A marine data double-check grading method based on constraint attention

Abstract

The application discloses a marine data double-check grading method based on constraint attention, and relates to the technical fields of marine data management and deep learning. The method comprises the following steps: obtaining multi-dimensional attribute information of marine data to be evaluated and quantifying the multi-dimensional attribute information into unified index values; grouping according to preset attribute categories, respectively inputting an attention model with a first preset weight range constraint, learning group index dynamic weights and outputting weighted scores of each category; inputting the weighted scores into a deep neural network for nonlinear fusion, outputting a total influence degree value after a second preset weight range constraint; and based on the total influence degree value, adopting a double-check strategy combining a hard rule threshold mapping and a neural network soft classification to determine a final grade. The application realizes full-process automation of marine data grading, avoids artificial subjectivity and fixed weight limitations, and improves grading accuracy, compliance and interpretability.

Description

Technical Field

[0001] This invention relates to the fields of marine data management and deep learning technology, and more specifically to a dual-verification and grading method for marine data based on constrained attention. Background Technology

[0002] Ocean observation, monitoring, and surveys have accumulated massive amounts of multi-source ocean data. Different operational scenarios, such as marine defense, resource development, scientific research and teaching, and public services, have different value requirements and security requirements for data. Therefore, it is necessary to quantify the value and classify the levels of ocean data.

[0003] Currently, common methods for classifying marine data mainly include manual weighting, the Analytic Hierarchy Process (AHP), and fuzzy comprehensive evaluation. These methods rely on the experience of domain experts and determine the classification by setting fixed indicator weights and classification thresholds. However, with the increasing volume, diversification, and dynamic changes in application scenarios of marine data, the limitations of traditional classification methods are becoming increasingly apparent: first, manually set weights and thresholds lack flexibility and are difficult to adapt to the characteristics and dynamic changes of different types of marine data; second, they struggle to capture the nonlinear correlations and complex coupling relationships between multiple indicators, resulting in insufficient accuracy of the classification results.

[0004] In recent years, some scholars have attempted to use methods such as backpropagation neural networks to conduct research on marine data classification. However, most existing studies only target single types and single scenarios of marine data, and have not yet introduced attention mechanisms and business rule constraints, resulting in a lack of interpretability of the weights learned by the models, making it difficult to meet audit requirements.

[0005] Therefore, there is an urgent need for a method that can adaptively learn indicator weights, capture multi-dimensional data correlations, conform to business rule requirements, and achieve dynamic classification of marine data. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide a constrained attention-based dual-verification classification method for ocean data that overcomes or at least partially solves the above problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] S1. Obtain multi-dimensional attribute information of the marine data to be evaluated and transform it into a unified quantitative indicator value; S2. The quantitative indicator values ​​are grouped according to multiple preset attribute categories, and then input into the attention model with business rule constraints. The dynamic weights of each indicator in each group are learned, and a first preset weight range constraint is applied. The weighted score corresponding to each attribute category is output. S3. Input the weighted scores corresponding to each attribute category into the deep neural network, learn the high-order correlation between each attribute category through multi-layer nonlinear transformation, apply the second preset weight range constraint, and output the total influence value. S4. Based on the total impact value, a dual verification strategy combining hard rule threshold mapping and neural network soft classification is adopted to determine the final level of the marine data.

[0009] Preferably, the multi-dimensional attribute information includes data volume, spatiotemporal coverage, time precision, spatial precision, data quality and processing level, domain type, business stage, user level, and usage frequency.

[0010] Preferably, the calculation method for the quantitative indicator value includes: Data volume quantification value: Using the original data volume as the independent variable, logarithmic operations are used to map data volumes of different magnitudes into quantification scores; Spatiotemporal coverage quantification value: Using time span and spatial coverage area as independent variables, the spatiotemporal coverage range is mapped to a quantification score through a weighted summation operation; Time precision quantization value: With time resolution as the independent variable, the time resolution is mapped to a quantization score through reciprocal operation. The higher the resolution, the higher the quantization score. Spatial precision quantization value: Using spatial resolution as the independent variable, the spatial resolution is mapped to a quantization score through reciprocal operation. The higher the resolution, the higher the quantization score. Data quality quantification: Using efficiency and completeness as independent variables, data quality is mapped to a quantification score through a weighted summation operation; Processing level quantification: Using the processing level as the independent variable, the processing level is mapped to a quantification score through linear operations; Usage frequency quantization: Using the original usage frequency as the independent variable, the usage frequency is mapped to a quantization score through logarithmic operation; Domain type quantization value: Maps the category to which it belongs to the corresponding quantization value according to the preset enumeration mapping table; Business phase quantification value: Map the category to the corresponding quantification value according to the preset enumeration mapping table; User-level quantization value: The category to which the user belongs is mapped to the corresponding quantization value according to the preset enumeration mapping table.

