A fishing behavior analysis method for fishing vessels based on AIS data

Abstract

The present invention discloses a method for analyzing fishing behavior of fishing vessels based on AIS data, including: performing abnormal processing on raw data to obtain a trajectory data sequence; constructing a data set based on the trajectory data sequence; and identifying the fishing state and operation type of the fishing vessel based on the data set. The present invention provides a method for identifying the fishing state and operation type of a fishing vessel, starting from using a deep learning algorithm to mine useful information in a data set, and constructing a complete set of experimental processes and standards including data processing, identification of the fishing state of a fishing vessel, and identification of the operation type of a fishing vessel, thereby providing data information guarantee for the sustainable development of the fishery.

Description

Technical Field

[0001] The present application belongs to the field of deep learning technology, and specifically relates to a method for analyzing fishing behavior of fishing vessels based on AIS data. Background Art

[0002] There are many problems in the fishery sector, such as excessive fishing intensity, illegal abuse of juvenile fish resources, poaching during the closed fishing season, and other behaviors that damage the fishery ecology. At the same time, fishery supervision is also a difficult task. For the sustainable development of fisheries, reasonable fishing and increased supervision have become urgent matters. Summary of the invention

[0003] The present invention proposes a fishing behavior analysis method for fishing vessels based on AIS data, aiming to identify the fishing status and operation type of fishing vessels through AIS fishing tag data, use deep learning algorithms to mine useful information in the data set, and construct a complete set of experimental processes and standards including data processing, fishing status identification and fishing operation type identification, so as to provide data information guarantee for the sustainable development of fisheries.

[0004] To achieve the above object, the present invention provides the following solution: a method for analyzing fishing behavior of fishing vessels based on AIS data, which is as follows:

[0005] S1, perform exception processing on the original data to obtain a trajectory data sequence;

[0006] S2. constructing a data set based on the trajectory data sequence;

[0007] S3, identifying the fishing status of the fishing vessel based on the data set;

[0008] S4. Identify the fishing vessel operation type based on the data set.

[0009] Preferably, the original data in S1 is data on fishing operations of different types of fishing vessels in a fixed sea area, and the fishing status is marked.

[0010] Preferably, the abnormal processing of the original data in S1 comprises the following steps:

[0011] S1.1, cleaning and noise reduction processing are performed on the raw data to obtain preprocessed data;

[0012] S1.2. Perform trajectory extraction and interpolation processing based on the preprocessed data to obtain the trajectory data sequence.

[0013] Preferably, the cleaning and noise reduction processing of the raw data in S1.1 to obtain the preprocessed data comprises the following steps:

[0014] S1.1.1. discarding data rows with invalid data in the original data;

[0015] S1.1.2. Delete the data row with null value in the original data;

[0016] S1.1.3. Delete the data rows with exactly the same data in the original data;

[0017] S1.1.4. Delete data values ​​in the original data that exceed the standard data range to obtain the preprocessed data.

[0018] Preferably, the step of performing the trajectory extraction and the interpolation processing based on the preprocessed data to obtain the trajectory data sequence in S1.2 includes the following steps:

[0019] S1.2.1, grouping the pre-processed data, and sorting each group of the pre-processed data in ascending order by timestamp to obtain historical trajectory data;

[0020] S1.2.2, cutting the historical trajectory data to obtain track segments;

[0021] S1.2.3, calculating the derivative features of the trajectory points in the historical trajectory data to obtain the time difference between the previous and next trajectory points;

[0022] S1.2.4. Cut the track segment according to the time difference to obtain a normal time difference;

[0023] S1.2.5. Delete the data of the trajectory points corresponding to the time difference less than the normal time difference, and perform interpolation to obtain the trajectory data sequence.

