Apparatus and method for predicting track using MLP(Multi-Layer Perceptron)-based time series prediction

Abstract

An apparatus and method for predicting a trajectory using MLP-based time series forecasting are disclosed. A trajectory prediction method using MLP-based time series prediction includes the steps of: collecting data from an Automatic Identification System (AIS) to construct AIS time series data including latitude (lat), longitude (lon), speed (sog), and azimuth (cog); performing Density-Based Spatial Clustering of Applications with Noise (DBSCAN) on feature vectors including latitude (lat), longitude (lon), speed (sog), and azimuth (cog) in the AIS time series data to group data points with similar region and movement characteristics into the same cluster; performing learning by applying independent MLPs in the time-axis and feature-axis directions, respectively, to the grouped AIS time series data using a Class Label-guided Multi-Layer Perceptron (CL-MLP) model; predicting a future trajectory within a preset prediction time range using the learned results; and performing post-processing on the predicted values ​​of the predicted future trajectory according to preset resolution criteria or speed and route constraints.

Description

Technology Field

[0001] The present invention relates to a track prediction device and method using MLP-based time series prediction. Background Technology

[0002] Automatic Identification System (AIS)-based track prediction is an essential technology for maritime safety, route optimization, collision avoidance, and logistics planning. Previously, time series models such as RNN, LSTM, and Transformer were utilized, but problems existed including the difficulty of learning long time series dependencies, increased computational load, and limitations in generalization performance.

[0003] The recently proposed TS-Mixer (Time Series Mixer) model is known to overcome these limitations through a time-axis-feature axis separation structure based on an MLP (Multi-Layer Perceptron) and to exhibit excellent prediction performance with a lightweight structure.

[0004] However, existing models have limitations in that they cannot reflect static location and behavior patterns. Prior art literature

[0005] Korean Patent Publication No. 10-2778555 (February 27, 2025) The problem to be solved

[0006] The present invention aims to provide a track prediction device and method using MLP-based time series prediction, which uses AIS time series data as input, predicts future tracks in fixed time units based on a TS-Mixer structure, and improves prediction accuracy and generalization performance by adding static characteristics through DBSCAN (Density-Based Spatial Clustering of Applications with Noise). means of solving the problem

[0007] According to one aspect of the present invention, a trajectory prediction method using MLP-based time series prediction performed by a trajectory prediction device is disclosed.

[0008] A trajectory prediction method using MLP-based time series prediction according to an embodiment of the present invention comprises: a step of collecting data from an Automatic Identification System (AIS) to construct AIS time series data including latitude (lat), longitude (lon), speed (sog), and azimuth (cog); a step of performing Density-Based Spatial Clustering of Applications with Noise (DBSCAN) on feature vectors including latitude (lat), longitude (lon), speed (sog), and azimuth (cog) in the AIS time series data to group data points having similar region and movement characteristics into the same cluster; a step of performing training by applying independent MLPs in the direction of the time axis and the feature axis, respectively, to the grouped AIS time series data using a Class Label-guided Multi-Layer Perceptron (CL-MLP) model; a step of predicting a future trajectory within a preset prediction time range using the trained results; and, regarding the predicted value of the predicted future trajectory, according to preset resolution criteria or speed and route constraints It includes a step of performing post-processing.

[0009] According to another aspect of the present invention, a trajectory prediction device using MLP-based time series prediction is disclosed.

