Marine ranch multi-level transaction system and method based on multi-level market architecture

Abstract

The invention belongs to the technical field of computers, and particularly relates to a marine ranch multi-level transaction system and method based on a multi-level market architecture, and the method comprises the steps: obtaining multi-source heterogeneous monitoring data, and carrying out the time-space alignment; constructing a dynamic space-time correlation graph; evaluating data quality by using a multi-modal diagram attention network, and distinguishing systematic anomalies and random faults; and a three-level market pricing mechanism is constructed based on the quality score, and transaction transparency and traceability are realized through an intelligent contract and a block chain. The quantized data quality score is used as a core pricing factor, and a three-level market architecture containing original data, verification information and decision analysis products is designed, so that the data value is directly linked with the credibility, the problem of pricing misaccuracy caused by non-transparent data quality in the prior art is solved, and the pricing accuracy is improved. And a fair and efficient data value discovery mechanism is formed.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically a multi-level trading system and method for marine ranching based on a multi-level market architecture. Background Technology

[0002] As the marine economy accelerates its transformation towards digitalization and intelligentization, marine ranches, as a new type of blue economy carrier integrating ecological aquaculture, resource conservation, and industrial integration, face an increasingly urgent need for data assetization and market-based trading. Against this backdrop, constructing a multi-tiered trading system covering the entire process of data collection, ownership confirmation, evaluation, pricing, and trading has become a key infrastructure for promoting the efficient circulation and value release of marine data elements. Such systems typically rely on the ability to integrate multi-source heterogeneous data, combined with blockchain, smart contracts, and big data analytics technologies, aiming to transform marine ranch data from raw information into tradable assets and support collaborative operations across different market levels.

[0003] The marine ranching transaction system, based on a multi-tiered market architecture, focuses on incorporating multi-dimensional data on marine environment, biological growth, supply chain, and transaction behavior into a unified governance framework. Through tiered rights confirmation and graded pricing mechanisms, it meets the data usage needs of diverse stakeholders, including government regulators, corporate operators, and research institutions. The core of this system lies in establishing a reliable, dynamic, and traceable data value assessment chain, enabling data quality, scarcity, and timeliness to be accurately mapped to transaction prices, thereby ensuring market fairness and efficient resource allocation.

[0004] While existing technologies have attempted to incorporate blockchain for data ownership verification, employ multiple pricing models to support transaction flexibility, and deploy security and regulatory modules to mitigate risks, significant shortcomings remain in key aspects of the data processing module—particularly the real-time anomaly detection mechanism of the data quality assessment unit. Specifically, current automated detection systems generally employ static thresholds or fixed rule models, making it difficult to dynamically adapt to the frequent sudden noise interference in marine ranch sensor data, such as instantaneous data distortion caused by tidal changes, equipment drift, or sudden increases in salinity. Such interference has a very high probability of being misjudged as genuine anomalies, resulting in a false alarm rate consistently exceeding 35%, severely distorting data availability assessment results. Furthermore, the pricing unit heavily relies on the quality assessment output; if the input is distorted, it will directly cause the reference price to deviate from its actual value, further amplifying pricing discrepancies in multi-level market scenarios and weakening the credibility and operational stability of the trading system.

[0005] Therefore, this invention provides a multi-level trading system and method for marine ranching based on a multi-level market architecture. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: a multi-level trading system and method for marine ranching based on a multi-level market architecture, the method comprising the following steps: Acquire multi-source heterogeneous monitoring data from both inside and outside the marine ranch. The multi-source heterogeneous monitoring data includes internal environmental data collected by a sensor network deployed in a preset monitoring area, external environmental data obtained through remote sensing satellites or meteorological service interfaces, and manually entered aquaculture operation data. The multi-source heterogeneous monitoring data is subjected to data preprocessing and spatiotemporal alignment operations. The operations include data cleaning, normalization, and unifying data from different sources and sampling frequencies onto a preset spatiotemporal grid reference through Kriging spatial interpolation and cubic spline temporal interpolation algorithms to form a standardized spatiotemporal data cube. Based on the standardized spatiotemporal data cube, a dynamic spatiotemporal correlation map characterizing the intrinsic correlation of the marine ranch environmental elements is constructed. In this map, the nodes represent independent sensors or divided geographic grid units, and the edges represent the spatial proximity relationship, hydrodynamic correlation, or historical data statistical correlation between nodes. The weights of the edges are dynamically updated based on the real-time water flow model and the historical covariance matrix. The dynamic spatiotemporal correlation map and real-time collected monitoring data are input into a data quality assessment engine driven by a pre-trained multimodal graph attention network. The engine scores each data point and identifies the anomaly type, outputting a data quality assessment report containing data quality scores, confidence levels, and anomaly labels. The anomaly labels are distinguished into systemic environmental anomalies and random sensor anomalies. Based on the data quality assessment report, the multi-level pricing and value discovery module is activated to generate differentiated basic reference prices for data assets of different quality levels. Combined with market supply and demand, data timeliness and information entropy, the listed transaction prices of the first-level raw data market, the second-level verification information market and the third-level decision analysis product market are calculated. The listed transaction price and corresponding asset information are published to the third-level market transaction execution engine. The engine matches transactions through a continuous two-way auction mechanism based on the entrustment instructions of both the buyer and seller, and records the transaction details. Each transaction details, including the identities of both parties, the transaction price, the hash digest of the data asset, and the corresponding data quality assessment report, are encapsulated into a transaction record and written into a distributed ledger through smart contracts and a blockchain notarization module to ensure the immutability and full traceability of the transaction.

