An intelligent design method for ocean observation stations based on convolutional neural networks

Abstract

The present invention discloses a method for intelligent design of ocean observation stations based on convolutional neural networks, which belongs to the field of intelligent station design technology, including obtaining and preprocessing multi-source ocean data, making a deep learning sample set, constructing a training convolutional neural network model, and using the trained convolutional neural network to perform ocean observation station reasoning. The method for intelligent design of ocean observation stations of the present invention fills the gap in the use of artificial intelligence methods to design ocean observation stations, and solves the problems of unstable artificial station design effects, difficult station design in complex sea areas, and easy fall into local optimum of traditional optimization algorithms in existing station design methods.

Description

Technical Field

[0001] The invention discloses an intelligent design method for ocean observation stations based on convolutional neural networks, belonging to the technical field of intelligent station design. Background Art

[0002] The traditional observation station design method has the following problems. First, the artificial design of stations is generally based on experience and statistical principles. The results are unstable and may not be the optimal solution. There is still room for further optimization. Secondly, for complex marine environments, such as the presence of complex irregular sub-mesoscale phenomena and the lack of an obvious dominant direction in the gradient of environmental field changes, it is very difficult to empirically analyze environmental characteristics and station design is very difficult. In addition, the method of searching for the optimal station design through traditional optimization algorithms such as genetic algorithms and greedy algorithms cannot solve the contradiction between limited computing power and global optimality in principle and practice. It is easy to fall into local optimal solutions, resulting in some unsatisfactory results. Summary of the invention

[0003] The purpose of the present invention is to provide a method for intelligent design of ocean observation stations based on convolutional neural networks to solve the problem that the results of intelligent design of observation stations are not ideal in the prior art.

[0004] An intelligent design method for ocean observation stations based on convolutional neural networks includes acquiring multi-source ocean data and making a deep learning sample set, building a training convolutional neural network model, and using the trained convolutional neural network to perform ocean observation station inference.

[0005] The convolutional neural network model includes five serial convolutional layers in sequence. The first convolutional layer has a size of 7×7, a channel of 64, and a stride of 2. The second convolutional layer includes a maximum pooling layer and two blocks. The maximum pooling layer has a size of 3×3 and a stride of 2. Each block includes two convolutional layers with a size of 3×3, a channel of 64, and a stride of 1. Residual connections are established between blocks.

[0006] The third convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and 128 channels. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks.

[0007] The fourth convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and channel 256. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks.

[0008] The fifth convolutional layer consists of two blocks, each block includes two convolutional layers of size 3×3 and channels 512. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks.

[0009] The fifth convolutional layer is followed by an average pooling layer and a fully connected layer with an output size of 1×160, which outputs the coordinates of the observation station.

[0010] The input of the residual connection is The ideal output is , After weighted operation, activation function and weighted operation in sequence, a residual map is formed. , and Perform feature fusion to form , input the activation function and output the result.

[0011] Training a convolutional neural network model involves:

[0012] S1 inputs multi-source ocean data into the convolutional neural network to obtain observation station information;

[0013] S2 uses the observation station information to perform virtual sampling in the marine environment field, and the sampling results are used as observation values;

[0014] S3 uses data assimilation algorithm to calculate the inverted environmental field through observation values ​​and station information;

[0015] S4 calculates the deviation between the inverted environment field and the original environment field, and uses the deviation as a loss function;

[0016] S5 uses the gradient descent method to find the partial derivatives of the loss function with respect to the neural network parameters, and updates the neural network parameters through gradient back propagation optimization;

[0017] S6 repeats steps S1 to S5 until the neural network training is completed.

[0018] Ocean observation station location reasoning involves using a trained convolutional neural network model, inputting multi-source ocean data, and outputting the intelligent design process of ocean observation stations.

[0019] The data assimilation algorithm is an algorithm for inferring the real environmental field through sampling of observation data, including stepwise correction method, Kriging interpolation method, optimal interpolation method, and variational assimilation method.

[0020] The deviation between the inverted environmental field and the original environmental field describes the difference between the two field distributions, including the root mean square error and the cross entropy.

[0021] The elements of multi-source ocean data include temperature data, salinity data and density data, and the types of multi-source ocean data include observational data, reanalysis data and numerical simulation data.

[0022] The deep learning sample set includes training set, validation set and test set.

