A multi-target association fusion method based on rules and prediction models

Abstract

The application relates to a multi-target association fusion method based on rules and a prediction model, and belongs to the field of target data association information fusion; prior knowledge rules based on a knowledge base are constructed; a set of data points required for model training is obtained by adopting a Douglas-Puke algorithm, the data amount is simplified while the geometric shape is reserved, and trajectory feature extraction is realized; a ship track prediction model based on a time convolution network (TCN) is constructed and is recorded as a TCN ship track prediction model; the extracted trajectory features are taken as inputs of the TCN ship track prediction model, and a TCN network prediction result is output; a target association method based on the prior knowledge rules and the TCN network prediction model is established, and targets are determined; the application realizes improvement of the accuracy of multi-target association fusion and ensures the real-time performance of the association fusion.

Description

Technical Field

[0001] This invention belongs to the field of target data association information fusion, and relates to a multi-target association fusion method based on rules and prediction models. Background Technology

[0002] Target association fusion technology stems from the urgent need of modern intelligent systems for efficient collaboration of multi-source heterogeneous data. With the widespread application of multiple sensors such as radar, infrared, and cameras in fields like autonomous driving, military surveillance, and robot navigation, single sensors are susceptible to interference from environmental noise, limited field of view, or data redundancy, making it difficult to achieve continuous and accurate target perception. This technology integrates observation data from different sensors through spatiotemporal alignment, probabilistic association, and deep learning, solving challenges such as time synchronization, spatial matching, and uncertainty elimination of cross-modal information. This improves the robustness of target recognition in complex scenarios, providing reliable data support for intelligent decision-making, and has become one of the core research directions in the field of multi-sensor information fusion.

[0003] Trajectory prediction technology aims to infer a target's possible future movement path by analyzing its historical motion state and environmental information. It incorporates various methods such as temporal modeling, machine learning, and probabilistic reasoning, aiming to address challenges such as motion uncertainty, interactive behavior modeling, and multimodal trajectory generation in complex dynamic scenarios. Trajectory prediction technology is widely used in fields such as autonomous driving obstacle avoidance, robot navigation, and traffic flow optimization, improving system safety and decision-making efficiency by predicting target intentions.

[0004] Deep learning, with its powerful multi-scale feature extraction and generalization capabilities, has become a key technology in trajectory prediction. By designing targeted deep neural network models, it is possible to effectively capture the long-term temporal dependencies, multimodal interaction features, and dynamic environmental constraints of target trajectories, significantly improving prediction accuracy. However, existing models often rely on complex network structures, leading to a surge in the number of parameters and increased computational complexity, making it difficult to meet the inference speed requirements in real-time scenarios. Therefore, there is an urgent need to find a balance between lightweight models and predictive performance. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a multi-objective association fusion method based on rules and prediction models, so as to improve the accuracy of multi-objective association fusion and ensure the real-time performance of association fusion.

[0006] The solution of the present invention is: A multi-objective association fusion method based on rules and prediction models includes: Construct prior knowledge rules based on a knowledge base; The Douglas-Puk algorithm is used to obtain the set of data points required for model training, which reduces the amount of data while preserving the geometric shape, and achieves trajectory feature extraction. A ship trajectory prediction model based on Temporal Convolutional Network (TCN) is constructed, denoted as TCN ship trajectory prediction model; the extracted trajectory features are used as input to TCN ship trajectory prediction model, and the TCN network prediction results are output. A target association method based on prior knowledge rules and TCN network prediction models is established to determine the target; Data correction based on fuzzy logic systems improves the accuracy of target determination.

[0007] In the aforementioned multi-objective association fusion method based on rules and prediction models, the prior knowledge rules are as follows: When historical data for the target is insufficient, the system retrieves performance knowledge of that type of ship from the knowledge base, and obtains the possible range of the ship based on this performance knowledge. The ship's center is then used as the predicted target's location. The system iterates through the observed data and the predicted ship's position data, calculating the distance between each point. A distance threshold is set to determine whether the association between two points is successful. If the calculated distance is greater than the threshold, the point is considered impossible to associate; if the distance is less than the threshold, the point is considered successfully associated. The established locations are recorded in the association database for subsequent fine-grained track association.

