Intelligent data management system and method suitable for central control system

Abstract

The invention discloses a data intelligent management system and method suitable for a central control system, relates to the technical field of data processing, and determines dynamic differentiation data management period information based on a hardware, data and business three-layer collaborative model in combination with acquired control parameters. According to the intelligent data management system and method suitable for the central control system, the matching relation between the hardware resource state and the process stage is analyzed through the hardware, data and service three-layer collaborative model, the management period of non-core data is automatically prolonged in the high-load stage of the system, the core data period is shortened in the low-load stage of the system, and the management efficiency is improved. Meanwhile, exclusive hardware resources are distributed for data with different priorities through a dynamic resource scheduling algorithm; therefore, resource waste caused by a fixed period and uniform resource allocation is avoided; secondly, mining precursor features in rare fault data by adopting an activation reasoning causal chain algorithm, and meanwhile, ensuring judgment accuracy in combination with a double-hardware channel secondary analysis mechanism;

Description

Technical Field

[0001] This invention relates to the field of data processing technology for central control systems, and in particular to a data intelligent management system and method suitable for central control systems. Background Technology

[0002] The central control system is the core hub for industrial production and energy management. It needs to process four types of key data in real time: system operation status data, user operation behavior data, equipment interaction response data, and system log record data. The quality of data management directly determines the system stability and production safety.

[0003] Current mainstream solutions adopt an architecture of fixed-cycle data collection, universal quality assessment, and independent operation of each stage. For example, a unified collection cycle is set, and four types of key data are collected through conventional industrial protocols. The collection process is not related to the status of hardware resources such as server CPU load and memory usage. Quality assessment adopts universal standards, and all types of data are evaluated according to the same requirements for completeness and accuracy, and it relies on correlation analysis that requires a large number of historical failure samples. The collected data is directly transmitted to the core database for storage. The analysis process uses a fixed CPU core and memory quota, and does not adjust the allocation of hardware resources according to the process stage. Hardware scheduling is completely independent of data processing and process requirements. However, during the implementation of the above technical solution, at least the following technical problems were discovered: First, fixed management cycles lead to resource mismatch and data failure. During periods of high system load, a unified data collection cycle can cause data collection delays exceeding reasonable thresholds due to insufficient hardware resources, while during periods of low load, hardware resources remain idle. Second, infrequent fault data assessments result in delayed early warnings. Due to limited historical sample sizes, existing solutions cannot uncover early warning characteristics for low-frequency faults, leading to early warning delays that exceed industrial safety requirements and hindering risk mitigation. Third, there is a disconnect between general quality benchmarks and industry process requirements. Unified standards impose excessively low requirements on core data and excessively high requirements on auxiliary data, posing risks of process misjudgment and wasting storage resources. To address these issues, we propose a data intelligent management system and method suitable for central control systems. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a data intelligent management system and method suitable for central control systems, solving the technical problems of fixed management cycles, failure in evaluating scarce fault data, and insufficient universality of quality benchmarks in existing data management systems for central control systems.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A data intelligent management method suitable for central control systems, comprising the following steps: Acquire control parameters from the central control system, covering multiple sets of source data, hardware resource status data, and industry process characteristic data; Based on a three-layer collaborative model of hardware, data, and business, and combined with the acquired control parameters, dynamic and differentiated data management cycle information is determined; the model is generated through joint training of historical industry fault data and hardware load data. Based on the dynamic differentiated data management cycle information, a data sample set is generated. The hardware processing capacity is adapted to the current and preset time period through a hardware resource pre-scheduling algorithm. The input of the scheduling algorithm includes the current hardware resource status data and industry process characteristic data. Based on an industry-specific dynamic quality benchmark library, the data sample set is integrated and analyzed to obtain data management quality assessment values; the industry-specific dynamic quality benchmark library has built-in industry process templates and is updated in real time with the system's running data through incremental learning; The activation inference causal chain algorithm is used to calculate the data management judgment value by combining the data management quality assessment value and scarce fault data; the activation inference causal chain algorithm extracts fault precursor features and models causal relationships. The system determines whether the data management judgment value is greater than the preset industry-specific management threshold. If it is, the data is stored in the core database of the central control system in a hardware optimization format, and a hardware resource optimization plan is generated. Otherwise, invalid data is removed and a quality optimization report is generated.

[0006] Preferably, the steps for acquiring data specifically include: The system synchronously acquires source data, hardware resource status data, and industry process characteristic data through the acquisition interface. The hardware resource status data includes processor utilization, memory usage, and storage I / O rate; the industry process characteristic data consists of parameters specific to the corresponding field; each set of source data includes system operating status, user operation behavior, equipment interaction response, and system log record data. Establish a three-dimensional mapping table relating data acquisition, hardware status, and process nodes, record the processor load value, response latency value, and associated process unit corresponding to each type of source data, and back them up regularly; A two-stage filtering mechanism is used to remove abnormal data. The first stage filters data whose processor utilization or response latency exceeds a preset reasonable range during data acquisition. The second stage filters invalid data that is not caused by hardware by verifying the range of process parameters. The effective data after two levels of filtering is associated with a three-dimensional association mapping table, which serves as the basic data for the three-layer collaborative model of hardware, data, and business.

[0007] Preferably, the steps for determining the dynamic differentiated data management cycle information are as follows: The data storage structure, access frequency, and process impact weight are obtained from the system operation status data; the process impact weight is set according to the category differences between core process-related data and auxiliary data. Extract the timestamp information and hardware resource consumption values ​​of key operating parameters from the system's operating status data; the hardware resource consumption values ​​include memory usage, storage bandwidth utilization, and CPU computation time during the data writing process. Based on the process impact weight and hardware resource consumption value, a four-dimensional correlation model is constructed on the basis of the three-layer collaboration model of hardware, data and business, including key parameter index, operation scenario identification, hardware load and process stage; and the first dynamic management cycle corresponding to the system operation status data is set based on the four-dimensional correlation model. A fusion model of LSTM and attention mechanism is adopted to predict the hardware resource usage trend within a preset time period; and the initial management cycle corresponding to user operation behavior, device interaction response and system log data is dynamically calibrated according to the load prediction results. The first dynamic management cycle and other management cycles after calibration are summarized to generate dynamic differentiated data management cycle information. This cycle information includes data acquisition interval, analysis frequency and storage update frequency, and is associated with the corresponding process safety level label. It is synchronized to the hardware resource pre-scheduling algorithm and the industry-specific dynamic quality benchmark library.

