Ship target path prediction method based on massive historical trajectory enhancement

Abstract

The invention discloses a ship target path prediction method based on massive historical trajectory enhancement. According to the method, the advantages of a historical data association analysis-based method and a real-time data pattern recognition-based method are combined, and meanwhile, rich knowledge in historical data and a prediction context in real-time data are considered for prediction. And the robustness is higher than that of only using a certain one. According to the method, knowledge implied in mass data is fully utilized by utilizing historical data, but the historical data is only used as an enhancement method of real-time prediction data and does not directly influence a prediction result, so that the requirement on diversity of the historical data is not high. According to the method, knowledge training of massive historical data is performed in advance offline, and only a small amount of real-time to-be-predicted trajectory data needs to be trained in the online stage of prediction like a common pattern recognition method, so that the performance of the method is far higher than that of a historical data association analysis method; and the performance is similar to that of a mode recognition method.

Description

Technical Field

[0001] The invention relates to a ship target path prediction technology based on massive historical trajectory data enhancement, and belongs to the field of information technology. Background Art

[0002] The path prediction problem of mobile targets is a widely studied problem. The prediction problem for targets such as ships and aircraft is affected by natural factors such as climate and human factors such as political culture. Compared with the path prediction of natural targets such as hurricanes and thunderstorms, it is more challenging. In addition, in real environments, due to the influence of sampling frequency and sampling noise of mobile target position sensors, path prediction is more susceptible to data noise.

[0003] The generalized mobile target path prediction problem can be divided into short-term and long-term predictions according to the size of the prediction time window, and can be divided into association analysis based on historical trajectory data and pattern recognition based on real-time streaming data according to the prediction algorithm. However, the path prediction for ship targets only uses either association analysis or pattern recognition methods, and the effect is not ideal. The main reasons are as follows: The prediction technology that uses historical data for association analysis is essentially a similar trajectory search, using the most similar samples in the historical data as a prediction of future trends. The prediction effect of this method is limited by the diversity of historical trajectory samples, and searching massive historical trajectory data is very time-consuming and computationally intensive. The prediction technology that directly performs real-time pattern recognition on the trajectory to be predicted has a considerable performance advantage, but it completely abandons the rich knowledge contained in massive historical data, and is more suitable for short-term prediction or natural mobile target path prediction problems that are insensitive to the natural and human environment.

[0004] Based on the above problems, it is necessary to propose a technical solution that combines the advantages of the two types of technologies for the motion path prediction problem of ship targets. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the prediction technology of the association analysis method consumes huge time and computing power when facing diverse sample data, while the pattern recognition method loses the rich knowledge in the historical data.

[0006] In order to solve the above technical problems, the technical solution of the present invention is to disclose a ship target path prediction method based on massive historical trajectory enhancement, which is characterized by comprising the following steps:

[0007] Step 1: Grid coding of the ship motion area to be predicted;

[0008] Step 2: Construct a grid encoding mapping of the original trajectory and the downsampled trajectory, which includes the following steps:

[0009] Step 2.1: Add Gaussian noise to the original trajectory coordinates of the historical data and downsample the coordinate sequence;

[0010] Step 2.2: Map the original trajectory and the noisy and downsampled samples from coordinates to region grid IDs;

[0011] Step 3: Use massive historical data samples to train the trajectory autoencoder for the ship's motion area;

[0012] Step 4: Use the pre-trained trajectory autoencoder model to upsample the predicted real-time ship motion trajectory;

[0013] Step 5: Using the enhanced real-time ship trajectory to be predicted to dynamically train the trajectory prediction model, specifically including the following steps:

[0014] Step 5.1: Construct a trajectory path prediction training dataset according to the prediction step length parameter;

[0015] Step 5.2: Use the momentum-based gradient descent algorithm Adam as the optimizer and soft dynamic time warping as the training target to train the trajectory timing prediction model;

[0016] Step 6: Use the trained model to predict the path of the target within the specified time window.

[0017] Preferably, the step 1 specifically includes the following steps:

[0018] Step 1.1: Divide the entire ship motion area from the upper left to the lower right into W*H rectangles according to the gridding parameter W*H and number them;

[0019] Step 1.2: Extract the historical data in the ship's motion area from the massive historical trajectory data, divide it into buckets according to the grids where it is located, and count the number of trajectory points in all grids;

[0020] Step 1.3: Keep a hotspot grid with a sufficient number of trajectory points according to the preset gridding parameters.

