An internal tidal wave spatial state inversion method, system, computer device and medium

Abstract

The application provides an internal tide wave space state inversion method, system, computer device and medium, and belongs to the field of marine environment analysis.The method comprises the following steps: obtaining sea temperature time series observation data of an internal tide wave region to be inverted, and extracting time series features of the internal tide wave to be inverted from the sea temperature time series observation data; mapping the time series features to obtain space state features of the internal tide wave to be inverted based on a pre-trained space-time feature correlation relationship, wherein the space-time feature correlation relationship is obtained by measuring the correlation matching of time series feature samples and space state feature samples; and reconstructing the spatial distribution of the internal tide wave to be inverted according to the space state features of the internal tide wave to be inverted, so as to obtain a spatial distribution inversion result of the internal tide wave to be inverted. The method infers the space state around the measurement point through the pre-trained space-time feature correlation relationship, so as to make up for the information loss of the measurement data in the spatial dimension, and ensure the accuracy of the spatial distribution inversion result of the internal tide wave.

Description

Technical Field

[0001] This invention belongs to the field of marine environmental analysis, specifically relating to a method, system, computer equipment, and medium for inverting the spatial state of internal tidal waves. Background Technology

[0002] Internal tidal waves are a key environmental factor affecting the operational safety of underwater vehicles and the laws governing underwater sound propagation. Accurately acquiring the three-dimensional distribution structure of ocean internal tidal waves is a fundamental guarantee for precise underwater target detection, covert navigation of underwater vehicles, and secure underwater acoustic communication. Due to the complex and variable nature of the underwater marine environment and its strong absorption of electromagnetic signals, it is currently difficult to achieve full-dimensional, high-density observation of the underwater marine environment. At present, we can only roughly estimate the spatial distribution of subsurface marine hydrological elements based on sparse underwater observation data, making it difficult to accurately construct the spatial state of small- to medium-scale oceanic processes such as internal tidal waves.

[0003] To address the significant difficulties and challenges faced in marine scientific research, existing technologies often employ multi-source observation data fusion, data assimilation, and artificial neural network techniques to comprehensively utilize observation data from different dimensions to construct a three-dimensional marine environmental state field, thereby accurately reconstructing the subsurface marine environmental state field. However, these construction methods typically require the observation data to contain sufficient spatial constraint information. Underwater vehicles, constrained by communication limitations and stealth requirements, can often only measure marine hydrological information at their location using onboard CTD and other observation equipment, unable to directly acquire state information such as temperature, salinity, and current velocity of the surrounding seawater. Constructing a three-dimensional marine environmental state field based on this observation data, lacking spatial constraint information, is particularly problematic due to the limited accuracy of the spatial distribution of internal tidal waves, making it difficult to guide the normal operation of underwater vehicles. Summary of the Invention

[0004] To address the problem that existing internal tidal space state inversion methods have low accuracy and cannot be used to guide the operation of underwater vehicles, this invention provides an internal tidal space state inversion method, system, computer equipment, and medium.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for inverting the spatial state of internal tidal waves includes: Acquire sea surface temperature time series observation data of the region to be inverted for internal tidal waves, and extract the time series characteristics of the internal tidal waves to be inverted from the sea surface temperature time series observation data; Based on the spatiotemporal feature correlation, the spatial state features of the internal tidal wave to be inverted are obtained by mapping the time series features. The spatiotemporal feature correlation is determined in advance through metric learning, with the optimization objective being to minimize the Euclidean distance between the matched spatial state features and the time series features in the sample, and maximize the Euclidean distance between the unmatched spatial state features and the time series features. The spatial distribution of the internal tidal wave to be inverted is reconstructed based on its spatial state characteristics, and the spatial distribution inversion result of the internal tidal wave to be inverted is obtained.

[0006] Optionally, in the spatial state inversion method for internal tidal waves provided by this invention, sea surface temperature time series observation data are processed through an internal tidal wave inversion model with a metric learning architecture to obtain the spatial distribution inversion result of the internal tidal waves to be inverted; the internal tidal wave inversion model includes a time autoencoder and a mapping module: Time series features were extracted from sea surface temperature time series observation data using a time autoencoder; The spatial state features are obtained by mapping the time series features through the mapping module.

