Multi-feature fusion-based multi-element marine observation data prediction method

Abstract

The application discloses a kind of based on multi-feature fusion multi-element marine observation data prediction method, belong to marine observation data intelligent processing technical field.Method includes: collection multi-element marine observation data and pre-processing, create analysis dataset;Extract data features, time series features and frequency domain features;Respectively using time convolution neural network, long short-term memory network and convolution neural network process three kinds of features, obtain corresponding feature vector;Attention mechanism is used to determine the attention score of each feature vector, and the fusion feature is obtained by fusion;Using auto-encoder to analyze and process the fusion feature;Output the prediction result of multi-element marine observation data.The multi-feature fusion based multi-element marine observation data prediction method provided by the application comprehensively considers multi-dimensional features, improves the accuracy and reliability of marine observation data prediction, and has good application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent processing of marine observation data, and particularly relates to a multi-element marine observation data prediction method based on multi-feature fusion. BACKGROUND

[0002] With the intensification of global climate change, the uncertainty of marine environmental change is increasingly significant, and at the same time, marine development activities continue to diversify, which makes efficient sea state information prediction a key technology to ensure maritime safety and support scientific exploration. Whether it is marine scientific research, ecological monitoring and other activities, it needs to rely on sea state prediction to plan observation equipment deployment and data collection window, and thus puts forward higher requirements for the fine prediction of parameters such as sea waves, wind speed, ocean currents, etc. Therefore, sea state prediction research is not only the intersection of marine science and meteorology, but also an important technical support to ensure maritime safety, promote economic development and cope with climate challenges.

[0003] From the perspective of technology development, marine prediction has achieved a leap from empirical prediction to numerical prediction, and numerical simulation, data assimilation and other technical means have been widely used in sea state prediction, which can predict the spatial distribution and temporal evolution of parameters such as sea waves, wind speed, water temperature and ocean currents. With the development of deep learning models and the improvement of computing power, marine prediction is gradually moving towards intelligence.

[0004] Currently, how to combine marine observation data analysis with artificial intelligence algorithms in depth to improve the accuracy and stability of observation data prediction has become a research focus. This direction has important significance in terms of early warning of marine disasters, ensuring safety of maritime operations and promoting innovation of marine prediction technology. SUMMARY

[0005] In view of the deficiencies in the related art, the purpose of the present application is to provide a multi-element marine observation data prediction method based on multi-feature fusion to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A multi-element marine observation data prediction method based on multi-feature fusion, comprising the following steps:

[0008] S1, collecting multi-element marine observation data and preprocessing to create an analysis data set;

[0009] S2, extracting data features, time sequence features and frequency domain features of the multi-element marine observation data;

[0010] S3, the time convolutional neural network, the long short-term memory network and the convolutional neural network are used to process data features, time sequence features and frequency domain features respectively to obtain data feature vectors, time sequence feature vectors and frequency domain feature vectors;

[0011] S4, the attention mechanism is used to fuse and process the data feature vectors, the time sequence feature vectors and the frequency domain feature vectors to determine attention scores of the data feature vectors, the time sequence feature vectors and the frequency domain feature vectors, and obtain fused features;

[0012] S5, the fused features are analyzed and processed by using the autoencoder;

[0013] S6, a prediction result of the multi-element marine observation data is output.

[0014] In some embodiments, step S1 specifically comprises:

[0015] S11, abnormal values in the multi-element marine observation data set are detected and removed, and the data is normalized to obtain a normalized data set X ;

[0016] S12, the data set is divided into a training set and a test set in a predetermined proportion ;

[0017] S13, the training set and the test set are respectively sliced by using a sliding window to obtain training data, test data and corresponding labels.

[0018] In some embodiments, in step S2, the data features are extracted by a graph neural network: the graph neural network comprises at least two nodes and at least one edge connected between the two nodes, each node represents an observation element, and each edge represents the association between the two nodes connected by the edge, a plurality of graph networks are constructed at the same time interval to form a data association model of the multi-element marine observation data set, and the output of the data association model is taken as the data features.

[0019] In some embodiments, in step S2, the time sequence features are extracted by an attention mechanism graph network: the attention mechanism graph network comprises at least two nodes and at least one edge connected between the two nodes, each node represents the multi-element observation data corresponding to each time point, and each edge corresponds to an attention score, the attention score represents the correlation degree of the two nodes connected by the edge, a plurality of attention mechanism graph networks are constructed to form a time sequence association model of the multi-element marine observation data set, and the output of the time sequence association model is taken as the time sequence features.

