A multi-source data feature fusion method based on underwater

Abstract

The present invention relates to the field of data fusion technology and discloses a multi-source underwater data feature fusion method. This method obtains and preprocesses ocean acoustic data collected by multiple sensors. The preprocessed ocean acoustic data is then fed into a trained denoising autoencoder to extract a preliminary feature representation. An energy valley optimization algorithm is used to optimize the K-means initial clustering centers to obtain an optimal initial clustering center configuration. A K-means algorithm is then run based on the optimal initial clustering center configuration to perform cluster analysis on the preliminary feature representation. A feature fusion strategy is selected based on the clustering results to obtain a comprehensive feature representation. Based on the comprehensive feature representation, a classification algorithm is used to output a target recognition result. This method significantly improves the accuracy, robustness, and efficiency of underwater target recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-source data fusion, and in particular to an underwater multi-source data feature fusion method. Background Art

[0002] The complex and ever-changing nature of the ocean environment poses a significant challenge to underwater target recognition. Frequent background noise from ship activity further complicates the task, making it difficult for a single sensor to achieve ideal recognition accuracy and data processing efficiency in this environment. To address these challenges, underwater multi-source sensor fusion technology has been extensively researched and has achieved significant results.

[0003] Traditional approaches primarily focus on linear dimensionality reduction, manifold learning, and deep learning. Linear dimensionality reduction techniques such as principal component analysis (PCA) and linear discriminant analysis (LDA) extract low-dimensional features through projection transformations. However, their linear assumptions make it difficult to characterize nonlinear correlations between multi-source data, such as acoustic and optical data, limiting their ability to express features in complex ocean scenarios. Manifold learning methods (such as isometric mapping and local linear embedding (LLE)) capture nonlinear structures by constructing local neighborhood relationships within the data. However, their computational complexity increases exponentially with the amount of data and they are sensitive to noise, easily causing manifold distortion in the presence of background noise from ships. Furthermore, deep learning methods based on autoencoders (such as denoising autoencoders (DAEs)) learn robust features through unsupervised reconstruction. However, their performance is highly dependent on the quality of the input data. If heterogeneous data is directly fed without sufficient preprocessing, the network is susceptible to noise contamination or fuzzy information, making it difficult to reliably extract discriminative features. During the feature fusion stage, the traditional K-means algorithm, due to its sensitivity to initial cluster centers, tends to produce suboptimal solutions in complex data distributions, resulting in unstable feature grouping results and, in turn, affecting the reliability of the fusion strategy. For example, if the initial center falls into a noisy area, the clustering process may mistakenly identify ship noise as the target feature cluster, resulting in the fused features containing false information. More importantly, existing fusion strategies are mostly fixed rules (such as directly concatenating multimodal features or global PCA dimensionality reduction) that cannot be dynamically adjusted based on clustering quality. Forced fusion in low-quality clustering scenarios can introduce noise features, significantly reducing classifier accuracy.

[0004] Therefore, there is an urgent need for a multi-source data feature fusion method based on underwater to improve the target recognition accuracy in complex marine environments. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides an underwater multi-source data feature fusion method, which can improve the underwater target recognition accuracy, enhance the ability to resist ship noise, and realize efficient and robust multi-source data fusion in complex marine environments through an adaptive feature fusion mechanism.

[0006] The present invention provides a method for fusion of underwater multi-source data features, comprising the following steps:

[0007] S1. Obtain ocean acoustic data collected by multi-source sensors and perform preprocessing;

[0008] The ocean hydroacoustic data includes acoustic data, optical data, physical oceanographic data and geological exploration data;

[0009] S2, input the preprocessed ocean acoustic data into the trained denoising autoencoder to extract preliminary feature representation;

[0010] S3, using the energy valley optimization algorithm to optimize the K-means initial cluster center and obtain the optimal initial cluster center configuration;

[0011] S4. Running a K-means algorithm based on the optimal initial cluster center configuration, performing cluster analysis on the preliminary feature representation, and selecting a feature fusion strategy based on the clustering results to obtain a comprehensive feature representation;

[0012] S5. Output the target recognition result through a classification algorithm based on the comprehensive feature representation.

[0013] Furthermore, in S1, the preprocessing specifically includes:

[0014] Denoising and normalizing the acoustic data;

[0015] performing deblurring and color correction on the optical data;

[0016] The physical oceanographic data and the geological exploration data are temporally and spatially aligned.

