Black tide path forecasting system based on multi-source data fusion

Abstract

The invention discloses a black tide path forecasting system based on multi-source data fusion, and belongs to the technical field of black tide path forecasting, and the system comprises a data acquisition subsystem which is used for acquiring multi-source data of a black tide path; the data processing subsystem is used for preprocessing the multi-source data of the black tide path to obtain a to-be-calculated data set; the model construction subsystem is used for constructing a path forecasting model based on the convolutional neural network with combinatorial attention and the bidirectional long-short term memory network; and the prediction subsystem is used for inputting the to-be-calculated data set into the path prediction model for prediction and generating a prediction result. The output of the system can directly support decision making in multiple fields of marine scientific research, commercial shipping, marine protection area management and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of Kuroshio path prediction, and in particular relates to a Kuroshio path prediction system based on multi-source data fusion. Background Art

[0002] The Kuroshio is a famous western boundary current, known for its high temperature, high salinity, narrow flow, strong flow velocity, and large flow. Every year, it transports a large amount of heat from the low latitudes of the North Pacific to the mid- and high-latitudes, which has an important impact on the climate and environment of neighboring countries and even the world. The Kuroshio will present multiple flow characteristics, mainly including the typical large bend path and the non-large bend path. The changes in the shape and position of these two typical paths have a significant impact on the climate, fisheries, navigation safety, and marine resources in the North Pacific region where the Kuroshio passes. Some previous studies have also pointed out that the changes in the Kuroshio path have an important impact on the precipitation of some cities along the eastern coast of my country and the Yangtze River Basin. Therefore, the prediction of the future Kuroshio path is of great significance.

[0003] With the development of ocean observation and assimilation technology, the available ocean element data are becoming increasingly rich, including field observation data, satellite altitude data and ocean reanalysis data. Therefore, data-driven methods have begun to be widely used in ocean analysis and prediction. However, there are relatively few studies on Kuroshio path prediction based on machine learning methods. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a Kuroshio path forecasting system based on multi-source data fusion to solve the above problems existing in the prior art.

[0005] To achieve the above object, the present invention provides a Kuroshio path forecasting system based on multi-source data fusion, comprising:

[0006] A data acquisition subsystem, used to obtain multi-source data of the Kuroshio path;

[0007] A data processing subsystem, connected to the data acquisition subsystem, for preprocessing the multi-source data of the Kuroshio path to obtain a data set to be calculated;

[0008] A model building subsystem, connected to the data processing subsystem, for building a path prediction model based on a convolutional neural network with combined attention and a bidirectional long short-term memory network;

[0009] The prediction subsystem is connected to the model building subsystem and is used to input the data set to be calculated into the path prediction model for prediction and generate a prediction result.

[0010] Preferably, the multi-source data of the Kuroshio path include: satellite data, buoy and ship observation data and model data;

[0011] The satellite data include sea surface temperature, sea surface height and sea surface salinity;

[0012] The buoy and ship observation data include deep temperature and salinity data collected by the Argo float system and ship observation data from ocean research institutions.

[0013] Preferably, the data processing subsystem includes:

[0014] A filtering module, used for smoothing the multi-source data of the Kuroshio path based on a Butterworth low-pass filter to obtain filtered data;

[0015] A normalization module, used for normalizing the filtered data to obtain normalized data, and then using a window segmentation technique to divide the normalized data to obtain a plurality of data segments;

[0016] The sorting module is used to arrange the data segments based on a time axis to generate the data set to be calculated.

[0017] Preferably, the model building subsystem includes:

[0018] A convolution construction module, used to construct a convolutional neural network with combined attention, and train the convolutional neural network with combined attention through a parameter output vector of a Transformer encoder to obtain a spatial feature extraction model;

[0019] A long short-term memory construction module is used to construct a bidirectional long short-term memory network, and train the bidirectional long short-term memory network through a sample set with a timestamp to obtain a temporal feature extraction model;

[0020] A connection module is used to connect the output of the spatial feature extraction model and the output of the temporal feature extraction model through a fully connected layer to generate the path prediction model.

