A method for identifying fishing vessel operation modes based on one-dimensional CNN and improved GRU network structure

Abstract

A method for identifying fishing vessel operation patterns based on a one-dimensional CNN and an improved GRU network structure is proposed. This method addresses the current challenges of optimizing convolutional neural networks (CNNs) for fishing vessel operation pattern identification tasks, as well as the problems of feature loss and long-term dependencies when directly using recurrent neural networks. The invention proposes a fishing vessel operation pattern identification model (1DCNN-RIGRU) based on a one-dimensional convolutional neural network (CNN) and an improved gated recurrent unit (GRU). This model fully extracts the local spatial features and temporal dependencies of fishing vessel trajectory data using a one-dimensional CNN and an improved GRU. Dropout and the RAdam optimizer are introduced to further optimize the model, ultimately achieving the identification of fishing vessel operation patterns, standardizing fishing behavior, and ensuring the sustainable and healthy development of the marine fishing industry.

Description

Technical Field

[0001] This invention relates to the field of image recognition, specifically to a method for recognizing fishing vessel operation methods based on a one-dimensional CNN and an improved GRU network structure. Background Technology

[0002] The marine fishing industry still faces numerous problems and an increasingly severe development situation due to unreasonable and insufficient supervision of fishing vessels. Illegal, unreported, and unregulated fishing (IUU) is a persistent and frequent occurrence. IUU not only damages fish resources and causes a rapid decline in fish biodiversity, but also severely damages the ecological environment on which fish and other marine life depend. This results in catches characterized by low nutritional levels, small size, and young age, thus affecting fishing vessel yields and profitability, dampening fishermen's enthusiasm for production, and hindering the sustainable development of marine fishery resources. Therefore, comprehensively controlling the fishing methods employed by fishing vessels, regulating their fishing behavior, and arranging rational and orderly fishing activities have become important research topics for ensuring the sustainable and healthy development of the marine fishing industry.

[0003] Fishing vessel operation mode identification can determine the operation mode adopted by a fishing vessel based on the spatiotemporal location data continuously transmitted by the onboard terminal installed on the fishing vessel. This helps to regulate fishing behavior, control the intensity of fishing operations, and perceive the situation and dynamic changes of the fishing grounds, providing strong support for fishing vessel supervision and the protection of fishery resources. In recent years, with the rapid development of the field of artificial intelligence, scholars at home and abroad have been conducting research on the problem of fishing vessel operation mode identification using traditional machine learning and deep learning methods. Shan Xiaoying et al. of Jiaxing University used a support vector machine (SVM) model optimized with the sparrow search algorithm (SSA) to identify fishing vessel fishing modes and achieved a high recognition rate. Yang Shenglong et al. of the Chinese Academy of Fishery Sciences constructed a three-layer bidirectional long short-term memory network (BiLSTM) to identify four different types of fishing vessels. Because this model considers and constructs 17 feature vectors, including the difference in flight phase and speed with surrounding fishing vessels, the recognition accuracy can reach over 91%. The above methods can all identify fishing vessels of various different operation types and have good effectiveness and scalability. However, traditional convolutional neural networks are more suitable for the field of image recognition. The application of convolutional neural networks in the task of identifying fishing vessel operation methods needs to be further optimized. At the same time, directly using recurrent neural networks will also have problems such as feature loss and long-term dependence. Therefore, further research and improvement are needed. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a method for identifying fishing vessel operation modes based on a one-dimensional CNN and an improved GRU network structure. This method mainly addresses the need for further optimization of convolutional neural networks in current fishing vessel operation mode identification tasks, as well as the problems of feature loss and long-term dependence when directly using recurrent neural networks.