[0021] Preferably, the step 3 specifically includes the following steps:

[0022] Step 3.1: Construct a trajectory autoencoder model, which consists of an encoder and a decoder. The encoder and the decoder are respectively composed of a recurrent neural network based on a long short-term memory unit. The encoder encodes the output of the last long short-term memory unit into a 128-dimensional embedding vector.

[0023] Step 3.2: The encoder reduces the dimension of the noisy and downsampled grid ID sequence into an embedding vector of fixed size; the encoder restores the embedding vector to the original grid ID sequence;

[0024] Step 3.3: Use negative log-likelihood as the training target, momentum-based gradient descent algorithm Adam as the optimizer, and use gradient clipping to avoid gradient explosion.

[0025] The present invention discloses a ship target path prediction technology based on massive historical trajectory data enhancement. Compared with the existing technical solutions, it has the following beneficial effects:

[0026] 1. The present invention combines the advantages of two methods: those based on historical data association analysis and those based on real-time data pattern recognition, and considers the rich knowledge in historical data and the prediction context in real-time data for prediction. It is more robust than using only one of them.

[0027] 2. The present invention makes full use of the knowledge implicit in massive data by using historical data. However, since the historical data is used only as an enhancement method for real-time prediction data and does not directly affect the prediction results, the diversity requirements for the historical data itself are not high.

[0028] 3. The present invention performs knowledge training on massive historical data offline in advance, and in the online stage of prediction, it only needs to train a small amount of real-time trajectory data to be predicted, just like ordinary pattern recognition methods. Therefore, the performance of the present invention is much better than that of historical data association analysis methods, and is similar to that of pattern recognition methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart;

[0030] Figure 2 A sample of historical trajectories is illustrated;

[0031] Figure 3 The overall model architecture of the autoencoder is shown;

[0032] Figure 4 The internal structure of the long short-term memory unit is shown;

[0033] Figure 5 This is the upsampling effect diagram of the trajectory by the autoencoder;

[0034] Figure 6 Schematic diagram of the effect of real-time ship target path prediction. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0036] The embodiment of the present invention discloses a ship target path prediction method based on massive historical trajectory enhancement, such as Figure 1 As shown, it includes: in the offline stage, the area to be predicted is gridded and hotspot grids are screened out; trajectory autoencoder training samples are constructed, historical trajectories are downsampled, and hotspot grid mapping sequence pairs are formed with the original trajectories; the dimension reduction encoder and restoration decoder of the historical trajectory are trained based on the autoencoder / decoder network architecture. In the online stage, the real-time trajectory of the ship is upsampled using the aforementioned trajectory autoencoder; path prediction training samples are constructed in real time for the enhanced real-time ship trajectory, and the real-time target path prediction model is trained based on the linear regression model of time series data and soft dynamic time warping as the training target; the model trained in real time is used to predict the motion path of the ship target in the next several time windows.

[0037] According to a cluster analysis and anomaly analysis method for ship trajectory data provided by the present invention, this embodiment combines a sample of historical ship trajectory data from April 2021 to August 2021 in a coastal waters (see the historical trajectory sample). Figure 2 ) to train the trajectory autoencoder, and introduce the implementation process of moving target path prediction based on historical trajectory data enhancement, which specifically includes the following steps:

[0038] Step 1: Grid coding is performed on the ship motion area in the area where path prediction is to be performed, as follows:

[0039] Step 1.1: Divide the entire space from the upper left to the lower right into W*H rectangles according to the gridding parameter W*H and number them.

[0040] Step 1.2: Extract the historical data in the area from the massive historical trajectory data, divide it into buckets according to the grids, and count the number of trajectory points in all grids.

[0041] Step 1.3: Keep a hotspot grid with a sufficient number of trajectory points according to the preset gridding parameters.

[0042] Step 2: Construct a grid encoding mapping of the original trajectory and the downsampled trajectory as follows:

[0043] Step 2.1: Add Gaussian noise to the original trajectory coordinates of the historical data and downsample the coordinate sequence.

[0044] Step 2.2: Map the original trajectory and the noisy and downsampled samples from coordinates to region grid IDs.

