapt to the track pattern of ship k;

[0120] When the sub-model is optimized, that is, when the personalized prediction model is optimized, the output vector is R k (t), which is used to predict the track of ship k at time t;

[0121] The adaptability metric function R k has a value range of 0 to 1. When the value of R k is close to 1, it indicates that the adaptability of the base model P to ship k is high and the prediction effect is good;

[0122] When the value of R k is close to 0, it means that the adaptability is poor and the prediction error is large. By optimizing to make R k approach 1, the effectiveness of the personalized prediction model can be ensured;

[0123] The normal distribution function, the expression is:

[0124]

[0125] where e represents the base of the natural logarithm and π represents the pi.

[0126] S5 combines the personalized prediction model with real-time data, and predicts the ship track through the large model, and outputs the predicted track; the real-time data input vector at time t is set as X(t), which includes sensor data, environmental conditions, navigation data and communication data at time t;

[0127] Based on the real-time data X(t) and R k (t), and at the same time combining the pure feature set Z of ship k k Construct the composite function G, the expression is:

[0128]

[0129] where Y k (t) represents the predicted track of ship k at time t, P k (t) represents the output vector of the personalized prediction model of ship k at time t, w iFor adjusting the importance of each sensor data in the comprehensive prediction, ReLU represents the activation function, K represents the convolution kernel for processing sensor data, B represents the matrix for adjusting the output of the personalized prediction model, C represents the matrix for adjusting the pure features, and μ k represents the mean of the real-time data, and represents the variance of the real-time data;

[0130] The value range of Yk(t) is from 0 to 1. When the value of Yk(t) is close to 0, it indicates a greater uncertainty in the predicted track. When the value of Yk(t) is close to 1, it indicates a higher confidence in the predicted track.

[0131] S6 establishes a safety decision-making assistance system, compares the predicted track with the navigation rules, and when the prediction result exceeds the safety threshold, triggers an alarm and provides safety suggestions to the ship;

[0132] Set the set of navigation rules as R, including but not limited to the channel width, center line, minimum safety distance, maximum speed limit, and no-sailing area;

[0133] Construct a safety evaluation function based on the predicted track of the ship, and the expression is:

[0134]

[0135] where S(t) represents the output value of the safety evaluation function at time t, R i represents the i-th navigation rule, Y k (t) i represents the value of the i-th dimension of the k-th predicted track at time t, n represents the number of navigation rules, γ represents the exponent for adjusting the influence degree of the deviation value. A larger γ will make the influence of a larger deviation on the overall safety evaluation more significant, τ represents the preset safety threshold, and min represents the minimum value to ensure that the output of the safety evaluation function does not exceed the preset safety threshold τ;

[0136] Based on the safety evaluation function and the safety threshold, set the alarm trigger function A(t), and the expression is:

[0137]

[0138] This formula means that when the value of S(t) is greater than the preset safety threshold τ, then A(t) returns 1 and triggers an alarm; otherwise, it returns 0 and does not trigger an alarm;

[0139] Design a safety suggestion function to provide safety suggestions SG k (t), and the expression is:

[0140]

[0141] where v cDenote the corrected velocity vector based on the predicted track deviation from the navigation rules, a a Denote the adjusted acceleration vector based on the predicted track deviation from the navigation rules, α k and β k Respectively denote the adjustment coefficients of the corrected velocity and the adjusted acceleration, and adjust the intensity of the suggestions according to the ship characteristics;

[0142] The value range of S(t) is [0, τ], where 0 means fully meeting the safety standards, and τ means just reaching the safety threshold;

[0143] The value range of A(t) is from 0 to 1, where 0 means not exceeding the safety threshold and no alarm needs to be triggered; 1 means the predicted track exceeds the safety threshold and an alarm is triggered.