[0017] Furthermore, the energy valley optimization algorithm is used to optimize the K-means initial cluster centers to obtain the optimal initial cluster center configuration, specifically including:

[0018] S31, setting the population size and the maximum number of iterations, and randomly generating multiple candidate initial cluster centers as the initial population;

[0019] S32, performing a particle splitting operation on a preset proportion of individuals in the population with the highest fitness ranking to generate child individuals; performing a particle merging operation on a preset proportion of individuals in the population with the lowest fitness ranking to generate new individuals; performing an energy decay operation on all individuals in the population to narrow the search range;

[0020] S33, sorting the individuals in the population according to the fitness evaluation index, and retaining the individuals with a preset proportion;

[0021] S34. Determine whether the maximum number of iterations has been reached. If so, output the current optimal initial cluster center configuration; if not, return to S32.

[0022] Furthermore, the K-means algorithm is run on the initial cluster center corresponding to each individual, and the sum of the intra-cluster distances is calculated as the fitness evaluation index.

[0023] Furthermore, the particle splitting operation includes applying Gaussian perturbation to each cluster center in the individual; and the particle merging operation includes performing weighted averaging on the cluster centers of two individuals.

[0024] Furthermore, the S4 specifically includes: S41, running the K-means algorithm based on the optimal initial cluster center configuration, performing cluster analysis on the preliminary feature representation, and obtaining cluster labels and cluster centers; S42, calculating the silhouette coefficient and the sum of the intra-cluster distances based on the cluster labels and cluster centers, and determining the cluster quality; S43, if the cluster quality meets the preset conditions, using the cluster center as part of the comprehensive feature representation; S44, if the cluster quality does not meet the preset conditions, performing principal component analysis dimensionality reduction on the feature grouping corresponding to each cluster label, and merging the reduced dimensionality features to generate the comprehensive feature representation.

[0025] Furthermore, the calculation formula of the silhouette coefficient is: ;

[0026] Among them, S N is the silhouette coefficient; s(i) represents the silhouette coefficient of sample point i; N represents the total number of samples.

[0027] The calculation formula for the sum of the intra-cluster distances is: ;

[0028] Where K is the number of clusters; WCSS is the sum of the distances within a cluster; WCSS k Represents a single cluster C k The intra-cluster distance and .

[0029] Furthermore, the calculation formula of the silhouette coefficient of the sample point i is: ;

[0030] Among them, a(i) represents the average distance between sample point i and other samples in the same cluster; b(i) represents the minimum average distance between sample point i and other clusters; C k Represents cluster C k The number of samples;

[0031] Furthermore, the single cluster C k The calculation formula of the sum of the intra-cluster distance is:

[0032] ;

[0033] in, Represents cluster C k the center of Represents sample x i The squared Euclidean distance to the cluster center.

[0034] Furthermore, in S43, taking the cluster center as a part of the comprehensive feature representation includes: splicing the cluster center and the preliminary feature representation according to the dimension;

[0035] In the S44, performing principal component analysis dimensionality reduction on the feature grouping corresponding to each cluster label includes: independently reducing the dimensionality of the feature representation of each group and retaining the principal component of the preset dimension.

[0036] The embodiments of the present invention have the following technical effects:

[0037] 1. The denoising autoencoder extracts preliminary feature representations from the raw data. Although these features have been denoised, they are still high-dimensional and may contain redundant information. Cluster analysis can group these preliminary feature representations into clusters with similar characteristics, reducing redundant information and revealing the underlying distribution patterns and structure of the data. This high-level grouping helps better understand the inherent relationships in the data.

[0038] 2. By using the Energy Valley Optimization (EVO) algorithm to optimize the initial cluster centers of K-Means, a global search is performed to find a more optimal initial configuration, avoiding local optimality and improving the accuracy and stability of clustering results. Traditional K-Means algorithms are very sensitive to the choice of initial cluster centers and are prone to falling into local optimal solutions, especially when data distribution is uneven or there is a lot of noise.

