Salinity drift correction method and system for buoy observation data

Abstract

The invention relates to the technical field of data processing, and discloses a salinity drift correction method and system for buoy observation data. The method comprises the steps that historical buoy observation data and current environment parameters are collected and preprocessed, and a multi-source heterogeneous data set is obtained; performing multi-dimensional environmental factor analysis on the multi-source heterogeneous data set to obtain a salinity influence factor weight; drifting detection and quantification are carried out on the buoy salinity data through the salinity influence factor weight, and an accumulated deviation integral value is obtained; performing multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data; performing dynamic space-time weight optimization on the preliminary correction data to obtain optimized correction data; and performing real-time performance evaluation and automatic threshold adjustment on the optimized correction data to obtain target correction data and a parameter updating strategy. According to the invention, the accuracy of salinity drift correction for buoy observation data is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a method and system for correcting salinity drift of buoy observation data. Background Art

[0002] As an important means of obtaining marine environmental data, the ocean buoy observation system plays a key role in global climate change research, marine resource development, and marine environmental protection. However, buoy sensors immersed in seawater for a long time will be affected by factors such as biological attachment, chemical corrosion, and mechanical wear, causing the measurement data to drift. Traditional salinity drift correction methods mainly rely on regular manual calibration and simple statistical models such as linear regression or moving average. Although these methods can correct salinity drift to a certain extent, they are difficult to adapt to the complex and changeable marine environment and different types of drift patterns.

[0003] The existing technology has the following shortcomings: first, traditional methods have difficulty in capturing the nonlinear and multi-scale characteristics of salinity drift, resulting in low correction accuracy; second, existing methods often ignore the temporal and spatial correlation and cannot fully utilize the useful information in adjacent buoys and historical data; third, the current correction algorithm lacks adaptive capabilities and is difficult to cope with emerging drift patterns and sudden environmental changes; finally, existing methods are inefficient in processing large-scale multi-source heterogeneous data and are difficult to meet the needs of real-time correction. These problems seriously restrict the quality and reliability of ocean observation data and affect the accuracy of marine scientific research and practical applications. Summary of the invention

[0004] The present application provides a method and system for correcting salinity drift of buoy observation data, which are used to improve the accuracy of salinity drift correction for buoy observation data.

[0005] In a first aspect, the present application provides a salinity drift correction method for buoy observation data, the salinity drift correction method for buoy observation data comprising: collecting and preprocessing historical buoy observation data and current environmental parameters to obtain a multi-source heterogeneous data set; performing multi-dimensional environmental factor analysis on the multi-source heterogeneous data set to obtain salinity influencing factor weights; performing drift detection and quantification on the buoy salinity data using the salinity influencing factor weights to obtain a cumulative deviation integral value; performing multi-scale correction calculations on the cumulative deviation integral value to obtain preliminary correction data; performing dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data; performing real-time performance evaluation and automatic threshold adjustment on the optimized correction data to obtain target correction data and parameter update strategy.

[0006] In combination with the first aspect, in a first implementation method of the first aspect of the present application, the historical buoy observation data and the current environmental parameters are collected and preprocessed to obtain a multi-source heterogeneous data set, including: timestamp alignment of the historical buoy observation data and the current environmental parameters to obtain time synchronization data; spatial gridding of the time synchronization data to obtain gridded observation data; wavelet transform denoising of the gridded observation data to obtain denoised observation data; missing value interpolation of the denoised observation data to obtain a complete observation data set; principal component analysis of the complete observation data set to obtain a reduced dimension feature set; outlier detection based on the interquartile range method on the reduced dimension feature set to obtain abnormal labeled data; abnormal sample elimination of the abnormal labeled data to obtain a cleaned data set; format unification of the cleaned data set to obtain a standardized multi-source heterogeneous data set; storing the standardized multi-source heterogeneous data set in a distributed database to obtain an indexed data set; data quality scoring of the indexed data set to obtain a multi-source heterogeneous data set with quality labels.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present application, the multi-dimensional environmental factor analysis is performed on the multi-source heterogeneous data set to obtain the salinity influencing factor weights, including: performing time series decomposition on the multi-source heterogeneous data set to obtain trend components; performing sliding window correlation analysis on the trend components to obtain a dynamic correlation coefficient matrix; performing spectral clustering on the dynamic correlation coefficient matrix to obtain environmental factor clustering results; performing principal component regression analysis on the environmental factor clustering results to obtain preliminary factor weights; applying Bootstrap resampling to the preliminary factor weights to obtain weight confidence intervals; fuzzifying the weight confidence intervals through an adaptive fuzzy inference system to obtain a fuzzy weight set; performing a hierarchical analysis process on the fuzzy weight set to obtain a hierarchical weight structure; applying cross-validation to the hierarchical weight structure to obtain validation weights; quantifying the contribution of the validation weights to obtain factor importance rankings; and performing nonlinear optimization on the factor importance rankings to obtain the salinity influencing factor weights.

[0008] In combination with the first aspect, in a third implementation method of the first aspect of the present application, the drift detection and quantification of the buoy salinity data is performed through the salinity influencing factor weight to obtain a cumulative deviation integral value, including: performing a sliding time window segmentation on the buoy salinity data to obtain time series segments; performing wavelet multi-resolution analysis on the time series segments to obtain a multi-scale decomposition result; performing trend extraction on the multi-scale decomposition result through empirical mode decomposition to obtain a drift trend curve; performing piecewise linear fitting on the drift trend curve to obtain a piecewise slope sequence; performing adaptive threshold segmentation on the piecewise slope sequence to obtain a drift event marker; performing pattern matching on the drift event marker to obtain a drift type identification result; performing cumulative deviation calculation on the drift type identification result to obtain an original cumulative deviation value; and smoothing the original cumulative deviation value to obtain the cumulative deviation integral value.

[0009] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, the multi-scale correction calculation is performed on the cumulative deviation integral value to obtain preliminary correction data, including: performing wavelet packet decomposition on the cumulative deviation integral value to obtain a multi-scale coefficient set; performing soft threshold processing on the multi-scale coefficient set through an adaptive threshold to obtain a denoising coefficient set; performing inverse wavelet packet transform on the denoising coefficient set to obtain a reconstructed deviation sequence; performing piecewise polynomial fitting on the reconstructed deviation sequence to obtain a correction function; smoothing the correction function through a sliding median filter to obtain a smoothed correction curve; performing a differential operation on the original salinity data and the smoothed correction curve to obtain a corrected salinity value; performing residual drift extraction on the corrected salinity value through variational mode decomposition to obtain the preliminary correction data.

[0010] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, the dynamic spatiotemporal weight optimization of the preliminary correction data to obtain optimized correction data includes: performing spatiotemporal grid division on the preliminary correction data to obtain gridded correction data; performing spatial autocorrelation analysis on the gridded correction data to obtain a local spatial dependence index; performing probability distribution estimation on the local spatial dependence index through variational Bayesian inference to obtain a spatial correlation model; performing dynamic spatial weight matching on the spatial correlation model to obtain a dynamic spatial weight matrix; performing time series decomposition on the preliminary correction data to obtain seasonal components; performing local polynomial regression on the seasonal components to obtain a time correlation curve; performing exponential smoothing on the time correlation curve to obtain a dynamic time weight vector; performing a tensor product operation on the dynamic spatial weight matrix and the dynamic time weight vector to obtain a spatiotemporal joint weight tensor; performing weighted average calculation on the preliminary correction data and the spatiotemporal joint weight tensor to obtain weight correction data; and performing edge-preserving smoothing on the weight correction data through an anisotropic diffusion filtering algorithm to obtain optimized correction data.