[0008] Preferably, the steps for generating the data sample set include: Based on the process priority determined by hardware resource status data and industry process characteristic data, a dynamic resource scheduling algorithm is used to allocate dedicated processing cores, memory quotas and IO bandwidth to each data management task. Among them, the allocation priority of core process data is higher than that of auxiliary data. The system monitors the storage I / O rate in real time, and initiates adaptive compression when the storage I / O rate falls below a preset threshold. Structured data is encoded using an improved dictionary encoding, while unstructured data is compressed using a streaming compression algorithm. An adaptive verification node is set in the data acquisition sequence, and the verification frequency is dynamically adjusted according to the process complexity. If the verification finds that the hardware load exceeds the preset threshold, the acquisition interval is extended and a dynamic resource scheduling algorithm is triggered to reallocate resources. The core process-related sample data is temporarily stored in the high-speed memory area, and the auxiliary data is temporarily stored in the SSD cache area; Integrate various types of data after resource allocation, compression, verification, and caching, and generate a data sample set according to the dynamic and differentiated data management cycle information; the sample set carries resource consumption tags, process association tags, and quality grade tags.

[0009] The preferred integration and analysis steps are as follows: The quality baseline is retrieved from the industry-specific dynamic quality benchmark library and matches the industry process characteristic data. The quality baseline includes quantitative indicators such as completeness, accuracy, and process relevance, and the baseline threshold is updated in combination with industry accident cases. The data sample set is distinguished by data type to obtain structured, semi-structured, and unstructured data results; A multimodal Transformer model was used to perform cross-type correlation analysis on the three types of data results to identify hidden abnormal correlation patterns. A dynamic time warping algorithm is introduced to calculate the continuity evaluation value of data information; if the similarity between the current batch data and the historical qualified batch data is lower than the preset similarity threshold, it is marked as breakpoint data and the breakpoint location and associated process operation are recorded; The weights of each quality dimension were determined using the analytic hierarchy process (AHP). The weights were calculated by weighting the hardware consumption coefficient and the process influence coefficient. Among them, the accuracy of data related to core equipment was given a higher weight than that of non-core equipment. By combining the quality baseline, correlation analysis results, continuity assessment value, and weighting, a data management quality assessment value is calculated, and a three-dimensional quality assessment report is generated.

[0010] Preferably, the steps for calculating the data management judgment value include: By activating inference algorithms to mine precursor features in scarce fault data, and combining them with causal inference models to determine coefficient values; among them, the coefficient of core precursor data is higher than that of indirect correlation data; scarce fault data comes from system log records and industry historical fault databases. The formula for calculating the optimized data quality assessment result is: Data quality assessment result = Data management quality assessment value × Data proportion × (Actual data volume / Average data volume of similar data) × Fault causal correlation coefficient × Process stage correction coefficient, where the data proportion is the proportion of each type of source data in the total data volume, and the process stage correction coefficient is set according to the differences in the categories of process stages. The total amount of calibration data is counted. If the calibration range exceeds the preset range threshold or involves core process parameters, it is automatically marked as a manual review item. After the manual review is completed, the calibration results and review comments are updated to the industry-specific dynamic quality benchmark library. The quality assessment results of various source data are weighted and summed according to their data proportions to obtain the final data management judgment value. If the data management judgment value is within the preset reasonable range of the industry-specific management threshold, a secondary analysis through dual hardware channels is initiated. The main channel recalculates the judgment value using the current hardware resources, while the backup channel calls the redundant computing nodes of the central control system to recalculate the judgment value. If the similarity between the two results reaches the preset threshold, the final judgment value is output. Otherwise, a re-integration analysis is triggered until the similarity between the two results reaches the threshold.

[0011] A data intelligent management system suitable for central control systems, the management system comprising: The acquisition module is used to acquire source data, hardware resource status data, and industry process characteristic data, and to establish a three-dimensional correlation mapping table from data acquisition and hardware status to process nodes. The determination module is used to determine the dynamic and differentiated data management cycle information based on the three-layer collaborative model of hardware, data and business and the collected data; The generation module generates a set of data samples through dynamic resource scheduling and adaptive compression based on the output dynamic differentiated data management cycle information, and also has a hierarchical cache management function; The analysis module, based on an industry-specific dynamic quality benchmark library and sample set, uses a multimodal Transformer model to output data management quality assessment values ​​and three-dimensional quality assessment reports. The calculation module uses the activation inference causal chain algorithm, combined with data management quality assessment value and scarce fault data, to calculate the data management judgment value, and outputs the judgment value for storage / removal decision; The judgment and processing module is used to determine whether the data management judgment value is greater than the industry-specific management threshold, perform storage or rejection operations, and generate a quality optimization report containing process parameter adjustment suggestions.

[0012] Preferably, the data acquisition module specifically includes: Multi-protocol acquisition unit with built-in timestamp synchronization is used to access source data, hardware resource status data and industry process characteristic data; The three-dimensional association unit establishes a three-dimensional association mapping table between data acquisition, hardware status and process nodes and automatically backs it up according to a preset cycle. The two-stage filtering unit first filters overload data based on hardware thresholds, then filters invalid data based on process parameter ranges, and associates valid data with a three-dimensional association mapping table. The process parameter monitoring interface, as a component of the data access node, can directly access and convert industrial sensor signals into analog and digital signals.

[0013] Preferably, the determining module specifically includes: The multi-dimensional information acquisition unit is used to collect information such as storage structure, access frequency, process influence weight, time stamp information of key operating parameters, and hardware resource consumption value from system operation status data. The four-dimensional model building unit, based on the data output by the multi-dimensional information acquisition unit, builds a four-dimensional correlation model of key parameter index, running scenario identification, hardware load, and process stage on the basis of the three-layer collaborative model of hardware, data and business. The first cycle setting unit sets the first dynamic management cycle based on the four-dimensional correlation model; The load prediction unit uses a fusion model of LSTM and attention mechanism to predict the hardware resource usage trend within a preset time period; The periodic calibration unit calibrates the initial management cycle based on the trend results output by the load prediction unit. The cycle integration unit is used to summarize each dynamic management cycle and generate dynamic differentiated data management cycle information; The process safety verification unit is used to perform process safety level label association operations, control the process data collection interval to meet process safety requirements, and add safety labels to periodic information.

[0014] Preferably, the generation module specifically includes: The dynamic resource allocation unit allocates processing cores, memory quotas, and I / O bandwidth based on hardware resource status data and process priorities through a dynamic resource scheduling algorithm. The intelligent compression unit selects the appropriate compression algorithm based on the data type and initiates adaptive compression processing. The adaptive verification unit adjusts the hardware load verification frequency according to the process complexity, triggering the dynamic resource allocation unit to reallocate resources. The tiered caching unit temporarily stores core process-related sample data in a high-speed memory area and auxiliary data in an SSD cache area. The sample integration unit integrates data from each stage to generate a data sample set, and adds resource consumption tags, process association tags, and quality grade tags to the sample set; The IO rate monitoring unit is used to monitor the storage IO rate in real time. When the rate is lower than a preset threshold, the intelligent compression unit is triggered to start compression processing.