[0024] Preferably, constructing the data set in S2 comprises the following steps:

[0025] S2.1, set the upper and lower limits of the sequence length;

[0026] S2.2, if the sequence length of the trajectory data sequence is lower than the lower limit, it is directly discarded;

[0027] S2.3, if the sequence length of the trajectory data sequence exceeds the upper limit, iterative sampling processing is performed to obtain a sampling data sequence whose sequence length is the upper limit;

[0028] S2.4. The trajectory data sequence and the sampling data sequence whose sequence lengths are between the upper limit and the lower limit are normalized by using deviation normalization to obtain the data set.

[0029] Preferably, identifying the fishing status of the fishing vessel in S3 comprises the following steps:

[0030] S3.1, combine the LSTM network with the CRF network to obtain the LSTM-CRF network;

[0031] S3.2. According to the data set, a conventional sequence tagging detection network and an LSTM-CRF network are selected to perform a prediction experiment, and the results of the prediction experiment are analyzed to obtain the LSTM-CRF network as the network model with the highest accuracy;

[0032] S3.3. Identify the fishing status of the fishing vessel through the LSTM-CRF network.

[0033] Preferably, identifying the fishing vessel operation type in S4 comprises the following steps:

[0034] S4.1. constructing an image dataset according to the dataset;

[0035] S4.2. According to the image data set, a conventional image classification processing network is selected to perform an image classification experiment, and the image classification experiment results are compared to obtain the best network model with the highest accuracy;

[0036] S4.3, embedding the C-CNN structure into the optimal network model, keeping other network parameters unchanged, setting the sampling cycle ratio, and performing a comparative experiment based on the image data set to obtain the optimal sampling cycle ratio;

[0037] S4.4. Identify the fishing vessel operation type by embedding the optimal network model of the C-CNN structure when the optimal sampling cycle ratio is used.

[0038] Preferably, constructing the image dataset in S4.1 comprises the following steps:

[0039] S4.1.1, distinguishing the data in the data set to obtain distinguished data, generating an image of the distinguished data according to the fishing coefficient, and obtaining an original track image;

[0040] S4.1.2, rotating and mirroring the original trajectory image to obtain an expanded trajectory image;

[0041] S4.1.3. Construct the image dataset according to the original trajectory image and the expanded trajectory image.

[0042] The beneficial effects of the present invention are: constructing a data set and using a deep learning algorithm to mine useful information from the data set, constructing an LSTM-CRF network model for identifying the fishing status of fishing vessels, using an image processing method and adopting a C-CNN structure for identifying the operation type of fishing vessels, and constructing a complete set of experimental processes and standards including data processing, fishing status identification of fishing vessels and fishing operation type identification, thereby providing data information guarantee for the sustainable development of fisheries.

[0043] This method has broad promotion space and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 It is a schematic diagram of the process of the present invention;

[0046] Figure 2 This is a schematic diagram of randomly selected historical trajectory data in an embodiment of the present invention;

[0047] Figure 3 It is a schematic diagram of time interval statistics and frequency distribution in trawling operation in an embodiment of the present invention;

[0048] Figure 4 Schematic diagram of comparison of track segments before and after interpolation in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of an LSTM-CRF network in an embodiment of the present invention;

[0050] Figure 6 This is a visual comparison diagram of the BP neural network prediction experimental results and the original labels in an embodiment of the present invention;

[0051] Figure 7 This is a comparison chart of the experimental results of the BP neural network and the LSTM network in the embodiment of the present invention;

[0052] Figure 8 This is a flow chart of the LSTM-CRF network algorithm in an embodiment of the present invention;

[0053] Fig. 9 This is a comparison chart of the labeling results of the LSTM network and LSTM-CRF network prediction experiments in an embodiment of the present invention;

[0054] Fig.10 A visualization diagram of the trajectory of a trawler in an embodiment of the present invention;

[0055] Fig.11 A visualization diagram of the trajectory of a gillnet fishing boat in an embodiment of the present invention;

[0056] Fig.12 Schematic diagram of the original trajectory image and the expanded trajectory image in an embodiment of the present invention;

[0057] Fig.13 Schematic diagram of the VGG-16 network model in an embodiment of the present invention;

[0058] Fig.14 Schematic diagram of C-CNN structure embedded in ResNet structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown, the present invention provides a fishing behavior analysis method for fishing vessels based on AIS data. The raw data of the present invention comes from five different types of fishing vessels, namely trawl, gillnet, purse seine, longline and trolling fishing vessels. The specific implementation process is as follows:

[0062] 1. Perform exception processing on the original data to obtain the trajectory data sequence. The method is as follows:

[0063] 1.1. Fishing status is marked for the data of fishing operations of different types of fishing vessels in fixed sea areas to obtain the original data. The key information of the original data is shown in Table 1.

