Business and financial fusion method and system based on data deep interaction and storage medium

Abstract

The invention discloses a business and financial fusion method and system based on data deep interaction, and a storage medium, and the method comprises the steps: collecting equipment operation data, financial cost data and sales prediction data in real time through an edge calculation node, constructing a dynamic classification pipeline through Apache NiFi, and carrying out the real-time collection of the data; in combination with DeltaLake hierarchical storage, structured data ACID transaction control and unstructured data key index extraction are realized; a multi-dimensional interaction model comprising a spatio-temporal feature extraction branch, a time sequence prediction branch and a cross-modal fusion branch is constructed, a graph neural network mining device is associated with topological features, a Transform model processes financial time sequence dependence, and a gating loop unit realizes business-financial feature dynamic weighted fusion; generating a multi-scene decision scheme through Monte Carlo tree search, issuing and executing the multi-scene decision scheme through an MQTT protocol, and constructing a feedback closed loop; a zero-trust architecture data channel and a block chain evidence storage mechanism are set to guarantee security, and the system comprises an edge computing layer, a cloud management layer, a security control layer and a man-machine collaboration layer.

Description

Technical Field

[0001] This invention relates to the field of financial management, and in particular to a business-finance integration method, system, and storage medium based on deep data interaction. Background Technology

[0002] Against the backdrop of the rapid development of Industry 4.0 and smart manufacturing technologies, the deep integration of business and financial data during enterprise digital transformation has become a core requirement for improving operational efficiency. With the widespread application of IoT, big data, and AI technologies, the need for data interaction between industrial production and financial management is becoming increasingly prominent. Enterprises urgently need to build intelligent systems to achieve dynamic collaboration between business and capital flows, thereby optimizing resource allocation and enhancing market competitiveness.

[0003] Existing technological frameworks still face multiple challenges in achieving business-finance data integration. Data generated by different business systems varies significantly in source, format, and update frequency, requiring substantial resources for cleaning and standardization during data integration. Furthermore, traditional data interaction models struggle to meet real-time requirements. Simultaneously, the information systems of business and finance departments often operate independently, leading to data silos that cause delays in early warning of cash flow risks and a lack of financial indicators to support production plan adjustments. In addition, existing analytical models often focus on single-dimensional data processing, lacking cross-modal feature fusion capabilities and failing to accurately capture the correlation between business dynamics and financial trends, thus limiting the in-depth mining of data value. Therefore, this paper proposes a business-finance integration method, system, and storage medium based on deep data interaction. Summary of the Invention

[0004] This invention provides the following technical solution: a business-finance integration method based on deep data interaction, comprising the following steps: S1. Multi-source heterogeneous data acquisition and dynamic preprocessing: By deploying edge computing nodes on the production site to collect equipment operation data, financial cost data and sales forecast data in real time, and using Apache NiFi to build a data flow pipeline, the system dynamically classifies sensor time-series data, ERP structured data and market unstructured data based on a business rule engine. A hierarchical storage architecture is built using Delta Lake to implement ACID transaction control for structured data and extract key financial indicators from unstructured data using a pre-trained natural language processing model. S2, Multi-dimensional Feature Interaction Modeling: Spatiotemporal feature extraction branch: Construct a graph neural network based on equipment operation data, mine the correlation topological features between production equipment, and generate equipment-level health index through graph pooling operation; Time series forecasting branch: Build a Transformer model with self-attention mechanism for financial data to capture the long-term and short-term dependencies between cash flow and costs; Cross-modal fusion branch: Gated loop units are used to dynamically assign weights to business features and financial features to generate a fusion feature matrix. The weighting strategy is based on the results of financial indicator sensitivity analysis. S3, Closed-loop decision optimization execution: Based on the fusion feature matrix, combined with sales forecast data and inventory cost data, a multi-scenario simulation scheme is generated through Monte Carlo tree search. The optimization results are sent to the production execution system via the MQTT protocol, and a decision-making effect feedback link is established. The model parameters are dynamically adjusted based on the actual execution results to form a continuous improvement closed loop. S4. End-to-end security control: A data interaction channel based on a zero-trust architecture is built through a security control layer. Dynamic access control is implemented for the integration process of business data and financial data. Blockchain technology is used to immutably store evidence of key decision-making nodes, and the stored evidence includes the decision context hash value.

[0005] Multi-source heterogeneous data acquisition and dynamic preprocessing stage: By deploying edge computing nodes on the production site, real-time data collection of equipment operation, financial cost, and sales forecasts is achieved, ensuring the timeliness and completeness of the raw data. Apache NiFi is used to build a data flow pipeline, which dynamically classifies and routes sensor time-series data, ERP structured data, and market unstructured data based on a predefined business rule engine. A hierarchical storage architecture is built using Delta Lake to implement ACID transaction control on structured data, ensuring consistency between data writing and querying; Key financial indicators are extracted from unstructured data using a pre-trained natural language processing model, and data standardization and semantic alignment are completed. Multi-dimensional feature interaction modeling stage: Spatiotemporal feature extraction branch: Graph neural network is constructed based on equipment operation data. The correlation topological features between production equipment are mined through node embedding learning, and equipment-level health index is generated by graph pooling operation; Time series forecasting branch: Build a Transformer model with self-attention mechanism for financial data, and capture the long-term and short-term dependencies between cash flow and costs through multi-head attention mechanism; Cross-modal fusion branch: Gated loop units are used to dynamically assign weights to business features and financial features to generate a fusion feature matrix. The weight assignment strategy is adaptively adjusted based on the results of financial indicator sensitivity analysis. Closed-loop decision-making optimization execution phase: Based on the fusion feature matrix, combined with sales forecast data and inventory cost data, a multi-scenario simulation solution is generated through the Monte Carlo tree search algorithm, covering decision-making dimensions such as capacity allocation and cost optimization. The optimization results are sent to the production execution system via the MQTT protocol, enabling lightweight transmission of instructions and real-time response. Establish a decision-making effect feedback loop, and dynamically adjust model parameters by analyzing the deviation between actual implementation results and simulation schemes to form a continuous improvement closed loop; End-to-end security control phase By building a data interaction channel based on a zero-trust architecture through a security control layer, dynamic access control is implemented for the integration process of business data and financial data, including identity authentication and fine-grained permission management. Blockchain technology is used to immutably store evidence of key decision-making nodes. The stored evidence includes the decision context hash value, timestamp, and operation log, ensuring audit traceability.

[0006] Preferably, the business rule engine further includes: The dynamic rule configuration module supports receiving account mapping rules defined by the finance department via RESTful API and updating them to the local rule base of the edge computing node in real time. The conflict detection submodule automatically triggers a manual arbitration process and stores the arbitration result in the rule base as a new entry when sensor data and ERP data have semantic conflicts at the field level. The rule version management unit establishes a blockchain evidence chain for historical rule change records, ensuring that rule traceability can be traced back to the approval process of specific business departments.

[0007] Dynamic rule configuration module: The system receives account mapping rules defined by the finance department via a RESTful API, parses the rule parameters, and generates a standardized configuration file. The configuration file is then synchronized to the local rule base of the edge computing node, and an incremental update mechanism is used to replace only the changed parts. After the update is completed, a rule hot loading process is triggered to ensure that business processing threads seamlessly switch to the new rules. Conflict detection submodule: When sensor data and ERP data have semantic conflicts at the field level, the metadata and context information of the conflicting fields are automatically extracted; a manual review process is initiated through a preset arbitration interface, and the operation log of the arbitrator is recorded; the arbitration result is converted into structured rule entries and written into the rule base after version verification. Rule version management unit: For each rule change operation, a timestamped version snapshot is generated, and the snapshot data is encapsulated into a storage block using blockchain technology; the signatures of the approval flow nodes are verified through smart contracts to build a complete traceability chain containing the applicant, approver, and changed content; and a query interface for searching by version hash value is provided.