[0011] Preferably, it also includes a feature standardization step, which scales each quantitative indicator value to a uniform numerical range.

[0012] Preferably, the preset multiple attribute categories include three main categories: inherent attributes, business attributes, and user attributes; The inherent attributes include six indicators: data volume, spatiotemporal coverage, temporal precision, spatial precision, data quality, and processing level. The business attributes include two indicators: domain type and business stage; The user attributes include two indicators: user level and usage frequency.

[0013] Preferably, the attention model is a scaled dot product attention model.

[0014] Preferably, the step of learning the dynamic weights of each indicator within each group and applying a first preset weight range constraint specifically includes: For each type of attribute's indicator features, first calculate the attention score for each indicator item. Then, the attention score is normalized to the original attention weight. Then Mapping to a preset weight range yields the final dynamic weights of each indicator. ; The formula for calculating the attention score is as follows:

[0015] in, Indicates the first Attention scores for each metric Indicates the first Query vector for each indicator Indicates the first The key vector of each index, Indicates the dimension of the key vector; The normalization formula for the original attention weights is:

[0016] in, Indicates the first The original attention weights of each metric, Indicates the first Attention scores for each metric To find the summation index, iterate through all indicators within the current attribute group; Indicates the number of metrics for the current attribute; The final dynamic weights of each indicator are obtained based on the preset weight range, including:

[0017] in, and The first The preset lower limit and preset upper limit of each indicator.

[0018] Preferably, the deep neural network is an embedded deep neural network, wherein: The weighted scores corresponding to each attribute category are concatenated as the initial input features. A non-linear transformation is then performed through multiple fully connected layers, and the global fusion weights for each attribute category are learned through the output layer. After normalization and constraint by a second preset weight range, the final total weight is obtained. The original influence value is calculated using the following formula:

[0019] in, The total number of attribute categories. For the first Weighted scores for each attribute category, For the first Global fusion weights for each attribute category; For the original influence value The total influence value is obtained by linear scaling.

[0020] Preferably, the second preset weight range constraint is: Inherent Attribute Global Fusion Weight The preset constraint range is 0.4 to 0.6; Global integration weight of business attributes The preset constraint range is 0.2 to 0.4; Global fusion weight of user attributes The preset constraint range is 0.1 to 0.3.

[0021] Preferably, step S4 specifically includes: The hard rule threshold mapping maps the total influence value to multiple preset levels according to a preset threshold; In the neural network soft classification, the total influence value and the weighted scores of each attribute category are concatenated into the classification head input vector. Through a fully connected layer and Softmax normalization, the probability distribution of each level is output, and the level with the highest probability is taken as the soft classification result. Overall rating determination: When the hard rule mapping rating result is consistent with the soft classification result, the rating is taken as the final rating; when the two are inconsistent, the soft classification result shall prevail and manual review shall be triggered.

[0022] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This invention achieves adaptive learning of indicator weights through a constrained attention model, replacing the traditional manually set fixed weights. This allows weight allocation to be dynamically adjusted based on the actual characteristics of each data set, avoiding the drawbacks of strong subjectivity and poor flexibility. By using a deep neural network to perform multi-layer nonlinear fusion of the scores of the three major attributes, it automatically captures the nonlinear correlations and higher-order coupling relationships between multiple indicators, overcoming the insufficient accuracy of traditional linear weighting methods. By introducing first and second preset weight range constraints into the attention model and deep neural network respectively, it ensures that the weights learned by the model always conform to marine data management specifications, solving the problems of poor interpretability and difficulty in meeting audit requirements of existing deep learning methods. Simultaneously, this invention employs a dual verification strategy combining hard rule threshold mapping and neural network soft classification for comprehensive grade determination. This ensures that the grading results conform to preset business threshold specifications, and the soft classification model adapts to changes in data distribution, improving the robustness of the grading results. This invention achieves full automation from indicator quantification to graded output without manual intervention, significantly improving the efficiency and accuracy of marine data grading and providing a reliable grading basis for efficient management, accurate sharing, secure control, marine scientific research, and strategic decision-making. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a flowchart of a dual-verification and grading method for ocean data based on constrained attention, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the embedded deep neural network RE-DNN model structure provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention discloses a dual-verification classification method for ocean data based on constrained attention, comprising: S1. Obtain multi-dimensional attribute information of the marine data to be evaluated and transform it into a unified quantitative indicator value; S2. Group the quantitative indicator values ​​according to multiple preset attribute categories, input them into the attention model with business rule constraints, learn the dynamic weights of each indicator in each group, apply the first preset weight range constraint, and output the weighted score corresponding to each attribute category. S3. Input the weighted scores corresponding to each attribute category into the deep neural network, learn the high-order correlation between each attribute category through multi-layer nonlinear transformation, apply the second preset weight range constraint, and output the total influence value. S4. Based on the total impact value, a dual verification strategy combining hard rule threshold mapping and neural network soft classification is adopted to determine the final grade of marine data.