[0023] Compared with the prior art, the present invention has the following beneficial effects: the intelligent design method for ocean observation stations of the present invention constructs and applies a deep neural network model based on convolutional neural networks, which fills the gap in the use of artificial intelligence methods to design ocean observation stations, and solves the problems in the existing station design methods, such as unstable manual station design effects, difficulty in station design in complex sea areas, and traditional optimization algorithms easily falling into local optimality. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a technical flow chart of the present invention;

[0025] Figure 2 It is a schematic diagram of the neural network model structure;

[0026] Figure 3 It is a schematic diagram of the structure of convolution module 1;

[0027] Figure 4 It is a schematic diagram of the structure of convolution module 2, convolution module 3, convolution module 4, and convolution module 5;

[0028] Figure 5 It is a schematic diagram of the residual connection structure;

[0029] Figure 6 This is a sample diagram of the sea surface temperature field;

[0030] Figure 7 This is a graph showing the change in loss over epoch in the embodiment;

[0031] Figure 8 is an example diagram of observation stations designed by the neural network in the embodiment;

[0032] Fig. 9 is the isotherm of the original sea surface temperature field in the first experiment;

[0033] Fig.10 is the isotherm of the virtual sampling inversion result obtained using uniform design in the first experiment;

[0034] Fig.11 is the isotherm of the virtual sampling inversion result obtained by genetic algorithm design in the first experiment;

[0035] Fig.12 is the isotherm of the virtual sampling inversion result obtained by the neural network design in the first experiment;

[0036] Fig.13 is the isotherm of the original sea surface temperature field in the second experiment;

[0037] Fig.14is the isotherm of the virtual sampling inversion result obtained in the second experiment using uniform design;

[0038] Fig.15 is the isotherm of the virtual sampling inversion result obtained by genetic algorithm design in the second experiment;

[0039] Fig.16 is the isotherm of the virtual sampling inversion result obtained by the neural network design in the second experiment;

[0040] Fig.17 It is a statistical comparison of the station design effects in the embodiments. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but 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.

[0042] An intelligent design method for ocean observation stations based on convolutional neural networks includes acquiring multi-source ocean data and making a deep learning sample set, building a training convolutional neural network model, and using the trained convolutional neural network to perform ocean observation station inference.

[0043] The convolutional neural network model includes five serial convolutional layers in sequence. The first convolutional layer has a size of 7×7, a channel of 64, and a stride of 2. The second convolutional layer includes a maximum pooling layer and two blocks. The maximum pooling layer has a size of 3×3 and a stride of 2. Each block includes two convolutional layers with a size of 3×3, a channel of 64, and a stride of 1. Residual connections are established between blocks.

[0044] The third convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and 128 channels. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks.

[0045] The fourth convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and channel 256. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks.

[0046] The fifth convolutional layer consists of two blocks, each block includes two convolutional layers of size 3×3 and channels 512. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks.

[0047] The fifth convolutional layer is followed by an average pooling layer and a fully connected layer with an output size of 1×160, which outputs the coordinates of the observation station.

[0048] The input of the residual connection is The ideal output is , After weighted operation, activation function and weighted operation in sequence, a residual map is formed. , and Perform feature fusion to form , input the activation function and output the result.

[0049] Training a convolutional neural network model involves:

[0050] S1 inputs multi-source ocean data into the convolutional neural network to obtain observation station information;

[0051] S2 uses the observation station information to perform virtual sampling in the marine environment field, and the sampling results are used as observation values;

[0052] S3 uses data assimilation algorithm to calculate the inverted environmental field through observation values ​​and station information;

[0053] S4 calculates the deviation between the inverted environment field and the original environment field, and uses the deviation as a loss function;

[0054] S5 uses the gradient descent method to find the partial derivatives of the loss function with respect to the neural network parameters, and updates the neural network parameters through gradient back propagation optimization;

[0055] S6 repeats steps S1 to S5 until the neural network training is completed.

[0056] Ocean observation station location reasoning involves using a trained convolutional neural network model, inputting multi-source ocean data, and outputting the intelligent design process of ocean observation stations.

[0057] The data assimilation algorithm is an algorithm for inferring the real environmental field through sampling of observation data, including stepwise correction method, Kriging interpolation method, optimal interpolation method, and variational assimilation method.

[0058] The deviation between the inverted environmental field and the original environmental field describes the difference between the two field distributions, including the root mean square error and the cross entropy.

[0059] The elements of multi-source ocean data include temperature data, salinity data and density data, and the types of multi-source ocean data include observational data, reanalysis data and numerical simulation data.

[0060] The deep learning sample set includes training set, validation set and test set.