[0020] ​​​In some embodiments, in step S2, the frequency domain features are extracted by wavelet transform: using Morlet wavelet as the base function, performing continuous wavelet transform on the multi-element marine observation data, obtaining the energy spectrum density by using the wavelet transform coefficients, and taking the energy spectrum density of the multi-element marine observation data as the frequency domain features.

[0021] In some embodiments, in step S3, the time convolutional neural network for processing the data features comprises a residual block and a first fully connected layer; the residual block comprises an expanded causal convolutional layer, a weight normalization layer, an activation function layer, a random inactivation layer and a 1*1 convolutional layer; the input data features are processed by the expanded causal convolutional layer, the weight normalization layer, the activation function layer and the random inactivation layer to form a first processing result; at the same time, the input data features are processed by the 1*1 convolutional layer to form a second processing result; then the first processing result and the second processing result are fused by a residual connection to obtain a residual block output result; the first fully connected layer performs feature mapping on the residual block output result to finally obtain a data feature vector.

[0022] In some embodiments, in step S3, the long short-term memory network for processing the time series features comprises a long short-term memory network layer and a second fully connected layer; the input time series features are processed by the input gate, the forget gate and the output gate of the long short-term memory network layer to obtain an output result; the second fully connected layer performs feature mapping on the output result of the long short-term memory network layer to finally obtain a time series feature vector.

[0023] In some embodiments, in step S3, the convolutional neural network for processing the frequency domain features comprises a convolutional layer, a pooling layer and a third fully connected layer; the input frequency domain features are processed by the convolutional layer to extract local features by weight sharing and local connection, are processed by the pooling layer to reduce the feature size and the parameter amount, and are finally mapped by the third fully connected layer to obtain a frequency domain feature vector with a required size.

[0024] In some embodiments, in step S5, the autoencoder comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer; the process from the input layer to the second hidden layer is an encoding process, which maps the input fusion feature data to a hidden space; the process from the second hidden layer to the output layer is a decoding process, which outputs prediction data with the same dimension as the data of the input layer.

[0025] In some embodiments, in step S5, a mean square error function is used as a loss function to constrain the optimization of the model parameters, so as to reduce the difference between the prediction data and the real data.

[0026] Compared with the prior art, the present application has the following advantages:

[0027] 1. The multi-feature fusion based multi-element marine observation data prediction method provided by the present application, in the aspect of feature extraction, through a graph neural network, multi-element correlation is mined, an attention mechanism is used to analyze time sequence dependence, and a wavelet transform is used to capture frequency domain characteristics, so that data characteristics are comprehensively extracted from multiple dimensions, the limitations of single characteristics in information description are made up, and sufficient information support is provided for accurate prediction.

[0028] 2. The multi-feature fusion based multi-element marine observation data prediction method provided by the present application, in the aspect of feature analysis, a special network is designed for different features: a time convolutional neural network is used to process data characteristics, a long short-term memory network is used to analyze time sequence characteristics, and a convolutional neural network is used to analyze frequency domain characteristics, so that the feature analysis accuracy is improved in a targeted manner, and the accuracy of data prediction is improved.

[0029] 3. In the multi-feature fusion based multi-element marine observation data prediction method provided by the present application, the attention mechanism is used for feature fusion, the attention scores of each feature vector are allocated, the weights are determined according to the contribution degrees of the feature vectors to the prediction task, the feature is reasonably fused, the disadvantages of fixed weights are avoided, and the stability of the prediction model is effectively enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0030] The drawings described herein are used to provide further understanding of the present application, constitute a part of the present application, and the illustrative embodiments of the present application and the descriptions thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0031] Figure 1 The method flowchart of one embodiment of the multi-feature fusion based multi-element marine observation data prediction method of the present application;

[0032] Figure 2 The data feature extraction module framework schematic diagram of one embodiment of the multi-feature fusion based multi-element marine observation data prediction method of the present application;

[0033] Figure 3 The time sequence feature extraction module framework schematic diagram based on the attention mechanism of one embodiment of the multi-feature fusion based multi-element marine observation data prediction method of the present application;

[0034] Figure 4 The residual block structure schematic diagram of the time convolutional neural network of one embodiment of the multi-feature fusion based multi-element marine observation data prediction method of the present application;

[0035] Figure 5 The long short-term memory network structure schematic diagram of one embodiment of the multi-feature fusion based multi-element marine observation data prediction method of the present application;

[0036] Figure 6A convolutional neural network-based frequency domain feature analysis network structure schematic diagram of an embodiment of the multi-feature fusion-based multi-element marine observation data prediction method of the present application is shown in the figure.