[0007] Optionally, the method for inverting the spatial state of internal tidal waves provided by the present invention further includes: Acquire sea surface temperature data of the training area and extract spatial state feature samples from the sea surface temperature data; Acquire sea surface temperature time series observation data of the training area, and extract time series feature samples from the sea surface temperature time series observation data; Based on metric learning, spatial state features and time series features are associated and matched to establish spatiotemporal feature relationships; Matched spatial state features and time series features are used as positive samples, and unmatched spatial state features and time series features are used as negative samples. The optimization objective is to minimize the Euclidean distance between spatial state features and time series features in the positive samples and maximize the Euclidean distance between spatial state features and time series features in the negative samples. The parameters of the untrained internal tidal wave inversion model are optimized to obtain an internal tidal wave inversion model with spatiotemporal feature correlation.

[0008] Optionally, the internal tidal wave inversion model further includes a spatial autoencoder module, which includes a spatial encoder. The spatial encoder includes a spatial convolutional layer, an Inception layer, and a fully connected layer connected in sequence. The internal tidal wave spatial state inversion method provided by this invention also includes: Wavelet analysis was performed on the sea surface temperature data layer by layer and point by point to extract the wavelet coefficients, and inverse transformation was performed to obtain the filtered sea surface temperature data; empirical orthogonal function analysis was performed on the filtered sea surface temperature data to obtain the feature vector of the spatial distribution state of sea surface temperature. Local features are extracted from the feature vector of the spatial distribution of sea surface temperature through a spatial convolutional layer. Spatial features of different scales are extracted from the local features through an Inception layer. The spatial features of different scales are flattened and compressed through a fully connected layer of a spatial encoder to obtain spatial state feature samples.

[0009] Optionally, the time autoencoder includes a time encoder, which includes a convolutional layer and a fully connected layer. The internal tidal wave spatial state inversion method provided by this invention further includes: Time-series local features are extracted from sea surface temperature time-series observation data through the convolutional layer of the time encoder; The time series local features are mapped to the feature space through the fully connected layer of the time encoder, and then compressed to obtain time series feature samples.

[0010] Optionally, the method for inverting the spatial state of internal tidal waves provided by the present invention further includes: The Euclidean distance between the spatial state features and time series features in the positive and negative samples is substituted into the contrastive loss function to obtain the contrastive loss value. With the contrastive loss value as the optimization objective, the temporal encoder and spatial encoder are iteratively updated through backpropagation and batch training.

[0011] Optionally, the mapping module includes a first fully connected layer, a Transformer encoding module, and a second fully connected layer connected in sequence. The internal tidal wave spatial state inversion method provided by this invention also includes: The time series features are linearly transformed and nonlinearly activated by the first fully connected layer of the mapping module, and then mapped to the feature space to obtain the time feature vector located in the feature space. The dependencies of temporal feature vectors are calculated using the Transformer encoding module based on a multi-head attention mechanism, resulting in a feature vector weighted by the association relationships. The second fully connected layer of the mapping module adjusts the dimension of the weighted feature vectors of the association relationship to obtain spatial state features that match the dimension of the time series features.

[0012] The present invention also provides an internal tidal wave spatial state inversion system, comprising: The time series feature extraction module is used to acquire the sea surface temperature time series observation data of the inner tidal wave region to be inverted, and extract the time series features of the inner tidal wave to be inverted from the sea surface temperature time series observation data. The spatial state feature mapping module is used to map time series features to obtain the spatial state features of the internal tidal wave to be inverted based on the spatiotemporal feature correlation. The spatiotemporal feature correlation is pre-determined through metric learning, with the optimization objective being to minimize the Euclidean distance between the matched spatial state features and time series features in the sample, and maximize the Euclidean distance between the unmatched spatial state features and time series features. The spatial distribution inversion module is used to reconstruct the spatial distribution of the internal tidal wave to be inverted based on its spatial state characteristics, and to obtain the spatial distribution inversion result of the internal tidal wave to be inverted.

[0013] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in an internal tidal wave spatial state inversion method.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any step of an internal tidal space state inversion method.

[0015] The method for inverting the spatial state of internal tidal waves provided by this invention has the following beneficial effects: The present invention provides a method for inverting the spatial state of internal tides by acquiring sea surface temperature time series observation data, extracting time series features that can characterize the dynamic changes of internal tides, mapping the time series features to spatial state features of internal tides, and then, based on the mapped spatial state features, upsampling and reconstructing the spatial distribution of internal tides, and finally outputting the spatial distribution inversion result of the internal tides to be inverted.

