A ship trajectory anomaly detection method based on trajectory reconstruction

Abstract

The present invention proposes a ship trajectory anomaly detection method based on trajectory reconstruction. First, the motion trajectory of the ship is converted into the form of an image vector, and then a convolutional autoencoder is used for modeling and training. The training goal of the autoencoder is to reconstruct the original trajectory data to the maximum extent during the decoding process. After the training is completed, the model can effectively identify trajectory anomalies by comparing the difference between the model input trajectory and the reconstructed trajectory. The present invention only needs to compare the difference between the original trajectory and the reconstructed trajectory, thereby greatly reducing the computational cost. In addition, the model of the present invention can automatically set the normal and abnormal trajectory distance thresholds by analyzing the proportion of abnormal trajectories in the training data set, avoiding the trouble of manually setting the threshold, and making the model easier to apply in various practical scenarios. Performance tests show that the present invention has good convergence performance during the training process, and the accuracy rate reaches a high standard of more than 90%.

Description

Technical Field

[0001] The invention belongs to the field of radar data processing. Background Art

[0002] The goal of trajectory anomaly detection is to identify abnormal patterns from a large amount of trajectory data, which can help detect or predict potential traffic risks. Trajectory anomalies are usually divided into two categories: 1) trajectories that deviate from other trajectories in space, and 2) time series of trajectory points that are significantly different from other trajectories. Trajectory anomalies indicate deviations from expected behavior and are usually related to specific or unusual events.

[0003] Taking ship traffic monitoring and management as an example, in the ship road network, local trajectory anomalies may reflect the occurrence of traffic accidents, bad weather or emergencies. Real-time detection and analysis of these trajectory anomalies can not only provide a basis for the maritime administration's traffic situation analysis, but also enable timely detection and early warning. Traditional trajectory anomaly detection methods mainly focus on trajectory similarity and cluster analysis. These methods define distance metrics to calculate the similarity between trajectories, cluster trajectories into several categories, and finally identify clusters with fewer trajectories as anomalies based on threshold settings. However, this method is often inefficient and may not effectively meet the real-time analysis needs of large-scale trajectory data. Ship trajectory data has the characteristics of uncertainty, sparsity, skewed distribution, and continuous updating, which makes real-time anomaly detection in trajectories more challenging. With the rise of deep learning methods, deep neural networks are increasingly used in trajectory data analysis and mining. Summary of the invention

[0004] In order to solve the problems of uncertainty, sparsity, skewed distribution and difficulty in continuous updating of ship trajectory data in the prior art, the present invention proposes a ship trajectory anomaly detection method based on trajectory reconstruction.

[0005] The technical solution for realizing the present invention includes:

[0006] Step 1: Trajectory transformation: First, complete the preprocessing of the target data, remove abnormal data and normalize the trajectory of the trajectory space;

[0007] Step 2: Trajectory reconstruction: Based on the cubic spline interpolation method, the possible missing trajectory points are supplemented, and the timestamp information is integrated into the ship's motion trajectory using the Mercator projection to complete a new complete ship motion trajectory image;

[0008] Step 3: Trajectory comparison: Based on the deep learning model, the similarities of the motion images of similar ships are summarized, their abnormal images are compared, and abnormal trajectories are detected.

[0009] Furthermore, in step 1, considering the spatial motion range of the ship, in order to make the tracks of different sizes and shapes reasonably accommodated in a unified bounding box or monitoring area, the cumulative percentage of the moving distance of each AIS position point relative to the minimum longitude and latitude starting point is calculated by the ship's travel distance in the horizontal (x-axis) and vertical (y-axis) directions, thereby mapping the position point into a predefined bounding box.

[0010] Furthermore, in step 2, the original trajectory data is resampled at fixed time intervals (preferably at fixed time intervals of 5 seconds) to form a data set with uniform time resolution. The spline interpolation method is used to reconstruct the timestamp points that may be damaged or missing in the original data. During the projection process, it is ensured that the timestamp points fall on the W×H grid of grid points, and finally a detailed grayscale image representation with a resolution of W×H is obtained.

