CN116861799A - Submarine topography inversion model generation and submarine topography inversion method based on residual errors - Google Patents
Submarine topography inversion model generation and submarine topography inversion method based on residual errors Download PDFInfo
- Publication number
- CN116861799A CN116861799A CN202311126318.XA CN202311126318A CN116861799A CN 116861799 A CN116861799 A CN 116861799A CN 202311126318 A CN202311126318 A CN 202311126318A CN 116861799 A CN116861799 A CN 116861799A
- Authority
- CN
- China
- Prior art keywords
- model
- residual
- layer
- inversion
- representing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000012876 topography Methods 0.000 title claims abstract description 63
- 230000005484 gravity Effects 0.000 claims abstract description 24
- 238000003062 neural network model Methods 0.000 claims abstract description 12
- 238000005457 optimization Methods 0.000 claims abstract description 10
- 238000011478 gradient descent method Methods 0.000 claims abstract description 5
- 210000002569 neuron Anatomy 0.000 claims description 50
- 238000012549 training Methods 0.000 claims description 25
- 230000006870 function Effects 0.000 claims description 17
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 238000011156 evaluation Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 8
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 7
- 238000000137 annealing Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 abstract description 14
- 238000012545 processing Methods 0.000 abstract description 14
- 238000004519 manufacturing process Methods 0.000 abstract 1
- 230000008569 process Effects 0.000 description 12
- 238000013461 design Methods 0.000 description 10
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000004880 explosion Methods 0.000 description 5
- 238000009825 accumulation Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 238000010276 construction Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000013178 mathematical model Methods 0.000 description 3
- 238000012706 support-vector machine Methods 0.000 description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000013535 sea water Substances 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention relates to a submarine topography inversion model generation and submarine topography inversion method based on residual errors, and belongs to the technical field of submarine topography inversion; the production method comprises the following steps: model determination: designing a basic residual error module contained in a residual error depth neural network model; setting a model input: taking the shortwave gravity anomaly component and the longitude and latitude of the control point as model input; model optimization: performing model optimization by adopting a first momentum gradient descent method; model implementation: model implementation was performed using a Keras framework. The inversion method should invert the model generated by the residual-based submarine topography inversion model generation method. The accuracy is improved, the calculation complexity is reduced, the automation processing is carried out, and the potential information of the gravitational field data is fully mined.
Description
Technical Field
The invention relates to a submarine topography inversion model generation and submarine topography inversion method based on residual errors, and belongs to the technical field of submarine topography inversion.
Background
In marine geology and geophysical research, inversion of the topography of the sea floor is a critical task that has profound effects on understanding the structure of the earth's crust, the internal structure of the earth, the distribution of ocean currents, and the marine ecosystem. At present, in the technology of carrying out submarine topography inversion by utilizing gravitational field data, a traditional mathematical model and an iterative algorithm are widely used. However, these methods have certain errors in the processing, which are mainly caused by error accumulation and data noise problems in the calculation process, and these problems result in poor accuracy of the inversion result. Meanwhile, due to high calculation complexity and long time consumption, the efficiency and instantaneity of submarine topography inversion are limited. Furthermore, the methods require specialized personnel to process and interpret data, and have the characteristics of high operation threshold and lack of automation and intelligence. Most importantly, these traditional methods often fail to fully exploit and exploit the complex nonlinear relationships in the data when processing gravitational field data, thereby limiting their application in complex scenarios.
Specifically, the conventional submarine topography inversion method has the following problems when the submarine topography inversion is performed by using gravitational field data:
1. the precision is not high: the traditional method has the problems of error accumulation and data noise in the process of processing and analyzing gravitational field data, which directly leads to low precision of submarine topography inversion results.
2. The calculation complexity is high: conventional inversion methods typically employ complex mathematical models and iterative algorithms for data processing and inversion, which makes the computation complex and time consuming.
3. The manual intervention is as follows: conventional inversion methods often require specialized personnel for data processing and interpretation, which requires significant human resources and is somewhat lacking in automation and intelligence.