[0015] (III) Beneficial Effects 1. By analyzing the matching relationship between hardware resource status and process stage through a three-layer collaborative model of hardware, data, and business, the management cycle of non-core data is automatically extended during high-load phases and shortened during low-load phases. Simultaneously, a dynamic resource scheduling algorithm allocates dedicated hardware resources to data of different priorities. This avoids resource waste caused by fixed cycles and uniform resource allocation. Specifically, when the system is under high load, the extended cycle of non-core data reduces hardware usage, while core data still receives sufficient resources to ensure timely processing. During low load, the shortened cycle of core data enhances data value, and idle hardware resources are utilized efficiently. Secondly, an activation-based inference causal chain algorithm is used to mine precursor features in sparse fault data. Combined with a dual-hardware channel secondary analysis mechanism, the accuracy of judgment is ensured. Therefore, it does not rely on a large number of fault samples; even with a small number of samples of a single type of fault, causal relationship modeling can still capture fault precursors.

[0016] 2. Construct an industry-specific dynamic quality benchmark library, customizing quality baselines for the process characteristics of different fields, and assigning higher quality weights to core equipment data through the analytic hierarchy process (AHP). Core process data is evaluated according to high benchmark requirements to ensure it meets production safety needs; auxiliary data is evaluated according to reasonable benchmarks to avoid excessive storage resource consumption. Furthermore, the benchmark library is updated in real-time with system operation data through incremental learning, adapting to new requirements after process adjustments. Moreover, two-level filtering removes hardware overload and invalid process data to ensure the validity of the original data. A layered caching mechanism is used in the sample generation stage to reduce the risk of data loss during transmission and storage. The judgment calculation stage uses dual hardware channels for secondary analysis to avoid misjudgments caused by single hardware failures. Attached Figure Description

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0018] Figure 1 This is an overall architecture diagram of an embodiment of the present invention; Figure 2 This is an overall flowchart of an embodiment of the present invention. Detailed Implementation

[0019] This application provides a data intelligent management system and method suitable for central control systems. It effectively solves the technical problems of fixed management cycles, ineffective evaluation of scarce fault data, and insufficient universality of quality benchmarks in existing central control system data management systems. By analyzing the matching relationship between hardware resource status and process stages through a three-layer collaborative model of hardware, data, and business, the system automatically extends the management cycle of non-core data during high-load periods and shortens the cycle of core data during low-load periods. Simultaneously, a dynamic resource scheduling algorithm allocates dedicated hardware resources to data of different priorities. This avoids resource waste caused by fixed cycles and uniform resource allocation. Specifically, when the system is under high load, the extended cycle of non-core data reduces hardware usage, while core data still receives sufficient resources to ensure timely processing. During low load, the shortened cycle of core data enhances data value, and idle hardware resources are utilized efficiently. Furthermore, an activation-based inference causal chain algorithm is used to mine precursor features in scarce fault data, combined with a dual-hardware channel secondary analysis mechanism to ensure accuracy. Therefore, it does not rely on a large number of fault samples; even with a small number of samples of a single type of fault, it can still capture fault precursors through causal relationship modeling.

[0020] Example: The technical solution in this application example effectively solves the technical problems of fixed management cycles, ineffective evaluation of scarce fault data, and insufficient universality of quality benchmarks in the data management system of existing central control systems. The overall idea is as follows: To address the problems existing in the prior art, this invention provides a data intelligent management method suitable for central control systems, the steps of which are as follows: S1. Acquire multiple sets of source data, hardware resource status data, and industry process characteristic data from the central control system; among which, hardware resource status data includes processor utilization, memory occupancy, and storage I / O rate; industry process characteristic data are parameters specific to the corresponding field; each set of source data includes system operating status data, user operation behavior data, equipment interaction response data, and system log record data; S2. Based on the hardware, data and business three-layer collaborative model, dynamic differentiated data management cycle information is determined according to the acquired source data, hardware resource status data and industry process characteristic data; the hardware, data and business three-layer collaborative model is jointly trained and generated by industry historical fault data and hardware load data, and the model parameters are optimized according to the preset cycle in combination with new fault cases. S3. Based on the determined dynamic differentiated data management cycle information, generate multiple batches of data sample sets; during the generation process, adapt the hardware processing capabilities of the current period and the preset time period through the hardware resource pre-scheduling algorithm. The input parameters of the hardware resource pre-scheduling algorithm include the current hardware resource status data and industry process characteristic data. S4. Based on an industry-specific dynamic quality benchmark library, multiple batches of data sample sets are integrated and analyzed to obtain data management quality assessment values; the industry-specific dynamic quality benchmark library has built-in multiple industry process adaptation templates, and is updated in real time with the system running data through incremental learning, with the update frequency synchronized with the dynamic differentiated data management cycle; S5. The activation inference causal chain algorithm is used to calculate the data management judgment value by combining the obtained data management quality assessment value and scarce fault data. The scarce fault data is the data in which the sample size of a single type of fault is within a preset small range. The activation inference causal chain algorithm improves the assessment accuracy under scarce data by extracting fault precursor features and modeling causal relationships. Fault precursor features are extracted from system log data and device interaction response data. S6. Determine whether the obtained data management judgment value is greater than the industry-specific management threshold; the industry-specific management threshold is calibrated based on the fault loss data within a preset period in this field, and is adjusted according to the preset cycle in conjunction with industry safety standards; S7. If the judgment result is yes, the acquired source data is stored in the core database of the central control system according to the hardware optimization format, and a hardware resource optimization scheme is generated; the hardware optimization format is that structured data is stored in columns and unstructured data is stored in blocks; the hardware resource optimization scheme is output to the hardware and business collaborative scheduling module. S8. If the judgment result is negative, invalid data in the acquired source data will be removed, and a quality optimization report containing process parameter adjustment suggestions will be generated and fed back to the data acquisition node and the operation and maintenance terminal respectively; the process parameter adjustment suggestions are generated based on industry process characteristic data.