[0010] A trajectory prediction device using MLP-based time series prediction according to an embodiment of the present invention comprises a memory for storing instructions and a processor for executing said instructions, wherein the instructions include: a step of collecting data from an Automatic Identification System (AIS) to construct AIS time series data including latitude (lat), longitude (lon), speed (sog), and azimuth (cog); a step of performing Density-Based Spatial Clustering of Applications with Noise (DBSCAN) on feature vectors including latitude (lat), longitude (lon), speed (sog), and azimuth (cog) in the AIS time series data to group data points having similar region and movement characteristics into the same cluster; a step of performing learning by applying independent MLPs in the direction of the time axis and the feature axis, respectively, to the grouped AIS time series data using a Class Label-guided Multi-Layer Perceptron (CL-MLP) model; a step of predicting a future trajectory within a preset prediction time range using the learned results; and the predicted future A trajectory prediction method is performed, comprising the step of performing post-processing on the predicted value of the trajectory according to preset resolution criteria or speed and route constraints. Effects of the invention

[0011] The track prediction device and method using MLP-based time series prediction according to an embodiment of the present invention use AIS time series data as input to predict future tracks in fixed time units based on a TS-Mixer structure, and can improve prediction accuracy and generalization performance by adding static characteristics through DBSCAN (Density-Based Spatial Clustering of Applications with Noise). Brief explanation of the drawing

[0012] FIG. 1 is a flowchart schematically illustrating a trajectory prediction method using MLP-based time series prediction performed by a trajectory prediction device according to an embodiment of the present invention. FIGS. 2 to 4 are drawings for explaining a trajectory prediction method using MLP-based time series prediction according to an embodiment of the present invention of FIG. 1. FIG. 5 is a diagram schematically illustrating the configuration of a trajectory prediction device using MLP-based time series prediction according to an embodiment of the present invention. Specific details for implementing the invention

[0013] As used in this specification, singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "composed" or "comprising" should not be interpreted as necessarily including all of the various components or steps described in the specification, and should be interpreted as meaning that some of the components or steps may be excluded, or that additional components or steps may be included. Furthermore, terms such as "...part," "module," etc., as used in the specification refer to a unit that processes at least one function or operation, which may be implemented in hardware or software, or a combination of hardware and software.

[0014] Hereinafter, various embodiments of the present invention will be described in detail with reference to the attached drawings.

[0015] FIG. 1 is a flowchart schematically illustrating a trajectory prediction method using MLP-based time series prediction performed by a trajectory prediction device according to an embodiment of the present invention, and FIGS. 2 to 4 are drawings for explaining the trajectory prediction method using MLP-based time series prediction according to an embodiment of the present invention of FIG. 1. Hereinafter, a trajectory prediction method using MLP-based time series prediction according to an embodiment of the present invention will be described with reference to FIGS. 2 to 4, with reference to FIGS. 1.

[0016] In step S110, the track prediction device collects data from the Automatic Identification System (AIS) and constructs AIS time series data including latitude (lat), longitude (lon), speed (sog), and azimuth (cog).

[0017] For example, the track prediction device uses AIS time series data It can be constructed. Here, T is the input time series length, and F consists of items of latitude (lat), longitude (lon), speed (sog), and azimuth (cog).

[0018] At this time, the track prediction device can perform a structured preprocessing procedure to ensure the quality and consistency of AIS-based ship track data.

[0019] First, AIS messages containing unrealistic speed values ​​are removed. In particular, if the speed over ground is 30 knots or higher, it is determined that it does not reflect the actual movement of the vessel and is excluded from the dataset. Additionally, if the vessel's navigational status indicates that it is anchored or moored, the corresponding data point is also removed as it is considered not to be active movement.

[0020] To maintain temporal continuity, if the time interval between consecutive AIS messages exceeds 2 hours, the vessel track is divided based on this. This ensures that each segment is a continuous movement section without large time intervals.

[0021] Segments that are too short to be meaningfully analyzed are removed. Specifically, track segments with fewer than 20 AIS messages or a total duration of less than 4 hours are excluded from analysis.

[0022] To improve data reliability, outlier detection is performed. The empirical velocity between consecutive locations is calculated using the great-circle distance and elapsed time, and AIS messages with a velocity exceeding 40 knots are considered outliers and removed.

[0023] Since AIS data has irregular sampling intervals, all track data is resampled at uniform time intervals of 10 minutes. To achieve this, linear interpolation is applied to estimate the position of the vessel at each fixed point in time, thereby unifying the temporal resolution.

