A marine industry and finance integration decision-making method, medium and system based on big data analysis

Abstract

The application provides a marine industry and finance fusion decision-making method, medium and system based on big data analysis, belongs to the technical field of big data analysis, and achieves the spatiotemporal alignment and collaborative feature domain classification of multi-source heterogeneous data by adopting a multiscale pattern Laplacian operator evolution dynamics atlas clustering algorithm, removes mixed noise through a fractional calculus complex time-varying diffusion dynamics denoising algorithm, compresses high-dimensional sparse features by using a self-encoder and t-SNE fusion dimension reduction technology, inputs an adaptive spatiotemporal feedback skip graph neural network model based on wave equation constraint to obtain a dynamic financial cost prediction result, and finally maps high-dimensional non-Gaussian industry and finance distribution to a standard Gaussian space by using a deep flow artificial intelligence model based on manifold alignment and performs reverse inverse transformation to output disaster prevention budget compilation and capitalization budget suggestions, thereby solving the technical problem that after multi-source heterogeneous marine industry and finance data are fused under the condition that the spatiotemporal scales do not match, false features are generated, and financial decision-making is misjudged.

Description

Technical Field

[0001] This invention belongs to the field of big data analytics technology, and more specifically, relates to a marine industry-finance integration decision-making method, medium, and system based on big data analytics. Background Technology

[0002] In the field of marine environmental monitoring and business-finance integration decision-making, existing technologies typically use satellite remote sensing, deep-sea buoy arrays, and equipment maintenance financial cost flows as data sources. Traditional linear interpolation or principal component analysis methods are used to perform spatiotemporal alignment and dimensionality reduction on multi-source data, and then combined with standard graph neural networks or regression models to output financial forecast results. While these existing technologies are applicable to scenarios with consistent data acquisition frequencies and similar spatial resolutions, in actual marine multi-source heterogeneous data acquisition conditions, the acquisition frequencies and spatial resolutions of satellite remote sensing, buoy arrays, and financial cost flows differ significantly. When linear interpolation is used to forcibly align these heterogeneous data, the spatiotemporal scale mismatch introduces spurious correlation features, or pseudo-features, during the interpolation process. In existing technologies, due to the spatiotemporal scale mismatch of multi-source heterogeneous data, the linear interpolation alignment process systematically introduces pseudo-features into the business-finance integration features, causing subsequent financial forecasting models to misjudge noise correlations as genuine business patterns, leading to systematic deviations in disaster prevention budget preparation and capitalization budget decisions. In other words, existing technologies have a technical problem: when multi-source heterogeneous marine financial data are fused under conditions of mismatched spatiotemporal scales, false features are generated, leading to misjudgments in financial decisions. Summary of the Invention

[0003] In view of this, the present invention provides a marine industry and finance fusion decision-making method, medium and system based on big data analysis, which can solve the technical problem in the prior art where false features are generated after the fusion of multi-source heterogeneous marine industry and finance data under conditions of mismatch in spatiotemporal scales, leading to misjudgment of financial decisions.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a marine industry-finance integrated decision-making method based on big data analysis, comprising the following steps: Collect satellite remote sensing, deep-sea buoy array, and shore three-dimensional current velocity and temperature fluctuation data, as well as equipment maintenance financial cost data. Utilize the evolutionary dynamics graph clustering algorithm based on multi-scale map Laplacian operator, and use physical energy dissipation as a geometric constraint to complete the spatiotemporal alignment and collaborative feature domain classification of multi-source heterogeneous data in the spectral space. A complex time-varying diffusion dynamics denoising algorithm based on nonlocal mean field and fractional calculus is applied to the spatiotemporally aligned multi-source heterogeneous data to remove Gaussian and non-Gaussian mixed noise in the deep-sea exploration data, resulting in a high signal-to-noise ratio fusion input dataset. High-dimensional sparse feature compression was performed on the high signal-to-noise ratio business and finance fusion input dataset using autoencoder and t-SNE fusion dimensionality reduction technology to construct a low-dimensional core business indicator set. The low-dimensional core business indicators set, along with the three-dimensional current velocity and temperature fluctuation data of the deep-sea buoy array and the financial cost flow of equipment maintenance, are input into an adaptive spatiotemporal feedback jump graph neural network model based on the wave equation constraint, and the output is a dynamic financial cost prediction result. High-dimensional non-Gaussian business and financial distribution data are input into a manifold-aligned deep flow model, mapped to a standard multivariate Gaussian distribution space through a reversible affine coupling transformation layer, and then an inverse transformation is performed to output disaster prevention budget preparation and capitalization budget recommendations.

[0005] Specifically, the autoencoder and t-SNE fusion dimensionality reduction technique involves the autoencoder encoding high-dimensional business feature vectors into a low-dimensional latent space, training the encoder with reconstruction error as the loss function to preserve the nonlinear topological structure, and then using the low-dimensional latent vector output by the encoder as input to execute the t-SNE algorithm to further optimize the low-dimensional distribution.

[0006] Specifically, the evolutionary dynamics graph clustering algorithm abstracts geographically scattered buoy observation points, satellite observation points, and financial reimbursement nodes into a dynamic spatiotemporal weighted graph. It constructs a multi-scale graph Laplacian operator on the graph structure, solves for eigenvalues ​​and eigenvectors, projects the high-dimensional spatiotemporal financial relationship onto the low-dimensional spectral space, and performs adaptive evolutionary clustering based on energy aggregation degree to complete the spatiotemporal alignment and collaborative feature domain classification of heterogeneous data.

[0007] The mathematical form of the pattern Laplace operator is: ,in Let the node degree matrix be... For the adjacency weight matrix, the multi-scale version uses... This is achieved by applying polynomial filter banks of different orders; the edge weights are calculated by jointly normalizing the spatiotemporal distance and frequency differences between each data source node.

[0008] The energy clustering threshold of the adaptive evolutionary clustering is determined by statistical analysis of the distribution of spectral space eigenvalues. The normalization coefficient ranges from 0.01 to 1.00, and the recommended range for the energy clustering threshold is 0.05 to 0.30.

[0009] Specifically, the denoising algorithm based on nonlocal mean field and fractional calculus of complex time-varying diffusion dynamics treats the original time series of noisy business as a virtual two-dimensional complex electromagnetic fluid field, and introduces fractional differential operators in both the time and frequency domains to construct a non-traditional partial differential diffusion equation. The diffusion coefficient is dynamically determined by the similarity of the nonlocal mean structure.

[0010] The recommended order range for the fractional differential operator is 0.3–0.9; the recommended window size for nonlocal mean structure similarity is 5–21 sampling points; and the recommended range for the convergence determination residual threshold of the collaborative feature domain is [missing information]. ~ .

[0011] The adaptive spatiotemporal feedback jump graph neural network model based on wave equation constraints consists of a spatial graph convolution branch and a time-gated recurrent unit branch, which are coupled through an adaptive spatiotemporal feedback jump mechanism. When the physical input index deviates from the normal threshold, the conventional intermediate hidden layer is skipped, and the Laplace operator physical wave equation constraint term is directly injected into the final financial loss prediction layer.

[0012] The spatial graph convolution branch uses four graph convolutional layers, with output feature dimensions of 64, 128, 128, and 256 respectively; the hidden state dimension of the temporal gated recurrent unit branch is 128; the recommended range for the time series window length is 24 to 168 time steps; and the recommended range for the number of samples per batch is 32 to 128.

[0013] In the adaptive spatiotemporal feedback jump mechanism, the normal threshold is determined by statistical analysis of historical buoy data, and the recommended range for the anomaly determination quantile is 1% to 5%; the constraint terms of the physical wave equation are superimposed on the prediction layer input in the form of residuals of the Laplace operator acting on the node features.

[0014] The adaptive dynamic attention adjustment function is defined as follows: ,in This represents the loss value of the validation set in the current round. For the current training round, Initial loss value; constraint weight coefficients With attention factor The relationship is .

