Ocean current velocity profile pattern recognition method based on DWT and joint-nmf

Abstract

This application belongs to the fields of marine engineering, marine environmental perception and intelligent recognition technology. Specifically, it discloses a method for ocean current velocity profile pattern recognition based on DWT and Joint-NMF, including: constructing and performing discrete wavelet decomposition on the ocean current velocity profile sample matrix to obtain sub-band components of different scales, using each sub-band component as a multi-view feature to construct a multi-view feature set; establishing a weighted joint non-negative matrix factorization model to perform co-decomposition on the feature matrices of each view in the multi-view feature set to obtain the low-dimensional representation matrix corresponding to each view; performing weighted fusion on the low-dimensional representation matrices corresponding to each view after co-decomposition based on the pattern structure similarity of each sub-band feature after discrete wavelet decomposition to obtain a fused feature representation; and performing spectral clustering analysis to obtain the ocean current velocity profile pattern recognition result. This application can improve the accuracy and stability of ocean current velocity profile pattern recognition in complex marine environments.

Description

Technical Field

[0001] This application belongs to the fields of marine engineering, marine environmental perception and intelligent recognition technology, and more specifically, relates to a method for recognizing ocean current velocity profiles. Background Technology

[0002] Ocean currents are a crucial dynamic element in the marine environment. Their variations are influenced by a variety of factors, including wind fields, tides, topography, temperature and salinity structure, and air-sea interactions, exhibiting significant spatiotemporal multi-scale characteristics and complex vertical structures. Ocean current velocity profiles reflect the motion and evolution of ocean currents at different depths, serving as essential foundational data for understanding ocean dynamic processes, conducting marine environmental monitoring, and supporting marine engineering operations. Therefore, effective analysis and pattern recognition of ocean current velocity profiles are of great significance for maritime operational safety, navigation support, and marine resource development.

[0003] Existing methods for analyzing ocean current velocity profiles mainly include direct statistical analysis, empirical rule discrimination, and traditional clustering identification. While these methods are applicable to simple scenarios, they still have limitations in complex marine environments. Ocean current velocity profile data typically exhibits high dimensionality, strong correlation, and significant non-stationarity. Direct identification within the original feature space is susceptible to noise interference and feature redundancy, leading to unclear pattern boundaries, category confusion, and unstable identification results. Furthermore, information at different scales within the ocean current velocity profile has different physical meanings. Existing methods mostly focus on single-scale analysis or employ simple splicing and equal weighting for multi-scale features, failing to adequately address the representation needs of both overall trends and local variations, thus affecting pattern recognition accuracy. In addition, existing methods often use traditional clustering methods based on Euclidean distance or simple similarity metrics to divide samples. When ocean current velocity profiles exhibit nonlinear structures, multi-scale information coupling, and blurred inter-class boundaries, traditional clustering methods struggle to fully reveal the potential structural relationships between samples, easily resulting in inaccurate category division, insufficient pattern discrimination ability, and low stability of clustering results. Meanwhile, the importance of features at different scales in describing ocean current structure is not consistent. Existing methods usually lack a mechanism to reasonably distinguish the degree of feature contribution based on data characteristics, which can easily lead to strong features dominating the results and weak features being masked. Different features may be consistent or complementary. If there is a lack of appropriate representation and fusion methods, it is difficult to balance noise suppression and effective information preservation.

[0004] Therefore, improving the accuracy and stability of ocean current velocity profile pattern recognition in complex marine environments is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a method for ocean current velocity profile pattern recognition based on discrete wavelet decomposition and weighted joint nonnegative matrix decomposition, which can effectively improve the accuracy and stability of ocean current velocity profile pattern recognition in complex marine environments.

[0006] To achieve the above objectives, in a first aspect, this application provides a method for ocean current velocity profile pattern recognition based on Discrete Wavelet Decomposition (DWT) and Weighted Joint Nonnegative Matrix Factorization (Joint-NMF), comprising the following steps: S10, normalize the ocean current velocity profile data and construct an ocean current velocity profile sample matrix; S20. Discrete wavelet decomposition is performed on the ocean current velocity profile sample matrix to obtain sub-band components at different scales. Each sub-band component is used as a multi-view feature to construct a multi-view feature set. S30. Establish a joint nonnegative matrix factorization model with weight constraints, and perform collaborative decomposition on the feature matrices of each view in the multi-view feature set to obtain the low-dimensional representation matrix corresponding to each view. S40. Based on the pattern structure similarity of each sub-band feature after discrete wavelet decomposition, the low-dimensional representation matrices corresponding to each view after co-decomposition are weighted and fused to obtain the fused feature representation. S50 performs spectral clustering analysis on the fused feature representation to obtain the ocean current velocity profile pattern recognition results.