[0008] Preferably, in the cross-modal fusion branch: The training process of the dynamic weight allocation mechanism adopts a progressive learning strategy. In the initial training phase, historical financial audit reports are used as supervision signals, and when the model converges, it switches to real-time business feedback data. The joint loss function includes a financial indicator sensitivity analysis term and an equipment downtime probability term. The sensitivity analysis term has its weighting coefficient determined by calculating the marginal effect of financial item fluctuations on net profit. After the fused feature matrix is ​​generated, it needs to pass through the integrity verification module of the security control layer, and use the SM3 hash algorithm to verify that the feature vector has not been tampered with.

[0009] Dynamic weight allocation mechanism: During the initial training phase, historical financial audit report text data is loaded, and semantic features are extracted using the BERT model as supervision signals. Once the model loss value converges to the threshold, the system switches to real-time business feedback data streams and dynamically adjusts the training sample weights using a sliding window mechanism. The contribution of the two data sources is periodically evaluated, and the weight allocation strategy is optimized. Joint loss function: A composite loss function is constructed, which includes a financial indicator sensitivity analysis term and an equipment downtime probability term. The sensitivity analysis term generates differentiable weighting coefficients by calculating the marginal effect of financial item fluctuations on net profit. The equipment downtime probability term introduces an LSTM network to predict equipment time-series data. The two losses are weighted and summed according to business priority. Fusion feature matrix verification: After generating the fused feature matrix, the integrity verification module of the security control layer is called; the SM3 hash algorithm is used to perform hash calculation on the feature vector, and the calculation result is compared with the stored value; if the verification fails, the feature isolation mechanism is triggered to block abnormal data from flowing into the decision layer.

[0010] Preferably, during the closed-loop decision optimization execution: The Monte Carlo tree search algorithm adopts a two-stage node expansion strategy. In the initial stage, it performs a fast traversal based on the historical scene library. When it enters the preset decision depth threshold, it switches to real-time data-driven simulation verification. A cost sensitivity coefficient is introduced when generating multi-scenario simulation solutions. This coefficient is dynamically adjusted based on the marginal contribution rate fed back from the financial module to ensure that the number of simulations of high-cost scenarios accounts for no less than 40% of the total number of simulations. Before the optimization results are issued, they need to pass the interface compatibility verification module of the production execution system. The verification content includes the device instruction set version, communication protocol encryption level, and abnormal interruption recovery strategy.

[0011] Monte Carlo Tree Search Algorithm: In the initial stage, typical business cases are loaded from the historical scenario library and quickly traversed using the UCT algorithm; when the search depth reaches a preset threshold, it switches to a real-time data-driven simulation verification mode to dynamically generate simulation environment parameters; and outputs a set of decision paths containing confidence intervals. Multi-scenario simulation solution generation workflow Calculate the cost sensitivity coefficient based on the marginal contribution rate fed back from the financial module, and adjust the sampling weight of high-cost scenarios; ensure that the number of simulations of high-cost scenarios accounts for no less than 40% of the total number of simulations, and optimize the allocation of computing resources through importance sampling techniques; Interface compatibility check: Before the optimization results are issued, the interface compatibility verification module of the production execution system is called to verify the matching between the equipment instruction set version and the target system, check whether the encryption level of the communication protocol meets the security baseline, and test the effectiveness of the recovery strategy under abnormal interruption scenarios.

[0012] Preferably, during the closed-loop decision optimization execution: The decision effect feedback link is constructed using a sliding time window mechanism, with the window length set to the most recent 3 complete production cycles. The model parameters are solidified only when the decision deviation rate of 2 consecutive windows is lower than the preset threshold. Gradient pruning protection is implemented during the model parameter adjustment process. When the parameter update exceeds three times the standard deviation of the historical mean, it automatically triggers a rollback to the most recent valid parameter version and generates an alarm event. The deviation analysis between the actual implementation results and the simulation plan includes three-dimensional evaluation indicators: deviation rate of financial indicators, fluctuation value of equipment utilization rate, and change in order delivery on-time rate.

[0013] Decision-making effect feedback loop: A sliding time window mechanism is adopted to continuously collect decision execution data from the three most recent complete production cycles; when the decision deviation rate of two consecutive windows is lower than the preset threshold, the model parameter solidification process is triggered to migrate the stable parameters to the production environment. Model parameter adjustment protection workflow During the parameter update process, a gradient pruning strategy is implemented. When the parameter update amount exceeds three times the standard deviation of the historical average, the system automatically rolls back to the most recent valid parameter version. Simultaneously, an alarm event containing a snapshot of abnormal parameters is generated and pushed to the operation and maintenance management platform. Deviation Analysis 3D Evaluation Workflow The actual results are compared and analyzed with the simulation plan to calculate the deviation rate of financial indicators, the fluctuation value of equipment utilization rate, and the change in order delivery timeliness rate; a deviation heat map containing three-dimensional evaluation values ​​is generated to provide a quantitative basis for subsequent optimization.

[0014] Preferably, in the end-to-end security control: The zero-trust architecture implementation includes a device fingerprint authentication submodule, which generates a unique device identifier by collecting the hardware serial number, BIOS version and boot log of the edge computing node; The dynamic access control policy adopts an attribute-based access control model. When a request to access financial data is detected, the system additionally verifies the mapping relationship between the operator's job role and the current business scenario's permissions. The blockchain evidence storage node deployment adopts a hybrid architecture, with key decision nodes stored on the consortium blockchain and ordinary operation logs stored on the private blockchain. The stored data includes operation timestamps, data hash values, and digital signature certificate serial numbers.

[0015] Zero-trust architecture implementation: When a device is connected, its hardware serial number, BIOS version, and boot log are collected and transmitted to the authentication server via an encrypted channel. A unique device identifier containing multiple hardware characteristics is generated and combined with a dynamic token to implement continuous authentication. An isolation process is automatically triggered for devices that fail to authenticate. Dynamic access control policy: When a request to access financial data is detected, the operator's job role information is extracted; the permission mapping table for the current business scenario is queried, and attribute-based access control (ABAC) decisions are implemented; a complete log containing the operation time, data range, and approval result is recorded. Blockchain evidence storage node deployment: Data from key decision-making nodes is stored on the consortium blockchain, while ordinary operation logs are stored on the private blockchain. The stored data includes operation timestamps, data hash values, and digital signature certificate serial numbers. Joint verification of the stored data is achieved through a cross-chain interaction protocol.

[0016] A business-finance integration system based on deep data interaction, employing the aforementioned business-finance integration method based on deep data interaction, includes: Edge computing layer: Deploy a lightweight feature extraction model, incorporate a data caching strategy based on business priorities, and automatically store critical device data to a local time-series database when the network is interrupted. The database uses a columnar storage format to optimize query performance. Cloud management layer: It includes an intelligent fusion analysis model training platform that supports configuring the number of graph neural network layers and attention heads for spatiotemporal feature extraction branches via a web interface, and provides a model performance monitoring dashboard; Security control layer: It integrates a hardware security module to perform encrypted calculations on the feature matrix during cross-modal fusion. Key management follows the NISTSP800-90A standard and supports key rotation strategy configuration. Human-machine collaboration layer: It provides configurable decision dashboards, supports drag-and-drop operations to create correlation analysis views between financial indicators and production parameters, and has a built-in adaptive early warning rule template library based on historical data.