[0015] Wherein, the attention factor The attention weight parameter ranges corresponding to the three intervals are as follows: Time to take ~ Time to take ~ Time to take ~ .

[0016] The deep flow model based on manifold alignment is composed of a series of reversible affine coupling transformation layers. Each reversible affine coupling transformation layer splits the input vector into two parts. One part is transformed by scaling and translation functions, while the other part remains unchanged. During the transformation, the probability density is continuously deformed according to the determinant of the Jacobian matrix.

[0017] The recommended range for the number of cascaded layers in the deep flow model is 8 to 24; the range of attention weight parameters for each interval of the adaptive dynamic attention adjustment function is determined by iterative analysis of the loss curves of multiple training experiments and validation sets; and the recommended range for the edge weight normalization coefficient is 0.01 to 1.00.

[0018] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described marine industry-finance integration decision-making method based on big data analysis.

[0019] A third aspect of the present invention provides a marine industry and finance integrated decision-making system based on big data analysis, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0020] This invention employs an evolutionary dynamics graph clustering algorithm based on the multi-scale graph Laplacian operator. It abstracts multi-source heterogeneous data into a dynamic spatiotemporal weighted graph, automatically searching for the intrinsic spatiotemporal correlation structure of each data source in the spectral space using physical energy dissipation as a geometric constraint. This eliminates the spatiotemporal scale assumption of linear interpolation during the alignment process, forming a collaborative feature domain in the graph space that possesses both physical and economic significance. This collaborative feature domain provides a feature foundation with dual physical and economic self-consistency for subsequent denoising, dimensionality reduction, and prediction models. It ensures that the features entering the financial decision-making model no longer contain spurious correlations introduced by spatiotemporal scale mismatch, thus establishing the disaster prevention budget and capitalization budget schemes output by the subsequent deep flow model on a true business-financial correlation structure. In summary, this invention solves the technical problem mentioned in the background art of generating spurious features after fusing multi-source heterogeneous marine business-financial data under conditions of spatiotemporal scale mismatch, leading to misjudgments in financial decisions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 A comparison chart of the time-series characteristics of business and financial data integration before and after noise reduction.

[0023] Figure 3 This is a chart comparing the predicted and actual values ​​of dynamic financial costs. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0025] like Figure 1 The diagram shown is a flowchart of a marine industry-finance integration decision-making method based on big data analysis, provided by the first aspect of this invention. This method includes the following steps: S01. Collect satellite remote sensing, deep-sea buoy array, and shore three-dimensional current velocity and temperature fluctuation data, as well as equipment maintenance financial cost data. Use the evolutionary dynamic map clustering algorithm based on multi-scale map Laplace operator to complete the spatiotemporal alignment and collaborative feature domain classification of multi-source heterogeneous data. S02. Apply a complex time-varying diffusion dynamics denoising algorithm based on nonlocal mean field and fractional calculus to the spatiotemporally aligned multi-source heterogeneous data to remove Gaussian and non-Gaussian mixed noise in the deep-sea exploration data and obtain a high signal-to-noise ratio integrated input dataset. S03. Use autoencoder and t-SNE fusion dimensionality reduction technology to compress high-dimensional sparse features into the high signal-to-noise ratio business and finance fusion input dataset, construct a low-dimensional core business indicator set, retain nonlinear topological structure features and remove redundant orthogonal bases. S04. Input the low-dimensional core business indicator set with the three-dimensional current velocity and temperature fluctuation data of the ocean buoy array and the financial cost of equipment maintenance into the adaptive spatiotemporal feedback jump graph neural network model based on the wave equation constraint, and output the dynamic financial cost prediction result. S05. Input the high-dimensional non-Gaussian business and financial distribution data into the deep flow model based on manifold alignment, map it to the standard multivariate Gaussian distribution space through the reversible affine coupling transformation layer, and then perform the inverse transformation to obtain the business and financial decision solution space distribution, and output disaster prevention budget compilation and capitalization budget recommendations.

[0026] The multi-source heterogeneous data refers to satellite remote sensing data, three-dimensional current velocity and temperature fluctuation data of deep-sea buoy arrays, and financial cost data of equipment maintenance. The three types of data have significant differences in acquisition frequency and spatial resolution, and direct fusion is prone to false features. The false features are false correlation features introduced by the mismatch of spatiotemporal scales during the interpolation process of heterogeneous data, which will increase the risk of misjudgment in financial decision-making.

[0027] The specific implementation of the evolutionary dynamics graph clustering algorithm based on the multi-scale graph Laplacian operator is as follows: First, geographically scattered buoy observation points, satellite observation points, and financial reimbursement nodes are abstracted into a dynamic spatiotemporal weighted graph, where nodes represent data sources and edge weights represent the spatiotemporal correlation strength. Then, a multi-scale graph Laplacian operator is constructed on the graph structure to simulate heat conduction and wave dissipation processes. By solving the eigenvalues ​​and eigenvectors of the operator, high-dimensional nonlinear spatiotemporal business financial relationships are projected onto a low-dimensional spectral space. Within the spectral space, the algorithm performs adaptive evolutionary clustering based on the energy concentration of feature spectral lines, automatically aligning and classifying heterogeneous data with different spatiotemporal scales into a collaborative feature domain. The collaborative feature domain is a unified low-dimensional feature space that simultaneously possesses physical and economic meaning, in which data with different acquisition frequencies and resolutions obtain a unified spatiotemporal reference. The graph Laplacian operator is a difference operator defined on the graph structure, and its mathematical form is... ,in Let the node degree matrix be... For the adjacency weight matrix, the multi-scale version uses... The algorithm employs polynomial filter banks of different orders. The edge weights are calculated by jointly normalizing the spatiotemporal distance and frequency differences between data source nodes, with a recommended normalization coefficient range of 0.01–1.00. This normalization coefficient was determined through iterative experimental analysis of alignment errors from multiple sets of measured ocean multi-source data. The energy aggregation threshold for adaptive evolutionary clustering was determined through statistical analysis of the spectral space feature value distribution, with a recommended range of 0.05–0.30 determined after multiple rounds of iterative experiments. The technical effect of the evolutionary dynamics map clustering algorithm based on the multi-scale map Laplace operator is as follows: This algorithm integrates spectral theory with the principle of physical wave dissipation, enabling the alignment process of multi-source heterogeneous data to no longer rely on the linear assumptions of traditional interpolation. Instead, it automatically searches for the intrinsic spatiotemporal correlation structure between data sources in the map space with physical energy dissipation as a geometric constraint. This fundamentally suppresses the problem of pseudo-features caused by heterogeneous interpolation, ensuring that the features relied upon for subsequent financial decisions possess both physical and economic self-consistency, significantly improving the authenticity and reliability of multi-source fusion features.

[0028] The specific implementation of the complex time-varying diffusion dynamics denoising algorithm based on nonlocal mean field and fractional calculus is as follows: The algorithm first treats the original noisy business data time series as a virtual two-dimensional complex electromagnetic fluid field; it introduces fractional-order differential operators simultaneously in the time and frequency domains to construct a non-traditional partial differential diffusion equation. The recommended range for the fractional order is 0.3–0.9, determined through iterative analysis of denoising experiments on multiple sets of simulated ocean noise data; the diffusion coefficient is dynamically determined by the nonlocal mean structure similarity, with a recommended range for the structure similarity window size being 5–21 sampling points, determined through experimental measurement of the signal-to-noise ratio improvement; high-order diffusion triggered in the noise region is rapidly and smoothly smoothed, while real business mutations and financial inflection point signals are completely preserved due to the protection of fractional-order long memory characteristics; the fractional-order differential operator is a non-integer differential operator, compared to integer-order operators... It possesses long memory characteristics, enabling it to distinguish between slowly drifting noise and real business mutation signals. The nonlocal mean structural similarity is calculated in the time series to determine the structural similarity between the current neighborhood and other global neighborhoods. Regions with high similarity have small diffusion coefficients, thus protecting the signal, while regions with low similarity have large diffusion coefficients, thus smoothing noise. The technical effect of the complex time-varying diffusion dynamics denoising algorithm based on nonlocal mean field and fractional calculus is as follows: This algorithm overcomes the limitations of traditional robust principal component analysis and Kalman filtering, which are only applicable to specific noise distributions. Through the fractional long memory effect, it removes Gaussian and non-Gaussian mixed noise while strictly protecting low-frequency mutation signals such as financial inflection points. This allows the data entering the subsequent financial decision-making model to achieve a dynamic balance between signal fidelity and noise suppression depth, fundamentally eliminating the risk of noise pollution misleading disaster prevention budget preparation.