[0027] The following is combined with Figure 1 Steps S1-S4 are explained in detail below: S1. Acquire and quantify multi-dimensional attribute information of ocean data. S1.1 Extract the original dataset to be evaluated from the marine database and organize and archive it according to a unified format.

[0028] In one embodiment, measured data from 10 buoy monitoring stations in a certain area are used as the research object. This buoy dataset covers continuous observation data from 2022 to 2024. The original dataset contains 10 indicators, namely: data volume, spatiotemporal coverage, temporal accuracy, spatial accuracy, data quality, processing level, usage frequency, domain type, business stage, and user level.

[0029] It should be noted that fields such as domain type, business stage, and user level in the original dataset are usually recorded in Chinese text format, such as maritime rights protection, disaster early warning, business pilot stage, and business routine operation stage. Due to differences in naming conventions among different data sources and business systems, the Chinese description of the same category may have multiple expressions. If we directly enumerate and map based on the original Chinese text, the mapping table will need to exhaustively enumerate a large number of non-standard expressions, which is not only difficult to maintain, but also prone to the problem of the same category being mapped to different quantification values, thus affecting the learning effect of subsequent attention models and deep neural networks and the consistency of the classification results.

[0030] Therefore, this step first standardizes the original Chinese text according to preset business rules, unifying text tags representing the same business meaning into a limited set of standard enumeration values. Specifically: text tags representing purposes such as safeguarding maritime rights and interests and disaster early warning are standardized as core domains; text tags representing purposes such as fisheries production, nearshore aquaculture, and resource development are standardized as important domains; text tags representing purposes such as scientific research, teaching, and public services are standardized as general domains; the pilot phase of business operations is unified as the pilot phase; and the normalized operation phase of business operations is unified as the normalized phase. The standardized enumeration values ​​are used as input for subsequent enumeration mapping tables, thereby ensuring the consistency, auditability, and automation of the quantification process.

[0031] In one embodiment, the raw data of buoy station #1 in a certain area is as follows: total data volume of 100TB; time span of 5 years; spatial coverage area of ​​500,000 square kilometers; time resolution of hourly; spatial resolution of 1 kilometer; data quality efficiency of 95% and integrity of 90%; processing level of 4; average number of visits per year (usage frequency) of 5,000; domain type standardized as "core domain"; business stage standardized as "normalization stage"; user level as provincial and ministerial-level scientific research institutions.

[0032] S1.2, Regularized Feature Quantization The purpose of this step is to convert all the raw data prepared in the previous step into a unified numerical score. Different quantification strategies are used for different indicator types: For continuous numerical indicators (the raw numerical fields corresponding to data volume, spatiotemporal coverage, time precision, spatial precision, data quality, processing level, and usage frequency), a preset deterministic calculation rule is used to map them into quantified scores; for enumerated indicators (domain type, business stage, user level), a preset enumeration mapping table is used to map their category to the corresponding quantified value. A specific example is shown below: (1) Quantification of continuous numerical indicators Data volume quantification: Using the original data volume as the independent variable, logarithmic operations are used to map data volumes of different magnitudes to quantification scores. The reason for using logarithmic operations is that ocean data volumes can span several orders of magnitude, from GB to PB. The logarithmic function can compress this exponential difference into a linear difference, making data of different magnitudes comparable in quantification scores. The total data volume is quantified as: log 10 (Total data volume) × 200; Substitute data for buoy #1: log 10 (100) × 20 = 2 × 20 = 40.