[0021] Preferably, the convolution building block comprises:

[0022] A teacher model building unit, used for building a Transformer encoder, training the Transformer encoder, and obtaining a parameter output vector of the Transformer encoder;

[0023] A network improvement unit, used for constructing a convolutional neural network, taking the parameter output vector of the Transformer encoder as the starting input of the convolutional neural network, and importing the combined attention module into the convolutional neural network to obtain the convolutional neural network with combined attention;

[0024] A training module is used to train the convolutional neural network with combined attention through a sample set to obtain the spatial feature extraction model.

[0025] Preferably, the training module includes:

[0026] A training set construction unit for obtaining a historical dataset of the Kuroshio path;

[0027] A pre-training unit, used for pre-training the convolutional neural network with combined attention through a BERT pre-training model to obtain a pre-training model;

[0028] A training unit is used to iteratively train the pre-trained model based on the training data set until the pre-trained model reaches the relationship between the automatic learning input parameters and the Kuroshio path prediction, and then terminate the iteration to obtain the spatial feature extraction model.

[0029] Preferably, in the long-short time construction module, the calculation formula for extracting periodic signal features by the bidirectional long-short time memory network model is:

[0030]

[0031]

[0032]

[0033] in, is the output vector of the forward LSTM hidden layer at time t, (t=1, 2, ..., N), which is the input vector x at the current time t and the forward LSTM output vector at the previous moment jointly determined;

[0034] in, is the output vector of the reverse LSTM hidden layer at time t, (t=1, 2, ..., N), which is the input vector x at the current time t and the reverse LSTM output vector of the previous moment jointly determined;

[0035] h t is the output of the BiLSTM model, ω t is the weight matrix of the forward LSTM output; v t is the weight matrix of the reverse LSTM output; b t is the bias of the weight matrix.

[0036] Preferably, the prediction subsystem comprises:

[0037] A calculation module, used for inputting the data set to be calculated into the path prediction model for prediction and generating the prediction result;

[0038] The interpretation module is used to interpret the prediction result to obtain a path prediction curve.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] The present invention uses CNN and Bi-LSTM networks with combined attention mechanisms, which can effectively learn complex dependencies from spatiotemporal data. The combined attention mechanism enables the model to better focus on the most informative parts of the data, thereby improving learning efficiency and prediction accuracy. The output of the system can directly support decision-making in multiple fields such as marine scientific research, commercial shipping, and marine protected area management. By providing more accurate predictions of the Kuroshio flow path, it can help relevant departments and organizations optimize operational plans and emergency response strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0042] Figure 1 Schematic diagram of a Kuroshio path forecasting system based on multi-source data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] Embodiment 1

[0046] like Figure 1 As shown, in this embodiment, a Kuroshio path forecasting system based on multi-source data fusion is provided, comprising:

[0047] A data acquisition subsystem, used to obtain multi-source data of the Kuroshio path;

[0048] A data processing subsystem, connected to the data acquisition subsystem, for preprocessing the multi-source data of the Kuroshio path to obtain a data set to be calculated;

[0049] A model building subsystem, connected to the data processing subsystem, for building a path prediction model based on a convolutional neural network with combined attention and a bidirectional long short-term memory network;

[0050] The prediction subsystem is connected to the model building subsystem and is used to input the data set to be calculated into the path prediction model for prediction and generate a prediction result.

[0051] Further optimizing the scheme, the multi-source data of the Kuroshio path includes: satellite data, buoy and ship observation data and model data;

[0052] The satellite data include sea surface temperature, sea surface height and sea surface salinity;

[0053] The buoy and ship observation data include deep temperature and salinity data collected by the Argo float system and ship observation data from ocean research institutions.

[0054] Further optimizing the scheme, the data processing subsystem includes:

[0055] A filtering module, used for smoothing the multi-source data of the Kuroshio path based on a Butterworth low-pass filter to obtain filtered data;

[0056] The calculation formula of Butterworth low-pass filter is:

[0057]

[0058] Among them, D o is the cutoff frequency of the filter, n is the order of the filter. In practical applications, D can be modified according to different filtering requirements. o and n value.

[0059] A normalization module, used for normalizing the filtered data to obtain normalized data, and then using a window segmentation technique to divide the normalized data to obtain a plurality of data segments;

[0060] The filtered data is normalized to normalized data using a linear function. In order to improve the network convergence speed and avoid neuron oversaturation, the input data is processed using the maximum and minimum normalization method to map the original signal data to [0,1].