[0064] Table 1

[0065]

[0066] 1.2. Clean and reduce noise on the original data to obtain preprocessed data. The specific process and rules are as follows:

[0067] Rule 1: discard the data rows marked with -1 in the original data, greatly compress the data, and reduce the running time and running memory for subsequent data processing;

[0068] Rule 2: Delete the data rows with null values ​​in the mmsi, timestamp, lat, lon, speed, is_fishing, and source columns in the original data;

[0069] Rule 3: Delete the rows in the original data where the timestamp column and mmsi column are exactly the same. Due to the garbled packet loss and retransmission mechanism during data transmission, there may be multiple rows for one data point. Therefore, it is necessary to deduplicate such points and keep only one row to prevent the accuracy of the model from being affected in subsequent experiments.

[0070] Rule 4: Delete values ​​in the original data that are outside the standard data range.

[0071] 1.3. Perform trajectory extraction and interpolation processing on the obtained preprocessed data to obtain a trajectory data sequence, as follows:

[0072] like Figure 2 As shown, in the original data, the trajectory data of different fishing vessels are grouped by the vessel ID of the fishing vessel, and the records in each group are sorted in ascending order according to the timestamp to obtain the historical trajectory data.

[0073] The historical trajectory data is divided into different voyages according to the time factor, the time difference between the previous and next trajectory points is calculated, and the trajectory points with a time difference greater than 4000 seconds are used as cutting points to cut the historical trajectory data into different track segments, and the track segments with a length less than 50 are deleted.

[0074] Calculate the derived features of the trajectory points in the historical trajectory data. The time difference is essential for trajectory segmentation, so it is necessary to calculate the time difference ΔT between the previous and next trajectory points:

[0075] ΔT=T1-T0

[0076] Among them, T1 represents the timestamp of the next trajectory point, and T0 represents the timestamp of the previous trajectory point.

[0077] The actual ground distances at different longitudes and latitudes are different. In order to eliminate this tiny effect, it is necessary to calculate the actual sea surface distance difference d. The Haversin formula uses a sine function to achieve high-precision calculation of the earth's sea level distance:

[0078]

[0079] haver sin(θ)=sin 2 (θ / 2)=(1-cos(θ)) / 2

[0080]

[0081] The track segment is cut by the time difference ΔT. The cutting principle is that ΔT needs to satisfy the actual fishing process, that is, two complete fishing processes of a fishing boat must be separated by a certain period of time. Therefore, if the time interval is greater than a certain value, it can be considered as two different fishing records.

[0082] like Figure 3 As shown in the figure, the time intervals between consecutive track points in each fishing boat's track segment are mostly distributed in 600-800 seconds, and 600-800 seconds is set as the normal time difference. Short time interval data are deleted, and a threshold of 5 times the interval of 800 seconds is set. The time difference within 4000 seconds is also considered as the track point of the same fishing trip, and the missing data is filled by interpolation.

[0083] like Figure 4 As shown, the data with a time difference within 600-4000 seconds is considered to be a complete track segment, and the data in this interval is interpolated. The present invention uses the Lagrange interpolation method to interpolate the missing values ​​in the track, so that the time difference between the previous and next track points is between 600-800 seconds, and a track data sequence with uniform records is obtained.