[0008] In the aforementioned multi-objective association fusion method based on rules and prediction models, the knowledge base is implemented as a database storing known information, including the basic types of ships and the specific performance parameters of each type of ship.

[0009] In the aforementioned multi-objective association fusion method based on rules and prediction models, the trajectory feature extraction method is as follows: Connect the first and last points of the trajectory with a straight line; measure the distance from all points to this line, and find the point with the maximum distance value. The point; will Compare with the set threshold t_res_old, if If the value is less than the threshold t_res_old, then other intermediate points of the trajectory will no longer be segmented, and feature extraction will be completed; if... If the value is greater than or equal to the threshold t_res_old, then retain it. The corresponding data points; Using the location as a boundary, the trajectory is divided into two segments. The above steps are repeated for each segment to continue the judgment until feature extraction is completed.

[0010] In the aforementioned multi-objective association fusion method based on rules and prediction models, the method for establishing a ship trajectory prediction model based on Temporal Convolutional Network (TCN) is as follows: S1. Let the input sequence of the ship trajectory prediction model of the Temporal Convolutional Network (TCN) be... ;in, The sequence number of a certain feature of a ship's track; The number of sequences corresponding to each feature; Represents ship track-related characteristics including longitude, latitude, speed, heading, and time interval; the output sequence is set as follows. ,in, Indicates longitude or latitude; S2, Calculate the causal convolutional layer under... Predicted values ​​of ship tracks at the location; S3. Based on the calculation of the causal convolutional filter, a dilated convolutional layer is further used to increase the field of view of the convolution, and the calculation is performed on... Predicted ship tracks at the location ; S4. Establish the calculation formula for the residual connection layer; S5. The TCN ship trajectory prediction model consists of causal convolutional layers, hollow convolutional layers, and residual connection layers.

[0011] In the aforementioned multi-objective association fusion method based on rule and prediction models, in step S2, the causal convolutional filter is set as... ; exist Predicted ship tracks at the location for:

[0012] In the formula, The size of the kernel; It is a convolution mapping function; This represents the current layer number of the convolutional network. The nth sequence number corresponding to each feature.

[0013] In the aforementioned multi-objective association fusion method based on rules and prediction models, in step S3, the predicted value... for:

[0014] In the formula, It is the expansion factor; This represents the size of the dilated convolution filter; It represents the direction of history.

[0015] In the aforementioned multi-objective association fusion method based on rules and prediction models, in step S4, the residual connection layer... for:

[0016] In the formula, For activation functions; It is the residual function; The input sequence is .

[0017] In the aforementioned multi-objective association fusion method based on rules and prediction models, the method for establishing the objective association method is as follows: In the early stages of observation, when historical data of the target is insufficient, the knowledge base is used to calculate the possible range of the target based on the performance knowledge of this type of ship, and the center of the range is used as the predicted location of the target. The distance calculation result is subtracted from the activity range distance before comparison. The distance is calculated one by one by traversing the observation data and the predicted ship position data, and a distance threshold is set to determine whether the association between two points is successful. Points with a calculated distance greater than the threshold are judged as points that cannot be associated, and points with a distance less than the threshold are judged as points that have been successfully associated. The locations of the points with established associations are recorded in the association database for subsequent fine association of tracks. In the later stages of observation, the target association is based on the prediction results of the constructed TCN model. The TCN prediction model determines the position of the target at the same time as the observation, and uses Euclidean distance calculation as a measure of the degree of target association. The distance between the target observation point and the prediction point data is calculated and compared to determine the target. When a target to be associated matches multiple known targets, the target with the smallest distance is taken as its final target; when no matching target is found, the target to be associated is regarded as a new target.

[0018] In the aforementioned multi-objective association fusion method based on rules and prediction models, the Euclidean distance calculation specifically involves:

[0019] In the formula, , These are vectors composed of the latitude and longitude coordinates of the observation point to be associated and the prediction point to be selected, respectively. The longitude of the observation point to be associated; To select the longitude of the target prediction point; The latitude of the observation point to be associated; To select the latitude of the target prediction point.