[0008] Preferably, the acquisition of internal environmental data specifically includes: acquiring water temperature, salinity, and depth data using temperature, salinity, and depth meters deployed at different water depths and geographical locations; acquiring three-dimensional flow velocity and direction data of the water using an underwater acoustic Doppler current profiler; acquiring chlorophyll concentration, dissolved oxygen content, and turbidity data of the water using an optical sensor array; and acquiring the pH value of the water using a pH sensor. The acquisition of external environmental data specifically includes: acquiring wind speed, wind direction, air pressure, rainfall, and solar radiation intensity data over the monitoring area using a meteorological data application programming interface; and acquiring sea surface temperature, sea surface height, and sea color remote sensing image data using a satellite remote sensing data interface. The aquaculture operation data includes feeding time, feeding amount, cage cleaning records, and disease prevention and control records.

[0009] Preferably, the construction of the dynamic spatiotemporal correlation map specifically includes: dividing the entire three-dimensional space of the marine ranch into multiple grid cells of equal volume, with each sensor or grid cell serving as a graph node; for any two nodes, the initial weight of their edges is determined by the weighted sum of the reciprocal of their spatial Euclidean distance and the Pearson correlation coefficient of historical data, with the weight coefficient set based on prior knowledge; during system operation, a four-dimensional variational assimilation system based on a mesoscale ocean model is introduced to update the hydrodynamic field in real time, and dynamically adjust the weight of the edges between nodes according to the water flow transmission path and velocity to reflect the actual impact of material and energy transport.

[0010] Preferably, the internal processing flow of the data quality assessment engine driven by the multimodal graph attention network specifically includes: First, the input dynamic spatiotemporal correlation graph and the temporal feature data of each node are aggregated through a graph convolutional network layer to generate a preliminary feature embedding vector that integrates spatial context information for each node. Secondly, the feature embedding vector sequence is input into a multi-head self-attention mechanism transformer encoding layer, which captures the temporal evolution pattern and long-range dependencies of each node and outputs a time-aware node state representation. Furthermore, the time-aware node state representation is input into a graph attention network layer. This network layer dynamically assigns different information weights to the neighboring nodes of the central node by calculating the attention coefficients between nodes, so that the model can focus on the associated nodes with the most diagnostic value in the current state, thereby enhancing the ability to identify weak abnormal signals. Finally, the final node representations aggregated by the graph attention network layer are fed into two parallel fully connected output layers. One output layer outputs a continuous value between zero and one as a data quality score through a sigmoid activation function, and the other output layer outputs a classification probability distribution through a flexible maximum activation function, corresponding to a preset anomaly category, including normal, sensor drift, biological attachment interference, signal line failure, red tide outbreak precursor, and cold wave intrusion.

[0011] Preferably, the specific pricing logic of the multi-tiered pricing and value discovery module is as follows: In the first-level raw data market, the asset price formula is: price equals the base value multiplied by the square of the data quality score, and then multiplied by a time-sensitivity decay factor. The base value is determined by the data type and scarcity, and the time-sensitivity decay factor is an exponential decay function with respect to the time difference of data generation. In the second-level verification information market, the asset is a high-quality subset of data or aggregated statistics confirmed by the evaluation engine. Its price is based on the first-level market price, with an additional authentication premium calculated based on confidence level and information gain.

[0012] In the third-level decision analysis product market, assets are derivative products generated based on high-quality data, such as suggestions for optimizing aquaculture density, disease early warning reports, or water quality change prediction models for the next week. Their prices are determined using a cost-plus pricing method, where costs include the cost of purchasing the data used, the resource consumption of the model calculations, and the preset profit margin.

[0013] Preferably, the smart contract and blockchain notarization module adopts a consortium blockchain architecture. The transaction execution logic built into the smart contract specifically stipulates that: when the buyer initiates a purchase request and pays an agreed amount of digital tokens to the contract address, the contract automatically locks the funds; the seller uploads the hash digest of the data asset to the contract; after an independent oracle service verifies the consistency between the hash digest of the data asset and the data quality assessment report, it triggers the contract to transfer the funds to the seller's address, grants access to the data asset to the buyer's address, and permanently records the complete transaction record on the blockchain as an event log.

[0014] The present invention also provides a multi-level trading system for marine ranching based on a multi-level market architecture, the system comprising: The multi-source heterogeneous data acquisition module is equipped with multiple physical sensor interfaces and network data interfaces, which are used to acquire the multi-source heterogeneous monitoring data from the internal sensor network of the marine ranch, external meteorological and remote sensing data service providers, and manual data entry terminals, respectively. The data preprocessing and spatiotemporal alignment module integrates a data normalization algorithm library, a kriging spatial interpolation engine, and a cubic spline temporal interpolation engine. It is used to perform standardized spatiotemporal alignment operations on the multi-source heterogeneous monitoring data to generate standardized spatiotemporal data cubes. The dynamic spatiotemporal correlation map construction module includes a spatial gridded unit and a hydrodynamic model interface, which is used to construct and dynamically update a dynamic spatiotemporal correlation map that characterizes the intrinsic correlation of marine ranch environmental elements according to preset rules and real-time hydrodynamic data. The multimodal graph attention network-driven data quality assessment engine has a deep neural network model at its core. The model is trained offline and can receive the dynamic spatiotemporal correlation graph and real-time monitoring data, and output a data quality assessment report containing data quality scores, confidence levels and anomaly labels. The multi-tiered pricing and value discovery module has an embedded set of tiered pricing algorithms, which are used to calculate and generate listing prices for data assets at different market levels based on the data quality assessment report. The third-level market trading execution engine includes an order book management unit and a matching unit, which are used to receive and process trading orders from market participants, execute continuous two-way auctions, and generate transaction records. The smart contract and blockchain evidence storage module is deployed on a distributed ledger network and contains predefined smart contract code to automatically execute transaction settlement and asset delivery, and to store all transaction records on the blockchain for evidence storage.

