Water affair multi-domain data dynamic cooperative processing method based on business scene driving

Abstract

The invention relates to a water affair multi-domain data dynamic cooperative processing method based on business scene driving, and belongs to the technical field of water affair data processing. Constructing a scene feature library and a scene-model mapping relation library of water affairs, performing scene type automatic identification based on the obtained terminal data, matching a correlation model, judging a priority and generating a collection strategy; a collection strategy is issued to each monitoring terminal through an edge gateway, a networking architecture of'center node-multi-domain federated node 'is constructed, a center node initiates a federated task and is connected with each federated node through an encrypted communication link, the federated nodes train a related local model based on local original data, and the local data is transmitted to the monitoring terminals through the edge gateway. Local parameters are calculated through real-time data and uploaded to a center node, the center node adopts a weighted average method to generate a global model in an aggregation mode, and a risk level is judged to generate a comprehensive strategy; according to the method, efficient cooperative processing of water affair multi-domain data is realized through a whole-process dynamic mechanism of scene identification-federated cooperation-linkage feedback.

Description

Technical Field

[0001] This invention relates to a dynamic collaborative processing method for multi-domain water affairs data driven by business scenarios, belonging to the field of water affairs data processing technology. Background Technology

[0002] With the continuous expansion of water systems, the volume of data is growing explosively, leading to increasingly higher demands on data processing and transmission. Existing water data processing technologies suffer from three major pain points: (1) Problem of rigid scenario adaptation: Traditional water data processing adopts the "fixed process + unified parameters" mode (such as collecting data at a frequency of 15 minutes / time and using the same analysis model, whether it is flood control emergency or daily monitoring). When the business scenario changes (such as a sudden rainstorm triggering a flood control emergency), the collection frequency and model parameters need to be manually reconfigured, and the response lag is ≥2 hours, which causes the emergency response to miss the best time and cannot adapt to the dynamic needs of smart water business. (2) Multi-domain data fragmentation problem: Water affairs data such as pollution control, flood control, and water supply belong to different management entities such as environmental protection bureaus, district and county water affairs stations, and water plants, and are stored in independent systems (e.g., pollution control data is stored in the environmental protection bureau's monitoring platform, and water supply data is stored in the water affairs bureau's pipeline system). The data is not interconnected and cannot be linked for analysis, which makes it impossible to identify cross-domain related risks such as "excessive pollution discharge affects the water supply of downstream water sources" and "rainstorms and floods cause water supply pipeline rupture". The analysis results are one-sided and cannot support overall decision-making. (3) Lack of data privacy protection: When cross-domain data collaboration is required, the original data needs to be transmitted (such as districts and counties reporting complete sewage monitoring data and residents' water consumption data to the city level). There is a risk of leakage of sensitive sewage information of enterprises and residents' privacy data, which does not meet the requirements of the Data Security Law and the Personal Information Protection Law for data security and restricts the collaborative application of multi-domain data. Summary of the Invention

[0003] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a dynamic collaborative processing method for multi-domain water affairs data based on business scenarios. Through a dynamic mechanism of "scenario recognition - federated collaboration - linkage feedback" throughout the entire process, efficient collaborative processing of multi-domain water affairs data can be achieved.

[0004] The technical solution adopted in this invention is as follows: A dynamic collaborative processing method for multi-domain water data based on business scenarios includes the following steps: S1. Construct a scene feature library and a scene-model mapping relationship library for water affairs. Based on the acquired terminal data, call the scene feature library to automatically identify the scene type. Based on the scene type, call the scene-model mapping relationship library to match and associate models to determine priority and generate a collection strategy. S2. The data collection strategy is distributed to each monitoring terminal through the edge gateway, dynamically adjusting the collection frequency and monitoring indicators of the monitoring terminal. The edge gateway preprocesses the data collected by the terminal, and the standardized data after preprocessing is stored in the in-memory database for subsequent federated analysis. The raw data collected is stored in the water data lake based on HDFS for subsequent traceability. S3. A network architecture of "central node - multi-domain federation node" is jointly constructed with relevant business nodes as the core. A key business node is set as the central node, and other relevant business nodes are federation nodes. The central node initiates federation tasks and connects each federation node through encrypted communication links. The federation nodes train relevant local models based on local raw data and calculate local parameters using real-time data. The local parameters are encrypted and uploaded to the central node. The transmission process only contains parameters and does not contain local raw data. The central node uses a weighted average method to aggregate and generate a global model, realize cross-domain data correlation analysis, determine risk level and generate comprehensive strategies. S4. Based on the comprehensive strategy obtained from the global model, generate customized tasks for each role and push them out. Each role provides real-time feedback on task handling and effects. The global model is recalculated and the strategy is adjusted until the risk is eliminated.