[0007] As a further preferred embodiment, the specific process of normalizing the ocean current velocity profile data and constructing the ocean current velocity profile sample matrix in step S10 is as follows: Let the original ocean current velocity profile sample matrix be... ,in N Indicates the number of samples. M This represents the depth dimension of a single ocean current velocity profile sample. Indicates the first i The sample at the th j Values ​​at depth positions; For each dimension of the ocean current velocity profile sample matrix, its minimum and maximum values ​​are calculated across all samples. A minimum-maximum normalization method is then used to linearly map the depth data of the ocean current velocity profile, resulting in the normalized ocean current velocity profile sample matrix. , , Representing the first j Minimum and maximum flow velocity values ​​among all samples at each depth location; The normalized ocean current velocity profile data are arranged according to the sample dimension and the profile dimension to construct an ocean current velocity profile sample matrix.

[0008] As a further preferred embodiment, in step S20, the specific process of performing discrete wavelet decomposition on the ocean current velocity profile sample matrix to obtain sub-band components at different scales and constructing a multi-view feature set is as follows: Discrete wavelet multilevel decomposition was performed on the one-dimensional ocean current velocity profile sequence corresponding to each sample in the ocean current velocity profile sample matrix, and the Daubechies wavelet was selected as the mother wavelet function. J Discrete wavelet decomposition is employed, using a two-level decomposition structure. Through discrete wavelet decomposition, the original ocean current velocity profile sequence is expressed as: ,in Indicates the first J The low-frequency approximation component of the layer is used to characterize the overall variation trend of ocean current velocity profiles. Indicates the first j High-frequency detail components are used to characterize local perturbation features and dynamic changes at different scales; Since the discrete wavelet decomposition is orthogonal, the different sub-band components and energies satisfy the following: ; Based on the fast wavelet reconstruction method, the sub-band components at each scale are restored to the same length as the original ocean current velocity profile sequence, resulting in: ,in Representing the i The two-layer low-frequency approximation components of each sample mainly represent the overall trend of change; Representing the i The high-frequency detail components of a single sample mainly represent small, rapidly changing local perturbations; Representing the i The second-level detail components of a sample typically correspond to mid-frequency variations, expressed as a ratio. More gradual, larger-scale local changes; The different scale sub-band components mentioned above are used as multi-view feature representations to construct a multi-view feature set. ,in, V Indicates the number of views. This represents the feature matrix corresponding to the v-th view.

[0009] As a further preferred embodiment, in step S30, a joint nonnegative matrix factorization model with weight constraints is established to perform collaborative decomposition of the multi-view features, specifically including: Calculate the energy proportion of each view feature and determine the corresponding weight, and calculate the energy value of the multi-view features obtained by discrete wavelet decomposition respectively: ,in Indicates the first v In the feature matrix of the nth view i The sample at the th jValues ​​at depth positions; The weights are determined based on the proportion of each view's feature energy to the total energy. , Representing the v The energy percentage of each view, and satisfying , .

[0010] As a further preferred embodiment, step S30, in which the joint nonnegative matrix factorization model with weight constraints is constructed, further includes: Input the feature matrices of each view into the joint nonnegative matrix factorization model, and perform the following decomposition for each view: ,in To share the basis matrix, For the first v The low-dimensional representation matrix corresponding to each view K Low-dimensional feature dimension; Under non-negativity constraints, satisfy .

[0011] As a further preferred approach, the joint nonnegative matrix factorization model is subjected to collaborative optimization based on weighted reconstruction error. An objective function is constructed by minimizing the multi-view weighted reconstruction error, and a regularization term is introduced into the objective function. ,in It is the Frobenius norm. and This is the regularization parameter.

[0012] As a further preferred approach, the objective function is iteratively optimized to achieve collaborative decomposition of the views under the constraint of a shared basis matrix, thereby obtaining a low-dimensional latent representation shared by multiple views. W and the low-dimensional feature representations corresponding to each view. Shared basis matrix Common structures used to characterize ocean current velocity profiles at different scales; low-dimensional representation of each view. It is used to characterize the cross-sectional representation at the corresponding scale, thereby realizing the fusion representation of information at different scales.