[0084] 2. Construct a data set based on the obtained trajectory data sequence. Since the sequence lengths of the trajectory data sequences obtained above vary, the longest sequence length contains hundreds of trajectory points, and the shortest sequence length contains only 10 trajectory points. Therefore, it is necessary to perform reasonable screening before constructing the data set. The specific method is as follows:

[0085] 2.1. Set the upper limit of sequence length l max =100 and lower limit l min =50;

[0086] 2.2. If the sequence length of the trajectory data sequence is lower than the lower limit l min =50, discard directly;

[0087] 2.3. If the sequence length of the trajectory data sequence exceeds the upper limit l max =100, iterative sampling is used to obtain a sampled data sequence with an upper limit of sequence length: the existing sample length is n, and the upper limit of the sample length of the data set is l max , n>l max , then we can sample (nl max ) samples.

[0088] 2.4. For trajectory data sequences and sampling data sequences whose sequence lengths are between the upper limit and the lower limit, deviation standardization is used to perform normalization processing to obtain a data set. The normalized data are all between [0 and 1].

[0089] 3. Identify the fishing status of the fishing vessel based on the obtained data set. The specific method is as follows:

[0090] 3.1. Combine the LSTM network with the CRF network to obtain the LSTM-CRF network model:

[0091] The recurrent neural network (RNN) in the deep neural network slowly developed from the BP neural network is a model for processing time series data. The RNN network shows better performance than the BP neural network in solving sequence problems, but it is powerless against the gradient vanishing and gradient exploding problems that occur in long sequences. Therefore, the LSTM network (Long Short-Term Memory Model LSTM) was born. The LSTM network introduces a memory cell based on the RNN network to solve the long-term dependency problem of sequence data.

[0092] The LSTM network relies on the powerful nonlinear fitting ability and sequence data memory ability of the neural network, while the CRF network uses a conditional state transition probability matrix that is statistically normalized globally, and then selects a sequence label with the highest sample probability through subsequent calculations. The LSTM network and the CRF network can complement each other's shortcomings by combining them.

[0093] like Figure 5 As shown in the figure, the bidirectional LSTM network combined with the CRF network constitutes an LSTM-CRF network. The LSTM-CRF network can take into account all the forward and reverse information of the sequence data, calculate all possible label sequences, and output the optimal label sequence after being constrained by the CRF layer.

[0094] 3.2. According to the obtained data set, BP neural network, LSTM network and LSTM-CRF network are selected for prediction experiments. According to the results of the prediction experiments, the network model with the highest accuracy is LSTM-CRF network. The specific process is as follows:

[0095] (1) BP neural network prediction experiment, as follows:

[0096] The present invention selects sequence data with a sequence length of 80 to carry out BP neural network prediction experiments. The hidden layer node setting interval of the BP neural network is [15-25]. The final experimental result is a labeled sequence composed of 0 and 1.

[0097] The data with a sequence length of 80 are selected from the dataset, and the training set, validation set, and test set are set in a ratio of 8:1:1. Taking the trawler dataset with the largest data volume as an example, comparative experiments are conducted on different activation functions and different numbers of nodes. The optimal number of hidden layer nodes is 20, and the optimal activation function is the Leaky ReLU activation function.

[0098] like Figure 6 As shown, Figure 6Figure (a) shows the original label, and Figure (b) shows the BP neural network prediction experimental results, where the gray points are fishing and the black points are non-fishing. It can be seen that the BP neural network is generally effective, and even the points in the continuous linear non-fishing track segment are misclassified.

[0099] (2) LSTM network prediction experiment, as follows:

[0100] The training process of the LSTM network is to input the feature items in the extracted track sequence into the network, calculate the output label items, compare the differences between the label items extracted from the actual track sequence and the calculated label items, and update the weights through error back propagation and gradient descent.

[0101] The goal of the LSTM network is the same as that of the BP neural network, that is, to mark the fishing behavior of each trajectory point in the trajectory sequence. The data set constructed by the present invention can all directly participate in the training of the LSTM network model, using the pytorch integrated library, which provides a processing method for the LSTM network for variable-length sequences.