[0021] In the above steps, the hardware, data, and business three-layer collaborative model aims to solve the problem of traditional models that separate hardware, data, and business. It mainly constructs a three-dimensional linkage framework of resources, data, and requirements. Its input layer includes hardware load characteristics, data attribute characteristics, and business process characteristics. The hidden layer constructs feature interaction paths through a random forest algorithm. The output layer consists of data management cycle adjustment coefficients and resource allocation weights. The specific model construction is as follows: During the data preparation phase, historical fault data from the industry over the past few years are collected. Each data point includes fault type, triggering conditions, hardware status snapshot, and process stage information. Hardware load data from the past few months is also collected. Fault data is tagged, such as hardware overload faults and process anomaly faults. Moving averages are used to smooth out instantaneous fluctuations in hardware load data. During the feature engineering phase, hardware load features, including CPU utilization and memory usage, data attribute features, including data type and access frequency, and business process features, including process stage and core level, are extracted. Redundant features are eliminated through correlation analysis, and effective features are retained. During the model training phase, a random forest model was constructed using Python's Scikit-learn library. The number of decision trees was set to 100, the maximum depth to 15 layers, and the minimum number of sample splits to 5. Accuracy, recall, and F1 score were used as evaluation metrics. The training set and test set were divided in a 7:3 ratio. The training iterations were 1000 rounds, and the model was validated every 100 rounds. Training was stopped when the accuracy, recall, and F1 score on the test set stabilized at 92%, 90%, and 91%, respectively. During the model optimization phase, incremental training is performed periodically using newly added fault data and hardware load data to update the decision tree weights and prevent model aging.

[0022] The iterative mechanism is designed for system self-optimization. This involves creating an iterative log table in the database to record the results of each decision feedback, such as hardware resource utilization, fault warning accuracy, and data quality score. The iteration trigger condition is set when the log accumulates for a period of time or when key indicators drop to a critical point. After triggering, the log data is extracted, the adaptation deviation between the model parameters and the benchmark thresholds is analyzed, the model feature weights are adjusted using the gradient descent method, and the benchmark thresholds are recalculated. For example, the core data integrity threshold is fine-tuned based on the average data quality of the past few months. An optimization report is then generated and the old parameters are overwritten.

[0023] After data acquisition is triggered, the multi-protocol acquisition terminal automatically wakes up and synchronously collects source data, hardware resource status data, and industry process characteristic data at preset time intervals. When the amount of effective data accumulated reaches a certain level, such as 100 records, the hardware, data, and business three-layer collaborative model is invoked to trigger the cycle adjustment process, generate dynamic differentiated cycle information, and store it in the database cycle table. According to the acquisition interval in the cycle table, the sample generation task is automatically started at each interval, converting the raw data into labeled samples and storing them in the sample library. When the number of samples in a batch in the sample library reaches 50, the industry-specific dynamic quality benchmark library is automatically retrieved, the quality assessment value is calculated, and a three-dimensional assessment report is generated. A 3D assessment report is generated, automatically linking rare fault data from the industry's historical fault database to obtain a data management judgment value. After the judgment value is generated, it is immediately compared with the industry-specific management threshold. If it exceeds the threshold, the source data is written to the core database in columnar or block storage format, and a hardware resource optimization algorithm is invoked to generate a solution and push it to the operation and maintenance terminal. Conversely, an invalid data removal algorithm is invoked to delete low-quality data, process parameter adjustment suggestions are generated, and pushed to the data acquisition node and the operation and maintenance terminal. Each time a decision feedback is completed, the feedback result is automatically recorded to the iteration log. When the iteration log accumulates 6 months of data, the model parameter optimization and benchmark library update tasks are triggered to realize the system's self-iteration.

[0024] In this embodiment, the steps for acquiring data specifically include: S101. Synchronously acquire source data, hardware resource status data, and industry process characteristic data through a multi-protocol compatible acquisition interface; the multi-protocol compatible acquisition interface has built-in parsers for various industrial protocols, supports parallel access of multiple protocols, and has the ability to synchronize timestamps with preset precision; S102. Establish a three-dimensional mapping table relating data acquisition, hardware status, and process nodes; this mapping table records the processor load value, response latency value, and associated process unit corresponding to each type of source data; the mapping table is automatically backed up to the core database at a preset period; S103. A two-level filtering mechanism is adopted to remove abnormal data; the first-level filtering rule is to filter data whose processor utilization rate exceeds the preset reasonable range or whose response delay exceeds the preset reasonable range during data acquisition; the second-level filtering rule is to filter invalid data that is not caused by hardware by verifying the range of process parameters. S104. Associate the effective data after two-level filtering with a three-dimensional association mapping table to provide basic data support for building a three-layer collaborative model of hardware, data and business.

[0025] In the above steps, the two-level filtering mechanism is a dual screening process, prioritizing hardware over business considerations. The first level of filtering targets the hardware environment during data acquisition, while the second level targets the business rationality of the data itself. The first-level filtering threshold is set based on industrial server safety operation standards: a CPU utilization threshold of 90% and a response latency threshold of 100 milliseconds, referencing the minimum response window for industrial real-time control; exceeding these limits results in the loss of real-time data value. The second-level filtering threshold is determined through industry standards and historical data statistics, such as a pressure threshold of 0.1 to 1.5 MPa for chemical reactors and a load threshold of ±10% of the rated value for power transmission lines. These thresholds are stored in an industry process characteristic table and support remote updates. During use, the first-level filtering obtains CPU utilization by reading the server's proc / loadavg file, calculates the difference between the acquisition time and the parsing completion time to obtain the response latency, and iterates through the data, marking overloaded data. The second-level filtering uses SQL queries to associate process unit IDs with threshold ranges, compares data values ​​with thresholds, marks invalid data, and finally outputs valid data.

[0026] The multi-protocol compatibility mechanism is designed for modular parser development and standardized time synchronization. At the hardware level, multi-protocol acquisition terminals are used, each with a built-in independent protocol processing chip. At the software level, each protocol parser is developed modularly using C language. The PROFINET parser supports real-time modes such as IRT, while the Modbus parser supports both RTU and TCP transmission modes, adapting to devices at different distances. The DCS-specific protocol parser is developed according to the protocol specifications of mainstream central control systems, ensuring data format compatibility. Time synchronization uses the PTP precise time protocol. The acquisition terminal has a built-in high-precision crystal oscillator and sends a time synchronization request to the server at regular intervals. The server returns a standard timestamp via the industrial real-time bus, and the acquisition terminal embeds the timestamp into the data frame header to control the time deviation of data acquired from different protocols and devices.

[0027] A three-dimensional mapping mechanism linking data, hardware, and business operations is established. The mapping table structure follows the third normal form to avoid data redundancy. Fields include: data ID as the primary key; data type as an enumeration value (system running status / user operation behavior, etc.); acquisition time as a timestamp; CPU load value as a floating-point number; response latency value as an integer; process unit ID as a foreign key linking to the process unit table; and backup flag as a boolean. The mapping table is created by executing the CREATE TABLE statement in the MySQL database, setting the process unit ID as a foreign key linking to the process unit table. Incremental backup strategy is adopted, and backup files are stored in a specified directory of the HDD. During data association, the process unit mapping table is queried by device ID to obtain the corresponding process unit ID, and the hardware status data at the time of acquisition is read simultaneously. All information is integrated and written to the mapping table.