[0061] The technical flow chart of the present invention is as follows: Figure 1 As shown, it includes making a deep learning sample set, building a convolutional neural network, training the convolutional neural network, and performing observatory inference.

[0062] The convolutional neural network model of the present invention is as follows Figure 2 As shown, Figure 2 The input field in it represents the input domain, avgpool represents average pooling, and the convolutional neural network model includes 5 serial convolutional layers in sequence. The first convolutional layer (convolutional module 1) has a size of 7×7, 64 channels, and a stride of 2. The second convolutional layer (convolutional module 2) includes a maximum pooling layer and two blocks. The maximum pooling layer has a size of 3×3 and a stride of 2. Each block includes two convolutional layers with a size of 3×3, 64 channels, and a stride of 1. Residual connections are established between blocks. The third convolutional layer (convolutional module 3) includes two blocks, each of which includes two convolutional layers with a size of 3×3 and a channel of 128. The convolutional layer of each block has a stride of 2 and a stride of 1, and a residual connection is established between blocks; the fourth convolutional layer (convolutional module 4) includes two blocks, each block includes two convolutional layers of size 3×3 and channels 256, the first convolutional layer in each block has a stride of 2 and a stride of 1, and a residual connection is established between blocks; the fifth convolutional layer (convolutional module 5) includes two blocks, each block includes two convolutional layers of size 3×3 and channels 512, the first convolutional layer in each block has a stride of 2 and a stride of 1, and a residual connection is established between blocks. After the fifth convolutional layer, the output layer is composed of an average pooling layer and a fully connected layer with an output size of 1×160, which outputs the coordinates of the observation station.

[0063] Convolution module 1 consists of a convolution layer and a pooling layer. In the embodiment, the convolution kernel size of the convolution layer is 7×7, the channel is 64, and the step size is 2; the pooling layer is a maximum pooling layer, and its kernel size is 3×3 and the step size is 2. The input of convolution module 1 is , the output is , as shown in formula (1):

[0064] (1);

[0065] In the formula, represents the operation function of convolution module 1, The number of rows, columns and channels are h=224, w=224, c=1 respectively; The shape of is h=56, w=56, c=64.

[0066] The structures of convolution module 2, convolution module 3, convolution module 4 and convolution module 5 are as follows: Figure 4 As shown, the module consists of four convolutional layers.

[0067] In convolutional layer module 2, the convolution kernel size of convolutional layer A, convolutional layer B, convolutional layer C, and convolutional layer D is 3×3, the channel is 64, and the stride is 1. The input of convolutional module 2 is , the output is , as shown in formula (2):

[0068] (2);

[0069] In the formula, Represents the operation function of convolution module 2, The shape of is h=56, w=56, c=64.

[0070] In the convolutional layer module 3, the convolution kernel size of convolutional layer A is 3×3, the channel is 128, and the stride is 2. The convolution kernel size of convolutional layer B, convolutional layer C, and convolutional layer D is 3×3, the channel is 128, and the stride is 1. The input of convolutional module 3 is , the output is , as shown in formula (3):

[0071] (3);

[0072] In the formula, represents the operation function of convolution module 3, The shape of is h=28, w=28, c=128.

[0073] In convolutional layer module 4, the convolution kernel size of convolutional layer A is 3×3, the channel is 256, and the stride is 2. The convolution kernel size of convolutional layer B, convolutional layer C, and convolutional layer D is 3×3, the channel is 256, and the stride is 1. The input of convolutional module 4 is , the output is , as shown in formula (4):

[0074] (4);

[0075] In the formula, represents the operation function of convolution module 4, The shape is h=14, w=14, c=256.

[0076] In convolutional layer module 5, the convolution kernel size of convolutional layer A is 3×3, the channel is 512, and the stride is 2. The convolution kernel size of convolutional layer B, convolutional layer C, and convolutional layer D is 3×3, the channel is 512, and the stride is 1. The input of convolutional module 5 is , the output is , as shown in formula (5):

[0077] (5);

[0078] In the formula, represents the operation function of convolution module 5, The shape of is h=7, w=7, c=512.

[0079] In this embodiment, the pooling layer uses an average pooling layer and a fully connected layer with an output dimension of 160 to output the observation station coordinates, as shown in formula (6):

[0080] (6);

[0081] In the formula, represents the pooling operation function, represents the fully connected operation function, is a vector of length 160.