[0011] In combination with the first aspect, in a sixth implementation method of the first aspect of the present application, the real-time performance evaluation and automatic threshold adjustment of the optimized correction data to obtain the target correction data and the parameter update strategy include: performing multi-dimensional spectrum analysis on the optimized correction data to obtain a frequency domain feature matrix; performing feature extraction on the frequency domain feature matrix to obtain the main feature vector; performing cluster analysis on the main feature vector to obtain data pattern grouping; performing time series correlation calculation on the data pattern grouping by a sliding window method to obtain a dynamic association matrix; based on the dynamic association matrix, constructing a multi-layer threshold network by an adaptive threshold algorithm to obtain a threshold candidate set; cross-validating the threshold candidate set to obtain an optimal threshold combination; performing continuous processing on the optimal threshold combination by an interpolation method to obtain a smooth threshold curve; based on the smooth threshold curve, performing multi-indicator evaluation on the correction effect to obtain a comprehensive performance score; performing time series prediction on the comprehensive performance score to obtain a performance trend prediction; according to the performance trend prediction, adjusting the correction parameters by an iterative optimization algorithm to obtain the target correction data and the parameter update strategy.

[0012] In a second aspect, the present application provides a salinity drift correction system for buoy observation data, the salinity drift correction system for buoy observation data comprising:

[0013] The acquisition module is used to collect and preprocess historical buoy observation data and current environmental parameters to obtain multi-source heterogeneous data sets;

[0014] An analysis module, used for performing multi-dimensional environmental factor analysis on the multi-source heterogeneous data set to obtain the weight of the salinity influencing factor;

[0015] A quantification module, used to perform drift detection and quantification on the buoy salinity data by using the salinity influencing factor weight to obtain a cumulative deviation integral value;

[0016] A correction module, used for performing multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data;

[0017] An optimization module, used for performing dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data;

[0018] The evaluation module is used to perform real-time performance evaluation and automatic threshold adjustment on the optimized correction data to obtain target correction data and parameter update strategy.

[0019] In the technical solution provided by the present application, by introducing the preprocessing of multi-source heterogeneous data sets and multi-dimensional environmental factor analysis, this method greatly improves the accuracy of the weights of salinity influencing factors, thereby laying a solid foundation for subsequent drift detection and correction. Secondly, the cumulative deviation integral value is used as a drift quantification indicator, combined with the sliding time window and wavelet multi-resolution analysis technology, so that this method can more accurately capture the drift characteristics on different time scales, significantly improving the sensitivity and accuracy of drift detection. Furthermore, the introduction of multi-scale correction calculations, especially the combination of wavelet packet decomposition and adaptive threshold processing, enables this method to effectively process drift components of different frequencies and amplitudes, greatly improving the accuracy and robustness of the correction. In addition, the application of the dynamic spatiotemporal weight optimization strategy, by integrating spatial autocorrelation analysis and time series decomposition technology, fully utilizes the spatiotemporal correlation of buoy observation data, and significantly enhances the spatial consistency and temporal continuity of the correction results. Finally, the introduction of real-time performance evaluation and automatic threshold adjustment mechanism enables the method to have the ability of adaptive learning and continuous optimization, and can effectively respond to the dynamic changes of the marine environment and emerging drift patterns, ensuring the long-term effectiveness and reliability of the correction system. The method of the present invention has achieved significant improvements in processing speed, correction accuracy and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0021] Figure 1 A schematic diagram of an embodiment of a method for correcting salinity drift of buoy observation data in an embodiment of the present application;

[0022] Figure 2 Schematic diagram of an embodiment of a salinity drift correction system for buoy observation data in an embodiment of the present application. DETAILED DESCRIPTION

[0023] Embodiments of the present application provide a method and system for correcting salinity drift of buoy observation data. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the method for correcting salinity drift of buoy observation data in the embodiment of the present application includes:

[0025] Step S101, collecting and preprocessing historical buoy observation data and current environmental parameters to obtain a multi-source heterogeneous data set;

[0026] Step S102, performing multidimensional environmental factor analysis on multi-source heterogeneous data sets to obtain salinity influencing factor weights;

[0027] Step S103, performing drift detection and quantification on the buoy salinity data by using the salinity influencing factor weight to obtain a cumulative deviation integral value;

[0028] Step S104, performing multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data;

[0029] Step S105, performing dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data;

[0030] Step S106: Perform real-time performance evaluation and automatic threshold adjustment on the optimized correction data to obtain target correction data and parameter update strategy.

[0031] It is understandable that the execution subject of the present application may be a salinity drift correction system for buoy observation data, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0032] Specifically, historical buoy observation data and current environmental parameters are collected and preprocessed to obtain a multi-source heterogeneous data set. In this process, the collected data include buoy ID, observation time, location, depth, original salinity, temperature, pressure and other parameters. At the same time, satellite remote sensing data and shipborne CTD data are also collected as references. In the preprocessing stage, the sliding time window technology is used to synchronize the data in time, and then the spatial standardization of the data is achieved through spatial grid division. Next, the wavelet transform denoising algorithm is applied to reduce the random noise in the data, and the missing values ​​are filled by the interpolation algorithm. Subsequently, the principal component analysis is used to reduce the data dimension, and the interquartile range method and the isolation forest algorithm are used to detect and remove outliers. Finally, the processed data is stored in a distributed database to form a standardized multi-source heterogeneous data set. Multidimensional environmental factor analysis is performed on the multi-source heterogeneous data set to obtain the weights of salinity influencing factors. This step first uses the time series decomposition technique to divide the data into trend, seasonal and random components. Then, the dynamic correlation coefficient matrix is ​​calculated by sliding window correlation analysis, and the environmental factors are clustered using the spectral clustering algorithm. Then, the principal component regression analysis was applied to obtain the preliminary factor weights, and the confidence intervals of the weights were calculated by the Bootstrap resampling technique. Subsequently, the weights were fuzzified using an adaptive fuzzy inference system, and a hierarchical weight structure was constructed through a hierarchical analysis process. Finally, the Shapley value calculation was applied to quantify the contribution of each factor, and the final salinity influencing factor weights were obtained through nonlinear optimization.

[0033] Then, the drift detection and quantification of the buoy salinity data are carried out by the weight of the salinity influencing factor to obtain the cumulative deviation integral value. In this process, the salinity data is first segmented into time series segments using the sliding time window technique, and each segment is multi-scale decomposed by wavelet multi-resolution analysis. Then, the drift trend curve is extracted by applying the empirical mode decomposition, and the segmented slope sequence is obtained by piecewise linear fitting. Then, the drift events are identified by using the adaptive threshold segmentation algorithm, and the drift pattern matching is performed by the dynamic time warping algorithm. Finally, the cumulative deviation is calculated, and the result is smoothed by the kernel density estimation method to obtain the final cumulative deviation integral value. The cumulative deviation integral value is multi-scale corrected to obtain the preliminary correction data. In this step, the cumulative deviation integral value is first decomposed into multiple scale coefficient sets using wavelet packet decomposition, and then soft threshold processing is performed by adaptive threshold to remove noise. Then, the deviation sequence is reconstructed by applying the inverse wavelet packet transform, and the correction function is obtained by piecewise polynomial fitting. Subsequently, the correction function is smoothed by sliding median filtering to obtain a smooth correction curve. Finally, the original salinity data and the smoothed correction curve are differentially operated, and the residual drift is extracted by variational mode decomposition to obtain the preliminary correction data. The preliminary correction data are dynamically optimized in time and space to obtain the optimized correction data. This process first divides the data into time and space grids, and then calculates the local spatial dependence index through spatial autocorrelation analysis. Next, the spatial correlation model is estimated using variational Bayesian inference, and the dynamic spatial weight matrix is ​​obtained by applying the adaptive kernel function. At the same time, the preliminary correction data are decomposed into time series, and the time correlation curve is obtained by local polynomial regression of the dynamic time window. Then, the time correlation curve is exponentially smoothed to obtain the dynamic time weight vector. Finally, the spatial weight matrix and the time weight vector are tensor-producted to obtain the spatiotemporal joint weight tensor, and the optimized correction data are obtained by weighted average calculation and anisotropic diffusion filtering.