[0016] In this invention, although the marine hydrological information observed by the underwater vehicle at its location lacks spatial information, the temporal resolution of the time series observation data is very high. By inferring the spatial state around the measurement point through pre-trained spatiotemporal feature correlations, the missing information in the spatial dimension of the measurement data is made up for, thus ensuring the accuracy of the spatial distribution inversion results of the internal tidal wave. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of an internal tidal wave spatial state inversion method provided by an embodiment of the present invention; Figure 2This is an example of a spatial autoencoder structure provided in an embodiment of the present invention, wherein, Figure 2 (a) is an example of a spatial encoder structure in a spatial autoencoder. Figure 2 (b) is an example of a spatial decoder structure in a spatial autoencoder. Figure 2 (c) is a specific structural example of the convolutional layer in a spatial autoencoder. Figure 2 (d) is a specific structural example of a multi-branch convolutional layer (Inception); Figure 3 This is an example of a time autoencoder structure provided in an embodiment of the present invention, wherein, Figure 3 (a) is an example of a time encoder structure in a time autoencoder. Figure 3 (b) is an example of a time decoder structure in a time autoencoder. Figure 3 (c) is a specific structural example of the convolutional layer in a temporal autoencoder; Figure 4 This is an example of a metric learning model structure diagram provided in an embodiment of the present invention; Figure 5 This is a flowchart example of a metric learning model for obtaining matching features, provided in an embodiment of the present invention. Figure 6 This is an example of inversion from time series to spatial distribution provided in an embodiment of the present invention; Figure 7 This is an example of sea surface temperature inversion in a certain sea area provided in an embodiment of the present invention, wherein, Figure 7 Examples of (a), (b), and (c) for the target field, temperature distribution, and inversion error in this sea area during summer are respectively. Figure 7 (d), (e), and (f) are examples of the target field, temperature distribution, and inversion error for the sea area in autumn. Detailed Implementation

[0019] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0020] Existing sea surface temperature (SST) retrieval technologies fall into two main categories: data assimilation techniques and intelligent retrieval techniques. However, regardless of the type of retrieval algorithm, they are all based on the assumption that the probability distribution of unknown variables has a simple spatial dependence on the observation data. This requires a high spatial sampling density of the observation data, while the time-varying information of the observation data is excluded as redundant data. However, when there is only one observation point in space, such as temperature observation on an anchored buoy, the spatial constraint information is completely missing, and these retrieval algorithms cannot perform spatial state retrieval and prediction.

[0021] The internal tidal wave spatial state inversion method provided by this invention makes full use of the high-density time-series information of single-point sea surface temperature observation data. Based on the mapping from the time-series characteristics of sea surface temperature observation data to the spatial characteristics of internal tidal waves, it constructs the spatial distribution state of internal tidal waves within a range of hundreds of kilometers around the observation point, realizing the perception of internal tidal wave spatial state information based on single-point observation data.

[0022] Example 1 This invention provides a method for inverting the spatial state of internal tidal waves, specifically as follows: Figure 1 As shown, it includes the following steps: Step 11: Obtain sea surface temperature (SST) time-series observation data for the area to be inverted (internal tidal wave), and extract the time-series features of the internal tidal wave from the SST time-series observation data. Specifically, the SST time-series observation data is processed using an internal tidal wave inversion model with a metric learning architecture to obtain the spatial distribution inversion results of the internal tidal wave. The internal tidal wave inversion model includes a time autoencoder and a mapping module: the time autoencoder extracts time-series features from the SST time-series observation data; the mapping module maps the time-series features to obtain spatial state features.

[0023] Step 12: Obtain sea surface temperature data for the training area and extract spatial state feature samples from the sea surface temperature data.

[0024] The internal tidal wave inversion model also includes a spatial autoencoder module, which includes a spatial encoder. The spatial encoder includes a spatial convolutional layer, an Inception layer, and a fully connected layer connected in sequence. Step 12 includes: Step 121: Perform wavelet analysis on the sea surface temperature data layer by layer and point by point to extract the wavelet coefficients, and perform inverse transformation to obtain the filtered sea surface temperature data; perform empirical orthogonal function analysis on the filtered sea surface temperature data to obtain the feature vector of the spatial distribution state of sea surface temperature.

[0025] Step 122: Extract local features from the feature vector of the spatial distribution of sea surface temperature through a spatial convolutional layer, extract spatial features of different scales from the local features through an Inception layer, and flatten and compress the spatial features of different scales through a fully connected layer of a spatial encoder to obtain spatial state feature samples.