[0144] This embodiment also provides a ship track prediction system based on a large model, including:

[0145] A collection module, responsible for collecting multimodal data from various sources, including unstructured text descriptions, historical tracks, environmental conditions, radar images, and sonar signals, and performing preprocessing such as denoising and missing value filling on these data;

[0146] A feature module, responsible for extracting features from the preprocessed data, converting text data into numerical vectors, extracting the speed, direction, and position change rate of the track, performing spectral analysis on environmental conditions, performing convolutional neural network processing on sensor data, and finally fusing all features through weighted averaging to form environmental perception features;

[0147] A detection module, responsible for building an anomaly detection model, identifying and filtering out abnormal data, building a meta-learning model using the meta-learning mechanism and pure features, quickly adapting to the behavior patterns of specific ships through sub-models, and outputting personalized prediction models;

[0148] A prediction module, responsible for combining the personalized prediction model with real-time data, integrating sensor data, environmental conditions, navigation data, and communication data using a composite function, performing ship track prediction through a large model, and outputting the predicted track and its confidence level;

[0149] A decision-making module, responsible for establishing a safety decision-making assistance system, comparing the predicted track with the navigation rules, calculating the safety level through a safety assessment function, triggering an alarm when the prediction result exceeds the safety threshold, and generating safety suggestions for corrected velocity and adjusted acceleration based on the deviation degree to help the ship avoid risks.

[0150] This embodiment also provides a computer device applicable to the case of a ship trajectory prediction method based on a large model, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the ship trajectory prediction method based on a large model as proposed in the above embodiment.

[0151] The computer device can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0152] This embodiment also provides a storage medium with a computer program stored thereon. When the program is executed by a processor, it implements the ship trajectory prediction method based on a large model as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM for short), Electrically Erasable Programmable Read-Only Memory (EEPROM for short), Erasable Programmable Read Only Memory (EPROM for short), Programmable Read-Only Memory (PROM for short), Read-Only Memory (ROM for short), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0153] In summary, the present invention: collects multi-modal data through multiple channels, uses advanced feature extraction techniques to fuse environmental perception information, constructs an anomaly detection model to enhance data quality, and utilizes a meta-learning mechanism to quickly adapt to specific ship behaviors, ultimately achieving high-precision and real-time ship trajectory prediction and safety decision-making assistance. The present invention focuses on solving the accuracy and safety problems of ship trajectory prediction in complex marine environments, aiming to improve the intelligent level and safety of ship navigation.

[0154] Example 2, referring to Table 1, is the second example of the present invention. To further verify the advancement of the present invention, experimental simulation data of a ship trajectory prediction method based on a large model is given.

[0155] To verify the effectiveness and superiority of the ship trajectory prediction method based on a large model proposed by the present invention, a three-month sea trial was conducted. A total of six different types of ships were selected as test objects, including three frigates, two destroyers, and one supply ship. During the experiment, ship trajectory data, environmental condition data, and sensor data were collected every day. Among them, the environmental condition data included wind speed, wind direction, sea current speed and direction, and seawater temperature; the sensor data included radar images and sonar signals. In addition, historical trajectory data and unstructured text description data about each ship were also collected, such as the performance parameters of the ship and past navigation logs.

[0156] In the preprocessing stage, the collected data was denoised and missing values were filled. The unstructured text description data was converted into numerical vectors through the BERT model. The historical trajectory data was used to extract the change rates of speed, direction, and position through the sliding window technique. The environmental condition data was subjected to spectral analysis through the discrete Fourier transform. The sensor data was subjected to feature extraction through a convolutional neural network (CNN). Then, the above features were fused to form an environmental perception feature vector.

[0157] Next, an anomaly detection model was constructed to identify and filter out abnormal data through a causal inference algorithm to ensure the accuracy of the subsequent prediction model. Using the meta-learning mechanism, a meta-learning model was constructed based on the pure features to quickly adapt to the behavior patterns of each ship and output a personalized prediction model. Finally, the personalized prediction model was combined with real-time data, and the ship trajectory was predicted through a large model, and a safety decision-making assistance system was established to evaluate the safety of the predicted trajectory.