[0028] After startup, the multi-protocol acquisition terminal scans the communication protocols of the access devices. It identifies the device protocol type by matching protocol header features: PROFINET protocol header is 0x0000, Modbus is 0x01, and DCS-specific protocol is 0x10. Based on the identification result, it calls the corresponding parser to parse the binary data of the PROFINET protocol, the RTU data of the Modbus protocol, and the custom format data of the DCS-specific protocol into a unified JSON format. Fields include data type, acquisition time, data value, device ID, and process unit ID. The acquisition terminal has a built-in high-precision crystal oscillator and sends a time synchronization request to the server. The server returns a standard timestamp, which the acquisition terminal embeds into each parsed data entry. According to the data structure, the system controls the time deviation of all data. A data, hardware, and business association table is created in the MySQL database, with fields including data ID, data type, acquisition time, CPU load value, response latency value, process unit ID, and backup flag. After parsing the JSON data, the CPU utilization during data acquisition is queried by reading the server's proc / loadavg file, the response latency value is the acquisition time minus the parsing completion time, and the associated process unit ID is matched with the process unit mapping table using the device ID. This information, along with the data itself, is written into the association table. When the number of data records in the association table reaches 100, a backup task is automatically triggered, copying the data to a backup table stored on the HDD, and marking the original table as backed up.

[0029] Next, the unfiltered data in the association table is traversed, and the CPU load value and response latency value are read. If the CPU load value exceeds 90% or the response latency value exceeds 100 milliseconds, it is marked as hardware overload data and moved to the abnormal data table; otherwise, it is marked as normal hardware data. The data values ​​and process unit IDs in the normal hardware data are read, and the parameter threshold range in the industry process feature table is queried according to the process unit ID. If the data value exceeds the threshold range, it is marked as invalid business data and moved to the abnormal data table; otherwise, it is marked as valid data. The ID of the valid data is associated with the three-dimensional label in the association table to generate tagged valid data, which is written to the valid data table for subsequent periodic adjustments and sample generation calls.

[0030] In this embodiment, the steps for determining the dynamic differentiated data management cycle information are as follows: S201. Obtain the storage structure, access frequency, and corresponding process impact weight of the system operation status data in the central control system database; the process impact weight is set according to the category differences between core process-related data and auxiliary data; S202. Extract the time stamp information and hardware resource consumption values ​​of key operating parameters from the system operating status data; the hardware resource consumption values ​​include the memory usage, storage bandwidth utilization, and CPU calculation time during the data writing process; S203. Based on the process impact weight and hardware resource consumption value, and on the basis of the three-layer collaborative model of hardware, data and business, construct a four-dimensional correlation model of key parameter index, operation scenario identification, hardware load and process stage; the input parameters of the four-dimensional correlation model also include industry process characteristic data. S204. Based on the four-dimensional correlation model, set the first dynamic management cycle corresponding to the system operation status data; the specific rule is to adjust the cycle duration according to the combination relationship between processor utilization and process stage. S205. An LSTM and attention mechanism fusion model is used to predict the hardware resource usage trend within a preset time period; the training data of this model includes hardware resource status data and corresponding process stage data within a preset historical time period. S206. Based on the load forecast results, the initial management cycles corresponding to user operation behavior data, device interaction response data, and system log record data are dynamically calibrated, and the calibration error is controlled within the preset allowable range. S207. Summarize the first dynamic management cycle and other management cycles after calibration to generate dynamic differentiated data management cycle information; this cycle information includes data acquisition interval, analysis frequency and storage update frequency, and each type of cycle parameter is associated with the corresponding process safety level label; S208. Synchronize the dynamic differentiated data management cycle information to the hardware resource pre-scheduling algorithm and the industry-specific dynamic quality benchmark library to ensure the cycle coordination of each stage of data processing.

[0031] In the above steps, the four-dimensional correlation model is used to address the shortcomings of the three-layer collaborative model in data refinement adaptation. By adding a data key dimension, it achieves more accurate periodic adjustments. Its input layer includes four-dimensional features: key parameter index, operating scenario identifier, hardware load, and process stage. The key parameter index is 1 for the core parameter of the binary classification feature and 0 for the auxiliary parameter; the operating scenario identifier is 0 for normal operation of the binary classification feature and 1 for maintenance mode; the hardware load is the CPU utilization rate of the continuous feature from 0 to 100%; and the process stage is 0 for stable binary classification feature and 1 for adjustment. The feature cross-layer uses a multinomial feature cross-layer algorithm to generate 16-dimensional high-order features, such as core parameter × high load, adjustment stage × normal operation, etc. The output layer is a single-value output, that is, the periodic adjustment coefficient corresponding to the feature combination is 0.6 to 1.5 times. During implementation, 100,000 data points were extracted from the effective data table, including four-dimensional features and corresponding manually labeled optimal periods. Discrete features were one-hot encoded, and continuous features were normalized and mapped to the 0-1 interval. Then, a fully connected neural network was built using the TensorFlow framework, with 16 nodes in the input layer, 32 and 16 nodes in the two hidden layers respectively, and 1 node in the output layer. The activation function was ReLU, the optimizer was Adam, the learning rate was 0.001, the loss function was mean squared error, and the training iterations were 500 times with a batch size of 32. When deploying the model, the trained model was saved in pb format for easy deployment in industrial environments, and the calling service was provided through a RESTful API interface.

[0032] This model integrates LSTM and an attention mechanism. Its input layer consists of hardware load data from the past six months, sampled at 1 second, with a sequence length of 300 (5 minutes of historical data). The LSTM layer comprises two layers, each with 64 hidden units, and uses dropout regularization (dropout rate = 0.2) to prevent overfitting and capture long-term dependencies in the time series. The attention mechanism employs Bahdanau attention, assigning weights to each time step of the LSTM layer's output to focus on critical periods of load fluctuations. The output layer is a fully connected layer that outputs a prediction of CPU utilization for the next 10 minutes. The measurements were taken at 1 second, resulting in 600 prediction points. During model construction, CPU usage time-series data from the past 6 months was extracted and divided into training and test sets in a 7:3 ratio. The training set used a sliding window method to generate input sequence windows of size 300. Training was then performed using the TensorFlow 2.10 framework with the Adam optimizer and an initial learning rate of 0.001. A learning rate decay strategy was adopted, decreasing by 10% every 100 epochs. The loss function was mean squared error. Training was iterated for 1000 epochs, stopping when the mean absolute error (MAE) on the test set did not exceed 2%. Finally, load data from different industries were used for testing.