[0034] Finally, the optimized correction data is subjected to real-time performance evaluation and automatic threshold adjustment to obtain the target correction data and parameter update strategy. This step includes calculating evaluation indicators such as the root mean square error, temporal consistency index, and spatial consistency index before and after correction. Then, the Bayesian optimization algorithm is used to automatically adjust the key threshold parameters in each step. Next, the autoencoder is used to identify new abnormal drift patterns and trigger the model update process. Finally, the incremental learning framework is used to gradually integrate new observation data and update the model parameters.

[0035] In the embodiment of the present application, by introducing the preprocessing of multi-source heterogeneous data sets and multi-dimensional environmental factor analysis, this method greatly improves the accuracy of the weights of salinity influencing factors, thereby laying a solid foundation for subsequent drift detection and correction. Secondly, the cumulative deviation integral value is used as a drift quantification indicator, combined with the sliding time window and wavelet multi-resolution analysis technology, so that the method can more accurately capture the drift characteristics on different time scales, significantly improving the sensitivity and accuracy of drift detection. Furthermore, the introduction of multi-scale correction calculations, especially the combination of wavelet packet decomposition and adaptive threshold processing, enables this method to effectively process drift components of different frequencies and amplitudes, greatly improving the accuracy and robustness of the correction. In addition, the application of the dynamic spatiotemporal weight optimization strategy, by integrating spatial autocorrelation analysis and time series decomposition technology, fully utilizes the spatiotemporal correlation of buoy observation data, and significantly enhances the spatial consistency and temporal continuity of the correction results. Finally, the introduction of real-time performance evaluation and automatic threshold adjustment mechanism enables the method to have the ability of adaptive learning and continuous optimization, and can effectively respond to the dynamic changes of the marine environment and emerging drift patterns, ensuring the long-term effectiveness and reliability of the correction system. The method of the present invention has achieved significant improvements in processing speed, correction accuracy and adaptability.

[0036] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0037] (1) Align the timestamps of historical buoy observation data and current environmental parameters to obtain time-synchronized data;

[0038] (2) spatially grid the time-synchronized data to obtain gridded observation data;

[0039] (3) Perform wavelet transform denoising on the gridded observation data to obtain denoised observation data;

[0040] (4) interpolate missing values ​​of the denoised observation data to obtain a complete observation data set;

[0041] (5) Perform principal component analysis on the complete observation data set to obtain a reduced dimension feature set;

[0042] (6) Perform outlier detection based on the interquartile range method on the reduced dimension feature set to obtain abnormal labeled data;

[0043] (7) Remove abnormal samples from abnormal labeled data to obtain a cleaned data set;

[0044] (8) The cleaned data set is formatted in a unified manner to obtain a standardized multi-source heterogeneous data set;

[0045] (9) Storing the standardized multi-source heterogeneous data sets into a distributed database to obtain an indexed data set;

[0046] (10) Perform data quality scoring on the indexed dataset to obtain a multi-source heterogeneous dataset with quality labels.

[0047] Specifically, the timestamps of historical buoy observation data and current environmental parameters are aligned to obtain time-synchronized data. This process unifies data of different time resolutions to the same time scale through a linear interpolation algorithm to ensure that all data points have the same time interval. For example, hourly buoy data and daily satellite data are unified to the hourly level. The time-synchronized data is spatially gridded to obtain gridded observation data. This step uses an equidistant grid method or an adaptive grid method to divide the observation area into several grid cells, and the data points in each grid cell are assigned the same spatial coordinates. In this way, data of different spatial resolutions can be unified to the same spatial scale, which is convenient for subsequent analysis.

[0048] Then, the gridded observation data is subjected to wavelet transform denoising to obtain the denoised observation data. The wavelet transform denoising process includes: selecting a suitable wavelet basis function (such as Daubechies wavelet), performing wavelet decomposition on the data, setting a threshold to process the wavelet coefficients (such as soft threshold or hard threshold method), and finally performing wavelet reconstruction. This step can effectively remove high-frequency noise in the data while retaining the important features of the signal.

[0049] Perform missing value interpolation on the denoised observation data to obtain a complete observation data set. Missing value interpolation uses multiple interpolation methods, including a combination of mean interpolation, regression interpolation, and multivariate interpolation. For time series data, time series interpolation methods such as ARIMA model interpolation can also be used. Perform principal component analysis on the complete observation data set to obtain a reduced dimension feature set. The steps of principal component analysis include: calculating the data covariance matrix, solving eigenvalues ​​and eigenvectors, selecting principal components, and projecting data into principal component space. By retaining the first few principal components with a large proportion of explained variance, data dimensionality reduction is achieved while retaining most of the information.

[0050] Perform outlier detection based on the interquartile range method on the reduced dimension feature set to obtain abnormal labeled data. The interquartile range method first calculates the first quartile (Q1) and the third quartile (Q3) of the data, and then calculates the interquartile range IQR = Q3-Q1. Data points that are less than Q1-1.5IQR or greater than Q3+1.5IQR are marked as potential outliers. Abnormal samples are removed from the abnormal labeled data to obtain a cleaned data set. This step determines the final outliers by setting a threshold or using a statistical test method (such as the Grubbs test) and removes them from the data set.

[0051] The cleaned data set is formatted in a unified way to obtain a standardized multi-source heterogeneous data set. Format unification includes operations such as data type conversion, unit unification, and naming standardization to ensure that data from different sources have consistent formats and structures.

[0052] Store standardized multi-source heterogeneous data sets in a distributed database to obtain an indexed data set. Distributed databases (such as Apache Cassandra or MongoDB) can efficiently store and retrieve large-scale data. By establishing appropriate indexes (such as time index and spatial index), data query efficiency can be significantly improved.

[0053] Finally, the indexed data set is scored for data quality to obtain a multi-source heterogeneous data set with quality labels. Data quality scoring is based on multiple indicators, including completeness, accuracy, consistency, and timeliness. The scoring method can use weighted average or machine learning models (such as random forests) to integrate various indicators and ultimately assign a quality score to each piece of data.