In addition, the prior art has the following disadvantages in processing gravitational field data:
1. neglecting the nonlinear relationship: traditional methods tend to ignore complex nonlinear relationships in the data when processing gravitational field data, which makes it impossible to fully mine and exploit the underlying information of the data.
2. Results are affected by subjective factors: conventional methods typically rely on human experience and expertise, and the results are susceptible to subjective factors, lacking in objectivity and consistency.
3. The efficiency and the instantaneity are poor: due to the high computational complexity of the conventional method, a large amount of computational resources and time are required, which limits the efficiency and real-time of the submarine topography inversion.
For example, chinese patent publication No. CN113963121a proposes a global submarine topography model building method and system, the method comprising: step 1: dividing and cutting the global sea area by adopting square grids with the side length of 2 degrees; step 2: based on the global sea area grids divided in the step 1, starting from the upper left corner, and recovering the 2 degree multiplied by 2 degree target sea area submarine topography data formed by cutting one by one according to a submarine topography model construction strategy of a multi-inversion method from left to right and from top to bottom; step 3: and (3) splicing and merging all the 2 degree multiplied by 2 degree target sea area submarine topography data obtained in the step (2) so as to obtain a global sea area submarine topography data set, namely a global submarine topography model. The invention takes the advantages of submarine topography model construction efficiency, inversion method applicability and sea surface gravity field inversion into consideration, and overcomes the defects of insufficient gravity input elements in the construction of the traditional global submarine topography model, weak independence performance of the topography model and single use of the inversion method. However, there is no description of the problem of solving the problem of error accumulation and data noise, nor of the drawbacks in processing gravitational field data.
For another example, chinese patent publication No. CN115081320a proposes a method and a system for generating a submarine topography model, which uses an improved CNN network to generate a mid-band submarine topography model, without ignoring the influence of nonlinear terms and higher order terms, and reducing the influence of uneven distribution and sparseness of band data in ship survey, thereby greatly improving the accuracy and precision of the submarine topography model. However, there is no reference to solving the problem of error accumulation and data noise, nor to solving the drawbacks in processing gravitational field data.
On the other hand, some machine learning algorithms, such as random forests, support Vector Machines (SVMs), etc., are also applied to the undersea terrain inversion technique. While these shallow machine learning methods improve the accuracy and efficiency of inversion to some extent, they are primarily suited to handle simple linear problems, have limited capacity for complex nonlinear problems, and cannot effectively mine deep nonlinear relationships between terrain and gravitational field data.
Therefore, the technical problems to be solved in the current field are: how to design a novel submarine topography inversion method, the method can overcome the problems, improve inversion precision and efficiency, reduce manual intervention, fully excavate and utilize potential information of gravitational field data, and particularly the complex nonlinear relation in the data; this is an important technical challenge facing current inversion of the seafloor topography.
Disclosure of Invention
The invention aims to solve the technical problems that: the method for generating the submarine topography inversion model and inverting the submarine topography based on the residual error aims to improve inversion accuracy, reduce calculation complexity, realize automatic processing, fully mine potential information of gravitational field data and improve inversion objectivity, consistency, efficiency and instantaneity.
The invention relates to a residual error-based submarine topography inversion model generation method, which comprises the following steps:
model determination: designing a basic residual error module contained in a residual error depth neural network model;
setting a model input: taking the shortwave gravity anomaly component and the longitude and latitude of the control point as model input;
model optimization: performing model optimization by adopting a first momentum gradient descent method;
model implementation: model implementation was performed using a Keras framework.
By taking the short wave gravity anomaly component and the longitude and latitude of the control point as model input, the complex nonlinear relation between complex terrain and gravity field data is excavated, the precision of submarine topography inversion is improved, and the real-time or efficient inversion requirement is met. The model optimization is carried out by adopting a first momentum gradient descent method, so that the problem of gradient explosion when a network deepens is avoided. The Keras framework is used for model realization, so that the efficiency and the instantaneity are improved, and the problems of high dependence on human resources and extra human cost are solved.
Preferably, the residual deep neural network model comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a first basic residual module, three second basic residual modules and three third basic residual modules, and neurons are contained in the first basic residual module (201), the second basic residual module (202) and the third basic residual module (203).