[0005] The technical solution adopted in this invention is:

[0006] A method for identifying fishing vessel operation patterns based on a one-dimensional CNN and an improved GRU network structure includes the following steps:

[0007] Step 1: Data preprocessing. The dataset consists of spatiotemporal trajectory points reported every 10 to 20 minutes by the Beidou shipborne terminal on Shandong-registered fishing vessels. This data includes the fishing vessel's Beidou ID, latitude and longitude, speed, time, and heading dynamic data. The parallelization features of Spark RDD are used to process the fishing vessel trajectory data stored in the database. Finally, the fishing vessel trajectory data obtained after data cleaning is used as the dataset. The trajectories of tracing, purse seine, and gillnetting operations each account for about one-third, and are divided into training set, validation set, and test set.

[0008] Step 2: Divide each trajectory into 64 segments. Statistically select 10 features for each fishing boat trajectory segment: mean longitude (lon_mean) and 75th percentile (lon_3 / 4), mean latitude (lon_mean) and 25th percentile (lon_1 / 4), latitude-longitude covariance matrix (xy_cov), mean speed due to time and space difference (skd_vmean), mean speed (v_mean), standard deviation of speed (v_std) and 75th percentile (v_3 / 4), and mean direction (d_mean). These are used as inputs to the model.

[0009] Step 3: Construct a 1DCNN-RTGRU model; use fully connected layers to perform linear transformations on the trajectory feature sequences during input and output, and add non-linear factors by using the LeakyReLU activation function before and after the one-dimensional convolutional layer; the model first receives the trajectory feature sequence as input, mines the spatial features of the data in the one-dimensional convolutional layer, and then inputs it into the recurrent layer to extract the temporal patterns between the data and improve the internal structure of the recurrent neural network. Finally, the retained information is input into the fully connected layer for linear transformation to obtain the final classification result;

[0010] Step 4: Continuously train and adjust the model parameters using the training and validation sets;

[0011] Step 5: Output the optimal model and identify the fishing vessel operation mode.

[0012] In step three, the 1DCNN-RIGRU model removes the reset gate from the regular GRU network model. Revised to:

[0013]

[0014] Replace the tanh function with the ReLU activation function and apply batch normalization; the parameters of the GRU model change as follows:

[0015] z i =σ(BN(W) (z) x i )+U (z) h i-1 )

[0016]

[0017]

[0018] Batch normalized BN(·) is defined as:

[0019]

[0020] In the above formula, μ k and σ k These are the mean and variance of the mini-batch, respectively; ε is a very small constant that serves to ensure numerical stability; and variables γ and λ are trainable parameters.

[0021] The convolutional layers of a 1DCNN slide in a fixed direction during operation.

[0022] With the convolution kernel set to 3 and the stride set to 1, the expression is as follows. In the formula, y i+1 x represents the output of the convolutional layer. i w represents the feature quantity before convolution. i+1 and b i+1 These represent the weight matrix and bias vector of the convolution kernel, respectively, and f represents the activation function.

[0023] The beneficial effects of this invention are as follows: This invention proposes a fishing vessel operation mode recognition model (1DCNN-RIGRU) based on a one-dimensional convolutional neural network (CNN) and an improved gated recurrent unit (GRU). This model fully extracts the local spatial features and temporal dependencies of fishing vessel trajectory data using a one-dimensional CNN and an improved GRU. Furthermore, it introduces the Dropout method and the RAdam optimizer to optimize the model, ultimately achieving the recognition of fishing vessel operation modes, standardizing fishing behavior, and ensuring the sustainable and healthy development of the marine fishing industry. Attached Figure Description

[0024] Figure 1 This is an example diagram of one-dimensional convolution in this invention.

[0025] Figure 2 This is a structural diagram of the improved GRU of the present invention.

[0026] Figure 3 This is a block diagram of the fishing vessel operation mode identification algorithm of the present invention. Detailed Implementation

[0027] The invention will be further described below with reference to the accompanying drawings: As shown in the figures, a method for identifying fishing vessel operation modes based on a one-dimensional CNN and an improved GRU network structure includes the following steps.