[0082] The residual connection is used as a pluggable module in convolution module 2, convolution module 3, convolution module 4 and convolution module 5. Figure 5 , the residual connection expresses the output as a linear superposition of the input and a nonlinear transformation of the input. The residual connection in the embodiment is shown in formula (7):

[0083] (7);

[0084] In the formula, represents the convolution operation function, is a nonlinear activation function, represents the input of the residual connection, Represents the output of the residual connection.

[0085] The structure of convolution module 1 is as follows Figure 3 shown.

[0086] The residual connection of the present invention is as follows Figure 5 As shown, the input is The ideal output is , After weighted operation, activation function and weighted operation in sequence, a residual map is formed. , and Perform feature fusion to form , input the activation function and output the result.

[0087] The present invention uses HYCOM sea surface temperature field data from 2005 to 2015 to produce a sample set with a total number of 100,000; samples are uniformly selected in 10 years with a step length of 1 month, thereby achieving sample time diversity; samples are randomly selected in mid- and low-latitude sea areas and a certain sea area at a ratio of 4:1, thereby achieving regional diversity of samples and strengthening the characteristic proportion of key areas; three different sizes of sea areas of 100km×100km, 200km×200km, and 300km×300km are selected at a ratio of 1:3:1, and then scaled to the same size, thereby achieving diversity in sea area scale; all sample sets are stratified and sampled at a ratio of 7:2:1 as training set, validation set, and test set.

[0088] The training consists of the following steps:

[0089] T1 inputs the ocean environment field into the convolutional neural network to obtain the observation station information. Specifically, in the embodiment, the two-dimensional ocean environment field is resampled into a 224×224 matrix, and the instance normalization is input into the convolutional neural network model constructed in step T2, and the model outputs the coordinates of 80 observation stations.

[0090] T2 uses the observation station information to perform virtual sampling on the environmental field, and the sampling result is used as the observation value. is the horizontal coordinate of the observation station, is the vertical coordinate of the observation station, in the original environmental field Search on Corresponding attribute value ,Will As the observations of virtual sampling.

[0091] T3 uses a data assimilation algorithm to calculate the inverted environmental field through observation values ​​and station information, wherein the data assimilation algorithm refers to an algorithm for inferring the real environmental field through sampling of observation data, including stepwise correction method, Kriging interpolation method, optimal interpolation method, and variational assimilation method. Specifically, in the embodiment, according to the station coordinates and observation values ​​of the virtual observation , using the ordinary Kriging method to invert the reconstructed environmental field .

[0092] T4 calculates the deviation between the inverted environment field and the original environment field, and uses the deviation as a loss function, wherein the deviation between the inverted environment field and the original environment field describes the difference between the two field distributions, including root mean square error and cross entropy. Specifically, in the embodiment, the root mean square error is calculated as the loss function, and the root mean square error The formula is as follows:

[0093] ;

[0094] In the formula, is the observed value at each pixel position of the original environment field, is the observed value at each pixel position of the reconstructed environment field, is the number of observations.

[0095] T5 uses the gradient descent method to find the partial derivative of the loss function with respect to the neural network parameters, and updates the neural network parameters through gradient back propagation optimization. Specifically, in the embodiment, the optimizer is Adam, the learning rate is 0.01, and the number of iterations is 256.

[0096] T6, repeat steps T1 to T5 until the neural network training is completed. Specifically, in the embodiment, the epoch is set to 400, and the experiment observes that the training error and the generalization error tend to converge after about 350 epochs. Therefore, it is considered that the neural network training is completed when the training is 400 epochs, and the epoch expression is a traversal of the data set.

[0097] In the embodiment, a convolutional neural network model trained in step T3 is used, and the input is an ocean environment field resampled to a size of 224×224. The model outputs the coordinates of 80 observation stations, and the running time is less than 1 minute. The accuracy of the reconstructed environment field is higher than the results of manual station design and traditional optimization algorithm design, thereby realizing rapid, batch, and optimized design of observation stations.

[0098] In the embodiment, the training samples are randomly sampled from the HYCOM sea surface temperature field data from 2005 to 2015, with a total number of 100000. The samples are evenly selected in 10 years with a step length of 1 month.

[0099] The sample of sea surface temperature field is a two-dimensional matrix, whose two dimensions are longitude and latitude, and the value represents the sea surface temperature. Figure 6 shown.

[0100] The neural network training is performed as described above. The loss on the training set and the validation set changes with epoch as follows: Figure 7 The experiment observed that both the training error and the generalization error tended to converge after about 350 epochs, so it was considered that the neural network training was completed after 400 epochs.

[0101] The sea surface temperature field is input into the neural network, and the coordinates of the observation stations output by the neural network are obtained. The stations designed by the neural network are drawn on the temperature field. Figure 8 shown.