[0054] For example, a marine observation project deployed 100 buoys, each of which recorded salinity data once an hour. During the one-month observation period, a total of 72,000 raw data records were collected (100 buoys × 24 hours × 30 days). First, all data were unified to the hour by timestamp alignment. Spatial grid division divided the observation area into 10×10 grids. Wavelet transform denoising used db4 wavelet, 3-level decomposition, and soft thresholding to reduce the noise level of the data. Missing value interpolation found that 2% of the data was missing, which was successfully interpolated using the ARIMA model. Principal component analysis retained the first 5 principal components, explaining 95% of the total variance. The interquartile range method detected 500 potential outliers, and 200 outlier samples were confirmed and eliminated after the Grubbs test. Data format unification unified the salinity unit to PSU (practical salinity unit). Standardized data was stored in the Apache Cassandra database, and a time and space composite index was established. Finally, the data quality score was scored on a scale of 0-100, taking into account the completeness (weight 0.3), accuracy (weight 0.4), consistency (weight 0.2) and timeliness (weight 0.1) of the data, and the average quality score was 85. This series of processing improved the availability and reliability of the data and laid a solid foundation for the subsequent salinity drift correction.

[0055] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0056] (1) Perform time series decomposition on multi-source heterogeneous data sets to obtain trend components;

[0057] (2) Perform sliding window correlation analysis on the trend component to obtain a dynamic correlation coefficient matrix;

[0058] (3) Perform spectral clustering on the dynamic correlation coefficient matrix to obtain the environmental factor clustering results;

[0059] (4) Conduct principal component regression analysis on the clustering results of environmental factors to obtain preliminary factor weights;

[0060] (5) Apply Bootstrap resampling to the preliminary factor weights to obtain the weight confidence intervals;

[0061] (6) Fuzzifying the weight confidence interval through an adaptive fuzzy inference system to obtain a fuzzy weight set;

[0062] (7) performing a hierarchical analysis process on the fuzzy weight set to obtain a hierarchical weight structure;

[0063] (8) Apply cross validation to the hierarchical weight structure to obtain validation weights;

[0064] (9) Quantify the contribution of the verification weight and obtain the factor importance ranking;

[0065] (10) Nonlinear optimization is performed on the ranking of factor importance to obtain the weights of salinity influencing factors.

[0066] Specifically, in the salinity drift correction method for buoy observation data, the multi-source heterogeneous data set is first decomposed into time series to obtain the trend component. This process uses the STL (Seasonal and Trend decomposition using Loess) method to decompose the time series data into three parts: trend, seasonality, and residual. The STL method estimates the trend and seasonal components by iteratively applying local weighted regression (LOESS), and finally obtains a smooth trend component.

[0067] The trend component is subjected to sliding window correlation analysis to obtain a dynamic correlation coefficient matrix. The sliding window correlation analysis calculates the Pearson correlation coefficient between each environmental factor and salinity in the window by sliding a fixed-size window on the time series. The selection of the window size requires a trade-off between time resolution and statistical significance. Usually, 30 days can be selected as the starting window size. After each sliding, a set of correlation coefficients will be obtained, and these coefficients will form a dynamic correlation coefficient matrix over time. Spectral clustering is performed on the dynamic correlation coefficient matrix to obtain the clustering results of environmental factors. Spectral clustering is a clustering method based on graph theory. First, a similarity matrix is ​​constructed, its Laplace matrix is ​​calculated, and then the Laplace matrix is ​​eigendecomposed. Finally, the eigenvectors are clustered using the K-means algorithm. This step can classify environmental factors with similar time change patterns into one category. The clustering results of environmental factors are subjected to principal component regression analysis to obtain preliminary factor weights. Principal component regression first performs principal component analysis on the independent variables (environmental factors), and then uses the principal components as new independent variables for multivariate linear regression. The regression coefficients are reversely transformed to obtain the weights of the original variables, i.e., preliminary factor weights.

[0068] Bootstrap resampling is applied to the preliminary factor weights to obtain weight confidence intervals. The Bootstrap method generates multiple sample sets by randomly sampling with replacement from the original data set, and repeats the principal component regression analysis on each sample set to obtain multiple groups of factor weights. By calculating the distribution characteristics of these weights (such as 2.5% and 97.5% quantiles), the confidence interval of each factor weight can be obtained. The weight confidence interval is fuzzified by the adaptive fuzzy inference system to obtain a fuzzy weight set. Based on the statistical characteristics of the weight confidence interval, the adaptive fuzzy inference system designs a fuzzy membership function (such as a triangular membership function or a Gaussian membership function) to convert the exact weight value into a fuzzy set. This step takes into account the uncertainty of weight estimation, making subsequent analysis more robust.

[0069] Perform the analytic hierarchy process on the fuzzy weight set to obtain a hierarchical weight structure. The analytic hierarchy process (AHP) constructs a judgment matrix and calculates the eigenvector to obtain the relative importance of each hierarchical factor. In this process, it is necessary to consider the relationship between environmental factors and the degree of influence on salinity to construct a multi-level weight structure. Cross-validation is applied to the hierarchical weight structure to obtain the validation weight. Cross-validation divides the data set into a training set and a validation set, and evaluates the generalization ability of the weight structure through multiple iterations of training and validation. Commonly used methods include k-fold cross-validation or leave-one-out cross-validation. During the validation process, the root mean square error (RMSE) or the coefficient of determination (R2) can be used as an evaluation indicator.

[0070] The contribution of the validation weights is quantified to obtain the factor importance ranking. The contribution quantification can be done using the Shapley value method, which takes into account the interaction between factors and calculates the average marginal contribution of each factor to the model prediction. The calculation of the Shapley value involves all possible factor combinations, and ultimately the importance score of each factor is obtained. The factor importance ranking is nonlinearly optimized to obtain the weights of the salinity influencing factors. Nonlinear optimization uses methods such as gradient descent or genetic algorithms to fine-tune the factor weights with the objective function of minimizing the prediction error. This step takes into account the possible nonlinear relationship between factors and further improves the accuracy of the weights.

[0071] For example, a marine observation project collected 5 years of buoy salinity data and observation data of 10 related environmental factors. First, the trend component of each time series was obtained by STL decomposition. A 30-day sliding window was used for correlation analysis to obtain a dynamic correlation coefficient matrix of 1825×10 (5 years×365 days, 10 factors). Spectral clustering divided the 10 environmental factors into 3 categories. Principal component regression analysis obtained preliminary factor weights, such as temperature 0.3, pressure 0.2, chlorinity 0.25, etc. The 95% confidence interval of each factor weight was obtained by 1000 Bootstrap resampling. The adaptive fuzzy inference system used a triangular membership function to convert the weights into fuzzy sets. The hierarchical analysis process constructed a 3-layer weight structure. The weight structure was evaluated by 10-fold cross validation, and the average RMSE was 0.05. Shapley value calculation showed that temperature, chlorinity and pressure were the three most important factors affecting salinity. Finally, a genetic algorithm was used for nonlinear optimization to obtain the final weights of salinity influencing factors: temperature 0.32, pressure 0.22, chlorinity 0.27, etc.

[0072] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0073] (1) Segment the buoy salinity data into sliding time windows to obtain time series segments;

[0074] (2) Perform wavelet multi-resolution analysis on the time series segments to obtain multi-scale decomposition results;

[0075] (3) Extract the trend of the multi-scale decomposition results through empirical mode decomposition to obtain the drift trend curve;

[0076] (4) Perform piecewise linear fitting on the drift trend curve to obtain a piecewise slope sequence;

[0077] (5) Adaptively threshold segment the segmented slope sequence to obtain drift event markers;

[0078] (6) performing pattern matching on drift event markers to obtain drift type identification results;

[0079] (7) Calculating the cumulative deviation of the drift type identification result to obtain the original cumulative deviation value;

[0080] (8) The original cumulative deviation value is smoothed to obtain the cumulative deviation integral value.

