A machine learning-based sea wave spectrum prediction method and system
By utilizing a machine learning-based wave spectrum prediction system, multimodal data processing and deep learning models are employed to address the spatiotemporal correlation and dynamic changes in wave spectrum prediction, achieving accurate wave spectrum prediction and providing safety assurance for shipping and marine engineering.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GOLDWIND PIONEER TECHNOLOGY (YANCHENG) CO LTD
- Filing Date
- 2025-08-27
- Publication Date
- 2026-04-10
AI Technical Summary
In existing technologies, the complexity of marine environmental data leads to the high complexity of traditional wave spectrum prediction methods. These methods have limited spatiotemporal correlation and cannot effectively cope with the efficient processing of environmental data on wave formation and changes. Therefore, traditional methods are unable to accurately predict wave spectra.
A wave spectrum prediction system based on machine learning is adopted, including data acquisition, reconstruction, encoding and prediction modules. A wave spectrum prediction method is constructed through multimodal data processing, spatiotemporal coding and deep learning models.
It enables accurate prediction of wave spectrum, reflecting the dynamic changes of waves in the spatiotemporal dimensions, and provides accurate state prediction for shipping and marine engineering.
Smart Images
Figure CN121093768B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of marine disasters, and particularly relates to a sea wave spectrum prediction method and system based on machine learning. BACKGROUND
[0002] In many fields such as marine engineering, shipping safety, and marine resource development, accurate prediction of sea wave spectrum is of great significance. Sea wave spectrum can comprehensively reflect the distribution of sea wave energy at different frequencies and directions, and is a core element for describing the characteristics of sea waves. For example, in the shipping industry, accurate sea wave spectrum prediction can help ships plan routes in advance and avoid dangerous sea wave areas, ensuring shipping safety; in marine engineering construction, such as the construction of offshore wind power platforms, the wind resistance of the structure needs to be designed according to the sea wave spectrum to ensure the stability and durability of the project.
[0003] Currently, in the prior art, the formation and change of sea waves are influenced by a variety of environmental factors, and traditional data processing methods are difficult to cope with such high complexity of environmental data, resulting in insufficient feature extraction and affecting prediction accuracy; and sea waves have obvious spatial distribution characteristics and time evolution rules, and the sea wave state of different sea areas has spatial correlation, and the sea wave of the same sea area also has time dependence at different times, and the traditional prediction method has limited ability in processing spatial and temporal correlation, and cannot fully mine the internal relationship of sea waves in the time and space dimensions, so that the prediction result is difficult to reflect the dynamic change of sea waves.
[0004] Therefore, the present application proposes a sea wave spectrum prediction method and system based on machine learning to solve the above problems. SUMMARY
[0005] The main purpose of the present application is to provide a sea wave spectrum prediction method and system based on machine learning to solve the problems raised in the above background.
[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a sea wave spectrum prediction system based on machine learning, the system comprising a data acquisition module, a data reconstruction module, a data encoding module and a prediction module;
[0007] The data acquisition module is used for collecting sea wave spectrum, wind speed, wind direction, sea surface temperature, atmospheric pressure change data and ocean salinity change data in real time, and pre-processing the collected multi-modal data;
[0008] The data reconstruction module is used for multi-dimensional reconstruction of the pre-processed multi-modal data to form a comprehensive feature vector, and the multi-dimensional reconstruction includes feature selection, data standardization, feature dimension reduction and dynamic weighted fusion;
[0009] The data encoding module performs spatio-temporal encoding processing on the comprehensive feature vector, and the spatio-temporal encoding processing includes extracting spatial features by using a convolutional neural network and extracting time features by using a recurrent neural network, and fusing to construct complete spatio-temporal representation;
[0010] The prediction module is configured to input the spatio-temporal representation after the spatio-temporal encoding processing into the trained prediction model to obtain a sea wave spectrum prediction result.
[0011] Preferably, the data acquisition module includes an acquisition unit and a preprocessing unit.
[0012] The acquisition unit includes a wave sensor, a meteorological sensor, a temperature sensor, a pressure sensor, and a salinity sensor.
[0013] The preprocessing unit is configured to clean and denoise the collected real-time data; the cleaning includes outlier detection and missing value processing, and the denoising uses filtering and smoothing methods.