[0102] The input layer of the LSTM network contains 5 neurons, each of which corresponds to 5 features with high correlation coefficients with fishing behavior selected from each trajectory point. The number of neurons in the output layer is 2, corresponding to the results of sequence labeling. The LSTM network predicts that there are only two cases of single-point results in the experimental sequence, namely 0 and 1. The Leaky-RELU function is selected as the activation function, the learning rate is set to 0.05, the batch is set to 128, and the maximum number of training iterations is set to 100. The data sets of five different types of fishing boats are divided into training sets, validation sets and test sets in a ratio of 8:1:1. The training set participates in the training calculation of the model. The validation set participates in a validation calculation after each model training to observe the experimental effect. The test set is used for the final test and does not participate in the model training process. A comparative experiment between the unidirectional LSTM network and the bidirectional LSTM network is performed for each data set. It can be seen that the bidirectional LSTM network is better than the unidirectional LSTM network. Set the early end training mark and record the best validation set result while training. If the validation set results of the next five consecutive validation sets are worse than the optimal validation set results, the training is terminated early and the network with the optimal parameters is recorded.

[0103] The number of hidden layer nodes in the LSTM network is set to [3-13]. 10% of the trawler data is selected to replace the entire data set for parameter selection experiments. The relationship between the accuracy of the predicted value and the true value of the LSTM network prediction experiment and the number of hidden layer nodes is obtained. When the trawler data set is selected, the optimal number of hidden layer nodes corresponding to the LSTM network is 14.

[0104] like Figure 7 As shown, Figure 7 Figure (c) shows the prediction experiment labeling results using the BP neural network, and Figure (d) shows the prediction experiment labeling results using the best bidirectional LSTM network. In Figure (d), many errors in the box are corrected, and other errors emerge. It can be seen that the main error is at the critical value. Overall, the results predicted by the LSTM network are more accurate.

[0105] (3) LSTM-CRF network prediction experiment, as follows:

[0106] The addition of the CRF network is to add probability constraints to the tag sequence calculated by the LSTM network to determine which result has the highest probability of occurring, thereby making the result more accurate.

[0107] like Figure 8 As shown in the figure, the data features of the LSTM-CRF network model are constructed from the bottom up, and are input into the LSTM-CRF network model to calculate the output results, which are expanded in time from left to right. Figure 8 One of the trajectory data was selected for analysis. The bottom layer of the LSTM-CRF network model is the input of data features. The data set comes from the data set constructed by the present invention. Each trajectory point input has 5 features, including normalized speed, longitude, latitude, time difference between previous and next records, and actual distance at sea level, which are 5 features with great relevance to fishing. These 5 features are used as the input of the model, and the data enters the LSTM network. The output is calculated and judged on the neurons of this layer. When the single-layer or double-layer neural network calculation is completed, the final output is a two-dimensional vector, in which the value represents the probability of each category of the two-class classification. At the same time, this output will be used as the input of the CRF network. After the probability calculation of the CRF network, a label sequence with the highest probability is finally determined.

[0108] from Figure 8 There is a slight discrepancy between the original label shown in Figure 1 and the predicted label calculated only by the LSTM network, which is manifested as a classification error at the end of the track segment, where the last digit in the original label is marked as 0, while the predicted label calculated by the LSTM network is as follows: Figure 8 The last digit of the LSTM network label is 1, and the calculation using only the LSTM model ends here. However, on this basis, a layer of CRF network is added. After the CRF network calculation, the last digit of the label is finally successfully corrected, and the result is the same as the original label.

[0109] The number of hidden layer nodes still needs to be determined on the chain structure of the LSTM part of the LSTM-CRF network. The approximate range of hidden layer nodes is [3-13]. The small sample trial and error method is used to experiment to obtain the optimal number of hidden layer nodes. The dimensions of the input layer and output layer remain unchanged. 10% of the corresponding data sets of five types of fishing boats are selected for the experiment, and the optimal number of nodes is finally obtained. The full data set of five different types of fishing boats is divided into training set, validation set and test set in a ratio of 8:1:1. The same experimental parameters as the LSTM network are used to obtain the LSTM-CRF experimental results.

