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

Abstract

The invention discloses a multi-element marine observation data prediction method based on multi-feature fusion, and belongs to the technical field of marine observation data intelligent processing. The method comprises the following steps: collecting and preprocessing multi-element marine observation data, and creating an analysis data set; extracting data features, time sequence features and frequency domain features; respectively processing the three types of features by adopting a time convolutional neural network, a long-short-term memory network and a convolutional neural network to obtain corresponding feature vectors; determining an attention score of each feature vector by using an attention mechanism, and fusing to obtain a fused feature; a self-encoder is adopted to analyze and process the fusion features; and outputting a prediction result of the multi-element marine observation data. According to the multi-feature fusion-based multi-element ocean observation data prediction method provided by the invention, multi-dimensional features are integrated, the accuracy and reliability of ocean observation data prediction are improved, and the method has a good application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent processing of ocean observation data, and in particular relates to a multi-factor ocean observation data prediction method based on multi-feature fusion. Background Art

[0002] As global climate change intensifies, the uncertainty of marine environmental changes becomes increasingly significant. At the same time, marine development activities continue to diversify, making efficient sea state information forecasting a key technology for ensuring maritime safety and supporting scientific exploration. Whether it's marine scientific research or ecological monitoring, all activities rely on sea state forecasting to plan the deployment of observation equipment and data collection windows, which in turn places higher demands on the refined prediction of parameters such as waves, wind speed, and ocean currents. Therefore, sea state forecasting research is not only a frontier at the intersection of marine science and meteorology, but also a vital technical support for ensuring maritime safety, promoting economic development, and responding to climate challenges.

[0003] From a technological perspective, ocean forecasting has made a significant leap from empirical to numerical forecasting. Numerical simulation and data assimilation are now widely used in ocean forecasting, enabling real-time prediction of the spatial distribution and temporal evolution of parameters such as waves, wind speed, water temperature, and ocean currents. Furthermore, with the development of deep learning models and improvements in computing power, ocean forecasting is gradually becoming intelligent.

[0004] Currently, a key research area is the integration of ocean observation data analysis with artificial intelligence algorithms to improve the accuracy and stability of observation data predictions. This research is crucial for early warning of marine disasters, ensuring safe operations at sea, and promoting innovation in ocean forecasting technologies. Summary of the Invention

[0005] In view of the shortcomings in the related art, the purpose of the present invention is to provide a multi-factor ocean observation data prediction method based on multi-feature fusion to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A multi-factor ocean observation data prediction method based on multi-feature fusion includes the following steps: S1. Collect and preprocess multi-factor ocean observation data to create analysis datasets; S2. Extract data features, time series features, and frequency domain features of multi-factor ocean observation data; S3. Use time convolution neural network, long short-term memory network and convolution neural network to process data features, time series features and frequency domain features respectively to obtain data feature vector, time series feature vector and frequency domain feature vector; S4. Use the attention mechanism to fuse the data feature vector, the time series feature vector, and the frequency domain feature vector, determine the attention scores of the data feature vector, the time series feature vector, and the frequency domain feature vector, and obtain the fused feature; S5. Use autoencoder to analyze and process the fusion features; S6. Output the prediction results of multi-factor ocean observation data.

[0007] In some embodiments, step S1 specifically includes: S11. Detect and eliminate multi-factor ocean observation datasets X The outliers in the data are normalized to obtain the normalized data set ; S12, according to the preset ratio Divide into training set and test set ; S13, use sliding window to analyze the training set and test set Perform slicing processing to obtain training data, test data and corresponding labels.

[0008] In some embodiments, in step S2, data features are extracted through 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 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 of a multi-element ocean observation data set, and the output of the data association model is used as the data feature.