[0061] The calculation formula for normalization is:

[0062]

[0063] x is the input signal value, x max 、x min is the maximum and minimum value of all signals, x normis the normalized value.

[0064] The sorting module is used to arrange the data segments based on the time axis to generate the data set to be calculated. The normalized data is divided into several data segments using the window segmentation technology. Most behavior recognition methods divide the time series data into smaller time windows through the window segmentation technology, and then extract features and recognize them in each time window.

[0065] Further optimizing the scheme, the model building subsystem includes:

[0066] A convolution construction module, used to construct a convolutional neural network with combined attention, and train the convolutional neural network with combined attention through a parameter output vector of a Transformer encoder to obtain a spatial feature extraction model;

[0067] This embodiment constructs a one-dimensional convolutional neural network with downsampling properties, as shown in 3. Layered convolution is used instead of maximum pooling to aggregate feature information. This method can reduce information loss. Convolution kernels of different sizes are used in the network to learn the long-term and short-term features of the input signal. As the number of network layers increases, the convolution kernel gradually becomes smaller, while the number of convolution channels and the convolution stride gradually increase. After testing, it is found that the recognition accuracy of the model is improved.

[0068] Table 1

[0069]

[0070] First, a 1×1 convolution layer is used to obtain the mapping of the input feature on the time signal, that is, the feature map on the time signal. Then, the feature map on the time signal is processed by a sigmoid function to obtain the time weight vector. The attention module uses a convolution layer to encode the feature information between local time signal segments to prevent excessive focus on the signal segments. Then, the first calibration feature is multiplied with the time weight vector to obtain the first calibration feature. Based on the residual method, the first calibration feature and the preprocessed data are added to obtain the spatial attention unit feature, that is, the input feature of the channel attention unit is obtained.

[0071] The working process of the channel attention module is divided into three steps: compression, excitation, multiplication and addition. The first is the input feature, which is the feature processed by the spatial attention module. Then the input spatial attention unit feature is compressed, that is, global average pooling and global maximum pooling, to obtain the global average and global maximum of each channel, followed by an excitation operation, using two fully connected layers and a sigmoid activation function, and adding the feature channels that have undergone global average pooling and global maximum pooling after passing through a multi-layer perceptron to obtain the final channel attention feature. Finally, the input feature and the final channel attention feature are multiplied to obtain the second calibration feature. Based on the residual method, the second calibration feature and the spatial attention unit feature are added to generate the combined attention module optimization feature.

[0072] The specific implementation steps of the combined attention module are:

[0073] The spatial attention unit first passes a 1×1 convolution layer through which the features to be processed are mapped on the time signal, that is, the feature projection on the time signal; then the feature mapping on the time signal is processed through the Sigmoid activation function to obtain the time weight vector. In addition to reducing the dimension of the input data, the 1×1 convolution layer can also enable the network to extract more in-depth features.

[0074] The spatial attention unit simultaneously encodes the feature information between local time signal segments through a one-dimensional convolutional layer for the features to be processed, i.e., preprocessed data, in order to prevent the attention unit from over-focusing on some signal segments. Then, the feature information between local time signal segments is multiplied by the time weight vector to obtain the first calibration feature.

[0075] The spatial attention unit introduces the residual idea to add the first calibration feature to the feature to be processed to obtain the input feature of the channel attention unit, that is, the spatial attention unit feature. The purpose of this is to avoid repeated calibration and cause the loss of some features.

[0076] The channel attention unit first uses the average pooling layer and the maximum pooling layer to perform pooling operations on the input features to obtain feature vectors. Because a single pooling loses too much information, the average pooling layer and the maximum pooling layer are connected in parallel, so that the deep features extracted are more comprehensive and richer.

[0077] After the global average feature vector and the global maximum feature vector are respectively transmitted to the multi-layer perceptron through the fully connected layer and the Sigmoid activation function, the channel attention feature is obtained by adding them together;

[0078] Multiply the channel attention feature with the input feature to obtain the second calibration feature.

[0079] The channel attention unit introduces the residual idea to add the second calibration feature to the input feature of the channel attention unit to obtain the optimized feature. The purpose of this is to avoid repeated calibration and cause the loss of some features.