[0015] Preferably, the multi-source heterogeneous data acquisition module includes: an IoT gateway that receives data from underwater sensor nodes via a long-range wireless communication protocol; an application programming interface (API) poller that periodically calls interfaces of third-party meteorological and oceanographic satellite data services to obtain external data; and a database interface for reading a relational database storing historical aquaculture operation records. All acquired data is appended with timestamps accurate to milliseconds and GPS coordinates.

[0016] Preferably, the hardware implementation of the multimodal graph attention network-driven data quality assessment engine is based on a dedicated computing server equipped with a tensor processing unit, and its software implementation is based on a deep learning framework. The pre-trained model weight file is stored in non-volatile memory and loaded into memory during system runtime to support real-time data evaluation inference.

[0017] Preferably, the third-level market trading execution engine provides a trading front-end interface based on the web socket protocol, allowing authorized users to view real-time market data, submit limit or market orders, and query historical transaction records and account asset status through a graphical interface.

[0018] The beneficial effects of this invention are as follows: By constructing a dynamic spatiotemporal correlation map and using a multimodal graph attention network for data quality assessment, this invention can deeply explore the physical and statistical correlations between multidimensional environmental data of marine ranches, elevating the analysis of isolated data points to a systematic state evolution analysis. Thus, when faced with sudden sensor noise and drastic fluctuations in the real environment, it can accurately distinguish between random equipment failures and systematic environmental events, significantly reducing the false positive rate of data anomaly detection from over 35% to below 5%, greatly improving the accuracy and reliability of data quality assessment. By using the quantified data quality score as the core pricing factor and designing a three-tier market structure that includes raw data, verification information, and decision analysis products, the value of data is directly linked to its credibility. This solves the problem of inaccurate pricing caused by the lack of transparency in data quality in existing technologies, and forms a fair and efficient data value discovery mechanism. By introducing smart contracts and blockchain notarization technology, this invention places the entire data transaction process, from quality assessment to price discovery and final settlement, under a transparent, tamper-proof, and automatically executed framework. This completely eliminates information asymmetry and credit risk in the transaction process, protects the legitimate rights and interests of all market participants, and promotes the safe and efficient circulation of marine ranch data assets. By establishing a multi-tiered trading system, not only has the original sensor data been revitalized, but also validated high-value information products and decision support services have been generated. This has provided a wealth of data assets that meet the specific needs of different stakeholders, such as marine ranch managers, research institutions, and financial and insurance institutions, opening up entirely new business models and value growth points. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram of the structure of a multi-level trading system for marine ranches based on a multi-level market architecture proposed in this invention; Figure 2 This is a flowchart illustrating a multi-level trading method for marine ranches based on a multi-level market architecture, as described in this invention. Figure 3 This is a logical flowchart of the multi-source heterogeneous data acquisition and spatiotemporal alignment processing in this invention; Figure 4 This is a logical flowchart of the dynamic spatiotemporal correlation graph construction and weight update mechanism in this invention; Figure 5 This is a diagram illustrating the hierarchical pricing logic framework of the multi-level pricing and value discovery module in this invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] like Figure 2 As shown, this invention provides a multi-level trading method for marine ranching based on a multi-level market architecture. It aims to solve the technical problem in existing technologies where data quality assessment units cannot dynamically adapt to sudden noise in marine ranching sensor data, leading to a high misjudgment rate in anomaly detection and consequently affecting pricing accuracy. This method constructs a dynamic spatiotemporal correlation graph and introduces a data quality assessment engine driven by a multimodal graph attention network to achieve accurate identification and characterization of data anomalies, distinguishing between systemic environmental events and random equipment failures. The quantified data quality score is embedded as a core pricing factor in a three-tier market trading architecture. Finally, smart contracts and blockchain technology ensure the transparency and traceability of the transaction process.

[0023] The method includes the following steps: S1, acquire multi-source heterogeneous monitoring data inside and outside the marine ranch; S2 performs data preprocessing and spatiotemporal alignment operations on multi-source heterogeneous monitoring data to form a standardized spatiotemporal data cube; S3, based on standardized spatiotemporal data cubes, constructs a dynamic spatiotemporal correlation map representing the intrinsic correlation of environmental elements in marine ranches; S4 inputs the dynamic spatiotemporal correlation map and real-time collected monitoring data into a pre-trained multimodal graph attention network-driven data quality assessment engine, and outputs a data quality assessment report containing data quality scores, confidence levels and anomaly labels; S5, based on the data quality assessment report, activates the multi-level pricing and value discovery module to generate differentiated basic reference prices for data assets of different quality levels, and calculates and forms the listing and trading prices of the first-level raw data market, the second-level verification information market and the third-level decision analysis product market; S6 publishes the listed transaction price and corresponding asset information to the secondary market transaction execution engine. Based on the orders from both buyers and sellers, it matches transactions through a continuous two-way auction mechanism and records transaction details. S7 encapsulates the details of each completed transaction into a transaction record and writes it into the distributed ledger through smart contracts and a blockchain notarization module.