[0009] In some embodiments, in step S2, the temporal features are extracted through an attention mechanism graph network: the attention mechanism graph network includes at least two nodes and at least one edge connecting the two nodes, each node represents the multi-factor observation data corresponding to each time point, each edge corresponds to an attention score, and the attention score represents the degree of correlation between the two nodes connected by the edge, and multiple attention mechanism graph networks are constructed to form a temporal association model of the multi-factor ocean observation data set, and the output of the temporal association model is used as the temporal feature.

[0010] In some embodiments, in step S2, frequency domain features are extracted by wavelet transform: using Morlet wavelet as basis function, continuous wavelet transform is performed on multi-factor ocean observation data, energy spectral density is obtained using wavelet transform coefficients, and the energy spectral density of the multi-factor ocean observation data is used as frequency domain features.

[0011] In some embodiments, in step S3, the temporal convolutional neural network for processing data features includes a residual block and a first fully connected layer; the residual block includes an expanded causal convolution layer, a weight normalization layer, an activation function layer, a random dropout layer and a 1×1 convolution layer; the input data features are processed by the expanded causal convolution layer, the weight normalization layer, the activation function layer and the random dropout layer to form a first-path processing result; at the same time, the input data features are convolved by the 1×1 convolution layer to form a second-path processing result; the first-path processing result and the second-path processing result are then fused through 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 ultimately obtain a data feature vector.

[0012] In some embodiments, in step S3, the long short-term memory network for processing time series features includes a long short-term memory network layer and a second fully connected layer; the input time series features are processed by the input gate, forget gate and 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, and finally obtains a time series feature vector.

[0013] In some embodiments, in step S3, the convolutional neural network for processing frequency domain features includes a convolution layer, a pooling layer, and a third fully connected layer; the input frequency domain features are extracted by the convolution layer through weight sharing and local connection to extract local features, the feature size and parameter amount are reduced by the pooling layer, and finally the processed features are mapped by the third fully connected layer to obtain a frequency domain feature vector of the required size.

[0014] In some embodiments, 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 an 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 a decoding process, which outputs predicted data of the same dimension as the input layer data.

[0015] In some embodiments, in step S5, a mean square error function is used as a loss function to constrain the optimization of model parameters to reduce the difference between the predicted data and the real data.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The multi-factor ocean observation data prediction method based on multi-feature fusion provided by the present invention, in terms of feature extraction, uses graph neural networks to mine multi-factor associations, attention mechanisms to analyze temporal dependencies, and wavelet transforms to capture frequency domain characteristics, comprehensively extracting data features from multiple dimensions, making up for the limitations of single features in describing information, and providing sufficient information support for accurate prediction.

[0017] 2. The multi-factor ocean observation data prediction method based on multi-feature fusion provided by the present invention designs exclusive networks for different features during feature analysis: time convolutional neural networks process data features, long-short-term memory networks analyze time series features, and convolutional neural networks analyze frequency domain features. The targeted design improves the accuracy of feature analysis and thus improves the accuracy of data prediction.

[0018] 3. In the multi-factor ocean observation data prediction method based on multi-feature fusion provided by the present invention, feature fusion adopts the attention mechanism to assign an attention score to each feature vector, determine the weight according to its contribution to the prediction task, realize reasonable feature fusion, avoid the disadvantages of fixed weights, and effectively enhance the stability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 This is a method flow chart of an embodiment of the multi-factor ocean observation data prediction method based on multi-feature fusion of the present invention; Figure 2 This is a schematic diagram of a data feature extraction module framework based on a graph neural network in accordance with an embodiment of a multi-factor ocean observation data prediction method based on multi-feature fusion according to the present invention; Figure 3 This is a schematic diagram of a temporal feature extraction module framework based on an attention mechanism in one embodiment of a multi-factor ocean observation data prediction method based on multi-feature fusion according to the present invention; Figure 4 This is a schematic diagram of the residual block structure of a temporal convolutional neural network according to an embodiment of the multi-factor ocean observation data prediction method based on multi-feature fusion of the present invention; Figure 5 A schematic diagram of the long short-term memory network structure of an embodiment of the multi-factor ocean observation data prediction method based on multi-feature fusion of the present invention; Figure 6 This is a schematic diagram of a frequency domain feature analysis network structure based on a convolutional neural network according to an embodiment of the multi-factor ocean observation data prediction method based on multi-feature fusion of the present invention; Figure 7 This is a schematic diagram of the autoencoder network structure of an embodiment of the multi-factor ocean observation data prediction method based on multi-feature fusion of the present invention. DETAILED DESCRIPTION