[0020] The advantages of this invention compared to the prior art are: (1) In the case of insufficient historical data, the present invention will complete the association and fusion task based on pre-set rules; in the case of sufficient historical data, the ship's historical trajectory is processed to extract the ship trajectory features. (2) The present invention establishes a ship trajectory prediction model and takes the ship trajectory features as input into the ship trajectory prediction model. The output result is the ship position predicted by the model. (3) The present invention uses the ship position predicted by the model as the basis for association fusion and matches it with the observation data according to the rules to complete the multi-objective association fusion task. Attached Figure Description

[0021] Figure 1 This is a flowchart of the multi-target correlation fusion process of the present invention; Figure 2 A basic flowchart for establishing the TCN ship trajectory prediction model of this invention; Figure 3 This is a flowchart illustrating the target association process of the present invention. Figure 4 This invention uses the weight function curve as the input membership function curve as a schematic diagram. Figure 5 This is a schematic diagram of the input membership function curve as a delta function curve for this invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the embodiments.

[0023] This invention provides a multi-objective association fusion method based on rules and a prediction model. When historical data is insufficient, the association fusion task is completed based on pre-set rules. When historical data is sufficient, the historical trajectories of ships are processed to extract ship trajectory features. These features are then input into a ship trajectory prediction model, and the output is the predicted ship position.

[0024] Multi-objective association fusion methods based on rules and prediction models, such as Figure 1 As shown, the specific steps include the following: Construct prior knowledge rules based on a knowledge base.

[0025] The knowledge base is a database that stores known information, including the basic types of ships and the specific performance parameters of each type of ship.

[0026] The rules of prior knowledge are: When historical data for the target is insufficient, the system retrieves performance knowledge of that type of ship from the knowledge base, and obtains the possible range of the ship based on this performance knowledge. The ship's center is then used as the predicted target's location. The system iterates through the observed data and the predicted ship's position data, calculating the distance between each point. A distance threshold is set to determine whether the association between two points is successful. If the calculated distance is greater than the threshold, the point is considered impossible to associate; if the distance is less than the threshold, the point is considered successfully associated. The established locations are recorded in the association database for subsequent fine-grained track association.

[0027] The Douglas-Puk algorithm is used to obtain the set of data points required for model training, which reduces the amount of data while preserving the geometry and achieves trajectory feature extraction.

[0028] The method for trajectory feature extraction is as follows: Connect the first and last points of the trajectory with a straight line; measure the distance from all points to this line, and find the point with the maximum distance value. The point; will Compare with the set threshold t_res_old, if If the value is less than the threshold t_res_old, then other intermediate points of the trajectory will no longer be segmented, and feature extraction will be completed; if... If the value is greater than or equal to the threshold t_res_old, then retain it. The corresponding data points; Using the location as a boundary, the trajectory is divided into two segments. The above steps are repeated for each segment to continue the judgment until feature extraction is completed.

[0029] A ship trajectory prediction model based on Temporal Convolutional Network (TCN) is constructed, denoted as TCN ship trajectory prediction model; the extracted trajectory features are used as input to TCN ship trajectory prediction model, and the TCN network prediction results are output.

[0030] like Figure 2 As shown, the method for establishing a ship trajectory prediction model based on Temporal Convolutional Network (TCN) is as follows: S1. Let the input sequence of the ship trajectory prediction model of the Temporal Convolutional Network (TCN) be... ;in, The sequence number of a certain feature of a ship's track; The number of sequences corresponding to each feature; Represents ship track-related characteristics including longitude, latitude, speed, heading, and time interval; the output sequence is set as follows. ,in, Indicates longitude or latitude.

[0031] S2, Calculate the causal convolutional layer under... The predicted value of the ship's trajectory at that location.

[0032] In S2, the causal convolution filter is set as ; exist Predicted ship tracks at the location for:

[0033] In the formula, This refers to the size of the kernel.

[0034] This is the convolution mapping function.

[0035] This represents the current layer number of the convolutional network.

[0036] The nth sequence number corresponding to each feature.