[0024] Finally, data points corresponding to vessels that are stationary or moving very slowly are filtered out. AIS messages with a ground speed of less than 0.5 knots are excluded from the analysis as they indicate stationary or minimal drift.

[0025] Through such preprocessing procedures, the ship track data applied in this invention is composed of a clean and consistent dataset suitable for marine engineering research and analysis of ship movement patterns.

[0026] In step S120, the track prediction device performs DBSCAN (Density-Based Spatial Clustering of Applications with Noise) on feature vectors including latitude (lat), longitude (lon), speed (sog), and azimuth (cog) from AIS time series data to group data points with similar region and movement characteristics into the same cluster.

[0027] Through this, a cluster ID is assigned as a static feature to each time series sample, and this cluster information is included in the input of a TS-Mixer-based prediction model and utilized for learning local and behavioral patterns.

[0028] For example, for each time series sample, the corresponding cluster ID can be assigned as a static feature as shown in the following mathematical formula.

[0029]

[0030] Cluster IDs are injected into the model as static input vectors, enabling the incorporation of local and behavioral patterns of the data into the learning process.

[0031] In this invention, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), a density-based clustering algorithm, is applied to effectively reflect the regional and behavioral patterns of AIS data.

[0032] DBSCAN identifies high-density regions in the data space as clusters, and the condition is that at least a minimum number of points (minPts) within a cluster must exist within a specific radius (ε, epsilon). As a result, DBSCAN does not require prior information regarding the shape or size of the clusters and can effectively separate noise and outliers.

[0033] Below, we will examine the core concepts and formula definitions of DBSCAN.

[0034] 1. ε-neighborhood

[0035] For a point p, the set of points contained within radius ε is defined as follows.

[0036]

[0037] Here, D is the data set, and dist(p, q) is the distance function between point p and point q. For example, it is the Euclidean distance.

[0038] 2. Core Point

[0039] The condition for point p to become a key point is that there are at least minPts points within the ε-neighborhood.

[0040]

[0041] 3. Directly Density-Reachable Connection

[0042] If point q is an ε-neighbor of core point p, then q is said to be directly and densely connected from p.

[0043] 4. Density-Reachable Connection

[0044] That point q is densely connected to point p means that starting from p, one can reach q by following a series of directly densely connected points.

[0045] Evaluation metrics include MSE (Mean Squared Error) and RMSE (Root Mean Square Error).

[0046] MSE is the average of the squared errors between the predicted value and the actual value, and it is an indicator that is sensitive to large errors.

[0047]

[0048] RMSE is the square root of MSE, an indicator that provides an error value consistent with actual units and is sensitive to outliers.

[0049]

[0050] In step S130, the track prediction device performs learning by applying independent MLPs in the time-axis and feature-axis directions, respectively, to the grouped AIS time series data using a CL-MLP (Class Label-guided Multi-Layer Perceptron) model.

[0051]

[0052] This structure can enable effective learning of time-series data while making it lightweight.

[0053]

[0054] In this invention, a Class Label-guided Multi-Layer Perceptron (CL-MLP) model based on a Times Series Mixer (TS-Mixer) structure was designed to effectively process AIS time series data. The TS-Mixer is a structure that maximizes expressiveness by independently mixing the time axis and feature axis of the time series input, and possesses characteristics that are more suitable for time series than existing MLPs.

[0055] CL-MLP is designed to achieve higher classification accuracy and generalization performance by inheriting the structural advantages of TS-Mixer while incorporating class information (class labels) into the network flow. In particular, by utilizing class information as an induction signal during the mixing process of time and variable axes, it enables the model to incorporate not only the time-series patterns of the input data but also the distinguishing characteristics between classes into the training.

[0056] Furthermore, this model automatically extracts fixed static features using the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm. By generating density-based clustering information from raw time-series data and adding this information to the model input, it enables the simultaneous learning of not only time-series patterns but also fixed structural characteristics.

