CN-122020398-A - AIS data-based automatic identification method and system for ship loitering behaviors

CN122020398ACN 122020398 ACN122020398 ACN 122020398ACN-122020398-A

Abstract

The invention provides an automatic identification method and an automatic identification system for ship loitering behaviors based on AIS data, which are characterized in that firstly, track segmentation, cleaning, resampling and space filtering are carried out on original AIS data to obtain a continuous navigation track segment, self-intersecting detection and multi-scale sliding window analysis are combined, a candidate loitering track is screened by utilizing an isolated forest and manually marked, a balance training set of BIO marks is constructed, multidimensional feature vectors comprising plane coordinates, accumulated distances, course period components, logarithmic navigational speeds and time stamps are calculated, training is carried out by adopting a double-branch deep learning sequence marking model, a Focal Loss and Dice Loss function is fused, the tracks to be identified are subjected to the same pretreatment, then the model is input, a point-by-point label is output, and finally, accurate loitering track segments are analyzed according to BIO rules, so that the high-precision, fine granularity and automatic positioning of ship loitering behaviors are realized, and the method is remarkably superior to the existing coarse granularity classification or threshold judgment method.

Inventors

TU ENMEI
XIE ZHIYE
FU XIANPING
HAN YI

Assignees

中远海运科技股份有限公司
大连海事大学

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. An automatic identification method for ship loitering behaviors based on AIS data is characterized by comprising the following steps: Acquiring original AIS data of a ship, extracting a plurality of key fields from the original AIS data, sorting the key fields according to time sequence to form an original AIS track sequence of the ship, and preprocessing the original AIS track sequence to divide the original AIS track sequence of the same ship into a plurality of historical track sections formed by continuous navigation points, wherein the preprocessing sequentially comprises track segmentation processing, data cleaning processing, resampling and difference processing and space filtering processing; Marking a training data set construction step, namely identifying a self-intersecting area in a history track section of a certain ship by adopting a self-intersecting point detection method, sliding along the history track section by adopting a multi-scale sliding window method according to a plurality of windows with different time intervals, calculating course change frequency, displacement distance and track tortuosity in each window, calculating the loiter score of each window according to the course change frequency, the displacement distance and the track tortuosity, taking the maximum value in the loiter score of each window as the final loiter score of the history track section, automatically screening candidate loiter tracks by adopting an isolated forest algorithm, verifying and screening the candidate loiter tracks to obtain a plurality of track sections, marking each track point in the loiter track section by adopting a BIO marking method, wherein the label comprises a starting point of the track section, a middle point of the loiter track section and a normal navigation point; Calculating a multidimensional feature vector of each track point in the labeling training data set based on the labeling training data set, and normalizing the multidimensional feature vector, wherein the multidimensional feature vector comprises an absolute plane coordinate feature, a course angle period component feature, a navigational speed transformation feature, a timestamp feature and an accumulated navigational distance feature; The model training step is that the normalized multidimensional feature vector is input into a deep learning sequence labeling model for training, and a mixed Loss function comprising a Focal Loss function and a Dice Loss function is adopted as a Loss function in the deep learning sequence labeling model training, so that a trained deep learning sequence labeling model is obtained; The behavior identification and output step comprises the steps of executing preprocessing and multidimensional feature vector calculation processing which are the same as those of an original AIS track sequence on a ship track sequence to be identified, inputting the obtained multidimensional feature vector after normalization processing into a trained deep learning sequence labeling model, wherein the trained deep learning sequence labeling model comprises a parallel double-branch structure, a sequence attention mechanism, a channel attention mechanism and a sequence semantic encoder, the double-branch structure comprises a spatial feature processing branch and a dynamic feature processing branch, the spatial feature processing branch is used for receiving normalized absolute plane coordinate features and accumulated distance features, the spatial feature processing branch sequentially passes through a one-dimensional convolution layer and the sequence attention mechanism to generate a spatial feature sequence representing a local track form, the dynamic feature processing branch is used for receiving normalized heading angle periodic component features, navigational speed transformation features and timestamp features, sequentially passes through a feedforward neural network and the sequence attention mechanism to generate a dynamic feature sequence representing dynamic navigation behavior, splicing the spatial feature sequence and the dynamic feature sequence in the channel dimension mechanism to generate a fusion feature sequence, and the channel attention adaptive fusion feature sequence is used for carrying out weighted feature sequence and self-adaptive to the fusion feature sequence, and outputting the weighted feature sequence to a navigation feature prediction starting point of the navigation behavior under a navigation point or a navigation point of a ship, so that the navigation point belongs to a normal navigation point is identified.