[0026] Specifically, in the training phase of the inland tidal wave inversion model, high-resolution ocean reanalysis data, such as sea surface temperature (SST) data for the training area, are first collected for the target or training region. Wavelet analysis is then performed layer-by-layer and point-by-point on this SST data. For example, the Morlet wavelet function is used to perform continuous wavelet transform on the temperature time series data point-by-point, extracting wavelet coefficients with fluctuation periods around 24 hours and 12 hours for inverse transform to filter out non-inland tidal wave signals. Next, empirical orthogonal function (EOF) analysis is performed on the filtered SST data to construct a feature vector of the SST spatial distribution. The SST spatial distribution samples are added, subtracted, and replaced, and the sensitivity of the feature vector to the samples is analyzed. Redundant samples are removed, and sensitive samples are introduced until the feature vector stabilizes, resulting in a complete sample library of the inland tidal wave spatial state.

[0027] After that, as Figure 2 As shown, this invention employs a deep convolutional network based on the Inception structure to establish a spatial autoencoder, thereby extracting the spatial features of the internal tidal wave. The spatial autoencoder consists of two parts: a spatial encoder and a spatial decoder. The spatial encoder comprises four convolutional layers and two fully connected layers. The first and second convolutional layers are both 3×3 spatial convolutions, extracting local detail features from the samples. The subsequent two convolutional layers both use the Inception module, extracting multi-scale features through parallel convolutions of different sizes and concatenating them along the channel dimension to enhance feature diversity. Furthermore, each convolutional layer is followed by a 1×1 convolutional kernel for channel dimensionality reduction, and a residual connection module is introduced to avoid gradient vanishing or exploding. Finally, the two fully connected layers of the spatial encoder process the data to obtain the feature representation of the internal tidal wave spatial state, i.e., the spatial state features.

[0028] The spatial decoder's structure is symmetrical to the spatial encoder, representing the inverse process of the spatial encoder. It is used to upsample and reconstruct the spatial distribution of internal tidal waves after the time-series features are mapped to the corresponding spatial state features. For example, when dealing with sea surface temperature (SST) time-series observation data for an internal tidal wave region to be inverted, the time encoder in the time autoencoder generates temporal features, the mapping model maps these temporal features to spatial features, and the spatial decoder decodes these spatial features into physical space, such as obtaining the spatial distribution of SST. During model training, the spatial encoder extracts features from the SST spatial distribution training data and works with the spatial decoder to optimize parameters. Furthermore, the spatial decoder adds a 1×1 convolutional layer before the final output to reconstruct the features into the original data space.

[0029] For example, a spatial encoder, such as Figure 2As shown in (a), the spatial data is processed sequentially through the first convolutional layer (Conv1), the normalization layer (Norm), the second convolutional layer (Conv2), and the normalization layer (Norm). Then, it is added element-wise with the original spatial data and input into the first multi-branch convolutional layer (Inception1), which is the first convolutional layer of the Inception module. Subsequently, it is processed sequentially through the normalization layer (Norm), the second multi-branch convolutional layer (Inception2), and the normalization layer (Norm). It is then added element-wise with the features before input to the first multi-branch convolutional layer. The result is processed sequentially through the first fully connected layer (FC1), the normalization layer (Norm), and the second fully connected layer (FC2) to obtain the spatial feature (SpaceFeature).

[0030] And spatial decoders, such as Figure 2 As shown in (b), the reverse process of the spatial encoder is as follows: spatial features are processed sequentially through the first fully connected layer (FC1), the normalization layer (Norm), the second fully connected layer (FC2), and the normalization layer (Norm). The processed result is processed by the first encoding module (Inceptrans1) that integrates Inception and Transformer, and then processed by the normalization layer (Norm), the second encoding module (Inceptrans2), and the normalization layer (Norm). The result is added to the features before the input to the first encoding module. The result is processed sequentially through the first transposed convolutional layer (Convtrans1), the normalization layer (Norm), the second transposed convolutional layer (Convtrans2), and the normalization layer (Norm). The result is added to the features before the transposed convolution, and then processed by the output convolutional layer (Convout) to obtain the spatial reconstruction result (SpaceRecon).

[0031] Among them, the convolutional layer of the spatial autoencoder is as follows: Figure 2 As shown in (c), the convolutional layer (Conv) can be designed to first process the data using a 3×3 convolutional kernel with p=1 and stride s=1, followed by a 1×1 convolutional kernel with p=0 and stride s=2. The Inception module integrated in the multi-branch convolutional layer is as follows... Figure 2 As shown in (d), the input features are processed in parallel through three branches: a 1×1 convolutional kernel with p=0 and stride s=1, a 1×1 convolutional kernel with p=0 and stride s=1 combined with a 3×3 convolutional kernel with p=1 and stride s=1, and a 1×1 convolutional kernel with p=0 and stride s=1 combined with a 5×5 convolutional kernel with p=2 and stride s=1. The results of each branch are concatenated to obtain the output of the Inception module.