[0033] The periodic calibration and integration mechanism employs a triple adjustment of initial period, load prediction, and feature adaptation. The initial period is set based on the update frequency of data types and business importance. User operation data updates frequently, so the initial period is set to 8 seconds; device interaction data updates at a moderate frequency, so the initial period is set to 7 seconds; log data updates at a low frequency, so the initial period is set to 12 seconds; and system operating status data, being core data, has an initial period set to 10 seconds. During use, a 10-minute load prediction curve from an LSTM and attention model is first obtained. If the predicted CPU utilization consistently exceeds 80% and the process stage is stable, all data periods are extended by 50%. If the predicted CPU utilization consistently falls below 40% and the process stage is in adjustment, the core data period is shortened by 40%, and the auxiliary data period is shortened by 20%. If the predicted load fluctuation is large with a variance exceeding 5%, the core data period remains unchanged, and the auxiliary data period is extended by 30%. During periodic integration, the calibrated periods are organized according to data type, and process safety level labels are added for core data (Level 1) and auxiliary data (Level 2), generating dynamically differentiated periodic information. This information is synchronized to the generation and analysis modules via the industrial real-time bus, with the synchronization frequency matching the periodic update frequency.

[0034] During the training and loading of the three-layer collaborative model, 1200 historical fault data records from the past few years and 15.552 million hardware load data records from the past 6 months were extracted from the database and divided into training and testing sets in a 7:3 ratio. A random forest model was built using the TensorFlow framework, with input features including fault type, hardware load value, process stage, and data type, and output as periodic adjustment coefficients. The training iterations were performed for 1000 rounds until the accuracy of the test set was no less than 90%. The trained model was saved as an h5 file and deployed to the server model directory. The model was automatically loaded into memory when the system started.

[0035] When constructing the four-dimensional correlation model, key parameters are extracted from the effective data table. Core parameters are marked as 1, auxiliary parameters as 0, operating scenario identifiers are 0 for normal operation, 1 for maintenance mode, hardware load value (i.e., CPU utilization), and stable process stage as 0, while adjustment is marked as 1. The four-dimensional features are then combined using a polynomial feature cross-validation algorithm to form higher-order features such as key parameters × hardware load and process stage × operating scenario, resulting in a 16-dimensional feature vector. The 16-dimensional feature vector is then input into the three-layer collaborative model to obtain the periodic adjustment coefficient for this feature combination.

[0036] During load prediction and periodic calibration, hardware load time-series data from the past few months are extracted and resampled at 1-second intervals to generate a time-series dataset. An LSTM and attention mechanism model is used to train the dataset to predict the CPU utilization trend for the next 10 minutes, outputting a prediction curve. An initial period is set according to the data type: 8 seconds for user operation data, 7 seconds for device interaction data, 12 seconds for log data, and 10 seconds for system operating status data. The initial period is adjusted based on the adjustment coefficients output by the prediction curve and the four-dimensional correlation model. If the predicted CPU utilization exceeds 80% and the process stage is stable, the period is extended by 50%; if the predicted CPU utilization is below 40% and the process stage is in adjustment, the period is shortened by 40%, with the calibration error controlled within 5%.

[0037] During the periodic integration and synchronization, the four types of data periods after calibration are integrated according to data type, period length and process safety label format. Core data is marked as primary label and auxiliary data as secondary label to generate dynamic differentiated period information. The period information is written into the periodic table in the database and pushed to the generation module and analysis module through the industrial real-time bus to ensure periodic coordination among the modules.

[0038] In this embodiment, the steps for generating multiple batches of data sample sets are as follows: S301. Based on the process priority determined by hardware resource status data and industry process characteristic data, a dynamic resource scheduling algorithm is used to allocate dedicated processing cores, memory quotas and IO bandwidth to each data management task; the process priority determination standard refers to the key process level system, and the allocation priority of core process data is higher than that of auxiliary data; S302. Real-time monitoring of storage I / O rate; when the storage I / O rate is lower than a preset threshold, automatic adaptive compression processing is initiated; structured data uses an improved dictionary encoding adapted to its type, while unstructured data uses a streaming compression algorithm based on process characteristics to ensure the integrity of fault precursor characteristics. S303. Set an adaptive verification node in the data acquisition sequence; dynamically adjust the verification frequency according to the difference in process complexity; if the verification finds that the hardware load exceeds the preset threshold, immediately extend the acquisition interval and trigger the dynamic resource scheduling algorithm to reallocate resources; S304. A hierarchical caching mechanism is adopted to store the collected data; core process-related sample data is temporarily stored in a high-speed memory area, and auxiliary data is temporarily stored in an SSD cache area to reduce the access pressure on the core database. S305. Integrate various types of data after resource allocation, compression, and verification caching, and generate multiple batches of data sample sets according to the dynamic differentiated data management cycle information; each batch of sample sets carries resource consumption tags, process association tags, and quality grade tags; these tags provide support for integrated analysis.

[0039] In the above steps, the dynamic resource scheduling mechanism is priority-oriented and load-aware. Priority determination refers to the ISO 9001 critical process level standard, dividing process units into core process units such as reactor control units and power line protection units, and auxiliary process units such as environmental monitoring units and auxiliary pump control units. The data priority corresponding to core process units is level 1, and auxiliary units are level 2. Weight calculation uses a weighted summation formula: Resource allocation weight = Process priority coefficient × Data processing complexity coefficient × Data access frequency coefficient. Where, the process priority coefficient is 1.0 for level 1 and 0.5 for level 2; the data processing complexity coefficient is 0.8 for structured data and 1.2 for unstructured data; and the data access frequency coefficient is 1.0 for high-frequency data and 1.0 for medium-frequency data. The resource allocation rule is as follows: Level 1 data is allocated more resources than Level 2 data. Resource scheduling adopts a weighted round-robin algorithm. When a Level 1 data processing task arrives, it can preempt the idle resources of Level 2 data. When the resources of a Level 2 data task are preempted, they will be automatically restored after the Level 1 task is completed. When in use, a resource scheduling daemon process is started to monitor the priority of data processing tasks and hardware resource usage in real time. It obtains the CPU and memory usage status by reading the proc / cpuinfo and proc / meminfo files, calculates the resource allocation ratio of each task according to the weight formula, and limits the upper limit of resource usage of each task through the cgroup mechanism of the Linux kernel to ensure accurate resource allocation.