[0033] Spatiotemporal coverage quantification: Using time span and spatial coverage area as independent variables, a weighted summation operation is used to map the spatiotemporal coverage range into a quantified score. The time span is based on a 10-year unit, and the spatial coverage area on a 1 million square kilometer unit, each accounting for 50% of the weight. The longer the time span and the larger the spatial coverage area, the stronger the representativeness and applicability of the data. The spatiotemporal coverage quantization value is: Spatiotemporal coverage quantization value = (time span / base unit) × 5 + (spatial coverage area / base unit) × 5; Substituting the data of buoy #1: (5 / 10) × 5 + (50 / 100) × 5 = 5.0.

[0034] Time precision quantization: Using time resolution as the independent variable, the time resolution is mapped to a quantization score through reciprocal calculation. The higher the resolution (i.e., the smaller the numerical value), the higher the quantization score. High time resolution data can support more refined time series analysis and event tracking; The time precision quantization value is: (1 / time resolution)×10, where the time resolution is in hours; Substitute the data of buoy #1: 1 / 1×10=10.

[0035] Spatial precision quantization: Using spatial resolution as the independent variable, the spatial resolution is mapped to a quantization score through a reciprocal operation. The higher the resolution (i.e., the smaller the numerical value), the higher the quantization score. High spatial resolution data can support more refined region analysis and target recognition; The spatial accuracy quantization value is: Spatial accuracy quantization value = (1 / spatial resolution) × 20, where the spatial resolution is in kilometers; Substitute the data of buoy #1: 1 / 1 × 20 = 20.

[0036] Data quality quantification: Using efficiency and completeness as independent variables, data quality is mapped to a quantified score through a weighted summation operation. Efficiency has a weight of 0.7, and completeness has a weight of 0.3, reflecting that in marine data quality assessment, efficiency is generally more important than completeness. The data quality quantification value is: (efficiency × 0.7 + integrity × 0.3) × 100, where efficiency and integrity are expressed in decimal form; Substituting the data of buoy #1: efficiency = 95% = 0.95; integrity = 90% = 0.90; data quality quantification value = (0.95 × 0.7 + 0.90 × 0.3) × 100 = (0.665 + 0.27) × 100 = 93.5.

[0037] Processing level quantification: Using the processing level as the independent variable, a linear operation maps the processing level to a quantified score. The processing level is an integer from 1 to 4, with higher levels indicating more complete standardization, quality control, and correction processes, resulting in greater usability. The processing level quantification value is: processing level × 25; Substitute the data for buoy #1: 4 × 25 = 100.

[0038] Usage frequency quantization: Using the original usage frequency as the independent variable, the usage frequency is mapped to a quantized score through logarithmic operations. The reason for using logarithmic operations is similar to the data volume; the number of accesses may range from tens to tens of thousands, and the logarithmic function can compress the difference in magnitude.

[0039] The average annual number of visits is quantified as: = log 10 (Usage frequency) / 4; Substitute data for buoy #1: log 10 (5000) ≈0.92 (2) Quantification of enumeration-type indicators Domain type quantification value: The domain type is mapped to its corresponding quantification value according to a preset enumeration mapping table. In one embodiment, core domain = 0.9, important domain = 0.6, and ordinary domain = 0.3. These values ​​reflect the differences in importance of different domain types in marine data management. The No. 1 buoy site is a core domain, with a quantification value of 0.9.

[0040] Business Stage Quantification Value: The category is mapped to the corresponding quantification value according to a preset enumeration mapping table. In this embodiment, collection = 0.3, transmission = 0.5, application = 0.8, and archiving = 0.2. The above assignments reflect the differences in application activity and sharing needs of data at different lifecycle stages. This buoy station is in the "normalization stage" (corresponding to the "application" stage), with a quantification value of 0.8.

[0041] User-level quantification values: These values ​​are mapped to the user's category using a pre-defined enumeration mapping table. In one embodiment, national level = 0.9, provincial / ministerial level = 0.7, municipal level = 0.5, and enterprise / individual level = 0.3. These values ​​reflect the differences in decision-making weight and scope of influence among different user levels in the use of marine data. The No. 1 buoy site is a provincial / ministerial level research institution, and its quantification value is 0.7.

[0042] Taking the data from buoy #1 as an example, the quantitative results of each indicator are shown in the table below:

[0043] S1.3, Feature Standardization Because the quantized values ​​of each indicator differ significantly in scale and range (processing level is 100, domain type is only 0.9), directly inputting these quantized values ​​into the neural network would cause indicators with larger values ​​to dominate during model training, while the influence of indicators with smaller values ​​would be overwhelmed, preventing the model from correctly learning the true contribution of each indicator. Therefore, it is necessary to standardize the quantized values ​​of each indicator.