[0037] Figure 7 An auto-encoder network structure schematic diagram of an embodiment of the multi-feature fusion-based multi-element marine observation data prediction method of the present application is shown in the figure. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0039] In the description of the present application, it should be understood that the terms “center”, “transverse”, “longitudinal”, “upper”, “lower”, “front”, “rear”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inner”, “outer” and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0040] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms “mounting”, “connection”, “connection” should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0041] Referring to the drawings Figures 1 to 7 An illustrative embodiment of the multi-feature fusion-based multi-element marine observation data prediction method proposed in the present application is shown in the figure, which is a multi-element marine observation data prediction method taking into account the data association characteristics, time sequence correlation characteristics and frequency domain characteristics. The extracted three features are used to realize effective prediction of data, and improve the accuracy and robustness of marine observation data prediction.

[0042] The multi-feature fusion-based multi-element marine observation data prediction method includes the following steps:

[0043] S1, collect multi-element marine observation data and pre-process to create an analysis data set;

[0044] S2, extract data features, time sequence features and frequency domain features of the multi-element marine observation data;

[0045] S3, respectively process the data features, time sequence features and frequency domain features by using a time convolution neural network, a long short-term memory network and a convolution neural network to obtain data feature vectors, time sequence feature vectors and frequency domain feature vectors;

[0046] S4, fuse and process the data feature vectors, time sequence feature vectors and frequency domain feature vectors by using an attention mechanism to determine attention scores of the data feature vectors, time sequence feature vectors and frequency domain feature vectors and obtain fused features;

[0047] S5, analyze and process the fused features by using an autoencoder;

[0048] S6, output a prediction result of the multi-element marine observation data.

[0049] In step S1, the collected multi-element marine observation data is from marine observation equipment such as buoys and satellite remote sensing data, and includes multiple observation elements such as temperature, wind speed and wave height.

[0050] Step S1 specifically includes:

[0051] S11, detect and remove outliers in the multi-element marine observation data set X , and normalize the data to obtain a normalized data set to reduce the dependence of the prediction model on the input scale and avoid the problem of gradient explosion;

[0052] S12, divide the data set into a training set and a test set in a predetermined ratio;

[0053] S13, use a sliding window to slice the training set and the test set respectively to obtain training data, test data and corresponding labels.

[0054] Wherein, the multi-element marine observation data set , n is the length of the data set, m is the number of observation elements.

[0055] In this embodiment, in step S12, the data set is divided into a training set and a test set in a ratio of 6:4. The training set is applied in the data prediction model training and parameter optimization process, and the test set Used to test the performance of the prediction model. In step S13, a sliding window of length 64 is used to test the training set. and test set By slicing the data into individual segments, training data can be obtained. With test data Using a sliding window of length 1 on the training set and test set Process and obtain training labels With test label .

[0056] In step S2, data features are extracted using a graph neural network: the graph neural network includes at least two nodes and at least one edge connecting the two nodes. Each node represents an observation, and each edge represents the correlation between the two nodes it connects to, i.e., the correlation between the two observations. Multiple graph networks are constructed at the same time interval to form a data association model for the multi-factor ocean observation dataset. The output of the data association model is used as the data features. .

[0057] See appendix Figure 2 This is a schematic diagram of a data feature extraction module framework based on a graph neural network. Two nodes without an edge connection represent no direct relationship and no direct information transfer. Because the relationships between different observed elements change over time, the system constructs features at equal time intervals. k Individual graph networks, the model itself allows for changes in the graph structure, thus effectively capturing different time points. The node relationships are described below. In this embodiment, the Pearson correlation coefficient is used to measure the correlation between nodes, and its specific calculation method is as follows:

[0058]

[0059] in, and Representing two types of observations, and This represents the mean of the two observed quantities.

[0060] Based on this, a graph neural network structure was built to connect the observed data, and the data features were obtained. .

[0061] In step S2, temporal features are extracted using an attention mechanism graph network. This network comprises at least two nodes and at least one edge connecting them. Each node represents the multi-factor observation data at each time point, and each edge corresponds to an attention score, which represents the correlation between the two nodes connected by the edge. Multiple attention mechanism graph networks are constructed to form a temporal correlation model for the multi-factor ocean observation dataset. The output of this model is then used as the temporal features. .