[0028] Step 1: Data preprocessing. The dataset consists of spatiotemporal trajectory points reported every 10 to 20 minutes by the Beidou shipborne terminal on Shandong-registered fishing vessels. This data includes the fishing vessel's Beidou ID, latitude and longitude, speed, time, and heading dynamic data. The parallelization features of Spark RDD are used to process the fishing vessel trajectory data stored in the database. Finally, the fishing vessel trajectory data obtained after data cleaning is used as the dataset. The trajectories of tracing, purse seine, and gillnetting operations each account for about one-third, and are divided into training set, validation set, and test set.

[0029] Step 2: Divide each trajectory into 64 segments. Statistically select 10 features for each fishing boat trajectory segment: mean longitude (lon_mean) and 75th percentile (lon_3 / 4), mean latitude (lon_mean) and 25th percentile (lon_1 / 4), latitude-longitude covariance matrix (xy_cov), mean speed due to time and space difference (skd_vmean), mean speed (v_mean), standard deviation of speed (v_std) and 75th percentile (v_3 / 4), and mean direction (d_mean). These are used as inputs to the model.

[0030] Step 3: Construct a 1DCNN-RIGRU model; Fully connected layers are used to perform linear transformations on the trajectory feature sequences during input and output, while LeakyReLU activation functions are used before and after the one-dimensional convolutional layers to add non-linear factors; The model first receives the trajectory feature sequences as input, mines the spatial features of the data in the one-dimensional convolutional layers, and then inputs them into the recurrent layers to extract the temporal patterns between the data, improving the internal structure of the recurrent neural network. Finally, the retained information is input into the fully connected layers for linear transformation to obtain the final classification result; 1DCNN refers to a one-dimensional convolutional neural network design.

[0031] Step 4: Continuously train and adjust the model parameters using the training and validation sets;

[0032] Step 5: Output the optimal model and identify the fishing vessel operation mode.

[0033] In step three, the 1DCNN-RIGRU model removes the reset gate from the regular GRU network model. Revised to:

[0034]

[0035] Replace the tanh function with the ReLU activation function and apply batch normalization; the parameters of the GRU model change as follows:

[0036] z i =σ(BN(W) (z) x i )+U (z) h i-1 )

[0037]

[0038]

[0039] Batch normalized BN(·) is defined as:

[0040]

[0041] In the above formula, μ k and σ k These are the mean and variance of the mini-batch, respectively; ε is a very small constant that serves to ensure numerical stability; and variables γ and λ are trainable parameters.

[0042] The convolutional layers of a 1DCNN slide in a fixed direction during operation.

[0043] With the convolution kernel set to 3 and the stride set to 1, the expression is as follows. In the formula, y i+1 x represents the output of the convolutional layer. iw represents the feature quantity before convolution. i+1 and b i+1 These represent the weight matrix and bias vector of the convolution kernel, respectively, and f represents the activation function.

[0044] The technical solution adopted by this invention to solve its technical problem is as follows: a fishing vessel operation mode recognition model (1DCNN-RIGRU) based on a one-dimensional convolutional neural network (CNN) and an improved gated recurrent unit (GRU) network is proposed. This model fully extracts the local spatial features and temporal dependencies of fishing vessel trajectory data using a one-dimensional CNN and an improved GRU. The Dropout method and the RAdam optimizer are introduced to optimize the model, ultimately achieving the recognition of fishing vessel operation modes, standardizing fishing behavior, and ensuring the sustainable and healthy development of the marine fishing industry.

[0045] (1) Dataset preparation

[0046] The dataset used for anomaly detection of fishing vessel trajectories consists of spatiotemporal trajectory points reported approximately every 10 to 20 minutes by Beidou shipborne terminals mounted on Shandong-registered fishing vessels. This data includes dynamic data such as the fishing vessel's Beidou ID, latitude and longitude, speed, time, and heading. However, the dynamic data does not contain static information such as the fishing vessel's operation type. Therefore, it is necessary to match the dynamic data with the static data stored in the database during the fishing vessel's registration based on the Beidou ID to form complete trajectory data. Ultimately, trajectory data generated during fishing outside the closed season was selected and preprocessed to obtain 18,000 fishing vessel trajectory data points as the dataset.