[0039] 3. Clustering allows for the design of specific feature fusion strategies for different clusters. For clusters with high clustering quality, the cluster centers are directly used as new features. For clusters with lower clustering quality, local dimensionality reduction techniques (such as PCA) are used to explore latent features, thereby enhancing their discriminability and expressiveness. Data within different clusters may represent different target categories or environmental characteristics, thus avoiding the information confusion caused by global fusion. The resulting comprehensive feature representation can better preserve key information and improve classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 This is a flowchart of a method for underwater multi-source data feature fusion provided by an embodiment of the present invention.

[0042] Figure 2 This is a flow chart of an energy valley optimization algorithm provided by an embodiment of the present invention.

[0043] Figure 3 The present invention provides an application of an underwater multi-source data feature fusion method in seabed shipwreck identification. DETAILED DESCRIPTION

[0044] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0045] Figure 1 This is a flow chart of a method for underwater multi-source data feature fusion provided by an embodiment of the present invention. Figure 1 , specifically including:

[0046] S1. Obtain ocean acoustic data collected by multi-source sensors and perform preprocessing;

[0047] Ocean hydroacoustic data includes acoustic data, optical data, physical oceanographic data, and geological exploration data.

[0048] In some embodiments, the acquisition of ocean acoustic data collected by multiple source sensors includes but is not limited to collecting data from sonar, optical sensors, temperature and salinity sensors, pressure sensors, magnetometers, and seismic exploration equipment.

[0049] The preprocessing stage is the foundation of the entire underwater multi-source data feature fusion method. The primary goal of this stage is to clean and standardize the data collected by different sensor types so that subsequent steps can more effectively extract features and perform analysis. The following preprocessing steps are specific to each type of data:

[0050] The acoustic data is denoised and normalized.

[0051] Specifically, denoising: Acoustic data often contains interference factors such as environmental noise and equipment noise. Denoising can be performed using a variety of technical means, such as frequency domain filtering (such as bandpass filtering to remove signals outside the frequency range of interest), time domain filtering (such as Kalman filtering for state estimation of dynamic systems), or automatic denoising using deep learning models such as convolutional neural networks (CNNs). Standardization: After denoising, the acoustic data needs to be standardized to ensure that all data has the same scale. Common practices include scaling the data to the [0, 1] range or calculating the Z score (i.e., subtracting the mean and dividing by the standard deviation), which helps improve the performance of subsequent algorithms.

[0052] The optical data is deblurred and color corrected.

[0053] Specifically, deblurring: Optical images may become unclear due to motion blur or other reasons. Image processing techniques such as Wiener filtering and blind deconvolution can be used to restore image quality. In addition, deep learning-based methods such as generative adversarial networks (GANs) can also be used to improve image clarity. Color correction: Due to changes in lighting conditions or differences in camera characteristics, the captured color information may be distorted. Color correction can be achieved through methods such as white balance adjustment and color mapping transformation, with the goal of making the colors in the image reflect the actual situation as realistically as possible.

[0054] The physical oceanographic data and the geological exploration data are temporally and spatially aligned.

[0055] Specifically, for physical oceanographic data (such as temperature and salinity distribution) and geological exploration data (such as seafloor topography), the first step is to determine the accuracy of their timestamps and geographic locations, and to ensure that these data correspond to each other in both time and space. If there are time delays or position deviations, they must be corrected through interpolation and extrapolation, so that different types of data can be compared and analyzed at the same time point and geographic coordinate system.

[0056] S2. Input the preprocessed ocean acoustic data into the trained denoising autoencoder to extract preliminary feature representation.

[0057] A denoising autoencoder is an unsupervised learning model designed to extract clean and representative feature representations from noisy data. Its basic structure consists of two parts: an encoder, which compresses the input data into a low-dimensional latent representation (i.e., feature representation) to preserve the data's core information; and a decoder, which reconstructs the low-dimensional latent representation back into the original data space, attempting to restore the original input as closely as possible.

[0058] The characteristic of DAE is that it artificially adds noise (such as Gaussian noise or random occlusion) to the input data, and then learns how to remove this noise and reconstruct the original data through training. This can enhance the model's ability to learn the inherent patterns of complex data.

[0059] In some embodiments, S2 specifically includes:

[0060] Preprocessed multi-source sensor data (such as acoustic, optical, physical oceanographic, and geological exploration data) is used as input. During the training phase, artificial noise (such as Gaussian white noise) can be added to the input data to simulate noise interference in real environments. This step is a key feature of DAE and improves the robustness of the model.