[0014] Preferably, the data reconstruction module includes a selection unit, a standard unit, a dimension reduction unit, and a weighted fusion unit.
[0015] The selection unit calculates the mutual information between each real-time feature and the sea wave spectrum target variable based on the preprocessed multi-modal data, filters out features with mutual information values higher than a set threshold, calculates the importance scores of the filtered features using a random forest model, and selects and retains the features with high scores.
[0016] Preferably, the standard unit performs data standardization processing on the retained features using the min-max standardization method, which maps the data to the [0, 1] interval.
[0017] Preferably, the dimension reduction unit calculates the covariance matrix between the data standardized features, performs eigenvalue decomposition on the covariance matrix to obtain eigenvalues and corresponding eigenvectors, sorts the eigenvalues according to their sizes, selects the top K largest eigenvalues and corresponding eigenvectors, constructs a projection matrix, and multiplies the original high-dimensional data with the projection matrix to obtain low-dimensional data, thereby realizing linear mapping of high-dimensional real-time data to low-dimensional space.
[0018] Preferably, the weighted fusion unit sets a threshold for key features based on historical data, constructs an ocean scene feature library by using the K-Means algorithm on historical scene data, calculates the cosine similarity between the real-time low-dimensional data and the scenes in the feature library when the real-time low-dimensional data is input, and assists in determining the current real-time scene.
[0019] According to the prior knowledge of the ocean scene, the weights and weight intervals of each feature in the low-dimensional data are initialized.
[0020] The real-time low-dimensional data is input into the trained prediction model to obtain a sea wave spectrum prediction value, an error between the sea wave spectrum prediction value and a real sea wave spectrum value is calculated, a weight is updated by adjusting and updating based on the error through a gradient descent algorithm and back propagation, and the weight update result is constrained based on a real-time scene recognition result to obtain a dynamic weight;
[0021] Based on the dynamic weight, feature information in the low-dimensional data is weighted and fused to obtain a comprehensive feature vector.
[0022] The prediction model is an LSTM model constructed based on deep learning and trained based on historical sea wave spectrum and associated data, and the training process aims to minimize the error between the sea wave spectrum prediction value and the real value.
[0023] Preferably, the data encoding module comprises a space unit, a time unit and a space-time fusion unit.
[0024] The space unit maps feature values corresponding to different latitude and longitude positions in the comprehensive feature vector to rows and columns of a space matrix according to a preset marine monitoring area latitude and longitude division, and constructs the space matrix.
[0025] A convolution kernel of k by K size suitable for capturing local features of marine data is selected, a convolution operation is performed on the space matrix, after the convolution operation, a maximum pooling processing is performed on the convolution output result, and a spatial feature is output.
[0026] Preferably, the time unit is used to arrange the spatial feature output by the space unit into a time sequence according to the time dimension of the marine data, and a long short-term memory network gating mechanism is used to perform feature extraction on the time sequence to output a time feature sequence containing the marine data.
[0027] Preferably, the space-time fusion unit is used to copy the spatial feature T times to form a spatial feature sequence with the same length as the time feature sequence.
[0028] The spatial feature sequence and the time feature sequence are spliced according to the feature dimension to obtain a space-time feature sequence, and global pooling of the maximum value in the time dimension is performed on the spliced space-time feature sequence to output a space-time representation.
[0029] A sea wave spectrum prediction method based on machine learning, comprising the following steps:
[0030] Step 1: Real-time acquisition of sea wave spectrum, meteorological and marine environment data by multi-source sensors, pre-processing of anomaly detection, missing value filling and filtering denoising, and output of standardized multi-modal data set;
[0031] Step 2: Key features are selected by mutual information and random forest, and after minimum-maximum standardization, a comprehensive feature vector is generated by PCA dimension reduction and dynamic weighted fusion.
[0032] Step three: map the comprehensive feature vector to a spatial matrix according to the latitude and longitude, and the convolutional neural network extracts the local spatial features; the recurrent neural network analyzes the time feature sequence, and after spatiotemporal splicing, global pooling fusion is performed to output the spatiotemporal representation;
[0033] Step four: input the spatiotemporal representation into the trained prediction model to obtain the sea wave spectrum prediction result.
