A fishing vessel type recognition method based on semantic feature vector
By combining Word2Vec word vector model and LightGBM classification model, the problems of low accuracy and insufficient feature extraction in fishing boat type recognition are solved, and high accuracy and stability of fishing boat type recognition is achieved, and the feature extraction process is simplified.
Patent Information
- Application Number
- CN202111435653.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-29
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-11-29
AI Technical Summary
In the prior art, the problem of low accuracy of fishing boat type recognition, insufficient feature extraction and complex feature construction, especially in deep learning and machine learning methods based on VMS trajectory data, there are low classification accuracy and complex feature engineering. question.
The fishing boat type recognition method based on Word2Vec word vector model is used to extract semantic feature vectors and combine LightGBM classification model. Through data preprocessing, semantic feature extraction and classification model training, the fishing boat operation type recognition is realized. This method includes track extraction and data repair in the data preprocessing stage, latitude, longitude, velocity direction and sea area grid feature extraction in the semantic feature extraction stage, and LightGBM model application in the classification model training stage.
It improves the classification accuracy of fishing boat type recognition, reduces the workload of manual feature extraction, enhances the stability and accuracy of the recognition system, and avoids the problems of poor computational complexity and poor interpretability in deep learning.
Smart Images

Figure CN114155446B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of marine vessel supervision, and in particular relates to a method for identifying fishing vessel types based on semantic feature vectors. The method extracts semantic feature vectors based on a Word2Vec word vector model and combines a LightGBM classification model to complete fishing vessel type identification. Background Art
[0002] Quantitative research on marine fisheries and monitoring of illegal fishing activities are key technologies for marine resource management. Current quantitative fishing research is based on a single fishing vessel operation type, but fishing vessels actually operate in a variety of ways, so identifying the type of fishing vessel operation based on the Vessel Monitoring System (VMS) trajectory is the first problem to be solved.
[0003] The satellite-based vessel monitoring system is an essential system for maritime navigation. It was originally used to enhance maritime navigation safety. As a means of monitoring fishing vessels, it also provides a new data source for fishery scientific research. VMS records mainly include real-time navigation data of fishing vessels such as vessel number, timestamp, location, and instant speed. With the development of big data and in-depth research on trajectory data mining, research based on VMS data has become an important means of fishery behavior mining and related applications.
[0004] In recent years, the classification of fishing vessel operation types based on VMS trajectory data has mainly focused on two directions: automatic feature extraction by neural networks combined with deep learning, and manual feature extraction combined with machine learning. Automatic feature extraction by neural networks combined with deep learning mainly includes converting fishing vessel trajectories into images and inputting them into convolutional neural networks (CNN) for image classification, and converting trajectories into sequence data and inputting them into recurrent neural networks (RNN) for sequence data modeling. However, converting fishing vessel trajectories into images and using convolutional neural networks will lose important features such as the speed and direction of the fishing vessel, and there are problems such as low classification accuracy and few classification categories; directly inputting VMS trajectory data into recurrent neural networks after serialization lacks the process of modeling the characteristics of fishing vessel trajectory data, and there are problems such as low classification accuracy. In general, training VMS trajectory data based on deep neural networks can realize feature engineering automation without the need to study manual feature design, but the training is complex, computationally intensive, time-consuming, and deep learning has poor interpretability. Manual feature extraction combined with machine learning mainly models classic features and extracts statistics such as the mean, median, and variance of features such as fishing boat speed, heading, and steering angle. This method does not take into account the spatiotemporal characteristics of fishing boat trajectory data and has the problem of insufficient feature extraction. In addition, studying different operating modes of fishing boats requires a lot of domain expert knowledge, and manually extracting high-order combined features from raw data consumes the researchers' energy, making it impossible to focus on studying algorithm models, and there is a problem of complex artificial feature construction. Summary of the invention
[0005] One purpose of the present invention is to solve the technical problem of complex artificially constructed features, insufficient feature extraction and low accuracy of deep learning classification. A fishing vessel type recognition method based on the Word2Vec word vector model to extract semantic feature vectors combined with the LightGBM classification model is proposed to identify the operation type of fishing vessels and provide a reference for further scientific research.