[0047] (2) Design of one-dimensional convolutional neural network

[0048] Due to the rapid development of computer vision, multidimensional convolutional neural networks (CNNs) have the widest range of applications, being widely used in image classification, face recognition, and other fields. However, fishing boat trajectories are time-series data. Although multidimensional CNNs are currently used for time-series data recognition tasks, this method compresses the dimensionality of the feature maps, leading to a significant loss of feature information from the original input data, ultimately affecting the recognition performance. Therefore, this application improves upon traditional CNNs by designing a one-dimensional convolutional neural network (1DCNN) to extract local features from the spatiotemporal location data of fishing boats, using only the convolutional layers. This method effectively avoids feature information loss, ensuring that the information input to the recurrent neural network in subsequent research is maximized, thereby fully mining data information and improving the recognition performance of fishing boat trajectory operation methods. 1DCNN and multidimensional CNNs share similar characteristics and processing methods; the key difference lies in the dimension of the input data and how the convolutional kernel slides between the input information. In 1DCNN, the convolutional layers slide in a fixed direction, extracting features from a sliding window through convolution operations, such as... Figure 1 As shown, the convolution kernel is set to 3 and the stride is set to 1; its expression is shown in formula (1), where y i+1 x represents the output of the convolutional layer. i w represents the feature quantity before convolution. i+1 and b i+1 These represent the weight matrix and bias vector of the convolution kernel, respectively, and f represents the activation function.

[0049]

[0050] (3) RIGRU network architecture design

[0051] Fishing vessel trajectory data is multidimensional and noisy, making direct application of recurrent neural networks (RNNs) ineffective for identifying fishing vessel operation methods. Therefore, this application adopts a GRU structure. However, when processing data with a GRU, the activation of two gates may exhibit redundancy. Considering that the current information needs to be given greater emphasis, r can be... i and z i Set it to a very small value. Similarly, if you want to place more emphasis on historical information, you can set r to a very small value. i and z i Set it to a large value. If you want to solve redundancy and improve computational efficiency, removing the reset gate is an effective method, because this will allow the GRU to be implemented with a more compact single-gate model.

[0052] Remove the reset door, Revised to:

[0053]

[0054] In GRU, tanh activation is crucial because its saturation slows down training and leads to the vanishing gradient problem. ReLU activation can be considered to effectively improve this limitation and accelerate network training. The update gate structure uses ReLU as the activation function, which... Revised to:

[0055]

[0056] The ReLU activation function is not commonly used in traditional RNN models because unbounded ReLU activation leads to numerical instability. An effective solution is to combine the ReLU activation function with batch normalization. The improved GRU structure is as follows: Figure 2 As shown.

[0057] The reset gate is removed from the structure, the tanh function is replaced with the ReLU activation function, and batch normalization is applied. The parameters of the GRU model change as follows:

[0058] z i =σ(BN(W) (z) x i )+U (z) h i-1 )

[0059]

[0060]

[0061] Batch normalized BN(·) is defined as:

[0062]

[0063] In the above formula, μ k and σ k These are the mean and variance of the mini-batch, respectively. ε is a small constant that ensures numerical stability, while variables γ and λ are trainable parameters. We call the improved GRU structure RIGRU.

[0064] To verify the effectiveness of the proposed fishing vessel operation behavior recognition algorithm, under the same experimental conditions, a partitioned test set not used in training was employed. Comparative experiments were conducted using the traditional machine learning model LightGBM, the deep learning model LSTM, and the ConvLSTM model (a parallel combination of CNN and LSTM). This part of the experiment added precision P, recall R, and f1 score F as evaluation criteria to the overall accuracy A, thereby revealing the performance of each algorithm for recognizing each type of fishing vessel operation behavior. The experimental results are shown in the table below.