[0034] The present application has the following beneficial effects:
[0035] 1. In the present application, by comprehensively collecting multi-modal data such as sea wave spectrum and wind speed, basic data for accurate prediction is provided, and preprocessing is performed, which lays a solid foundation for subsequent mining of sea wave features from complex environmental data. Based on mutual information, the strongly associated features of sea wave spectrum are selected from multi-modal data, and the importance of the features is further quantified using random forest. Key features are selected in combination with marine scenarios, and the selected key features are standardized to eliminate dimensional differences, so that different marine data can participate in model operation cooperatively. Then, through the covariance matrix and eigenvalue decomposition, the high-dimensional marine data is reduced, which effectively deals with the complex characteristics of sea waves.
[0036] 2. In the present application, the spatial unit constructs a spatial matrix based on the comprehensive feature vector according to the latitude and longitude, which is consistent with the spatial distribution characteristics of sea waves. The convolutional neural network is used for multi-layer convolution to extract spatial features from local to global, and the maximum pooling retains key information to accurately capture the spatial correlation of sea waves. The time unit relies on the LSTM gating mechanism to mine the dynamic evolution and long-time dependence of sea wave spectrum over time. The spatiotemporal fusion unit splices and pools the data to fuse the spatiotemporal features and form a complete spatiotemporal representation, which fully excavates the internal relationship in the spatiotemporal dimension of sea waves and provides high-quality input for the prediction model that can reflect the dynamic changes of sea waves. This allows the model to accurately grasp the change rule of sea wave spectrum in space and time, solves the problem of weak spatiotemporal correlation in traditional methods and inability to reflect the dynamic changes of sea waves, and provides accurate sea wave state prediction for shipping and marine engineering. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 The flowchart of the sea wave spectrum prediction method based on machine learning of the present application;
[0038] Figure 2 The framework diagram of the sea wave spectrum prediction system based on machine learning of the present application. DETAILED DESCRIPTION
[0039] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. 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 in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.
[0040] Embodiment 1: please refer to Figure 1 and Figure 2 The present application provides a technical solution: a sea wave spectrum prediction system based on machine learning, the system comprising a data acquisition module, a data reconstruction module, a data encoding module and a prediction module;
[0041] The data acquisition module is used for collecting sea wave spectrum, wind speed, wind direction, sea surface temperature, atmospheric pressure change data and ocean salinity change data in real time, and pre-processing the collected multi-modal data;
[0042] The data reconstruction module is used for multi-dimensional reconstruction of the pre-processed multi-modal data to form a comprehensive feature vector, and the multi-dimensional reconstruction includes feature selection, data standardization, feature dimension reduction and dynamic weighted fusion;
[0043] The data encoding module performs spatio-temporal encoding processing on the comprehensive feature vector, and the spatio-temporal encoding processing includes extracting spatial features by using a convolutional neural network and extracting time features by using a recurrent neural network, and fusing to build a complete spatio-temporal representation;
[0044] The prediction module is used for inputting the spatio-temporal representation after spatio-temporal encoding processing into a trained prediction model to obtain a sea wave spectrum prediction result.
[0045] The data acquisition module comprises an acquisition unit and a pre-processing unit;
[0046] The acquisition unit comprises a wave sensor, a meteorological sensor, a temperature sensor, a pressure sensor and a salinity sensor;
[0047] The pre-processing unit is used for cleaning and denoising the collected real-time data; the cleaning includes outlier detection and missing value processing, and the denoising uses filtering and smoothing methods.
[0048] The data reconstruction module comprises a selection unit, a standard unit, a dimension reduction unit and a weighted fusion unit;
[0049] The selection unit calculates the mutual information between each real-time feature and the sea wave spectrum target variable based on the pre-processed multi-modal data, filters out features with mutual information values higher than a set threshold, and then calculates the importance scores of the filtered features by using a random forest model, and selects and retains the features with high scores.
[0050] Specifically, the mutual information between the real-time features and the sea wave spectrum target variable is calculated according to the following formula:
[0051]
[0052] Where p(x, y) is the joint probability distribution of the real-time feature X and the sea wave spectrum Y, p(x) represents the marginal probability distribution of the real-time feature X, and p(y) represents the marginal probability distribution of the sea wave spectrum Y.