[0062] See appendix Figure 3 This is a schematic diagram of the temporal feature extraction module framework based on the attention mechanism. Each time point is treated as a node in the attention mechanism graph network. The entire model describes the relationships between all nodes, and the features output by the temporal correlation model are... It is expressed as follows:

[0063]

[0064] in, This represents the Sigmoid activation function. Represents a node The number of neighboring nodes, Attention score, used to characterize a node To the node The degree of relevance is calculated using the following method:

[0065]

[0066] in, This represents the algorithm for concatenating two node variables. This represents a vector of learnable parameters, and LeakyReLU is the activation function. This allows us to construct time points... The attention mechanism graph network structure between other time points needs to be built. Individual graph network representation Timing correlation between time points.

[0067] In step S2, frequency domain features are extracted using wavelet transform: Morlet wavelets are used as basis functions to perform continuous wavelet transform on the multi-element ocean observation data; the energy spectral density is obtained using the wavelet transform coefficients; and the energy spectral density of the multi-element ocean observation data is used as the frequency domain feature. .

[0068] Using Morlet wavelets as basis functions, it is defined as follows:

[0069]

[0070] in, The center frequency.

[0071] The Morlet wavelet has a localization property and can accurately capture the transient characteristics in the signal. The Morlet wavelet is used as a base function to perform continuous wavelet transform on the multi-element ocean observation data, and the calculation method is as follows:

[0072]

[0073] wherein, denotes a scale factor, denotes a translation factor, denotes a base function and a complex conjugate.

[0074] The energy spectrum density is obtained by using the wavelet transform coefficient , and the specific calculation method is as follows:

[0075]

[0076] In this embodiment, the energy spectrum density of the multi-element ocean observation data is used as a frequency domain feature for subsequent feature analysis.

[0077] In step S3, the time convolutional neural network for processing the data features includes a residual block and a first full connection layer; the residual block includes a dilated causal convolutional layer, a weight normalization layer, an activation function layer, a dropout layer and a 1*1 convolutional layer; the input data features are processed by the dilated causal convolutional layer, the weight normalization layer, the activation function layer and the dropout layer to form a first path processing result; at the same time, the input data features are processed by the 1*1 convolutional layer to form a second path processing result; then the first path processing result and the second path processing result are fused by a residual connection to obtain a residual block output result; the first full connection layer performs feature mapping on the residual block output result to finally obtain a data feature vector .

[0078] Referring to the accompanying Figure 4In this embodiment, the temporal convolutional neural network model used to process data features includes two residual blocks and one fully connected layer. The activation function layer of the residual blocks is chosen to be the ReLU activation function. The dilated causal convolutional layer has a better understanding of the contextual information of the input data while ensuring the directionality of information flow, making it more suitable for the analysis of time-series data. The weight normalization layer is used to speed up computation and reduce path dependence in gradient descent. The ReLU activation function enhances the nonlinearity of the network, making the network sparsity-oriented. The random deactivation layer is used to prevent overfitting. The residual connections in the residual blocks can alleviate the gradient vanishing problem and promote the training of deeper networks. The 1×1 convolutional layer is used to ensure the consistency between input and output.

[0079] In step S3, the Long Short-Term Memory (LSTM) network used to process temporal features includes an LTM network layer and a second fully connected layer. The input temporal features are processed by the input gate, forget gate, and output gate of the LTM network layer to obtain the output result. The second fully connected layer performs feature mapping on the output result of the LTM network layer to finally obtain the temporal feature vector. .

[0080] In this embodiment, the Long Short-Term Memory (LSTM) network used to process temporal features includes three LSTM network layers and one fully connected layer. Each LSTM layer comprises a memory cell and three gating mechanisms: an input gate, a forget gate, and an output gate. Its structure is as follows: Figure 5 As shown. The Long Short-Term Memory (LSTM) network layer is processed by the LeakyReLU function, and the second fully connected layer is processed by the Tanh function. Memory cells are the core storage units of LSTM, responsible for the transmission of long-term information. The forget gate is based on the current input. And the previous hidden state Determines the state of the previous cell The retention ratio is calculated as follows:

[0081]

[0082] in, This represents the activation value of the forget gate. This represents the Sigmoid activation function. and These represent the weight matrix and bias of the forget gate, respectively.