[0065]

[0066]

[0067] As shown in the table, when using various experimental models to identify fishing vessel operation methods, trawls and gillnets outperform purse seines in all evaluation metrics. This is mainly because fishing vessels have more flexible and varied movement patterns during purse seine operations, making their movement patterns less significant and thus affecting the model's recognition performance for this type of operation. In the comparative experiment, the algorithm designed in this application improved the accuracy of fishing vessel operation method identification by 2% compared to the LightGBM model, indicating that the deep learning algorithm model proposed in this application can extract data information more fully for classification tasks. Compared to using only a CNN model, it improved the accuracy by 3%, demonstrating the advantage of using recurrent neural networks to process time-series data and playing a crucial role in the task of identifying fishing vessel operation methods. Compared to using LSTM and ConvLSTM models, it improved the accuracy by 1.3% and 0.7% respectively. This is because the algorithm proposed in this application adopts a cascaded fusion of one-dimensional convolutional neural networks and gated recurrent unit networks, maximizing the retention of useful information while fully mining the spatiotemporal patterns of fishing vessel trajectory data. Based on the above experiments and analysis, it is shown that the 1DCNN-RIGRU model proposed in this application has a more effective recognition effect in the task of identifying fishing vessel operation methods.

[0068] This specific implementation uses the spatiotemporal trajectory data reported every 10 to 20 minutes by the Beidou shipborne terminal on a Shandong-registered fishing vessel as an example. To address the problems of inapplicability and low accuracy when using convolutional neural networks and recurrent neural networks alone to identify fishing vessel operation behavior using deep learning methods, a fishing vessel operation mode identification algorithm based on the 1DCNN-SAGRU model is proposed, such as... Figure 3 As shown,

[0069] (1) Data preprocessing. The parallelization features of Spark RDD were used to process the fishing vessel trajectory data stored in the database. Finally, 18,000 fishing vessel trajectory data obtained after data cleaning were used as the dataset, of which trawl, purse seine, and gillnet trajectories each accounted for about one-third.

[0070] (2) Input Feature Selection. Each fishing boat trajectory obtained after data preprocessing contains a large amount of trajectory point data. Directly using the original parameters as input will cause information loss in the loop layer. In this chapter, each trajectory is divided into 64 segments, and 10 features of each fishing boat trajectory segment are selected statistically: longitude mean (lon_mean) and 75th percentile (lon_3 / 4), latitude mean (lon_mean) and 25th percentile (lon_1 / 4), longitude and latitude covariance matrix (xy_cov), spatiotemporal difference mean velocity (skd_vmean), velocity mean (v_mean), velocity standard deviation (v_std) and 75th percentile (v_3 / 4), and direction mean (d_mean) as input to the model.

[0071] (3) Standardization. To address the differences in the dimensions and magnitudes of various features, ensuring that different features have the same scale, this chapter uses standard score standardization (Z-Score), calculated as follows:

[0072]

[0073] In the formula, u is the sample mean, and σ represents the standard deviation of the sample data. Standardization can improve the convergence speed and output accuracy of the network.

[0074] (4) Divide all feature sequences into a training set (80%), a validation set (10%), and a test set (10%). The training set is used to train the model and determine the parameters, the validation set is used to determine the network structure and adjust the hyperparameters of the model, and the test set is reserved for the final evaluation of the model.

[0075] (5) Construct a 1D CNN-RIGRU model. The model structure is as follows: Figure 3 As shown, fully connected layers are used to perform linear transformations on the trajectory feature sequences during input and output. Simultaneously, LeakyReLU activation functions are used before and after the one-dimensional convolutional layers to add non-linear factors, improving the model's robustness and convergence speed. The model first receives the trajectory feature sequence as input, mines spatial features in the one-dimensional convolutional layer, and then inputs it into a recurrent layer to extract temporal patterns between data points, improving the internal structure of the recurrent neural network. Finally, the retained information is input into a fully connected layer for linear transformation to obtain the final classification result.

[0076] (6) Introducing the Dropout mechanism to optimize the model. When the training error of the prediction model is small, but the test error is large and much larger than the training error, the prediction result of this model is called overfitting. By adding the Dropout method, a portion of the nodes can be randomly deactivated with a certain probability during training, weakening the connection between nodes, improving the generalization ability of the model, and preventing overfitting.