[0005] The scenario feature library described in step S1 of the above method collects and stores key identification information of core water affairs business scenarios, specifically including scenario unique ID, scenario name, trigger condition set, and associated business domain. The scenario-model mapping relationship library stores the precise correspondence information between scenarios and dedicated analysis models, specifically including mapping ID, scenario unique ID, model unique identifier, model type, model core parameter range, and model adaptation scenario priority.

[0006] The principle for prioritizing scenarios is based on a comprehensive assessment of three dimensions: urgency, scope of impact, and risk of loss. Specifically: Level 1: Emergency Response (e.g., flood control emergency, pollution accident response): sudden events affecting ≥3 areas or potentially causing significant personal injury or property damage; Level 2: Important Regulatory Scenario (e.g., routine water quality monitoring of water sources, abnormal pipeline pressure early warning): affecting specific key areas or potentially causing substantial economic losses; Level 3: Routine Scenario (e.g., daily water conservation analysis, general area flow statistics): no urgent risk or only used for data statistics and archiving. After prioritization, computing power and storage resources are allocated preferentially to Level 1 and Level 2 scenarios. 70% of system computing power and storage resources are allocated to Level 1 scenarios, 50% to Level 2 scenarios, and the remaining resources are allocated to Level 3 scenarios to ensure rapid response for high-priority scenarios.

[0007] In step S2, the policy is distributed to each monitoring terminal through the edge gateway. The edge gateway completes the terminal parameter configuration and feeds back the configuration results to the node platform.

[0008] In step S3, the central node is the water management node, and the federated nodes include the pollution control management node, the water supply management node, and the meteorological management node; the edge gateway is only responsible for the distribution of collection strategies, data preprocessing and distribution, and does not participate in the networking of federated nodes.

[0009] The encryption described in step S3 is SM4 national cryptographic algorithm encryption.

[0010] In step S3, the local model is a submodule of the global model. The input data format (such as data dimension, field type, timestamp precision) is consistent with the global model, and the output local parameters (gradient value, weight coefficient) dimension is completely matched with the global model parameter dimension, ensuring that the central node can be directly aggregated. The training of the local model of each federation node needs to be based on the local raw collected data to build a basic framework, and then call the real-time standardized data in the memory database for parameter calibration. The process of generating a global model using the weighted average method: ① The central node sets weight coefficients based on the data quality (data integrity ≥95% is considered qualified, and the higher the percentage of qualified data, the higher the weight) and business importance (pollution control / water supply nodes have higher weights than meteorological nodes) of each federation node. (n is the number of federated nodes), and satisfies (Example: Pollution control management node) Water supply management nodes Meteorological management nodes ); ② The central node receives local parameters uploaded by each federation node. ③ Calculate the global parameter P using the formula: .