[0037] S3. Based on the calculation of the causal convolutional filter, a dilated convolutional layer is further used to increase the field of view of the convolution, and the calculation is performed on... Predicted ship tracks at the location .

[0038] In S3, the predicted value for:

[0039] In the formula, It is the expansion factor; This represents the size of the dilated convolution filter; It represents the direction of history.

[0040] S4. Establish the calculation formula for the residual connection layer.

[0041] In S4, the residual connection layer for:

[0042] In the formula, For activation functions; It is the residual function; The input sequence is .

[0043] S5. The TCN ship trajectory prediction model consists of causal convolutional layers, hollow convolutional layers, and residual connection layers.

[0044] A target association method based on prior knowledge rules and TCN network prediction model is established to determine the target.

[0045] The method for establishing the target association method is as follows: In the early stages of observation, when historical data on the target is insufficient, a knowledge base is used to calculate the possible range of the target based on the performance knowledge of this type of ship, with the center of the range as the predicted target location. The distance calculation result is compared with the distance of the activity range before comparison. The distance is calculated one by one by traversing the observation data and the predicted ship position data, and a distance threshold is set to determine whether the association between two points is successful. Points with a calculated distance greater than the threshold are judged as points that cannot be associated, and points with a distance less than the threshold are judged as points that have been successfully associated. The locations with established associations are recorded in the association database for subsequent fine association of tracks.

[0046] In the later stages of observation, the target association is based on the prediction results of the constructed TCN model. The TCN prediction model determines the position of the target at the same time as the observation, and uses Euclidean distance calculation as a measure of the degree of target association. The distance between the target observation point and the prediction point data is calculated and compared to determine the target.

[0047] The Euclidean distance is calculated as follows:

[0048] In the formula, , These are vectors composed of the latitude and longitude coordinates of the observation point to be associated and the prediction point to be selected, respectively. The longitude of the observation point to be associated; To select the longitude of the target prediction point; The latitude of the observation point to be associated; To select the latitude of the target prediction point.

[0049] When a target to be associated matches multiple known targets, the target with the smallest distance is chosen as its final associated target; if no matching target is found, the target to be associated is considered a new target. The target association process is as follows: Figure 3 As shown.

[0050] Data correction based on fuzzy logic systems improves the accuracy of target determination.

[0051] To further improve the reliability of the correlation fusion results, a gradual attenuation factor α is established. The observed data are first corrected using the attenuated error before fusion. Using fuzzy sets to define the parameter range, rather than strict numerical values, offers greater flexibility. Furthermore, using membership functions to handle fuzzy information avoids simplifying the problem to binary logic, providing better tolerance to noise in the input data. Compared to other complex machine learning methods, fuzzy logic methods typically offer better interpretability.

[0052] First, the fuzzer will input the precise set... The fuzzy input set is mapped using membership functions. The Gaussian, sigmoid, and bell-shaped membership functions used in subsequent algorithms are expressed as follows:

[0053] in The standard deviation is α. In the problem of choosing α, the input of the fuzzy logic system includes two dimensions: weight and delta. Weight represents the proportion of each dimension's error to the total error. Here, a Gaussian membership function is used, and there are a total of 3 input fuzzy sets. Delta represents the difference between the position at the previous time step and the position at the current time step, which simultaneously has high and low orbit information. Here, a Sigmoid membership function and a bell-shaped membership function are used, and there are a total of 3 input fuzzy sets.

[0054] Two-dimensional membership functions such as Figure 4 , Figure 5 As shown.

[0055] Input the multi-target data to be correlated and fused, along with historical trajectories, to obtain the correlation and fusion results.

[0056] The multi-objective data to be correlated and fused, along with historical trajectory data, are input into the model through the inference interface. The model returns the prediction results, and further correlation and fusion of multiple objects can be carried out based on the results.

[0057] Example This invention provides a multi-objective association fusion method based on rules and prediction models to improve the accuracy of multi-objective association fusion and ensure its real-time performance. The specific steps are as follows: Step 1: Constructing Prior Knowledge Rules The construction of the ship prior knowledge database requires obtaining relevant data, and the methods for obtaining this information include, but are not limited to: Public sources: international ship registration documents, technical specifications, maritime reports, industry standards, etc.