[0061] Build a denoising autoencoder model and design the network structure: Encoder: Typically composed of multiple fully connected layers or convolutional layers, gradually reducing the data dimension. Decoder: Symmetrical to the encoder, gradually restoring the data dimension to the original input size. Activation function: Common activation functions include ReLU (Rectified Linear Unit) and tanh, which are used to introduce nonlinear transformations. Output layer: Select the appropriate output format based on the input data type. For continuous data (such as acoustic signals), use a linear activation function; for image data (such as optical images), use a sigmoid or tanh activation function.

[0062] The mean square error (MSE) is used as the loss function to measure the difference between the reconstructed data and the original data. The loss function formula is:

[0063] ;

[0064] Among them, y i It is the original data; is the reconstructed data; M is the number of samples.

[0065] Model training: Training data is divided into training and validation sets to ensure the model's generalization capabilities. Noisy data is input, and the encoder generates a low-dimensional feature representation. The decoder attempts to reconstruct the original data from the low-dimensional feature representation. The error is calculated based on the loss function, and the model parameters are updated using the backpropagation algorithm.

[0066] After training is complete, the encoder is fixed and the decoder is removed. The preprocessed ocean acoustic data is fed into the trained encoder to generate a low-dimensional, preliminary feature representation. This feature representation effectively captures the essential information in the data while removing redundancy and noise.

[0067] It's worth noting that for different types of sensor data (such as acoustic and optical), you can train separate denoising autoencoders, or you can combine all the data and train a single model. If you train them independently, you can ultimately form a preliminary feature representation by concatenating the output features of each encoder.

[0068] S3, using the energy valley optimization algorithm to optimize the K-means initial cluster center and obtain the optimal initial cluster center configuration;

[0069] The Energy Valley Optimization Algorithm is a global optimization algorithm based on evolutionary principles. It gradually searches for the optimal solution by simulating the natural processes of particle splitting, merging, and energy decay. Its core concept is to optimize the objective function (fitness evaluation metric) by manipulating individuals in the population (such as splitting, merging, and energy decay).

[0070] In this solution, EVO is used to optimize the initial cluster center configuration of K-means, aiming to find a set of initial cluster centers that maximize K-means algorithm performance. This effectively overcomes the traditional K-means algorithm's sensitivity to initial values ​​and avoids the problem of randomly generated cluster centers easily falling into local optimal solutions.

[0071] In some embodiments, see Figure 2 , S3 specifically includes:

[0072] S31, setting the population size and the maximum number of iterations, and randomly generating multiple candidate initial cluster centers as the initial population;

[0073] It is worth noting that you also need to set a preset ratio to select the proportion of individuals to split, merge, and retain. For example, the top 20% of individuals are split, the bottom 20% are merged, and the top 50% are retained for the next generation.

[0074] S32, performing a particle splitting operation on a preset proportion of individuals in the population with the highest fitness ranking to generate child individuals; performing a particle merging operation on a preset proportion of individuals in the population with the lowest fitness ranking to generate new individuals; performing an energy decay operation on all individuals in the population to narrow the search range;

[0075] Optionally, the particle splitting operation includes applying Gaussian perturbation to each cluster center in the individual; and the particle merging operation includes performing weighted averaging on the cluster centers of two individuals.

[0076] For example,

[0077] Particle splitting operation:

[0078] Perform a split operation on the individuals ranked in the top 20% of fitness. Apply Gaussian perturbation to each cluster center to generate new sub-individuals. The formula is:

[0079] ;

[0080] Among them, x is the original cluster center, is Gaussian noise with mean 0 and standard deviation σ.

[0081] Particle merging operation:

[0082] Merge the individuals in the bottom 20% of the fitness ranking. Merge method: Take the weighted average of the cluster centers of two individuals to generate a new individual. The formula is:

[0083] ;

[0084] Among them, x new is a new individual, x1 and x2 are the cluster centers of the two individuals to be merged, and α is the weight coefficient (usually 0.5).

[0085] Energy decay operation:

[0086] Perform energy decay on all individuals to narrow the search range. The formula is:

[0087] ;

[0088] in, is the updated cluster center position after energy decay operation; x center is the center position of the current population, x is the cluster center position of the current individual; β is the decay factor (usually less than 1).