[0011] The beneficial effects of this invention are: (1) This invention adopts a dual-database linkage of scene feature library and scene-model mapping relationship library, which breaks through to realize full-process automation of "scene recognition → priority determination → collection strategy generation → terminal configuration". In response to the core contradiction between the dynamic changes of water business scenarios and the rigidity of traditional data processing process, by pre-setting multiple scenario trigger rules and model mapping relationships, the collection frequency, monitoring indicators and analysis model parameters can be automatically adjusted according to different business needs such as flood control emergency and pollution control linkage, completely getting rid of the dependence on fixed process; the data collection strategy adjustment time when switching scenarios is shortened from 2 hours to 10 seconds, and the emergency scenario handling efficiency is improved by 720%, ensuring that emergency business such as flood control and pollution control is "discovered early and handled early"; (2) This invention takes the "central node - multi-domain federated node" distributed collaborative architecture as the core, combined with the SM4 national cryptographic encryption transmission link, to realize the innovative collaborative mode of "local retention of original data and cross-domain transmission of model parameters". It focuses on the industry pain points of fragmented water affairs data and privacy protection among multiple entities. Without leaking enterprise sewage discharge data and residents' water privacy, it completes cross-domain data correlation analysis, which not only breaks down data silos, but also fully complies with the privacy protection requirements of the Data Security Law. It breaks down multi-domain data barriers, improves the cross-domain risk identification rate, and avoids hidden risks such as "excessive sewage discharge affecting water supply" and "waterlogging causing pipeline rupture". Through federated learning, it realizes "data does not move, model moves", and each domain only transmits model parameters (not original data). Combined with the SM4 national cryptographic encryption link, the risk of original data leakage is reduced to 0, which complies with the requirements of national data security regulations. (3) This invention positions the edge gateway as the core of “terminal-platform” collaboration, and innovatively realizes the integrated function of “rapid policy instruction issuance + localized data preprocessing + dual-path data distribution”. In response to the pain points of high platform transmission pressure and data redundancy in traditional solutions, the edge gateway completes preprocessing operations such as null value completion and format correction, and stores real-time data in memory database and raw data in HDFS data lake. This not only ensures the millisecond-level response of core business, but also reduces the transmission and storage pressure of the backend platform. The computing power and storage resources are dynamically allocated according to the priority of the scenario (e.g., 70% of the computing power is used in the first-level scenario and 30% in the third-level scenario), avoiding the waste of resources in the traditional “full load operation of all scenarios”. (4) By constructing a closed-loop mechanism of "global model analysis → role-based task customization → handling feedback → strategy iteration", multiple roles such as law enforcement personnel, water plant operation and maintenance, and residents are included in the handling process. In response to the problem of "disconnect between analysis and handling" in traditional data processing, differentiated tasks (such as law enforcement work orders, operation and maintenance instructions, and resident reminders) can be automatically generated based on cross-domain risk levels. Data collection and analysis strategies can be dynamically adjusted through real-time feedback on handling effects, realizing full automation from risk identification to elimination and significantly improving the task completion rate. The entire process from risk identification to task push to effect feedback does not require manual intervention, reducing the workload of operation and maintenance personnel and avoiding human error. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0013] The present invention will be further described below with reference to specific embodiments.

[0014] Example 1 is based on the scenario of "a sudden exceedance of sewage discharge standards in a certain city triggering a 'pollution control-water supply linkage scenario'".

[0015] A dynamic collaborative processing method for multi-domain water data based on business scenarios includes the following steps: S1. Construct a scene feature library and a scene-model mapping relationship library for water affairs. Based on the acquired terminal data, call the scene feature library to automatically identify the scene type, and call the scene-model mapping relationship library based on the scene type to match and associate models, determine priorities, and generate a collection strategy: The scenario feature library collects and stores key identification information for core water business scenarios, including a unique scenario ID, scenario name (such as flood control emergency scenario, pollution control-water supply linkage scenario, and daily water conservation monitoring scenario), trigger condition set (multi-dimensional combination of conditions, such as "rainfall ≥ 50 mm / h and water level ≤ 0.5 m below warning value and there is a flood-prone area downstream" or "sewage outlet COD value exceeds the threshold by 3 times and there is a drinking water source within 2 km downstream"), and associated business domains (such as flood control domain, pollution control domain + water supply domain, and water conservation domain).

[0016] The scenario-model mapping relationship library collects and stores precise correspondence information between scenarios and dedicated analysis models, including mapping ID, scenario unique ID, model unique identifier, model type (such as rainstorm flooding diffusion prediction model, pollutant migration-water supply risk correlation model, water conservation analysis model), model core parameter range, and model adaptation scenario priority.

[0017] The municipal water affairs big data platform (the core hub platform, responsible for the integration, decision-making and scheduling of data throughout the entire process) obtains two types of data in real time through the API interface: ① COD data uploaded by the sewage outlet monitoring terminal (transmitted via MQTT protocol) (current value 180mg / L, exceeding the preset threshold (60mg / L) by 3 times); ② water source location data issued by the geographic information platform (there is an urban drinking water source 2km downstream of the sewage outlet).