[0058] Intelligence analysis: Intelligence gathered through satellite imagery, aerial reconnaissance, underwater surveillance, and other means.

[0059] Historical data: past navigation records, incident files, and intelligence shared with other national organizations.

[0060] Building a priori database for ships is an ongoing and complex process that requires verifying the accuracy of the data, ensuring the timeliness of the information, and continuously updating the database to reflect the latest changes in ship technology.

[0061] To maintain the high quality and reliability of the database, a regular audit and update mechanism is required to filter new information, verify old information, and correct errors or outdated entries.

[0062] Specifically, the rule strategy for association fusion is to use the speed of each ship at its last observed position × (current time - ship's last observed time) as the ship's possible activity range, and then determine which ship's activity range the ship is in.

[0063] Step 2 Trajectory Feature Extraction The specific steps of the algorithm are as follows: This invention uses the Douglas-Puk algorithm to process the ensemble data points required for model training, reducing the amount of data while preserving the geometric shape. The basic idea of ​​the algorithm is: connect the first and last points of the trajectory with a straight line, calculate the distance metric from all points to this line, and find the point with the maximum distance value. The point will Compare with the set threshold t_res_old, if If the value is less than the threshold t_res_old, then other intermediate points of the trajectory will no longer be segmented, and feature extraction is complete; if If the value is greater than or equal to the threshold t_res_old, then retain it. Using the corresponding data points as a boundary, the trajectory is divided into two segments, and the above steps are repeated for these two segments.

[0064] (1) Connect the beginning and end of the trajectory to be processed with a straight line and set this line as the initial simplified line; (2) Find the distances between all points and the simplified line, and find the point with the largest distance from the simplified line. ; (3) Compare with the threshold t_res_old, if If <t_res_old, then all midpoints on this curve are deleted. If ≥ t_res_old, then the line segment is divided into two parts; (4) Repeat steps (1) to (3) until the distance from all points to the simplified line is less than the threshold t_res_old.

[0065] Step 3: Trajectory Prediction Model Design Convolutional operations in Temporal Convolutional Networks (TCNs) offer the advantage of parallel computation, processing information from multiple locations simultaneously and thus reducing training time. Furthermore, TCNs utilize residual connections to mitigate the vanishing gradient problem through direct information transfer across layers. Therefore, this invention constructs a ship trajectory prediction model based on Temporal Convolutional Networks (TCNs). Let the input sequence of the TCN model be... ,in, The sequence number of a certain feature of a ship's track; The number of sequences corresponding to each feature. This represents ship track characteristics such as longitude, latitude, speed, heading, and time interval. The output sequence is... ,in, Indicates longitude or latitude.

[0066] Among them, the causal convolutional network is a unidirectional structure, that is, for The predicted trajectory value at any given time depends only on The values ​​of relevant factors at and before time point. Let the filter for causal convolution be... Therefore, in The predicted value of the ship's trajectory at that location is:

[0067] In the formula, The size of the kernel; It is a convolution mapping function; This represents the current layer number of the convolutional network. It is the nth sequence number corresponding to each feature.

[0068] Causal convolution relies on prior data for prediction, leading to a large network depth when the data is long, resulting in complex network structure, poor real-time performance, and difficulty in ensuring real-time performance during ship navigation. Therefore, dilated convolution is used to broaden the convolutional scope. The calculation of dilated convolution is as follows:

[0069] In the formula, It is the expansion factor; The size of the filter; It represents the direction of history.

[0070] The relationship between the receptive field of convolution and hyperparameters such as the dilation factor and kernel size is as follows:

[0071] In the formula, It is an expanding group; Indicates the number of layers below the new layer; Indicates the number of convolutional layers; This indicates the receptive field of view of the convolution.