[0032] Furthermore, during the model training phase, it is necessary to compress the number of encoded features as much as possible while maintaining reconstruction accuracy, thereby achieving coefficient representation of spatial features, for example, using... The parameters are used as a measurement standard, as shown in formula (1): (1) in, For the i-th sample, the temperature value or sea surface temperature data value. This is a reconstructed value of the temperature or sea surface temperature for the sample. The value is the average of all samples. Through hyperparameter screening experiments, while ensuring... Under conditions where the number of features output by the encoder is greater than 90%, the number of features is compressed to the minimum.

[0033] Step 13: Obtain the sea surface temperature time series observation data of the training area, and extract time series feature samples from the sea surface temperature time series observation data.

[0034] The temporal autoencoder includes a temporal encoder, which includes a convolutional layer and a fully connected layer. Step 13 includes: Step 131: Extract local time series features from sea surface temperature time series observation data through the convolutional layer of the time encoder.

[0035] Step 132: Map the local features of the time series to the feature space through the fully connected layer of the time encoder, and compress them to obtain the time series feature samples.

[0036] Specifically, considering the large amount of redundant information contained in time series observation data, this invention uses a time autoencoder to filter out redundant information and extract the effective information from the time series data. The time autoencoder consists of a time encoder and a time decoder, with the time decoder being symmetrical to the time encoder and representing the inverse process of the time encoder. For example, during the model training phase, the time decoder also optimizes parameters by working in conjunction with the time encoder. Specifically, such as... Figure 3 As shown, the temporal encoder includes three convolutional layers and one fully connected layer. Each of the three convolutional layers consists of a combination of a 1×3 convolution and a 1×1 convolution. Its function is to extract local features of the time series and adjust the dimensions. Finally, the temporal features are mapped to a high-dimensional feature space through a fully connected layer. In addition, a residual connection module is introduced in each layer of the encoder to alleviate the gradient vanishing problem.

[0037] For example, such as Figure 3As shown in (a), in the time encoder, time data is first processed by the first convolutional layer (Conv1) and the normalization layer (Norm) in sequence, and then added to the features before convolution. Thus, the first convolutional layer (Conv1), the second convolutional layer (Conv2), and the third convolutional layer (Conv3) work together with the normalization layer (Norm) to complete three convolutional processes. After that, the time feature is obtained by processing through the first fully connected layer (FC1).

[0038] like Figure 3 As shown in (b), the temporal decoder is the reverse process of the temporal encoder. For the temporal feature, it is processed sequentially through a first fully connected layer (FC1) and a normalization layer (Norm), then sequentially through a first transposed convolutional layer (Convtrans1) and a normalization layer (Norm), and added to the feature before the transposed convolution. This process is repeated three times, and finally processed through a second fully connected layer (FC2) to obtain the temporal data. The convolutional layers of the temporal autoencoder are as follows: Figure 3 As shown in (c), the convolutional layer (Conv) can be designed to first be processed by a 1×3 convolutional kernel with p=1 and stride s=1, and then processed by a 1×1 convolutional kernel with p=0 and stride s=2.

[0039] Furthermore, similar to the model training and parameter tuning of spatial autoencoders, temporal autoencoders also employ similar methods to ensure that effective features are fully extracted. Parameters are used as a standard to measure accuracy, and hyperparameter screening experiments are conducted to ensure accuracy. Under conditions where the number of features output by the encoder is greater than 90%, the number of features is compressed to the minimum.

[0040] Step 14: Based on metric learning, spatial state features and time series features are associated and matched to establish spatiotemporal feature correlation. Matched spatial state features and time series features are used as positive samples, and unmatched spatial state features and time series features are used as negative samples. The optimization objective is to minimize the Euclidean distance between spatial state features and time series features in the positive samples and maximize the Euclidean distance between spatial state features and time series features in the negative samples. The parameters of the untrained intidal wave inversion model are optimized to obtain an intidal wave inversion model with spatiotemporal feature correlation.

[0041] Step 14 includes: Step 141: Substitute the Euclidean distance between the spatial state features and time series features in the positive and negative samples into the contrastive loss function to obtain the contrastive loss value.

[0042] Step 142: Using the contrastive loss value as the optimization objective, iteratively update the temporal encoder and spatial encoder through backpropagation and batch training.

[0043] Specifically, the spatial state features and time-series features extracted by spatial and temporal autoencoders are not necessarily dependent on each other. To achieve the inversion from time series to spatial distribution state, the features need to be filtered and replaced during the model training phase, extracting features with temporal-spatial dependencies and minimizing interference from irrelevant features. A specific internal tidal wave inversion model with spatiotemporal correlation matching capabilities is shown below. Figure 4 As shown, a fully connected layer is added after the spatial encoder to compress the spatial features to the same dimension as the temporal features, and the matching relationship between the temporal and spatial features is measured by a contrastive loss function.