[0006] The technical solution adopted by the present invention is a fishing boat type recognition method based on extracting semantic feature vectors based on the Word2Vec word vector model and combining with the LightGBM classification model, comprising the following steps:
[0007] S1, data preprocessing stage:
[0008] 1-1 Collect the historical VMS track data of the fishing vessel with the fishing vessel operation type mark and convert it into Track = {points i (lat, lon, speed, time, dir, type)|i=1,2,...,N} format, where Track represents the set of track points of a single fishing boat, pointsi represents the trajectory point at the i-th moment, lat represents latitude, lon represents longitude, speed represents instantaneous speed, time represents the time when the trajectory point is collected, dir represents the current direction of the fishing vessel, N represents the maximum moment, and type represents the type of fishing vessel operation, including: trawl type, purse seine type and gillnet type;
[0009] 1-2 Using a track extraction algorithm based on the ship position track for the data processed in step 1-1, extract the complete track of the fishing boat entering and leaving the port;
[0010] 1-3 The voyage trajectory of the fishing boat extracted in step 1-2 is repaired by using a fishing boat trajectory repair algorithm based on speed constraints to obtain the fishing boat trajectory data after voyage extraction and trajectory repair;
[0011] 1-4 According to steps 1-1 to 1-3, multiple fishing boat repair voyage trajectory sets are constructed, which are represented as fishing boat trajectory dataset Data = {T v |v=1,2,···,V}, where T v represents the trajectory set of the vth fishing boat after repair, and V represents the number of fishing boats;
[0012] S2. Extract semantic features from the fishing boat trajectory dataset Data using a word embedding-based semantic feature extraction model; the word embedding-based semantic feature extraction model includes a word embedding model based on longitude and latitude, a word embedding model based on speed direction, a word embedding model based on sea area grid, and a concatenation layer; the longitude and latitude word embedding model extracts features from longitude and latitude; the speed direction word embedding model extracts features from speed and direction; the sea area grid word embedding model extracts features from the sea area where the fishing boat passes;
[0013] S3, classification model training stage:
[0014] Firstly, the trajectory feature vectors obtained in the S2 semantic feature extraction stage are divided into training set and test set. Then, the training set is input into the LightGBM classifier for training, and the test set is used to test the LightGBM classifier. Finally, the trained LightGBM classifier is used to realize VMS fishing vessel trajectory recognition of fishing vessel operation type.
[0015] Another object of the present invention is a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.
[0016] Yet another object of the present invention is a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described is implemented.
[0017] The beneficial effects of the present invention are:
[0018] 1. The present invention adopts a feature extraction method based on word embedding, which can actively extract high-order features of fishing boat trajectories, avoid the problem of insufficient feature extraction in manually designed features, improve classification accuracy, eliminate the need for manual feature extraction, and reduce the workload of personnel.
[0019] 2. The present invention uses an effective stable classification model, designs a track extraction algorithm and a data repair algorithm to process the VMS fishing boat track, effectively reduces the impact of data noise on the classification results, and improves the stability of the recognition system. In addition, the application of neural networks to the recognition of fishing boat operation types improves the accuracy of classification compared to traditional classification methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 This is the process of the fishing vessel type identification method based on semantic feature vectors of the present invention.
[0021] Figure 2 This is a diagram illustrating the word embedding model based on latitude and longitude.
[0022] Figure 3 An illustration of the word embedding model based on speed direction.
[0023] Figure 4 This is a diagram illustrating the word embedding model based on ocean grids. DETAILED DESCRIPTION
[0024] The technical solution provided by the present invention is further described below in conjunction with the accompanying drawings.