[0044] In one embodiment, the Min-Max standardization method is used to scale the quantified values ​​of each indicator to a uniform numerical range [0,1], making all indicators numerically comparable. The standardization formula is:

[0045] in, The raw value of a certain indicator item. and These are the minimum and maximum values ​​of the indicator in the current dataset, respectively. Essentially, this formula maps the minimum value to 0, the maximum value to 1, and the intermediate values ​​to a proportional distribution.

[0046] The minimum and maximum values ​​of each indicator in the current dataset (10 buoy stations) are shown in the table below:

[0047] Taking buoy #1 as an example, the standardized values ​​of each indicator are as follows: data volume 0.40, spatiotemporal coverage 0.05, temporal precision 0.10, spatial precision 0.20, data quality 0.94, processing level 1.00, domain type 1.00, business stage 1.00, user level 0.78, and usage frequency 0.38. All indicators are within the range of [0,1], which meets the format requirements of the neural network model for input data.

[0048] S2. Grouping by preset attribute categories and learning the weights within each group based on an attention mechanism. The purpose of this step is to divide the 10 standardized indicators into three groups according to their attribute categories: Inherent attribute group (6 indicators: data volume, spatiotemporal coverage, temporal precision, spatial precision, data quality, and processing level); Business attribute group (2 indicators: domain type, business stage); User attribute groups (two metrics: user level and usage frequency) are input into a constrained scaling dot product attention model. Within each group, the weights of each metric are automatically adjusted based on the characteristics of the data itself, and a first preset weight range constraint is applied to ensure that the adjusted weights conform to the preset business specifications.

[0049] In traditional grading methods, the weights of each indicator are pre-set and fixed by experts, such as 30% for data volume and 20% for accuracy, making it impossible to flexibly adjust according to the actual situation of each dataset. When faced with a dataset of extremely high accuracy, the "spatial accuracy" indicator should receive a higher weight, but traditional methods cannot achieve this adaptive adjustment.

[0050] This invention employs a constrained scaled dot product attention CSDA model to solve this problem.

[0051] S2.1 Feature Dimension Mapping For each group's indicator features, a linear layer maps each indicator to a query vector. and key vector :

[0052] in, The standardized indicator feature matrix, To perform a linear transformation of the query vector into a weight matrix, This is a linear transformation of the key vectors into a weight matrix. The purpose of this mapping is to generate two sets of vector representations for each indicator, facilitating subsequent calculations of the attention levels between indicators. and It learns automatically through model training, without the need for manual settings.

[0053] S2.2 Attention Score Calculation For each metric within each group, calculate its attention score:

[0054] in, For the first i Attention scores for each metric For the first i Query vector for each indicator For the first The key vector of each index, is the dimension of the key vector.

[0055] pass and The dot product operation measures the first... The degree of attention each indicator receives in the current dataset. A larger dot product indicates a stronger correlation between the indicator and other indicators, or a more prominent unique characteristic, and thus it should be assigned a higher weight. Divide by This is to scale the dot product result to prevent the gradient from vanishing in the subsequent Softmax function due to excessively large values.

[0056] In this embodiment, the inherent attribute group =6, Business attribute group and user attribute group =2.

[0057] S2.3 Attention Weight Normalization Convert the attention scores into raw attention weights ranging from 0 to 1, ensuring that the sum of the weights of all metrics under the same attribute is 1:

[0058] in, For the first The original attention weights of each metric, Indicates the first Attention scores for each metric To calculate the sum index, iterate through all metrics within the current attribute group, meaning the denominator is the exponential sum of the attention scores of all metrics within the same group; This indicates the number of indicators for the current attribute. By converting the attention score into a probability distribution, the sum of the weights of each indicator within the same attribute group is equal to 1, ensuring comparability.

[0059] It should be noted that what is obtained at this time These are weights that the model learns freely based on data characteristics, without any business constraints being imposed.

[0060] S2.4, First Preset Weight Range Constraint The above data-driven raw attention weights Mapped to a preset weight range:

[0061] in, For the first The final dynamic weight of each indicator and The first The preset lower and upper limits for the weights of each indicator. These preset lower and upper limits are pre-set by business experts according to marine data management specifications and stored in the system, remaining fixed when processing each data point.

[0062] This step is explained as follows: Although the model can freely learn the weights of each indicator based on data characteristics, the final output weights must fall within a reasonable range defined by business requirements. Taking data quality indicators as an example, the lower limit of their weight is set at 20%, meaning that even if the data itself is of poor quality, it should still occupy a fundamental position in the classification process. The upper limit is 50%, meaning that even if the data is of extremely high quality, it should not be the sole determining factor. This constraint ensures that the classification results are both data-adaptive and comply with industry standards for marine data management.