[0018] The municipal water affairs big data platform calls the "scenario feature library" (which stores preset trigger conditions) to compare and determine: it meets the rule of "sewage outlet COD exceeds the threshold and there is a water source downstream → pollution control-water supply linkage scenario", and the priority of this scenario is "level one (emergency scenario)", which occupies 70% of the system's computing resources.

[0019] The municipal water affairs big data platform prioritizes scenarios, calls the "scenario-model mapping relationship library," matches the "pollutant migration-water supply risk association model," and generates the following collection strategies: ① Sewage outlet monitoring terminal: the collection frequency is adjusted from 5 minutes / time to 30 seconds / time, and a "ammonia nitrogen concentration" collection item is added; ② Water source monitoring terminal: the collection frequency is adjusted from 10 minutes / time to 1 minute / time, and a backup water quality sensor is enabled; ③ Meteorological terminal: real-time acquisition of wind direction and rainfall data (1 time / 5 minutes).

[0020] S2. The data collection strategy is distributed to each monitoring terminal via the edge gateway, dynamically adjusting the collection frequency and monitoring indicators of the terminals. The edge gateway preprocesses the data collected by the terminals, and the standardized preprocessed data is stored in an in-memory database for subsequent federated analysis. The raw data collected is stored in a water data lake based on HDFS for subsequent traceability. The data collection strategy is distributed to each monitoring terminal through an edge gateway (deployed around sewage outlets and water sources). The edge gateway completes the terminal parameter configuration within 10 seconds and feeds back the configuration results to the municipal water affairs big data platform (the feedback includes configuration status (success / failure), terminal unique ID, adjusted collection parameters (frequency / indicator), and feedback time).

[0021] The edge gateway preprocesses the data collected by the terminals (including null value completion, format correction, invalid and garbled character removal, and analog signal to digital signal conversion). Standardized data that meets the real-time analysis requirements after preprocessing is stored in an in-memory database for subsequent federated analysis. The raw data collected by the terminals that has been initially screened by the edge gateway but has not undergone in-depth processing (i.e., raw data streams directly output by terminals such as sensors and video surveillance, including structured numerical data and unstructured video / log data) is stored in a water data lake based on HDFS for subsequent data traceability and review.

[0022] S3. A network architecture of "central node - multi-domain federated nodes" is jointly constructed with relevant business nodes as the core. A key business node is set as the central node, and other relevant business nodes are federated nodes. The central node initiates federated tasks and connects each federated node through encrypted communication links. The federated nodes train relevant local models based on local raw data and calculate local parameters using real-time data. The local parameters are then encrypted and uploaded to the central node. The transmission process only contains the parameters and does not include the local raw data. The central node uses a weighted average method to aggregate and generate a global model, realize cross-domain data correlation analysis, determine risk levels, and generate comprehensive strategies. The central node is the municipal water affairs management node (deployed based on the municipal water affairs big data platform), and the federated nodes include pollution control management nodes (deployed by the Environmental Protection Bureau), water supply management nodes (deployed by water plants), and meteorological management nodes (deployed by the Meteorological Bureau). Using the municipal water affairs bureau platform as the "central node," federated tasks are automatically initiated. Through a communication link encrypted with the SM4 national cryptographic algorithm, three types of "federated nodes" are connected: ① Pollution control management node (deployed by the Environmental Protection Bureau, storing historical COD data from sewage outlets and enterprise pollution discharge registration information); ② Water supply management node (deployed by water plants, storing water source quality standards and water supply network pressure data); ③ Meteorological management node (deployed by the Meteorological Bureau, storing rainfall forecast data for the next 6 hours).

[0023] (1) Local model training and parameter determination Each federated node trains a local model based on local data, and the output local parameters (gradient values, weight coefficients) follow these principles for determination: Weighting coefficients are determined based on a dual dimension of "data quality + business relevance." Data quality is assessed using "data integrity ≥ 95%" as the passing standard; the higher the percentage of qualified data, the higher the weight. Business relevance is determined by "the degree of core relevance to the current scenario (pollution control-water supply linkage)," with pollution control and water supply nodes having a higher relevance than meteorological nodes. In this embodiment, the weighting coefficient is set as follows: for pollution control management nodes... Water supply management nodes Meteorological management nodes (satisfy ); Gradient value: Calculated based on the goodness of fit of the local model during training. The higher the goodness of fit (i.e., the better the model's prediction matches the local historical data), the closer the gradient value is to 1, indicating stronger reliability of the local parameters. The gradient is calculated using the gradient descent algorithm during model training. First, model parameters are initialized based on local historical data, and a loss function is constructed. The partial derivatives of the loss function with respect to the model parameters are iteratively calculated using gradient descent. In each iteration, the parameters are adjusted to reduce the loss. When the loss function value falls below a preset threshold, iteration stops, and the output gradient value at this point is the final gradient of the current local model.