[0017] The data acquisition and edge preprocessing stages correspond to the edge computing layer: Collect multi-source heterogeneous data through IoT devices deployed on the production site, including equipment operating parameters, financial transaction records and environmental monitoring data; The lightweight feature extraction model cleans the raw data in real time and dynamically adjusts the data caching strategy based on business priorities, separating and storing high-priority device status data from low-frequency financial data. When a network interruption is detected, the columnar storage engine of the local time-series database is automatically triggered to build an index structure by device ID-timestamp dimension to ensure the integrity of critical business data. The cloud-based intelligent analytics phase corresponds to the cloud-based management layer: Preprocessed data from the edge is uploaded to the cloud via a secure transmission protocol, where it undergoes spatiotemporal feature fusion by a graph neural network model. The model training platform dynamically constructs feature extraction branches based on the number of graph neural network layers (adjustable from 3 to 8 layers) and the number of multi-head attention mechanisms (configurable from 4 to 16 heads) configured by the user through the web interface. The real-time monitoring dashboard displays model training loss values, feature fusion entropy values, and inference latency metrics, and supports threshold alarm functions. The security control phase corresponds to the following security control layer: During the transmission of cross-modal feature matrices, the hardware security module uses the AES-256-GCM encryption algorithm to encapsulate the feature vectors; The key management system generates deterministic random bit sequences based on the NISTSP800-90A standard and supports automatic rotation of encryption keys on a 24-hour cycle. Implement role-based access control lists and record complete data operation audit logs; The human-machine collaborative decision-making stage corresponds to the human-machine collaboration layer: The decision dashboard loads pre-configured financial-production parameter association templates and supports the creation of dynamic association views through drag-and-drop operations; The adaptive warning engine automatically calculates the anomaly detection threshold based on historical data distribution, and triggers an alarm when the actual value deviates from the benchmark value by more than 3 standard deviations. Users can trace the feature fusion path through the interactive interface and view the business semantic interpretation of specific data points.

[0018] Preferably, the local time-series database of the edge computing layer: Query performance is optimized by adopting a columnar storage format, and time-series fields such as equipment vibration data and temperature data are compressed in segments. The compression algorithm incorporates a hybrid mode of run-length encoding and dictionary encoding, using run-length encoding for consecutive repeating values ​​and dictionary mapping for discrete symbolic values; When the network is restored, the data synchronization module automatically verifies the timestamps of the local cache and the cloud data, and only uploads incremental difference data.

[0019] After receiving time-series fields such as equipment vibration data and temperature data, the local time-series database uses a columnar storage format to vertically partition the data, storing fields of the same type in a continuous physical space. For the timestamp and numerical columns of the time series field, segmented compression processing is performed according to a preset time window: run-length encoding is applied to the continuous repeating numerical sequence to record the number of repetitions and the baseline value; for discrete symbolic fields (such as device status identifiers), an encoding table is generated through dictionary mapping to replace the original symbol values ​​with short codes; When the network connection is restored, the data synchronization module reads the timestamp sequence of the locally cached data, compares it with the corresponding timestamp in the cloud database, locates the difference range using a binary search algorithm, extracts only the incremental data within that range, encapsulates it into a transmission packet, and performs a one-way upload operation.

[0020] Preferably, the adaptive early warning rule template library of the human-machine collaboration layer includes: The threshold generator based on time series decomposition uses the STL decomposition algorithm to automatically identify the trend, seasonal and residual terms in the data. When the warning boundary conditions are updated, the paper document with the CFO's handwritten signature is identified by OCR technology and converted into structured rules and stored in the template library. Once the warning is triggered, the system automatically generates a draft response plan that includes abnormal data fragments, potentially affected subjects, and suggested countermeasures.

[0021] The threshold generator calls the Seasonal Decomposition Time Series Algorithm (STL) to perform a three-level decomposition on historical financial data: extracting the trend term through local weighted regression, identifying the periodic seasonal term through Fourier transform, and calculating the fluctuation range of the remaining residual term through moving average. When the CFO submits a paper rule change document with handwritten signatures, the optical character recognition module analyzes the document layout, locates the signature area, extracts the handwritten text, and converts it into structured JSON format after matching key parameters using regular expressions. Once the warning is triggered, the rules engine extracts abnormal data fragments from the time-series database, combines them with the knowledge graph to associate the affected accounting subjects, and calls the natural language generation template to output a draft resolution that includes quantitative impact analysis, historical case references, and operational suggestions.

[0022] A business-finance converged storage medium based on deep data interaction, comprising the aforementioned business-finance converged system based on deep data interaction, including: Storage media architecture design; This storage medium adopts a hierarchical storage architecture, including a read-only storage area and a writable storage area. The read-only storage area stores the pre-trained cross-modal fusion model parameter file and the business rule engine configuration template, while the writable storage area supports dynamic loading of the production site data fingerprint library and financial account mapping table. The medium has a built-in hardware-level encryption chip that transparently encrypts the parameters of the cross-modal fusion model. Key management follows the GM / T0028-2014 standard and supports dynamic distribution of one key per device. Adaptive and extended functionality: It supports remote updates of the business rule engine configuration template via OTA. The update package uses asymmetric encryption signature and only allows firmware upgrades signed by the authorized server. The built-in model version management module can roll back to historical version parameter files; It provides multi-format data interfaces, supporting data reading and writing via OPCUA, MQTT, and RESTfulAPI protocols. Interface parameters are dynamically bound through configuration files.

[0023] Storage medium initialization phase: After the system is powered on, the hardware-level encryption chip (compliant with GM / T0028-2014 standard) is automatically activated, and transparent encryption protection is implemented on the cross-modal fusion model parameter file fixed in the read-only storage area, generating a dynamic encryption key bound to the device's unique identifier; The writable storage area loads the production site data fingerprint database and financial account mapping table, dynamically binds multi-format data interface parameters through configuration files, and initializes the OPCUA, MQTT and RESTfulAPI protocol stack. Data interaction and processing stage: When an external system submits a business request through a multi-format data interface, the storage medium first verifies the compliance of the interface protocol: If it is the OPCUA protocol, resolve the node address mapping relationship; If it is an MQTT protocol, the subscription topic is associated with the business rule engine configuration template; If it is a RESTful API protocol, validate the validity of the HTTP method and URI path; The business rules engine dynamically loads the production site data fingerprint database from the writable storage area, and performs feature extraction and semantic fusion on the input data by combining the parameters of the pre-trained cross-modal fusion model. The financial account mapping table generates financial codes based on the data fusion results and writes them into the writable storage area temporary buffer according to predefined rules. Adaptive update and maintenance phase: When an OTA upgrade command is received, the storage medium verifies the validity of the asymmetric cryptographic signature of the upgrade package: If the signature verification passes, decrypt and overwrite the business rules engine configuration template; If the verification fails, the model version management module will be triggered to roll back to the historical version parameter file; The hardware-level encryption chip periodically polls the key management server and updates the encryption key through a one-machine-one-key dynamic distribution mechanism to ensure the security of cross-modal fusion model parameter file transmission.