[0057] Through such structural improvements, CL-MLP demonstrates superior performance compared to existing models in various time series classification problems, and it has been experimentally confirmed that it is capable of effectively handling environments where data distributions are complex or non-linear.

[0058] The Feature Mixing (FM) block is the input matrix It is a 2-layer Residual MLP that operates on a row-by-row basis. This block shares the same weights (parameters) for all rows, and each row X j,* It is designed to perform feature transformation on. Each row X j,* The following operations are applied to .

[0059]

[0060] Here, , , σ(·) is a non-linear activation function (e.g., ReLU), Drop(·) is Dropout applied, Norm(·) is Layer Normalization, L is the time series length (number of rows), and C is the feature dimension.

[0061] If a change in the output dimension is required (C≠H), that is, when projecting the feature dimension onto H, it is modified as follows.

[0062]

[0063] Here,

[0064] Conditional Feature Mixing (CFM) blocks are static characteristics input sequence Consider together with. When processing time series data, this block transforms hidden features by utilizing static features as conditions.

[0065] After extending static characteristics to fit the time series, perform the following operations.

[0066]

[0067] Here, Expand L (S) replicates S for a time series length L and Expand to, is the result of concatenating X and V along the feature axes, and is a linear transformation that projects static characteristics onto the time-series hidden dimension.

[0068] In step S140, the trajectory prediction device predicts the future trajectory within a preset prediction time range using the learned results.

[0069] In other words, the track prediction device can regressively predict future position, velocity, and direction items from prediction time point T+1 to T+σ. The prediction results can be calculated in latitude and longitude coordinates or Haversine distance-based coordinates.

[0070] In step S150, the track prediction device performs post-processing on the predicted value of the predicted future track according to preset resolution criteria or speed and route constraints.

[0071] Here, the predicted future trajectory based on the post-processed prediction value can be linked to a visualization tool or a Vessel Traffic Service (VTS).

[0072] As a result of conducting comparative experiments with various time series forecasting models, it was confirmed that the CL-MLP model proposed in this invention exhibited the best performance. Mean Squared Error (MSE) and Root Mean Square Error (RMSE) were used as key evaluation indicators for evaluating the model's performance.

[0073] In addition, Figures 3 and 4 visually compare the actual prediction results of each model, and it can be seen that the CL-MLP model performs the most stable and accurate predictions, even though the overall prediction accuracy tends to decline in sharp turning sections.

[0074]

[0075] FIG. 5 is a diagram schematically illustrating the configuration of a trajectory prediction device using MLP-based time series prediction according to an embodiment of the present invention.

[0076] Referring to FIG. 5, a track prediction device using MLP-based time series prediction according to an embodiment of the present invention includes a processor (10), a memory (20), a communication unit (30), and an interface unit (40).

[0077] The processor (10) may be a CPU or a semiconductor device that executes processing instructions stored in memory (20).

[0078] The memory (20) may include various types of volatile or non-volatile memory media. For example, the memory (20) may include ROM, RAM, etc.

[0079] For example, memory (20) can store instructions for performing a track prediction method using MLP-based time series prediction according to an embodiment of the present invention.

[0080] The communication unit (30) is a means for transmitting and receiving data with other devices through a communication network.

[0081] The interface section (40) may include a network interface and a user interface for connecting to a network.

[0082] Meanwhile, the components of the aforementioned embodiments can be easily identified from a process perspective. That is, each component can be identified as a respective process. Furthermore, the processes of the aforementioned embodiments can be easily identified from the perspective of the device components.