[0081] Specifically, the buoy salinity data is segmented by sliding time windows to obtain time series fragments. This process divides the long time series into several overlapping short time series fragments by setting a fixed-size time window (such as 30 days) and sliding it over the entire time series, each time moving a fixed step size (such as 1 day). This segmentation method not only retains the temporal continuity of the data, but also facilitates the capture of local drift characteristics. Wavelet multi-resolution analysis is performed on the time series fragments to obtain multi-scale decomposition results. Wavelet multi-resolution analysis uses the discrete wavelet transform (DWT) method, selects appropriate wavelet basis functions (such as Daubechies wavelets), and performs multi-level decomposition on each time series fragment. This process decomposes the signal into approximate coefficients and detail coefficients of different frequencies, so that the changing characteristics of salinity data can be analyzed at different time scales. The trend of the multi-scale decomposition results is extracted by empirical mode decomposition (EMD) to obtain a drift trend curve. EMD is an adaptive signal processing method that decomposes complex signals into a finite number of intrinsic mode functions (IMFs) and a residual term. In this method, EMD is used to extract the long-term trend component, namely the drift trend curve, from the results of wavelet multi-resolution analysis. This step can effectively separate the drift trend and short-term fluctuations in salinity data.

[0082] The drift trend curve is piecewise linearly fitted to obtain a piecewise slope sequence. The piecewise linear fitting sets a fixed length of time period (such as 7 days), performs linear regression on the trend curve in each time period, and calculates the slope. These slopes constitute a time series that reflects the rate of change of the drift trend. The piecewise slope sequence is adaptively thresholded to obtain drift event markers. The adaptive threshold segmentation method is based on the OTSU algorithm, which automatically determines the optimal threshold by maximizing the inter-class variance. The part of the slope sequence that exceeds the threshold is marked as a potential drift event, thereby achieving preliminary detection of drift events. Pattern matching is performed on the drift event markers to obtain drift type recognition results. The pattern matching uses the dynamic time warping (DTW) algorithm to compare the detected drift events with predefined typical drift patterns (such as linear drift, step drift, periodic drift, etc.). The DTW algorithm can handle nonlinear scaling of time series, thereby identifying different types of drift patterns.

[0083] The drift type identification results are cumulatively calculated to obtain the original cumulative deviation value. The cumulative deviation calculation is based on the identified drift type, and the corresponding integration method is used for each type. For example, for linear drift, the trapezoidal integration method can be used; for step drift, the step function integration is used. This step quantifies the cumulative effect of drift on salinity measurement. The original cumulative deviation value is smoothed to obtain the cumulative deviation integral value. The smoothing process uses the kernel density estimation method, selects the appropriate kernel function (such as Gaussian kernel) and bandwidth parameters, and performs non-parametric smoothing on the original cumulative deviation value. This step reduces the impact of noise and obtains a more stable and reliable cumulative deviation integral value.

[0084] For example, a buoy was deployed in an ocean observation project to continuously monitor salinity data for one year with a sampling frequency of once per hour. First, a 30-day sliding window with a step size of 1 day was used to segment the data for one year (8760 data points) to obtain 336 time series segments. A 4-level Daubechies-4 wavelet decomposition was applied to each segment to obtain coefficients for 5 bands. Then, EMD was applied to these coefficients to extract the IMF components representing the drift trend. A linear fit was performed on the extracted trend curve every 7 days to obtain 52 slope values. An adaptive threshold of 0.02 PSU / day was calculated using the OTSU algorithm, and 15 potential drift events were identified. These events were classified into 8 linear drifts, 5 step drifts, and 2 periodic drifts by comparing them with the predefined patterns using the DTW algorithm. The cumulative deviation was calculated for each drift type, such as using the trapezoidal integration method for linear drift. Finally, the original cumulative deviation values ​​were smoothed using a Gaussian kernel with a bandwidth of 7 days to obtain the final cumulative deviation integral value sequence.

[0085] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0086] (1) Perform wavelet packet decomposition on the accumulated deviation integral value to obtain a multi-scale coefficient set;

[0087] (2) soft threshold processing is performed on the multi-scale coefficient set through an adaptive threshold to obtain a denoising coefficient set;

[0088] (3) Perform inverse wavelet packet transform on the denoising coefficient set to obtain the reconstructed deviation sequence;

[0089] (4) performing piecewise polynomial fitting on the reconstructed deviation sequence to obtain a correction function;

[0090] (5) Smoothing the correction function by sliding median filtering to obtain a smooth correction curve;

[0091] (6) performing a differential operation on the original salinity data and the smoothed calibration curve to obtain a corrected salinity value;

[0092] (7) The residual drift of the corrected salinity value is extracted through variational mode decomposition to obtain preliminary correction data.

[0093] Specifically, the accumulated deviation integral value is decomposed by wavelet packet to obtain a multi-scale coefficient set. Wavelet packet decomposition is an extension of wavelet transform. It decomposes not only the low-frequency part but also the high-frequency part, providing a finer frequency division. The specific steps include selecting a suitable wavelet basis function (such as Daubechies wavelet), determining the number of decomposition layers, and then recursively decomposing the signal to obtain coefficients of different scales and frequencies. The multi-scale coefficient set is subjected to soft threshold processing by an adaptive threshold to obtain a denoising coefficient set. The calculation of the adaptive threshold is based on the statistical characteristics of each subband, such as using the universal threshold method proposed by Donoho, the threshold T = σ√(2log(N)), where σ is the standard deviation of the noise and N is the signal length. Soft threshold processing sets the coefficients whose absolute values ​​are less than the threshold to zero, and subtracts the threshold from the coefficients whose absolute values ​​are greater than the threshold. This method is smoother than the hard threshold and can better maintain the continuity of the signal.

[0094] Perform an inverse wavelet packet transform on the denoising coefficient set to obtain a reconstructed deviation sequence. The inverse wavelet packet transform is the inverse process of wavelet packet decomposition. The denoised signal is finally reconstructed by reorganizing the processed coefficients in sequence. This step restores the deviation sequence in the time domain and removes the influence of high-frequency noise. Perform piecewise polynomial fitting on the reconstructed deviation sequence to obtain the correction function. Piecewise polynomial fitting divides the entire sequence into several time periods, and each section is fitted using a polynomial of a different order. The specific implementation includes determining the segmentation points (which can be based on the changing characteristics of the signal or a fixed time interval), selecting the appropriate polynomial order (usually 2-3 orders are used), and then fitting using the least squares method. This method can capture the nonlinear characteristics of drift in different time periods.

[0095] The correction function is smoothed by sliding median filtering to obtain a smooth correction curve. Sliding median filtering is a nonlinear filtering method that slides a fixed-size window on the time series and takes the median of the data in the window as the output each time. This method can effectively remove mutation points, maintain edge features, and make the correction curve smoother and more continuous.

[0096] The original salinity data and the smoothed correction curve are differentially operated to obtain the corrected salinity value. The differential operation is essentially to subtract the smoothed correction curve from the original salinity data to eliminate the influence of drift. This step directly applies the correction information obtained by the previous calculation to achieve a preliminary correction of the original data. The residual drift of the corrected salinity value is extracted by variational mode decomposition (VMD) to obtain preliminary corrected data. VMD is an adaptive signal decomposition method that decomposes the signal into a set of band-limited intrinsic mode functions (IMFs). VMD estimates all modes simultaneously by solving a variational problem and has good noise robustness. Here, VMD is used to extract the residual drift components that may exist in the corrected data and further refine the correction results.