The design of the residual depth neural network model ensures that the model has enough complexity and can capture the complex relationship between the submarine topography and the shortwave gravity anomaly
Preferably, each basic residual module sequentially comprises two fully connected layers FC with a ReLU activation function, one Dropout layer and one Add layer with an activation function.
Specifically, the form of the base residual block is expressed as:
FC1 →ReLU→FC2 →Dropout→Add→ReLU；
wherein, FC1 and FC2 represent two fully connected layers, reLU represents the ReLU activation function carried by the two fully connected layers, dropout represents the Dropout layer, add represents the Add layer, and ReLU represents the ReLU activation function carried by the Add layer.
Preferably, a propagation formula between all connection layers FC in the base residual module is as follows:
；
in the method, in the process of the invention,representing the output of the layer 1 ith neuron to the layer 2 jth neuron,/>Input representing layer 1, i neuron of the neural network,/i>Representing the weights of the layer 1 ith neuron to the layer 2 jth neuron,/>Representing the bias coefficients of the layer 1 ith neuron to the layer 2 jth neuron.
Preferably, the propagation formula between each layer of the basic residual module is as follows:
；
in the method, in the process of the invention,representing the final output of the module,/>Representing Add layer->Representing the output of the layer 2 ith neuron to the Add layer jth neuron,/>Input representing layer 1, i neuron of the neural network,/i>Representing the weights of the i-th neuron of layer 1 to the j-th neuron of the Add layer,/->Representing the bias coefficients of the layer 1 ith neuron to the Add layer jth neuron.
Preferably, the setting model input is expressed as:
；
in the method, in the process of the invention,representing model input, ++>Representing longitude and latitude of control point->Representing a short wave gravity anomaly component.
Such model input settings allow our model to learn features of the seafloor terrain from geographic location and gravity anomaly data.
Preferably, the model optimization includes:
setting an optimizer: adam optimizers were chosen for model training.
Setting a learning rate and an adjustment scheme: the learning rate adjustment scheme is a cosine annealing scheme; the cosine annealing scheme dynamically adjusts the learning rate according to the training progress, so that the model training is more stable and effective; the specific learning rate adjustment formula is as follows:
；
where lr is the learning rate, T is the current iteration number, and T is the total iteration number;
determining batch and iteration times: setting the minimum batch and iteration times of each training;
selecting a loss function: the mean square error is selected as a loss function, and the specific formula is as follows:
；
wherein y is p Is the predicted value, y t Is the true value and N is the number of samples.
Setting evaluation indexes: the root mean square error is selected as an evaluation index.
Preferably, the model implementation implements a residual depth neural network model under a Keras framework.
Specifically, firstly, a model is built layer by using a Sequential API of Keras, then the model is compiled by utilizing a rule method, an Adam optimizer and a mean square error loss function are set, and meanwhile, an evaluation index is set as a root mean square error; finally, performing model training by using a fit method, and setting training batches and iteration times; through the steps, the residual depth neural network model is realized under the Keras framework.
According to the residual-based submarine topography inversion method, the model generated by the residual-based submarine topography inversion model generation method is used for inverting the submarine topography.
Compared with the prior art, the invention has the following beneficial effects:
1. higher precision: due to the residual structure design of the model and the complex input data processing method, the invention can learn the complex relation between the submarine topography and the gravity data more accurately, thereby realizing the inversion of the submarine topography with higher precision. The model structure and the data processing mode are designed, so that the model can effectively learn the complex characteristics of the data, and the problem of gradient explosion is avoided, which cannot be achieved by the prior art.
2. The calculation efficiency is high: through the design of the neural network and the residual structure, the model can realize rapid forward propagation and backward propagation, and the calculation efficiency is greatly improved. The method is characterized in that the design of the model can reduce the gradient explosion problem during network training, and the training process of the model is more efficient through reasonable structural design and selection of an optimizer.
3. Reducing manual intervention: the invention is based on the automatic design of deep learning, reduces the dependence on professional staff, improves the degree of automation and intellectualization, and reduces the manual intervention. This is because the parameters of the model can be automatically updated by self-optimization in the model training process, and manual participation is not required.