[0024] Specific training and parameter output for each federation node: ① Pollution control management node: Based on the local sewage outlet's historical COD data for the past year and enterprise sewage discharge registration information, a "pollutant diffusion rate model" is trained with a data fit of 97%, outputting local parameters (gradient value: 0.82, weight coefficient: 0.35), calculating that "the sewage outlet with excessive COD will diffuse to the downstream water source within 4 hours"; ② Water supply management node: Based on the local water source's water quality tolerance standard and historical water supply network pressure data, a "water source pollution impact model" is trained with a data fit of 96%, outputting local parameters (gradient value: 0.65, weight coefficient: 0.35), calculating that "COD values ​​exceeding 80 mg / L will lead to the shutdown of the water source"; ③ Meteorological management node: Based on the local 6-hour rainfall forecast data, a "rainfall dilution impact model" is trained with a data fit of 92%, outputting local parameters (gradient value: 0.12, weight coefficient: 0.3), calculating that "if there is no rainfall in the next 6 hours, pollutant diffusion will not be diluted by rainwater".

[0025] Each node uploads its local parameters to the central node via an SM4 encrypted link. The transmission process only includes the parameters and does not include the original data.

[0026] (2) Weighted average method to aggregate and generate global model The central node uses a weighted average method to aggregate local parameters. The specific process and calculation formula are as follows: Parameter validation: First, validate the consistency of the local parameters uploaded by each federated node (ensure that the gradient values ​​and weight coefficients are consistent and can be directly used in the calculation). Weighted calculation: Based on preset weight coefficients, global parameters (gradient value P, weight coefficient Ω) are calculated using a formula: Global gradient values: ; Global weight coefficients: ; Model Construction: A "pollution control-water supply linkage global model" was constructed based on global parameters (P=0.5505, Ω=0.335). Substituting the global parameters (gradient P=0.5505, weight Ω=0.335) into the core formula of the "pollutant migration-water supply risk correlation model," for example: Predicted COD value of the water source = Current COD value of the sewage outlet × Diffusion coefficient × Global gradient P × Global weight Ω + Initial COD value of the water source. Substituting real-time data, the predicted COD value of the water source is calculated and compared with the water source's safety threshold to determine the risk level. Combining this with the "risk level - impact range" mapping rule in the global model, the final conclusion is reached, and a comprehensive strategy is output: ① Risk level: Red (the COD of the water source will rise to 92 mg / L after 4 hours, exceeding the safety threshold); ② Impact range: Water supply to three downstream residential communities will decrease by 30%; ③ Key recommendations: Switch to a backup water source in advance and implement emergency sealing of the sewage outlet.

[0027] S4. Based on the comprehensive strategy obtained from the global model, generate customized tasks for each role and push them out. Each role provides real-time feedback on task handling and effects. The global model is recalculated and the strategy is adjusted until the risk is eliminated. Customized task push for multiple roles: Based on the conclusions of the global model, the platform generates customized tasks for three types of roles: ① Law enforcement personnel: The task of "going to the sewage outlet for on-site law enforcement within 1 hour" is pushed through the mobile APP, with the COD exceeding the standard evidence chain (not the original data, only including the period of exceeding the standard and the numerical trend graph) and the coordinates of the sewage outlet; ② Water plant operation and maintenance: The instruction of "switching to the backup water source within 3 hours" is pushed through the SCADA system, with the water quality prediction curve of the water source attached; ③ Residents: The reminder of "the water quality in some areas may fluctuate in the next 6 hours, it is recommended to store water" is pushed through SMS, with the water service hotline attached.