[0025] See also Figure 1 The present invention provides a method for identifying fishing boat types based on extracting semantic feature vectors using a Word2Vec word vector model and combining it with a LightGBM classification model. Figure 1 The figure shows the overall architecture of the method. Generally speaking, the present invention includes three major steps: step S1, data preprocessing stage: processing the fishing vessel trajectory data through track extraction algorithm and data repair algorithm; step S2, semantic feature extraction stage: designing a semantic feature extraction model based on word embedding, using the data set of step S1 to extract features, and obtaining a semantic feature vector; step S3, classification model training stage: using the LightGBM classification model to train the semantic feature vector obtained in step S2 to realize the recognition of the fishing vessel operation type.
[0026] Step S1 data preprocessing stage includes: to solve the problem of large amount of historical VMS trajectory data, a track extraction algorithm based on ship position trajectory is proposed to process the historical VMS trajectory data and extract the trajectory of fishing boats entering and leaving the port; a data repair algorithm based on speed constraint is proposed to process the abnormal speed and abnormal position in the fishing boat trajectory and reduce the impact of data noise on fishing boat classification.
[0027] Figure 1 The data preprocessing stage provides a framework flow chart of the above step S1 in the method for identifying fishing boat types based on the Word2Vec word vector model extracting semantic feature vectors combined with the LightGBM classification model, which further includes the following steps:
[0028] 1-1 Collect the historical VMS track data of the fishing vessel with the fishing vessel operation type mark and convert it into Track = {points i (lat, lon, speed, time, dir, type)|i=1,2,...,N} format, where Track represents the track point set of a single fishing boat, points i represents the trajectory point at the i-th moment, lat represents latitude, lon represents longitude, speed represents instantaneous speed, time represents the time when the trajectory point is collected, dir represents the current direction of the fishing boat, N represents the maximum moment, and type represents the type of fishing boat operation, including: trawl type, purse seine type and gillnet type;
[0029] 1-2 The data processed in step 1-1 is subjected to a track extraction algorithm based on the ship position track to extract the complete track of the fishing boat entering and leaving the port; specifically:
[0030] 1-2-1: The port area is abstracted as a polygon, represented by port = {p k (lat, lon)|k=1,2,...,K}, the coordinates of the port centroid are calculated by the polygon centroid formula;
[0031] Where port indicates the specific port, p k represents the boundary points of the port, k represents the order of the boundary points of the port, and K represents the number of boundary points;
[0032] 1-2-2: For each track point, the current location information of the fishing boat is obtained according to the latitude lat and longitude lon, and then the distance D between the current location of the fishing boat and the centroid of the port is calculated according to the distance formula between two points i If D i ≤ threshold D0, then D i Updated to 0, otherwise unchanged; the threshold D0 is the radius of the maximum outer circle of the port area;
[0033] 1-2-3: Traverse all track points in turn and identify the departure and arrival points for each track point. Specifically:
[0034] Determine the current trajectory points i The distance D i Is it 0? If D i = 0, then continue to determine the distance D of the trajectory point at the next moment i+1 Is it 0? If D i+1 ≠0, then the trajectory points are considered i The position is the departure point, if D i+1 = 0, it is considered that the current fishing boat has not moved; if D i ≠0, then continue to the distance D of the next trajectory point i+1 Is it 0? If D i+1 = 0, then the trajectory points are considered i The position is the port entry point, if D i+1 ≠0, it is considered that the current fishing boat is in a moving state;
[0035] 1-2-4: According to the departure and arrival points of the fishing boat, the fishing boat track Track is intercepted to obtain a set of multiple voyage tracks of the fishing boat
[0036] Trajectory={trajectoryj(pointsi(shiptrackId,lat,lon,speed,time,dir,type)i=1,2,...,N)|j=1,2,...,M};
[0037] Trajectory is the set of voyage trajectories of a single fishing vessel. j is the jth voyage track of the current fishing vessel, shiptrackId represents the unique identification ID of the fishing vessel voyage track, and M represents the number of voyage tracks of a single fishing vessel.