[0097] For example, a marine observation project used buoys to continuously monitor salinity data for one year, with a sampling frequency of once an hour. First, the calculated cumulative deviation integral value sequence (8760 data points) was subjected to a 4-level wavelet packet decomposition, and the coefficients of 16 frequency bands were obtained using the Daubechies-4 wavelet. Then, the threshold of each sub-band was calculated using the universal threshold method, and the coefficients were soft-thresholded. For example, for a certain sub-band, the calculated threshold was 0.05 PSU. The inverse wavelet packet transform was used to reconstruct the denoised deviation sequence. Next, the reconstructed sequence was divided into a time period of 30 days, a total of 12 segments, and a third-order polynomial was used to fit each segment to obtain 12 piecewise functions. The fitting function was smoothed using a sliding median filter with a 24-hour window to obtain a continuous correction curve. The original salinity data was subtracted from this correction curve to obtain the preliminary corrected salinity value. Finally, the VMD algorithm was used to decompose the corrected salinity value into 3 IMFs and a residual, where the residual was regarded as a potential residual drift and removed from the data.

[0098] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0099] (1) Performing spatiotemporal grid division on the preliminary correction data to obtain gridded correction data;

[0100] (2) Perform spatial autocorrelation analysis on the gridded correction data to obtain local spatial dependence indicators;

[0101] (3) The probability distribution of local spatial dependency indicators is estimated through variational Bayesian inference to obtain a spatial correlation model;

[0102] (4) Perform dynamic spatial weight matching on the spatial correlation model to obtain a dynamic spatial weight matrix;

[0103] (5) Perform time series decomposition on the preliminary corrected data to obtain seasonal components;

[0104] (6) Perform local polynomial regression on the seasonal component to obtain the time correlation curve;

[0105] (7) performing exponential smoothing on the time correlation curve to obtain a dynamic time weight vector;

[0106] (8) Performing a tensor product operation on the dynamic spatial weight matrix and the dynamic temporal weight vector to obtain a spatiotemporal joint weight tensor;

[0107] (9) performing weighted average calculation on the preliminary correction data and the spatiotemporal joint weight tensor to obtain weight correction data;

[0108] (10) The weight correction data is subjected to edge-preserving smoothing processing by anisotropic diffusion filtering algorithm to obtain optimized correction data.

[0109] Specifically, the preliminary correction data are divided into time and space grids to obtain gridded correction data. This process divides the observation area into regular grids, and each grid cell contains data points within a specific time and space range. The selection of grid size requires a balance between spatial resolution and data density. Usually, an adaptive grid method can be used to dynamically adjust the grid size according to the data distribution. Spatial autocorrelation analysis is performed on the gridded correction data to obtain local spatial dependence indicators. Spatial autocorrelation analysis uses Moran's I index or Geary's C index to calculate the degree of correlation between each grid cell and its neighboring cells. This step quantifies the spatial dependence of salinity data and provides a basis for subsequent spatial weight calculations.

[0110] The probability distribution of local spatial dependency indicators is estimated by variational Bayesian inference to obtain the spatial correlation model. The variational Bayesian method approximates the posterior distribution by minimizing the KL divergence and can effectively handle complex probability models. Here, it is used to estimate the probability distribution of spatial dependency indicators and capture the uncertainty and variability of spatial correlation. Dynamic spatial weight matching is performed on the spatial correlation model to obtain a dynamic spatial weight matrix. This step assigns weights to each grid cell based on the estimated spatial correlation model. Commonly used methods include inverse distance weighting or Gaussian kernel-based weighting. The weight matrix reflects the degree of mutual influence between different spatial locations.

[0111] At the same time, the preliminary corrected data is decomposed into a time series to obtain a seasonal component. The time series decomposition uses the STL (Seasonal and Trend decomposition using Loess) method to decompose the time series into trend, season, and residual components. Here we focus on the seasonal component, which reflects the periodic changes in salinity data. Local polynomial regression is performed on the seasonal component to obtain a time correlation curve. Local polynomial regression slides a fixed-size window on the time series and fits a low-order polynomial (usually 1-3 orders) to the data in each window. This method can capture local nonlinear trends and adapt to the dynamic changes of time correlation.

[0112] Exponential smoothing is performed on the time correlation curve to obtain a dynamic time weight vector. Exponential smoothing gives higher weights to recent data, and the weights of distant data decay exponentially. This processing method takes into account the recent correlation in time, making the weight distribution more reasonable. The dynamic spatial weight matrix and the dynamic time weight vector are tensor-producted to obtain a spatiotemporal joint weight tensor. The tensor product operation combines the weight information of the spatial and temporal dimensions to form a high-dimensional weight structure, which fully considers the spatiotemporal correlation of the data.

[0113] The weighted average calculation of the preliminary correction data and the spatiotemporal joint weight tensor is performed to obtain the weight correction data. This step applies the spatiotemporal joint weight to the preliminary correction data, realizing data adjustment based on spatiotemporal correlation. Finally, the weight correction data is edge-preservingly smoothed by the anisotropic diffusion filter algorithm to obtain the optimized correction data. The anisotropic diffusion filter can maintain important edge information while smoothing the data, avoiding information loss caused by over-smoothing.

[0114] For example, a marine observation project deployed 25 buoys in a 100km×100km sea area. Each buoy recorded salinity data once an hour for a one-year observation period. First, the observation area was divided into 10×10 grids, with each grid cell being 10km×10km. The Moran's I index was calculated for each grid cell to obtain the local spatial dependence index. The variational Bayesian method was used to estimate the distribution model of exponentially decaying spatial correlation. Based on this model, a dynamic spatial weight matrix was constructed. For example, the weight of adjacent grids was 0.5, and the weight of grids separated by one grid was 0.25. The time series data was decomposed by STL to extract the seasonal component. The seasonal component was fitted using a local polynomial regression with a 30-day window to obtain the temporal correlation curve. An exponential smoothing process with a parameter of 0.2 was applied to generate a dynamic temporal weight vector. The 10×10 spatial weight matrix was tensor-producted with the 8760-dimensional (number of hours in a year) temporal weight vector to obtain a 10×10×8760 spatiotemporal joint weight tensor. The preliminary correction data are weighted averaged using this weight tensor, and then anisotropic diffusion filtering is applied (the diffusion coefficient is set to 0.1) to obtain the final optimized correction data.

[0115] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0116] (1) Perform multi-dimensional spectrum analysis on the optimized correction data to obtain a frequency domain feature matrix;

[0117] (2) Extract features from the frequency domain feature matrix to obtain the main feature vector;

[0118] (3) Cluster analysis is performed on the main feature vectors to obtain data pattern grouping;

[0119] (4) Using the sliding window method to calculate the temporal correlation of the data pattern groups, a dynamic correlation matrix is ​​obtained;

[0120] (5) Based on the dynamic correlation matrix, a multi-layer threshold network is constructed through an adaptive threshold algorithm to obtain a threshold candidate set;

[0121] (6) Cross-validate the threshold candidate set to obtain the optimal threshold combination;

[0122] (7) The optimal threshold combination is processed continuously by interpolation method to obtain a smooth threshold curve;

[0123] (8) Based on the smooth threshold curve, the correction effect is evaluated by multiple indicators to obtain a comprehensive performance score;

[0124] (9) Perform time series prediction on the comprehensive performance score to obtain performance trend prediction;

[0125] (10) According to the performance trend prediction, the correction parameters are adjusted through an iterative optimization algorithm to obtain the target correction data and parameter updating strategy.