[0053] Meanwhile, the mutual information threshold is verified and set through experiments, and features satisfying the condition of mutual information I(X; Y) being greater than or equal to the mutual information threshold are screened out.
[0054] The sea wave spectrum is taken as the prediction target, and the features screened out by the mutual information are taken as the input to train a random forest composed of multiple decision trees. The random forest measures the importance score of a feature by the contribution of the feature to the splitting of a decision tree.
[0055] The features are sorted in descending order of the importance scores, and the top G features are selected, where the value of G is based on the marine scene.
[0056] The standard unit performs data standardization processing on the retained features, and adopts the minimum-maximum standardization method. The minimum-maximum standardization maps the data to the interval [0, 1].
[0057] The dimension reduction unit calculates the covariance matrix of the data standardized features, performs eigenvalue decomposition on the covariance matrix, obtains the eigenvalues and corresponding eigenvectors, sorts the eigenvalues in descending order, selects the top K largest eigenvalues and corresponding eigenvectors, constructs a projection matrix, and multiplies the original high-dimensional data by the projection matrix to obtain low-dimensional data, thereby realizing linear mapping of high-dimensional real-time data to a low-dimensional space.
[0058] Specifically, the covariance matrix of the features is calculated according to the following formula:
[0059]
[0060] Where V represents a feature matrix of n samples, represents a feature mean vector, and ∑ represents a covariance matrix.
[0061] The covariance matrix ∑ is subjected to eigenvalue decomposition to obtain eigenvalues λ1≥ λ2≥...≥ λ d and corresponding eigenvectors v1, v2,..., v d , where d is the original feature dimension.
[0062] The weighted fusion unit sets a threshold of the key feature based on historical data, and constructs an ocean scene feature library by a K-Means algorithm on historical scene data, calculates the cosine similarity of the real-time low-dimensional data with the scenes in the feature library to assist in determining the current real-time scene when the real-time low-dimensional data is input;
[0063] Specifically, the scenes corresponding to the top 3 cluster centers in terms of the highest cosine similarity are selected as candidate scenes; and the candidate scenes are verified based on the threshold of the key feature, for example, if the real-time wind speed of the key feature is greater than 10 m / s, the candidate of the breeze scene is excluded, and finally the scene with the highest similarity and meeting the threshold constraint is output as the current real-time scene.
[0064] According to the prior knowledge of the ocean scene, the weight and weight interval of each feature in the low-dimensional data are initialized;
[0065] The real-time low-dimensional data is input into the trained prediction model to obtain a sea wave spectrum prediction value, the error between the sea wave spectrum prediction value and the real value is calculated, the weight is updated by the gradient descent algorithm based on the error, and the weight update result is constrained in combination with the real-time scene recognition result to obtain a dynamic weight;
[0066] The gradient descent algorithm calculation formula is:
[0067]
[0068] wherein, η represents a learning rate, represents the weight of feature i at time t, e represents the error between the sea wave spectrum prediction value and the real value, represents the updated weight;
[0069] At the same time, the updated weight is adjusted according to the current real-time scene, so that it is forced to be within the interval set by the scene priori, if it exceeds the set interval, the maximum value of the interval is selected, if it is less than the set interval, the minimum value of the interval is selected, and the final dynamic weight is obtained.
[0070] The feature information in the low-dimensional data is weighted and fused based on the dynamic weight to obtain a comprehensive feature vector.
[0071] The prediction model is an LSTM model based on deep learning and trained by historical sea wave spectrum and related data, and the training process aims to minimize the error between the sea wave spectrum prediction value and the real value.
[0072] The LSTM model is a mature technology in the field and is not the inventive step of the patent, and will not be described here.
[0073] In this embodiment, by comprehensively collecting multi-modal data such as sea wave spectrum and wind speed, basic data for accurate prediction is provided, and through outlier detection and missing value processing, data impurities are removed, noise is removed through filtering and smoothing, data quality is improved, and a solid foundation is laid for subsequent mining of sea wave characteristics from complex environmental data. The selection unit first selects the features with strong correlation with the sea wave spectrum from the multi-modal data based on mutual information, finds out the key features that affect the change of the sea wave spectrum, and discards redundant information. Then, the random forest is used to further quantify the importance of the features, and the key features are selected in combination with the marine scene. The selected key features are standardized to eliminate the dimension difference, so that different marine data can be used in model operation. Then, the high-dimensional marine data is reduced through covariance matrix and eigenvalue decomposition, the main thread of the sea wave spectrum change is grasped, the calculation burden is reduced, and the key features are retained, which effectively cope with the complex characteristics of the sea wave.