[0158] Specifically, as shown in Table 1:

[0159] Table 1 Experimental Record Table

[0160]

[0161] Through the analysis of the above data, it can be clearly seen that the ship trajectory prediction method of the present invention performs excellently in terms of prediction accuracy, with an average accuracy rate reaching 92%, far higher than about 80% of the traditional methods. The abnormal data filtering rate also reaches a satisfactory level, averaging 95%, effectively improving the data quality and reducing the impact of noise on the prediction model. The feature extraction efficiency averages 88%, indicating that this method can quickly and accurately extract useful information from multi-modal data and provide rich inputs for the prediction model.

[0162] The prediction delay remains at a low level, averaging 0.13 seconds, which is crucial for real-time prediction and ensures the immediate availability of the prediction results. The number of times of exceeding the safety threshold reflects the effectiveness of the safety decision-making assistance system. On average, each ship only has 3 times of exceeding the limit, and the safety recommendation execution rate is as high as 97%, indicating that this system can timely and accurately identify risks and provide effective avoidance measures.

[0163] In summary, the ship trajectory prediction method based on the large model proposed by the present invention shows obvious advantages in terms of prediction accuracy, data processing ability and safety. Compared with the prior art, it not only improves the accuracy and real-time performance of prediction, but also enhances the safety of ship navigation, having important practical value and innovative significance.

[0164] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A ship trajectory prediction method based on a large model, characterized in that: including, collecting multimodal data through multiple channels; the multimodal data includes unstructured text descriptions, historical tracks, environmental conditions, radar images, and sonar signals; extracting features from the multimodal data respectively and outputting fused environmental perception features; constructing an anomaly detection model based on the fused environmental perception features, inputting the fused environmental perception feature vectors into the anomaly detection model, identifying and filtering out abnormal data through a causal inference algorithm, and outputting pure features; constructing a meta-learning model using a meta-learning mechanism and pure features, inputting the pure features into the meta-learning model, and outputting a personalized prediction model through a sub-model that quickly adapts to the specific behavior patterns of a particular ship; combining the personalized prediction model with real-time data, and predicting the ship's track through a large model, and outputting a predicted track; establishing a safety decision-making assistance system, comparing the predicted track with navigation rules, and when the prediction result exceeds the safety threshold, triggering an alarm and providing safety suggestions to the ship; The specific steps for the output of pure features are as follows: Using a multivariate Gaussian distribution, information entropy, and combining the latent causal variable model LCM in causal inference to construct an anomaly detection model, and the expression of the anomaly detection model is: where A(Z) represents the output of the anomaly detection model, represents the integral symbol, μ represents the mean vector, Σ represents the covariance matrix, N(Z; μ, Σ) represents the multivariate Gaussian distribution function, ι(Z) represents the information filtering function, exp(-ι(Z)) represents the exponential function, du is the differential element, representing the infinitesimal change of the integral variable μ; The expression of the information filtering function ι(Z) is: Among them, p i represents the probability of the occurrence of the i-th feature, n represents the number of features in Z j where log represents the natural logarithm function. When p i is smaller, the absolute value of log is larger. λ represents the regularization parameter, denotes the summation over all terms from 1 to m, and m represents the number of features in Z, j Zj represents the actual observed value of the j-th feature in Z, Ej represents the expected value of the j-th feature, and q represents the exponential operation; The specific steps for the output of the personalized prediction model are as follows: Setting the meta-learning structure to consist of a shared base model and a series of task sub-models; Setting each ship as k, and then constructing a sub-model to quickly adapt to the specific behavior pattern of ship k through historical ship track data; Based on the ship and the base model P, constructing an adaptability metric function, and the expression is: Among them, R k (P, Z k ) represents the reliability measure of ship k under the pure feature Z k and the prediction P. P represents the prediction result of the basic model, and Z k represents the set of pure features of ship k, T represents the total number of time steps, and d k (t) represents the deviation at time t, α k represents a constant, β k represents another constant, and μ d represents the mean of the deviation, represents the variance of the deviation, N represents the normal distribution function, represents the sum of the probability densities of calculating the normal distribution within the entire deviation space; Using the adaptability metric function as the objective function R k , optimize the sub-model; When the sub-model is optimized, that is, when the personalized prediction model is optimized, the output vector is R k (t).