4. Efficiency is improved: due to high calculation efficiency and reduced manual intervention, the invention can greatly improve the efficiency of submarine topography inversion. The model is high in calculation efficiency, training and prediction can be automatically performed, manual intervention is not needed, and inversion speed is greatly improved.
Drawings
FIG. 1 is a network structure diagram of a residual deep neural network model according to the present invention;
FIG. 2 is a schematic diagram of a basic residual module according to the present invention;
FIG. 3 is a graph of the inversion results of the RDNN model of the present invention;
FIG. 4 is a GGM model inversion result graph;
FIG. 5 shows the RDNN model checking result, wherein a is the deep sea error quantity checking result, and b is the deep sea deep checking result;
fig. 6 shows GGM model checking results, where a is deep sea error quantity checking result and b is deep sea deep checking result.
In the figure: 1. an input layer; 2. a hidden layer; 3. an output layer; 201. a first base residual module; 202. a second base residual module; 203. and a third base residual module.
Detailed Description
In order to more clearly illustrate the embodiments or technical solutions of the present invention, the following description of the embodiments or technical solutions will be briefly introduced, and it is obvious that the following description is only some embodiments of the present invention, and other embodiments may be obtained according to these embodiments without inventive effort to those of ordinary skill in the art.
Example 1
In this embodiment, we designed and implemented a model of the inversion of the seafloor terrain based on a Residual Depth Neural Network (RDNN). In order to solve the limitations of the prior art, such as low prediction accuracy, high calculation complexity, excessive dependence on manual operation and low inversion efficiency, the scheme mainly adopts the following technical means:
the following are 10 steps of reconstructing a residual depth neural network-based seafloor terrain inversion model (i.e., a residual-based seafloor terrain inversion model generation method):
1. designing a network structure: first, we designed a residual depth neural network model consisting of 7 basic residual modules. As shown in fig. 1, the residual deep neural network model includes an input layer 1, a hidden layer 2 and an output layer 3, where the hidden layer 2 includes a first basic residual module 201, three second basic residual modules 202 and three third basic residual modules 203, the first basic residual module 201 is a module including 32 neurons, the second basic residual module 202 is a module including 64 neurons, and the third basic residual module 203 is a module including 128 neurons. This design allows our model to be of sufficient complexity to capture complex relationships between seafloor topography and short wave gravity anomalies.
2. Determining module content: each base residual module includes two fully connected layers (FCs) with a ReLU activation function, one Dropout layer, and one Add layer with an activation function. As shown in fig. 2, the form of the module may be expressed as:
FC1 →ReLU→FC2 →Dropout→Add→ReLU
the specific propagation process in the residual block is as follows:
(1) propagation process between fully connected layers
The method comprises the steps of carrying out a first treatment on the surface of the In (1) the->Representing the output of the layer 1 ith neuron to the layer 2 jth neuron,/>Input representing layer 1, i neuron of the neural network,/i>Representing the weights of the layer 1 ith neuron to the layer 2 jth neuron,/>Representing the bias coefficients of the layer 1 ith neuron to the layer 2 jth neuron.
(2) Propagation of jump connections (i.e. propagation between layers)
The method comprises the steps of carrying out a first treatment on the surface of the In (1) the->Representing the final output of the module,/>Representing Add layer->Representing the output of the layer 2 ith neuron to the Add layer jth neuron,/>Input representing layer 1, i neuron of the neural network,/i>Representing the weights of the i-th neuron of layer 1 to the j-th neuron of the Add layer,/->Representing the bias coefficients of the layer 1 ith neuron to the Add layer jth neuron.
3. Calculating the shortwave gravity anomaly component: first we calculate the short wave gravity anomaly component of the control point using the following formula:
the method comprises the steps of carrying out a first treatment on the surface of the In (1) the->Control point for measuring water depth of ship>An abnormal component of gravity at the location,is the gravitational constant; />Is the density difference constant of sea water and sea bottom; />For control point->Deep sea in the region, deep sea in the region>For reference sea depth, the maximum water depth value is taken.