[0013] As a further preferred embodiment, in step S40, the low-dimensional features after collaborative decomposition are weighted and fused based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition, specifically including two fusion methods: When the features of each sub-band are similar in overall trend and structural form, a weighted summation method is used for fusion to obtain a unified feature representation: When the features of each sub-band differ in structural morphology or local variation characteristics, a weighted splicing method is used for fusion to obtain a joint feature representation: ; The weighted summation result is merged. Or weighted splicing and fusion results Z As the final fusion feature representation.

[0014] As a further preferred option, step S50 specifically includes: After standardizing the fused feature representations, the similarity between samples is calculated, and a similarity matrix is ​​constructed using radial basis functions. ,in It is a scale parameter. Indicates the first i One fused feature vector; Construct a degree matrix based on the similarity matrix. D , D The specific formula is as follows: ; Constructing a symmetric normalized Laplace matrix L : ,in Represents the identity matrix; For the normalized Laplace matrix L Find the minimum r The eigenvectors corresponding to the eigenvalues ​​constitute the feature matrix. U : ,in r Indicates the number of cluster categories; For the feature matrix U The spectral embedding matrix is ​​obtained by normalizing the rows. The K-means clustering method is used to divide the spectral embedding matrix into samples, and the category labels of each ocean current velocity profile sample are obtained, thereby obtaining the ocean current velocity profile pattern recognition results and realizing the automatic classification of different ocean current structure patterns.

[0015] Secondly, this application provides an ocean current velocity profile pattern recognition system for implementing the method described in any one of the above, comprising: The normalization and matrix construction module is used to normalize ocean current velocity profile data and construct ocean current velocity profile sample matrices. The discrete wavelet decomposition module is used to perform discrete wavelet decomposition on the ocean current velocity profile sample matrix to obtain sub-band components at different scales. Each sub-band component is used as a multi-view feature to construct a multi-view feature set. The joint nonnegative matrix factorization module is used to establish a joint nonnegative matrix factorization model with weight constraints, and to perform collaborative decomposition of the feature matrices of each view in the multi-view feature set to obtain the low-dimensional representation matrix corresponding to each view. The weighted fusion module is used to perform weighted fusion of the low-dimensional representation matrices corresponding to each view after collaborative decomposition based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition, so as to obtain the fused feature representation. The spectral clustering analysis module is used to perform spectral clustering analysis on the fused feature representation to obtain the pattern recognition results of ocean current velocity profiles.

[0016] The beneficial effects of the technical solution provided in this application are as follows: By performing discrete wavelet decomposition on the ocean current velocity profile sample matrix, independent and complementary multi-scale sub-band components can be obtained as multi-view features. Among them, low-frequency components characterize the overall flow structure, while high-frequency components characterize local disturbances and transient changes, thereby effectively separating multi-scale dynamic information in the ocean current velocity profile and enhancing the multi-scale feature representation capability. By establishing a joint nonnegative matrix decomposition model with weight constraints to perform collaborative decomposition of multi-view features, low-dimensional representations of each view can be extracted under shared basis matrix constraints, which helps to reduce the interference of noise and feature redundancy on the recognition results. Disturbances are eliminated; based on the pattern structure similarity of each sub-band feature after discrete wavelet decomposition, the low-dimensional representation matrix after co-decomposition is weighted and fused, which can make full use of the complementary information of features at different scales, suppress the noise superposition effect while maintaining the overall structural consistency, and improve the stability of feature representation. Finally, spectral clustering analysis is performed on the fused feature representation, which can fully reveal the potential structural relationship between samples. It can effectively improve the accuracy and stability of ocean current velocity profile pattern recognition in complex marine environments, thus providing effective technical support for marine environmental monitoring, navigation safety assurance and marine engineering operations, and has good practicality and application prospects. Attached Figure Description