[0028] Real-time feedback on the effectiveness of the measures and adjustments to the strategy: ① After on-site rectification by law enforcement personnel, photos of the "sewage outlet sealing" are uploaded via the APP (stored in OSS object storage). After the platform obtains the photos, it instructs the sewage outlet monitoring terminal to restore the collection frequency to 5 minutes / time; ② The water plant completes the switch to the backup water source and sends back a "successful switch" signal. The platform instructs the water source monitoring terminal to restore the collection frequency to 10 minutes / time; ③ Two hours later, the global model is recalculated: the COD of the sewage outlet drops to 45mg / L, the COD of the water source remains at 38mg / L, the risk level drops to "green (safe)," and the linkage task is automatically terminated.

[0029] The above is a further description of the present invention in conjunction with specific embodiments, and the scope of protection of the present invention is not limited thereto.

Claims

1. A dynamic collaborative processing method for multi-domain water data based on business scenarios, characterized in that, The steps include the following: S1. Construct a scene feature library and a scene-model mapping relationship library for water affairs. Based on the acquired terminal data, call the scene feature library to automatically identify the scene type. Based on the scene type, call the scene-model mapping relationship library to match and associate models to determine priority and generate a collection strategy. S2. The data collection strategy is distributed to each monitoring terminal through the edge gateway, dynamically adjusting the collection frequency and monitoring indicators of the monitoring terminal. The edge gateway preprocesses the data collected by the terminal, and the standardized data after preprocessing is stored in the in-memory database for subsequent federated analysis. The raw data collected is stored in the water data lake based on HDFS for subsequent traceability. S3. A network architecture of "central node - multi-domain federation node" is jointly constructed with relevant business nodes as the core. A key business node is set as the central node, and other relevant business nodes are federation nodes. The central node initiates federation tasks and connects each federation node through encrypted communication links. The federation nodes train relevant local models based on local raw data and calculate local parameters using real-time data. The local parameters are encrypted and uploaded to the central node. The transmission process only contains parameters and does not contain local raw data. The central node uses a weighted average method to aggregate and generate a global model, realize cross-domain data correlation analysis, determine risk level and generate comprehensive strategies. S4. Based on the comprehensive strategy obtained from the global model, generate customized tasks for each role and push them out. Each role provides real-time feedback on task handling and effects. The global model is recalculated and the strategy is adjusted until the risk is eliminated.

2. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, The scenario feature library mentioned in step S1 collects and stores key identification information of core water business scenarios, specifically including scenario unique ID, scenario name, trigger condition set, and associated business domain.

3. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, The scenario-model mapping relationship library mentioned in step S1 stores the precise correspondence information between scenarios and dedicated analysis models, specifically including mapping ID, scenario unique ID, model unique identifier, model type, model core parameter range, and model adaptation scenario priority.

4. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, The principle for determining the priority in step S1 is based on a comprehensive assessment of three dimensions: urgency, scope of impact, and risk of loss. Specifically, the scenarios are categorized as follows: Level 1: Emergency Scenarios; Level 2: Important Regulatory Scenarios; and Level 3: Routine Scenarios.

5. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, Once the priorities are determined, computing power and storage resources are allocated preferentially to first- and second-level scenarios.

6. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, In step S2, the policy is distributed to each monitoring terminal through the edge gateway. The edge gateway completes the terminal parameter configuration and feeds back the configuration results to the node platform.

7. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, In step S3, the central node is the water management node, and the federated nodes include the pollution control management node, the water supply management node, and the meteorological management node. The edge gateway is only responsible for issuing collection policies, preprocessing and distributing data, and does not participate in the networking of federated nodes.

8. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, The encryption described in step S3 is SM4 national cryptographic algorithm encryption.

9. The method for dynamic collaborative processing of multi-domain water data based on business scenario driving according to claim 1, characterized in that, In step S3, the local model is a sub-module of the global model. The input data format is consistent with that of the global model, and the output local parameter dimensions are completely matched with the global model parameter dimensions, ensuring that the central node can be directly aggregated. The training of the local model of each federation node needs to be based on the local raw data to build a basic framework, and then call the real-time standardized data in the memory database for parameter calibration. The process of generating a global model using the weighted average method: ① The central node sets weight coefficients based on the data quality and business importance of each federated node. Let n be the number of federated nodes, and satisfy the condition that n is the number of federated nodes. ② The central node receives local parameters uploaded by each federation node. ③ Calculate global parameters using formulas .

Images