[0074] In embodiment 2, based on the basis of embodiment 1, the data encoding module includes a space unit, a time unit and a space-time fusion unit.
[0075] The space unit divides the marine monitoring area according to the preset latitude and longitude, maps the feature values corresponding to different latitude and longitude positions in the comprehensive feature vector to the rows and columns of the space matrix, and constructs the space matrix.
[0076] A convolution kernel with a size of k by K is selected to capture the local features of the marine data, and a convolution operation is performed on the space matrix. After the convolution operation, a maximum pooling processing is performed on the convolution output result, and the spatial features are output.
[0077] The specific operation steps of performing the convolution operation on the space matrix are as follows:
[0078] At least two convolution layers are used for feature extraction, each layer includes multiple convolution kernels with a size of k by K, and K is an odd number such as 3 and 5. The operation formula is:
[0079]
[0080] Wherein, I l represents the input feature matrix of the lth layer, K l represents the convolution kernel weight of the lth layer, b l (c) represents the bias term of the cth channel of the lth layer, σ represents the ReLU activation function, and O l represents the output feature of the lth layer.
[0081] At the same time, a maximum pooling operation is performed after each convolution layer, and a 2x2 pooling window is used. The calculation formula is:
[0082] P l (i,j,c)=max{O l (2i,2j,c),O l(2i+1,2j,c),O l (2i,2j+1,c),O l (2i+1,2j+1,c)};
[0083] In the formula, P l represents the spatial feature matrix after the maximum pooling of the lth layer, O l represents the feature matrix of the lth layer, c represents the channel dimension of the feature matrix, i and j represent the row and column indexes of the feature matrix after the pooling, and 2i and 2j represent the starting coordinates of the pooling window on the input feature matrix.
[0084] The three-dimensional feature matrix output by the last layer of pooling is flattened into a one-dimensional spatial feature vector.
[0085] The time unit is used to arrange the spatial features output by the spatial unit into a time sequence according to the time dimension of the marine data, and the gating mechanism of the long short-term memory network is used to perform feature extraction on the time sequence, and output a time feature sequence containing the marine data.
[0086] The gating mechanism of the long short-term memory network is used to perform feature extraction on the time sequence, which specifically includes:
[0087] Input gate: controls the fusion ratio of the current time step feature Q t and the historical hidden state h t-1 , and the formula is:
[0088] i t =σ(W xi ·Q t + W hi ·h t-1 +b i );
[0089] Forget gate: screens the information to be retained in the historical cell state c t-1 , for example, the influence of the previous storm on the current sea wave spectrum, and the formula is:
[0090] f t =σ(W xf ·Q t +W hf ·h t-1 +b f );
[0091] Cell state update: fuse historical information and current input to generate new cell state, formula is:
[0092]
[0093] Output gate: determines the current cell state c tThe time feature to be output is as follows:
[0094]
[0095] wherein W xi , W hi , W xf , W hf , W xc , W hc , W xo , W ho represent the weight matrix of the LSTM gating unit, b i , b f , b c and b o represent the bias term, sigma is a sigmoid activation function, tanh is a hyperbolic tangent activation function, and is a element-wise multiplication operation; through the above operations, a time feature sequence containing time dynamic characteristics is output;
[0096] The space-time fusion unit is used to copy the spatial feature T times to form a spatial feature sequence with the same length as the time feature sequence.
[0097] The spatial feature sequence and the time feature sequence are spliced according to the feature dimension to obtain a space-time feature sequence, and global pooling of the maximum value in the time dimension is performed on the spliced space-time feature sequence to output a space-time representation.