[0072] Causal convolution and dilated convolution can achieve good prediction results in shallow neural networks. However, in ship trajectory prediction, a large amount of trajectory-related data is generated, and shallow neural networks are insufficient to fully learn the patterns in this vast amount of data. Therefore, deep learning networks are needed to fully learn from historical trajectory data. However, as the network deepens, it can lead to network degradation. Therefore, a residual network module is introduced to optimize the network model. The residual network uses a "skip connection" operation that skips one or more layers to avoid the performance degradation caused by deeper networks. The calculation formula for the residual module is as follows:

[0073] In the formula, For activation functions; It is the residual function; The input sequence is .

[0074] Finally, the input trajectory features are processed through causal convolutional layers, dilated convolutional layers, and residual connection layers, and the output TCN network prediction results are processed through 1×1 convolutional layers.

[0075] The specific process of the constructed ship track TCN prediction model is as follows: the track data (including timestamp, longitude, latitude, speed, and heading) is preprocessed, mainly by filling in missing values, and the continuous time series data is divided into sliding windows of fixed length as model input; in the programming implementation, a network architecture containing multi-layer causal convolution and dilated convolution is implemented using PyTorch, and the receptive field of the model is expanded by introducing dilated convolutions with different dilation rates, thereby capturing feature dependencies over a longer time span; in order to overcome the gradient vanishing or exploding problem in deep networks, residual connections are added to the model, allowing information to be directly transmitted across several layers, thereby improving the stability of training.

[0076] Step 4: Trajectory Prediction Model Training To train the ship trajectory prediction model, this embodiment selects and processes a publicly available AIS dataset. Specifically, the discrete data in the dataset is organized into complete trajectory data based on time and ship number. All acquired AIS datasets are processed using the above method, and the processing results form the ship trajectory prediction dataset.

[0077] To complete the training of the ship trajectory prediction model, the above dataset was divided into a training set and a test set, with a specific dataset ratio of 7:3. The training set was used to train the ship trajectory prediction model, while the test set was used to test the effectiveness of the ship trajectory prediction model.

[0078] This embodiment uses PyTorch as the deep learning framework, specifically version 1.11.0. Correspondingly, Python 3.10 is used as the Python programming language interpreter to run the embodiment's code. The embodiment utilizes a GPU for hardware acceleration; the GPU used is an NVIDIA GeForce RTX 4090Ti.

[0079] The training of the example model does not use a pre-trained model. The network model parameters are set to random values, the mean squared error is used as the loss function to measure the prediction error, and the Adam optimization algorithm is used for weight updates.

[0080] The initial learning rate was set to 0.0001, and the number of training epochs was set to 200. The training model results were saved in pt format.

[0081] Step 5: Target association based on rules and predictions Due to incomplete target observations, insufficient historical data in the initial observation phase can lead to difficulties in determining target associations. Therefore, a solution is proposed to address this issue. When historical target data is insufficient, after obtaining the maximum operational radius of each ship using a knowledge base, geographic information system (GIS) technology is employed to draw a circular area with the corresponding radius centered on the current location of each ship on a map. This circular area represents the theoretical maximum reachable range of that ship within a specific time period.

[0082] The global nearest neighbor (GNN) data association method is used as the fusion method in the rule matching stage. GNN is a single-hypothesis association method. Its main idea is that for each new set of measurements, the globally nearest measurement is assigned to the existing track, and a new hypothetical track is created for measurements that are not assigned. This algorithm first needs to set a tracking gate, the center of which is the predicted position of the currently tracked target. A range is set based on the probability of accepting the correct measurement; measurements within this range become candidate measurements, thus limiting the number of measurements participating in the association decision. Let the measurement value be... Then the candidate measurement needs to satisfy the following formula, where S represents the information covariance matrix and T represents the design threshold:

[0083] If there is no correlation conflict between the candidate measurements, that is, there is only one candidate measurement in the gate, then the correlation is successful. However, if there are multiple measurements in the gate of a certain track or a measurement is located in the gate of multiple tracks, a conflict will occur. In this case, it is necessary to select the candidate measurement with the smallest generalized statistical distance as the correlation measurement and use the measurement to update the state of the track.