[0024] In summary, compared with the prior art, the present invention provides a business-finance integration method, system, and storage medium based on deep data interaction, which has the following beneficial effects: This invention systematically addresses the technical shortcomings of traditional technical frameworks in areas such as data integration, real-time response, cross-modal analysis, and security by constructing a business-finance integration system with deep data interaction. Specifically, in the data acquisition and preprocessing stage, the collaborative mechanism between edge computing nodes and the Apache NiFi data flow pipeline enables real-time dynamic classification of equipment operation data, financial cost data, and sales forecast data. Compared to the static mapping rules that rely on manual configuration in existing technologies, this method's dynamic classification strategy based on a business rule engine can adapt to changes in data format, significantly reducing data cleaning costs. DeltaLake's hierarchical storage architecture implements ACID transaction control for structured data, ensuring the consistency and integrity of financial data. For unstructured market data, pre-trained natural language processing models can automatically extract key financial indicators, breaking through the excessive reliance of traditional methods on structured data. For the decision-making and execution phase, the closed-loop optimization mechanism constructed in this invention enables real-time dynamic response to business adjustments. The multi-scenario simulation scheme generated based on Monte Carlo tree search provides a quantitative evaluation basis for decision-making. The decision distribution and feedback link supported by the MQTT protocol allows model parameters to be dynamically adjusted according to the actual execution effect. Compared with the traditional open-loop decision-making mode, this closed-loop mechanism shortens the production plan adjustment cycle and improves the response speed of cash flow risk warning. In addition, the configurable decision dashboard and adaptive warning rule template library provided by the human-machine collaboration layer allow business personnel to create a view of the relationship between financial indicators and production parameters through drag-and-drop operation, further enhancing the visualization and interpretability of the decision-making process. At the feature modeling level, the multi-dimensional interaction mechanism proposed in this invention effectively overcomes the limitations of single-dimensional analysis. The spatiotemporal feature extraction branch mines the topological features associated with production equipment through graph neural networks and generates equipment-level health indices by combining graph pooling operations, thereby achieving accurate modeling of the dynamic correlation of the production system. The time-series prediction branch uses a Transformer model with a self-attention mechanism to process financial data. Compared with traditional LSTM networks, this model can more effectively capture the long-term and short-term dependencies between cash flow and costs. The cross-modal fusion branch realizes the dynamic weight allocation of business features and financial features through gated recurrent units. Its weight strategy is based on the results of financial indicator sensitivity analysis, ensuring the dominance of key financial indicators in the fusion process. This multi-branch collaborative modeling approach, combined with the configuration of model layers and attention heads supported by the cloud management layer, enables the analysis model to simultaneously handle spatiotemporal correlations, time-series dependencies, and cross-modal interactions, significantly improving the accuracy of correlation analysis between business dynamics and financial trends. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the business-finance integration method based on deep data interaction according to the present invention.

[0026] Figure 2 This is a schematic diagram of the business-finance integration method based on deep data interaction of the present invention.

[0027] Figure 3 This is a schematic diagram of the business-finance integrated storage medium based on deep data interaction according to the present invention. Detailed Implementation

[0028] 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.

[0029] Please see Figure 1 This invention provides a technical solution, a business-finance integration method based on deep data interaction, comprising the following steps: S1. Multi-source heterogeneous data acquisition and dynamic preprocessing: By deploying edge computing nodes on the production site to collect equipment operation data, financial cost data and sales forecast data in real time, and using Apache NiFi to build a data flow pipeline, the system dynamically classifies sensor time-series data, ERP structured data and market unstructured data based on a business rule engine. A hierarchical storage architecture is built using Delta Lake to implement ACID transaction control for structured data and extract key financial indicators from unstructured data using a pre-trained natural language processing model. S2, Multi-dimensional Feature Interaction Modeling: Spatiotemporal feature extraction branch: Construct a graph neural network based on equipment operation data, mine the correlation topological features between production equipment, and generate equipment-level health index through graph pooling operation; Time series forecasting branch: Build a Transformer model with self-attention mechanism for financial data to capture the long-term and short-term dependencies between cash flow and costs; Cross-modal fusion branch: Gated loop units are used to dynamically assign weights to business features and financial features to generate a fusion feature matrix. The weighting strategy is based on the results of financial indicator sensitivity analysis. S3, Closed-loop decision optimization execution: Based on the fusion feature matrix, combined with sales forecast data and inventory cost data, a multi-scenario simulation scheme is generated through Monte Carlo tree search. The optimization results are sent to the production execution system via the MQTT protocol, and a decision-making effect feedback link is established. The model parameters are dynamically adjusted based on the actual execution results to form a continuous improvement closed loop. S4. End-to-end security control: A data interaction channel based on a zero-trust architecture is built through a security control layer. Dynamic access control is implemented for the integration process of business data and financial data. Blockchain technology is used to immutably store evidence of key decision-making nodes, and the stored evidence includes the decision context hash value.

[0030] Multi-source heterogeneous data acquisition and dynamic preprocessing stage: By deploying edge computing nodes on the production site, real-time data collection of equipment operation, financial cost, and sales forecasts is achieved, ensuring the timeliness and completeness of the raw data. Apache NiFi is used to build a data flow pipeline, which dynamically classifies and routes sensor time-series data, ERP structured data, and market unstructured data based on a predefined business rule engine. A hierarchical storage architecture is built using Delta Lake to implement ACID transaction control on structured data, ensuring consistency between data writing and querying; Key financial indicators are extracted from unstructured data using a pre-trained natural language processing model, and data standardization and semantic alignment are completed. By collaborating with edge computing nodes and Apache NiFi, real-time acquisition and classification of multi-source heterogeneous data can be achieved, reducing the impact of data latency on business decisions. DeltaLake's tiered storage architecture, combined with ACID transaction control, effectively solves the read-write conflict problem of traditional data lakes and improves data consistency. Natural language processing models automatically extract financial indicators from unstructured data, reducing human intervention and improving data utilization. Multi-dimensional feature interaction modeling stage: Spatiotemporal feature extraction branch: Graph neural network is constructed based on equipment operation data. The correlation topological features between production equipment are mined through node embedding learning, and equipment-level health index is generated by graph pooling operation; Time series forecasting branch: Build a Transformer model with self-attention mechanism for financial data, and capture the long-term and short-term dependencies between cash flow and costs through multi-head attention mechanism; Cross-modal fusion branch: Gated loop units are used to dynamically assign weights to business features and financial features to generate a fusion feature matrix. The weight assignment strategy is adaptively adjusted based on the results of financial indicator sensitivity analysis. The joint modeling of graph neural networks and Transformer models simultaneously captures the spatiotemporal correlation features of equipment and the time series patterns of finance, thereby improving the accuracy of the prediction model. The dynamic weight allocation mechanism of the gated loop unit makes the integration of business characteristics and financial characteristics more in line with the sensitivity of actual business, and enhances the interpretability of the model and the credibility of decision-making. Closed-loop decision-making optimization execution phase: Based on the fusion feature matrix, combined with sales forecast data and inventory cost data, a multi-scenario simulation solution is generated through the Monte Carlo tree search algorithm, covering decision-making dimensions such as capacity allocation and cost optimization. The optimization results are sent to the production execution system via the MQTT protocol, enabling lightweight transmission of instructions and real-time response. Establish a decision-making effect feedback loop, and dynamically adjust model parameters by analyzing the deviation between actual implementation results and simulation schemes to form a continuous improvement closed loop; Monte Carlo tree search generates multi-scenario simulation solutions, expanding the decision-making exploration space and improving the robustness of the solutions to uncertainty; The MQTT protocol decouples command issuance from the production system, reducing system maintenance costs. The feedback link dynamic parameter tuning mechanism forms a continuous improvement closed loop, enabling the system to have adaptive optimization capabilities and reducing reliance on manual parameter tuning; End-to-end security control phase By building a data interaction channel based on a zero-trust architecture through a security control layer, dynamic access control is implemented for the integration process of business data and financial data, including identity authentication and fine-grained permission management. Blockchain technology is used to immutably store evidence of key decision-making nodes. The stored evidence includes the decision context hash value, timestamp, and operation log to ensure audit traceability. Zero-trust architecture builds dynamic access control, effectively defending against the risk of internal data misuse and reducing the probability of data leakage; Blockchain-based evidence storage technology ensures the traceability of the decision-making process, enhances business compliance, and meets audit and regulatory requirements.