[0083] Input gate generates candidate memory content And determine the update ratio of new information, the specific calculation method is as follows:

[0084]

[0085]

[0086] wherein, denotes the activation value of the input gate, and denotes the weight matrix, and denotes the bias. A new cell state is obtained based on the above calculation, which is specifically defined as:

[0087]

[0088] The output gate decides which information from the current cell state is output to the hidden state , and the specific calculation method is:

[0089]

[0090]

[0091] wherein, denotes the activation value of the output gate, and are the weight matrix and bias of the output gate, respectively. The LSTM uses the cell state as the core memory carrier, and uses the three gating mechanisms of the forget gate, the input gate and the output gate, which can effectively learn the dependence relationship across long time steps in the time series data.

[0092] In step S3, the convolutional neural network for processing the frequency domain features includes a convolutional layer, a pooling layer and a third fully connected layer; the input frequency domain features are extracted by the convolutional layer through weight sharing and local connection to extract local features, the feature size and parameter quantity are reduced by the pooling layer, and finally the processed features are mapped by the third fully connected layer to obtain the frequency domain feature vector of the required size.

[0093] In this embodiment, the specific structure of the convolutional neural network for processing the frequency domain features includes 3 convolutional layers, 2 pooling layers and 1 fully connected layer, as shown in Figure 6 . The convolution kernels of the 3 convolutional layers are (3, 1), (3, 1) and (1, 1), and the activation function is the ReLU function. The convolutional layer has the characteristics of weight sharing and local connection, which reduces the model parameters while extracting local features. The pooling layer reduces the size of the features, reduces the parameter quantity of the model, and can also prevent overfitting. The third fully connected layer maps the features processed by the convolutional layer and the pooling layer to obtain the frequency domain feature vector of the required size.

[0094] Data feature vector Temporal feature vectors and frequency domain eigenvectors After processing with an attention mechanism, the attention scores of the three feature vectors are determined, resulting in the fused features. .

[0095] In this embodiment, the first fully connected layer of the convolutional neural network, the second fully connected layer of the long short-term memory network, and the third fully connected layer of the convolutional neural network, the first, second, and third before each fully connected layer are only used to distinguish the network structure to which they belong and have no other meaning.

[0096] In step S5, the autoencoder includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. The process from the input layer to the second hidden layer is the encoding process, mapping the input fused feature data to the latent space. The process from the second hidden layer to the output layer is the decoding process, outputting predicted data with the same dimension as the input layer data. The first and third hidden layers are processed by the ReLU activation function, while the second hidden layer and the output layer are processed by the Sigmoid activation function. Furthermore, in step S5, the mean squared error function is used as the loss function to constrain the optimization of the model parameters, thereby reducing the difference between the predicted data and the true data.

[0097] In this embodiment, a 5-layer autoencoder (AE) is used to fuse the features of multi-element ocean observation data. Analysis shows that the network structure of the autoencoder is as follows: Figure 7 As shown, the autoencoder network consists of an input layer, three hidden layers, and an output layer. It fuses the input features. The predicted data is obtained after processing by the encoder and decoder. The specific process can be represented as follows:

[0098]

[0099] by Indicates the number of network layers, when hour, , and These represent the first and second parts of the autoencoder. The layer's activation function, weight matrix, and bias. The activation function... The settings are different on each floor. and For ReLU function, and This is the Sigmoid function.

[0100] In the prediction model parameter optimization process, the mean square error (MSE) is used as the loss function, which is defined as follows:

[0101]

[0102] wherein, denotes the length of the predicted data, and denote the real observation data and the predicted data, respectively. After multiple iterations of training, the loss is constantly smaller until convergence, and the prediction model parameters reach the optimal setting. The prediction method can effectively predict multi-element marine observation data.

[0103] In the above exemplary embodiments, the multi-feature fusion based multi-element marine observation data prediction method forms a complete technical system from feature extraction, analysis to fusion through a multi-feature fusion strategy: comprehensive extraction of multi-dimensional data features is achieved by using graph neural networks, attention mechanisms and wavelet transforms, combined with specific network targeted analysis, and then dynamic fusion through the attention mechanism, effectively improving the accuracy and stability of multi-element marine observation data prediction, and being suitable for marine weather prediction, marine disaster warning and marine energy development and other fields.

[0104] Finally, it should be noted that: the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to.