[0126] Specifically, in the salinity drift correction method for buoy observation data, the optimized correction data is first subjected to multi-dimensional spectrum analysis to obtain a frequency domain feature matrix. This process uses short-time Fourier transform (STFT) or wavelet transform to convert time series data into the time-frequency domain. STFT performs Fourier transform on the data in each window by sliding a window on the time series to obtain the spectrum characteristics that change over time. Feature extraction is performed on the frequency domain feature matrix to obtain the main eigenvector. Feature extraction uses principal component analysis (PCA) or independent component analysis (ICA) methods. PCA calculates the covariance matrix of the feature matrix, solves its eigenvalues ​​and eigenvectors, and selects several eigenvectors with the largest contribution rate as the main features. Then, cluster analysis is performed on the main eigenvectors to obtain data pattern grouping. Cluster analysis uses K-means or hierarchical clustering algorithms. The K-means algorithm distributes data points into K clusters through iterative optimization, so that the sum of squared distances within the cluster is minimized. This step classifies similar data patterns for subsequent analysis. The sliding window method is used to calculate the time series correlation of data pattern groups to obtain a dynamic correlation matrix. The sliding window method sets a fixed-size time window (such as 30 days), slides on the time series, and calculates the correlation coefficient between each pattern group in the window. This method can capture the dynamic correlation relationship between data patterns. Based on the dynamic correlation matrix, a multi-layer threshold network is constructed by an adaptive threshold algorithm to obtain a threshold candidate set. The adaptive threshold algorithm uses the OTSU method or the maximum inter-class variance method to automatically determine multiple threshold levels according to the distribution characteristics of the dynamic correlation matrix. These thresholds form a network structure, reflecting the multi-level characteristics of data association. The threshold candidate set is cross-validated to obtain the optimal threshold combination. The cross-validation uses the k-fold cross-validation method to divide the data set into k parts, each time using k-1 parts as the training set and 1 part as the validation set, and the cycle is k times. The optimal combination is selected by comparing the performance of different threshold combinations on the validation set. The optimal threshold combination is continuous processed by the interpolation method to obtain a smooth threshold curve. The interpolation method can choose spline interpolation or polynomial interpolation to convert discrete threshold points into continuous curves. This can provide a suitable threshold for any time point, enhancing the adaptability of the method. Based on the smooth threshold curve, the correction effect is evaluated by multiple indicators to obtain a comprehensive performance score. The evaluation indicators include root mean square error (RMSE), correlation coefficient (R2), consistency index (CI), etc. Multiple indicators are combined into an overall performance score through weighted average or fuzzy comprehensive evaluation methods. Time series prediction is performed on the comprehensive performance score to obtain a performance trend prediction. The prediction method can use the ARIMA model or the long short-term memory network (LSTM). The ARIMA model makes predictions by analyzing the autocorrelation and moving average characteristics of the time series, while the LSTM can capture long-term dependencies and is suitable for complex nonlinear prediction tasks.Finally, according to the performance trend prediction, the correction parameters are adjusted through the iterative optimization algorithm to obtain the target correction data and parameter update strategy. Iterative optimization can use the gradient descent method or genetic algorithm. The algorithm continuously adjusts the correction parameters according to the predicted performance trend until the preset convergence condition or the upper limit of the number of iterations is reached. For example, a marine observation project used buoys to continuously monitor salinity data for one year, with a sampling frequency of once an hour. First, STFT analysis was performed on 8760 data points, and a 513×256 frequency domain feature matrix was obtained using a 1024-point Hanning window. The first 10 principal components were extracted as the main eigenvectors through PCA. The main eigenvectors were clustered using the K-means algorithm (K=5) to obtain 5 data modes. A 30-day sliding window was used to calculate the correlation coefficients between the 5 modes to form a 5×5×337 dynamic correlation matrix. A 3-layer threshold network was constructed using the OTSU method to obtain 27 threshold candidate combinations. The optimal threshold combination was selected through 5-fold cross validation. Cubic spline interpolation was applied to the optimal threshold combination to obtain a smooth threshold curve. Based on the threshold curve, the three indicators of RMSE, R2 and CI were calculated, and the comprehensive performance score was obtained by weighted average (weights were 0.4, 0.3, and 0.3 respectively). The performance score for the next 30 days was predicted using the ARIMA (2,1,2) model. Finally, the correction parameters were iteratively adjusted using the gradient descent method with the predicted performance trend as the objective function. After 100 iterations, the final target correction data and parameter update strategy were obtained. This process realizes the dynamic evaluation and adaptive optimization of the salinity drift correction method, ensuring the continuous improvement of the correction effect.

[0127] The above describes the salinity drift correction method for buoy observation data in the embodiment of the present application. The following describes the salinity drift correction system for buoy observation data in the embodiment of the present application. Figure 2 In one embodiment of the present application, a system for correcting salinity drift of buoy observation data includes:

[0128] The acquisition module 201 is used to collect and preprocess the historical buoy observation data and current environmental parameters to obtain a multi-source heterogeneous data set;

[0129] An analysis module 202 is used to perform multi-dimensional environmental factor analysis on the multi-source heterogeneous data set to obtain salinity influencing factor weights;

[0130] A quantification module 203 is used to perform drift detection and quantification on the buoy salinity data by using the salinity influencing factor weight to obtain a cumulative deviation integral value;

[0131] A correction module 204 is used to perform multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data;

[0132] An optimization module 205 is used to perform dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data;

[0133] The evaluation module 206 is used to perform real-time performance evaluation and automatic threshold adjustment on the optimized correction data to obtain target correction data and parameter update strategy.

[0134] Through the synergy of the above components, by introducing the preprocessing of multi-source heterogeneous data sets and multi-dimensional environmental factor analysis, this method greatly improves the accuracy of the weights of salinity influencing factors, thus laying a solid foundation for subsequent drift detection and correction. Secondly, the cumulative deviation integral value is used as a drift quantification indicator, combined with the sliding time window and wavelet multi-resolution analysis technology, so that this method can more accurately capture the drift characteristics on different time scales, significantly improving the sensitivity and accuracy of drift detection. Furthermore, the introduction of multi-scale correction calculations, especially the combination of wavelet packet decomposition and adaptive threshold processing, enables this method to effectively process drift components of different frequencies and amplitudes, greatly improving the accuracy and robustness of the correction. In addition, the application of the dynamic spatiotemporal weight optimization strategy, by integrating spatial autocorrelation analysis and time series decomposition technology, makes full use of the spatiotemporal correlation of buoy observation data, and significantly enhances the spatial consistency and temporal continuity of the correction results. Finally, the introduction of real-time performance evaluation and automatic threshold adjustment mechanism enables the method to have the ability of adaptive learning and continuous optimization, and can effectively respond to the dynamic changes of the marine environment and emerging drift patterns, ensuring the long-term effectiveness and reliability of the correction system. The method of the present invention has achieved significant improvements in processing speed, correction accuracy and adaptability.