[0031] The business rules engine further includes: The dynamic rule configuration module supports receiving account mapping rules defined by the finance department via RESTful API and updating them to the local rule base of the edge computing node in real time. The conflict detection submodule automatically triggers a manual arbitration process and stores the arbitration result in the rule base as a new entry when sensor data and ERP data have semantic conflicts at the field level. The rule version management unit establishes a blockchain evidence chain for historical rule change records, ensuring that rule traceability can be traced back to the approval process of specific business departments.

[0032] Dynamic rule configuration module: The system receives account mapping rules defined by the finance department via a RESTful API, parses the rule parameters, and generates a standardized configuration file. The configuration file is then synchronized to the local rule base of the edge computing node, and an incremental update mechanism is used to replace only the changed parts. After the update is completed, a rule hot loading process is triggered to ensure that business processing threads seamlessly switch to the new rules. It enables real-time linkage between financial rules and edge computing nodes, shortens the rule deployment cycle to minutes, eliminates the risk of errors in manual configuration, and supports dynamic collaboration of cross-departmental business rules; Conflict detection submodule: When sensor data and ERP data have semantic conflicts at the field level, the metadata and context information of the conflicting fields are automatically extracted; a manual review process is initiated through a preset arbitration interface, and the operation log of the arbitrator is recorded; the arbitration result is converted into structured rule entries and written into the rule base after version verification. Establish a closed-loop mechanism for handling data conflicts to reduce the rate of data inconsistency across systems, while transforming human decision-making experience into reusable rule assets to enhance audit traceability capabilities. Rule version management unit: For each rule change operation, a timestamped version snapshot is generated, and the snapshot data is encapsulated into a storage block using blockchain technology; the signatures of the approval flow nodes are verified through smart contracts to build a complete traceability chain containing the applicant, approver, and change content; and a query interface for searching by version hash value is provided. Ensure that the rule change process is immutable, meet compliance audit requirements, and support quick rollback to any historical version to reduce operational risks in rule management.

[0033] In the cross-modal fusion branch: The training process of the dynamic weight allocation mechanism adopts a progressive learning strategy. In the initial training phase, historical financial audit reports are used as supervision signals, and when the model converges, it switches to real-time business feedback data. The joint loss function includes a financial indicator sensitivity analysis term and an equipment downtime probability term. The sensitivity analysis term has its weighting coefficient determined by calculating the marginal effect of financial item fluctuations on net profit. After the fused feature matrix is ​​generated, it needs to pass through the integrity verification module of the security control layer, and use the SM3 hash algorithm to verify that the feature vector has not been tampered with.

[0034] Dynamic weight allocation mechanism: During the initial training phase, historical financial audit report text data is loaded, and semantic features are extracted using the BERT model as supervision signals. Once the model loss value converges to the threshold, the system switches to real-time business feedback data streams and dynamically adjusts the training sample weights using a sliding window mechanism. The contribution of the two data sources is periodically evaluated, and the weight allocation strategy is optimized. By balancing historical experience with real-time feedback, the model's adaptability to dynamic business scenarios is improved, and prediction bias caused by data distribution shifts is reduced. Joint loss function: A composite loss function is constructed, which includes a financial indicator sensitivity analysis term and an equipment downtime probability term. The sensitivity analysis term generates differentiable weighting coefficients by calculating the marginal effect of financial item fluctuations on net profit. The equipment downtime probability term introduces an LSTM network to predict equipment time-series data. The two losses are weighted and summed according to business priority. Strengthen the model's responsiveness to key financial indicators, while controlling the risk exposure of the production system to achieve a balanced optimization of economic benefits and equipment reliability; Fusion feature matrix verification: After generating the fused feature matrix, the integrity verification module of the security control layer is called; the SM3 hash algorithm is used to perform hash calculation on the feature vector, and the calculation result is compared with the stored value; if the verification fails, the feature isolation mechanism is triggered to block abnormal data from flowing into the decision layer; To prevent feature data from being tampered with during transmission or processing, ensure the integrity of decision-making basis, and enhance the system's ability to resist data poisoning attacks.

[0035] Closed-loop decision optimization execution: The Monte Carlo tree search algorithm adopts a two-stage node expansion strategy. In the initial stage, it performs a fast traversal based on the historical scene library. When it enters the preset decision depth threshold, it switches to real-time data-driven simulation verification. A cost sensitivity coefficient is introduced when generating multi-scenario simulation solutions. This coefficient is dynamically adjusted based on the marginal contribution rate fed back from the financial module to ensure that the number of simulations of high-cost scenarios accounts for no less than 40% of the total number of simulations. Before the optimization results are issued, they need to pass the interface compatibility verification module of the production execution system. The verification content includes the device instruction set version, communication protocol encryption level, and abnormal interruption recovery strategy.

[0036] Monte Carlo Tree Search Algorithm: In the initial stage, typical business cases are loaded from the historical scenario library and quickly traversed using the UCT algorithm; when the search depth reaches a preset threshold, it switches to a real-time data-driven simulation verification mode to dynamically generate simulation environment parameters; and outputs a set of decision paths containing confidence intervals. Balancing the efficiency of utilizing historical experience and real-time data, while ensuring the breadth of exploration, improves the simulation accuracy of key decision-making nodes; Multi-scenario simulation solution generation workflow Calculate the cost sensitivity coefficient based on the marginal contribution rate fed back from the financial module, and adjust the sampling weight of high-cost scenarios; ensure that the number of simulations of high-cost scenarios accounts for no less than 40% of the total number of simulations, and optimize the allocation of computing resources through importance sampling techniques; Focus on in-depth analysis of high-risk business scenarios to avoid excessive resource dispersion and improve decision-making robustness in extreme situations; Interface compatibility check: Before the optimization results are issued, the interface compatibility verification module of the production execution system is called to verify the matching between the equipment instruction set version and the target system, check whether the encryption level of the communication protocol meets the security baseline, and test the effectiveness of the recovery strategy under abnormal interruption scenarios. Eliminate technical gaps at the decision-making and execution level, ensure seamless integration of optimization instructions with the production environment, and reduce system integration risks.

[0037] Closed-loop decision optimization execution: The decision effect feedback link is constructed using a sliding time window mechanism, with the window length set to the most recent 3 complete production cycles. The model parameters are solidified only when the decision deviation rate of 2 consecutive windows is lower than the preset threshold. Gradient pruning protection is implemented during the model parameter adjustment process. When the parameter update exceeds three times the standard deviation of the historical mean, it automatically triggers a rollback to the most recent valid parameter version and generates an alarm event. The deviation analysis between the actual implementation results and the simulation plan includes three-dimensional evaluation indicators: deviation rate of financial indicators, fluctuation value of equipment utilization rate, and change in order delivery on-time rate.