[0084] The commonly used definition of generalized statistical distance is as follows:

[0085] In the later stages of observation, target association is based on the trajectory predicted by the TCN model constructed in step 2 according to the target's historical location results. The TCN prediction model determines the target's position at the same time as the observation, and Euclidean distance is used as a measure of the degree of target association. The global nearest neighbor algorithm is used to calculate and compare the distance between the target's observed points and predicted points. Euclidean distance is used to measure the actual distance between two vectors in space. It is not affected by the vector dimension or the number of samples. When all elements in the vector have the same dimensions, Euclidean distance has a good effect on judgment. The Euclidean distance is calculated as follows:

[0086] In the formula, , These are vectors composed of the latitude and longitude coordinates of the observation point to be associated and the prediction point to be selected.

[0087] If the target to be associated matches multiple known targets, the one with the smallest distance is taken as its final target; if the target to be associated does not find a matching target, it is regarded as a new target.

[0088] The target association process is as follows: Figure 3 As shown.

[0089] Step 6: Data Correction Based on Fuzzy Logic System To further improve the reliability of the correlation fusion results, a gradual attenuation factor α is established. The observed data are first corrected using the attenuated error before fusion. Using fuzzy sets to define the parameter range, rather than strict numerical values, offers greater flexibility. Furthermore, using membership functions to handle fuzzy information avoids simplifying the problem to binary logic, providing better tolerance to noise in the input data. Compared to other complex machine learning methods, fuzzy logic methods typically offer better interpretability.

[0090] First, the fuzzer will input the precise set... The fuzzy input set is mapped using membership functions. The Gaussian, sigmoid, and bell-shaped membership functions used in subsequent algorithms are expressed as follows:

[0091] in The standard deviation is α. In the problem of choosing α, the input of the fuzzy logic system includes two dimensions: weight and delta. Weight represents the proportion of each dimension's error to the total error. Here, a Gaussian membership function is used, and there are a total of 3 input fuzzy sets. Delta represents the difference between the position at the previous time step and the position at the current time step, which simultaneously has high and low orbit information. Here, a Sigmoid membership function and a bell-shaped membership function are used, and there are a total of 3 input fuzzy sets.

[0092] The two-dimensional membership function is shown in the appendix. Figure 4 , Figure 5 As shown.

[0093] Step 7: Input the multi-target data to be correlated and fused, and historical trajectories, to obtain the correlation and fusion results.

[0094] The multi-objective data to be correlated and fused, along with historical trajectory data, are input into the model through the inference interface. The model returns the prediction results, and further correlation and fusion of multiple objects can be carried out based on the results.

[0095] This invention provides a multi-objective association fusion method based on rules and a prediction model. In this method, when historical data is insufficient, the association fusion task is completed based on pre-set rules; when historical data is sufficient, the historical trajectories of ships are processed to extract ship trajectory features. These features are input into a ship trajectory prediction model, and the model's predicted ship position is the output. The predicted ship position serves as the basis for association fusion, and is matched with the observed data according to the rules to complete the multi-objective association fusion task.

[0096] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A multi-objective association fusion method based on rules and prediction models, characterized in that: include: Construct prior knowledge rules based on a knowledge base; The Douglas-Puk algorithm is used to obtain the set of data points required for model training, which reduces the amount of data while preserving the geometric shape, and achieves trajectory feature extraction. A ship trajectory prediction model based on Temporal Convolutional Network (TCN) is constructed, denoted as TCN ship trajectory prediction model; the extracted trajectory features are used as input to TCN ship trajectory prediction model, and the TCN network prediction results are output. A target association method based on prior knowledge rules and TCN network prediction model is established to determine the target.

2. The multi-objective association fusion method based on rules and prediction models according to claim 1, characterized in that: The prior knowledge rule is as follows: When historical data for the target is insufficient, the performance knowledge of that type of ship is obtained from the knowledge base, and the possible range of the ship is obtained based on the performance knowledge of that type of ship; and the center of the ship is used as the predicted location of the target. The system iterates through the observed data and the predicted ship position data, calculating the distance between each point. A distance threshold is set to determine whether the association between two points is successful. If the calculated distance is greater than the threshold, the points are considered impossible to associate; if the distance is less than the threshold, the points are considered successfully associated. The established location records are added to the association database for subsequent detailed tracking.