[0017] Figure 1 This is a flowchart of the ocean current velocity profile pattern recognition method based on discrete wavelet decomposition and weighted joint nonnegative matrix decomposition provided in this application; Figure 2 This is a heat map of a sample matrix of northward ocean current velocity profiles in a certain location after normalization, as provided in an embodiment of this application. Figure 3 This is a detailed comparison between the heatmaps of each sub-band component after discrete wavelet decomposition and the heatmap of the original sample matrix provided in the embodiments of this application; wherein, (a) is the heatmap of the original sample matrix, (b) is the heatmap of the D1 matrix of discrete wavelet decomposition, (c) is the heatmap of the D2 matrix of discrete wavelet decomposition, and (d) is the heatmap of the A2 matrix of discrete wavelet decomposition. Figure 4These are the minimum-maximum normalized heatmaps of each sub-band component after discrete wavelet decomposition provided in the embodiments of this application; wherein, (a) is the minimum-maximum normalized heatmap of the D1 matrix, (b) is the minimum-maximum normalized heatmap of the D2 matrix, and (c) is the minimum-maximum normalized heatmap of the A2 matrix. Figure 5 This is a comparison chart of the pattern recognition performance of the proposed method and the real pattern. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] Existing methods for recognizing ocean current velocity profiles are susceptible to noise interference, feature redundancy, and multi-scale information coupling, resulting in insufficient feature representation, inadequate recognition accuracy, and insufficient stability, making it difficult to meet the application requirements in complex marine environments. To address this, this application proposes an ocean current velocity profile pattern recognition method based on discrete wavelet decomposition and weighted joint nonnegative matrix decomposition. Through multi-scale decomposition, multi-view collaborative decomposition, and adaptive fusion, this method achieves effective recognition of ocean current velocity profile patterns, improving the accuracy and stability of pattern recognition in complex marine environments.

[0020] Please refer to Figure 1 , Figure 1 This is a flowchart of ocean current velocity profile pattern recognition based on discrete wavelet decomposition and weighted joint nonnegative matrix decomposition provided in this application, specifically including the following steps: S1: Normalize the ocean current velocity profile data and construct an ocean current velocity profile sample matrix.

[0021] The collected ocean current velocity profile data are preprocessed, and the original ocean current velocity profile sample matrix is ​​set as follows:

[0022] in, N Indicates the number of samples. M This represents the depth dimension of a single ocean current velocity profile sample. Indicates the first i The sample at the th j The value is taken at each depth position.

[0023] For each dimension of the ocean current velocity profile sample matrix, its minimum and maximum values ​​are calculated across all samples. A minimum-maximum normalization method is then used to linearly map the depth data of the ocean current velocity profile, resulting in the normalized ocean current velocity profile sample matrix.

[0024] The normalized ocean current velocity profile data are arranged according to the sample dimension and the profile dimension to construct an ocean current velocity profile sample matrix, which serves as the unified input data for subsequent discrete wavelet decomposition, multi-view feature construction, and ocean current velocity profile pattern recognition.

[0025] refer to Figure 2 This is a heatmap of the ocean current velocity profile sample matrix after normalization processing of the northward ocean current velocity profile of a certain location. The vertical axis represents the depth layer, and the horizontal axis represents the time sample on a daily scale. The color bars represent the velocity values ​​at different depths in the sample.

[0026] S2: Perform discrete wavelet decomposition on the ocean current velocity profile sample matrix to obtain sub-band components at different scales and construct multi-view features.

[0027] Discrete wavelet multilevel decomposition was performed on the one-dimensional ocean current velocity profile sequence corresponding to each sample in the obtained ocean current velocity profile sample matrix. The Daubechies wavelet, which has compact support and good orthogonality, was selected as the mother wavelet function. J Discrete wavelet decomposition, preferably employing a two-level decomposition structure, is used to simultaneously extract high-frequency details, mid-frequency details, and low-frequency trend information. Through discrete wavelet decomposition, the original ocean current velocity profile sequence is expressed as:

[0028] in, Indicates the first J The low-frequency approximation component of the layer is used to characterize the overall variation trend of ocean current velocity profile; Indicates the first j High-frequency detail components are used to characterize local perturbation features and dynamic changes at different scales.

[0029] Since the discrete wavelet decomposition is orthogonal, the different sub-band components and energies satisfy the following:

[0030] These characteristics enable sub-bands at different scales to be independent and complementary in their information representation. The low-frequency components characterize the overall flow structure, while the high-frequency components characterize local disturbances and transient changes. This allows for the effective characterization of the complex characteristics of ocean current velocity profiles from multiple scales, providing a foundation for subsequent multi-view feature fusion and ocean current velocity profile pattern recognition.