[0080] A long short-term memory construction module is used to construct a bidirectional long short-term memory network, and train the bidirectional long short-term memory network through a sample set with a timestamp to obtain a temporal feature extraction model;

[0081] A connection module is used to connect the output of the spatial feature extraction model and the output of the temporal feature extraction model through a fully connected layer to generate the path prediction model.

[0082] Further optimizing the scheme, the convolution building block includes:

[0083] A teacher model building unit, used for building a Transformer encoder, training the Transformer encoder, and obtaining a parameter output vector of the Transformer encoder;

[0084] A network improvement unit, used for constructing a convolutional neural network, taking the parameter output vector of the Transformer encoder as the starting input of the convolutional neural network, and importing the combined attention module into the convolutional neural network to obtain the convolutional neural network with combined attention;

[0085] A training module is used to train the convolutional neural network with combined attention through a sample set to obtain the spatial feature extraction model.

[0086] Further optimizing the scheme, the training module includes:

[0087] A training set construction unit for obtaining a historical dataset of the Kuroshio path;

[0088] A pre-training unit, used for pre-training the convolutional neural network with combined attention through a BERT pre-training model to obtain a pre-training model;

[0089] A training unit is used to iteratively train the pre-trained model based on the training data set until the pre-trained model reaches the relationship between the automatic learning input parameters and the Kuroshio path prediction, and then terminate the iteration to obtain the spatial feature extraction model.

[0090] Using the parameter output vector of the Transformer encoder as the starting input of the convolutional neural network with combined attention;

[0091] The convolutional neural network with combined attention is trained using the training set. During the training process, the model will learn the correlation between the Kuroshio path predictions and try to classify the data. The model parameters are updated through the back-propagation algorithm and optimizer (such as Adam, SGD, etc.) to minimize the loss function (such as cross entropy loss).

[0092] Use the validation set to evaluate the performance of the model. Indicators such as accuracy, precision, recall, F1 score, etc. can be used to evaluate the performance of the model on the classification task. Adjust the model based on the evaluation results, such as adjusting hyperparameters and model structure, to improve model performance.

[0093] Use the test set to verify the performance of the trained and evaluated model on new data. Ensure that the model has good generalization ability and can accurately classify unseen data. In practical applications, apply the trained model to the classification task of Kuroshio path prediction to achieve learning and classification of the relevance of Kuroshio path prediction. Monitor the performance of the model in practical applications, and continuously optimize and improve the model based on feedback to adapt to data changes and application needs.

[0094] Further optimizing the scheme, in the long-short time construction module, the calculation formula for extracting periodic signal features by the bidirectional long-short time memory network model is:

[0095]

[0096]

[0097]

[0098] in, is the output vector of the forward LSTM hidden layer at time t, (t=1, 2, ..., N), which is the input vector x at the current time t and the forward LSTM output vector at the previous moment jointly determined;

[0099] in, is the output vector of the reverse LSTM hidden layer at time t, (t=1, 2, ..., N), which is the input vector x at the current time t and the reverse LSTM output vector of the previous moment jointly determined;

[0100] h t is the output of the BiLSTM model, ω t is the weight matrix of the forward LSTM output; v t is the weight matrix of the reverse LSTM output; b t is the bias of the weight matrix.

[0101] The calculation formula of the SaveReLU activation function is:

[0102]

[0103] e is a constant with a value range of [0,0.5], which ensures that the derivative of SaveReLU is always non-zero, and then the gradient of the part of the input less than zero can be calculated during the back propagation process, avoiding the problem of neuron "death".

[0104] The model training adopts the cross entropy loss function, and the key hyperparameter settings are shown in Table 2:

[0105] Table 2

[0106]

[0107] Use the test data sample to evaluate the performance of the model. The specific method is to calculate the recognition accuracy, precision, recall rate, and F1 value of the model for various actions in the test data.

[0108] Further optimizing the scheme, the prediction subsystem includes:

[0109] A calculation module, used for inputting the data set to be calculated into the path prediction model for prediction and generating the prediction result;

[0110] The interpretation module is used to interpret the prediction result to obtain a path prediction curve.