2. The method for predicting ship tracks based on a large model according to claim 1, wherein: The specific steps for collecting multimodal data through multiple channels are as follows: Collecting unstructured text description data, historical track data, and environmental condition data from multiple sources; Collecting radar images and sonar signals through sensors; Performing preprocessing of denoising and missing value filling on the collected data.

3. The method for predicting ship trajectories based on large models according to claim 2, characterized in that: The specific steps for extracting features from the multimodal data respectively and outputting fused environmental perception features are as follows: Using the BERT model to convert the unstructured text description into a numerical vector to obtain a text feature vector, and the expression is: T = BERT(wo); where T represents the text feature obtained after being processed by the BERT model, and wo represents the input text sequence; Using a sliding window technique to extract the change rates of the speed, direction, and position of the track to obtain a track feature vector, setting the sliding window size as W and the step size of the sliding window as s, and the expression is: Among them, H represents the track feature, and i represents the starting time point of the window. represents the summation operation starting from time point i and ending at time point i + W - 1, v t represents the velocity vector at time t, θ t represents the direction angle at time t, x t represents the position coordinate at time t, x t+s -x t represents the displacement vector of the position between time t and time t + s; The environmental condition data is subjected to spectral analysis using the discrete Fourier transform, and the meteorological data at time t is set as m t , and the ocean current data at time t is o t , and the environmental condition feature expression is extracted as: E = (DFT(m), DFT(o)); where E represents the environmental condition feature, and DFT represents the discrete Fourier transform; Using a convolutional neural network CNN to extract features from the sensor data, setting the two-dimensional matrix of the sensor data as so, and the expression is: S = CNN(so) = ReLU(K * so); where S represents the sensor feature, K represents the CNN filter kernel, ReLU represents the activation function, and * represents the convolution operation; Using a weighted average method to fuse the extracted text features, track features, environmental condition features, and sensor features, and the expression is: Z = α T T + α H H + α E E + α S S; α T +α H +α E +α S ＝1； Among them, Z represents the fused environmental perception feature, α T , α H , α E and α S respectively represent the weights of text features, track features, environmental condition features, and sensor features.

4. The method for predicting ship tracks based on a large model according to claim 3, wherein: Combining the personalized prediction model with real-time data, and performing ship trajectory prediction through a large model to output a predicted trajectory. The specific steps are as follows: Set the real-time data input vector at time t as X(t); Based on real-time data X(t) and R k (t), and at the same time combining the pure feature set Z of the ship k k Construct a composite function G, and the expression is: Among them, Y k (t) represents the predicted track of ship k at time t, P k (t) represents the output vector of the personalized prediction model of ship k at time t, w i is used to adjust the importance of each sensor data in the comprehensive prediction, ReLU represents the activation function, K represents the convolution kernel, B represents the matrix for adjusting the output of the personalized prediction model, C represents the matrix for adjusting the pure features, μ k represents the mean of the real-time data, represents the variance of the real-time data.