[0024] In step S1, multi-source heterogeneous monitoring data from both inside and outside the marine ranch are acquired. This multi-source heterogeneous monitoring data includes three types of sources: the first type is internal environmental data collected by a sensor network deployed within a pre-defined monitoring area; the second type is external environmental data acquired through remote sensing satellites or meteorological service interfaces; and the third type is manually entered aquaculture operation data. As mentioned earlier, the acquisition of internal environmental data specifically includes: obtaining water temperature, salinity, and depth data through temperature, salinity, and depth gauges deployed at different water depths and geographical locations; obtaining three-dimensional flow velocity and direction data of the water through an underwater acoustic Doppler current profiler; obtaining chlorophyll concentration, dissolved oxygen content, and turbidity data of the water through an optical sensor array; and obtaining the pH value of the water through a pH sensor. All internal sensors transmit data to the IoT gateway via a long-range wireless communication protocol, with the data packets embedded with timestamps accurate to milliseconds and GPS coordinates. As mentioned earlier, the acquisition of external environmental data specifically includes: periodically calling third-party meteorological services through meteorological data application programming interfaces to obtain data on wind speed, wind direction, air pressure, rainfall, and solar radiation intensity over the monitoring area; and acquiring remote sensing image data of sea surface temperature, sea surface height, and sea color through satellite remote sensing data interfaces. Aquaculture operation data is entered by aquaculture farm managers through dedicated terminals, including feeding time, feeding amount, cage cleaning records, and disease prevention and control records. Each record is associated with a unique operator identification and operation timestamp. All three types of data are immediately placed in a buffer queue after collection, awaiting further processing.

[0025] In step S2, data preprocessing and spatiotemporal alignment operations are performed on the multi-source heterogeneous monitoring data, such as... Figure 3 As shown. This operation first performs data cleaning, removing values ​​that are clearly outside the physically reasonable range, such as salinity below zero or above 45 parts per thousand, or temperature below zero degrees Celsius or above 50 degrees Celsius. Second, it performs normalization, mapping data of different dimensions to the zero-to-one interval. The normalization formula uses the minimum-maximum scaling method, meaning that for any variable x, its normalized value is: ; in: and These are the global minimum and maximum values ​​of the variable in the historical database, respectively. Then, spatiotemporal alignment is performed to unify data from different sources and sampling frequencies onto a preset spatiotemporal grid reference. Spatial alignment uses the Kriging spatial interpolation algorithm, taking known sensor locations and their measurements as input, to estimate environmental parameter values ​​on regular grid nodes within the three-dimensional space of the entire marine ranch; temporal alignment uses a cubic spline temporal interpolation algorithm to resample the time series of all variables onto a unified five-second interval time axis. Understandably, after the above processing, all data is organized into a four-dimensional tensor structure, with dimensions of longitude, latitude, depth, and time, forming a standardized spatiotemporal data cube. This cube is stored in a high-performance time-series database, supporting fast slice queries by arbitrary spatiotemporal windows.

[0026] In step S3, based on the standardized spatiotemporal data cube, such as Figure 4 As shown, a dynamic spatiotemporal correlation map is constructed. The nodes of this map represent independent sensors or divided geographic grid units. Specifically, the entire three-dimensional space of the marine ranch is divided into a cubic grid with sides of ten meters, and the center point of each grid serves as a graph node. Simultaneously, the locations of all physical sensors are retained as additional nodes. If a sensor is located within a grid, a one-to-one mapping is established between that sensor node and the corresponding grid node. Edges in the graph represent spatial proximity, hydrodynamic correlations, or historical data statistical correlations between nodes. Initial edge weights are determined. Determined by the following formula: ; in Let be the Euclidean distance between nodes i and j. The Pearson correlation coefficient between the two for similar environmental parameters (such as temperature) over the past thirty days is given by α and β, which are preset weighting coefficients with values ​​of 0.6 and 0.4, respectively. During system operation, a four-dimensional variational assimilation system based on a mesoscale ocean model is introduced. This system receives the latest ocean current observation data every ten minutes and outputs the full-field hydrodynamic field, including the velocity vectors of each grid point. The weights of the edges between nodes are dynamically adjusted based on the current propagation path and velocity: if node j is downstream of node i and the current velocity is greater than 0.1 m / s, the weight is increased. The weighting of the graph increases proportionally to the flow rate; conversely, it decreases. This mechanism ensures that the graph can reflect the actual impact of matter and energy transport in real time, making the correlation physically interpretable.