[0111] Those skilled in the art will appreciate that embodiments of the present invention may be provided as systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0112] The present invention is described with reference to the flowcharts and / or block diagrams of computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0113] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A Kuroshio path forecasting system based on multi-source data fusion, characterized in that: include: A data acquisition subsystem, used to obtain multi-source data of the Kuroshio path; A data processing subsystem, connected to the data acquisition subsystem, for preprocessing the multi-source data of the Kuroshio path to obtain a data set to be calculated; A model building subsystem, connected to the data processing subsystem, for building a path prediction model based on a convolutional neural network with combined attention and a bidirectional long short-term memory network; The prediction subsystem is connected to the model building subsystem and is used to input the data set to be calculated into the path prediction model for prediction and generate a prediction result.

2. The Kuroshio path forecasting system based on multi-source data fusion according to claim 1 is characterized in that: The multi-source data of the Kuroshio path include: satellite data, buoy and ship observation data and model data; The satellite data include sea surface temperature, sea surface height and sea surface salinity; The buoy and ship observation data include deep temperature and salinity data collected by the Argo float system and ship observation data from ocean research institutions.

3. The Kuroshio path forecasting system based on multi-source data fusion according to claim 1 is characterized in that: The data processing subsystem comprises: A filtering module, used for smoothing the multi-source data of the Kuroshio path based on a Butterworth low-pass filter to obtain filtered data; A normalization module, used for normalizing the filtered data to obtain normalized data, and then using a window segmentation technique to divide the normalized data to obtain a plurality of data segments; The sorting module is used to arrange the data segments based on a time axis to generate the data set to be calculated.

4. The Kuroshio path forecasting system based on multi-source data fusion according to claim 1 is characterized in that: The model building subsystem includes: A convolution construction module, used to construct a convolutional neural network with combined attention, and train the convolutional neural network with combined attention through a parameter output vector of a Transformer encoder to obtain a spatial feature extraction model; A long short-term memory construction module is used to construct a bidirectional long short-term memory network, and train the bidirectional long short-term memory network through a sample set with a timestamp to obtain a temporal feature extraction model; A connection module is used to connect the output of the spatial feature extraction model and the output of the temporal feature extraction model through a fully connected layer to generate the path prediction model.

5. The Kuroshio path forecasting system based on multi-source data fusion according to claim 4 is characterized in that: The convolution building block includes: A teacher model building unit, used for building a Transformer encoder, training the Transformer encoder, and obtaining a parameter output vector of the Transformer encoder; A network improvement unit, used for constructing a convolutional neural network, taking the parameter output vector of the Transformer encoder as the starting input of the convolutional neural network, and importing the combined attention module into the convolutional neural network to obtain the convolutional neural network with combined attention; A training module is used to train the convolutional neural network with combined attention through a sample set to obtain the spatial feature extraction model.

6. The Kuroshio path forecasting system based on multi-source data fusion according to claim 5 is characterized in that: The training module includes: A training set construction unit for obtaining a historical dataset of the Kuroshio path; A pre-training unit, used for pre-training the convolutional neural network with combined attention through a BERT pre-training model to obtain a pre-training model; A training unit is used to iteratively train the pre-trained model based on the training data set until the pre-trained model reaches the relationship between the automatic learning input parameters and the Kuroshio path prediction, and then terminate the iteration to obtain the spatial feature extraction model.

7. The Kuroshio path prediction system based on multi-source data fusion according to claim 1 is characterized in that: In the long-short time construction module, the calculation formula for extracting periodic signal features by the bidirectional long-short time memory network model is: in, is the output vector of the forward LSTM hidden layer at time t, (t=1, 2, ..., N), which is the input vector x at the current time t and the forward LSTM output vector at the previous moment jointly determined; in, is the output vector of the reverse LSTM hidden layer at time t, (t=1, 2, ..., N), which is the input vector x at the current time t and the reverse LSTM output vector of the previous moment jointly determined; h t is the output of the BiLSTM model, ω t is the weight matrix of the forward LSTM output; v t is the weight matrix of the reverse LSTM output; b t is the bias of the weight matrix.

8. The Kuroshio path prediction system based on multi-source data fusion according to claim 1 is characterized in that: The prediction subsystem includes: A calculation module, used for inputting the data set to be calculated into the path prediction model for prediction and generating the prediction result; The interpretation module is used to interpret the prediction result to obtain a path prediction curve.

Images