[0063] S2.5 Weighted Score Calculation Calculate the weighted score for each attribute category using the constrained dynamic weights: (Weighted score based on inherent attributes); (Weighted score based on business attributes); (Weighted score based on user attributes); in, The score of the dataset in terms of its inherent attribute dimensions; The weights are the dynamic weights of the six inherent attributes after constraints. It is the standardized feature value of the six inherent attributes (data volume, spatiotemporal coverage, temporal precision, spatial precision, data quality, and processing level); The dataset is scored in terms of business attribute dimensions; The dynamic weights are constrained by two indicators of business attributes. The standardized feature values ​​are the two indicators of business attributes (domain type and business stage). The dataset scores in terms of user attributes. Dynamic weights for two user attribute metrics; The standardized feature values ​​of two user attributes (user level and usage frequency).

[0064] In one embodiment, the weighted scores of the three attributes of buoy site #1, calculated using the CSDA model, are as follows: =0.5087, =1.00, =0.62.

[0065] S3. Obtain the total influence value by fusing the scores of each attribute based on a deep neural network. Scoring the three independent attributes output by the CSDA model ( , , Deep integration is required. Complex nonlinear relationships may exist between the three attributes: even if a data point has a low score in its inherent attributes (small data volume, average accuracy), its overall data value may still be extremely high if its business attribute score is extremely high (related to core defense). This effect cannot be accurately expressed by simple linear weighted summation and must be automatically learned through the nonlinear transformation of deep neural networks.

[0066] Will , , Concatenate as initial input features , Input embedded deep neural network.

[0067] like Figure 2 As shown, RE-DNN first inputs 3D features The first layer of embedded features is obtained by mapping the features to 32-dimensional intermediate features through a fully connected layer and then performing ReLU activation. :

[0068] in, The weight matrix is ​​3×32. It is a 1×32 bias vector.

[0069] Then The feature interaction information is further extracted through the feature embedding layer. While maintaining the 32-dimensionality, the second layer of embedded features is obtained after ReLU activation. :

[0070] in, The weight matrix is ​​32×32. It is a 1×32 bias vector.

[0071] Then The feature is compressed into a 16-dimensional deep fusion feature through a recursive embedding layer and then activated by ReLU. :

[0072] in, The weight matrix is ​​32×16. This is a 1×16 bias vector. ReLU is a modified linear activation function used to introduce non-linear fitting capabilities, enabling the network to learn the complex relationships between the three main attributes. For deep fusion features (1×16 dimensions).

[0073] Features from deep fusion are obtained through the output layer. Learn the global fusion weights of the three main attributes:

[0074] in, The original weight vector (1×3 dimensions) is unnormalized. The weight output layer matrix is ​​16×3. It is a 1×3 bias vector.

[0075] Perform Softmax normalization on the original weights so that the sum of the three weights is 1:

[0076] At this time, it was obtained W The weights are entirely data-driven and have not yet been subject to any global business constraints.

[0077] To prevent the model from exhibiting extreme weight distributions when learning global weights, such as inherent attributes accounting for 95% and business attributes accounting for only 2%, a second preset weight range constraint needs to be applied to limit the global fusion weights of each attribute within a preset reasonable range:

[0078] in, The function will The value is limited to [ Within the interval, less than Time to take greater than Time to take .

[0079] The above constraint ranges are set based on the following: Inherent attributes, as the physical foundation of the data (data volume, accuracy, quality, etc.), should always occupy a dominant position; therefore, the lower limit of their weight is set at 0.4 (not less than 40%). Business attributes reflect the sensitivity of the domain to which the data is involved and have high business volatility; therefore, the weight range is set at 0.2 to 0.4 (20% to 40%). User attributes reflect the actual popularity of the data and serve as a reference auxiliary indicator; their weight is relatively low, so the upper limit is set at 0.3 (not exceeding 30%). These preset constraint ranges are also pre-set by business experts according to marine data management specifications and stored in the system.

[0080] The constrained weights are normalized again to ensure that the sum of the three global weights is 1:

[0081] And it satisfies:

[0082] In this embodiment, the global fusion weights of the three attributes of buoy site #1, calculated by the RE-DNN model, are as follows: =0.52, =0.28, =0.20, all of which fall within their respective preset intervals and sum to 1.

[0083] Calculate the original total impact value using the following formula. :

[0084] Substitute the data:

[0085] In one embodiment, the original total influence is linearly scaled and mapped to the range of 0 to 10 to obtain the total influence value. : .