[0089] S33, sorting the individuals in the population according to the fitness evaluation index, and retaining the individuals with a preset proportion;

[0090] In some embodiments, a K-means algorithm is run on the initial cluster center corresponding to each individual, and the sum of the intra-cluster distances is calculated as a fitness evaluation indicator.

[0091] The individuals in the population are sorted according to their fitness values, and a preset proportion of excellent individuals are retained to enter the next generation population.

[0092] S34. Determine whether the maximum number of iterations has been reached. If so, output the current optimal initial cluster center configuration; if not, return to S32.

[0093] If the current number of iterations reaches Tmax, the current optimal initial cluster center configuration is output; otherwise, return to S32 to continue the evolution operation.

[0094] S4, running a K-means algorithm based on the optimal initial cluster center configuration, performing cluster analysis on the preliminary feature representation, selecting a feature fusion strategy based on the clustering results, and obtaining a comprehensive feature representation; specifically comprising: S41, running a K-means algorithm based on the optimal initial cluster center configuration, performing cluster analysis on the preliminary feature representation, and obtaining cluster labels and cluster centers;

[0095] Specifically, we used the optimal initial cluster center configuration obtained through the Energy Valley Optimization (EVO) algorithm in S3 to run the K-means algorithm. We performed cluster analysis on the preliminary feature representation data and obtained the cluster label for each sample point and the cluster center of each cluster.

[0096] S42, calculating the silhouette coefficient and the sum of the intra-cluster distances based on the cluster labels and cluster centers, and determining the clustering quality;

[0097] In some embodiments, the calculation formula of the silhouette coefficient is: ;

[0098] Among them, S N is the silhouette coefficient; s(i) represents the silhouette coefficient of sample point i; N represents the total number of samples.

[0099] Furthermore, the calculation formula of the silhouette coefficient of the sample point i is: ;

[0100] Among them, a(i) represents the average distance between sample point i and other samples in the same cluster; b(i) represents the minimum average distance between sample point i and other clusters; C k Represents cluster C k The number of samples;

[0101] In some embodiments, the calculation formula for the sum of the intra-cluster distances is:

[0102] ;

[0103] Where K is the number of clusters; WCSS is the sum of the distances within a cluster; WCSS k Represents a single cluster C k The intra-cluster distance and .

[0104] Furthermore, the single cluster C k The calculation formula of the sum of the intra-cluster distance is:

[0105] ;

[0106] in, Represents cluster C k the center of Represents sample x iThe squared Euclidean distance to the cluster center.

[0107] S43. If the clustering quality meets a preset condition, use the cluster center as part of the comprehensive feature representation;

[0108] If the clustering quality (assessed by the silhouette coefficient and the sum of intra-cluster distances) meets the pre-defined criteria, the current clustering result is considered reliable. The cluster centers of each cluster are concatenated with the preliminary feature representation by dimension to form a new comprehensive feature representation. For example, if the preliminary feature representation is an n-dimensional vector and the cluster centers have m dimensions, the final comprehensive feature representation will be an n+m-dimensional vector.

[0109] S44: If the clustering quality does not meet the preset condition, perform principal component analysis and dimensionality reduction on the feature grouping corresponding to each cluster label, and merge the features after dimensionality reduction to generate the comprehensive feature representation.

[0110] If the clustering quality does not meet the preset criteria, further processing is required to improve the quality of the feature representation. Principal component analysis (PCA) is performed independently on the feature groups corresponding to each cluster label, retaining the principal components of the preset dimensionality. PCA is a dimensionality reduction technique that converts raw high-dimensional data into low-dimensional data through a linear transformation while preserving the data's key information as much as possible. After performing PCA dimensionality reduction on each group, all reduced features are combined to generate a comprehensive feature representation.

[0111] Optionally, the preset condition is set to evaluate whether the clustering quality meets the expected standard, thereby deciding whether to adopt the current clustering result. In the present invention, clustering quality is evaluated by two indicators: silhouette coefficient and sum of intra-cluster distances.

[0112] Silhouette coefficient S N The value range is [-1, 1]. A value close to 1 indicates that the sample is very similar to other samples in the same cluster and significantly different from other clusters, indicating a good clustering effect. A value close to 0 indicates that the sample may be at the edge of the cluster, indicating a moderate clustering effect. A negative value indicates that the sample may have been incorrectly assigned to the wrong cluster.