[0027] In step S4, the dynamic spatiotemporal correlation graph and the real-time collected monitoring data are input into a data quality assessment engine driven by a multimodal graph attention network. The core of this engine is a deep neural network model, and its internal processing flow includes four sub-steps. First, the input graph structure G=(V,E,W) and each node... The sequence of temporal feature vectors: ; Neighborhood information aggregation is performed using a two-layer graph convolutional network. The output of the first graph convolutional layer is: ; Where N(i) is the set of neighbors of node i. The first layer is a learnable weight matrix, and σ is the modified linear unit activation function. The second layer further aggregates information from the neighbors to obtain a feature embedding vector that initially incorporates spatial context information. ; Secondly, The sequence input is fed into a multi-head self-attention transformer encoding layer. This layer contains eight parallel attention heads, each of which computes a query, key, and value matrix and outputs a weighted sum. The sums are then concatenated and passed through a feedforward network to capture the temporal evolution patterns and long-range dependencies of the nodes, outputting a time-aware node state representation. ; Again, The input is fed into a graph attention network layer. This layer calculates the attention coefficients between node i and its neighbor j: ; in This represents vector concatenation. For learnable attention vectors, For shared weight matrices. (Through...) Dynamically assign neighbor information weights to the central node, so that the model focuses on the associated node with the most diagnostic value in the current state; Finally, the aggregated final node representations are fed into two parallel fully connected output layers. The first output layer contains a single neuron that outputs a data quality score through a sigmoid activation function. ∈[0,1]; The second output layer contains six neurons, corresponding to six abnormal categories: normal, sensor drift, bio-attachment interference, signal line failure, red tide outbreak precursor, and cold wave invasion. The classification probability distribution is output through a flexible maximum activation function. ; Data quality assessment report by Composition, in which For confidence level, This is an abnormal label.

[0028] In step S5, based on the data quality assessment report, such as Figure 5 As shown, activate the multi-tiered pricing and value discovery module; This module calculates the listed trading price for each of the three market tiers: In the primary raw data market, the formula for calculating the asset price P1 is: ; in The basic value is determined by the data type (for example, temperature data is 0.05 yuan per record, and chlorophyll concentration data is 0.2 yuan per record). Δt is the data quality score, Δt is the time difference between the current time and the data generation time (in hours), and λ is the timeliness decay coefficient, with a value of 0.02. In the Tier 2 verification information market, assets are data subsets or their aggregated statistics (such as daily averages or standard deviations) that have been verified by an evaluation engine to have a quality score higher than 0.9. Their price: ; Where C is the confidence level (i.e., ), Information gain is defined as the reduction in Shannon entropy of the subset of data relative to the original dataset, and γ is the premium coefficient with a value of 1.5. In the Level 3 decision analytics product market, assets are derivative products generated based on high-quality data. Its price ,in This is the sum of the procurement costs of all the original data used. The cost of computing resources consumed for model calculations (billed per GPU hour), π is the preset profit margin, typically 20%. All price calculation results are updated in real time and published to the trading engine along with the corresponding asset's metadata (including data source, time range, spatial coverage, quality report hash, etc.).

[0029] In step S6, the listed transaction price and corresponding asset information are published to the third-tier market trading execution engine. This engine maintains three independent order books, each corresponding to one of the three market tiers. Each order book records all unexecuted limit orders, sorted by price priority and then time priority. When a buyer submits a market order, the engine immediately starts matching from the best sell order until the order quantity is met; when a limit order is submitted, if a matching counterparty order exists, it is executed immediately; otherwise, it is placed pending. After a successful match, a transaction record is generated, including the transaction time, anonymous addresses of both the buyer and seller, transaction price, transaction volume, and unique asset identifier. The engine pushes real-time market data and transaction information to all online clients via WebSockets Protocol (WSPP) to ensure market transparency.

[0030] In step S7, the details of each completed transaction are encapsulated into a transaction record and written to the distributed ledger via a smart contract and blockchain notarization module. The transaction record includes the identities of both parties (processed with asymmetric encryption), the transaction price, a hash digest of the data asset (generated using the SHA-256 algorithm), and the digital signature of the corresponding data quality assessment report. The smart contract is deployed on the consortium blockchain, and its execution logic stipulates: after the buyer initiates a purchase request and transfers the agreed amount of digital tokens to the contract address, the contract automatically locks the funds; the seller then uploads the hash digest of the data asset to the contract; an independent oracle service retrieves the corresponding data quality assessment report from an off-chain database and verifies the validity of its digital signature and its consistency with the hash digest; after successful verification, the contract automatically transfers the funds to the seller's address, simultaneously sending the data asset access key to the buyer's address via asymmetric encryption, and permanently recording the complete transaction record as an event log on the blockchain. Any party can query the complete transaction history at any time via the transaction hash, ensuring full traceability.

[0031] like Figure 1 As shown, to support the above method, this invention also provides a corresponding system. This system includes a multi-source heterogeneous data acquisition module, a data preprocessing and spatiotemporal alignment module, a dynamic spatiotemporal correlation graph construction module, a data quality assessment engine driven by a multimodal graph attention network, a multi-level pricing and value discovery module, a three-tier market transaction execution engine, and a smart contract and blockchain notarization module. Each module interacts with the other through standardized application programming interfaces (APIs) and message queue services. It adopts a distributed architecture deployed on a cloud-edge-device integrated platform, with edge nodes responsible for data acquisition and real-time preprocessing, and cloud nodes handling complex computations and transaction services, ensuring low latency and high reliability.