[0044] For example, for Time data Processed by a time encoder, the result is... Time characteristics , and for Spatial data positive samples and spatial data negative samples All were processed by a spatial encoder to obtain... Positive sample space features and negative sample space features Then, the positive sample space features and negative sample space features All are processed through a fully connected layer (FC) to obtain Dimensional positive sample space features and negative sample space features Then, the time characteristics... With positive sample space features and negative sample space features Perform contrastive loss calculation; upon confirming a match, [the following will be done]. Time characteristics As a time embedding vector, Positive sample space features As a spatial embedding vector.

[0045] The specific process of spatiotemporal feature association matching using metric learning networks is as follows: Figure 5As shown, the temporal and spatial samples in the sample space are mapped one-to-one. Sample pairs that match in time and space are positive samples, and those that do not match are negative samples. Euclidean distance is used to measure the similarity between the temporal and spatial features of the sample pairs. A contrastive loss function is used to establish the contrast relationship between positive and negative sample pairs. Through batch training and repeated sampling and pairing, the encoder parameters are optimized, reducing the Euclidean distance between the temporal and spatial feature vectors of positive sample pairs and increasing the distance between the temporal and spatial feature vectors of negative sample pairs. This achieves the extraction of temporally and spatially related features and the replacement of unrelated features. Specific loss functions are shown in formulas (2), (3), (4), (5), and (6):

[0046] (2) (3) (4) (5) (6) in, The spatial sea surface temperature obtained by decoding using spatial coding features at time i. Let i be the time series at time i. Let i be the actual space sea temperature at time i. Let be the spatial sea surface temperature at time i, and n be the total number of sample pairs. For the threshold, This represents the positive and negative sample loss ratio. The parameters of the temporal and spatial autoencoders are updated using this metric, ensuring that the generated temporal and spatial features are dependent on each other. This guarantees the reliability of the mapping from temporal to spatial features.

[0047] Step 15: Based on the spatiotemporal feature correlation, the time series features are mapped to obtain the spatial state features of the internal tidal wave to be inverted. The spatiotemporal feature correlation is determined in advance through metric learning, with the optimization objective being to minimize the Euclidean distance between the matched spatial state features and the time series features in the samples, and maximize the Euclidean distance between the unmatched spatial state features and the time series features.

[0048] The mapping module includes a first fully connected layer, a Transformer encoding module, and a second fully connected layer connected in sequence. Step 15 includes: Step 151: Perform linear transformation and nonlinear activation on the time series features through the first fully connected layer of the mapping module to map the time series features into the feature space, thereby obtaining the time feature vector located in the feature space.

[0049] Step 152: Calculate the dependencies of temporal feature vectors using the Transformer encoding module based on the multi-head attention mechanism to obtain the feature vectors weighted by the association relationships.

[0050] Step 153: Adjust the dimensions of the weighted feature vectors of the association relationship through the second fully connected layer of the mapping module to obtain spatial state features that match the dimensions of the time series features.

[0051] Specifically, the above processing establishes a matching relationship between temporal and spatial features. After optimizing the parameters of the corresponding spatial and temporal autoencoders, considering that spatiotemporal features are not directly equal, it is necessary to further establish a mapping relationship between temporal and spatial features, thereby achieving the inversion from a time series to a spatial distribution state. The specific inversion process is as follows: Figure 6 As shown, the mapping module consists of two fully connected modules and a Transformer encoding module. A multi-head attention module is added to the Transformer module to enhance the weight ratio between the main associated features.

[0052] For example, the temporal embedding vector output by the temporal decoder is processed by a FeatureMapNet subnetwork consisting of two fully connected modules and a Transformer encoding module. Features of time mapping to space And processed by the spatial decoder to obtain Spatial prediction results In the Transformer encoding module, the input vector is processed based on a multi-head attention mechanism, and then processed through residual linking and normalization (Add&Norm). After that, it is processed through a feedforward network, and then processed through residual linking and normalization (Add&Norm) again. This process is repeated three times before the output is given.

[0053] Step 16: Reconstruct the spatial distribution of the internal tidal wave to be inverted based on its spatial state characteristics to obtain the spatial distribution inversion result of the internal tidal wave to be inverted.