[0011] Furthermore, in step 3, in order to speed up the calculation, the trajectory image is first normalized to a fixed size of 80x80 pixels. The activation function selected by the deep learning algorithm is ReLU, which improves the nonlinear expression ability of the model, alleviates the problem of gradient disappearance, and has higher calculation efficiency.

[0012] A ship trajectory anomaly detection method based on trajectory reconstruction proposed in the present invention not only reduces the necessary computational load, but also shows excellent detection performance in real-time scenarios. By converting trajectory data into information-rich images and integrating speed features, a new height in accuracy is achieved. The trajectory reconstruction model developed by the present invention is in the leading position in trajectory anomaly detection, with an overall accuracy of 91%. The recall rate of 92% ensures a high probability of accurate identification and trajectory reconstruction. In addition, the high accuracy of 93% significantly reduces the risk of misclassification. The F1 score of 92% indicates that the model has achieved a successful balance between precision and recall, minimizing false positives and false negatives. The visualization of the training process also shows that the trajectory reconstruction model exhibits a complex learning curve and a coherent loss reduction process. The similar accuracy and loss trends in training and validation further indicate that the model does not suffer from overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is the anomaly detection workflow diagram.

[0014] Figure 2 This is a schematic diagram of the original track.

[0015] Figure 3 Schematic diagram of the original track grid.

[0016] Figure 4 Trajectory recognition accuracy. DETAILED DESCRIPTION

[0017] The present invention proposes a ship trajectory anomaly detection method based on trajectory reconstruction. Based on the ship dynamic data from radar, AIS and other equipment data, the convolutional autoencoder is used to convert the ship's motion trajectory into the form of an image vector, and then modeling and training are performed. By comparing the difference between the original trajectory and the reconstructed trajectory, the trajectory anomaly is found, thereby greatly reducing the computational cost. By converting the trajectory data into an information-rich image and integrating the speed features, and using the trajectory reconstruction model for comparison, the detection has reached a new height in accuracy.

[0018] The technical solution of the present invention is further explained below in conjunction with the accompanying drawings and embodiments.

[0019] The main steps of the present invention include: trajectory conversion, trajectory reconstruction, and trajectory comparison. The preferred implementation process is described as follows:

[0020] 1. Trajectory conversion is to convert the original ship trajectory into information-rich 2D images. These images are crucial in the feature learning process as the input of the trajectory reconstruction model. This conversion ensures that important trajectory features, such as trajectory shape and velocity information, can be effectively captured and learned by the trajectory reconstruction model.

[0021] The focus of the trajectory conversion study is to accurately capture and reflect two core features of the ocean trajectory pattern: the contour of the trajectory and the speed of the ship. Considering the spatial motion range of the ship, the trajectory data is normalized so that trajectories of different sizes and shapes can be reasonably accommodated in a unified bounding box or monitoring area. Specifically, the distance traveled by the ship in the horizontal (x-axis) and vertical (y-axis) directions is measured, and measures are taken to ensure the consistency of the journey size. Then, the normalized representation of each track point on the trajectory image is ensured by calculating the cumulative percentage of the movement distance of each ship position point relative to the minimum longitude and latitude starting point. Finally, these ship position points are mapped into a predefined bounding box to form an image representation of the trajectory. Considering that the speed of most types of ships is usually in the range of 0 to 22 knots, the speed is divided into several intervals of 2 knots. This division enables the speed information in the trajectory image to be represented by different grayscale values, with one grayscale level for every 2 knots. In the final image, the speed values ​​of the ship at different positions are intuitively and accurately displayed, laying a solid foundation for the subsequent analysis using the trajectory reconstruction model.