[0032] The multi-source heterogeneous data acquisition module serves as the system's data input entry point, equipped with an IoT gateway (using Huawei OceanConnect IoT Gateway), a multi-protocol data interface adapter, and a manual data entry terminal. The IoT gateway establishes a stable connection with underwater sensor nodes via the LoRaWAN long-range wireless communication protocol, receiving real-time internal environmental data reported by devices such as CTD (Conductivity, Temperature, Depth) meters and acoustic Doppler current profilers. Data packets are Base64 encoded and include millisecond-accurate timestamps and GPS coordinates. The multi-protocol data interface adapter integrates protocols such as HTTP / HTTPS and WebSocket, periodically (every 30 seconds) calling third-party meteorological service APIs (such as the China Meteorological Data Network API) and satellite remote sensing data interfaces (such as the National Satellite Ocean Application Center interface) to acquire external environmental data such as wind speed and sea surface temperature. The manual data entry terminal uses a customized web page, allowing aquaculture farm managers to enter data such as feeding records and disease prevention records. Each record requires operator authentication with a password and an operation timestamp. This module has a built-in data buffer queue (using RabbitMQ), which formats the three types of data uniformly before pushing them to the data preprocessing module, while simultaneously retaining a 72-hour backup of the original data locally.

[0033] The data preprocessing and spatiotemporal alignment module is deployed on an edge computing node (using an Intel Xeon E-2274G processor) and integrates the Scikit-learn data normalization algorithm library, a Kriging spatial interpolation engine, and a cubic spline temporal interpolation engine. This module continuously monitors the data stream from the acquisition module and first performs a data cleaning process: outlier data points are removed using preset physical threshold rules (such as salinity 0-45‰, temperature 0-50℃), and missing values ​​are filled using the neighborhood mean. Then, the minimum-maximum scaling method is called to perform normalization processing, mapping all data to the [0,1] interval, and the normalization parameters are synchronized to the cloud parameter server in real time. In the spatiotemporal alignment stage, the Kriging spatial interpolation engine uses a 10m×10m×10m three-dimensional grid as a reference, combining sensor coordinates and measurement values ​​to estimate grid node parameters. The cubic spline temporal interpolation engine resamples all time-series data to a unified time axis with 5-second intervals. The processed data is organized as a four-dimensional tensor of longitude-latitude-depth-time, i.e., a standardized spatiotemporal data cube. This cube is stored in the InfluxDB high-performance time-series database, supporting millisecond-level response times for queries based on spatiotemporal range slices.

[0034] The dynamic spatiotemporal correlation map construction module establishes real-time communication with the cloud-based mesoscale ocean model service (using POM mode), and includes spatial meshing units and a dynamic weight adjustment engine. The spatial meshing units divide the three-dimensional space of the marine ranch into 10-meter-sided cubic grids, assigning a unique identifier to each grid, and mapping physical sensor locations to independent nodes. If a sensor is located within a grid, a node association mapping is established. The dynamic weight adjustment engine first calculates the initial weights of the edges using the formula: Achievement, where the Pearson correlation coefficient Calculations are based on historical data from the past 30 days; subsequently, hydrodynamic field data from a mesoscale ocean model is received every 10 minutes, and weights are adjusted according to flow direction and velocity: when node j is downstream of node i and the flow velocity is >0.1 m / s, the weight is increased proportionally to the flow velocity, and vice versa. The graph structure is updated in real time and stored in the Neo4j graph database, and synchronized to the data quality assessment engine through a subscription mechanism.

[0035] A multimodal graph attention network-driven data quality assessment engine is deployed on a dedicated computing server equipped with a Tensor Processing Unit (TPU). Developed based on the PyTorch deep learning framework, the model weight files are stored on SSD non-volatile memory and loaded into memory upon system startup. The engine continuously receives dynamic spatiotemporal correlation graphs and real-time monitoring data. It first aggregates neighborhood information through a two-layer graph convolutional network: the first layer takes graph node features and weights as input, activates them via a Corrected Linear Unit (ReLU), and outputs preliminary features; the second layer further aggregates neighborhood information to generate spatial context feature embedding vectors. The feature vector sequence is then input into the encoding layer of an 8-head self-attention transformer to capture temporal evolution patterns and long-range dependencies, outputting time-aware node representations; the graph attention network layer calculates attention coefficients. Dynamically assign neighbor weights to enhance anomaly detection capabilities; The final output is achieved through two parallel fully connected layers: a sigmoid activation function outputs the quality score in the [0,1] interval, and a softmax activation function outputs the probability distribution of six anomalies. The evaluation report (including score, confidence level, and anomaly label) is digitally signed and pushed to the multi-level pricing module and synchronized to the off-chain database.

[0036] The multi-tiered pricing and value discovery module embeds a tiered pricing algorithm engine, deployed on a cloud application server, and establishes bidirectional communication with the time-series database and trading engine. This module receives data quality assessment reports in real time, first querying the built-in basic value dictionary (e.g., temperature data 0.05 yuan / record, chlorophyll concentration 0.2 yuan / record), and then calculating the first-tier market price by combining the quality score and time-related decay factor. ; For data subsets or aggregate statistics with a quality score > 0.9, calculate the information gain (Shannon entropy reduction) and then add an authentication premium to obtain the second-tier price: ; For decision analysis derivative products, the hourly cost of GPUs is obtained by calling the computing resource monitoring service via an interface. This cost is then factored in, along with data procurement costs and a pre-set 20% profit margin. ; Calculate the third-level price. After the pricing result is bound to the asset metadata (source, spatiotemporal range, quality report hash), it is pushed to the asset release queue of the third-level market transaction execution engine.