[0054] In summary, the spatial state inversion method for internal tidal waves provided by this invention uses a triplet metric learning approach to express the time series characteristics and spatial distribution characteristics of seawater temperature under internal tidal wave conditions in a special special space, thereby realizing the cross-latitudinal mapping of seawater temperature time series and spatial distribution caused by internal tidal waves.

[0055] On the one hand, the internal tidal wave spatial state inversion method provided by this invention enables the perception of internal tidal wave spatial state information based on single-point observation data. Although the general distribution of internal tidal wave spatial state can be obtained through historical data, the spatial state of internal tidal waves exhibits a certain time-varying nature due to short-term changes in seawater temperature structure. Therefore, it is impossible to perceive the short-term state of internal tidal wave spatial distribution through point observation using existing techniques. This invention establishes a mapping relationship between temporal features and spatial features, thereby enabling the perception of internal tidal wave spatial state based on single-point observation data.

[0056] On the other hand, the internal tidal wave spatial state inversion method provided by this invention enables the effective utilization of high-frequency single-point time-series observation data. The scarcity of in-situ oceanographic data has always been a key issue limiting the development of marine science, but a large amount of high-frequency single-point time-series observation data is considered redundant and is largely discarded. This invention, by establishing a mapping relationship between temporal and spatial features, can effectively utilize this previously discarded "redundant" data to infer the surrounding spatial state, thus compensating to some extent for the lack of spatial observation information.

[0057] Furthermore, the internal tidal wave spatial state inversion method provided by this invention was tested in a 2°×2° sea area. The model training data used was 0.1°×0.1° sea surface temperature reanalysis data, and the time series observation frequency was once per hour. An example of the inversion of sea surface temperature distribution caused by internal tidal waves is shown below. Figure 7 As shown, the left column represents the target field given by the reanalysis data, the middle column represents the temperature distribution obtained from the inversion based on single-point time-series data, and the right column represents the inversion error. Inversion experiments were conducted in winter and summer, with the observation point used for inversion located at the center of the region. The time series used for inversion was 120 hours long. It can be seen that the morphology, phase, and propagation direction of the sea surface temperature spatial distribution caused by the internal tidal wave were accurately inverted.

[0058] Example 2 The present invention also provides an internal tidal wave spatial state inversion system, comprising: The time series feature extraction module is used to acquire the sea surface temperature time series observation data of the inner tidal wave region to be inverted, and extract the time series features of the inner tidal wave to be inverted from the sea surface temperature time series observation data. The spatial state feature mapping module is used to map the time series features to obtain the spatial state features of the internal tidal wave to be inverted based on the spatiotemporal feature correlation. The spatiotemporal feature correlation is determined in advance through metric learning, with the optimization objective being to minimize the Euclidean distance between the matched spatial state features and the time series features in the sample, and maximize the Euclidean distance between the unmatched spatial state features and the time series features. The spatial distribution inversion module is used to reconstruct the spatial distribution of the internal tidal wave to be inverted based on its spatial state characteristics, and to obtain the spatial distribution inversion result of the internal tidal wave to be inverted.

[0059] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of an internal tidal wave spatial state inversion method. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0060] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions, on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of an internal tidal wave spatial state inversion method. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0061] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for inverting the spatial state of internal tidal waves, characterized in that, include: Acquire sea surface temperature time series observation data of the region to be inverted for internal tidal wave, and extract the time series features of the internal tidal wave to be inverted from the sea surface temperature time series observation data; The spatial state features of the internal tidal wave to be inverted are obtained by mapping the time series features based on the spatiotemporal feature correlation. The spatiotemporal feature correlation is determined in advance through metric learning, with the optimization objective being to minimize the Euclidean distance between the matched spatial state features and the time series features in the sample, and maximize the Euclidean distance between the unmatched spatial state features and the time series features. The spatial distribution of the internal tidal wave to be inverted is reconstructed based on the spatial state characteristics of the internal tidal wave to be inverted, and the spatial distribution inversion result of the internal tidal wave to be inverted is obtained.

2. The method for inverting the spatial state of internal tidal waves according to claim 1, characterized in that, The sea surface temperature time series observation data are processed by an internal tidal wave inversion model based on a metric learning architecture to obtain the spatial distribution inversion results of the internal tidal waves to be inverted; the internal tidal wave inversion model includes a time autoencoder and a mapping module: The time series features are extracted from the sea surface temperature time series observation data using a time autoencoder; The spatial state features are obtained by mapping the time series features through the mapping module.