[0038] 1-3 The voyage trajectory of the fishing boat extracted in step 1-2 is repaired using a fishing boat trajectory repair algorithm based on speed constraints to obtain the fishing boat trajectory data after voyage extraction and trajectory repair; specifically:
[0039] 1-3-1: Trajectory of fishing vessels entering and leaving the port j , under the premise of a confidence interval of 95%, the minimum instantaneous speed speed is obtained based on statistical methods min and maximum instantaneous speed max And the minimum average speed meanspeed between two adjacent trajectory points min and the maximum average speed meanspeed max ;
[0040] 1-3-2: Determine whether the instantaneous speed of each trajectory point falls within [speed min ,speed max ], if not, the instantaneous speed of the current trajectory point is considered to be an abnormal value, and the instantaneous speed is updated to a null value; otherwise, the instantaneous speed of the current trajectory point is considered to be a normal value, and no update operation is performed;
[0041] 1-3-3: Traverse the trajectory point set of the fishing boat
[0042] Use the haversine formula to calculate the points of two adjacent trajectory points after processing in step 1-3-2 i 、points i+1 The distance between them is then used to calculate the average speed meanspeed between the two trajectory points. (i,i+1) ; For the average speed meanspeed (i, i +1) Determine whether it falls within [meanspeed min , meanspeed max ], if not, then the trajectory point points i The longitude and latitude of the current track point are considered abnormal values, and the longitude and latitude are updated to null values. Otherwise, the longitude and latitude of the current track point are considered normal values and no update operation is performed;
[0043] 1-3-4: Use polynomial interpolation to process the trajectory after step 1-3-3 j ={points i (shiptrackId, lat, lon, speed, time, dir, type)|i=1, 2, ..., N} to complete the empty values and obtain the repaired voyage track t_repair j ={pointsi(shiptrackId,lat,lon,speed,time,dir)|i=1,2,...,N}, and then get the repaired voyage track set
[0044] T=
[0045] {t_repairj(pointsi(shiptrackId,lat,lon,speed,time,dir,type)i=1,2,...,N)|j=1,2,...,M}; where t_repair j It represents the j-th voyage track of a fishing vessel after repair.
[0046] 1-4 According to steps 1-1 to 1-3, multiple fishing boat repair voyage trajectory sets are constructed, which are represented as fishing boat trajectory dataset Data = {T v |v=1,2,...,V}, where T v represents the trajectory set of the vth fishing boat after repair, and V represents the number of fishing boats;
[0047] See also Figure 1 In the semantic feature extraction stage, the method for identifying fishing boat types based on the Word2Vec word vector model extracting semantic feature vectors combined with the LightGBM classification model includes the following steps:
[0048] Step S2: extracting semantic features from the fishing boat trajectory dataset Data using a word embedding-based semantic feature extraction model; wherein the word embedding-based semantic feature extraction model includes a word embedding model based on longitude and latitude, a word embedding model based on speed direction, a word embedding model based on sea area grids, and a splicing layer;
[0049] 2-1 Extract features of longitude and latitude based on the longitude and latitude word embedding model;
[0050] like Figure 2 The latitude and longitude word embedding model includes a geohash encoding feature extraction module, a gradient encoding feature extraction module, and a splicing layer;
[0051] The geohash coding feature extraction module is used to perform geohash coding on the voyage track and extract the geohash coding features to obtain the distributed feature vector of the voyage track corresponding to the shiptrackId; specifically:
[0052] (1) The latitude lat and longitude lon of each track point in the fishing boat track data set Data are encoded by the geohash algorithm to obtain the hash code D of the longitude and latitude of the fishing boat hashcode ;
[0053] (2) For D hashcode The TfidfVectorizer algorithm in the scikit-learn toolkit is used to calculate and convert it into a Tf-IDF feature matrix. Then, the TruncatedSVD algorithm is used to perform text topic analysis on the matrix to obtain the hash code topic vector of the voyage track corresponding to the shiptrackId.