[0037] The third-tier market trading execution engine, serving as the core trading service, is deployed in a highly available cluster (using a master-slave backup architecture). It comprises an order book management unit and a matching unit, supporting over 1000 orders per second. The order book management unit maintains three independent in-memory order books, each corresponding to one of the third-tier markets. Limit orders are sorted according to a "price priority, time priority" principle, using a red-black tree structure to ensure efficient querying. The matching unit listens for order submission events in real time. Market orders are immediately matched starting with the best counterparty. Limit orders are executed immediately if a matching counterparty exists; otherwise, they are placed and the order book is updated. Upon successful matching, a transaction record is generated, including the transaction time, anonymous address, price, volume, and unique asset identifier. Real-time market data is also pushed to all clients via the WebSocket protocol. The engine also provides a RESTful API interface, allowing users to query order status, historical transaction records, and account assets. The front-end uses Vue.js to develop a graphical interface, enabling order submission and market data viewing.

[0038] The smart contract and blockchain evidence storage module is deployed on a Hyperledger Fabric consortium blockchain architecture, with nodes including transaction partner nodes, regulatory nodes, and oracle nodes. The smart contracts are written in Go and include built-in transaction execution logic. That is, when the buyer submits a purchase request, the digital tokens need to be transferred to the contract address, and the contract locks the funds after the balance is verified; after receiving the notification, the seller uploads the SHA-256 hash digest of the data asset to the contract; the independent oracle node extracts the corresponding data quality assessment report from the off-chain database and verifies the validity of the digital signature of the report and its consistency with the hash digest; Upon successful verification, the contract automatically executes the fund transfer (transferring tokens to the seller's address) and sends the data access key to the buyer's address via asymmetric encryption. Simultaneously, the transaction details (including encrypted identity, price, hash digest, and evaluation report) are written to the blockchain as an event log. The consortium blockchain uses the PBFT consensus mechanism, with a block generation interval of 3 seconds. All transaction records are immutable, and both parties can query the complete process via the transaction hash on the blockchain explorer. Regulatory nodes have full data access permissions to ensure transaction compliance.

[0039] The entire system's modules communicate loosely via Kafka message queues and RESTful APIs, and employ Prometheus + Grafana for end-to-end monitoring, providing real-time monitoring and alerts for module operating status, data transmission latency, and transaction matching efficiency. The system supports horizontal scaling; adding new sensor types allows for expansion of the acquisition module interfaces through configuration files, and adding new anomaly categories allows for updating the evaluation engine through model fine-tuning tools, meeting the dynamic expansion needs of marine ranching trading scenarios.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-level trading method for marine ranching based on a multi-level market architecture, characterized in that, include: S1: Acquire multi-source heterogeneous monitoring data inside and outside the marine ranch. The multi-source heterogeneous monitoring data includes internal environmental data collected by a sensor network deployed in a preset monitoring area, external environmental data obtained through remote sensing satellites or meteorological service interfaces, and manually entered aquaculture operation data. S2: Perform data preprocessing and spatiotemporal alignment operations on the multi-source heterogeneous monitoring data. The operations include data cleaning, normalization, and unifying data from different sources and sampling frequencies onto a preset spatiotemporal grid reference through Kriging spatial interpolation and cubic spline temporal interpolation algorithms to form a standardized spatiotemporal data cube. S3: Based on the standardized spatiotemporal data cube, construct a dynamic spatiotemporal correlation map characterizing the intrinsic correlation of the marine ranch environmental elements. The nodes of the map represent independent sensors or divided geographic grid units, and the edges of the map represent the spatial proximity relationship, hydrodynamic correlation, or historical data statistical correlation between nodes. The weights of the edges are dynamically updated according to the real-time water flow model and the historical covariance matrix. S4: Input the dynamic spatiotemporal correlation map and the real-time collected monitoring data into a pre-trained multimodal graph attention network-driven data quality assessment engine. The engine scores each data point and identifies the anomaly type, and outputs a data quality assessment report containing data quality score, confidence level, and anomaly label. The anomaly label is distinguished into systemic environmental anomalies and random sensor anomalies. S5: Based on the data quality assessment report, activate the multi-level pricing and value discovery module to generate differentiated basic reference prices for data assets of different quality levels, and calculate and form the listing transaction prices of the first-level raw data market, the second-level verification information market, and the third-level decision analysis product market by combining market supply and demand, data timeliness and information entropy. S6: The listed transaction price and corresponding asset information are published to the third-level market transaction execution engine. The engine matches transactions through a continuous two-way auction mechanism based on the entrustment instructions of both the buyer and seller, and records the transaction details. S7: Encapsulate the details of each transaction, including the identities of both parties, the transaction price, the hash digest of the data asset, and the corresponding data quality assessment report, into a transaction record. Then, write it into the distributed ledger through a smart contract and blockchain notarization module to ensure that the transaction is immutable and fully traceable.

2. The multi-level trading method for marine ranching based on a multi-level market architecture according to claim 1, characterized in that, The data preprocessing and spatiotemporal alignment operations performed on the multi-source heterogeneous monitoring data include: Remove data points that clearly exceed the physical reasonable range and perform normalization using the minimum-maximum scaling method; The Kriging spatial interpolation algorithm is used to map spatially unevenly distributed data to regular three-dimensional grid nodes; The cubic spline time interpolation algorithm is used to resample time series with different sampling frequencies to a unified time axis; The processed data is organized into a four-dimensional tensor structure of longitude, latitude, depth, and time as the standardized spatiotemporal data cube.