3. The method for inverting the spatial state of internal tidal waves according to claim 2, characterized in that, Before mapping the time series features to obtain the spatial state features of the internal tidal wave to be inverted based on the pre-trained spatiotemporal feature correlation, the method further includes: Acquire sea surface temperature data of the training area, and extract spatial state feature samples from the sea surface temperature data; Obtain sea surface temperature time series observation data of the training area, and extract time series feature samples from the sea surface temperature time series observation data; Based on metric learning, the spatial state features and time series features are associated and matched to establish spatiotemporal feature correlations. The matched spatial state features and time series features are used as positive samples, and the unmatched spatial state features and time series features are used as negative samples. The optimization objective is to minimize the Euclidean distance between the spatial state features and time series features in the positive samples and maximize the Euclidean distance between the spatial state features and time series features in the negative samples. The parameters of the untrained internal tidal wave inversion model are optimized to obtain the internal tidal wave inversion model with the spatiotemporal feature correlation.

4. The method for inverting the spatial state of internal tidal waves according to claim 3, characterized in that, The internal tidal wave inversion model further includes a spatial autoencoder module, which comprises a spatial encoder. The spatial encoder includes a spatial convolutional layer, an Inception layer, and a fully connected layer connected sequentially. The spatial state feature samples extracted from the sea surface temperature data include: Wavelet analysis is performed on the sea surface temperature data layer by layer and point by point to extract wavelet coefficients, and inverse transformation is performed to obtain filtered sea surface temperature data; empirical orthogonal function analysis is performed on the filtered sea surface temperature data to obtain the feature vector of the spatial distribution state of sea surface temperature. The spatial convolutional layer extracts local features from the feature vector of the spatial distribution of sea surface temperature, the Inception layer extracts spatial features of different scales from the local features, and the fully connected layer of the spatial encoder flattens and compresses the spatial features of different scales to obtain the spatial state feature samples.

5. The method for inverting the spatial state of internal tidal waves according to claim 4, characterized in that, The time autoencoder includes a time encoder, which comprises a convolutional layer and a fully connected layer. Extracting time-series feature samples from the sea surface temperature time-series observation data includes: The time series local features are extracted from the sea surface temperature time series observation data through the convolutional layer of the time encoder; The time series local features are mapped to the feature space through the fully connected layer of the time encoder, and then compressed to obtain the time series feature samples.

6. The method for inverting the spatial state of internal tidal waves according to claim 5, characterized in that, The parameter optimization of the untrained internal tidal wave inversion model, with the optimization objective being to minimize the Euclidean distance between the spatial state features and time series features in the positive samples and maximize the Euclidean distance between the spatial state features and time series features in the negative samples, includes: The Euclidean distance between the spatial state features and time series features in the positive and negative samples is substituted into the contrastive loss function to obtain the contrastive loss value. Using the contrastive loss value as the optimization objective, the temporal encoder and spatial encoder are iteratively updated through backpropagation and batch training.

7. The method for inverting the spatial state of internal tidal waves according to claim 4, characterized in that, The mapping module comprises a first fully connected layer, a Transformer encoding module, and a second fully connected layer connected in sequence. Based on pre-trained spatiotemporal feature correlations, it maps the time series features to obtain the spatial state features of the internal tidal wave to be inverted, including: The time series features are linearly transformed and nonlinearly activated by the first fully connected layer of the mapping module, and the time series features are mapped to the feature space to obtain the time feature vector located in the feature space. The Transformer encoding module calculates the dependencies of the temporal feature vectors based on a multi-head attention mechanism to obtain a feature vector weighted by the association relationships. The second fully connected layer of the mapping module adjusts the dimension of the weighted feature vector of the association relationship to obtain the spatial state feature that matches the dimension of the time series feature.

8. An internal tidal wave spatial state inversion system, characterized in that, include: The time series feature extraction module is used to acquire the sea surface temperature time series observation data of the inner tidal wave region to be inverted, and extract the time series features of the inner tidal wave to be inverted from the sea surface temperature time series observation data. The spatial state feature mapping module is used to map the time series features to obtain the spatial state features of the internal tidal wave to be inverted based on the spatiotemporal feature correlation. The spatiotemporal feature correlation is determined in advance through metric learning, with the optimization objective being to minimize the Euclidean distance between the matched spatial state features and the time series features in the sample, and maximize the Euclidean distance between the unmatched spatial state features and the time series features. The spatial distribution inversion module is used to reconstruct the spatial distribution of the tidal wave to be inverted based on the spatial state characteristics of the tidal wave to be inverted, and to obtain the spatial distribution inversion result of the tidal wave to be inverted.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the internal tidal space state inversion method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to execute the steps of the internal tidal space state inversion method according to any one of claims 1 to 7.

Images