3. The multi-objective association fusion method based on rules and prediction models according to claim 2, characterized in that: The knowledge base is implemented as a database that stores known information, including the basic types of ships and the specific performance parameters of each type of ship.

4. The multi-objective association fusion method based on rules and prediction models according to claim 2, characterized in that: The method for extracting trajectory features is as follows: Connect the first and last points of the trajectory with a straight line; measure the distance from all points to this line, and find the point with the maximum distance value. The point; will Compare with the set threshold t_res_old, if If the value is less than the threshold t_res_old, then other intermediate points of the trajectory will no longer be segmented, and feature extraction will be completed; if... If the value is greater than or equal to the threshold t_res_old, then retain it. The corresponding data points; Using the location as a boundary, the trajectory is divided into two segments. The above steps are repeated for each segment to continue the judgment until feature extraction is completed.

5. The multi-objective association fusion method based on rule and prediction model according to claim 4, characterized in that: The method for establishing a ship trajectory prediction model based on Temporal Convolutional Network (TCN) is as follows: S1. Let the input sequence of the ship trajectory prediction model of the Temporal Convolutional Network (TCN) be... ;in, The sequence number of a certain feature of a ship's track; The number of sequences corresponding to each feature; Represents ship track-related characteristics including longitude, latitude, speed, heading, and time interval; the output sequence is set as follows. ,in, Indicates longitude or latitude; S2, Calculate the causal convolutional layer under... Predicted values ​​of ship tracks at the location; S3. Based on the calculation of the causal convolutional filter, a dilated convolutional layer is further used to increase the field of view of the convolution, and the calculation is performed on... Predicted ship tracks at the location ; S4. Establish the calculation formula for the residual connection layer; S5. The TCN ship trajectory prediction model consists of causal convolutional layers, hollow convolutional layers, and residual connection layers.

6. The multi-objective association fusion method based on rule and prediction model according to claim 5, characterized in that: In S2, the causal convolution filter is set as ; exist Predicted ship tracks at the location for: In the formula, The size of the kernel; It is a convolution mapping function; This represents the current layer number of the convolutional network. The nth sequence number corresponding to each feature.

7. The multi-objective association fusion method based on rules and prediction models according to claim 6, characterized in that: In S3, the predicted value for: In the formula, It is the expansion factor; This represents the size of the dilated convolution filter; It represents the direction of history.

8. The multi-objective association fusion method based on rule and prediction model according to claim 7, characterized in that: In S4, the residual connection layer for: In the formula, For activation functions; It is the residual function; The input sequence is .

9. The multi-objective association fusion method based on rules and prediction models according to claim 8, characterized in that: The method for establishing the target association method is as follows: In the early stages of observation, when historical data on the target is insufficient, the knowledge base is used to calculate the possible range of the target based on the performance knowledge of this type of ship, and the center of the range is used as the predicted location of the target; the distance to the calculated result is subtracted from the activity range distance. The distance between the observed data and the predicted ship position data is calculated one by one. A distance threshold is set to determine whether the association between two positions is successful. If the calculated distance is greater than the threshold, the position is judged as impossible to be associated. If the distance is less than the threshold, the position is judged as successfully associated. The positions with established association are recorded in the association database for subsequent fine association of the track. In the later stages of observation, the target association is based on the prediction results of the constructed TCN model. The TCN prediction model determines the position of the target at the same time as the observation, and uses Euclidean distance calculation as a measure of the degree of target association. The distance between the target observation point and the prediction point data is calculated and compared to determine the target. When a target to be associated matches multiple known targets, the target with the smallest distance is taken as its final target. If no matching target is found for the target to be associated, the target to be associated is considered a new target.

10. The multi-objective association fusion method based on rules and prediction models according to claim 9, characterized in that: The Euclidean distance calculation is specifically as follows: In the formula, , These are vectors composed of the latitude and longitude coordinates of the observation point to be associated and the prediction point to be selected, respectively. The longitude of the observation point to be associated; To select the longitude of the target prediction point; The latitude of the observation point to be associated; To select the latitude of the target prediction point.

Images