[0029] The specific process of the dimensionality reduction technology combining the autoencoder and t-SNE is as follows: The autoencoder first encodes the high-dimensional business feature vector into a low-dimensional latent space, and then the decoder reconstructs the original features. The encoder is trained using the reconstruction error as the loss function to retain the nonlinear topological structure. Subsequently, the low-dimensional latent vector output by the encoder is used as input, and the t-SNE algorithm is executed to further optimize the distribution of each sample point in the low-dimensional space, so that the original high-dimensional nonlinear topological relationship is retained in the low-dimensional core business indicator set. The recommended dimension range of the low-dimensional core business indicator set is 8 to 32, which is determined by comparing the loss values ​​of subsequent prediction models under different dimensions through multiple rounds of experiments. The nonlinear topological structure feature is the nearest neighbor relationship between high-dimensional data points determined by the nonlinear manifold. After dimensionality reduction, this relationship should remain consistent in the low-dimensional space. The redundant orthogonal basis is a high-dimensional feature component with extremely low contribution to the business prediction target, which is identified and removed after being analyzed and removed by the autoencoder reconstruction error weight.

[0030] The adaptive spatiotemporal feedback jump graph neural network model based on wave equation constraints is based on deep learning. Its specific structure is as follows: the model consists of a spatial graph convolutional branch and a time-gated recurrent unit branch, coupled through an adaptive spatiotemporal feedback jump mechanism. The spatial graph convolutional branch takes a low-dimensional core business indicator set and a weighted graph of the sea area topology as input, and uses four graph convolutional layers to extract topological correlation features between sea areas layer by layer. The output feature dimensions of each graph convolutional layer are 64, 128, 128, and 256 respectively, and the layers are connected by batch normalization and activation functions. Within each graph convolutional layer, neurons are selectively connected using a sparse adjacency matrix, transmitting features only between node pairs with edges in the graph topology, avoiding redundant computation caused by full connections. At the CUDA stream allocation level, the forward computation of the four graph convolutional layers is allocated to four independent CUDA streams, numbered sequentially. For Stream-0 to Stream-3, parallel execution between layers is achieved. At the memory allocation level, the node feature matrix and adjacency matrix of the graph convolutional layer are allocated to GPU memory to accelerate matrix multiplication. The time-gated loop unit branch takes the time series of buoy array shore three-dimensional current velocity and temperature fluctuation data and equipment maintenance financial cost flow as input, and uses an improved gated loop unit for processing. The hidden state dimension of the gated loop unit is 128, and the recommended time series window length is 24–168 time steps, determined through multi-period analysis experiments of measured ocean time series data. At the memory allocation level, the gated loop... The hidden state tensor of the loop unit is allocated to memory and fused with the graph convolution features in GPU memory through an asynchronous copy mechanism. At the CUDA thread block allocation level, the size of the graph convolution matrix multiplication thread block is set to 128, and the size of the gated loop unit matrix operation thread block is set to 64. At the data loop batch allocation level, time-series data is divided into batches according to a sliding window, with the recommended batch size ranging from 32 to 128 samples, determined through iterative experiments on training convergence speed and memory usage. The specific implementation of the adaptive spatiotemporal feedback jump mechanism is as follows: when the model detects that the physical input index deviates from the normal threshold, it skips the normal input. The intermediate hidden layer directly injects the physical wave equation constraint terms with the Laplace operator into the final financial loss prediction layer; the normal threshold is determined by statistical analysis of historical buoy data, with a recommended anomaly quantile range of 1% to 5%, determined after multiple rounds of experiments comparing prediction errors; the physical wave equation constraint terms are superimposed on the prediction layer input in the form of residuals from the Laplace operator acting on node features, with a recommended constraint weight coefficient range of 0.01 to 0.10, determined after multiple rounds of training experiments comparing validation set losses; the constraint weight coefficients are compared with the attention factor output by the adaptive dynamic attention adjustment function. There exists a linear functional relationship, specifically the constraint weight coefficients. , The value range is determined by the adaptive dynamic attention adjustment function; the steps for establishing the training dataset of the adaptive spatiotemporal feedback jump graph neural network model based on wave equation constraints specifically include: collecting historical data of buoy array three-dimensional current velocity and temperature fluctuations at the shore in different sea areas and seasons, synchronizing the corresponding period's equipment maintenance financial cost flow records, aligning the data with timestamps and labeling outliers to form input-output sample pairs; performing data augmentation on the sample pairs, including temporal random pruning, noise injection, and random masking of the sea area topology map; the steps for training the adaptive spatiotemporal feedback jump graph neural network model based on wave equation constraints specifically include: using the root mean square error loss function to calculate the error between the predicted financial cost value and the actual value, superimposing the residual loss of the physical wave equation constraint term to form the overall loss function; using the adaptive dynamic attention adjustment function The attention weights are dynamically adjusted based on the current validation set loss trend; the mini-batch stochastic gradient descent method is used to backpropagate and update all model parameters, and the iteration round is determined by the early stopping strategy based on the validation set loss; the technical effect of the adaptive spatiotemporal feedback jump graph neural network model based on the wave equation constraint is as follows: the model embeds the hydrodynamic wave equation as an internal physical constraint into the graph neural network, so that financial cost prediction no longer relies purely on data fitting, but performs feature search under the constraint of physical laws, overcoming the defect of pure black box networks that are prone to getting trapped in local optima in the high-dimensional sparse scenario of big data in the marine environment; the adaptive feedback jump mechanism enables the network to bypass the intermediate layer and directly activate the physical constraint path when it detects physical anomalies, ensuring computational efficiency while making the prediction results strictly self-consistent in physics, and improving the robustness and credibility of dynamic financial cost prediction in the variable marine environment.

[0031] The adaptive dynamic attention adjustment function is defined as: attention factor Loss value of the current round of validation set Current training round Compared with the initial loss value Joint calculation, the formula is: ;when At that time, the range of values ​​for the attention weight parameter is: ~ To address the unstable feature alignment phase in the early stages of training, a larger attention weight is used to strengthen physical constraint guidance; when At that time, the range of values ​​for the attention weight parameter is: ~ This corresponds to the stable convergence phase in the middle of training; when At that time, the range of values ​​for the attention weight parameter is: ~ This corresponds to the fine-tuning stage in the later stages of training; the attention weight ranges for each interval were determined through iterative analysis of the loss curves of multiple training experiments and validation sets.