[0020] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0021] In the description of the present invention, it should be understood that the terms "center", "transverse", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0022] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood broadly. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to direct connections, indirect connections through an intermediary, or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0023] See attached Figures 1 to 7 , gives an illustrative embodiment of the multi-factor ocean observation data prediction method based on multi-feature fusion proposed in the present invention. This method is a multi-factor ocean observation data prediction method that takes into account data association characteristics, time series correlation characteristics and frequency domain characteristics. It uses the three extracted features to achieve effective data prediction and improve the accuracy and robustness of ocean observation data prediction.

[0024] The multi-factor ocean observation data prediction method based on multi-feature fusion includes the following steps: S1. Collect and preprocess multi-factor ocean observation data to create analysis datasets; S2. Extract data features, time series features, and frequency domain features of multi-factor ocean observation data; S3. Use time convolution neural network, long short-term memory network and convolution neural network to process data features, time series features and frequency domain features respectively to obtain data feature vector, time series feature vector and frequency domain feature vector; S4. Use the attention mechanism to fuse the data feature vector, the time series feature vector, and the frequency domain feature vector, determine the attention scores of the data feature vector, the time series feature vector, and the frequency domain feature vector, and obtain the fused feature; S5. Use autoencoder to analyze and process the fusion features; S6. Output the prediction results of multi-factor ocean observation data.

[0025] In step S1, the sources of the collected multi-factor ocean observation data are ocean observation equipment such as buoys and satellite remote sensing data, including multiple observation factors such as temperature, wind speed, and wave height.

[0026] Step S1 specifically includes: S11. Detect and eliminate multi-factor ocean observation datasets X The outliers in the data are normalized to obtain the normalized data set , to reduce the prediction model's dependence on the input scale and avoid the gradient explosion problem; S12, according to the preset ratio Divide into training set and test set ; S13, use sliding window to analyze the training set and test set Perform slicing processing to obtain training data, test data and corresponding labels.

[0027] Among them, the multi-factor ocean observation dataset , n is the length of the dataset, m is the number of observation elements.

[0028] In this embodiment, in step S12, the data set is divided into Divide into training set and test set . Training set Applied in the process of data prediction model training and parameter optimization, the test set Used to test the performance of the prediction model. In step S13, a sliding window with a length of 64 is used to test the training set. and test set Slice the data separately to obtain training data. With test data ; Use a sliding window of length 1 to train the and test set Process and obtain training labels With test tag .

[0029] In step S2, data features are extracted through a graph neural network: a 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, that is, the correlation between the two observations. Multiple graph networks are constructed at the same time interval to form a data association model of a multi-factor ocean observation dataset, and the output of the data association model is used as the data feature. .

[0030] See attached Figure 2 , is a schematic diagram of the data feature extraction module framework based on graph neural network. In the figure, two nodes without edge connection represent no direct connection and no direct information transmission. As time changes, the connection between different observation elements will change. Constructing at the same time interval k The model itself allows the graph structure to change, thus effectively capturing different time points. In this embodiment, the Pearson correlation coefficient is used to measure the correlation between nodes, and its specific calculation method is as follows:

[0031] in, and Represents two observation quantities, and represents the mean of the two observations.

[0032] On this basis, a graph neural network structure between observation data is built, and data features are obtained .