[0083] In addition, the technical details described above may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either individually or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiments, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. Hardware devices may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

[0084] The embodiments of the present invention described above are disclosed for illustrative purposes only, and those skilled in the art with ordinary knowledge of the present invention may make various modifications, changes, and additions within the spirit and scope of the present invention, and such modifications, changes, and additions should be considered to fall within the scope of the following claims. Explanation of the symbols

[0085] 10: Processor 20: Memory 30: Communications Department 40: Interface section

Claims

Claim 1 A trajectory prediction method using MLP-based time series prediction performed by a trajectory prediction device comprises: a step of collecting data from an AIS (Automatic Identification System) to construct AIS time series data including latitude (lat), longitude (lon), speed (sog), and azimuth (cog); a step of performing DBSCAN (Density-Based Spatial Clustering of Applications with Noise) on feature vectors including latitude (lat), longitude (lon), speed (sog), and azimuth (cog) in the AIS time series data to group data points with similar region and movement characteristics into the same cluster and generate cluster labels corresponding to each cluster; a step of learning the AIS time series data using a CL-MLP (Class Label-guided Multi-Layer Perceptron) model by using the cluster labels as class labels for supervised learning; wherein the CL-MLP model comprises a first multi-layer perceptron network that learns the time-axis direction time series pattern of the AIS time series data and A trajectory prediction method using MLP-based time series prediction, comprising: a second multilayer perceptron network that learns the correlation between variables in the feature-axis direction of AIS time series data, and a step of predicting a future trajectory within a preset prediction time range by combining the learning results of the first multilayer perceptron network and the second multilayer perceptron network; and a step of performing post-processing on the predicted value of the predicted future trajectory according to preset resolution criteria or speed and route constraints. Claim 2 A method for predicting a trajectory using MLP-based time series prediction, wherein the step of configuring the AIS time series data in claim 1 comprises removing data with a speed of 30 knots or more, removing data indicating that the navigation status of the vessel is anchored or moored, segmenting the vessel trajectory based on cases where the time interval between consecutive AIS messages exceeds 2 hours, excluding from analysis trajectory segments where the number of AIS messages is less than 20 or the total duration is less than 4 hours, resampling all trajectory data at uniform time intervals of 10 minutes, and excluding AIS messages with a speed of less than 0.5 knots from analysis. Claim 3 A trajectory prediction method using MLP-based time series prediction, wherein the generating step in claim 1 is characterized by assigning a cluster ID as a static characteristic to each time series sample through the grouping. Claim 4 A trajectory prediction method using MLP-based time series prediction, characterized in that, in claim 1, the CL-MLP model is based on a TS-Mixer (Times Series Mixer) structure, and the TS-Mixer is a structure that maximizes expressiveness by independently mixing the time axis and the characteristic axis of the time series input. Claim 5 A trajectory prediction method using MLP-based time series prediction, wherein the predicting step in claim 1 predicts items of future position, speed, and direction, and calculates the prediction result in latitude and longitude coordinates or Haversine distance-based coordinates. Claim 6 In a trajectory prediction device using MLP-based time series prediction, a memory for storing instructions; and includes a processor that executes the above instructions, wherein the instructions comprise: a step of collecting data from an AIS (Automatic Identification System) to construct AIS time series data including latitude (lat), longitude (lon), speed (sog), and azimuth (cog); a step of performing DBSCAN (Density-Based Spatial Clustering of Applications with Noise) on feature vectors including latitude (lat), longitude (lon), speed (sog), and azimuth (cog) in the AIS time series data to group data points with similar region and movement characteristics into the same cluster and generate cluster labels corresponding to each cluster; and a step of learning the AIS time series data using a CL-MLP (Class Label-guided Multi-Layer Perceptron) model by using the cluster labels as class labels for supervised learning; wherein the CL-MLP model comprises a first multi-layer perceptron network that learns the time-axis direction time series pattern of the AIS time series data and the AIS time series data A trajectory prediction device using MLP-based time series prediction, characterized by performing a trajectory prediction method comprising: a second multilayer perceptron network that learns correlations between variables in the direction of a feature axis, and a step of predicting a future trajectory within a preset prediction time range by combining the learning results of the first multilayer perceptron network and the second multilayer perceptron network; and a step of performing post-processing on the predicted value of the predicted future trajectory according to preset resolution criteria or speed and route constraints.

Images