2. The automatic identification method for the loitering behavior of the ship based on AIS data, which is characterized by further comprising the steps of performing super-parameter adjustment and model selection by adopting a pre-divided verification set in the deep learning sequence labeling model training process to obtain a trained and adjusted deep learning sequence labeling model, and evaluating the performance of the trained and adjusted deep learning sequence labeling model by using a divided test set through evaluation indexes, wherein the evaluation indexes comprise average intersection ratio, detection rate and error rate.
3. The automatic identification method for loitering behaviors of a ship based on AIS data according to claim 1, wherein in the data acquisition and preprocessing step, the preprocessing specifically comprises: the method comprises the steps of firstly carrying out track segmentation processing, namely dividing an original AIS track sequence of adjacent track points of the same ship, wherein the time interval of the original AIS track sequence exceeds a preset time threshold value, forming a plurality of track segments with continuous time, carrying out point-by-point cleaning on each track segment, removing track points which comprise null values, abnormal values, coordinate jumps or navigation states which are moored and moored, carrying out resampling and difference processing on the cleaned track segments by adopting a combination of spherical linear difference and standard linear difference, and finally carrying out space filtering processing, namely filtering track points positioned in a inland river or a navigation channel by using a land mask, and only keeping track points of an open water area.
4. The automatic identification method for vessel loitering behavior based on AIS data according to claim 1, wherein in the labeling training data set construction step, selecting a plurality of normal track segments from the normal running track of the vessel based on a K-means clustering algorithm specifically comprises: And dividing a plurality of normal navigation modes by adopting a K-means clustering algorithm based on the clustering feature vector, extracting normal track segments from each normal navigation mode, and finally obtaining a plurality of normal track segments.
5. The automatic identification method for the loitering behavior of the ship based on AIS data according to claim 1, wherein in the multi-dimensional feature vector calculation step, the absolute plane coordinate feature comprises the steps of taking a starting point of an original AIS track sequence as a coordinate origin, calculating a geographic distance between adjacent track points based on longitude and latitude coordinates of the adjacent track points and by utilizing a HAVERSINE formula, calculating an azimuth angle between the adjacent track points based on the longitude and latitude coordinates of the adjacent track points by utilizing an azimuth angle formula, converting the geographic distance and the azimuth angle into relative coordinate increment under a two-dimensional plane rectangular coordinate system by utilizing a triangular decomposition method, and obtaining absolute plane coordinates of each track point relative to the origin by accumulating the relative coordinate increment; calculating the accumulated geographic distance from the current track point to the track sequence starting point, and taking the accumulated geographic distance as an accumulated navigation distance feature; Converting a course angle into a course angle sine component and a course angle cosine component to be used as course angle periodic component characteristics; the speed transformation feature comprises that the speed is subjected to logarithmic transformation to obtain the logarithmic transformed speed as the speed transformation feature; The time stamp feature comprises the step of normalizing the time stamp of each track point relative to the reference to a [0,1] interval by taking the starting time of the track sequence as the reference to obtain the time stamp feature.
6. The automatic identification method for ship loitering behavior based on AIS data according to claim 1, wherein in the behavior identification and output step, the sequential attention mechanism comprises a parallel sequential pooling path and a local convolution path, the sequential pooling path generates sequential attention weights through global average pooling and 1×1 convolution, the local convolution path extracts local context features through 3×1 one-dimensional convolution, and a final time attention weight map is generated through a cross-space learning mechanism after two paths are fused and position-coded.
7. An AIS data-based automatic identification system for ship loitering behaviors is characterized by comprising a data acquisition and preprocessing module, a labeling training data set construction module, a multidimensional feature vector calculation module, a model training module and a behavior identification and output module which are connected in sequence, The data acquisition and preprocessing module acquires original AIS data of a ship, extracts a plurality of key fields from the original AIS data, sequences the key fields according to time sequence to form an original AIS track sequence of the ship, and preprocesses the original AIS track sequence to divide the original AIS track sequence of the same ship into a plurality of historical track sections formed by continuous navigation points, wherein the preprocessing sequentially comprises track segmentation processing, data cleaning processing, resampling and difference processing and space filtering processing; The marking training data set construction module is used for identifying a self-intersecting region in a history track section of a certain ship by adopting a self-intersecting point detection method, sliding along the history track section by adopting a multi-scale sliding window method according to a plurality of windows with different time intervals, calculating course change frequency, displacement distance and track tortuosity in each window, calculating the loiter score of each window according to the course change frequency, the displacement distance and the track tortuosity, taking the maximum value in the