[0033] In step S2, the temporal features are extracted through an attention mechanism graph network: the attention mechanism graph network includes at least two nodes and at least one edge connecting the two nodes, each node represents the multi-factor observation data corresponding to each time point, each edge corresponds to an attention score, and the attention score represents the degree of correlation between the two nodes connected by the edge. Multiple attention mechanism graph networks are constructed to form a temporal association model of the multi-factor ocean observation dataset, and the output of the temporal association model is used as the temporal feature. .

[0034] See attached Figure 3 , which is a schematic diagram of the temporal feature extraction module framework based on the attention mechanism. Each time point is regarded as a node in the attention mechanism graph network. The entire model describes the relationship between all nodes. The features output by the temporal correlation model It is expressed as follows:

[0035] in, represents the Sigmoid activation function, Representation node The number of adjacent nodes, Represents the attention score, used to characterize the node To Node The specific calculation method is as follows:

[0036] in, represents the concatenation algorithm of two node variables, Represents a vector of learnable parameters, and LeakyReLU is the activation function. This builds the time point The attention mechanism graph network structure between other time points needs to be built Graph Network Representation The temporal correlation between time nodes.

[0037] In step S2, frequency domain features are extracted by wavelet transform: using Morlet wavelet as basis function, continuous wavelet transform is performed on multi-factor ocean observation data, and energy spectrum density is obtained by using wavelet transform coefficients. The energy spectrum density of multi-factor ocean observation data is used as frequency domain features. .

[0038] Taking Morlet wavelet as the basis function, it is defined as:

[0039] in, is the center frequency.

[0040] Morlet wavelet has localized characteristics and can accurately capture transient features in the signal. Using Morlet wavelet as the basis function, continuous wavelet transform is performed on multi-factor ocean observation data. The calculation method is as follows:

[0041] in, represents the scale factor, represents the translation factor, Representation basis function The complex conjugate of .

[0042] Using wavelet transform coefficients to obtain energy spectrum density , the specific calculation method is as follows:

[0043] In this embodiment, the energy spectrum density of multi-factor ocean observation data is used as the frequency domain feature , which is used in subsequent feature analysis.

[0044] In step S3, the temporal convolutional neural network for processing 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 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 random dropout layer to form a first-path processing result; at the same time, the input data features are convolved by the 1×1 convolutional layer to form a second-path processing result; the first-path processing result and the second-path processing result are then fused through a residual connection to obtain a residual block output result; the first fully connected layer performs feature mapping on the residual block output result, and finally obtains a data feature vector .

[0045] See attached Figure 4 In this embodiment, the network model of the temporal convolutional neural network for processing data features includes 2 residual blocks and 1 fully connected layer. Among them, the activation function layer of the residual block is selected as the ReLU activation function (ReLUActivation Function). The dilated causal convolution layer has a better understanding of the contextual information of the input data, while ensuring the directionality of the information flow, and is more suitable for the analysis of time series data; the weight normalization layer is used to speed up the operation and reduce the path dependency in the gradient descent; the ReLU activation function enhances the nonlinearity of the network and makes the network sparse; the random inactivation layer is used to prevent the model from overfitting. The residual connection in the residual block can alleviate the gradient disappearance problem and promote the training of deeper networks, among which the 1×1 convolution layer is used to ensure the consistency of input and output.

[0046] In step S3, the long short-term memory network for processing time series features includes a long short-term memory network layer and a second fully connected layer; the input time series 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, and finally obtains the time series feature vector .

[0047] In this embodiment, the long short-term memory network for processing time series features includes three long short-term memory network layers (LSTM) and one fully connected layer. Among them, the long short-term memory network layer includes memory cells (Cell State) and three gating mechanisms: input gate (Input Gate), forget gate (ForgetGate) and output gate (OutputGate). Its structure is as follows Figure 5As shown in the figure. The long short-term memory network layer is processed by the LeakyReLU function, and the second fully connected layer is processed by the Tanh function. The memory cell is the core storage unit of LSTM, responsible for the transmission of long-term information. The forget gate is based on the current input. and the previous hidden state Determine the previous cell state The retention ratio is calculated as:

[0048] in, represents the activation value of the forget gate, represents the Sigmoid activation function, and Represent the weight matrix and bias of the forget gate respectively.