[0102] In order to test the effect of the observation stations designed by the neural network, the designed stations were used to perform virtual sampling on the original sea surface temperature field, and the sea surface temperature field was reconstructed from the virtual observations using the Kriging interpolation method and compared with the original sea surface temperature field. As a comparison, the stations designed by the convolutional neural network were compared with the uniformly designed stations and the genetic algorithm designed stations. Among them, the uniformly designed stations refer to making the observation stations as evenly distributed as possible in a grid, and the genetic algorithm designed stations refer to using the conventional genetic algorithm of the python geatpy library for station design. The comparison of the neural network station design results with the original sea surface temperature field and the station design results of other methods, as shown in the first experiment. Fig. 9 , Fig.10 , Fig.11 and Fig.12 As shown in the figure, it can be seen that the virtual sampling inversion results of the neural network station design reflect the shape and direction of the isotherms well. Fig.13 , Fig.14 , Fig.15 and Fig.16 As shown in the figure, it can be seen that the virtual sampling inversion results of the neural network station design reflect the high and low temperature areas and shapes of the sea area well. The statistical distribution comparison of the root mean square error of the neural network station design results and the station design results of other methods is shown in the figure. Fig.17 As shown, the average root mean square errors of the virtual sampling inversion results of uniform design, genetic algorithm design, and neural network design and the original sea surface temperature field are 0.30, 0.22, and 0.18, respectively. It can be seen that the effect of the neural network station design in the embodiment is significantly better than that of the uniform design and the genetic algorithm design.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent design of ocean observation stations based on convolutional neural networks, characterized in that: This includes acquiring multi-source ocean data and creating deep learning sample sets, building and training convolutional neural network models, and using trained convolutional neural networks to perform ocean observation station location inference; The convolutional neural network model includes five serial convolutional layers in sequence. The first convolutional layer has a size of 7×7, a channel of 64, and a stride of 2. The second convolutional layer includes a maximum pooling layer and two blocks. The maximum pooling layer has a size of 3×3 and a stride of 2. Each block includes two convolutional layers with a size of 3×3, a channel of 64, and a stride of 1. Residual connections are established between blocks. The third convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and 128 channels. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks. The fourth convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and channel 256. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks. The fifth convolutional layer consists of two blocks, each of which consists of two convolutional layers of size 3×3 and 512 channels. The stride of the first convolutional layer in each block is 2, and the stride of the convolutional layer is 1. Residual connections are established between blocks. The fifth convolutional layer is followed by an average pooling layer and a fully connected layer with an output size of 1×160, which outputs the coordinates of the observation station. Training a convolutional neural network model involves: S1 inputs multi-source ocean data into the convolutional neural network to obtain observation station information; S2 uses the observation station information to perform virtual sampling in the marine environment field, and the sampling results are used as observation values; S3 uses data assimilation algorithm to calculate the inverted environmental field through observation values ​​and station information; S4 calculates the deviation between the inverted environment field and the original environment field, and uses the deviation as a loss function; S5 uses the gradient descent method to find the partial derivatives of the loss function with respect to the neural network parameters, and updates the neural network parameters through gradient back propagation optimization; S6 repeats steps S1 to S5 until the neural network training is completed; The input of the residual connection is The ideal output is , After weighted operation, activation function and weighted operation in sequence, a residual map is formed. , and Perform feature fusion to form , input the activation function and output the result; The data assimilation algorithm is an algorithm for inferring the real environmental field through sampling of observation data, including stepwise correction method, Kriging interpolation method, optimal interpolation method, and variational assimilation method.

2. According to claim 1, a method for intelligent design of ocean observation stations based on convolutional neural networks is characterized in that: Ocean observation station location reasoning involves using a trained convolutional neural network model, inputting multi-source ocean data, and outputting the intelligent design process of ocean observation stations.

3. The method for intelligent design of ocean observation stations based on convolutional neural networks according to claim 2 is characterized in that: The deviation between the inverted environmental field and the original environmental field describes the difference between the two field distributions, including the root mean square error and the cross entropy.

4. The method for intelligent design of ocean observation stations based on convolutional neural networks according to claim 3 is characterized in that: The elements of multi-source ocean data include temperature data, salinity data and density data, and the types of multi-source ocean data include observational data, reanalysis data and numerical simulation data.

5. The method for intelligent design of ocean observation stations based on convolutional neural networks according to claim 4 is characterized in that: The deep learning sample set includes training set, validation set and test set.

Images