[0110] like Fig. 9 As shown, Fig. 9 Figure (e) shows the labeling results of the prediction experiment using the LSTM network, and Figure (f) shows the labeling results of the prediction experiment using the LSTM-CRF network. It can be seen that the points that were previously misidentified, such as the turning section, have been successfully corrected by the smoother parts after being constrained by the CRF layer, but there are still slight errors at the turning points.

[0111] In summary, the LSTM-CRF network model is superior to the other two models in sequence labeling, and the LSTM-CRF network is used in the recognition of the fishing status of fishing vessels.

[0112] 4. Identify the fishing vessel operation type based on the obtained data set. The specific method is as follows:

[0113] Fishing boat operation types can be subdivided into about 10 categories, but among them, trawl and gillnet fishing boats account for the majority of the total number of fishing boats. In the process of identifying the operation types of fishing boats, the present invention selects trawl and gillnet fishing boats and generates images of the fishing tracks of trawl and gillnet fishing boats. Fig.10 As shown in Figure 2, the changes in the speed and course of the trawler are relatively stable. Fig.11 As shown in the figure, the fishing trajectory of the gillnet fishing boat has the feature of repeating back and forth in a small range, which shows that there are differences in the trajectory of the two types of fishing boats. The embodiment of the present invention visualizes the trajectory of the two types of fishing boats, and intuitively sees the difference between the trajectories of the two types of fishing boats, laying the foundation for the possibility of using the trajectory image method to classify the operation type of fishing boats. The specific implementation process of the present invention to identify the operation type of fishing boats is as follows:

[0114] 4.1. Construct an image dataset based on the constructed dataset. The specific process is as follows:

[0115] The data of different fishing vessels are distinguished according to the fishing vessel ID, and the data of different voyages of the same fishing vessel are distinguished. When the fishing coefficient exceeds 0.5, it means that there is large-scale fishing behavior in the track. The track image is generated for all fishing vessels with fishing coefficients greater than or equal to 0.5. The track image is a track line diagram of the longitude and latitude of each voyage. The size of each image is unified as 200 pixels high × 200 pixels wide, abbreviated as 200×200.

[0116] like Fig.12 As shown, for each original trajectory image ( Fig.12 (g)), rotate 90° counterclockwise (corresponding to Fig.12 (h)), rotate 180° counterclockwise (corresponding to Fig.12 (i)), rotate 270° counterclockwise (corresponding to Fig.12 (j) in the middle) to form three extended trajectory images, and the original trajectory image is horizontally mirrored (corresponding to Fig.12 (k)) and vertical mirror processing (corresponding to Fig.12 After the two extended trajectory images are obtained, the errors that depend on the absolute position of the trajectory can be eliminated to a certain extent through these two operations. The original trajectory images and the extended trajectory images constitute the image datasets of two different types of fishing vessels. Finally, a total of 75,632 images of two types of fishing vessels are generated, including 46,131 images of trawlers and 29,501 images of gillnetters. From the distribution of data volume, there are more types of trawlers, but they are not too different from gillnetters, so binary classification research can be carried out on data distribution balance.

[0117] 4.2. According to the constructed image dataset, AlexNet network, VGG-16 network and ResNet-34 network were selected for image classification experiments. The image classification experimental results were compared to obtain the best network model with the highest accuracy. The specific process is as follows:

[0118] (1) AlexNet network image classification experiment, the specific process is as follows:

[0119] The AlexNet network uses multiple convolution, pooling, and downsampling operations to perform image classification. Based on the fact that the image data used in the embodiment of the present invention is a trajectory contour grayscale image, compression is performed on the number of channels to increase the calculation speed. The final AlexNet network model framework process and parameters used in the present invention are shown in Table 2.