Then, useCalculating +.>In the formula->Sea depth is measured for the control point ship:
；
the space interpolation method obtains the long-wave gravity data of the inversion area regular grid, and further calculates the long-wave gravity grid data of the points without ship measuring the water depth, and the method is as follows:
the method comprises the steps of carrying out a first treatment on the surface of the The short wave gravity data is used as one of inputs when the RDNN model predicts sea depths.
4. Setting a model input: and taking the calculated short wave gravity anomaly component and the longitude and latitude of the control point as the input of the model. The model input can be expressed as:
the method comprises the steps of carrying out a first treatment on the surface of the Such an arrangement allows our model to learn the characteristics of the seafloor terrain from the geographic location and gravity anomaly data.
5. Setting an optimizer: we selected Adam optimizers for model training. The Adam optimizer is an effective first momentum gradient descent method and can automatically adjust the learning rate.
6. Setting a learning rate and an adjustment scheme: we set the learning rate to 0.001 and the learning rate adjustment scheme is the cosine annealing scheme. The cosine annealing scheme can dynamically adjust the learning rate according to the training progress, so that model training is more stable and effective. The specific learning rate adjustment formula is as follows:
the method comprises the steps of carrying out a first treatment on the surface of the Where lr is the learning rate, T is the current iteration number, and T is the total iteration number.
7. Determining batch and iteration times: we set the minimum batch per training to 128 samples and the number of iterations to 40. Such an arrangement may ensure that there are enough samples to participate in each training while limiting the time of the training to prevent overfitting.
8. Selecting a loss function: we choose the mean square error (mse) as the loss function. The mean square error is an average value of squares of differences between the predicted value and the true value, and can accurately reflect the accuracy of model prediction. The specific formula is as follows:
the method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Is a predictive value->Is the true value and N is the number of samples.
9. Setting evaluation indexes: we selected root mean square error (rmse) as an evaluation index. The root mean square error is the square root of the mean square error, and can reflect the error of model prediction more intuitively.
10. Implementation model: after all parameters and structure are determined, we next need to implement the model. Specifically, we choose to use a kensflow-based Keras framework for model implementation. The Keras framework takes Tensorflow as the back end, provides an advanced programming interface for the realization of a deep learning algorithm, and can run seamlessly on a CPU and a GPU. Features of the framework include user friendliness, modularity and combinability, ease of expansion, and the like. By utilizing the characteristics, the RDNN model can be conveniently and quickly realized.
In this process, we first build a model layer by layer using the Sequential API of Keras, then compile the model using the rule method, set the Adam optimizer and mean square error (mse) loss function we selected in step 5 and step 8, and set the evaluation index as root mean square error (rmse). Finally, we use fit method to perform model training, set the training batch and iteration number as the value determined in step 7.
Through the steps, the RDNN model is realized under the Keras framework, so that the model design and training process is concise and efficient, and the operation and the debugging are easy, and an effective technical scheme for realizing the efficient submarine topography inversion is provided.
Through the steps, the scheme can effectively improve the prediction precision, reduce the calculation complexity, reduce the manual operation and improve the inversion efficiency, thereby solving the problems in the prior art.
In this example, we used the RDNN model and the conventional Gravity Method (GGM) to perform the following in the seafloor topography inversion experiment of Ma Liya nano-sea ditches:
RDNN model: as shown in fig. 3 and 4, the RDNN model uses a residual deep neural network that learns complex relationships between gravity anomaly data and the seafloor terrain using a deep learning method. In the submarine topography inversion experiment of the Malaysia sea ditch, through the evaluation of the actually measured check point water depth, the root mean square error of the RDNN model is 128.98m, good precision is shown, and good consistency is shown with the ship measurement check water depth.
GGM model: in contrast to the RDNN model, the GGM model is a gravity-geological based model for terrain inversion. Under the same experimental conditions, the root mean square error of the GGM model is 150.14m, and the accuracy is inferior to that of the RDNN model.
Fig. 5 and 6 are graphs of the distribution density of the scatter of the sea floor topography errors calculated using the ship's survey depth check points, and it can be found that the RDNN model results are more concentrated, closer to the ship's survey sea depth, and more similar. The correlation coefficient between the ship's survey sea depth and RDNN sea depth is 0.997, which is higher than 0.994 of GGM sea depth.