[0038] Decision-making effect feedback loop: A sliding time window mechanism is adopted to continuously collect decision execution data from the three most recent complete production cycles; when the decision deviation rate of two consecutive windows is lower than the preset threshold, the model parameter solidification process is triggered to migrate the stable parameters to the production environment. Establish a model iteration mechanism based on long-term effects to avoid overfitting caused by short-term fluctuations and improve the industrial-grade stability of the decision-making model; Model parameter adjustment protection workflow During the parameter update process, a gradient pruning strategy is implemented. When the parameter update amount exceeds three times the standard deviation of the historical average, the system automatically rolls back to the most recent valid parameter version. Simultaneously, an alarm event containing a snapshot of abnormal parameters is generated and pushed to the operation and maintenance management platform. To prevent parameter oscillations in the model due to abnormal data, ensure the continuity of decision-making services, and provide traceable records of anomaly handling; Deviation Analysis 3D Evaluation Workflow The actual results were compared and analyzed with the simulation plan. The deviation rate of financial indicators, the fluctuation value of equipment utilization rate, and the change of order delivery on-time rate were calculated. A deviation heatmap containing three-dimensional evaluation values ​​was generated to provide a quantitative basis for subsequent optimization. Establish a multi-dimensional decision-making quality assessment system to accurately identify and optimize shortcomings, and guide cross-departmental collaborative improvements.

[0039] In end-to-end security control: The zero-trust architecture implementation includes a device fingerprint authentication submodule, which generates a unique device identifier by collecting the hardware serial number, BIOS version and boot log of the edge computing node; The dynamic access control policy adopts an attribute-based access control model. When a request to access financial data is detected, the system additionally verifies the mapping relationship between the operator's job role and the current business scenario's permissions. The blockchain evidence storage node deployment adopts a hybrid architecture, with key decision nodes stored on the consortium blockchain and ordinary operation logs stored on the private blockchain. The stored data includes operation timestamps, data hash values, and digital signature certificate serial numbers.

[0040] Zero-trust architecture implementation: When a device is connected, its hardware serial number, BIOS version, and boot log are collected and transmitted to the authentication server via an encrypted channel. A unique device identifier containing multiple hardware characteristics is generated and combined with a dynamic token to implement continuous authentication. An isolation process is automatically triggered for devices that fail to authenticate. Build a fine-grained device identity management system to effectively prevent the risk of unauthorized device access and improve the security of the edge computing environment; Dynamic access control policy: When a request to access financial data is detected, the operator's job role information is extracted; the permission mapping table for the current business scenario is queried, and attribute-based access control (ABAC) decisions are implemented; a complete log containing the operation time, data range, and approval result is recorded. To achieve dynamic adaptation of permissions, avoid the problem of excessive or insufficient permissions, and meet the principle of least privilege access to financial data; Blockchain evidence storage node deployment: Key decision-making node data is stored on a consortium blockchain, while ordinary operation logs are stored on a private blockchain. The stored data includes operation timestamps, data hash values, and digital signature certificate serial numbers. Joint verification of the stored data is achieved through a cross-chain interaction protocol. Balancing evidence storage performance with compliance requirements, key data is verified through multi-party consensus, and ordinary logs are efficiently stored, thereby enhancing end-to-end audit capabilities.

[0041] Please see Figure 2 A business-finance integration system based on deep data interaction, employing the aforementioned business-finance integration method based on deep data interaction, includes: Edge computing layer: Deploy a lightweight feature extraction model, incorporate a data caching strategy based on business priorities, and automatically store critical device data to a local time-series database when the network is interrupted. The database uses a columnar storage format to optimize query performance. Cloud management layer: It includes an intelligent fusion analysis model training platform that supports configuring the number of graph neural network layers and attention heads for spatiotemporal feature extraction branches via a web interface, and provides a model performance monitoring dashboard; Security control layer: It integrates a hardware security module to perform encrypted calculations on the feature matrix during cross-modal fusion. Key management follows the NISTSP800-90A standard and supports key rotation strategy configuration. Human-machine collaboration layer: It provides configurable decision dashboards, supports drag-and-drop operations to create correlation analysis views between financial indicators and production parameters, and has a built-in adaptive early warning rule template library based on historical data.

[0042] The data acquisition and edge preprocessing stages correspond to the edge computing layer: Collect multi-source heterogeneous data through IoT devices deployed on the production site, including equipment operating parameters, financial transaction records and environmental monitoring data; The lightweight feature extraction model cleans the raw data in real time and dynamically adjusts the data caching strategy based on business priorities, separating and storing high-priority device status data from low-frequency financial data. When a network interruption is detected, the columnar storage engine of the local time-series database is automatically triggered to build an index structure by device ID-timestamp dimension to ensure the integrity of critical business data. Columnar storage format reduces the response time for time-series data queries to sub-second levels, improving query efficiency compared to traditional row storage solutions; The business priority caching strategy improves the data integrity rate of critical equipment data storage and significantly reduces the data backhaul pressure after network recovery; Lightweight model deployment reduces computing power consumption at the edge and is compatible with various industrial gateway hardware platforms; The cloud-based intelligent analytics phase corresponds to the cloud-based management layer: Preprocessed data from the edge is uploaded to the cloud via a secure transmission protocol, where it undergoes spatiotemporal feature fusion by a graph neural network model. The model training platform dynamically constructs feature extraction branches based on the number of graph neural network layers (adjustable from 3 to 8 layers) and the number of multi-head attention mechanisms (configurable from 4 to 16 heads) configured by the user through the web interface. The real-time monitoring dashboard displays model training loss values, feature fusion entropy values, and inference latency metrics, and supports threshold alarm functions. Configurable graph neural network structures shorten model iteration cycles and adapt to the needs of business-finance integration scenarios in different industries; Real-time monitoring dashboards reduce model maintenance costs and enable transparent management of the feature fusion process; Multimodal feature fusion improves the accuracy of financial forecasting, outperforming single data source analysis solutions. The security control phase corresponds to the following security control layer: During the transmission of cross-modal feature matrices, the hardware security module uses the AES-256-GCM encryption algorithm to encapsulate the feature vectors; The key management system generates deterministic random bit sequences based on the NISTSP800-90A standard and supports automatic rotation of encryption keys on a 24-hour cycle. Implement role-based access control lists and record complete data operation audit logs; The hardware-level encryption module keeps the increase in end-to-end data transmission latency to within 5ms, balancing security and real-time performance. The key management system, which conforms to NIST standards, has passed Level 3 Information Security Protection Certification and meets financial-grade data security requirements. The dynamic key rotation strategy reduces the risk of brute-force attacks by 90% and enhances the security of cross-domain data interaction. The human-machine collaborative decision-making stage corresponds to the human-machine collaboration layer: The decision dashboard loads pre-configured financial-production parameter association templates and supports the creation of dynamic association views through drag-and-drop operations; The adaptive warning engine automatically calculates the anomaly detection threshold based on historical data distribution, and triggers an alarm when the actual value deviates from the benchmark value by more than 3 standard deviations. Users can trace the feature fusion path through the interactive interface and view the business semantic interpretation of specific data points; Visualized correlation analysis enables business personnel to build analytical models independently, improving efficiency and reducing reliance on the IT department; The adaptive early warning mechanism reduces the false alarm rate and optimizes alarm quality compared to the fixed threshold scheme. Interactive feature backtracking shortens the time for locating the root cause of anomalies and improves decision-making response speed.

[0043] Local time-series database for edge computing layer: Query performance is optimized by adopting a columnar storage format, and time-series fields such as equipment vibration data and temperature data are compressed in segments. The compression algorithm incorporates a hybrid mode of run-length encoding and dictionary encoding, using run-length encoding for consecutive repeating values ​​and dictionary mapping for discrete symbolic values; When the network is restored, the data synchronization module automatically verifies the timestamps of the local cache and the cloud data, and only uploads incremental difference data.