3. The multi-level trading method for marine ranching based on a multi-level market architecture according to claim 1, characterized in that, Constructing the dynamic spatiotemporal correlation graph includes: The three-dimensional space of the marine ranch is divided into grid cells of equal volume, with each sensor or grid cell serving as a graph node; The initial weights of the edges are determined by the weighted sum of the reciprocal of the spatial Euclidean distance between nodes and the historical Pearson correlation coefficient. A four-dimensional variational assimilation system based on a mesoscale ocean model is introduced to dynamically adjust the weights of the edges between nodes according to the real-time water flow velocity and direction, so as to reflect the actual impact of material and energy transport.

4. The multi-level trading method for marine ranching based on a multi-level market architecture according to claim 1, characterized in that, The dynamic spatiotemporal correlation map and real-time collected monitoring data are input into a data quality assessment engine driven by a multimodal graph attention network, including: The neighborhood information of each node is aggregated by graph convolutional network layers to generate a feature embedding vector that initially integrates spatial context information. The feature embedding vector sequence is input into the transformer encoding layer of the multi-head self-attention mechanism to output a time-aware node state representation; The time-aware node state representation is input into the graph attention network layer, and the neighbor information weights are dynamically allocated by calculating the attention coefficients between nodes, thereby enhancing the ability to identify weak abnormal signals. The final node representation is fed into two parallel fully connected output layers. One outputs a data quality score between zero and one, and the other outputs a classification probability distribution corresponding to a preset anomaly category.

5. A multi-level trading method for marine ranching based on a multi-level market architecture as described in claim 4, characterized in that, The preset anomaly categories include normal, sensor drift, bio-attachment interference, signal line failure, red tide outbreak precursors, and cold wave intrusion.

6. The multi-level trading method for marine ranching based on a multi-level market architecture according to claim 1, characterized in that, The multi-tiered pricing and value discovery module calculates the listed transaction price, including: In the primary raw data market, the price equals the base value multiplied by the square of the data quality score, and then multiplied by the timeliness decay factor, which is an exponential decay function with respect to the time difference of data generation. In the second-tier verification information market, the price is based on the first-tier market price plus an authentication premium calculated based on confidence level and information gain; In the third-level decision analysis product market, prices are determined using a cost-plus pricing method, where costs include the cost of purchasing the data used, the resource consumption of the model calculations, and the preset profit margin.

7. The multi-level trading method for marine ranching based on a multi-level market architecture according to claim 1, characterized in that, The third-tier market transaction execution engine executes a continuous two-way auction mechanism, including: Maintain three separate order books, each corresponding to one of the three market tiers; Orders with price limits will be sorted according to the principles of price priority and time priority. Market orders are matched one by one, starting with the best counterparty, until the order quantity is met. Generate transaction records that include the transaction time, anonymous addresses of both the buyer and seller, transaction price, transaction volume, and unique identifier of the asset.

8. The multi-level trading method for marine ranching based on a multi-level market architecture according to claim 1, characterized in that, The smart contract and blockchain notarization module performs transaction settlement and notarization, including: The funds are automatically locked after the buyer transfers the agreed amount of digital tokens to the contract address; The seller uploads a hash summary of the data asset to the contract; An independent oracle service verifies the consistency between the hash digest and the data quality assessment report; Once verified, funds are automatically transferred to the seller's address, the buyer is granted data access rights, and the complete transaction record is permanently written to the blockchain as an event log.

9. A multi-level trading system for marine ranching based on a multi-level market architecture, applicable to the multi-level trading method for marine ranching based on a multi-level market architecture as described in any one of claims 1-8, characterized in that, include: A multi-source heterogeneous data acquisition module is used to acquire multi-source heterogeneous monitoring data inside and outside the marine ranch. The multi-source heterogeneous monitoring data includes internal environmental data, external environmental data, and aquaculture operation data. The data preprocessing and spatiotemporal alignment module is used to perform data cleaning, normalization, kriging spatial interpolation and cubic spline temporal interpolation on the multi-source heterogeneous monitoring data to generate a standardized spatiotemporal data cube. The dynamic spatiotemporal correlation map construction module is used to construct and dynamically update a dynamic spatiotemporal correlation map that characterizes the intrinsic correlation of marine ranch environmental elements based on preset rules and real-time hydrodynamic data. A multimodal graph attention network-driven data quality assessment engine is used to receive the dynamic spatiotemporal correlation graph and real-time monitoring data, and output a data quality assessment report containing data quality scores, confidence levels, and anomaly labels. The multi-tiered pricing and value discovery module is used to calculate the listing price for data assets at different market tiers based on the data quality assessment report. The third-level market transaction execution engine is used to receive transaction orders, execute continuous two-way auctions, and generate transaction records. The smart contract and blockchain evidence storage module is used to automate transaction settlement, asset delivery, and record transaction information on the blockchain.

10. A multi-level trading system for marine ranching based on a multi-level market architecture as described in claim 9, characterized in that, The multimodal graph attention network-driven data quality assessment engine is used for: Feature embedding vectors are generated by aggregating neighborhood information through graph convolutional network layers. The transformer coding layer outputs a time-aware node state representation through a multi-head self-attention mechanism; The ability to identify anomalies is enhanced by dynamically assigning neighbor information weights through a graph attention network layer. The data quality score and the probability distribution of anomaly categories are output through two parallel fully connected output layers.

Images