[0049] Input gate generates candidate memory content , and determines the update ratio of new information. The specific calculation method is:

[0050]

[0051] in, represents the activation value of the input gate, and represents the weight matrix, and Represents the bias. Based on the above calculation, the new cell state is obtained. , which is specifically defined as:

[0052] The output gate determines the current cell state What information is output to the hidden state? The specific calculation method is:

[0053]

[0054] in, represents the activation value of the output gate, and are the weight matrix and bias of the output gate, respectively. LSTM uses the cell state as the core memory carrier and utilizes three gating mechanisms: forget gate, input gate, and output gate. It can effectively learn dependencies in time series data that span long time steps.

[0055] In step S3, the convolutional neural network for processing frequency domain features includes a convolution layer, a pooling layer, and a third fully connected layer; the input frequency domain features are extracted by the convolution layer through weight sharing and local connection, the feature size and parameter amount 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. .

[0056] In this embodiment, the specific structure of the convolutional neural network for processing frequency domain features includes 3 convolutional layers, 2 pooling layers and 1 fully connected layer, such as Figure 6 As shown. The convolution kernels of the three convolutional layers are (3,1), (3,1) and (1,1), and the activation function is the ReLU function. The convolution 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 number of model parameters, and can also prevent overfitting. The third fully connected layer maps the features processed by the convolution layer and the pooling layer to obtain the frequency domain feature vector of the required size. .

[0057] Data feature vector , time series feature vector and frequency domain eigenvectors After the attention mechanism is processed, the attention scores of the three feature vectors are determined to obtain the fusion feature .

[0058] 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.

[0059] 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 into the latent space. The process from the second hidden layer to the output layer is the decoding process, which outputs predicted data with the same dimensions as the input layer data. The first and third hidden layers are activated by the ReLU function, while the second hidden layer and the output layer are activated by the Sigmoid function. In addition, in step S5, the mean square error function is used as the loss function to constrain the optimization of model parameters to minimize the difference between the predicted data and the actual data.

[0060] In this embodiment, a 5-layer autoencoder (AE) is used to fusion features of multi-factor ocean observation data. For analysis, the network structure of the autoencoder is as follows Figure 7As shown. The autoencoder network includes an input layer (InputLayer), three hidden layers (HiddenLayer) and an output layer (OutputLayer). For the input fusion feature , after being processed by the encoder and decoder, the predicted data is obtained , the specific process can be expressed as:

[0061] by Indicates the number of network layers. hour, 、 and Represents the first The activation function, weight matrix and bias of the layer. Among them, the activation function The settings on each layer are different. and is the ReLU function, and is the Sigmoid function.

[0062] In the process of optimizing the prediction model parameters, the mean square error (MSE) is used as the loss function, which is defined as follows:

[0063] in, Indicates the length of the predicted data, and Represent the real observation data and predicted data respectively. After multiple iterations of training, the loss The prediction model parameters are optimized and the prediction method can effectively predict multi-factor ocean observation data.

[0064] In the above-mentioned illustrative embodiment, the multi-factor ocean observation data prediction method based on multi-feature fusion forms a complete technical system from feature extraction, analysis to fusion through a multi-feature fusion strategy: it uses graph neural networks, attention mechanisms and wavelet transforms to comprehensively extract multi-dimensional features of data, combines targeted analysis with exclusive networks, and then dynamically fuses them through attention mechanisms, effectively improving the accuracy and stability of multi-factor ocean observation data prediction, and is suitable for multiple fields such as marine meteorological forecasting, marine disaster warning and marine energy development.

[0065] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0066] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the same. Although the present invention has been described in detail with reference to preferred embodiments, persons skilled in the art should understand that the specific implementation methods of the present invention may still be modified or some technical features may be replaced by equivalents without departing from the spirit of the technical solutions of the present invention, and all of these should fall within the scope of the technical solutions claimed for protection by the present invention.