[0120] Table 2

[0121]

[0122] The AlexNet network activation function uses the Relu activation function, the classification function uses the Softmax function, and the loss function uses the classic categorical cross entropy function (Categorical CrossEntropy). A deep convolutional neural network is built using the pytorch integrated development library, and the data set is divided into a training set, a validation set, and a test set according to an 8; 1: 1 structure. During the training, an early end flag is set. After each training is completed, the validation set is used to evaluate the results, and the optimal model parameters are updated at the same time. If the validation set error continues to rise, the training is stopped and the optimal model is used as the final trained model. The present invention uses an exponentially decayed learning rate for training, with 128 images as a batch for 100 rounds, and the initial learning rate is set to 0.1. After 20 rounds, the learning rate decays to 0.01, and after 40 rounds, the learning rate decays to 0.001. The experimental results of AlexNet network image classification show that the accuracy of the optimal test set of the Adam optimizer reaches about 85%, the number of training rounds is 64, and the total training time is 5190.4sec.

[0123] (2) VGG-16 network image classification experiment, the specific process is as follows:

[0124] like Fig.13 As shown in the figure, the VGG-16 network contains 5 convolutional layers and 3 fully connected layers. The first two convolutional layers have two convolution operations, and the third to fifth convolutional layers have three convolution operations. The convolutional layers use 3×3 convolution kernels, the convolution step size is 1, the padding is 1, and most of the maximum pooling layers use 2×2 pools and 2×2 padding. Fig.13 The black box in the middle represents the image convolution activation composite operation. It can be clearly seen that there are 13 black boxes representing 13 convolution operations, the dotted box represents the maximum pooling operation, the long dashed box represents the fully connected layer and the corresponding activation operation, and the thin dashed box represents the output layer.

[0125] The VGG-16 network image classification experiment uses the same training ratio size, maximum number of iterations, and exponentially decayed learning rate as the AlexNet network image classification experiment, and uses the pytorch integrated development library to build the VGG-16 network. The results of the VGG-16 network image classification experiment show that the Adam optimizer's optimal test set accuracy reached about 90%, the number of training rounds was 62, and the total training time was 4219.8 seconds. It can be seen that the VGG-16 network has a higher accuracy in image classification experiments than the AlexNet network, and the total training time and number of training rounds have both decreased.

[0126] (3) ResNet-34 network image classification experiment, the specific process is as follows:

[0127] The configuration of the ResNet-34 network structure is shown in Table 3.

[0128] Table 3

[0129]

[0130] The pytorch integrated development library was used to build the ResNet-34 network, and the same training ratio size, maximum number of iterations, and exponentially decayed learning rate as the AlexNet network image classification experiment were used. The results of the ResNet-34 network image classification experiment showed that the Adam optimizer's optimal test set accuracy reached about 92%, the number of training rounds was 69, and the total training time was 6639.7 seconds. It can be seen that the ResNet-34 network has a higher accuracy in image classification experiments than the VGG-16 network, but the total training time and number of training rounds have increased.

[0131] In summary, among the AlexNet network, VGG-16 network and ResNet-34 network, the Adam optimizer of the ResNet-34 network has the highest accuracy rate on the optimal test set.

[0132] 4.3, such as Fig.14 As shown in the figure, the C-CNN structure is embedded in the ResNet-34 network, that is, the convolution module in the ResNet-34 network structure is replaced by the C-CNN structure, other network parameters remain unchanged, the sampling cycle ratio is set to 1 / 4, 1 / 9 and 1 / 16, and comparative experiments are performed using image datasets.

[0133] The comparative experimental results show that the Adam optimizer has the highest optimal test set accuracy of 93.8% when the sampling cycle ratio is 1 / 9, so the optimal sampling cycle ratio is 1 / 9.

[0134] The C-CNN structure with a sampling cycle ratio of 1 / 9 is embedded in the ResNet-34 network to identify the operation types of fishing vessels.