[0032] The specific implementation of the manifold-aligned deep flow model is as follows: the model is composed of a series of reversible affine coupling transformation layers in a deep cascade. The input is a high-dimensional non-Gaussian business distribution vector composed of sonar cloud image features, robotic arm power consumption parameters, variable costs, and capitalization budgets. Each reversible affine coupling transformation layer splits the input vector into two parts: one part is transformed by scaling and translation functions, while the other part remains unchanged. The scaling and translation functions are implemented by fully connected subnetworks. During the transformation process, the probability density is continuously transformed according to the determinant of the Jacobian matrix, mapping the non-Gaussian distribution to the standard multivariate Gaussian distribution space layer by layer. In the reverse decision-making process, the standard distribution is mapped back to the real business decision solution space through the inverse transformation of each layer. The recommended range for the number of cascaded layers is 8 to 24, determined through multiple rounds of experimental comparison of log-likelihood loss at different numbers of layers. The determinant of the Jacobian matrix is ​​the input of the reversible transformation layer. The determinant of the partial derivative matrix with respect to the input is used to accurately track the change in probability density during the transformation process. The standard multivariate Gaussian distribution space is a multidimensional normal distribution space with a mean of zero vector and a covariance of identity matrix. Sampling and probability density evaluation in this space have low computational cost. The technical effect of the deep flow model based on manifold alignment is that the model establishes an accurate bidirectional mapping between the real non-Gaussian distribution of business and financial data and the standard Gaussian space through invertible transformation, so that the probability structure of the financial decision solution space is rigorously described mathematically, overcoming the defect of traditional prediction models that ignore the non-Gaussianity of the probability distribution of business and financial data. In the inverse transformation stage, decision-makers can systematically explore disaster prevention budget and capitalization budget schemes under different risk scenarios by sampling and inverse mapping in the standard Gaussian space, so that the business and financial integration decision has probabilistic interpretability and global optimality.

[0033] The establishment of the collaborative feature domain depends on the convergence accuracy of the eigenvalue decomposition of the pattern Laplacian operator, and the recommended range for the convergence determination residual threshold is [insert range here]. ~ The determination was made through iterative analysis of eigenvalue decomposition experiments on multiple sets of dynamic spatiotemporal weighted graphs; the business meaning of each indicator in the low-dimensional core business indicator set must be manually mapped and interpreted in combination with domain knowledge after dimensionality reduction to ensure that the subsequent financial budget compilation has business interpretability; the three-dimensional current velocity at the shore is the ocean current velocity component measured by the buoy array in the direction of the three-dimensional spatial coordinate axis, and the shore is a joint expression of time and shore-based reference coordinate system; the capitalized budget is a budget item that records the purchase and installation expenditure of marine exploration equipment into the asset account rather than the current period expense in accordance with accounting standards.

[0034] The specific implementation of step S01 is as follows: First, the multispectral image data stream from the satellite remote sensing system, the time series of three-dimensional current velocity and temperature fluctuations output by the deep-sea buoy array, and the financial cost flow records from the equipment maintenance financial reimbursement system are respectively accessed. These three types of data differ significantly in acquisition frequency and spatial resolution. Satellite remote sensing data has wide spatial coverage but low temporal resolution; buoy array data has a high acquisition frequency but only covers discrete spatial nodes; and the financial cost flow records are recorded in an event-driven manner. Therefore, the three cannot be directly spliced. To address this, the geographically scattered buoy observation points, satellite observation grid points, and financial reimbursement event nodes are abstracted into a dynamic spatiotemporal weighted graph. Nodes in the graph represent various data sources, and the edge weights between nodes are calculated by jointly normalizing the spatiotemporal distance between nodes and the differences in acquisition frequency. The recommended normalization coefficient range is 0.01–1.00. A multi-scale Laplacian operator is constructed on this graph structure, with the mathematical form as follows: ,in Let the node degree matrix be... For the adjacency weight matrix, the multi-scale version uses... This is achieved by applying polynomial filter banks of different orders. Then... Perform eigenvalue decomposition, and the recommended range for the residual threshold to determine convergence is: ~ The obtained eigenvectors are used as the spectral basis to project the high-dimensional nonlinear spatiotemporal economic relationship into a low-dimensional spectral space. Within the spectral space, adaptive evolutionary clustering is performed based on the energy concentration of the characteristic spectral lines. The recommended energy concentration threshold range is 0.05–0.30. Heterogeneous data with different spatiotemporal scales are automatically aligned and classified into a collaborative feature domain. This collaborative feature domain is a unified low-dimensional feature space that possesses both physical and economic significance, providing a unified spatiotemporal reference for data with different acquisition frequencies and resolutions.

[0035] The specific implementation of step S02 is as follows: The collaborative feature domain data output in step S01 is regarded as a virtual two-dimensional complex electromagnetic fluid field. Fractional-order differential operators are introduced simultaneously in the time and frequency domains to construct a non-traditional partial differential diffusion equation. The recommended order of the fractional-order differential operator is 0.3 to 0.9, determined through iterative analysis of denoising experiments on simulated ocean noise data. The diffusion coefficient is dynamically determined by the non-local mean structural similarity, that is, the structural similarity between the current neighborhood and other global neighborhoods is calculated in the time series. The recommended structural similarity window size is 5 to 21 sampling points. For noise regions with low similarity, a larger diffusion coefficient is used to quickly smooth the noise; for regions with high similarity but real business mutations or financial inflection points, a smaller diffusion coefficient is used to protect signal integrity. The long memory characteristic of the fractional-order differential operator enables the algorithm to distinguish between slowly drifting noise and real business mutation signals, overcoming the limitation of traditional Kalman filtering being only applicable to Gaussian noise distributions. Finally, a high signal-to-noise ratio business-finance fusion input dataset is output for use in subsequent steps.

[0036] The specific implementation of step S03 is as follows: The input is the high signal-to-noise ratio (SNR) business-finance fusion input dataset output from step S02. This dataset has high dimensionality and contains a large number of sparse and redundant features. First, an autoencoder is trained. The encoder compresses the high-dimensional business-finance feature vectors into a low-dimensional latent space. The decoder reconstructs the original features from the low-dimensional latent vectors. The encoder is trained by backpropagation using the reconstruction error as the loss function, so that the latent space retains the nonlinear topological structure features. After the encoder converges, the t-SNE algorithm is executed using its output low-dimensional latent vectors as input. By minimizing the difference in conditional probability distribution between sample pairs in the high-dimensional and low-dimensional spaces, the distribution of each sample point in the low-dimensional space is further optimized, so that the original high-dimensional nonlinear topological relationship is preserved in the low-dimensional core business indicator set. The recommended dimensionality range for the low-dimensional core business indicator set is 8 to 32. High-dimensional feature components with extremely low contribution are identified by the autoencoder reconstruction error weight analysis and then removed as redundant orthogonal bases. After dimensionality reduction, each low-dimensional indicator needs to be manually mapped and interpreted in conjunction with domain knowledge to ensure that the subsequent financial budget compilation has business interpretability.

[0037] The specific implementation of step S04 is as follows: The model consists of a spatial graph convolutional branch and a time-gated recurrent unit branch. The spatial graph convolutional branch takes a low-dimensional core business indicator set and a weighted graph of the sea area topology as input, and uses four graph convolutional layers to extract the topological correlation features between sea areas layer by layer. The output feature dimensions of each layer are 64, 128, 128, and 256, respectively. The layers are connected by batch normalization and activation functions, and features are only transmitted between node pairs with edges in the graph topology. At the parallel computing level, the forward computation of the four graph convolutional layers is distributed to four independent computational flows to achieve parallel execution between layers. The node feature matrix and adjacency matrix are allocated to GPU memory to accelerate matrix multiplication. The time-gated recurrent unit branch takes the time series of buoy array shore three-dimensional current velocity, temperature fluctuation data, and equipment maintenance financial cost flow as input. The hidden state dimension of the gated recurrent unit is 128, the recommended range of time series window length is 24 to 168 time steps, and the time series data is divided into batches according to the sliding window, with the recommended range of sample size per batch being 32 to 128. The two branches are coupled through an adaptive spatiotemporal feedback jump mechanism. When the physical input index deviates from the normal threshold determined by historical buoy data statistics (the recommended range for anomaly detection quantiles is 1% to 5%), the conventional intermediate hidden layer is skipped, and the constraint terms of the Laplace operator physical wave equation are directly injected into the final financial loss prediction layer in the form of residuals, with the constraint weight coefficients... , Adaptive dynamic attention adjustment function Calculation. The training uses the sum of the root mean square error loss and the physical wave equation constraint residual loss as the overall loss function, and the iteration rounds are determined by the validation set loss early stopping strategy.