[0022] 2. Trajectory reconstruction proposes to create information-rich trajectory images by integrating timestamp information into the image generation process, thus successfully creating information-rich trajectory images. When constructing these images, the classic Mercator projection is used to map the timestamp points in the trajectory to Cartesian coordinates, ensuring that the geometric position information of the trajectory retains its original accuracy and intuitiveness. This allows for more accurate and direct interpretation and analysis of spatial data in subsequent steps. Figure 2 As shown in the figure, the process of generating informative trajectory images can be divided into three clear steps: First, the original trajectory data is resampled using a fixed time interval of 5 seconds to form a dataset with uniform temporal resolution. Then, in order to improve the continuity and integrity of the trajectory, the spline interpolation method is applied to reconstruct the timestamp points that may be damaged or missing in the original data. Finally, during the projection process, it is ensured that the timestamp points fall on the W×H grid of grid points, and finally a detailed grayscale image representation with a resolution of W×H is obtained, which lays a solid foundation for subsequent analysis.

[0023] 3. In the trajectory comparison, the trajectory images are normalized to a fixed size of 80x80 pixels. This is mainly to ensure the consistency of the model input and speed up the processing time. In the training of the deep learning model, 30 training epochs were selected and the batch size was set to 8. The purpose of this approach is to find an appropriate balance between ensuring sufficient learning and avoiding overfitting. The selected activation function is ReLU. Compared with the traditional s-type activation function, ReLU not only improves the nonlinear expression ability of the model and alleviates the problem of gradient disappearance, but also has higher computational efficiency. Therefore, it guarantees the performance of the model while accelerating the convergence speed of the model. In order to build a model with good anti-overfitting ability, the dropout technique is used in the training process. The principle of Dropout is to randomly disable a certain proportion of neuron outputs during training, forcing the model to learn based on less information, thereby improving the generalization ability of the model. In addition, the softmax function is used in the final convolutional layer. Since the softmax function is particularly suitable for dealing with multi-class problems, it is able to generate a normalized probability distribution that clearly represents the predicted probability of each category.

[0024] In order to provide the model with high-quality representation of ship activities, 200 representative images were selected from each category. The resolution of the images was set to 64x64. The dataset was randomly split into 80% for training and 20% for testing, including different types of trajectory images. The trajectories were split into equally sized time windows of length 8. This particular time window was chosen because ship movement patterns take time to manifest.

Claims

1. A ship trajectory anomaly detection method based on trajectory reconstruction, characterized in that: Step 1: Trajectory transformation: First, complete the preprocessing of the target data, remove abnormal data and normalize the trajectory of the trajectory space; Step 2: Trajectory reconstruction: Based on the cubic spline interpolation method, the possible missing trajectory points are supplemented, and the timestamp information is integrated into the ship's motion trajectory using the Mercator projection to complete a new complete ship motion trajectory image; Step 3: Trajectory comparison: Based on the deep learning model, the similarities of the motion images of similar ships are summarized, their abnormal images are compared, and abnormal trajectories are detected.

2. The ship trajectory anomaly detection method based on trajectory reconstruction according to claim 1 is characterized in that: The step 1 further comprises: calculating the cumulative percentage of the moving distance of each AIS position point relative to the minimum longitude and latitude starting point according to the spatial movement range of the ship through the travel distance of the ship in the horizontal x-axis and vertical y-axis directions, and mapping the position point into a predefined bounding box.

3. The ship trajectory anomaly detection method based on trajectory reconstruction according to claim 1 is characterized in that: The step 2 also includes: resampling the original trajectory data at a fixed time interval to form a data set with a uniform time resolution; applying a spline interpolation method to reconstruct the timestamp points that may be damaged or missing in the original data; during the projection process, ensuring that the timestamp points fall on a W×H grid of grid points, and finally obtaining a detailed grayscale image representation with a resolution of W×H.

4. The ship trajectory anomaly detection method based on trajectory reconstruction according to claim 1 is characterized in that: The step 3 also includes: first, the trajectory image is normalized to a fixed size of 80x80 pixels, and the activation function selected by the deep learning algorithm is ReLU.

Images