[0044] After receiving time-series fields such as equipment vibration data and temperature data, the local time-series database uses a columnar storage format to vertically partition the data, storing fields of the same type in a continuous physical space. For the timestamp and numerical columns of the time series field, segmented compression processing is performed according to a preset time window: run-length encoding is applied to the continuous repeating numerical sequence to record the number of repetitions and the baseline value; for discrete symbolic fields (such as device status identifiers), an encoding table is generated through dictionary mapping to replace the original symbol values ​​with short codes; When the network connection is restored, the data synchronization module reads the timestamp sequence of the locally cached data, compares it with the corresponding timestamp in the cloud database, locates the difference range through a binary search algorithm, extracts only the incremental data within the range, encapsulates it into a transmission packet, and performs a one-way upload operation. Columnar storage improves the response speed of time-series data range queries by reducing disk I / O operations and increasing cache utilization. Hybrid compression mode combines the advantages of run-length encoding and dictionary encoding, reducing CPU decoding overhead and storage space usage while maintaining a high compression ratio; The incremental synchronization mechanism avoids full data retransmission, significantly reducing network bandwidth consumption and cloud storage redundancy.

[0045] Adaptive early warning rule template library for human-machine collaboration layer: The threshold generator based on time series decomposition uses the STL decomposition algorithm to automatically identify the trend, seasonal and residual terms in the data. When the warning boundary conditions are updated, the paper document with the CFO's handwritten signature is identified by OCR technology and converted into structured rules and stored in the template library. Once the warning is triggered, the system automatically generates a draft response plan that includes abnormal data fragments, potentially affected subjects, and suggested countermeasures.

[0046] The threshold generator calls the Seasonal Decomposition Time Series Algorithm (STL) to perform a three-level decomposition on historical financial data: extracting the trend term through local weighted regression, identifying the periodic seasonal term through Fourier transform, and calculating the fluctuation range of the remaining residual term through moving average. When the CFO submits a paper rule change document with handwritten signatures, the optical character recognition module analyzes the document layout, locates the signature area, extracts the handwritten text, and converts it into structured JSON format after matching key parameters using regular expressions. After the warning is triggered, the rule engine extracts abnormal data fragments from the time series database, combines them with the knowledge graph to associate the affected accounting subjects, and calls the natural language generation template to output a draft disposal plan that includes quantitative impact analysis, historical case references and operational suggestions. The STL decomposition algorithm removes data noise, improving the accuracy of dynamic threshold generation and reducing false alarms caused by seasonal fluctuations. The OCR and rule parsing process realizes a digital closed loop for paper-based approvals, shortening the rule update cycle; The automated draft generation mechanism shortens the response time for finance personnel and reduces the risk of omissions in human analysis.

[0047] Please see Figure 3 A business-finance converged storage medium based on deep data interaction, and based on the aforementioned business-finance converged system based on deep data interaction, comprising: Storage media architecture design; This storage medium adopts a hierarchical storage architecture, including a read-only storage area and a writable storage area. The read-only storage area stores the pre-trained cross-modal fusion model parameter file and the business rule engine configuration template, while the writable storage area supports dynamic loading of the production site data fingerprint library and financial account mapping table. The medium has a built-in hardware-level encryption chip that transparently encrypts the parameters of the cross-modal fusion model. Key management follows the GM / T0028-2014 standard and supports dynamic distribution of one key per device. Adaptive and extended functionality: It supports remote updates of the business rule engine configuration template via OTA. The update package uses asymmetric encryption signature and only allows firmware upgrades signed by the authorized server. The built-in model version management module can roll back to historical version parameter files; It provides multi-format data interfaces, supporting data reading and writing via OPCUA, MQTT, and RESTfulAPI protocols. Interface parameters are dynamically bound through configuration files.

[0048] Storage medium initialization phase: After the system is powered on, the hardware-level encryption chip (compliant with GM / T0028-2014 standard) is automatically activated, and transparent encryption protection is implemented on the cross-modal fusion model parameter file fixed in the read-only storage area, generating a dynamic encryption key bound to the device's unique identifier; The writable storage area loads the production site data fingerprint database and financial account mapping table, dynamically binds multi-format data interface parameters through configuration files, and initializes the OPCUA, MQTT and RESTfulAPI protocol stack. Data interaction and processing stage: When an external system submits a business request through a multi-format data interface, the storage medium first verifies the compliance of the interface protocol: If it is the OPCUA protocol, resolve the node address mapping relationship; If it is an MQTT protocol, the subscription topic is associated with the business rule engine configuration template; If it is a RESTful API protocol, validate the validity of the HTTP method and URI path; The business rules engine dynamically loads the production site data fingerprint database from the writable storage area, and performs feature extraction and semantic fusion on the input data by combining the parameters of the pre-trained cross-modal fusion model. The financial account mapping table generates financial codes based on the data fusion results and writes them into the writable storage area temporary buffer according to predefined rules. Adaptive update and maintenance phase: When an OTA upgrade command is received, the storage medium verifies the validity of the asymmetric cryptographic signature of the upgrade package: If the signature verification passes, decrypt and overwrite the business rules engine configuration template; If the verification fails, the model version management module will be triggered to roll back to the historical version parameter file; The hardware-level encryption chip periodically polls the key management server and updates the encryption key through a one-machine-one-key dynamic distribution mechanism to ensure the security of cross-modal fusion model parameter file transmission; By designing physically isolated read-only and writable storage areas, the core model parameters and dynamic business data are decoupled and stored, preventing data tampering in the production environment from causing model failure, while also supporting rapid iterative updates of the financial account mapping table. It has a built-in encryption chip that conforms to the GM / T0028-2014 standard, which transparently encrypts the parameters of the cross-modal fusion model. Combined with a one-device-one-key dynamic key distribution mechanism, it effectively resists side-channel attacks and man-in-the-middle theft risks and meets the Level 3 security requirements of the Information Security Protection 2.0 standard. OTA upgrade packages use asymmetric encrypted signature verification, allowing only firmware updates signed by authorized servers to prevent unauthorized firmware injection; the model version management module supports rapid rollback of historical versions to ensure business continuity is not affected by abnormal upgrades. It supports three mainstream industrial protocols: OPCUA, MQTT and RESTfulAPI. It can dynamically bind interface parameters through configuration files to adapt to the data interaction specifications of different manufacturers' equipment, thereby reducing system integration costs and protocol conversion delays. The business rules engine configuration template supports remote hot updates, allowing adjustments to the financial subject mapping logic without downtime. This adapts to changes in accounting standards or production processes, enhancing the agile response capabilities of the enterprise's business-finance integration system.