[0054] (3) For D hashcode Use Word2vec word vector embedding to represent the distributed feature vector and get the hash code feature vector of the shiptrackId corresponding to the voyage track
[0055] The gradient coding feature extraction module is used to perform longitude and latitude gradient coding on the voyage track and extract the features of the longitude and latitude gradient coding to obtain the gradient coding theme vector of the voyage track corresponding to the shiptrackId. Specifically:
[0056] (1) Calculate the longitude and latitude gradients for each repaired voyage track in the fishing vessel track dataset Data, and then merge the data to obtain the encoding Data of the longitude and latitude gradients of the fishing vessel grad ;
[0057] (2) Data grad The TfidfVectorizer algorithm in the scikit-learn toolkit is used to calculate and convert it into a Tf-IDF feature matrix. Then, the TruncatedSVD algorithm is used to perform text topic analysis on the matrix to obtain the gradient encoded topic vector of the shiptrackId corresponding to the voyage track.
[0058] The splicing layer is used to splice the outputs of the geohash encoding feature extraction module and the gradient encoding feature extraction module to obtain a feature vector extracted based on the longitude and latitude word embedding model.
[0059] 2-2 As Figure 3 The speed direction feature is extracted based on the speed direction word embedding model; specifically:
[0060] 2-2-1: Encode the speed speed and direction dir in the fishing boat trajectory dataset Data into a string to obtain the string-encoded trajectory D str ={c o (shiptrackId,speed_dir)|o=1,2,…,O}, speed_dir is the string concatenation of speed and direction, and o is the number of track points in the fishing vessel track dataset;
[0061] 2-2-2: For D str Use Word2vec word vector embedding to represent the distributed feature vector. According to the fishing boat track sequence corresponding to shiptrackId, obtain the distributed feature vector corresponding to shiptrackId.
[0062] 2-3 As Figure 4 Based on the sea area grid word embedding model, feature extraction is performed on the sea area where fishing boats pass through; specifically:
[0063] 2-3-1: First, convert the lat and lon in the fishing boat trajectory dataset Data into Mercator coordinates, then divide the sea area where the fishing boat passes into grids in fine granularity and number the grids in sequence;
[0064] 2-3-2: According to the sea area grids that a single fishing boat passes through, each track point in each fishing boat voyage track is converted into a grid number sequence to obtain the sub-sea area grid sequence of each fishing boat The grid number represents the track of the β-th fishing vessel voyage;
[0065] 2-3-3: Data of the sea area grid sequence dataset serialize Input the Word2vec word vector model for training, and finally get the grid coded feature vector of a single fishing boat according to shiptraclId
[0066] 2-4 concatenate the feature vectors extracted by the three word embedding models to form a serial merge vector, and obtain the feature vector extracted by the word embedding semantic feature extraction model;
[0067]
[0068] See also Figure 1 In the classification model training stage, step S3 of the fishing boat type recognition method based on the Word2Vec word vector model extracting semantic feature vectors combined with the LightGBM classification model of the present invention includes the following steps:
[0069] Step S3: Use the LightGBM classification model to train the semantic feature vector obtained in step S24 to realize the recognition of the fishing vessel operation type, which further includes the following steps:
[0070] 3-1: First, the trajectory feature vector obtained in the S2 semantic feature extraction stage is divided into training set and test set according to 8:2; then the training set is input into the LightGBM classifier for training, and the LightGBM classifier is tested using the test set; finally, the trained LightGBM classifier is used to realize the VMS fishing vessel trajectory to identify the fishing vessel operation type.
[0071] The following is an introduction to the experiment and analysis process of identifying the ship type in a specific application example of the method of the present invention:
[0072] In order to solve the problems of complex artificially constructed features, insufficient feature extraction, and low accuracy of deep learning classification of VMS trajectory data, the present invention proposes a semantic feature extraction model based on word embedding. The model encodes the longitude, latitude, speed and direction features into text sequences and converts them into semantic feature extraction, and uses a deep neural network model to extract features. This method does not require artificial feature engineering, can fully express the temporal features and semantic features of VMS data, inputs the extracted features into the classification model, and experiments verify that the classification results have high accuracy. The feature extraction of the word embedding semantic feature extraction model is called Embedding-based feature extraction in this article, and the artificially constructed feature engineering is called Classic-based feature extraction in this article. This article adopts the Embedding-LightGBM model for fishing vessel type recognition. In order to facilitate the comparison of the classification effects of XGboost, Catboost and LightGBM, this article combines the feature extraction model and the classification model to compare the effects.