loiter score of each window as the final loiter score of the history track section, automatically screening candidate loiter tracks by adopting an isolated forest algorithm, verifying and screening the candidate loiter tracks to obtain a plurality of loiter track sections, marking each track point in the loiter track section by adopting a BIO marking method, wherein the label comprises a starting point of the loiter track section, a middle point of the loiter track section and a normal navigation point; the multi-dimensional feature vector calculation module is used for calculating a multi-dimensional feature vector of each track point in the labeling training data set based on the labeling training data set and carrying out normalization processing on the multi-dimensional feature vector, wherein the multi-dimensional feature vector comprises an absolute plane coordinate feature, a course angle period component feature, a navigational speed transformation feature, a timestamp feature and an accumulated navigational distance feature; The model training module inputs the normalized multidimensional feature vector into a deep learning sequence labeling model for training, and adopts a mixed Loss function comprising a Focal Loss function and a Dice Loss function as a Loss function in the deep learning sequence labeling model training to obtain a trained deep learning sequence labeling model; The behavior recognition and output module is used for executing preprocessing and multidimensional feature vector calculation processing which are the same as those of the original AIS track sequence on the ship track sequence to be recognized, and inputting the multidimensional feature vector obtained by normalization processing into a trained deep learning sequence labeling model; the trained deep learning sequence labeling model comprises a parallel double-branch structure, a sequence attention mechanism, a channel attention mechanism and a sequence semantic encoder, wherein the double-branch structure comprises a spatial feature processing branch and a dynamic feature processing branch, the spatial feature processing branch is used for receiving normalized absolute plane coordinate features and accumulated navigation distance features, generating a spatial feature sequence representing a local track form sequentially through a one-dimensional convolution layer and the sequence attention mechanism, the dynamic feature processing branch is used for receiving normalized course angle periodic component features, navigation speed conversion features and time stamp features, sequentially passing through a feedforward neural network and the sequence attention mechanism and generating a dynamic feature sequence representing dynamic navigation behaviors, the spatial feature sequence and the dynamic feature sequence are spliced in a channel dimension to generate a fusion feature sequence, the channel attention mechanism carries out self-adaptive weighting on the fusion feature sequence, inputs the weighted fusion feature sequence into the sequence semantic encoder to carry out context modeling, and outputs a prediction tag of each track point belonging to a navigation behavior starting point, a navigation behavior middle point or a normal navigation point, and further realizes automatic recognition of the ship navigation behavior.
8. The automatic identification system for the loitering behavior of the ship based on AIS data, which is characterized in that in the model training module, super-parameter adjustment and model selection are further carried out by adopting a pre-divided verification set in the deep learning sequence labeling model training process to obtain a trained and adjusted deep learning sequence labeling model, the performance of the trained and adjusted deep learning sequence labeling model is evaluated through an evaluation index by using a divided test set, and the evaluation index comprises average intersection ratio, detection rate and error rate.
9. The automatic identification system for vessel loitering behavior based on AIS data according to claim 7, wherein in the labeling training data set construction module, selecting a plurality of normal track segments from the normal running track of the vessel based on a K-means clustering algorithm specifically comprises: And dividing a plurality of normal navigation modes by adopting a K-means clustering algorithm based on the clustering feature vector, extracting normal track segments from each normal navigation mode, and finally obtaining a plurality of normal track segments.
10. The automatic identification system for vessel loitering behavior based on AIS data according to claim 7, wherein the absolute plane coordinate feature comprises calculating a geographic distance between two adjacent track points based on longitude and latitude coordinates of the two adjacent track points by using a HAVERSINE formula with a starting point of an original AIS track sequence as a coordinate origin, calculating an azimuth angle between the two adjacent track points based on longitude and latitude coordinates of the two adjacent track points by using an azimuth angle formula, converting the geographic distance and the azimuth angle into relative coordinate increment in a two-dimensional plane rectangular coordinate system by using a triangular decomposition method, and obtaining absolute plane coordinates of each track point relative to the origin by accumulating the relative coordinate increment; calculating the accumulated geographic distance from the current track point to the track sequence starting point, and taking the accumulated geographic distance as an accumulated navigation distance feature; Converting a course angle into a course angle sine component and a course angle cosine component to be used as course angle periodic component characteristics; the speed transformation feature comprises that the speed is subjected to logarithmic transformation to obtain the logarithmic transformed speed as the speed transformation feature; The time stamp feature comprises the step of normalizing the time stamp of each track point relative to the reference to a [0,1] interval by taking the starting time of the track sequence as the reference to obtain the time stamp feature.