[0049] This solution systematically addresses the technical shortcomings of traditional technology frameworks in areas such as data integration, real-time response, cross-modal analysis, and security by constructing a business-finance integration system with deep data interaction. Specifically, in the data acquisition and preprocessing stage, the collaborative mechanism between edge computing nodes and the Apache NiFi data flow pipeline enables real-time dynamic classification of equipment operation data, financial cost data, and sales forecast data. Compared to the static mapping rules that rely on manual configuration in existing technologies, this method's dynamic classification strategy based on a business rule engine can adapt to changes in data format, significantly reducing data cleaning costs.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A business and finance integration method based on data deep interaction, characterized in that, Comprising the following steps: S1, multi-source heterogeneous data collection and dynamic preprocessing: Real-time collection of equipment operation data, financial cost data and sales forecast data through edge computing nodes deployed in the production site, construction of data flow pipeline using Apache NiFi, dynamic classification of sensor time series data, ERP structured data and market unstructured data based on business rule engine; DeltaLake is used to build a hierarchical storage architecture, ACID transaction control is implemented for structured data, and key financial indicators are extracted from pre-trained natural language processing models for unstructured data; S2, multi-dimensional feature interaction modeling: Temporal and spatial feature extraction branch: based on equipment operation data, graph neural network is constructed to mine the correlation topological features between production equipment, and device-level health index is generated through graph pooling operation; Time series prediction branch: a Transformer model with self-attention mechanism is established for financial data to capture the long-term and short-term dependence relationship between cash flow and cost; Cross-modal fusion branch: a gated recurrent unit is used to dynamically allocate weights to business features and financial features to generate a fusion feature matrix, and the weight allocation strategy is based on financial indicator sensitivity analysis results; S3, closed-loop decision optimization execution: Based on the fusion feature matrix, combined with sales forecast data and inventory cost data, multiple scenario simulation schemes are generated through Monte Carlo tree search; The optimization results are sent to the production execution system through the MQTT protocol, and a decision effect feedback link is established to dynamically adjust the model parameters according to the actual execution results, forming a continuous improvement closed loop; S4, full-link security control: Through the security control layer, a data interaction channel based on zero trust architecture is constructed, dynamic access control is implemented for the fusion process of business data and financial data, and blockchain technology is used for non-tamperable evidence of key decision nodes, and the evidence content includes decision context hash value.

2. The business and finance integration method based on data depth interaction according to claim 1, characterized in that: The business rule engine further comprises: Dynamic rule configuration module, which supports receiving subject mapping rules defined by the financial department through RESTful API and updating the local rule library of the edge computing node in real time; Conflict detection submodule, when semantic conflicts occur between sensor data and ERP data at the field level, automatically trigger the manual arbitration process and store the arbitration results in the rule library as new entries; Rule version management unit, a blockchain evidence chain is established for historical rule change records to ensure that the rule can be traced back to the specific business department approval flow.

3. The business and finance integration method based on data depth interaction according to claim 1, characterized in that: In the cross-modal fusion branch: The training process of the dynamic weight allocation mechanism adopts a progressive learning strategy, using historical financial audit reports as supervision signals in the initial training stage, and switching to real-time business feedback data after the model converges; The joint loss function includes financial indicator sensitivity analysis items and device downtime probability items, and the sensitivity analysis items determine the weight coefficient by calculating the marginal effect of financial subject fluctuations on net profit; After the fusion feature matrix is generated, it needs to pass through the integrity verification module of the security control layer to verify that the feature vector has not been tampered with using the SM3 hash algorithm.

4. The business and finance integration method based on data depth interaction according to claim 1, characterized in that: In the closed-loop decision optimization execution: The Monte Carlo tree search algorithm adopts a two-stage node expansion strategy. In the initial stage, a fast traversal is performed based on a historical scenario library. When the preset decision depth threshold is reached, the algorithm switches to real-time data-driven simulation verification. A cost sensitivity coefficient is introduced when generating the multi-scenario simulation scheme. The coefficient is dynamically adjusted by the marginal contribution rate fed back by the financial module, ensuring that the simulation times of high-cost scenarios account for no less than 40% of the total simulation amount. Before the optimization results are issued, they need to pass through the interface compatibility verification module of the production execution system. The verification content includes device instruction set version, communication protocol encryption level, and abnormal interruption recovery strategy.

5. The business and finance integration method based on data depth interaction according to claim 1, characterized in that: In the closed-loop decision optimization and execution: A sliding time window mechanism is used to build the decision effect feedback link. The window length is set to the last 3 complete production cycles. The model parameters are only solidified when the decision deviation rate of the last 2 windows is below the preset threshold. Gradient clipping protection is implemented during the model parameter adjustment process. When the parameter update amount exceeds 3 times the standard deviation of the historical mean, it automatically rolls back to the latest valid parameter version and generates an alarm event. The deviation analysis between actual execution results and simulation schemes includes three-dimensional evaluation indicators: financial indicator deviation rate, device utilization rate fluctuation value, and order delivery punctuality rate change.

6. The business and finance integration method based on data depth interaction according to claim 1, characterized in that: In the full-link security control: The zero-trust architecture implementation includes a device fingerprint authentication sub-module, which generates a unique device identifier by collecting the hardware serial number, BIOS version, and startup log of the edge computing node. The dynamic access control policy uses an attribute-based access control model. When a financial data access request is detected, the system verifies the operator's role and the permission mapping relationship with the current business scenario. The blockchain storage node deployment uses a hybrid architecture. Critical decision nodes are stored in a consortium chain, and ordinary operation logs are stored in a private chain. The stored data includes operation timestamp, data hash value, and digital signature certificate serial number.

7. A business and finance integration system based on data deep interaction, adopting the business and finance integration method based on data deep interaction according to any one of claims 1-6. It includes: Edge computing layer: Deploy lightweight feature extraction models. Built-in data caching strategy based on business priority. Automatically store critical device data in local time series database when network is interrupted. Database uses columnar storage format to optimize query performance. Cloud management layer: Includes intelligent fusion analysis model training platform. Supports configuring the number of graph neural network layers and attention heads for spatiotemporal feature extraction branches through Web interface. Provides model performance monitoring dashboard. Security control layer: Integrates hardware security modules to perform encrypted calculations on feature matrices during cross-modal fusion. Key management follows NIST SP800-90A standards, supporting key rotation strategy configuration. Human-machine collaboration layer: Provides configurable decision dashboards. Supports establishing correlation analysis views of financial indicators and production parameters through drag-and-drop operations. Built-in adaptive early warning rule template library based on historical data.

8. The industry-finance integration system based on data deep interaction according to claim 7, characterized in that: The local time series database of the edge computing layer: Uses columnar storage format to optimize query performance. Implements segmented compression on time series fields of device vibration data and temperature data. The compression algorithm includes a hybrid mode of run-length encoding and dictionary encoding. Run-length encoding is used for continuous repeated values, and dictionary mapping is used for discrete symbol values. When the network is restored, the data synchronization module automatically checks the timestamps of the local cache and the cloud data, and only uploads the incremental difference data.

9. The industry-finance integration system based on data deep interaction according to claim 7, characterized in that: The adaptive early warning rule template library of the man-machine collaboration layer: The threshold generator based on time series decomposition adopts the STL decomposition algorithm to automatically identify the trend item, seasonal item and residual item in the data. When the early warning boundary condition is updated, the OCR technology is used to identify the paper documents signed by the chief financial officer and convert them into structured rules and store them in the template library. After the early warning is triggered, the system automatically generates a draft of the disposal plan containing the abnormal data segment, the possible affected subjects and the suggested countermeasures.

10. A data deep interaction based industry and finance integration storage medium based on the data deep interaction based industry and finance integration system according to any one of claims 7-9, characterized in that, It includes: The storage medium adopts a hierarchical storage architecture and contains a read-only storage area and a writable storage area. The read-only storage area solidifies the pre-trained cross-modal fusion model parameter file and the business rule engine configuration template. The writable storage area supports dynamic loading of production site data fingerprint library and financial subject mapping table. The storage medium is equipped with a hardware-level encryption chip, which implements transparent encryption of the cross-modal fusion model parameters. The key management complies with the GM / T0028-2014 standard and supports one-machine-one-key dynamic distribution. The storage medium supports remote updating of the business rule engine configuration template through OTA. The update package uses asymmetric encryption signature, and only the firmware upgrade issued by the authorized server is allowed. The built-in model version management module can roll back to the historical version parameter file. It provides multiple format data interfaces, supports OPCUA, MQTT and RESTfulAPI three kinds of protocol data read and write, and the interface parameters are dynamically bound through configuration file.

Images