[0038] The specific implementation of step S05 is as follows: The input is a high-dimensional non-Gaussian business distribution vector composed of sonar cloud image features, robotic arm power consumption parameters, variable costs, and capitalization budgets. The deep flow model is composed of 8 to 24 layers of reversible affine coupling transformation layers in a deep cascade. Each layer splits the input vector into two parts: one part is transformed by scaling and translation functions implemented through a fully connected subnetwork, while the other part remains unchanged. During the transformation process, the probability density is continuously deformed according to the determinant of the Jacobian matrix, mapping the real non-Gaussian business distribution to the standard multivariate Gaussian distribution space layer by layer. In the reverse decision-making stage, the sampling points in the standard multivariate Gaussian distribution space are mapped back to the real business decision solution space through the inverse transformation of each layer, systematically exploring disaster prevention budget compilation and capitalization budget schemes under different risk scenarios. The accurate tracking of the determinant of the Jacobian matrix ensures that the change in probability density during the transformation process is measurable, enabling a rigorous mathematical description of the probability structure of the decision solution space and outputting budget recommendations with probabilistic interpretability.

[0039] It should be noted that the key technologies of this invention include: an evolutionary dynamics graph clustering algorithm based on multi-scale graph Laplace operator, which combines graph theory with the principle of physical wave dissipation, transforming the basis for heterogeneous data alignment from linear distance to spectral space clustering degree under physical energy dissipation, thus eliminating spurious features caused by spatiotemporal scale mismatch from the source; a complex time-varying diffusion dynamics denoising algorithm based on fractional calculus, which utilizes fractional-order long memory effect to simultaneously suppress Gaussian and non-Gaussian mixed noise under the adjustment of dynamic diffusion coefficient, and strictly protects low-frequency abrupt signals such as financial inflection points; and a deep flow model based on manifold alignment, which establishes an accurate bidirectional mapping between the real non-Gaussian financial distribution and the standard Gaussian space through reversible affine coupling transformation, enabling the decision solution space to have global explorability at the probabilistic level. The synergistic effect of the three key technologies is that graph alignment provides a real and authentic feature foundation for subsequent stages, fractional-order denoising further improves signal fidelity, and the deep flow model outputs a business and financial decision-making scheme with a strict probabilistic structure under the guarantee of high-quality feature input. All three are indispensable and together constitute a complete technical closed loop from raw multi-source heterogeneous data to reliable business and financial fusion decision-making.

[0040] It should be noted that in the context of high-dimensional and sparse big data in the dynamic marine environment, when abnormal disturbances occur in the physical environment (such as sudden changes in ocean current speed or unexpected equipment downtime leading to sharp fluctuations in financial costs), purely data-driven deep learning prediction models are prone to getting trapped in local optima in the high-dimensional and sparse feature space. This results in a systematic deviation between the predicted financial costs and the actual values, thus misleading disaster prevention budgeting. The reason for this technical problem is that in high-dimensional and sparse scenarios, the effective training sample density is insufficient to support a purely data-driven model in fully exploring the global optimum in the feature space. When abnormal disturbances occur in the physical environment that exceed the historical sample distribution range, the model lacks embedded constraints on physical laws and cannot correctly generalize the causal relationship between abnormal physical inputs and financial losses. This leads to a local optimum determined by local training samples, resulting in physically inconsistent predictions. Common solutions to this technical problem include increasing the amount of training data or introducing regularization techniques to alleviate overfitting. However, the acquisition cost of multi-source ocean data is extremely high, and historical anomaly event samples are scarce, making it difficult to simply increase the amount of data in practice. Traditional regularization methods only constrain model parameters statistically and cannot embed physical laws such as fluid dynamics as hard constraints into the network structure. Therefore, the physical consistency of prediction results cannot be guaranteed under abnormal physical conditions. This invention effectively solves this technical problem. The adaptive spatiotemporal feedback jump graph neural network model based on wave equation constraints embeds the fluid dynamics wave equation as a Laplace operator residual constraint term into the prediction layer, enabling financial cost prediction to perform feature search under physical constraints, rather than purely relying on data fitting. When the adaptive spatiotemporal feedback jump mechanism detects that the physical input index exceeds the normal threshold determined by historical buoy data statistics, the model automatically skips the conventional intermediate hidden layer and directly activates the physical constraint path, injecting the physical wave equation constraint term into the final prediction layer. The constraint weight coefficient changes linearly with the attention factor output by the adaptive dynamic attention adjustment function. Larger constraint weights are used in the early stages of training to strengthen physical guidance, while the constraint weights are reduced in the later stages of training to finely fit data details, thus achieving a dynamic balance between physical consistency and data fitting accuracy. This mechanism enables the model to correctly generalize the causal relationship between abnormal physical inputs and financial losses even in high-dimensional sparse scenarios where historical abnormal samples are scarce, relying on physical laws to constrain the model. This overcomes the defect of pure black-box networks getting stuck in local optima under such conditions, and significantly improves the robustness and reliability of dynamic financial cost prediction in the variable marine environment.

[0041] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described marine industry-finance integration decision-making method based on big data analysis.

[0042] A third aspect of the present invention provides a marine industry and finance integration decision-making system based on big data analysis, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0043] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in its adoption of a multi-source heterogeneous data alignment framework based on spectral theory and the principle of physical wave dissipation, and on this basis, it constructs a complete information processing link from raw data to business decision-making. First, the reason why existing linear interpolation methods produce pseudo-features is that they implicitly assume that each data source has linear continuity in space and time. However, the essential differences in acquisition frequency and spatial resolution between satellite remote sensing, buoy arrays, and financial cost flow make this assumption invalid under actual working conditions. This invention abstracts each data source node and its spatiotemporal correlation into a dynamic spatiotemporal weighted graph, constructs a multi-scale graph Laplace operator on the graph structure, and projects the high-dimensional nonlinear spatiotemporal business relationship to a low-dimensional spectral space by solving the eigenvalues ​​and eigenvectors of this operator. In the spectral space, the basis for data alignment is no longer linear distance, but energy concentration determined by the physical wave dissipation process. This allows heterogeneous data with different spatiotemporal scales to be automatically classified into a cooperative feature domain under the constraint of physical self-consistency, thus suppressing the generation of pseudo-features from the source. Secondly, after obtaining highly realistic business and financial integration features, this invention employs a complex time-varying diffusion dynamics denoising algorithm based on fractional calculus, utilizing the fractional-order long memory effect to distinguish slowly drifting noise from real financial inflection point signals, further ensuring the signal-to-noise ratio of the data entering the prediction model. Subsequently, the dimensionality reduction technology combining autoencoders and t-SNE compresses the feature dimensions while preserving the nonlinear topological structure, ensuring that the low-dimensional core business indicator set faithfully reflects the manifold structure of the original data. Building upon this, an adaptive spatiotemporal feedback jump graph neural network model constrained by the wave equation uses fluid dynamics physical laws as embedded constraints, enabling financial cost prediction to perform feature search under the guidance of physical laws, avoiding the pure data-driven model from getting trapped in local optima in high-dimensional sparse scenarios. Finally, a deep flow model based on manifold alignment establishes an accurate bidirectional mapping between the real non-Gaussian business and financial distribution and the standard Gaussian space through reversible affine coupling transformation, providing a rigorous mathematical description of the probability structure of the decision solution space and supporting the systematic exploration of budget schemes under different risk scenarios. The above steps are logically sequential, with each step's output providing higher-quality input for the next step. The self-consistency of the entire chain ensures the reliability of the final business and financial decisions.