[0098] In this embodiment, the spatial unit constructs a spatial matrix based on the comprehensive feature vector according to the latitude and longitude, which is consistent with the spatial distribution characteristics of the sea wave. The spatial features are extracted from local to global by using multi-layer convolution of the convolutional neural network, and the key information is retained by using max-pooling to accurately capture the spatial correlation of the sea wave. The time unit relies on the LSTM gating mechanism to sort the spatial feature time sequence, and the input gate, the forget gate and the like cooperate to mine the dynamic evolution and long-time dependence of the sea wave spectrum over time. The space-time fusion unit splices and pools the processing to fuse the space-time features to form a complete space-time representation, fully mines the internal relationship in the space-time dimension of the sea wave, provides a high-quality input reflecting the dynamic changes of the sea wave for the prediction model, and enables the model to accurately grasp the change rule of the sea wave spectrum in space-time. Based on the LSTM model of deep learning, the model is trained based on the historical sea wave spectrum and related data. The gating mechanism effectively processes the time sequence data according to the dynamic change characteristics of the sea wave in space-time, learns the complex relationship between the marine environment and the sea wave spectrum, and outputs the sea wave spectrum prediction result after inputting the encoded space-time representation. The model uses the learned rule to solve the problem that the traditional method has weak space-time correlation and cannot reflect the dynamic changes of the sea wave, and provides accurate sea wave state prediction for shipping and marine engineering.
[0099] The application discloses a sea wave spectrum prediction method and system based on machine learning, which comprehensively collects multi-modal data such as sea wave spectrum and wind speed to provide basic data for accurate prediction, and removes data impurities through anomaly value detection and missing value processing, and denoising through filtering and smoothing to improve data quality, lay a solid foundation for subsequent mining of sea wave characteristics from complex environmental data, and select units to first screen strong correlation features from multi-modal data with sea wave spectrum based on mutual information, find out key features that affect the change of sea wave spectrum, discard redundant information, and then further quantify feature importance by using a random forest, select key features in combination with a marine scene, and standardize the selected key features to eliminate dimension differences and enable different marine data to cooperatively participate in model operation, then reduce high-dimensional marine data through a covariance matrix and eigenvalue decomposition to grasp the main thread of sea wave spectrum change, reduce the calculation burden while retaining key features, and effectively cope with complex characteristics of sea wave spectrum.
[0100] The spatial unit constructs a spatial matrix based on the comprehensive feature vector according to longitude and latitude, matches the spatial distribution characteristics of sea wave, uses a convolutional neural network to perform multi-layer convolution to extract spatial features from local to global, retains key information through max-pooling, accurately captures the spatial correlation of sea wave, and the time unit relies on the LSTM gating mechanism to sort out the time sequence of spatial features, cooperates input gates, forgetting gates and the like, and mines the dynamic evolution and long-time dependence of sea wave spectrum with time.
[0101] It should be noted that, in the present document, the terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.
[0102] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.
Claims
1. A machine learning based sea wave spectrum prediction system, characterized by, The system comprises a data acquisition module, a data reconstruction module, a data coding module and a prediction module; The data acquisition module is used for collecting sea wave spectrum, wind speed, wind direction, sea surface temperature, atmospheric pressure change data and ocean salinity change data in real time, and pre-processing the collected multi-modal data; The data reconstruction module is used for multi-dimensional reconstruction of the pre-processed multi-modal data to form a comprehensive feature vector, wherein the multi-dimensional reconstruction comprises feature selection, data standardization, feature dimension reduction and dynamic weighted fusion; The data coding module performs spatio-temporal coding processing on the comprehensive feature vector, wherein the spatio-temporal coding processing comprises extracting spatial features by using a convolutional neural network and extracting time features by using a recurrent neural network, and fusing and constructing complete spatio-temporal representation; The prediction module is used for inputting the spatio-temporal representation after the spatio-temporal coding processing into a trained prediction model to obtain a sea wave spectrum prediction result. The data coding module comprises a spatial unit, a temporal unit and a spatio-temporal fusion unit; The spatial unit divides the comprehensive feature vector according to the preset marine monitoring region latitude and longitude, maps the feature values corresponding to different latitude and longitude positions in the comprehensive feature vector to the rows and columns of a spatial matrix, and constructs a spatial matrix; A convolution kernel with a size of k by K that is adapted to the local feature capture requirement of marine data is selected, a convolution operation is performed on the spatial matrix, after the convolution operation, a max-pooling processing is performed on the convolution output result, and spatial features are outputted; The temporal unit is used for arranging the spatial features outputted by the spatial unit into a time sequence according to the time dimension of marine data, and performing feature extraction on the time sequence by using a gating mechanism of a long short-term memory network to output a time feature sequence containing marine data; The spatio-temporal fusion unit is used for copying the spatial features T times to form a spatial feature sequence with the same length as the time feature sequence; The spatial feature sequence and the time feature sequence are spliced according to the feature dimension to obtain a spatio-temporal feature sequence, and a global pooling of the maximum value in the time dimension is performed on the spliced spatio-temporal feature sequence to output spatio-temporal representation.