[0045] Step 3: Use massive historical data samples to train the trajectory autoencoder for the area. The specific steps are as follows:

[0046] Step 3.1: Construct a trajectory autoencoder model, which consists of an encoder and a decoder. The encoder and decoder are composed of recurrent neural networks based on long short-term memory units. The encoder encodes the output of the last long short-term memory unit into a 128-dimensional embedding vector. The overall model architecture of the autoencoder is shown in Figure 3 , the internal structure of the long short-term memory unit can be seen in Figure 4 .

[0047] Step 3.2: The encoder reduces the dimension of the noisy and downsampled grid ID sequence into an embedding vector of a fixed size; the encoder restores the embedding vector to the original grid ID sequence.

[0048] Step 3.3: Use negative log-likelihood as the training target, momentum-based gradient descent algorithm Adam as the optimizer, and use gradient clipping to avoid gradient explosion.

[0049] Step 4: Use the pre-trained autoencoder model to upsample the real-time ship motion trajectory to be predicted. The upsampling effect of the autoencoder on the trajectory is shown in Figure 5 , the yellow one is the real-time trajectory to be predicted, and the blue one is the upsampled trajectory.

[0050] Step 5: Use the enhanced real-time ship trajectory to be predicted to dynamically train the trajectory prediction model. The specific steps are as follows:

[0051] Step 5.1: Construct a trajectory path prediction training dataset according to the prediction step size parameter.

[0052] Step 5.2: Use the momentum-based gradient descent algorithm Adam as the optimizer and soft dynamic time warping as the training target to train the trajectory timing prediction model.

[0053] Step 6: Use the trained model to predict the target path within the specified time window. Figure 6 , the blue trajectory segment is the trajectory segment to be predicted, the blue dot is the real path of the target in the next five time windows, and the yellow dot is the prediction result of the prediction algorithm.

Claims

1. A ship target path prediction method based on massive historical trajectory enhancement, characterized in that: The following steps are involved: Step 1: Grid coding of the ship motion area to be predicted; Step 2: Construct a grid encoding mapping of the original trajectory and the downsampled trajectory, which includes the following steps: Step 2.1: Add Gaussian noise to the original trajectory coordinates of the historical data and downsample the coordinate sequence; Step 2.2: Map the original trajectory and the noisy and downsampled samples from coordinates to region grid IDs; Step 3: Use massive historical data samples to train the trajectory autoencoder for the ship's motion area; Step 4: Use the pre-trained trajectory autoencoder model to upsample the predicted real-time ship motion trajectory; Step 5: Dynamically train the trajectory prediction model using the enhanced real-time ship trajectory to be predicted, which specifically includes the following steps: Step 5.1: Construct a trajectory path prediction training dataset according to the prediction step length parameter; Step 5.2: Use the momentum-based gradient descent algorithm Adam as the optimizer and soft dynamic time warping as the training target to train the trajectory timing prediction model; Step 6: Use the trained model to predict the path of the target within the specified time window.

2. A method for predicting a ship target path based on massive historical trajectory enhancement as claimed in claim 1, characterized in that: The step 1 specifically comprises the following steps: Step 1.1: Divide the entire ship motion area from the upper left to the lower right into W*H rectangles according to the gridding parameter W*H and number them; Step 1.2: Extract the historical data in the ship's motion area from the massive historical trajectory data, divide it into buckets according to the grids where it is located, and count the number of trajectory points in all grids; Step 1.3: Keep a hotspot grid with a sufficient number of trajectory points according to the preset gridding parameters.

3. A method for predicting a ship target path based on massive historical trajectory enhancement as claimed in claim 2, characterized in that: The step 3 specifically comprises the following steps: Step 3.1: Construct a trajectory autoencoder model, which consists of an encoder and a decoder. The encoder and the decoder are respectively composed of a recurrent neural network based on a long short-term memory unit. The encoder encodes the output of the last long short-term memory unit into a 128-dimensional embedding vector. Step 3.2: The encoder reduces the dimension of the noisy and downsampled grid ID sequence into an embedding vector of fixed size; the encoder restores the embedding vector to the original grid ID sequence; Step 3.3: Use negative log-likelihood as the training target, momentum-based gradient descent algorithm Adam as the optimizer, and use gradient clipping to avoid gradient explosion.

Images