[0044] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0045] The specific implementation of step S01 is as follows: The buoy observation points, satellite observation points, and financial reimbursement nodes scattered throughout the sea area are abstracted into a dynamic spatiotemporal weighted graph, denoted as... ,in Let be a set of nodes, and let the number of nodes be denoted as . , Let be the set of edges. This is the adjacency weight matrix, and its elements are... Represents a node With nodes The spatiotemporal correlation strength between nodes. Node degree matrix. For a diagonal matrix, the formula for calculating the diagonal elements is as follows: ; In the formula, For nodes The degree value, For nodes With nodes Edge weights between them Let be the total number of nodes. The graph Laplace operator is defined as follows: ; In the formula, For the pattern Laplace matrix, Let the node degree matrix be... This is the adjacency weight matrix. The formula for calculating the edge weights is as follows: ; In the formula, For nodes With nodes The edge weights between them are a dimensionless similarity measure. For nodes With nodes The spacetime distance between them For nodes The sampling frequency, For nodes The sampling frequency, For the spatiotemporal distance normalized scale parameter, and Dimensions consistent The normalized scaling parameter for frequency differences, and The dimensions are consistent, thus ensuring that the exponential term is dimensionless as a whole. and The joint normalization coefficient, empirically ranging from 0.01 to 1.00, was determined through iterative experimental analysis of the alignment errors of multiple sets of measured multi-source ocean data. The multi-scale map Laplace operator is obtained by... The implementation is achieved by applying polynomial filter banks of different orders, as described below: ; In the formula, For multi-scale pattern Laplacian operator, This represents the highest order of the polynomial filter, with an empirical value of 3 to 5. For the first The filter coefficients are obtained through training and fitting. For the Laplace matrix of the pattern Power of 1. (The last part is incomplete and likely refers to a power-law or similar term.) The solution for eigenvalues ​​and eigenvectors is as follows: ; In the formula, For the first 1 eigenvalue, For the corresponding feature vector, For feature serial number. (The preceding text appears to be incomplete and requires further context.) The projection matrix is ​​composed of eigenvectors. Projecting high-dimensional spatiotemporal business financial relationships onto a low-dimensional spectral space, as follows: ; In the formula, The characteristic matrix projected onto the low-dimensional spectral space. This is the transpose of the projection matrix. The original multi-source heterogeneous data feature matrix is ​​composed of satellite remote sensing data, buoy three-dimensional flow velocity and temperature fluctuation data, and equipment maintenance financial cost flow data, spliced ​​in node order. The formula for calculating energy concentration in the spectral space is expressed as follows: ; In the formula, For the first The energy concentration of each characteristic spectral line is a dimensionless ratio. For the first 1 eigenvalue, For the first 1 eigenvalue, is the dimension of the spectral space. When When the energy concentration threshold is greater than the threshold value, the data sources associated with the corresponding characteristic spectral lines are grouped into the same co-occurrence feature domain. The empirical threshold value ranges from 0.05 to 0.30, determined through statistical analysis of the spectral space eigenvalue distribution and multiple rounds of iterative experiments. The empirical range for the convergence criterion residual threshold is... ~ After iterative analysis of eigenvalue decomposition experiments on multiple sets of dynamic spatiotemporal weighted graphs, the residual calculation formula is expressed as follows: ; In the formula, The eigenvalue decomposition convergence residuals, and Let be the vectors at both ends of the aforementioned characteristic equation. Since they have the same dimensions, the difference norm is meaningful.

[0046] The specific implementation of step S02 is as follows: the original time series of the noisy business is regarded as a virtual two-dimensional complex electromagnetic fluid field, denoted as... ,in For the real component, corresponding to the business measurement signal, The imaginary component corresponds to the financial transaction signal; after normalization, both components have the same dimensions. For spatial coordinates, Using time as the coordinate, The unit is the imaginary unit. The complex time-varying diffusion kinetics equation is expressed as follows: ; In the formula, For the fractional differential operator in the time direction, the order is... The empirical value range is 0.3 to 0.9, determined through iterative analysis of denoising experiments on multiple sets of simulated ocean noise data. For the fractional gradient operator in spatial directions, the order is... The empirical value range is 0.3 to 0.9. The nonlocal mean diffusion coefficient is denoted by , and the dimensionless adjustment factor is denoted by . For random disturbance terms, and The dimensions are consistent. The diffusion coefficient is dynamically determined by the nonlocal mean structure similarity, as described below: ; In the formula, coordinates The diffusion coefficient at that location, For The neighborhood vector centered on the center For other positions in the sequence The neighborhood vectors at the location are both taken from The real and imaginary parts of the local sequence fragments have consistent dimensions. The neighborhood window size empirically ranges from 5 to 21 sampling points and is determined by experimentally measuring the signal-to-noise ratio improvement. The structural similarity smoothing parameter is... The dimensions are consistent, thus ensuring that the exponential term is dimensionless overall. Regions with high structural similarity. Small values ​​are used to protect the signal, and regions with low structural similarity are also included. A larger value smooths out the noise. The formula for calculating the improvement in signal-to-noise ratio after denoising is as follows: ; In the formula, The signal-to-noise ratio improvement is expressed in units of [value missing]. , For signal power, The noise power after denoising. The noise power before denoising is given. All three have the same dimension, and their ratio is dimensionless.

[0047] The specific implementation of step S03 is as follows: The autoencoder encoding process is described as follows: ; In the formula, The low-dimensional latent vector output by the encoder. For encoder activation function, Here is the encoder weight matrix. For high signal-to-noise ratio business and financial fusion input data feature vectors, For encoder bias vector, and The dimensions are consistent, therefore they can be added together. The decoding and reconstruction process is described as follows: ; In the formula, To reconstruct the feature vector, and Dimensions consistent For the decoder activation function, This is the decoder weight matrix. Let be the decoder bias vector. The reconstruction error loss function is expressed as follows: ; In the formula, To reconstruct the error loss, For the first Original feature vectors of each sample For the first Reconstruct feature vectors from each sample. The total number of samples. Redundant orthogonal bases are identified through reconstruction error weighting analysis, and the weighting formula is expressed as follows: ; In the formula, For the first The contribution weights of each feature component, The first eigenvector in the original feature vector One portion, To reconstruct the error loss for the first Partial derivatives of each eigencomponent, Feature components with absolute values ​​below a set threshold are identified as redundant orthogonal bases and removed. The empirical threshold is one-tenth of the mean absolute value of the partial derivatives of the reconstruction error loss with respect to all feature components. The t-distributed random nearest neighbor embedding algorithm is executed on the input. The conditional probability of two points in high-dimensional space is expressed as follows: ; In the formula, Samples in a high-dimensional latent space With samples Let be the conditional probability of the neighbors. , , Samples ,sample ,sample latent vectors, For the sample The Gaussian kernel bandwidth centered on the center, and The dimensions are consistent. The joint probability expression of the two points where low-dimensional core business indicators are concentrated is as follows: ; In the formula, A concentrated sample of low-dimensional core business indicators With sample The joint probability, , , , Samples ,sample ,sample ,sample The coordinate vectors in the low-dimensional core business indicator set. The optimization objective is to minimize the relative entropy between the high-dimensional and low-dimensional distributions, as stated below: ; In the formula, For relative entropy loss, For the joint probability in high-dimensional space, by and Obtained by symmetric averaging. Samples in a high-dimensional latent space With samples Let be the conditional probability of the neighbors. This represents the joint probability in a low-dimensional space. The empirical range of the dimensions of the low-dimensional core business indicator set is 8 to 32, determined through multiple rounds of experimental comparison of the loss values ​​of subsequent prediction models under different dimensions.