[0113] The intra-cluster distance sum (WCSS) represents the sum of the squared distances from each cluster's internal sample points to its center. A smaller value indicates a closer cluster. WCSS generally decreases as the number of clusters (K) increases, but the rate of decrease slows down.

[0114] The following are the specific preset conditions for these two indicators:

[0115] 1. Pre-conditions for comprehensive silhouette coefficient and WCSS:

[0116] Set two independent thresholds S Nmin and WCSSmax ,Only when the two indicators are met at the same time, the clustering quality is considered to meet the preset conditions.

[0117] 2. Weighted scoring method:

[0118] The silhouette coefficient and WCSS are weighted and scored to comprehensively evaluate the clustering quality;

[0119] For example, define a comprehensive scoring function:

[0120] ;

[0121] w1 and w2 are weight coefficients used to balance the importance of the two indicators.

[0122] Set a comprehensive score threshold Score min , if the calculated score is greater than or equal to Score min , the clustering quality is considered to meet the preset conditions.

[0123] Clustering quality is assessed by calculating the silhouette coefficient and summed intra-cluster distances based on cluster labels and cluster centers. This determines whether to use the cluster centers as part of a comprehensive feature representation or to perform feature fusion using PCA dimensionality reduction. This strategy considers both the effectiveness of the clustering results and the control of feature dimensionality, helping to improve both the effectiveness of the feature representation and computational efficiency. The comprehensive feature representation derived from optimized cluster analysis can better describe the essential characteristics of the sample, significantly improving the classifier's precision and recall when using these features for target recognition.

[0124] S5. Output the target recognition result through a classification algorithm based on the comprehensive feature representation.

[0125] Using comprehensive feature representation as input, the target recognition results are output through advanced classification algorithms (such as support vector machine SVM, random forest RF, etc.).

[0126] The comprehensive feature representation obtained after the previous steps (S1 to S4) has been preprocessed, denoised, and extracted using an autoencoder. K-means clustering analysis is optimized, and an appropriate feature fusion strategy is selected based on clustering quality. Depending on the specific application scenario and requirements, different classification algorithms can be used for object recognition.

[0127] The present invention starts with acquiring ocean hydroacoustic data collected by multi-source sensors, and then undergoes a series of data preprocessing, feature extraction, optimized cluster analysis and feature fusion, and finally uses a classification algorithm to output target recognition results. This can improve the accuracy and efficiency of data processing, thereby enhancing the understanding and monitoring capabilities of the marine environment.

[0128] For example, identifying submarine shipwrecks in complex marine environments, such as Figure 3 As shown, sensors from different countries collect data: a sonar array captures seafloor topography and sound waves reflected from objects. An optical camera captures images of the seafloor (blurred by suspended particles). A magnetometer detects magnetic field anomalies from metallic materials (possibly iron components of a shipwreck). Temperature and salinity sensors record environmental parameters to help eliminate hydrothermal interference.

[0129] Preprocessing: Acoustic data: Frequency-domain filtering is used to remove ship engine noise and normalize the data to a uniform decibel range. Optical images: A blind deconvolution algorithm is used to restore the shipwreck's outline and correct for the blue tint caused by water depth. Multi-source alignment: Magnetometer data is aligned with sonar coordinates to ensure spatial and temporal matching of data within the same area.

[0130] Optimize the initial cluster centers using Energy Valley Optimization (EVO): Set K = 3 (intended targets: shipwrecks, rocks, and hydrothermal vents). EVO generates 20 candidate initial centers. After 10 iterations of splitting (perturbing the shipwreck characteristics) and merging (merging similar centers of rocks and hydrothermal vents), the optimal initial center is found. This prevents K-means from misclassifying the shipwreck into two clusters due to random initialization.

[0131] Cluster analysis and feature fusion. Cluster quality assessment: The calculated silhouette coefficient is 0.65 (> the preset threshold of 0.6), and the sum of intra-cluster distances is lower than expected, indicating that the clustering is valid. Feature fusion strategy: The three cluster centers (shipwreck, rock, and hydrothermal) are added as new features and combined with the preliminary features to form a 173-dimensional comprehensive feature.