[0086] S4. Perform dual-verification grading based on the total impact value. Total influence value This is then translated into specific classification results. The classification employs a dual-verification strategy: first, hard rule mapping is performed based on preset thresholds, and then soft classification is performed by the neural network classification head. Only when the results of both methods are consistent is the final classification determined. Specifically: S4.1, Hard Rule Mapping Hierarchy Hard rule mapping is the most direct grading method: it presets a fixed threshold based on business specifications and directly assigns the total impact value. Mapped to the corresponding level.

[0087] In one embodiment, the following preset threshold is used to... The mapping is assigned to the corresponding level, and this is denoted as the hard rule mapping level result. :

[0088] The above thresholds are set based on the operational specifications for marine data classification management. Core data ( This involves significant rights and interests, requiring the highest level of protection; important data (7≤ <9) Involving important maritime rights and key operations, requiring strict management; data applications (5≤) are advisable. <7) Data of high value but not core, which can be shared through an application and approval process; publicly available data (3≤ <5) Data that has some value but poses no security risk can be proactively disclosed; publicly disclosed data ( (Low value or no risk at all, can be fully open.)

[0089] In this embodiment, buoy station #1 = , falling within 5≤ For intervals <7, the hard rule classification result is suitable for application.

[0090] S4.2, Soft classification model grading Soft classification models, serving as auxiliary validations for hard rule mappings, output probability distributions for each level through a neural network classification head. This allows them to adapt to changes in data distribution and avoid limitations imposed by hard rules at boundary values ​​(such as...). =9.01 or Possible misjudgments around (=8.99).

[0091] Construct a classification head input feature and fuse the total influence score with the scores of each attribute:

[0092] This input feature includes both the comprehensive grading signal reflected by the total score and the original distribution information of the scores of each attribute, enabling the classification head to make a more comprehensive grading judgment.

[0093] By using a fully connected layer and Softmax normalization, we obtain five levels of probability distributions:

[0094]

[0095] in, This is the unnormalized original score vector (1×5 dimensions) output by the fully connected layer of the classification head. The classification layer weight matrix is ​​4×5. The classification layer bias vector is 1×5. for through The probability distribution vectors of each level obtained after function normalization satisfy the condition that the sum of the probabilities of each level is 1.

[0096] In this embodiment, input After setting the value to [6.685, 0.5087, 1.0, 0.62], the probability distribution of the classification header output is P = [0.02, 0.08, 0.85, 0.04, 0.01] (corresponding to core, important, suitable for application, publicly available, and publicly available, respectively). The probability of suitable for application is 0.85, which is the highest among all levels. Therefore, the soft classification result is suitable for application.

[0097] S4.3 Comprehensive Grade Determination When the hard rule mapping result is consistent with the soft classification result, that level is taken as the final level; when the two are inconsistent, the soft classification result shall prevail, and manual review shall be triggered.

[0098] The formula for determining the overall grade is:

[0099] in, The overall rating result, Take the level index with the highest probability and map it as core / important / applicable / public / public. In this embodiment, both are considered "suitable for application," therefore no manual review is required, and the final classification is "suitable for application." This classification result means that the data from Buoy Station #1 can be used in accordance with regulations through the relevant data application and approval process.

[0100] If other buoy sites are processed using the above procedure, corresponding classification results can be obtained based on the differences in indicators of each site, thus achieving dynamic and accurate classification of marine data.

[0101] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0102] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dual-verification and grading method for ocean data based on constrained attention, characterized in that, Includes the following steps: S1. Obtain multi-dimensional attribute information of the marine data to be evaluated and transform it into a unified quantitative indicator value; S2. The quantitative indicator values ​​are grouped according to multiple preset attribute categories, and then input into the attention model with business rule constraints. The dynamic weights of each indicator in each group are learned, and a first preset weight range constraint is applied. The weighted score corresponding to each attribute category is output. S3. Input the weighted scores corresponding to each attribute category into the deep neural network, learn the high-order correlation between each attribute category through multi-layer nonlinear transformation, apply the second preset weight range constraint, and output the total influence value. S4. Based on the total impact value, a dual verification strategy combining hard rule threshold mapping and neural network soft classification is adopted to determine the final level of the marine data.

2. The marine data dual verification and classification method according to claim 1, characterized in that, The multi-dimensional attribute information includes data volume, spatiotemporal coverage, temporal precision, spatial precision, data quality, processing level, domain type, business stage, user level, and usage frequency.

