CN-121980162-A - KAN-LSTM-based target track prediction method

Abstract

The embodiment of the invention provides a KAN-LSTM-based target track prediction method, which comprises the following steps of S1, obtaining target observation data containing D-dimensional characteristics, S2, preprocessing the data, removing abnormal values in the target observation data, and supplementing missing values, S3, normalizing the data, scaling the data of each dimension of the target observation data to a range of [0,1], S4, constructing a KAN-LSTM-based target track prediction network model, completing model training, S5, inputting the target observation data into the target track prediction network model, and obtaining a target motion state prediction result at the next moment. The embodiment of the invention effectively improves the fitting capability of the model to the nonlinear relation, enhances the learning capability of the model to the complex data mode, provides a new solution for processing the marine target track prediction task, and shows higher accuracy and reliability in the dynamic behavior prediction of the marine target.

Inventors

• SUN RUIZE
• WANG XUELIANG
• WANG YANJIE
• LI ZHENG
• LIU JIE
• WANG JING

Assignees

• 中国船舶集团有限公司系统工程研究院

Dates

Publication Date

20260505

Application Date

20251219

Claims (8)

1. 1. A KAN-LSTM-based target track prediction method is characterized by comprising the following steps: step S1, obtaining target observation data containing D-dimensional characteristics; s2, preprocessing data, removing abnormal values in target observation data, and complementing missing values; step S3, data normalization, namely scaling the data of each dimension of the target observation data to the range of [0,1 ]; S4, constructing a KAN-LSTM-based target track prediction network model, and completing model training; And S5, inputting the target observation data into a target track prediction network model, and obtaining a target motion state prediction result at the next moment.
2. 2. The method for predicting the target track based on KAN-LSTM as recited in claim 1, wherein the outliers in the target observation data are eliminated by judging the preset conditions.
3. 3. The method of claim 1, wherein in step S2, the missing values are filled in by using an average value of a search method.
4. 4. The method for predicting target trajectories based on KAN-LSTM as set forth in claim 1, wherein in step S3, data normalization is performed by using a range transform method.
5. 5. The method for predicting a target trajectory based on KAN-LSTM of claim 1, wherein the KAN-LSTM-based target trajectory prediction network model comprises, The feature extraction layer adopts a residual KAN network structure, and input data is subjected to feature extraction through KAN designed by residual; the time sequence modeling layer adopts a K-LSTM structure, and the characteristics extracted by the characteristic extraction layer extract time sequence information through a K-LSTM module; And the regression output layer adopts an MLP dense connection layer structure, and the output of the time sequence modeling layer is mapped to an output state space through the MLP dense connection layer to obtain a target motion state prediction result at the next moment.
6. 6. The method for predicting a target trajectory based on KAN-LSTM of claim 5, wherein the model training process comprises, Forward propagation, obtaining normalized time sequence data of a batch from a data loader, inputting the normalized time sequence data into a model, and obtaining a predicted output through residual errors KAN, K-LSTM and MLP in sequence Loss calculation, calculating the batch forecast using a defined loss function Loss from the true value y; Counter-propagating, calculating gradients of all parameters of the model by using the loss function; updating parameters, and adjusting all model parameters by using the calculated gradient by an optimizer according to an updating rule of the calculated gradient; The above steps are repeated until the complete training set is traversed, completing one round, after which the model performance is evaluated on the validation set.
7. 7. The method for predicting a target trajectory based on KAN-LSTM of claim 6, wherein the loss function uses MSE loss.
8. 8. The method for predicting a target trajectory based on KAN-LSTM of claim 6, wherein the optimizer is an Adam optimizer.

Description

KAN-LSTM-based target track prediction method Technical Field The invention relates to the technical field of target track prediction in water, in particular to a KAN-LSTM-based target track prediction method. Background The ocean is not only a main channel of trade circulation, but also an important field for protecting the national territory safety and maintaining the national rights and interests, and is important for the economic development and national defense industry of China. The method can effectively predict the track of the water surface or underwater target so as to analyze the intention and the aggressiveness of the target, and has important significance for the offshore national defense industry. However, the ocean situation is various, the environment is complex, and many negative effects are brought to ocean target detection. The data rule of the ocean target track is difficult to model by using a traditional method due to factors such as weather, wind waves, noise and the like, so that the accuracy and reliability of the dynamic behavior prediction of the ocean target are poor. Disclosure of Invention In view of the above problems in the prior art, an embodiment of the present invention provides a KAN-LSTM-based target trajectory prediction method, so as to solve the technical problems of poor accuracy and reliability of dynamic behavior prediction of a marine target in the prior art. The embodiment of the invention provides a KAN-LSTM-based target track prediction method, which comprises the following steps: step S1, obtaining target observation data containing D-dimensional characteristics; s2, preprocessing data, removing abnormal values in target observation data, and complementing missing values; step S3, data normalization, namely scaling the data of each dimension of the target observation data to the range of [0,1 ]; S4, constructing a KAN-LSTM-based target track prediction network model, and completing model training; And S5, inputting the target observation data into a target track prediction network model, and obtaining a target motion state prediction result at the next moment. In an embodiment, the outliers in the target observation data are removed by a preset condition judgment. In one embodiment, in step S2, the missing values are filled with the average value of the interpolation method. In one embodiment, in step S3, the data normalization is implemented by using a range transform method. In one embodiment, the KAN-LSTM based target trajectory prediction network model comprises, The feature extraction layer adopts a residual KAN network structure, and input data is subjected to feature extraction through KAN designed by residual; the time sequence modeling layer adopts a K-LSTM structure, and the characteristics extracted by the characteristic extraction layer extract time sequence information through a K-LSTM module; And the regression output layer adopts an MLP dense connection layer structure, and the output of the time sequence modeling layer is mapped to an output state space through the MLP dense connection layer to obtain a target motion state prediction result at the next moment. In one embodiment, in step S4, the model training process includes, Forward propagation, obtaining normalized time sequence data of a batch from a data loader, inputting the normalized time sequence data into a model, and obtaining a predicted output through residual errors KAN, K-LSTM and MLP in sequence Loss calculation, calculating the batch forecast using a defined loss functionLoss from the true value y; Counter-propagating, calculating gradients of all parameters of the model by using the loss function; updating parameters, and adjusting all model parameters by using the calculated gradient by an optimizer according to an updating rule of the calculated gradient; The above steps are repeated until the complete training set is traversed, completing one round, after which the model performance is evaluated on the validation set. In an embodiment, the loss function employs MSE loss. In one embodiment, the optimizer employs an Adam optimizer. Compared with the prior art, the target track prediction method based on the KAN-LSTM has the advantages that the KAN network is introduced while the advantages of the LSTM in structural design are maintained, the fitting capacity of the model to nonlinear relations is effectively improved, the learning capacity of the model to complex data modes is enhanced, a new solution is provided for processing marine target track prediction tasks, higher accuracy and reliability are shown in dynamic behavior prediction of marine targets, and research and application progress of related fields are promoted. Drawings FIG. 1 is a schematic diagram of a KAN-LSTM-based target track prediction network model according to a KAN-LSTM-based target track prediction method according to an embodiment of the present invention; Fig. 2 is a schematic diagram of a time seque