[0135] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0136] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for correcting salinity drift of buoy observation data, characterized in that: The salinity drift correction method for buoy observation data comprises: Collect and preprocess historical buoy observation data and current environmental parameters to obtain multi-source heterogeneous data sets; Performing multidimensional environmental factor analysis on the multi-source heterogeneous data sets to obtain salinity influencing factor weights; Drift detection and quantification of the buoy salinity data are performed using the salinity influencing factor weights to obtain a cumulative deviation integral value; Performing multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data; Performing dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data; Real-time performance evaluation and automatic threshold adjustment are performed on the optimized correction data to obtain target correction data and parameter update strategy.

2. The salinity drift correction method for buoy observation data according to claim 1, characterized in that: The historical buoy observation data and current environmental parameters are collected and preprocessed to obtain a multi-source heterogeneous data set, including: Performing time stamp alignment on the historical buoy observation data and the current environmental parameters to obtain time synchronization data; Performing spatial grid division on the time-synchronized data to obtain gridded observation data; Performing wavelet transform denoising on the gridded observation data to obtain denoised observation data; Interpolating missing values ​​of the noise-reduced observation data to obtain a complete observation data set; Performing principal component analysis on the complete observation data set to obtain a dimension-reduced feature set; Performing outlier detection based on the interquartile range method on the dimension reduction feature set to obtain abnormal labeled data; Eliminating abnormal samples from the abnormally labeled data to obtain a cleaned data set; Performing format unification processing on the cleaned data set to obtain a standardized multi-source heterogeneous data set; Storing the standardized multi-source heterogeneous data set in a distributed database to obtain an indexed data set; The indexed data set is scored for data quality to obtain a multi-source heterogeneous data set with quality labels.

3. The salinity drift correction method for buoy observation data according to claim 1, characterized in that: The multi-dimensional environmental factor analysis is performed on the multi-source heterogeneous data set to obtain the weight of the salinity influencing factor, including: Performing time series decomposition on the multi-source heterogeneous data sets to obtain trend components; Performing sliding window correlation analysis on the trend components to obtain a dynamic correlation coefficient matrix; Performing spectral clustering on the dynamic correlation coefficient matrix to obtain environmental factor clustering results; Perform principal component regression analysis on the environmental factor clustering results to obtain preliminary factor weights; Bootstrap resampling is performed on the preliminary factor weights to obtain weight confidence intervals; Fuzzifying the weight confidence interval by an adaptive fuzzy inference system to obtain a fuzzy weight set; Performing a hierarchical analysis process on the fuzzy weight set to obtain a hierarchical weight structure; Applying cross validation to the hierarchical weight structure to obtain validation weights; Quantifying the contribution of the verification weights to obtain a ranking of factor importance; The importance ranking of the factors is nonlinearly optimized to obtain the weights of the salinity influencing factors.

4. The salinity drift correction method for buoy observation data according to claim 1, characterized in that: The method of performing drift detection and quantification on the buoy salinity data by using the salinity influencing factor weight to obtain a cumulative deviation integral value includes: The buoy salinity data is segmented into sliding time windows to obtain time series segments; Performing wavelet multi-resolution analysis on the time series fragments to obtain a multi-scale decomposition result; Performing trend extraction on the multi-scale decomposition result through empirical mode decomposition to obtain a drift trend curve; Performing piecewise linear fitting on the drift trend curve to obtain a piecewise slope sequence; Performing adaptive threshold segmentation on the segmented slope sequence to obtain drift event markers; Performing pattern matching on the drift event mark to obtain a drift type identification result; Performing cumulative deviation calculation on the drift type identification result to obtain an original cumulative deviation value; The original cumulative deviation value is smoothed to obtain the cumulative deviation integral value.

5. The salinity drift correction method for buoy observation data according to claim 1, characterized in that: The performing multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data includes: Performing wavelet packet decomposition on the accumulated deviation integral value to obtain a multi-scale coefficient set; Performing soft threshold processing on the multi-scale coefficient set through an adaptive threshold to obtain a denoising coefficient set; Performing an inverse wavelet packet transform on the denoising coefficient set to obtain a reconstructed deviation sequence; Performing piecewise polynomial fitting on the reconstructed deviation sequence to obtain a correction function; Smoothing the correction function by sliding median filtering to obtain a smooth correction curve; Performing a differential operation on the original salinity data and the smoothed correction curve to obtain a corrected salinity value; The residual drift of the corrected salinity value is extracted by variational mode decomposition to obtain the preliminary correction data.

6. The salinity drift correction method for buoy observation data according to claim 1, characterized in that: The step of performing dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data includes: Performing space-time grid division on the preliminary correction data to obtain gridded correction data; Performing spatial autocorrelation analysis on the gridded correction data to obtain a local spatial dependence index; Probability distribution estimation of the local spatial dependency index is performed through variational Bayesian inference to obtain a spatial correlation model; Performing dynamic spatial weight matching on the spatial correlation model to obtain a dynamic spatial weight matrix; Performing time series decomposition on the preliminary corrected data to obtain seasonal components; Performing local polynomial regression on the seasonal component to obtain a time correlation curve; Performing exponential smoothing on the time correlation curve to obtain a dynamic time weight vector; Performing a tensor product operation on the dynamic spatial weight matrix and the dynamic temporal weight vector to obtain a spatiotemporal joint weight tensor; Performing weighted average calculation on the preliminary correction data and the spatiotemporal joint weight tensor to obtain weight correction data; The weight correction data is subjected to edge-preserving smoothing processing by an anisotropic diffusion filtering algorithm to obtain optimized correction data.

7. The salinity drift correction method for buoy observation data according to claim 1, characterized in that: The real-time performance evaluation and automatic threshold adjustment of the optimized correction data to obtain target correction data and parameter update strategy include: Performing multi-dimensional spectrum analysis on the optimized correction data to obtain a frequency domain feature matrix; Performing feature extraction on the frequency domain feature matrix to obtain a main feature vector; Performing cluster analysis on the main feature vectors to obtain data pattern grouping; Performing time series correlation calculation on the data pattern groups by a sliding window method to obtain a dynamic correlation matrix; Based on the dynamic association matrix, a multi-layer threshold network is constructed by an adaptive threshold algorithm to obtain a threshold candidate set; Cross-validating the threshold candidate set to obtain an optimal threshold combination; The optimal threshold combination is processed continuously by interpolation method to obtain a smooth threshold curve; Based on the smooth threshold curve, a multi-index evaluation is performed on the correction effect to obtain a comprehensive performance score; Performing time series forecasting on the comprehensive performance score to obtain a performance trend forecast; According to the performance trend prediction, the correction parameters are adjusted through an iterative optimization algorithm to obtain the target correction data and the parameter update strategy.

8. A salinity drift correction system for buoy observation data, characterized in that: Used to perform the salinity drift correction method for buoy observation data according to any one of claims 1 to 7, the salinity drift correction system for buoy observation data comprises: The acquisition module is used to collect and preprocess historical buoy observation data and current environmental parameters to obtain multi-source heterogeneous data sets; An analysis module, used for performing multi-dimensional environmental factor analysis on the multi-source heterogeneous data set to obtain the weight of the salinity influencing factor; A quantification module, used to perform drift detection and quantification on the buoy salinity data by using the salinity influencing factor weight to obtain a cumulative deviation integral value; A correction module, used for performing multi-scale correction calculation on the accumulated deviation integral value to obtain preliminary correction data; An optimization module, used for performing dynamic spatiotemporal weight optimization on the preliminary correction data to obtain optimized correction data; The evaluation module is used to perform real-time performance evaluation and automatic threshold adjustment on the optimized correction data to obtain target correction data and parameter update strategy.

Images