2. The machine learning based sea wave spectrum prediction system of claim 1, wherein, The data acquisition module comprises an acquisition unit and a preprocessing unit; The acquisition unit comprises a wave sensor, a meteorological sensor, a temperature sensor, a pressure sensor and a salinity sensor; The preprocessing unit is used for cleaning and denoising the collected real-time data; the cleaning comprises outlier detection and missing value processing, and the denoising adopts filtering and smoothing methods.
3. The machine learning based sea wave spectrum prediction system of claim 1, wherein, The data reconstruction module comprises a selection unit, a standard unit, a dimension reduction unit and a weighted fusion unit; The selection unit calculates the mutual information between each real-time feature and the sea wave spectrum target variable based on the pre-processed multi-modal data, and screens out features with mutual information values higher than a set threshold; Then, the importance scores of the screened features are calculated by using a random forest model, and the features with high scores are selected and reserved.
4. The machine learning based sea wave spectrum prediction system of claim 3, wherein, The standard unit performs data standardization processing on the reserved features by using a min-max standardization method, which maps the data to the [0, 1] interval.
5. The machine learning based sea wave spectrum prediction system of claim 3, wherein, The dimension reduction unit calculates the covariance matrix between the features of the data standardization processing, performs eigenvalue decomposition on the covariance matrix, obtains eigenvalues and corresponding eigenvectors, sorts the eigenvalues according to the size, selects the first K largest eigenvalues and the corresponding eigenvectors, constructs a projection matrix, and multiplies the original high-dimensional data with the projection matrix to obtain low-dimensional data, thereby realizing linear mapping of high-dimensional real-time data to a low-dimensional space.
6. The machine learning based sea wave spectrum prediction system of claim 3, wherein, The weighted fusion unit sets a threshold value of a key feature based on historical data, constructs an ocean scene feature library by a K-Means algorithm based on historical scene data, calculates the cosine similarity of real-time low-dimensional data with the scenes in the feature library when the real-time low-dimensional data is input, and assists in determining the current real-time scene; According to the prior knowledge of the ocean scene, the weight and the weight interval of each feature in the low-dimensional data are initialized; The real-time low-dimensional data is input into the trained prediction model to obtain a sea wave spectrum prediction value, the error between the sea wave spectrum prediction value and the real value is calculated, the weight is updated by backward propagation adjustment based on the error through a gradient descent algorithm, and the dynamic weight is obtained by combining the real-time scene recognition result with the weight update result; Based on the dynamic weight, the feature information in the low-dimensional data is weighted and fused to obtain a comprehensive feature vector. The prediction model is an LSTM model constructed based on deep learning and trained based on historical sea wave spectrum and associated data, and the training process aims to minimize the error between the sea wave spectrum prediction value and the real value.
7. A machine learning based sea wave spectrum prediction method, referring to any one of claims 1-6, a machine learning based sea wave spectrum prediction system, characterized in that, The method comprises the following steps: Step one: real-time acquisition of sea wave spectrum, meteorological and marine environment data through multi-source sensors, preprocessing of anomaly detection, missing value filling and filtering denoising, and output of standardized multi-modal data set; Step two: screening of key features by mutual information and random forest, generation of a comprehensive feature vector by PCA dimension reduction and dynamic weighted fusion after minimum-maximum standardization; Step three: mapping of the comprehensive feature vector to a spatial matrix according to latitude and longitude, extraction of local spatial features by a convolutional neural network, analysis of time feature sequences by a recurrent neural network, global pooling fusion after spatio-temporal concatenation, and output of spatio-temporal representation; Step four: input of the spatio-temporal representation into the trained prediction model to obtain a sea wave spectrum prediction result.
Citation Information
Patent Citations
Marine hydrometeorological data fusion processing method based on deep learning fusion
CN120067973A