[0132] Classifier training: Use random forest to train historical annotated data (200 groups of shipwrecks and 500 groups of non-shipwrecks), input 173-dimensional comprehensive features, and output shipwreck identification results.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source data feature fusion method based on underwater, characterized in that: The steps include: S1. Obtain ocean acoustic data collected by multi-source sensors and perform preprocessing; The ocean hydroacoustic data includes acoustic data, optical data, physical oceanographic data and geological exploration data; The physical oceanographic data include: seawater temperature and salinity distribution data; the geological exploration data include: seabed topography; S2, input the preprocessed ocean acoustic data into the trained denoising autoencoder to extract preliminary feature representation; S3. Use the energy valley optimization algorithm to optimize the K-means initial cluster center to obtain the optimal initial cluster center configuration, specifically including: S31, setting the population size and the maximum number of iterations, and randomly generating multiple candidate initial cluster centers as the initial population; S32, performing a particle splitting operation on a preset proportion of individuals in the population with the highest fitness ranking to generate child individuals; performing a particle merging operation on a preset proportion of individuals in the population with the lowest fitness ranking to generate new individuals; performing an energy decay operation on all individuals in the population to narrow the search range; S33, sorting the individuals in the population according to the fitness evaluation index, and retaining the individuals with a preset proportion; S34, determining whether the maximum number of iterations has been reached, if so, outputting the current optimal initial cluster center configuration; if not, returning to S32; S4. Running a K-means algorithm based on the optimal initial cluster center configuration, performing cluster analysis on the preliminary feature representation, and selecting a feature fusion strategy based on the clustering results to obtain a comprehensive feature representation; S5. Output the target recognition result through a classification algorithm based on the comprehensive feature representation.

2. The underwater multi-source data feature fusion method according to claim 1, characterized in that: In S1, the preprocessing specifically includes: Denoising and normalizing the acoustic data; performing deblurring and color correction on the optical data; The physical oceanographic data and the geological exploration data are temporally and spatially aligned.

3. The underwater multi-source data feature fusion method according to claim 1, characterized in that: The K-means algorithm is run on the initial cluster center corresponding to each individual, and the sum of the intra-cluster distances is calculated as the fitness evaluation index.

4. The underwater multi-source data feature fusion method according to claim 1, characterized in that: The particle splitting operation includes applying Gaussian perturbation to each cluster center in an individual; and the particle merging operation includes performing weighted averaging on the cluster centers of two individuals.

5. The underwater multi-source data feature fusion method according to claim 1, characterized in that: Said S4 specifically includes: S41, running a K-means algorithm based on the optimal initial cluster center configuration, performing cluster analysis on the preliminary feature representation, and obtaining cluster labels and cluster centers; S42, calculating the silhouette coefficient and the sum of the intra-cluster distances based on the cluster labels and cluster centers, and determining the clustering quality; S43. If the clustering quality meets a preset condition, use the cluster center as part of the comprehensive feature representation; S44: If the clustering quality does not meet the preset condition, perform principal component analysis and dimensionality reduction on the feature grouping corresponding to each cluster label, and merge the features after dimensionality reduction to generate the comprehensive feature representation.

6. The underwater multi-source data feature fusion method according to claim 5, characterized in that: The calculation formula of the silhouette coefficient is: ; Among them, S N is the silhouette coefficient; s(i) represents the silhouette coefficient of sample point i; N represents the total number of samples; The calculation formula for the sum of the intra-cluster distances is: ; Where K is the number of clusters; WCSS is the sum of the distances within a cluster; WCSS k Represents a single cluster C k The intra-cluster distance and .

7. The underwater multi-source data feature fusion method according to claim 6, characterized in that: The calculation formula of the silhouette coefficient of the sample point i is: ; Where a(i) represents the average distance from sample point i to other samples in the same cluster; b(i) represents the minimum average distance from sample point i to other clusters; C k Represents cluster C k The number of samples.

8. The underwater multi-source data feature fusion method according to claim 6, characterized in that: The single cluster C k The calculation formula of the sum of the intra-cluster distance is: ； in, Represents cluster C k the center of Represents sample x i The squared Euclidean distance to the cluster center.

9. The underwater multi-source data feature fusion method according to claim 5 is characterized in that: In said S43, taking said cluster center as a part of the comprehensive feature representation includes: splicing said cluster center and said preliminary feature representation according to dimension; In the S44, performing principal component analysis dimensionality reduction on the feature grouping corresponding to each cluster label includes: independently reducing the dimensionality of the feature representation of each group and retaining the principal component of the preset dimension.

Images