By the comparison, the RDNN model is superior to the traditional GGM model in the accuracy of submarine topography inversion, and the consistency with the ship survey and inspection water depth is higher, so that the RDNN model has the superiority in submarine topography inversion. This advantage comes from the network structure and data processing method of the RDNN model, which enables it to better capture the complex nature of the data and to predict it more accurately.
Example 2
The embodiment discloses a submarine topography inversion method based on residual errors, which is to invert a model generated by the submarine topography inversion model generation method based on residual errors described in the embodiment 1, and can improve inversion precision, reduce calculation complexity, reduce manual intervention, improve inversion efficiency, and solve the following main defects in the prior art:
1. shallow machine learning methods such as random forests and SVMs can only deal with simple linearity problems, and cannot mine nonlinear relations between complex terrain and gravitational field data, so that the accuracy of submarine terrain inversion is limited.
2. The deep learning-based submarine topography inversion method has the problem that gradient explosion occurs when a network deepens, namely when the learning rate of a front hidden layer is lower than that of a rear hidden layer, the classification accuracy is lowered along with the increase of the number of hidden layers.
3. The existing submarine topography inversion method is high in calculation complexity, and the submarine topography inversion is carried out by adopting a complex mathematical model and an iterative algorithm, so that calculation is time-consuming and not efficient enough.
4. The prior art needs to rely on professionals for data processing and interpretation, and lacks the characteristics of automation and intelligence, resulting in high dependence on human resources and additional human cost.
5. The traditional submarine topography inversion method usually ignores complex nonlinear relations in the data when processing gravitational field data, has low inversion efficiency and cannot meet real-time or efficient inversion requirements.
In summary, the invention discloses a submarine topography inversion model generation and submarine topography inversion method based on residual errors, which improves the performance of the model by utilizing a residual error structure and effectively solves the problem of gradient explosion in a depth network by the residual error structure. This design enables our model to learn effectively the complex relationship between the seabed terrain and gravity data while maintaining a deeper level. This is the core structure of the model, which has a decisive influence on the performance of the model.
The method for generating the submarine topography inversion model and inverting the submarine topography based on the residual error provided by the invention is described in detail. The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to facilitate an understanding of the method of the present invention and its core ideas. It should be noted that it will be apparent to those skilled in the art that various modifications and adaptations of the invention can be made without departing from the principles of the invention and these modifications and adaptations are intended to be within the scope of the invention as defined in the following claims.
Claims (10)
1. The method for generating the submarine topography inversion model based on the residual error is characterized by comprising the following steps of:
model determination: designing a basic residual error module contained in a residual error depth neural network model;
setting a model input: taking the shortwave gravity anomaly component and the longitude and latitude of the control point as model input;
model optimization: performing model optimization by adopting a first momentum gradient descent method;
model implementation: model implementation was performed using a Keras framework.
2. The method for generating the residual-based submarine topography inversion model according to claim 1, wherein the residual depth neural network model comprises an input layer (1), a hidden layer (2) and an output layer (3), the hidden layer (2) comprises a first basic residual module (201), three second basic residual modules (202) and three third basic residual modules (203), and neurons are contained in the first basic residual module (201), the second basic residual modules (202) and the third basic residual modules (203).
3. The method for generating a residual-based seafloor terrain inversion model according to claim 1, wherein each of said basic residual modules comprises, in order, two fully connected layers FC with ReLU activation functions, one Dropout layer, and one Add layer with activation functions.
4. A method of generating a residual-based seafloor terrain inversion model as claimed in claim 3, wherein the propagation formula between fully connected layers FC in the basic residual module is as follows:
in (1) the->Representing the output of the layer 1 ith neuron to the layer 2 jth neuron,input representing layer 1, i neuron of the neural network,/i>Representing the weights of the layer 1 ith neuron to the layer 2 jth neuron,/>Representing the bias coefficients of the layer 1 ith neuron to the layer 2 jth neuron.