[0048] The specific implementation of step S04 is as follows: the spatial graph convolution branch... The layer propagation formula is expressed as follows: ; In the formula, For the first Layer node feature matrix, For the first The layer node feature matrix, where the input of layer 0 is the concatenation result of the low-dimensional core business indicator set and the node features of the marine topology weighted map. To add the adjacency matrix after self-connection, for The corresponding degree matrix, For the first Layer trainable weight matrix, The activation function has output feature dimensions of 64, 128, 128, and 256 for each layer, respectively. The formula for updating the hidden state of the time-gated recurrent unit branch is expressed as follows: ; ; ; ; In the formula, To update the door, For the reset gate, both are dimensionless gating coefficients, with values ​​ranging from 0 to 1. for The real-time input of buoy flow rate, temperature, and financial cost flow sequence data. This is the hidden state from the previous moment. This represents the hidden state at the current moment, with a dimension of 128. As a candidate hidden state, and Dimensions consistent , , For the input weight matrix, , , The hidden state weight matrix is... , , For bias vectors, For element-wise multiplication, For gated activation functions, This is the hyperbolic tangent activation function. The constraint terms of the physical wave equation are expressed as follows: ; In the formula, The physical wave equation constraint residuals applied to the nodal characteristics, and Dimensions consistent For the pattern Laplace matrix, This is the hidden state vector output by the time-gated loop unit branch. The final output of the financial loss prediction layer is expressed as follows: ; In the formula, This is the result of dynamic financial cost forecasting. The output features are from the 4th layer of the spatial graph convolution branch. For feature concatenation, the concatenated vector is mapped to a fully connected layer. The same dimensions ensure that addition is valid. To constrain the weight coefficients, and the attention factor There exists a linear functional relationship, expressed as follows: ; In the formula, To constrain the weighting coefficients, empirical values ​​range from 0.01 to 0.10; these are dimensionless coefficients. The attention factor, output by the adaptive dynamic attention adjustment function, is also a dimensionless quantity. The criteria for the adaptive spatiotemporal feedback jump mechanism are stated as follows: ; In the formula, for Real-time physical input indicators The average of historical buoy data, and Dimensions consistent The standard deviation of historical buoy data, and Dimensions consistent , where is the standard deviation multiple corresponding to the anomaly detection quantile, and is a dimensionless coefficient. The empirical range for the anomaly detection quantile is 1%–5%, determined after multiple rounds of experiments comparing prediction errors. When the condition is met, the conventional intermediate hidden layer is skipped and the quantile is directly... Injected into the prediction layer. The overall loss function is expressed as follows: ; In the formula, For the overall loss function, For the first The true value of financial costs for each sample For the first The projected financial costs for each sample are consistent in scale. Let be the total number of samples. The adaptive dynamic attention adjustment function is expressed as follows: ; In the formula, The attention factor is a dimensionless quantity. This represents the loss value of the validation set in the current round. Since the initial loss values ​​are the same in dimension, the ratio is dimensionless. For the current training round, This is an exponential term that decays with each training epoch; it is also a dimensionless quantity. When At that time, the empirical range of values ​​for the attention weight parameter is: ~ ;when At that time, the empirical range of values ​​for the attention weight parameter is: ~ ;when At that time, the empirical range of values ​​for the attention weight parameter is: ~ The range of each interval was determined through iterative analysis of the loss curves of multiple training experiments and validation sets.

[0049] The specific implementation of step S05 is as follows: the high-dimensional non-Gaussian distribution vector is denoted as... ,in Features of sonar cloud images These are the power consumption parameters for the robotic arm. For variable costs, For capitalization budgeting. The reversible affine coupling transformation layer decomposes the input vector into and The two parts, with the forward transformation described below: ; ; In the formula, and These are the two vector parts of the output after transformation. This is a scaling function; the output is a dimensionless exponential term. The translation function outputs the same as... Both have the same dimensions and are implemented using fully connected subnetworks. This is an element-wise product. The determinant of the transformed Jacobian matrix is ​​expressed as follows: ; In the formula, The determinant of the Jacobian matrix of the transform layer output with respect to the input. The output vector of this layer is... and Assembled by splicing The input vector for this layer is... and Assembled by splicing The output vector of the scaling function Each component. The probability density transformation relationship after multi-level cascading is expressed as follows: ; In the formula, Let be the probability density of the original business distribution vector. For the standard multivariate Gaussian distribution space and The corresponding mapping point, for The probability density at that location, The number of cascaded layers is empirically determined to be between 8 and 24, and this value was determined through multiple rounds of experiments comparing the log-likelihood loss at different numbers of layers. and The first The output and input vectors of the layer transform. The log-likelihood loss function is expressed as follows: ; In the formula, For log-likelihood loss, For the first A high-dimensional non-Gaussian distribution vector of the sample. Let be the total number of samples. The inverse transform is expressed as follows: ; ; In the formula, and The business decision solution space vector recovered by the inverse transformation consists of two parts, which are sampled in the standard multivariate Gaussian distribution space. Then, the inverse transformation is performed layer by layer to obtain the solution space samples corresponding to the disaster prevention budget compilation and capitalization budget recommendations.

[0050] To better understand and implement this invention, a specific application scenario of the invention is provided below as an example 2: This example takes a marine industry-finance integration decision-making task in a certain deep-sea area as the background, and the data comes from continuous observation records of this sea area from January 2022 to December 2023. The dataset includes time series of three-dimensional current velocity and temperature fluctuations from 12 deep-sea buoy nodes, corresponding satellite remote sensing multispectral image sequences, and concurrent equipment maintenance financial cost records, covering a total of 730 days. The current velocity sampling frequency is 1 time / hour, the satellite image temporal resolution is 1 scene / day, and the financial cost records are event-driven.

[0051] First, step S01 is executed, abstracting the 12 buoy nodes, satellite observation grid points, and financial reimbursement event nodes into a dynamic spatiotemporal weighted graph. The edge weights between nodes are calculated using a joint normalization of spatiotemporal distance and frequency difference, with a normalization coefficient of 0.15. A multi-scale graph Laplacian operator is constructed on the graph structure, with polynomial filters set to four orders: 1, 2, 4, and 8. The Laplacian matrix is ​​then processed... Perform eigenvalue decomposition, and set the convergence criterion residual threshold to [value missing]. The energy aggregation threshold in the spectral space was set to 0.12. Adaptive evolutionary clustering classified the three types of heterogeneous data into a unified collaborative feature domain. After alignment, the spatiotemporal reference of each data source was unified, eliminating the pseudo-features introduced by direct interpolation.

[0052] Then, step S02 is executed, mapping the collaborative feature domain data into a virtual two-dimensional complex field. The fractional-order differential operator is set to order 0.6, and the non-local mean structure similarity window size is set to 11 sampling points. Complex time-varying diffusion dynamics denoising is performed on all 730 days of time-series data. Gaussian noise and non-Gaussian burst noise are effectively suppressed. At the same time, the inflection point signals of financial costs caused by unplanned equipment shutdowns at multiple locations are completely preserved due to the fractional-order long memory effect. A high signal-to-noise ratio business-finance fusion input dataset is output. The feature comparison between the original data and the denoised data is as follows: Figure 2 As shown.

[0053] Next, step S03 is executed, where the high signal-to-noise ratio (SNR) business-finance fusion input dataset is input into the autoencoder. The encoding layer progressively compresses the original 256-dimensional feature vector into a 16-dimensional latent space. During training, the reconstruction error is used as the loss function, and the 16-dimensional latent vector is output after convergence. This latent vector is then used as input to execute the t-SNE algorithm, and the final dimension of the low-dimensional core business indicator set is set to 16. After dimensionality reduction, the 16 indicators are manually mapped and interpreted one by one, combining knowledge from marine engineering and finance, to clarify the corresponding physical or financial quantities, ensuring the business interpretability of subsequent budget compilation. The business meaning mapping results of each low-dimensional core business indicator are shown in Table 1. Table 1. Mapping Table of Business Meaning of Low-Dimensional Core Business Indicators

[0054] Then, step S04 is executed, inputting the 16-dimensional low-dimensional core business indicator set, along with the original buoy time series and financial cost flow data, into an adaptive spatiotemporal feedback jump graph neural network model constrained by the wave equation. The spatial graph convolutional branch uses four graph convolutional layers, with output feature dimensions of 64, 128, 128, and 256 respectively; the hidden state dimension of the temporally gated recurrent unit branch is 128, the time series window length is 72 time steps, and the number of samples per batch is 64. During training, the adaptive dynamic attention adjustment function adjusts the loss based on the current validation set. Compared with the initial loss and training rounds Calculate attention factor Dynamically adjust constraint weight coefficients In the 18th round, When the value drops below 0.3, the constraint weights enter the fine-tuning stage, and the value is [value to be filled in]. The early stopping strategy based on validation set loss was triggered in round 47, ending model training. A comparison of the model's dynamic financial cost prediction curve output on the test set with the actual value is shown below. Figure 3 As shown.