[0073] The experimental platform for ship type identification in this application is a server with Intel(R)Xeon(R)CPU E5-2630 v4@2.20GHz (20 cores) and 128GB memory. The experimental test data of the present invention are the VMS trajectories of some fishing vessels in the East China Sea fishing grounds of China from October 2018 to October 2020. These trajectories are recorded by the Beidou satellite system with a time resolution of 10min, and a total of 2236 registered fishing vessels are recorded for fishing activities. The total data set includes 22913 trajectories, and the training set and test set are divided into 8:2. Among them, there are 7624 purse seine trajectories, 7737 trawl trajectories, and 7552 gillnet trajectories.
[0074] Table 1
[0075]
[0076]
[0077] Note: Embedding—XGBoost represents the combination of word embedding-based semantic feature extraction model and XGBoost classification model.
[0078] Based on the comparative experimental results of the East China Sea fishing boat dataset, the F1-measure, precision and recall of the word embedding classification model can reach above 0.95, and the F1-measure, precision and recall of the classic feature classification model are above 0.88. These two sets of comparative experiments prove that the classification performance of the method based on word embedding-based feature extraction and classifier combination is higher than that of the method based on classic feature modeling. Since the principle of decision tree classification is to divide the categories based on feature importance, it is proved that the improvement of classification performance based on the Embedding-LightGBM model does not depend on the ability of the classifier, but the feature extraction model based on word embedding can fully extract effective features compared with the modeling based on classic features. As shown in Table 1, the comparison results of the three classification models based on word embedding feature extraction and classic feature modeling show that the overall classification performance of the three classification models is not much different. The LightGBM classifier model has higher performance than the other two classification methods, and the training time is lower, which fully proves the accuracy of using LightGBM as a classifier.
[0079] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.
Claims
1. A fishing vessel type recognition method based on semantic feature vectors, Features The method comprises the following steps: S1, data preprocessing stage: 1-1 Collect the historical VMS track data of the fishing vessel with the fishing vessel operation type mark and convert it into Track = {points i (lat, lon, speed, time, dir, type)|i=1,2,...,N}, where Track represents the track point set of a single fishing boat, points i represents the trajectory point at the i-th moment, lat represents latitude, lon represents longitude, speed represents instantaneous speed, time represents the time when the trajectory point is collected, dir represents the current direction of the fishing vessel, N represents the maximum moment, and type represents the type of fishing vessel operation, including: trawl type, purse seine type and gillnet type; 1-2 The data processed in step 1-1 is subjected to a track extraction algorithm based on the ship position track to extract the complete track of the fishing boat entering and leaving the port; specifically: 1-2-1: The port area is abstracted as a polygon, represented by port = {p k (lat, lon)|k=1,2,...,K}, the coordinates of the port centroid are calculated by the polygon centroid formula; where port represents a specific port, p k represents the boundary points of the port, k represents the order of the boundary points of the port, and K represents the number of boundary points; 1-2-2: For each track point, the current location information of the fishing boat is obtained according to the latitude lat and longitude lon, and then the distance D between the current location of the fishing boat and the centroid of the port is calculated according to the distance formula between two points i If D i ≤Threshold D 0 , then D i Update to 0, otherwise it remains unchanged; the threshold D 0 is the radius of the largest outer circle of the port area; 1-2-3: Traverse all track points in turn and identify the departure and arrival points for each track point. Specifically: Determine the current trajectory points i The distance D i Is it 0? If D i = 0, then continue to determine the distance D of the trajectory point at the next moment i+1 Is it 0? If D i+1 ≠0, then the trajectory points are considered i The position is the departure point, if D i+1 = 0, it is considered that the current fishing boat has not moved; if D i ≠0, then continue to the distance D of the next trajectory point i+1 Is it 0? If D i+1 = 0, then the trajectory points are considered i The position is the port entry point, if D i+1 ≠0, it is considered that the current fishing boat is in a moving state; 1-2-4: According to the departure and arrival points of the fishing boat, the fishing boat track Track is intercepted to obtain a set of multiple voyage tracks of the fishing boat Trajectory={trajectory j (points i (shiptrackId,lat,lon,speed,time,dir,type)i=1,2,...,N)|j=1,2,…,M}; Trajectory is the set of voyage trajectories of a single fishing vessel. j is the jth voyage track of the current fishing vessel, shiptrackId represents the unique ID of the voyage track of the fishing vessel, and M represents the number of voyage tracks of a single fishing vessel; 1-3 The voyage trajectory of the fishing boat extracted in step 1-2 is repaired using a fishing boat trajectory repair algorithm based on speed constraints to obtain the fishing boat trajectory data after voyage extraction and trajectory repair; specifically: 1-3-1: Trajectory of fishing vessels entering and leaving the port j , get the minimum instantaneous speed speed min and maximum instantaneous speed max And the minimum average speed meanspeed between two adjacent trajectory points min and the maximum average speed meanspeed max ; 1-3-2: Determine whether the instantaneous speed of each trajectory point falls within [speed min ,speed max ], if not, the instantaneous speed of the current trajectory point is considered to be an abnormal value, and the instantaneous speed is updated to a null value; otherwise, the instantaneous speed of the current trajectory point is considered to be a normal value, and no update operation is performed; 1-3-3: Traverse the trajectory point set trajectory of the fishing boat: Use the haversine formula to calculate the points of two adjacent trajectory points after processing in step 1-3-2 i 、points i+1 The distance between them is then used to calculate the average speed meanspeed between the two trajectory points. (i,i+1) ; For the average speed meanspeed (i,i+1) Determine whether it falls within [meanspeed min , meanspeed max ], if not, then the trajectory point points i The longitude and latitude of the current track point are considered abnormal values, and the longitude and latitude are updated to null values. Otherwise, the longitude and latitude of the current track point are considered normal values and no update operation is performed; 1-3-4: Use polynomial interpolation to process the trajectory after step 1-3-3 j ={points i (shiptrackId, lat, lon, speed, time, dir, type)|i=1,2,...,N} to complete the empty values and obtain the repaired voyage track t_repair j = {pointsi(shiptrackId, lat, lon, speed, time, dir)|i=1,2,...,N}, and then get the repaired voyage track set T={t_repair j (points i (shiptrackId, lat, lon, speed, time, dir, type)i=1,2,...,N)|j=1,2,...,M}; where t_repair j represents the j-th voyage track of a fishing vessel after repair; 1-4 According to steps 1-1 to 1-3, multiple fishing boat repair voyage trajectory sets are constructed, which are represented as fishing boat trajectory dataset Data = {T v |v=1,2,…,V}, where T v represents the trajectory set of the vth fishing boat after repair, and V represents the number of fishing boats; S2. Extract semantic features from the fishing boat trajectory dataset Data using a word embedding-based semantic feature extraction model; the word embedding-based semantic feature extraction model includes a word embedding model based on longitude and latitude, a word embedding model based on speed direction, a word embedding model based on sea area grid, and a concatenation layer; the longitude and latitude word embedding model extracts features from longitude and latitude; the speed direction word embedding model extracts features from speed and direction; the sea area grid word embedding model extracts features from the sea area where the fishing boat passes; The latitude and longitude-based word embedding model includes a geohash coding feature extraction module, a gradient coding feature extraction module and a splicing layer; The geohash coding feature extraction module is used to perform geohash coding on the voyage track and extract the geohash coding features to obtain the distributed feature vector of the voyage track corresponding to the shiptrackId; specifically: (1) The latitude lat and longitude lon of each track point in the fishing boat track data set Data are encoded by the geohash algorithm to obtain the hash code D of the longitude