[0105] The above embodiments are only used to illustrate the technical solutions of the present application rather than limiting them. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the specific implementation of the present application can be modified or some technical features can be replaced by equivalent ones without departing from the spirit of the present application. They should all be included in the technical solution range of the present application.

Claims

1. A method for predicting multi-element ocean observation data based on multi-feature fusion, characterized in that, Includes the following steps: S1. Collect multi-element ocean observation data and preprocess it to create an analysis dataset; S2. Extract the data features, temporal features, and frequency domain features of the multi-element ocean observation data; wherein, the data features are extracted using a graph neural network, which includes at least two nodes and at least one edge connecting the two nodes. Each node represents an observation element, and each edge represents the correlation between the two nodes it connects. Multiple graph networks are constructed at the same time interval to form a data association model for the multi-element ocean observation dataset, and the output of the data association model is used as the data features; the temporal features are extracted using an attention mechanism graph network, which includes at least two nodes and at least one edge connecting the two nodes. Each node represents the multi-element observation data corresponding to each time point, and each edge corresponds to an attention score, which represents the correlation between the two nodes connected by the edge. Multiple attention mechanism graph networks are constructed to form a temporal association model for the multi-element ocean observation dataset, and the output of the temporal association model is used as the temporal features; the frequency domain features are extracted using wavelet transform. S3. The data features, temporal features, and frequency domain features are processed by a temporal convolutional neural network, a long short-term memory network, and a convolutional neural network, respectively, to obtain data feature vectors, temporal feature vectors, and frequency domain feature vectors; S4. Use the attention mechanism to fuse the data feature vector, time-series feature vector, and frequency domain feature vector, determine the attention scores of the data feature vector, time-series feature vector, and frequency domain feature vector, and obtain the fused features; S5. Use an autoencoder to analyze and process the fused features; S6. Output the prediction results of multi-element marine observation data.

2. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 1, characterized in that, Step S1 specifically includes: S11. Detect and remove multi-element ocean observation datasets. X The outliers were identified, and the data was normalized to obtain a normalized dataset. ; S12. Divide the dataset according to the preset ratio. Divided into training set and test set ; S13. Use a sliding window to process the training set respectively. and test set The data is sliced ​​to obtain training data, test data, and corresponding labels.

3. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 1, characterized in that, In step S2, extracting the frequency domain features through wavelet transform specifically includes: using Morlet wavelet as the basis function, performing continuous wavelet transform on the multi-element ocean observation data, obtaining the energy spectral density using the wavelet transform coefficients, and using the energy spectral density of the multi-element ocean observation data as the frequency domain feature.

4. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 1, characterized in that, In step S3, the temporal convolutional neural network used to process the data features includes a residual block and a first fully connected layer. The residual block includes a dilated causal convolutional layer, a weight normalization layer, an activation function layer, a random deactivation layer, and a 1×1 convolutional layer. The input data features are processed by the dilated causal convolutional layer, the weight normalization layer, the activation function layer, and the random deactivation layer to form a first processing result. Simultaneously, the input data features are convolved by the 1×1 convolutional layer to form a second processing result. Then, the first processing result and the second processing result are fused through a residual connection to obtain the residual block output result. The first fully connected layer performs feature mapping on the residual block output result to finally obtain the data feature vector.

5. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 1, characterized in that, In step S3, the long short-term memory network used to process the temporal features includes a long short-term memory network layer and a second fully connected layer; the input temporal features are processed by the input gate, forget gate and output gate of the long short-term memory network layer to obtain the output result; the second fully connected layer performs feature mapping on the output result of the long short-term memory network layer to finally obtain the temporal feature vector.

6. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 1, characterized in that, In step S3, the convolutional neural network used to process the frequency domain features includes a convolutional layer, a pooling layer, and a third fully connected layer. The input frequency domain features are extracted by the convolutional layer through weight sharing and local connections, the pooling layer reduces the feature size and number of parameters, and finally the third fully connected layer maps the processed features to obtain a frequency domain feature vector of the required size.

7. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 1, characterized in that, In step S5, the autoencoder includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; the process from the input layer to the second hidden layer is the encoding process, which maps the input fused feature data to the latent space; the process from the second hidden layer to the output layer is the decoding process, which outputs predicted data of the same dimension as the input layer data.

8. The multi-feature fusion-based multi-element ocean observation data prediction method according to claim 7, characterized in that, In step S5, the mean squared error function is used as the loss function to constrain the optimization of the model parameters in order to reduce the difference between the predicted data and the actual data.

Images