5. The method for predicting the ship's track based on a large model according to claim 4, wherein: The establishment of a safety decision-making assistance system, which compares the predicted trajectory with the navigation rules. When the prediction result exceeds the safety threshold, an alarm is triggered and safety suggestions are provided to the ship. The specific steps are as follows: Set the set of navigation rules as R; Construct a safety assessment function based on the predicted ship trajectory, and the expression is: Among them, S(t) represents the output value of the security assessment function at time t, and R i represents the i-th navigation rule, and Y k (t) i represents the value of the i-th dimension of the k-th predicted trajectory at time t, n represents the number of navigation rules, γ represents the exponent, τ represents the preset security threshold, and min represents the minimum value; Based on the safety assessment function and the safety threshold, set the alarm trigger function A(t), and the expression is: This formula means that when the value of S(t) is greater than the preset safety threshold τ, A(t) returns 1, indicating that an alarm needs to be triggered; otherwise, it returns 0, indicating that no alarm is triggered; Design a security advice function to provide security advice SG k (t), and the expression is: Among them, v c represents the corrected velocity vector based on the deviation of the predicted track from the navigation rules, a a represents the adjusted acceleration vector based on the deviation of the predicted track from the navigation rules, α k and β k respectively represent the adjustment coefficients of the corrected velocity and the adjusted acceleration.

6. A ship trajectory prediction system based on a large model, based on the ship trajectory prediction method based on a large model according to any one of claims 1 to 5, characterized in that, Including: A collection module, responsible for collecting multimodal data through multiple channels; the multimodal data includes unstructured text descriptions, historical trajectories, environmental conditions, radar images, and sonar signals; A feature module, responsible for extracting features from multimodal data respectively and outputting fused environmental perception features; A detection module, responsible for constructing an anomaly detection model, identifying and filtering out abnormal data, using a meta-learning mechanism and pure features to construct a meta-learning model, and quickly adapting to the behavior patterns of specific ships through sub-models, and outputting a personalized prediction model; Output pure features. The specific steps are as follows: Use a multivariate Gaussian distribution, information entropy, and combine the latent causal variable model LCM in causal inference to construct an anomaly detection model. The expression of the anomaly detection model is: where A(Z) represents the output of the anomaly detection model, represents the integral symbol, μ represents the mean vector, Σ represents the covariance matrix, N(Z; μ, Σ) represents the multivariate Gaussian distribution function, ι(Z) represents the information filtering function, exp(-ι(Z)) represents the exponential function, du is the differential element, representing the infinitesimal change of the integral variable μ; The expression of the information filtering function ι(Z) is: Among them, p i represents the probability of the i-th feature occurring, n represents the number of features in Z j in, log represents the natural logarithm function. When p i is smaller, the absolute value of log is larger. λ represents the regularization parameter, represents the sum over all terms from 1 to m. m represents the number of features in Z j represents the actual observed value of the j-th feature in Z, represents the expected value of the j-th feature, and q represents the exponential operation; Output a personalized prediction model. The specific steps are as follows: Set the meta-learning structure to consist of a shared base model and sub-models for a series of tasks; Set each ship as k, and then construct a sub-model to quickly adapt to the specific behavior pattern of ship k through historical ship trajectory data; Based on the ship and the base model P, construct an adaptability metric function, and the expression is: Among them, R k (P, Z k ) represents the reliability metric of ship k under the pure feature Z k and the prediction P. P represents the prediction result of the basic model, and Z k represents the set of pure features of ship k. T represents the total number of time steps, and d k (t) represents the deviation at time t. α k represents a constant, and β k represents another constant. μ d represents the mean of the deviation, represents the variance of the deviation, and N represents the normal distribution function. represents calculating the sum of the probability densities of the normal distribution within the entire deviation space; Using the adaptability metric function as the objective function R k , optimize the sub-model; When the sub-model is optimized, that is, when the personalized prediction model is optimized, the output vector is R k (t); A prediction module, responsible for combining the personalized prediction model with real-time data, performing ship trajectory prediction through a large model, and outputting a predicted trajectory; A decision-making module, responsible for establishing a safety decision-making assistance system, comparing the predicted trajectory with the navigation rules, and when the prediction result exceeds the safety threshold, triggering an alarm and providing safety suggestions to the ship.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the large model-based ship trajectory prediction method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the large model-based ship trajectory prediction method according to any one of claims 1 to 5.

Images