[0077] (7) Select the RAdam optimizer to operate on the model during the training process. RAdam (Rectified Adam) is a variant of the Adam algorithm that can automatically and dynamically adjust the adaptive learning rate, effectively preventing the Adam optimizer from getting stuck in local optima in the early stages of training.

[0078] (8) Employ the cross-entropy loss function. Cross-entropy represents the difference between the true probability distribution and the predicted probability distribution, and is commonly used in classification tasks. Its calculation method is as follows:

[0079]

[0080] Where y is the true probability that sample x belongs to a certain class, f(x) is the predicted score of sample belonging to a certain class, s represents the softmax function, and L represents the loss result that measures the difference.

[0081] (9) Use the training set and validation set to continuously train and adjust the model parameters.

[0082] (10) Output the optimal model and identify the fishing vessel operation mode.

[0083] The embodiments described with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention. The embodiments should not be considered as limiting the invention, but any improvements made based on the spirit of the invention should be within the scope of protection of the invention.

Claims

1. A method for identifying fishing vessel operation modes based on a one-dimensional CNN and an improved GRU network structure, characterized in that: Includes the following steps, Step 1: Data preprocessing. The dataset consists of spatiotemporal trajectory points reported every 10 to 20 minutes by the Beidou shipborne terminal on Shandong-registered fishing vessels. This data includes the fishing vessel's Beidou ID, latitude and longitude, speed, time, and heading dynamic data. The parallelization features of SparkRDD are used to process the fishing vessel trajectory data stored in the database. Finally, the cleaned fishing vessel trajectory data is used as the dataset, in which trawl, purse seine, and gillnet trajectories each account for about one-third, and are divided into training set, validation set, and test set. Step 2: Divide each trajectory into 64 segments. Statistically select 10 features for each fishing boat trajectory segment: mean longitude (lon_mean) and 75th percentile (lon_3 / 4), mean latitude (lon_mean) and 25th percentile (lon_1 / 4), latitude-longitude covariance matrix (xy_cov), mean speed due to time and space difference (skd_vmean), mean speed (v_mean), standard deviation of speed (v_std) and 75th percentile (v_3 / 4), and mean direction (d_mean). These are used as inputs to the model. Step 3: Construct a 1DCNN-RIGRU model; use fully connected layers to perform linear transformations on the trajectory feature sequences during input and output, and add non-linear factors by using the LeakyReLU activation function before and after the one-dimensional convolutional layer; The model first receives the trajectory feature sequence as input, mines the spatial features of the data in a one-dimensional convolutional layer, and then inputs it into a recurrent layer to extract the temporal patterns between the data and improve the internal structure of the recurrent neural network. Finally, the retained information is input into a fully connected layer for linear transformation to obtain the final classification result. Step 4: Continuously train and adjust the model parameters using the training and validation sets; Step 5: Output the optimal model and identify the fishing vessel operation mode; In step three, the 1DCNN-RIGRU model removes the reset gate from the regular GRU network model. Revised to: Replace the tanh function with the ReLU activation function and apply batch normalization; the parameters of the GRU model change as follows: z i =σ(BN(W (z) x i )+U (z) h i-1 ) Batch normalized BN(·) is defined as: In the above formula, μ k and σ k These are the mean and variance of the mini-batch, respectively; ε is a very small constant that serves to ensure numerical stability; and variables γ and λ are trainable parameters.

2. The method for identifying fishing vessel operation modes based on a one-dimensional CNN and an improved GRU network structure according to claim 1, characterized in that: The convolutional layers of a 1DCNN slide in a fixed direction during operation.

3. The method for identifying fishing vessel operation modes based on a one-dimensional CNN and an improved GRU network structure according to claim 1, characterized in that: With the convolution kernel set to 3 and the stride set to 1, the expression is as follows. In the formula, y i+1 x represents the output of the convolutional layer. i w represents the feature quantity before convolution. i+1 and b i+1 These represent the weight matrix and bias vector of the convolution kernel, respectively, and f represents the activation function.

Images