and latitude of the fishing boat hashcode ; (2) For D hashcode The TfidfVectorizer algorithm in the scikit-learn toolkit is used to calculate and convert it into a Tf-IDF feature matrix. Then, the TruncatedSVD algorithm is used to perform text topic analysis on the matrix to obtain the hash code topic vector of the voyage track corresponding to the shiptrackId. (3) For D hashcode Use Word2vec word vector embedding to represent the distributed feature vector and get the hash code feature vector of the voyage track corresponding to the shiptrackId The gradient coding feature extraction module is used to perform longitude and latitude gradient coding on the voyage track and extract the features of the longitude and latitude gradient coding to obtain the gradient coding theme vector of the voyage track corresponding to the shiptrackId. Specifically: (1) Calculate the longitude and latitude gradients for each repaired voyage track in the fishing vessel track dataset Data, and then merge the data to obtain the encoding Datagrad of the longitude and latitude gradients of the fishing vessel; (2) Datagrad is calculated using the TfidfVectorizer algorithm in the scikit-learn toolkit to convert it into a Tf-IDF feature matrix. The matrix is then analyzed using the TruncatedSVD algorithm to obtain the gradient encoded topic vector of the shiptrackId corresponding to the voyage track. The splicing layer is used to splice the outputs of the geohash encoding feature extraction module and the gradient encoding feature extraction module to obtain a feature vector extracted based on the longitude and latitude word embedding model; The feature extraction of speed and direction is performed based on the speed direction word embedding model; specifically: 2-2-1: Encode the speed speed and direction dir in the fishing boat trajectory dataset Data into a string to obtain the string-encoded trajectory D str ={c o (shiptrackId, speed_dir)|o=1,2,…,O}, speed_dir is the string concatenation of speed and direction, and o is the number of track points in the fishing vessel track dataset; 2-2-2: For D str Use Word2vec word vector embedding to represent the distributed feature vector. According to the fishing boat track sequence corresponding to shiptrackId, obtain the distributed feature vector corresponding to shiptrackId. The feature extraction of the fishing boat passing through the sea area based on the sea area grid word embedding model is specifically: 2-3-1: Convert the lat and lon in the fishing boat trajectory dataset Data into Mercator coordinates, then divide the sea area where the fishing boat passes into grids in fine granularity and number the grids in sequence; 2-3-2: According to the sea area grids that a single fishing boat passes through, each track point in each fishing boat voyage track is converted into a grid number sequence to obtain the sub-sea area grid sequence of each fishing boat The grid number represents the track of the β-th fishing vessel voyage; 2-3-3: Data of the sea area grid sequence dataset serialize Input the Word2vec word vector model for training, and finally get the grid coded feature vector of a single fishing boat based on shiptrackId S3, classification model training stage: Firstly, the trajectory feature vectors obtained in the S2 semantic feature extraction stage are divided into training set and test set. Then, the training set is input into the LightGBM classifier for training, and the test set is used to test the LightGBM classifier. Finally, the trained LightGBM classifier is used to realize VMS fishing vessel trajectory recognition of fishing vessel operation type.
2. A method for identifying fishing vessel types based on semantic feature vectors as claimed in claim 1, Features The splicing layer is used to splice the feature vectors extracted by the word embedding model based on longitude and latitude, the word embedding model based on speed direction, and the word embedding model based on sea area grid to form a series merged vector, and obtain the feature vector extracted by the word embedding semantic feature extraction model; 3. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 2.
4. A computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in any one of claims 1 to 2 is implemented.
Citation Information
Patent Citations
Trawler behavior discrimination method based on multi-step clustering
CN110033051A
Fishing ship fishing mode prediction method based on Stacking algorithm
CN112733935A
Short-time fishing effort distribution prediction method based on fishing sequential relationship
CN113360544A
Performing semantic analyses of user-generated textual and voice content
US20180293978A1
