CN-122020111-A - Track identification method based on TCN-axis attention

Abstract

The invention relates to a track recognition method based on TCN-axis attention, which comprises the steps of collecting a track feature sequence of a target ship, constructing a track recognition model to generate a track recognition result according to the track feature sequence, wherein the track recognition model comprises a time convolution network for extracting first time sequence features from the track feature sequence, an axis attention mechanism module connected with the output end of the time convolution network and used for calculating attention along a time axis and a feature axis respectively according to the first time sequence features so as to generate a second time sequence feature containing global information, and a prediction module connected with the output end of the axis attention mechanism module and used for generating the track recognition result according to the second time sequence features. The invention can improve the efficiency and the precision of track identification.

Inventors

• LIN YIXING
• ZHANG LEFENG
• ZHU MINGWU
• ZHAO YAO

Assignees

• 上海戎慧智能科技有限公司

Dates

Publication Date

20260512

Application Date

20251224

Claims (8)

1. 1. A TCN-axis attention-based track recognition method, comprising: acquiring a track characteristic sequence of a target ship; constructing a track recognition model to generate a track recognition result according to the track feature sequence; the track recognition model comprises: a time convolution network for extracting a first timing feature from the track feature sequence; the axis attention mechanism module is connected with the output end of the time convolution network and is used for calculating attention along a time axis and a characteristic axis according to the first time sequence characteristic so as to generate a second time sequence characteristic containing global information; And the prediction module is connected with the output end of the shaft attention mechanism module and is used for generating a track recognition result according to the second time sequence characteristic.
2. 2. The method of claim 1, wherein the prediction module comprises a fully connected layer for track prediction based on the second timing characteristic.
3. 3. The method of claim 2, wherein the prediction module further comprises a LogSoftmax layer for behavior classification based on the trajectory prediction output by the full connection layer.
4. 4. The method of claim 1, wherein the track features include speed, heading, longitude, and latitude.
5. 5. The method of claim 1, wherein the temporal convolution network comprises a plurality of sequential stacked timing blocks, each timing block comprising two convolution layers, each convolution layer comprising a causal convolution and an dilation convolution.
6. 6. The method of claim 1, wherein the track recognition model is jointly trained using a multitasking learning framework of track prediction tasks and behavior recognition tasks.
7. 7. The method of claim 1, wherein the track recognition model, when trained, adaptively optimizes model parameters based on a loss weight of homodyne uncertainty.
8. 8. The method of claim 1, wherein the track feature sequence is extracted from AIS data.

Description

Track identification method based on TCN-axis attention Technical Field The invention relates to the technical field of track identification, in particular to a track identification method based on TCN-axis attention. Background The track identification is a core technology of behavior analysis and safety monitoring in maritime fields, the traditional method (such as LSTM and CNN) has the bottlenecks of insufficient capture of long time sequence dependence, low calculation efficiency and the like, and the transform-based model can model global dependence, but the calculation complexity grows along with the square of the sequence length, so that the large-scale track data is difficult to adapt. Therefore, an integrated scheme that combines long time-series modeling capability, computational efficiency, and recognition accuracy is needed. Disclosure of Invention The invention aims to provide a track identification method based on TCN-axis attention, which can improve the efficiency and the accuracy of track identification. The technical scheme adopted by the invention for solving the technical problems is to provide a track identification method based on TCN-axis attention, which comprises the following steps: acquiring a track characteristic sequence of a target ship; constructing a track recognition model to generate a track recognition result according to the track feature sequence; the track recognition model comprises: a time convolution network for extracting a first timing feature from the track feature sequence; the axis attention mechanism module is connected with the output end of the time convolution network and is used for calculating attention along a time axis and a characteristic axis according to the first time sequence characteristic so as to generate a second time sequence characteristic containing global information; And the prediction module is connected with the output end of the shaft attention mechanism module and is used for generating a track recognition result according to the second time sequence characteristic. Further, the prediction module includes a fully connected layer, where the fully connected layer is configured to perform track prediction according to the second timing characteristic. Further, the prediction module further comprises a LogSoftmax layer, which is used for classifying the behaviors according to the track prediction result output by the full-connection layer. Further, the track features include speed, heading, longitude, and latitude. Further, the temporal convolution network includes a plurality of sequentially stacked timing blocks, each timing block including two convolution layers, each convolution layer including a causal convolution and an dilation convolution. Furthermore, the track recognition model adopts a multi-task learning framework of track prediction tasks and behavior recognition tasks to carry out joint training. Further, when the track recognition model is trained, the model parameters are optimized in a self-adaptive mode based on the loss weight of the homodyne uncertainty; further, the track characteristic sequence is extracted from AIS data. Advantageous effects Compared with the prior art, the method has the advantages and positive effects that the efficiency and precision of track identification are improved by combining the advantages of the long time sequence dependence capturing capability of TCN and the efficient global modeling of the axis attention, on one hand, the parallel processing capability of TCN and the efficient characteristic capturing of the axis attention mechanism reduce training and predicting time of a model, so that the model can process a large-scale data set in real time, the response speed and instantaneity of a system are improved, on the other hand, the constructed model is simple in structure, small in parameter quantity, overfitting risk is reduced, the generalization capability on a new data set is improved, the model can be better adapted to different data set actual conditions, and has wider application prospects, and in addition, the sea-air target identification technology based on AIS data can provide technical standards and solutions which can be referred to the industry, promote cooperative and information sharing, the efficiency of sea-surface targets can be improved, and economic benefits can be improved due to delayed navigation caused by collision risks are reduced. Drawings FIG. 1 is a flow chart of an embodiment of the present invention; FIG. 2 is a schematic diagram of a track recognition model according to an embodiment of the present invention. Detailed Description The application will be further illustrated with reference to specific examples. It is to be understood that these examples are illustrative of the present application and are not intended to limit the scope of the present application. Furthermore, it should be understood that various changes and modifications can be made by one s