[0135] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A fishing behavior analysis method for fishing vessels based on AIS data, characterized in that: The following steps are involved: S1, perform exception processing on the original data to obtain a trajectory data sequence; S2. constructing a data set based on the trajectory data sequence; S3, identifying the fishing status of the fishing vessel based on the data set; S4. Identifying the type of fishing vessel operation based on the data set; Constructing the data set in S2 includes the following steps: S2.1, set the upper and lower limits of the sequence length; S2.2, if the sequence length of the trajectory data sequence is lower than the lower limit, it is directly discarded; S2.3, if the sequence length of the trajectory data sequence exceeds the upper limit, iterative sampling processing is performed to obtain a sampling data sequence whose sequence length is the upper limit; S2.4, performing normalization processing on the trajectory data sequence and the sampling data sequence whose sequence length is between the upper limit and the lower limit by using deviation normalization to obtain the data set; The step of identifying the fishing vessel operation type in S4 includes the following steps: S4.

1. constructing an image dataset according to the dataset; S4.

2. According to the image data set, a conventional image classification processing network is selected to perform an image classification experiment, and the image classification experiment results are compared to obtain the best network model with the highest accuracy; S4.3, embedding the C-CNN structure into the optimal network model, keeping other network parameters unchanged, setting the sampling cycle ratio, and performing a comparative experiment based on the image data set to obtain the optimal sampling cycle ratio; S4.4, identifying the fishing vessel operation type by embedding the optimal network model of the C-CNN structure when the optimal sampling cycle ratio is used; Constructing the image dataset in S4.1 includes the following steps: S4.1.1, distinguishing the data in the data set to obtain distinguished data, generating an image of the distinguished data according to the fishing coefficient, and obtaining an original track image; S4.1.2, rotating and mirroring the original trajectory image to obtain an expanded trajectory image; S4.1.

3. Construct the image dataset according to the original trajectory image and the expanded trajectory image.

2. The method for analyzing fishing behavior of fishing vessels based on AIS data according to claim 1, characterized in that: The original data in S1 is data on fishing operations of different types of fishing boats in a fixed sea area, and the fishing status is marked.

3. The method for analyzing fishing behavior of fishing vessels based on AIS data according to claim 1, characterized in that: The abnormal processing of the original data in S1 includes the following steps: S1.1, cleaning and noise reduction processing are performed on the raw data to obtain preprocessed data; S1.

2. Perform trajectory extraction and interpolation processing based on the preprocessed data to obtain the trajectory data sequence.

4. The method for analyzing fishing behavior of fishing vessels based on AIS data according to claim 3, characterized in that: The cleaning and noise reduction processing of the raw data in S1.1 to obtain the pre-processed data includes the following steps: S1.1.

1. discarding data rows with invalid data in the original data; S1.1.

2. Delete the data row with null value in the original data; S1.1.

3. Delete the data rows with exactly the same data in the original data; S1.1.

4. Delete data values ​​in the original data that exceed the standard data range to obtain the preprocessed data.

5. The method for analyzing fishing behavior of fishing vessels based on AIS data according to claim 3, characterized in that: The step S1.2 of performing the trajectory extraction and the interpolation processing based on the preprocessed data to obtain the trajectory data sequence includes the following steps: S1.2.1, grouping the pre-processed data, and sorting each group of the pre-processed data in ascending order by timestamp to obtain historical trajectory data; S1.2.2, cutting the historical trajectory data to obtain track segments; S1.2.3, calculating the derivative features of the trajectory points in the historical trajectory data to obtain the time difference between the previous and next trajectory points; S1.2.

4. Cut the track segment according to the time difference to obtain a normal time difference; S1.2.

5. Delete the data of the trajectory points corresponding to the time difference less than the normal time difference, and perform interpolation to obtain the trajectory data sequence.

6. The method for analyzing fishing behavior of fishing vessels based on AIS data according to claim 1, characterized in that: The step of identifying the fishing state of the fishing vessel in S3 includes the following steps: S3.1, combine the LSTM network with the CRF network to obtain the LSTM-CRF network; S3.

2. According to the data set, a conventional sequence tagging detection network and an LSTM-CRF network are selected to perform a prediction experiment, and the results of the prediction experiment are analyzed to obtain the LSTM-CRF network as the network model with the highest accuracy; S3.

3. Identify the fishing status of the fishing vessel through the LSTM-CRF network.

Images