[0031] Based on the fast wavelet reconstruction method, the sub-band components at each scale are restored to the same length as the original ocean current velocity profile sequence, resulting in:

[0032] Using the sub-band components of different scales as multi-view feature representations, a multi-view feature set is constructed:

[0033] in, V Indicates the number of views. Indicates the first v The feature matrix corresponding to each view, such as Right now .

[0034] refer to Figure 3 This is the result of a two-level decomposition based on Daubechies wavelet for the one-dimensional ocean current velocity profile sequence corresponding to each sample in the ocean current velocity profile sample matrix. Among them, (a) is the normalized ocean current velocity profile sample matrix; (b) is the sub-band component D1 of the first level decomposition, representing high-frequency details; (c) is the sub-band component D2 of the second level decomposition, representing mid-frequency details; and (d) is the remaining low-frequency approximate component after the second level decomposition, representing the low-frequency trend of the ocean current.

[0035] S3: Establish a joint nonnegative matrix factorization model with weight constraints to perform collaborative decomposition of the multi-view features.

[0036] (1-1): Calculate the energy proportion of each view feature and determine the corresponding weight. Calculate the energy values ​​of the multi-view features obtained by discrete wavelet decomposition respectively:

[0037] The weights are determined based on the proportion of each view's feature energy to the total energy.

[0038] in, Representing the v The energy percentage of each view, and satisfying , .

[0039] (1-2): Constructing a joint nonnegative matrix factorization model with weight constraints Input the feature matrices of each view into the joint nonnegative matrix factorization model, and perform the following decomposition for each view:

[0040] in, To share the basis matrix, For the first v The low-dimensional representation matrix corresponding to each view K It is a low-dimensional feature dimension.

[0041] Under nonnegativity constraints:

[0042] (1-3): Collaborative optimization based on weighted reconstruction error An objective function is constructed by minimizing the multi-view weighted reconstruction error, and a regularization term is introduced into the objective function to improve model stability.

[0043] in, It is the Frobenius norm. and The regularization parameter is set to 1. .

[0044] (1-4): Multi-view collaborative decomposition and low-dimensional feature extraction By iteratively optimizing the objective function, the views achieve collaborative decomposition under the constraint of a shared basis matrix, thereby obtaining a low-dimensional latent representation shared by multiple views. and the low-dimensional feature representations corresponding to each view. Shared basis matrix Common structures used to characterize ocean current velocity profiles at different scales; low-dimensional representation of each view. It is used to characterize the cross-sectional representation at the corresponding scale, thereby realizing the fusion representation of information at different scales.

[0045] Two weighted fusion methods were designed to achieve multi-view feature fusion. The appropriate fusion method was selected based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition.

[0046] S4: Based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition, the low-dimensional features after collaborative decomposition are weighted and fused. Two weighted fusion methods were designed to achieve multi-view feature fusion. The appropriate fusion method was selected based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition.

[0047] (2-1): Weighted summation and fusion When the features of each sub-band are similar in overall trend and structural form, a weighted summation method is used for fusion to obtain a unified feature representation:

[0048] This fusion method linearly combines low-dimensional features from different views, which can suppress noise superposition effects while maintaining overall structural consistency and improve the stability of fused features.

[0049] (2-2): Weighted splicing and fusion When there are significant differences in the structural morphology or local variation features of the sub-bands, a weighted splicing method is used for fusion to obtain a joint feature representation:

[0050] This fusion method splices and expands the low-dimensional representations of different views along the feature dimension, which can fully preserve the complementary information between features at different scales and improve feature representation capability. The weighted summation fusion result... Or weighted splicing and fusion results As the final fused feature representation, it is used for subsequent ocean current velocity profile pattern recognition.

[0051] refer to Figure 4 The image is obtained by min-max normalization of the three high, medium and low frequency sub-band components D1, D2 and A2 after discrete wavelet decomposition. As can be seen from the image, the structural morphology of sub-bands D1 and D2 is similar. Therefore, this embodiment adopts a weighted summation fusion method.

[0052] S5: Perform spectral clustering analysis on the fused feature representation to obtain the ocean current velocity profile pattern recognition results, as follows: (3-1): Constructing a fusion feature similarity matrix Spectral clustering is used to perform cluster analysis on the fused feature representations. After standardizing the fused feature representations, the similarity between samples is calculated, and a similarity matrix is ​​constructed. The similarity matrix is ​​constructed here using radial basis functions:

[0053] in It is a scale parameter. Indicates the first i A fused feature vector.