5. A method of generating a residual-based seafloor terrain inversion model as claimed in claim 3, wherein the propagation formula between layers of the basic residual module is as follows:
in (1) the->Representing the final output of the module,/>Representing Add layer->Representing the output of the layer 2 ith neuron to the Add layer jth neuron,/>Input representing layer 1, i neuron of the neural network,/i>Representing the weights of the i-th neuron of layer 1 to the j-th neuron of the Add layer,/->Representing the bias coefficients of the layer 1 ith neuron to the Add layer jth neuron.
6. The residual-based seafloor terrain inversion model generation method of claim 1, wherein the set model input is represented as:
in (1) the->Representing model input, ++>Representing longitude and latitude of control point->Representing a short wave gravity anomaly component.
7. The residual-based seafloor terrain inversion model generation method of claim 1, wherein the model optimization comprises:
setting an optimizer: selecting an Adam optimizer for model training;
setting a learning rate and an adjustment scheme: the learning rate adjustment scheme is a cosine annealing scheme; the cosine annealing scheme dynamically adjusts the learning rate according to the training progress, so that the model training is more stable and effective;
determining batch and iteration times: setting the minimum batch and iteration times of each training;
selecting a loss function: selecting a mean square error as a loss function;
setting evaluation indexes: the root mean square error is selected as an evaluation index.
8. The method for generating the residual-based submarine topography inversion model according to claim 1, wherein the model is realized by firstly constructing the model layer by using a Sequential API of Keras, then compiling the model by using a compatibility method, setting an Adam optimizer and a mean square error loss function, and setting an evaluation index as a root mean square error; finally, performing model training by using a fit method, and setting training batches and iteration times; the residual depth neural network model is implemented under the Keras framework.
9. The method of generating a residual-based seafloor terrain inversion model according to claim 2, wherein the first base residual module (201) is a module comprising 32 neurons, the second base residual module (202) is a module comprising 64 neurons, and the third base residual module (203) is a module comprising 128 neurons.
10. A residual-based seafloor terrain inversion method, characterized in that the model generated by the residual-based seafloor terrain inversion model generation method according to any one of claims 1-9 is applied to invert seafloor terrain.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311126318.XA CN116861799B (en) | 2023-09-04 | 2023-09-04 | Submarine topography inversion model generation and submarine topography inversion method based on residual errors |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311126318.XA CN116861799B (en) | 2023-09-04 | 2023-09-04 | Submarine topography inversion model generation and submarine topography inversion method based on residual errors |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116861799A true CN116861799A (en) | 2023-10-10 |
| CN116861799B CN116861799B (en) | 2023-11-24 |
Family
ID=88219381
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311126318.XA Active CN116861799B (en) | 2023-09-04 | 2023-09-04 | Submarine topography inversion model generation and submarine topography inversion method based on residual errors |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116861799B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117113857A (en) * | 2023-10-23 | 2023-11-24 | 自然资源部第一海洋研究所 | Full-connection depth neural network model and method for inverting submarine topography |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2020020872A (en) * | 2018-07-30 | 2020-02-06 | 国立研究開発法人情報通信研究機構 | Discriminator, learnt model, and learning method |
| CN113887342A (en) * | 2021-09-16 | 2022-01-04 | 上海工业自动化仪表研究院有限公司 | Equipment fault diagnosis method based on multi-source signals and deep learning |
| CN114137624A (en) * | 2021-10-27 | 2022-03-04 | 中国海洋大学 | Method and system for inverting submarine topography based on satellite altimeter |
| CN115130662A (en) * | 2022-08-12 | 2022-09-30 | 电子科技大学 | Power distribution network time-varying topological state estimation method based on transfer learning |
| CN115327451A (en) * | 2022-08-23 | 2022-11-11 | 中国科学技术大学 | NV color center magnetometer magnetic field calculation method based on deep learning and application |