3. The marine data dual verification and classification method according to claim 2, characterized in that, The methods for calculating quantitative indicator values ​​include: Data volume quantification value: Using the original data volume as the independent variable, logarithmic operations are used to map data volumes of different magnitudes into quantification scores; Spatiotemporal coverage quantification value: Using time span and spatial coverage area as independent variables, the spatiotemporal coverage range is mapped to a quantification score through a weighted summation operation; Time precision quantization value: With time resolution as the independent variable, the time resolution is mapped to a quantization score through reciprocal operation. The higher the resolution, the higher the quantization score. Spatial precision quantization value: Using spatial resolution as the independent variable, the spatial resolution is mapped to a quantization score through reciprocal operation. The higher the resolution, the higher the quantization score. Data quality quantification: Using efficiency and completeness as independent variables, data quality is mapped to a quantification score through a weighted summation operation; Processing level quantification: Using the processing level as the independent variable, the processing level is mapped to a quantification score through linear operations; Usage frequency quantization: Using the original usage frequency as the independent variable, the usage frequency is mapped to a quantization score through logarithmic operation; Domain type quantization value: Maps the category to which it belongs to the corresponding quantization value according to the preset enumeration mapping table; Business phase quantification value: Map the category to the corresponding quantification value according to the preset enumeration mapping table; User-level quantization value: The category to which the user belongs is mapped to the corresponding quantization value according to the preset enumeration mapping table.

4. The marine data dual verification and classification method according to claim 1, characterized in that, It also includes a feature standardization step, which scales the values ​​of each quantitative indicator to a uniform numerical range.

5. The marine data dual verification and classification method according to claim 1, characterized in that, The preset multiple attribute categories include three main categories: inherent attributes, business attributes, and user attributes; The inherent attributes include six indicators: data volume, spatiotemporal coverage, temporal precision, spatial precision, data quality, and processing level. The business attributes include two indicators: domain type and business stage; The user attributes include two indicators: user level and usage frequency.

6. The marine data dual verification and classification method according to claim 1, characterized in that, The attention model is a scaled dot product attention model.

7. The marine data dual verification and classification method according to claim 6, characterized in that, The process of learning the dynamic weights of each indicator within each group and applying a first preset weight range constraint specifically includes: For each type of attribute's indicator features, first calculate the attention score for each indicator item. Then, the attention score is normalized to the original attention weight. Then Mapping to a preset weight range yields the final dynamic weights of each indicator. ; The formula for calculating the attention score is as follows: in, Indicates the first Attention scores for each metric Indicates the first Query vector for each indicator Indicates the first The key vector of each index, Indicates the dimension of the key vector; The normalization formula for the original attention weights is: in, Indicates the first The original attention weights of each metric, Indicates the first Attention scores for each metric To find the summation index, iterate through all indicators within the current attribute group; Indicates the number of metrics for the current attribute; The final dynamic weights of each indicator are obtained based on the preset weight range, including: in, and The first The preset lower limit and preset upper limit of each indicator.

8. The marine data dual verification and classification method according to claim 1, characterized in that, The deep neural network is an embedded deep neural network, wherein: The weighted scores corresponding to each attribute category are concatenated as the initial input features. A non-linear transformation is then performed through multiple fully connected layers, and the global fusion weights for each attribute category are learned through the output layer. After normalization and constraint by a second preset weight range, the final total weight is obtained. The original influence value is calculated using the following formula: in, The total number of attribute categories. For the first Weighted scores for each attribute category, For the first Global fusion weights for each attribute category; For the original influence value The total influence value is obtained by linear scaling.

9. The marine data dual verification and classification method according to claim 8, characterized in that, The second preset weight range constraint is: Inherent Attribute Global Fusion Weight The preset constraint range is 0.4 to 0.6; Global integration weight of business attributes The preset constraint range is 0.2 to 0.4; Global fusion weight of user attributes The preset constraint range is 0.1 to 0.

3.

10. The marine data dual verification and classification method according to claim 1, characterized in that, Step S4 specifically includes: The hard rule threshold mapping maps the total influence value to multiple preset levels according to a preset threshold; In the neural network soft classification, the total influence value and the weighted scores of each attribute category are concatenated into the classification head input vector. Through a fully connected layer and Softmax normalization, the probability distribution of each level is output, and the level with the highest probability is taken as the soft classification result. Overall rating determination: When the hard rule mapping rating result is consistent with the soft classification result, the rating is taken as the final rating; when the two are inconsistent, the soft classification result shall prevail and manual review shall be triggered.

Images