[0054] (3-2): Constructing the normalized Laplacian matrix and the characteristic matrix Based on the similarity matrix S Construct the degree matrix D , D The specific formula is as follows:

[0055] Further construct the symmetric normalized Laplace matrix L have:

[0056] in, Represents the identity matrix.

[0057] For the normalized Laplace matrix L Find the minimum rThe eigenvectors corresponding to the eigenvalues ​​constitute the feature matrix. U :

[0058] in, r This represents the number of cluster categories. For the feature matrix... U The spectral embedding matrix is ​​obtained by normalizing the rows.

[0059] Finally, the K-means clustering method is used to divide the samples into categories based on the spectral embedding matrix, thereby obtaining the category labels of each ocean current velocity profile sample and thus obtaining the ocean current velocity profile pattern recognition results, realizing the automatic division of different ocean current structure patterns.

[0060] refer to Figure 5 This is a comparison between the final result of the proposed method for ocean current velocity profile pattern recognition and the actual pattern. With a clustering number of 6 patterns, the proposed method achieves a final pattern recognition accuracy of 70.98%, enabling rapid identification of current ocean current velocity profile patterns while maintaining a certain level of accuracy.

[0061] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for recognizing ocean current velocity profile patterns based on DWT and Joint-NMF, characterized in that, Includes the following steps: S10, normalize the ocean current velocity profile data and construct an ocean current velocity profile sample matrix; S20. Discrete wavelet decomposition is performed on the ocean current velocity profile sample matrix to obtain sub-band components at different scales. Each sub-band component is used as a multi-view feature to construct a multi-view feature set. S30. Establish a joint nonnegative matrix factorization model with weight constraints, and perform collaborative decomposition on the feature matrices of each view in the multi-view feature set to obtain the low-dimensional representation matrix corresponding to each view. S40. Based on the pattern structure similarity of each sub-band feature after discrete wavelet decomposition, the low-dimensional representation matrices corresponding to each view after co-decomposition are weighted and fused to obtain the fused feature representation. S50 performs spectral clustering analysis on the fused feature representation to obtain the ocean current velocity profile pattern recognition results.

2. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 1, characterized in that, In step S10, the specific process of normalizing the ocean current velocity profile data and constructing the ocean current velocity profile sample matrix is ​​as follows: Let the original ocean current velocity profile sample matrix be... ,in N Indicates the number of samples. M This represents the depth dimension of a single ocean current velocity profile sample. Indicates the first i The sample at the th j Values ​​at depth positions; For each dimension of the ocean current velocity profile sample matrix, its minimum and maximum values ​​are calculated across all samples. A minimum-maximum normalization method is then used to linearly map the depth data of the ocean current velocity profile, resulting in the normalized ocean current velocity profile sample matrix. ,in , Representing the first j Minimum and maximum flow velocity values ​​among all samples at each depth location; The normalized ocean current velocity profile data are arranged according to the sample dimension and the profile dimension to construct an ocean current velocity profile sample matrix.

3. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 1, characterized in that, In step S20, the specific process of performing discrete wavelet decomposition on the ocean current velocity profile sample matrix to obtain sub-band components at different scales and constructing a multi-view feature set is as follows: Discrete wavelet multilevel decomposition was performed on the one-dimensional ocean current velocity profile sequence corresponding to each sample in the ocean current velocity profile sample matrix, and the Daubechies wavelet was selected as the mother wavelet function. J Discrete wavelet decomposition is employed, using a two-level decomposition structure. Through discrete wavelet decomposition, the original ocean current velocity profile sequence is expressed as: ,in Indicates the first J The low-frequency approximation component of the layer is used to characterize the overall variation trend of ocean current velocity profiles. Indicates the first j High-frequency detail components are used to characterize local perturbation features and dynamic changes at different scales; Since the discrete wavelet decomposition is orthogonal, the different sub-band components and energies satisfy the following: ； Based on the fast wavelet reconstruction method, the sub-band components at each scale are restored to the same length as the original ocean current velocity profile sequence, resulting in: ,in Representing the i Two-layer low-frequency approximation components of each sample; Representing the i High-frequency detail components of a single sample; Representing the i The second-level detail components of each sample; The different scale sub-band components mentioned above are used as multi-view feature representations to construct a multi-view feature set. ,in, V Indicates the number of views. This represents the feature matrix corresponding to the v-th view.

4. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 1, characterized in that, In step S30, a joint nonnegative matrix factorization model with weight constraints is established to perform collaborative decomposition of the multi-view features, specifically including: Calculate the energy proportion of each view feature and determine the corresponding weight, and calculate the energy value of the multi-view features obtained by discrete wavelet decomposition respectively: ,in Indicates the first v In the feature matrix of the nth view i The sample at the th j Values ​​at depth positions; The weights are determined based on the proportion of each view's feature energy to the total energy. , Representing the v The energy percentage of each view, and satisfying , .

5. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 1, characterized in that, In step S30, constructing the joint nonnegative matrix factorization model with weight constraints further includes: Input the feature matrices of each view into the joint nonnegative matrix factorization model, and perform the following decomposition for each view: ,in To share the basis matrix, For the first v The low-dimensional representation matrix corresponding to each view K Low-dimensional feature dimension; Under non-negativity constraints, satisfy .

6. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 5, characterized in that, The joint nonnegative matrix factorization model is subjected to collaborative optimization based on weighted reconstruction error. An objective function is constructed by minimizing the multi-view weighted reconstruction error, and a regularization term is introduced into the objective function: ,in It is the Frobenius norm. and This is the regularization parameter.

7. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 6, characterized in that, The objective function is iteratively optimized to achieve collaborative decomposition of the views under the constraint of a shared basis matrix, resulting in a low-dimensional latent representation shared by multiple views. W and the low-dimensional feature representations corresponding to each view. Shared basis matrix Common structures used to characterize ocean current velocity profiles at different scales; low-dimensional representation of each view. It is used to characterize the cross-sectional representation at the corresponding scale, thereby realizing the fusion representation of information at different scales.

8. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 1, characterized in that, In step S40, the low-dimensional features after collaborative decomposition are weighted and fused based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition. This fusion includes two methods: When the features of each sub-band are similar in overall trend and structural form, a weighted summation method is used for fusion to obtain a unified feature representation: When the features of each sub-band differ in structural morphology or local variation characteristics, a weighted splicing method is used for fusion to obtain a joint feature representation: ; The weighted summation result is merged. Or weighted splicing and fusion results Z As the final fusion feature representation.

9. The ocean current velocity profile pattern recognition method based on DWT and Joint-NMF as described in claim 1, characterized in that, Step S50 specifically includes: After standardizing the fused feature representations, the similarity between samples is calculated, and a similarity matrix is ​​constructed using radial basis functions. ,in It is a scale parameter. Indicates the first i One fused feature vector; Construct a degree matrix based on the similarity matrix. D , D The specific formula is as follows: ; Constructing a symmetric normalized Laplace matrix L : ,in Represents the identity matrix; For the normalized Laplace matrix L Find the minimum r The eigenvectors corresponding to the eigenvalues ​​constitute the feature matrix. U : ,in r Indicates the number of cluster categories; For the feature matrix U The spectral embedding matrix is ​​obtained by normalizing the rows. The K-means clustering method is used to divide the spectral embedding matrix into samples, and the category labels of each ocean current velocity profile sample are obtained, thereby obtaining the ocean current velocity profile pattern recognition results and realizing the automatic division of different ocean current structure patterns.

10. A current velocity profile pattern recognition system based on DWT and Joint-NMF, characterized in that, To implement the method of any one of claims 1 to 9, comprising: The normalization and matrix construction module is used to normalize ocean current velocity profile data and construct ocean current velocity profile sample matrices. The discrete wavelet decomposition module is used to perform discrete wavelet decomposition on the ocean current velocity profile sample matrix to obtain sub-band components at different scales. Each sub-band component is used as a multi-view feature to construct a multi-view feature set. The joint nonnegative matrix factorization module is used to establish a joint nonnegative matrix factorization model with weight constraints, and to perform collaborative decomposition of the feature matrices of each view in the multi-view feature set to obtain the low-dimensional representation matrix corresponding to each view. The weighted fusion module is used to perform weighted fusion of the low-dimensional representation matrices corresponding to each view after collaborative decomposition based on the pattern structure similarity of the features of each sub-band after discrete wavelet decomposition, so as to obtain the fused feature representation. The spectral clustering analysis module is used to perform spectral clustering analysis on the fused feature representation to obtain the pattern recognition results of ocean current velocity profiles.

Images