Description

AIS data-based automatic identification method and system for ship loitering behaviors Technical Field The invention relates to the technical field of maritime monitoring and intelligent shipping, in particular to an automatic identification method and system for ship loitering behaviors based on AIS data. Background The ship AIS track sequence data refers to dynamic and static information including position, navigational speed and heading acquired by a ship Automatic Identification System (AIS), and the dynamic and static information are connected in space coordinates according to time sequence to form a track sequence, so that the actual navigational path and the movement state of the ship are completely reflected. Loitering behavior of a ship refers to a movement pattern in which the ship frequently changes heading in a certain geographical range during the course of the ship, resulting in a track exhibiting high redundancy and lacking an explicit forward trend. Loitering tracks are commonly found on types of ships such as fishing boats, tugboats and law enforcement boats, reflect specific activity intentions, such as fishing operation, towing operation, sea area patrol, waiting for berthing or illegal activities, and timely and accurately identify loitering behavior of the ships, navigation safety guarantee in maritime, abnormal early warning and other tasks. In contrast, cargo vessels and tankers and other transportation vessels generally follow a preset efficient route, and any unnecessary loitering behavior can lead to transportation delay, fuel waste and rising operation cost, which affect the economic benefit. Therefore, the ship loitering behavior is automatically identified and positioned based on AIS track data, and the method has important significance in improving maritime supervision efficiency and guaranteeing navigation safety. The existing common ship AIS track loitering behavior recognition method mainly comprises an unsupervised method and a supervised method, wherein 1) in the unsupervised recognition method, judgment is generally carried out by relying on a preset track characteristic threshold value or a segmentation rule, and the method mainly comprises typical technical paths such as a space-time grid analysis method, a track refining analysis method, a sliding window analysis method and the like. The space-time grid analysis method performs gridding treatment through AIS data, converts the track points into grid coordinates, and extracts grid features to identify the loitering behavior of the ship. The track refining analysis method uses a Douglas-Peucker algorithm to compress the track and reserves key points to reduce redundant data. Calculating the curvature of the key points, identifying wave points (local maximum points of the curvature), and finally judging the loitering behavior of the ship through angle constraint and a wave point number threshold value. The sliding window analysis method traverses the track through a multi-scale time window, calculates quantization indexes built based on space-time characteristics in the window, and realizes loitering judgment by using an anomaly detection algorithm such as an isolated forest. However, although the unsupervised method does not need to annotate data, the unsupervised method is highly dependent on a manually set threshold value or a track segmentation rule, and generally only can judge whether the whole track presents the loitering characteristic, and cannot automatically and accurately locate a specific period and a specific position where loitering occurs. 2) In the supervised recognition method, a plurality of loitering modes are considered to be set at first, and then tracks conforming to the preset loitering modes are marked according to AIS track shapes so as to form a data set training classification model. For example, in the prior art, it is proposed to convert a detected loitering track segment into a standardized gray level image, construct a Convolutional Neural Network (CNN) model, and automatically learn morphological features of a track through operations such as multi-layer convolution and pooling, so as to perform classification and identification on four typical loitering track shapes such as a disordered fold-back shape, a lasso shape, a regular reciprocation shape and a random coil shape. Although the existing supervised method can identify specific loitering forms, the core function of the existing supervised method is remained in classification of the whole shape of the loitering track in the past, but not in accurate point-by-point positioning of loitering fragments, so that the actual requirements can not be well met. In summary, the existing method for identifying the loitering behavior of the AIS track has double limitations. In the detection aspect, the unsupervised method is highly dependent on a manually preset threshold value or a segmentation rule, and can only carry out coarse granularity judg