Claims

1. A multi-factor ocean observation data prediction method based on multi-feature fusion, characterized in that: The steps include: S1. Collect and preprocess multi-factor ocean observation data to create analysis datasets; S2. extracting data features, time series features, and frequency domain features of the multi-factor ocean observation data; S3, using a time convolutional neural network, a long short-term memory network and a convolutional neural network to process the data features, time series features and frequency domain features respectively to obtain a data feature vector, a time series feature vector and a frequency domain feature vector; S4. Use the attention mechanism to fuse the data feature vector, the time series feature vector, and the frequency domain feature vector, determine the attention scores of the data feature vector, the time series feature vector, and the frequency domain feature vector, and obtain the fused feature; S5. Use autoencoder to analyze and process the fusion features; S6. Output the prediction results of multi-factor ocean observation data.

2. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: Step S1 specifically includes: S11. Detect and eliminate multi-factor ocean observation datasets X The outliers in the data are normalized to obtain the normalized data set ; S12, according to the preset ratio Divide into training set and test set ; S13, use sliding window to analyze the training set and test set Perform slicing processing to obtain training data, test data and corresponding labels.

3. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: In step S2, the data features are extracted through a graph neural network: the graph neural network includes at least two nodes and at least one edge connecting the two nodes, each of the nodes represents an observation element, and each of the edges 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 of a multi-element ocean observation data set, and the output of the data association model is used as the data feature.

4. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: In step S2, the temporal features are extracted through an attention mechanism graph network: the attention mechanism graph network includes at least two nodes and at least one edge connecting the two nodes, each of the nodes represents the multi-factor observation data corresponding to each time point, each of the edges corresponds to an attention score, and the attention score represents the degree of correlation between the two nodes connected by the edge. Multiple attention mechanism graph networks are constructed to form a temporal association model of the multi-factor ocean observation data set, and the output of the temporal association model is used as the temporal feature.

5. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: In step S2, the frequency domain features are extracted by wavelet transform: using Morlet wavelet as basis function, continuous wavelet transform is performed on the multi-factor ocean observation data, energy spectrum density is obtained using wavelet transform coefficients, and the energy spectrum density of the multi-factor ocean observation data is used as the frequency domain features.

6. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: In step S3, the temporal convolutional neural network for processing the data features includes a residual block and a first fully connected layer; the residual block includes an expanded causal convolution layer, a weight normalization layer, an activation function layer, a random dropout layer and a 1×1 convolution layer; the input data features are processed by the expanded causal convolution layer, the weight normalization layer, the activation function layer and the random dropout layer to form a first-path processing result; at the same time, the input data features are convolved by the 1×1 convolution layer to form a second-path processing result; then the first-path processing result and the second-path processing result are fused through 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.

7. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: In step S3, the long short-term memory network used to process the time series features includes a long short-term memory network layer and a second fully connected layer; the input time series features are processed by the input gate, forget gate and 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, and finally obtains a time series feature vector.

8. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is characterized in that: In step S3, the convolutional neural network used to process the frequency domain features includes a convolution layer, a pooling layer and a third fully connected layer; the input frequency domain features are extracted by the convolution layer through weight sharing and local connection. The feature size and parameter amount are reduced by the pooling layer, and finally the processed features are mapped by the third fully connected layer to obtain a frequency domain feature vector of the required size.

9. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 1 is 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 an encoding process, which maps the input fusion feature data to the latent space; the process from the second hidden layer to the output layer is a decoding process, which outputs predicted data of the same dimension as the input layer data.

10. The multi-factor ocean observation data prediction method based on multi-feature fusion according to claim 9 is characterized in that: In step S5, the mean square error function is used as the loss function to constrain the optimization of the model parameters to reduce the difference between the predicted data and the real data.

Images