[0055] Finally, step S05 is executed, where the sonar cloud image features, robotic arm power consumption parameters, variable costs, and capitalization budget vectors are combined to form a high-dimensional non-Gaussian business and financial distribution vector. This vector is then input into a deep flow model based on manifold alignment, with 16 cascaded layers. The model uses 16 layers of reversible affine coupling transformation to map the real non-Gaussian business and financial distribution layer by layer to a standard multivariate Gaussian distribution space. Subsequently, 100 schemes are sampled in the standard space for low-risk, medium-risk, and high-risk scenarios, respectively. After an inverse transformation, these schemes are mapped back to the real decision solution space, outputting disaster prevention budget preparation and capitalization budget recommendations for the three scenarios, as shown in Table 2. Table 2 Summary of Budget Recommendations for Business-Finance Integration under Three Risk Scenarios

[0056] The technological advancements of this invention compared to traditional methods are reflected in the following aspects. Traditional methods rely on linear interpolation for aligning multi-source heterogeneous data. This linear assumption systematically introduces spurious features under conditions of spatiotemporal scale mismatch, contaminating all subsequent analysis stages. In contrast, this invention uses physical energy dissipation as a geometric constraint to perform alignment in the spectral space, fundamentally eliminating the mechanism for spurious feature generation. The features processed in subsequent stages possess both physical and economic self-consistency. Regarding noise processing, traditional Kalman filtering is effective against Gaussian noise but has limited ability to handle non-Gaussian noise, and it cannot strictly protect low-frequency abrupt signals while denoising. The fractional-order diffusion dynamics algorithm of this invention, through the synergistic effect of long memory effect and dynamic diffusion coefficient, suppresses mixed noise while protecting financial inflection point signals, simultaneously improving signal fidelity and noise suppression depth in a dynamic balance. At the decision-making level, traditional prediction models output point prediction results and lack a description of the non-Gaussianity of the probability distribution of business and financial data. The deep flow model of this invention establishes an accurate bidirectional mapping between the real non-Gaussian distribution and the standard Gaussian space, enabling decision-makers to systematically sample and inversely map all risk scenarios in the standard space, thus realizing a fundamental shift from point prediction to a probabilistic decision solution space.

[0057] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A marine industry-finance integrated decision-making method based on big data analysis, characterized in that, Includes the following steps: Collect satellite remote sensing, deep-sea buoy array, and shore three-dimensional current velocity and temperature fluctuation data, as well as equipment maintenance financial cost data. Utilize the evolutionary dynamics graph clustering algorithm based on multi-scale map Laplacian operator, and use physical energy dissipation as a geometric constraint to complete the spatiotemporal alignment and collaborative feature domain classification of multi-source heterogeneous data in the spectral space. A complex time-varying diffusion dynamics denoising algorithm based on nonlocal mean field and fractional calculus is applied to the spatiotemporally aligned multi-source heterogeneous data to remove Gaussian and non-Gaussian mixed noise in the deep-sea exploration data, resulting in a high signal-to-noise ratio fusion input dataset. High-dimensional sparse feature compression was performed on the high signal-to-noise ratio business and finance fusion input dataset using autoencoder and t-SNE fusion dimensionality reduction technology to construct a low-dimensional core business indicator set. The low-dimensional core business indicators set, along with the three-dimensional current velocity and temperature fluctuation data of the deep-sea buoy array and the financial cost flow of equipment maintenance, are input into an adaptive spatiotemporal feedback jump graph neural network model based on the wave equation constraint, and the output is a dynamic financial cost prediction result. High-dimensional non-Gaussian business and financial distribution data are input into a manifold-aligned deep flow model, mapped to a standard multivariate Gaussian distribution space through a reversible affine coupling transformation layer, and then an inverse transformation is performed to output disaster prevention budget preparation and capitalization budget recommendations.

2. The marine industry-finance integration decision-making method based on big data analysis according to claim 1, characterized in that, The aforementioned dimensionality reduction technique combining autoencoder and t-SNE involves the autoencoder encoding high-dimensional business feature vectors into a low-dimensional latent space, training the encoder using reconstruction error as the loss function to preserve the nonlinear topological structure, and then using the low-dimensional latent vector output by the encoder as input to execute the t-SNE algorithm to further optimize the low-dimensional distribution.

3. The marine industry-finance integration decision-making method based on big data analysis according to claim 2, characterized in that, The evolutionary dynamics graph clustering algorithm specifically abstracts geographically scattered buoy observation points, satellite observation points, and financial reimbursement nodes into a dynamic spatiotemporal weighted graph. It constructs a multi-scale graph Laplacian operator on the graph structure, solves for eigenvalues ​​and eigenvectors, projects the high-dimensional spatiotemporal financial relationship onto the low-dimensional spectral space, and performs adaptive evolutionary clustering based on energy aggregation degree to complete the spatiotemporal alignment and collaborative feature domain classification of heterogeneous data.

4. The marine industry-finance integration decision-making method based on big data analysis according to claim 3, characterized in that, The mathematical form of the pattern Laplace operator is: ,in Let the node degree matrix be... For the adjacency weight matrix, the multi-scale version uses... This can be achieved by applying polynomial filter banks of different orders.

5. The marine industry-finance integration decision-making method based on big data analysis according to claim 4, characterized in that, The deep flow model based on manifold alignment is composed of a series of reversible affine coupling transformation layers in deep cascade. Each reversible affine coupling transformation layer splits the input vector into two parts. One part is transformed by scaling and translation functions, while the other part remains unchanged. During the transformation process, the probability density is continuously deformed according to the determinant of the Jacobian matrix.

6. The marine industry-finance integration decision-making method based on big data analysis according to claim 5, characterized in that, The proposed denoising algorithm based on nonlocal mean field and fractional calculus for complex time-varying diffusion dynamics specifically treats the original noisy time series as a virtual two-dimensional complex electromagnetic fluid field, and introduces fractional differential operators in both the time and frequency domains to construct a non-traditional partial differential diffusion equation. The diffusion coefficient is dynamically determined by the similarity of the nonlocal mean structure.

7. The marine industry-finance integration decision-making method based on big data analysis according to claim 6, characterized in that, The adaptive spatiotemporal feedback jump graph neural network model based on wave equation constraints consists of a spatial graph convolution branch and a time-gated recurrent unit branch, which are coupled through an adaptive spatiotemporal feedback jump mechanism. When the physical input index deviates from the normal threshold, the conventional intermediate hidden layer is skipped, and the Laplace operator physical wave equation constraint term is directly injected into the final financial loss prediction layer.

8. The marine industry-finance integration decision-making method based on big data analysis according to claim 7, characterized in that, The spatial graph convolution branch uses four graph convolutional layers, with output feature dimensions of 64, 128, 128, and 256 respectively; the hidden state dimension of the temporal gated recurrent unit branch is 128, the time series window length ranges from 24 to 168 time steps, and the number of samples per batch ranges from 32 to 128.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the marine industry and finance integration decision-making method based on big data analysis as described in any one of claims 1-8.

10. A marine industry-finance integrated decision-making system based on big data analysis, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images