| WO2023042045A1 (en) * | 2021-09-16 | 2023-03-23 | 3M Innovative Properties Company | Convolution attention network for multi-label clinical document classification |
-
2023
- 2023-09-04 CN CN202311126318.XA patent/CN116861799B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2020020872A (en) * | 2018-07-30 | 2020-02-06 | 国立研究開発法人情報通信研究機構 | Discriminator, learnt model, and learning method |
| CN113887342A (en) * | 2021-09-16 | 2022-01-04 | 上海工业自动化仪表研究院有限公司 | Equipment fault diagnosis method based on multi-source signals and deep learning |
| WO2023042045A1 (en) * | 2021-09-16 | 2023-03-23 | 3M Innovative Properties Company | Convolution attention network for multi-label clinical document classification |
| CN114137624A (en) * | 2021-10-27 | 2022-03-04 | 中国海洋大学 | Method and system for inverting submarine topography based on satellite altimeter |
| CN115130662A (en) * | 2022-08-12 | 2022-09-30 | 电子科技大学 | Power distribution network time-varying topological state estimation method based on transfer learning |
| CN115327451A (en) * | 2022-08-23 | 2022-11-11 | 中国科学技术大学 | NV color center magnetometer magnetic field calculation method based on deep learning and application |
Non-Patent Citations (2)
| Title |
|---|
| 李遥;李烨;: "基于深度残差网络的DeepFM点击率预测模型", 软件导刊, no. 03 * |
| 高培贤;魏立线;刘佳;刘明明;: "基于密集连接网络的图像隐写分析", 计算机工程与应用, no. 15 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117113857A (en) * | 2023-10-23 | 2023-11-24 | 自然资源部第一海洋研究所 | Full-connection depth neural network model and method for inverting submarine topography |
| CN117113857B (en) * | 2023-10-23 | 2024-01-30 | 自然资源部第一海洋研究所 | Full-connection depth neural network model and method for inverting submarine topography |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116861799B (en) | 2023-11-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109709603B (en) | Seismic horizon identification and tracking method and system | |
| CN116861799B (en) | Submarine topography inversion model generation and submarine topography inversion method based on residual errors | |
| US20210026031A1 (en) | Dolomite reservoir prediction method and system based on well and seismic combination, and storage medium | |
| US10439594B2 (en) | Actually-measured marine environment data assimilation method based on sequence recursive filtering three-dimensional variation | |
| Park et al. | Handling conflicting multiple objectives using Pareto-based evolutionary algorithm during history matching of reservoir performance | |
| Karssenberg et al. | Conditioning a process-based model of sedimentary architecture to well data | |
| CN110031895B (en) | Multipoint geostatistical stochastic inversion method and device based on image stitching | |
| KR101917734B1 (en) | Tsunami forecast method using numerical model and scenario database | |
| Agarwal et al. | Statistical field estimation for complex coastal regions and archipelagos | |
| CN111722283B (en) | Stratum velocity model building method | |
| CN114896564A (en) | Transient electromagnetic two-dimensional Bayesian inversion method adopting self-adaptive Thiessen polygon parameterization | |
| CN116663428B (en) | Neural network-based face wave imaging joint inversion method | |
| FR3040809A1 (en) | ||
| CN110441815B (en) | Simulated annealing Rayleigh wave inversion method based on differential evolution and block coordinate descent | |
| CN108133286B (en) | Underground water multi-target calculation method based on ground settlement substitution model | |
| CN114286279A (en) | Sound field obtaining method of three-dimensional sound field model based on global matrix coupling normal wave | |
| CN114139437A (en) | Method and system for inverting submarine topography by using satellite height measurement data | |
| CN114488301A (en) | Method for predicting porosity of tight sandstone reservoir | |
| Xie et al. | Seismic inverse modeling method based on generative adversarial networks | |
| CN112649869A (en) | Reservoir characteristic parameter prediction method and system based on GA-WNN | |
| Gao et al. | Combining the genetic algorithms with BP Neural Network for GPS height Conversion | |
| CN117709207B (en) | Method, device, equipment and medium for designing measurement wiring of multi-beam measurement ship | |
| CN113219534B (en) | Pre-stack depth migration speed quality control method and device, medium and electronic equipment | |
| US20210389493A1 (en) | Modeling presence and quality of original organic materials in a region | |
| CN115267932A (en) | Improved gravity geological method for recovering sea-land junction area seabed terrain |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |