Ocean data reconstruction method for generating missing features based on conditional information

Abstract

The invention relates to an ocean data reconstruction method for generating missing features based on conditional information. The ocean data reconstruction method comprises the following steps: dividing time periods and constructing a sample set; building a missing feature generation model; training and verifying a missing feature generation model; building an ocean data reconstruction model; and training and verifying the ocean data reconstruction model. According to the constructed missing feature generation model, controllable condition information is input as a constraint, the generator is guided to generate the ocean element field, and compared with an original generator adopting random disturbance, convergence is easier, and the generated ocean element field is more real and more practical; in the training process, the design of combining confrontation loss and L1 loss is adopted, so that the stability and efficiency of the model and the quality and precision of the generated element field are improved.

Description

Technical Field

[0001] The present invention relates to the field of marine science and machine learning technology, and in particular to a method for reconstructing marine data by generating missing features based on conditional information. Background Art

[0002] In response to the problem that "the traditional linear method model is too simple and the numerical model method is complex to calculate, which makes it difficult to effectively integrate multi-source heterogeneous marine environmental data", the invention patent number is CN202211050583.X, and the name is "Three-dimensional marine environmental field intelligent fusion method and system based on deep learning". By constructing a deep learning network model, it realizes the intelligent fusion of multi-source marine data, thereby generating a three-dimensional marine element fusion field that is closer to observation. However, due to the multi-source, heterogeneous, and multi-dimensional characteristics of marine environmental data, the effective years of the data used for fusion are different. The above method works well in the period after the 1990s when marine data is rich, but there are certain shortcomings in the early history.

[0003] Due to the late appearance of ocean satellites, early ocean field observation data was relatively scarce, and the types and number of ocean observations that can be assimilated were limited, resulting in poor accuracy and consistency of three-dimensional gridded data sources before the 1990s. High-resolution, high-precision high-quality data sources such as HYCOM and SODA started relatively late, and the intelligent fusion method lacks high-quality data sources covering the early history, which brings great challenges to the historical reconstruction of ocean elements. There are two common ways to deal with the problem of lack of high-quality fusion data sources in the early history: one is to retrain the intelligent fusion model in the early history, and the input features do not include the missing high-quality data sources; the other is to use the climatological element field of the high-quality data source to supplement the missing time period. However, the accuracy of the ocean element reconstruction field obtained by these two methods has certain limitations and cannot meet the needs of refined scientific research. Summary of the invention

[0004] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide an ocean data reconstruction method for generating missing features based on conditional information, so as to solve one or more problems in the prior art.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] The method for reconstructing ocean data based on conditional information to generate missing features includes the following steps:

[0007] Time period division and sample set construction;

[0008] Building a missing feature generation model;

[0009] Training and validation of missing feature generation models;

[0010] Building ocean data reconstruction models;

[0011] Training and validation of ocean data reconstruction models.

[0012] Furthermore, the time period division and sample set construction include the following steps:

[0013] Divide the feature missing period P1 and the feature complete period P2;

[0014] The sample set DS1 is constructed by the feature complete period P2 and the sample set DS2 is constructed by the feature missing period P1.

[0015] Furthermore, the missing feature generation model building includes the following steps:

[0016] Construct a generator G to integrate the known fusion data source ODS n Convert to generated fusion data source

[0017] Construct the discriminator D to generate the fusion data source and real fusion data sources Perform feature extraction and calculate the probability of being true;

[0018] Loss function design.

[0019] Furthermore, the loss function in the loss function design is designed based on adversarial loss, and the loss function is shown in the following formula:

[0020] Loss(f G ,f D )=Loss L1 (f G )+λLoss GAN (f G ,f D )

[0021] Where: λ is a hyperparameter that controls the adversarial loss Loss GAN (f G ,f D ) and L1 loss Loss L1 (f G )’s contribution to the overall optimization goal.

[0022] Furthermore, the adversarial loss calculation formula is as follows:

[0023] Loss GAN (f G ,f D )=E[logf D (ODS real)]+E[log(1-f D (ODS gen ))]

[0024] Where: E is the expectation operator, f D (i) represents the probability obtained after the data is input into the discriminator D.

[0025] Furthermore, the training and verification of the missing feature generation model is based on the fusion data source generated by comparing the sample set DS2 of the P1 period and real fusion data sources The root mean square error RMSE and the coefficient of determination R 2 .

[0026] Furthermore, the ocean data reconstruction model building includes the following steps:

[0027] Freeze generator G;

[0028] Feature extraction;

[0029] Data reconstruction.

[0030] Furthermore, the feature extraction includes feature extraction of fused data sources. Feature extraction of sea surface environmental elements and temporal and spatial information

[0031] Furthermore, the data reconstruction is based on the global fusion reconstruction of the feature Feat1 of the fusion data source and the feature Feat2 of the sea surface environment elements and spatiotemporal information to obtain the reconstructed ocean element field OD fusion .

[0032] Furthermore, the training and verification of the ocean data reconstruction model is based on the sample set DS1 of the P2 period, and the ocean element field OD is reconstructed by comparison. fusion The observation element field OD corresponding to the sample set DS2 obs The root mean square error RMSE and the coefficient of determination R 2 .

[0033] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0034] (1) The missing feature generation model constructed by the present invention guides the generator to generate an ocean feature field by inputting controllable conditional information as constraints. Compared with the original generator using random perturbations, it is easier to converge, and the generated ocean feature field is more realistic and more in line with reality. During the training process, a design combining adversarial loss and L1 loss is adopted to improve the stability and efficiency of the model as well as the quality and accuracy of the generated feature field.

[0035] (ii) Freeze the parameters of the trained generator and connect it to the ocean data reconstruction model. The feature extraction of 3D fusion data source is performed; the sea surface environmental elements and temporal and spatial information are added to the fusion process in the form of an associated network. This model structure design can learn different degrees of information in different dimensions, effectively avoiding the problem of one-dimensional and two-dimensional information loss due to the network being too deep, thereby making the reconstruction of the ocean feature field more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic flow chart of a method for reconstructing ocean data by generating missing features based on conditional information according to an embodiment of the present invention is shown.

[0037] Figure 2 A schematic diagram showing the composition of sample set DS1 and sample set DS2 of an ocean data reconstruction method for generating missing features based on conditional information according to an embodiment of the present invention.

[0038] Figure 3 The structure diagram of the missing feature generation model FFG-cGAN of the ocean data reconstruction method for generating missing features based on conditional information in an embodiment of the present invention is shown.

[0039] Figure 4 The structure diagram of the ocean data reconstruction model FFG-ODF-Net of the ocean data reconstruction method for generating missing features based on conditional information according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the present invention clearer, the ocean data reconstruction method based on conditional information to generate missing features proposed by the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structure, proportion, size, etc. illustrated in the accompanying drawings of this specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so they have no technical substantive significance. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effect that the present invention can produce and the purpose that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0041] See also Figure 1 , an ocean data reconstruction method for generating missing features based on conditional information, comprising the following steps:

[0042] Step 1: Time period division and sample set construction.

[0043] Step 1.1: Divide the feature missing period P1 and the feature complete period P2.

[0044] According to the time distribution of multi-source ocean data, it is divided into a feature missing period, that is, the time period when the fused data source is missing in the early history, recorded as P1, and a feature complete period, that is, the time period when the fused data source is complete in the recent history, recorded as P2.

[0045] Step 1.2: Construct sample set DS1 through feature complete period P2 and construct sample set DS2 through feature missing period P1.

[0046] Please continue reading Figure 2 , for sample set DS1, ODS n represents the nth set of fused data sources that both P1 and P2 have, st represents the corresponding spatiotemporal information of the temperature-salinity profile processed by the data feature generation model within the sample, i.e., the complete feature period, Indicates the mth set of fused data sources that P2 has but P1 does not have.

[0047] Furthermore, for sample set DS2, represents the mth set of fused data sources generated by the missing feature generation model FFG-cGAN, SSx represents the corresponding sea surface environmental elements of the temperature-salinity profile processed by the data fusion model during the sample feature missing period, OD obs It represents the corresponding observation element field of the temperature-salinity profile processed by the data fusion model during the sample, i.e., the feature missing period.

[0048] Step 2: Build the missing feature generation model.

[0049] The missing feature generation model FFG-cGAN model is based on the maximum-minimum game theory. It uses the generator G and the discriminator D to compete with each other to generate "fake data" that cannot be distinguished from the real fusion data source.

[0050] Please continue reading Figure 1 and Figure 3 , the missing feature generation model building includes the following steps:

[0051] Step 2.1: Build the generator G.

[0052] The goal of the generator G is to transform the known fusion data source ODS n That is, the controllable condition information mapping as input is converted into the generated fusion data source Make it difficult for the discriminator D to distinguish whether the input fusion data source is real or generated by the generator G.

[0053] Step 2.2: Build the discriminator D.

[0054] Generated fusion data source and real fusion data sources Feature extraction is performed and the probability of being true is calculated to maximize the ability to distinguish between real data and generated data, thereby improving the effect of adversarial training.

[0055] Step 2.3: Loss function design.

[0056] The loss function in the loss function design is designed based on adversarial loss, and the loss function of the missing feature generation model FFG-cGAN is shown in the following formula 1:

[0057] Loss(f G ,f D )=Loss L1 (f G )+λLoss GAN (f G ,f D ) (1)

[0058] Where: λ is a hyperparameter that controls the adversarial loss Loss GAN (f G ,f D ) and L1 loss Loss L1 (f G )’s contribution to the overall optimization goal.

[0059] Furthermore, the adversarial loss calculation formula is shown in Formula 2 below:

[0060] Loss GAN (f G ,f D )=E[logf D (ODS real )]+E[log(1-f D (ODS gen ))] (2)

[0061] Where: E is the expectation operator, f D (i) represents the probability obtained after the data is input into the discriminator D.

[0062] Step 3: Training and validation of feature-deficient generation model.

[0063] The training and verification of the missing feature generation model is based on the sample set DS1 of the P2 period to train the missing feature generation model FFG-cGAN built in step 2, and compare the generated fusion data source and real fusion data sources The root mean square error RMSE and the coefficient of determination R 2 , and then verify the model effect.

[0064] Step 4: Build the ocean data reconstruction model.

[0065] Please continue reading Figure 1 and Figure 4 The construction of the ocean data reconstruction model includes the following steps:

[0066] Step 4.1: Freeze the generator G.

[0067] Based on the generator G of the missing feature generation model FFG-cGAN trained in step 2, freeze its parameters and build the ocean data reconstruction model FFG-ODF-Net.

[0068] Step 4.2: Feature extraction.

[0069] Combining sample set DS1 and sample set DS2, the feature extraction includes feature extraction of fusion data source Feature extraction of sea surface environmental elements SSx and temporal and spatial information st Two processes. The feature extraction of fusion data source is to merge the known data sources ODS n and the generated fusion data source obtained by generator G in sample set DS2 Splicing is performed to extract the features of the 3D fusion data source. The feature extraction of the sea surface environmental element SSx and the spatial and temporal information st It is to add the sea surface environmental elements SSx and spatiotemporal information st into the fusion process in the form of an associated network.

[0070] Step 4.3: Data reconstruction.

[0071] The process in step 4.2 above The extracted feature is recorded as Feat1, and the process f D2 The extracted features are recorded as Feat2. The data reconstruction is based on the global fusion reconstruction of the features Feat1 of the fusion data source and the features Feat2 of the sea surface environmental elements and temporal and spatial information to obtain the reconstructed ocean element field OD fusion .

[0072] Step 5: Training and validation of ocean data reconstruction model.

[0073] Based on the sample set DS2 of the P1 period, the ocean data reconstruction model FFG-ODF-Net built in step 4 is trained. By comparing the reconstructed ocean feature field OD fusion The observation element field OD corresponding to the sample set DS2 obsThe root mean square error RMSE and the coefficient of determination R 2 , and then verify the model effect.

[0074] End the process.

[0075] Specifically, multi-source sea temperature data from 1994 to 2017 are used to train and verify the missing feature generation model FFG-cGAN in the method described in the present invention. Data from 1994, 2006 and 2017 are selected as the test set, and the remaining data are the training and verification set. The known fusion data sources, i.e., the controllable condition information, are the sea temperatures of CORA, ORAS5, BCC-CSM2-MR, CESM2, HYCOM climate state and SODA climate state, and the output is the sea temperatures of ECCO, SODA and HYCOM. The indicators on the test set are shown in Table 1 below.

[0076] Table 1

[0077]

[0078] From Table 1 above, we can conclude that and The overall error is small, and the determination coefficient R 2 All of them are higher than 0.98, indicating that the missing feature generation model FFG-cGAN of the method described in the present invention can better fit the real fusion data source.

[0079] Furthermore, multi-source sea temperature data from 1958 to 1981 are used to train and verify the ocean data reconstruction model FFG-ODF-Net in the method described in the present invention. Data from 1958, 1970 and 1981 are selected as the test set, and the remaining data are the training and verification sets. Regarding the effectiveness of ocean data reconstruction, this embodiment sets up several groups of experiments for comparison. The experimental settings and indicators are shown in Table 2 below.

[0080] Table 2

[0081]

[0082] As shown in Table 2 above, by comparing Case A with Case B to Case E, it is found that the average RMSE of the sea temperature reconstruction result when the generated fusion data source is used is smaller than that when the generated fusion data source is not used, indicating that the high-quality fusion data source generated in addition has a significant improvement effect on the sea temperature reconstruction effect. By comparing Case B to Case E, it is found that the accuracy of the sea temperature reconstruction results obtained by using different generated data sources is also different. The accuracy of the sea temperature data is SODA>HYCOM>ECCO. Therefore, the accuracy of the corresponding label of the generated fusion data source is positively correlated with the accuracy of the final sea temperature reconstruction result. Case E showed the best performance in all experiments, indicating that generating more high-quality fusion data sources can further improve the accuracy of sea temperature data reconstruction.

[0083] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for reconstructing ocean data by generating missing features based on conditional information, characterized in that: The following steps are involved: Time period division and sample set construction; Building a missing feature generation model; Training and validation of missing feature generation models; Building ocean data reconstruction models; Training and validation of ocean data reconstruction models.

2. The method for reconstructing ocean data based on conditional information to generate missing features as claimed in claim 1, characterized in that: The time period division and sample set construction include the following steps: Divide the feature missing period P1 and the feature complete period P2; The sample set DS1 is constructed by the feature complete period P2 and the sample set DS2 is constructed by the feature missing period P1.

3. The method for reconstructing ocean data based on conditional information to generate missing features as claimed in claim 1, characterized in that: The missing feature generation model building includes the following steps: Construct a generator G to integrate the known fusion data source ODS n Convert to generated fusion data source Construct the discriminator D to generate the fusion data source and real fusion data sources Perform feature extraction and calculate the probability of being true; Loss function design.

4. The method for reconstructing ocean data based on conditional information to generate missing features as claimed in claim 3, characterized in that: The loss function in the loss function design is designed based on adversarial loss, and the loss function is shown in the following formula: Loss(f G ,f D )=Loss L1 (f G )+λLoss GAN (f G ,f D ) Where: λ is a hyperparameter that controls the adversarial loss Loss GAN (f G ,f D ) and L1 loss Loss L1 (f G )’s contribution to the overall optimization goal.

5. The method for reconstructing ocean data based on conditional information to generate missing features as claimed in claim 4, characterized in that: The adversarial loss calculation formula is as follows: Loss GAN (f G ,f D )=E[logf D (ODS real )]+E[log(1-f D (ODS gen ))] Where: E is the expectation operator, f D (i) represents the probability obtained after the data is input into the discriminator D.

6. The method for reconstructing ocean data based on conditional information to generate missing features as claimed in claim 1, characterized in that: The training and verification of the missing feature generation model is based on the sample set DS1 of the P2 period, and the fusion data source generated by comparison and real fusion data sources The root mean square error RMSE and the coefficient of determination R 2 .

7. The method for reconstructing ocean data based on conditional information to generate missing features as claimed in claim 1, characterized in that: The ocean data reconstruction model construction includes the following steps: Freeze generator G; Feature extraction; Data reconstruction.

8. The method for reconstructing ocean data based on conditional information to generate missing features according to claim 7, characterized in that: The feature extraction includes feature extraction of fused data sources Feature extraction of sea surface environmental elements and temporal and spatial information 9. The method for reconstructing ocean data by generating missing features based on conditional information according to claim 8, characterized in that: The data reconstruction is based on the global fusion reconstruction of the feature Feat1 of the fusion data source and the feature Feat2 of the sea surface environment elements and temporal and spatial information to obtain the reconstructed ocean element field OD fusion .

10. The method for reconstructing ocean data by generating missing features based on conditional information according to claim 1, characterized in that: The training and verification of the ocean data reconstruction model is based on the sample set DS2 of the P1 period, and the ocean element field OD is reconstructed by comparison. fusion The observation element field OD corresponding to the sample set DS2 obs The root mean square error RMSE and the coefficient of determination R 2 .

Images