[0040] The adaptive compression mechanism adapts to data types and is aware of I / O load. The compression algorithm is selected based on data structure characteristics: structured data, such as parameter tables and CSV files, uses LZ77 improved encoding, with optimizations including adding a process feature dictionary containing high-frequency process parameters, such as reactor pressure and power transmission load; unstructured data, such as log text and binary files, uses a process feature-based streaming compression algorithm. The core of this algorithm is to set a key feature retention list, including error codes, anomaly description keywords, and process parameter values. During compression, regular expression matching is used to identify the contents of the list, compressing only non-critical content to ensure the integrity of fault precursor features; the I / O rate trigger threshold is set to 100MB. Based on the standard performance parameters of industrial-grade SSDs, when the IO rate is consistently below 100MB / s for three consecutive times, it indicates that the storage system is under high load, and compression is initiated. When the IO rate is above 150MB / s, the storage load is low, and compression is stopped to reduce CPU consumption. During implementation, an IO rate monitoring process is started, reading the proc / diskstats file every 100 milliseconds to calculate the SSD's IO rate (IOPS) × average IO size, and determining whether to trigger compression. The compression task uses multi-threaded parallel processing, with 4 threads for structured data compression and 2 threads for unstructured data compression to avoid latency caused by single-threaded processing.

[0041] The adaptive verification mechanism adapts to process complexity and cyclically controls hardware load. The verification frequency is set based on process complexity. Complex processes involve multiple devices operating in cyclical control, such as a chemical reactor with dual temperature and pressure control, where data fluctuations are large; verification is performed every two data acquisitions. Stable processes involve a single device operating in a steady state, such as an auxiliary pump running at a constant speed, where data is stable; verification is performed every five data acquisitions. The hardware load threshold is set at 85% CPU utilization and 80% memory utilization. Based on long-term stable operation tests of industrial servers, this threshold can prevent system lag caused by hardware overload. During use, the process complexity identifier in the process unit table is read to determine the verification frequency for each data point. Verification is triggered according to the frequency, and the CPU and memory usage status is read. If the threshold is exceeded, the acquisition interval is immediately extended by 20% via the acquisition interval API. At the same time, a resource reallocation request is sent to the dynamic resource scheduling process to recalculate the resource allocation weight and allocate more resources to the current data task until the hardware load drops below the threshold.

[0042] The tiered caching mechanism uses access frequency-based caching and an LRU (Least Recently Used) replacement strategy. Cache levels are divided into high-speed DDR5 memory cache and SSD cache. Core data (Level 1 data) is stored in the high-speed memory cache due to its high access frequency; auxiliary data (Level 2 data) is stored in the SSD cache due to its low access frequency. The cache replacement strategy uses the LRU algorithm, maintaining a linked list of cached data access times. When the cache is full, the least recently used data at the tail of the list is evicted, ensuring that high-frequency data is always in the cache. In practice, a cache management program is written to implement data cache writing, reading, and eviction functions. When data is generated, it is written to the corresponding cache level according to its priority, and the access time is recorded. When reading data, the cache is checked first; if the cache hits, the data is returned directly; otherwise, it is read from the database and written to the cache. When the cache capacity reaches a set limit, the LRU eviction mechanism is triggered, deleting the least recently used data.

[0043] The label generation model adds standardized labels to samples to facilitate subsequent quality assessment and analysis. Labels are categorized into resource consumption labels, process-related labels, and quality level labels. Resource consumption labels record CPU utilization to 0.1%, memory usage to MB, and IO bandwidth utilization to 1% during data processing, generated by reading resource monitoring logs from the data processing process. Process-related labels indicate the corresponding process unit ID, process stage, and core level of the data, retrieved from the process unit table. Quality level labels are based on preliminary verification results, indicating whether the hardware load is normal and whether the data is within the process threshold range, marking data as pending analysis or potentially abnormal. The "potentially abnormal" label indicates data that, while not exceeding the process threshold, shows abnormal hardware load or significant data fluctuations, such as three consecutive data fluctuations exceeding 5%. In practice, a label generation program is built, integrating resource monitoring logs, process unit information, and preliminary verification results to generate labels in JSON format. These labels are then stored in association with the sample data, and the label fields serve as indexes to the sample database, improving query efficiency.

[0044] When implementing dynamic resource allocation, the process unit ID in the valid data is read, and the process priority table is queried. The core process unit has a priority of 1, and the auxiliary one has a priority of 2. Each data item is marked with a priority tag. In the server resource management module, resource allocation rules are set so that the resources allocated to data with priority 1 are greater than those allocated to data with priority 2. The resource scheduling daemon is started to monitor the priority of data processing tasks in real time. When multiple tasks are concurrent, resources are allocated according to priority order. The task with priority 1 can preempt the idle resources of the task with priority 2.

[0045] During adaptive compression, an IO monitoring tool is activated to read the SSD's IO rate in real time by reading the proc / diskstats file and recording the rate value periodically. If the IO rate recorded for three consecutive times is below 100MB / s, a compression task is automatically triggered; if the IO rate is above 150MB / s, the compression task is stopped. For structured data such as system operating status parameter tables, the LZ77 improved encoding program is called, and a process feature dictionary containing high-frequency process parameters is loaded to encode data values ​​into short codes, with the compression rate controlled at 45%. For unstructured data such as log text, a streaming compression program is called, and a key feature retention list is set, including error codes, exception descriptions, and process parameter keywords. During compression, the contents of the list are skipped to ensure the integrity of key features.

[0046] When implementing adaptive verification and tiered caching, the process complexity flag corresponding to the process unit ID is read. If the process complexity flag is 1, it is 0. If the process complexity flag is 1, the verification frequency is set to once every 2 data acquisitions. If the process complexity flag is 0, it is set to once every 5 data acquisitions. At the set frequency, the CPU utilization and memory usage are read. If the CPU utilization exceeds 85% or the memory usage exceeds 80%, the data acquisition interval is immediately extended by 20%, and the resource scheduling process is triggered to reallocate resources. The sample data of priority 1 is written to the cache area of ​​DDR5 high-speed memory with a cache size of 32GB, and the sample data of priority 2 is written to the cache area of ​​SSD with a cache size of 64GB. The LRU strategy is used to replace the cached data.

[0047] During sample integration and tag generation, the compressed structured and unstructured data are uniformly converted into JSON format, with fields including sample ID, data type, priority, compression flag, acquisition time, and process unit ID. Three types of tags are added to each piece of data in the uniform format: resource consumption tags record CPU and memory usage during processing, process association tags indicate the corresponding process unit, and quality level tags indicate samples to be analyzed or suspected anomalies based on preliminary verification results. Samples generated within the same acquisition interval are divided into batches according to the acquisition interval in the dynamic differential cycle information, and each batch is assigned a unique batch ID and written to the sample database.

[0048] In this embodiment, the integration and analysis steps specifically include: S401. Retrieve the quality baseline corresponding to the current process stage from the industry-specific dynamic quality benchmark library, which matches the industry process characteristic data; the quality baseline includes quantitative indicators such as completeness, accuracy, and process correlation, and updates the baseline threshold according to a preset period based on industry accident cases; S402. Differentiate the data types of the generated multi-batch data sample sets to obtain structured data results, semi-structured data results, and unstructured data results; S403. Use the multimodal Transformer model to perform cross-type data correlation analysis on the three types of data results to identify hidden abnormal correlation patterns; S404. Calculate the data information continuity evaluation value; introduce a dynamic time warping algorithm to compare the trend consistency between the current batch data and the historical qualified batch data; if the similarity is lower than the preset similarity threshold, mark it as breakpoint data and record the breakpoint location and possible associated process operations; S405. The weighting of each quality dimension is determined using the analytic hierarchy process (AHP). The weighting is calculated by weighting the hardware consumption coefficient and the process influence coefficient. The accuracy of data related to core equipment has a higher weight than that of non-core equipment. The weighting can be adjusted in real time to respond to process changes. S406. The data management quality assessment value is calculated by combining the quality baseline, correlation analysis results, continuity assessment value and weighting degree; S407. Generate a 3D quality assessment report; this report includes data quality dimension scores, details of anomaly correlation patterns, a list of breakpoint data, and process impact analysis; the 3D quality assessment report is simultaneously fed back to the hardware and business collaborative scheduling module and the process control platform, providing support for the generation of hardware optimization schemes and process parameter adjustment suggestions.

[0049] The multimodal Transformer model is designed to address the correlation analysis problem of different types of data, including structured, semi-structured, and unstructured data. The model structure employs an Encoder-Decoder architecture. The input layer contains feature representations for three types of data: structured data is converted to 128-dimensional vectors through an embedding layer; semi-structured JSON data is converted to 128-dimensional vectors through key-value pair embeddings; and unstructured text data is converted to 128-dimensional vectors using a BERT pre-trained model. The encoding layer contains six Transformer encoder layers, each with eight attention heads, employing a self-attention mechanism to capture correlation features across data types. The decoding layer outputs a correlation strength matrix with dimensions of sample number × sample number and anomaly correlation pattern labels. During model construction, the input layer first... We collected 1 million industrial multimodal data points, including 300,000 structured, 300,000 semi-structured, and 400,000 unstructured data points. We labeled normal and abnormal associations, such as pressure fluctuations, response delays, and error codes as abnormal associations. Then, we fine-tuned the BERT model using an industrial data corpus to adapt it to the text features of industrial scenarios. Next, we input the feature vectors of the above three types of data into a Transformer encoder, using the association strength prediction error and abnormal pattern classification error as the joint loss function. The optimizer was AdamW, the learning rate was 0.0001, and the training iterations were 800 epochs. The batch size was set to 64, and the accuracy of association analysis on the test set reached 94%. During model deployment, the model was converted to ONNX format, and TensorRT was used to accelerate inference and ensure real-time performance.

[0050] The hierarchical analysis weight model is used to prioritize quality assessments. It constructs a three-tiered structure: a target layer, a criterion layer, and an indicator layer. The target layer represents data management quality assessment values; the criterion layer includes hardware consumption criteria and process impact criteria; and the indicator layer includes indicators such as completeness, accuracy, and process relevance. During weight calculation, multiple industry experts are invited to pairwise compare and score the importance of the criterion and indicator layers using a 1-9 scale, where 1 represents equal importance and 9 represents extreme importance. A judgment matrix is ​​then constructed, and the largest eigenvalue and eigenvector are calculated. A consistency check is performed, with CR < 0. 1. If the test passes, the normalized feature vector becomes the weight coefficient; if the test fails, experts are invited to re-evaluate until it passes; finally, the weights of the criterion layer are determined as follows: hardware consumption coefficient 0.4, process influence coefficient 0.6; the weights of the indicator layer are as follows: accuracy of core equipment data 0.8, accuracy of non-core equipment data 0.3, completeness 0.2, process correlation 0.3, and other indicator weights are allocated proportionally. When using, the weight coefficients are stored in the weight configuration table of the benchmark library, which can be adjusted remotely by experts. After adjustment, the consistency test is automatically recalculated to ensure that the weights are reasonable.

[0051] The industry-specific dynamic quality benchmark database mechanism is industry-adapted and incrementally updated. The benchmark database structure includes an industry table, a process stage table, a baseline table, and an update log table. The industry table stores basic information for different industries, such as chemical and power; the process stage table stores the process stage divisions for each industry, such as heating, reaction, and cooling; the baseline table stores the completeness, accuracy, and process correlation of quality thresholds corresponding to each industry and process stage; and the update log table records the baseline adjustment history. Furthermore, core data completeness is no less than 95%, accuracy no less than 98%, and process correlation no less than 0.7; auxiliary data completeness is no less than 90%, accuracy no less than 95%, and process correlation no less than 0.5. The update cycle is set to 3 months, extracting qualified and accident data from the past 3 months, calculating the statistical distribution of each quality indicator, and comparing it with the current baseline threshold. If the deviation between the indicator mean and the threshold exceeds 5%, the threshold is adjusted to the 95th percentile. After an update, an update log is recorded, indicating the reason for the adjustment, such as new accident cases or process adjustments, ensuring baseline traceability.

[0052] The DTW continuity calculation mechanism is used to solve the time scaling problem of different batches of data. The Dynamic Time Warping algorithm finds the optimal matching path between two time series and calculates their similarity. In practice, 100 historical qualified batches of data are selected as reference sequences, and the current batch of data is used as the query sequence. The two sequences are normalized to eliminate the influence of dimensions. A distance matrix is ​​constructed, with elements representing the Euclidean distance between corresponding points in the two sequences. A dynamic programming algorithm is used to find the optimal matching path, calculate the sum of distances along the path, and obtain a similarity interval of 0 to 1 after normalization. If the similarity is lower than 85%, it is marked as a breakpoint, and the timestamp of the breakpoint location and possible reasons are recorded. The algorithm is implemented using Python's dtw library, with a window constraint window_size=50 set to reduce computation and ensure real-time performance.

[0053] The detailed implementation process is divided into four parts: construction of an industry-specific dynamic quality benchmark library, multimodal correlation analysis, continuity calculation and weight determination, and quality assessment value calculation and report generation. Specifically, in constructing the industry-specific dynamic quality benchmark library, 1 million qualified data entries and 50,000 accident data entries from the past three years are extracted and categorized by industry type, such as chemical, power, and intelligent manufacturing, to construct industry datasets. For each industry dataset, the statistical distribution of indicators such as completeness, accuracy, and process correlation is calculated. The 95th percentile is used as the core baseline threshold, and the 90th percentile is used as the auxiliary data baseline threshold, generating an industry-specific baseline table. The baseline update cycle is set to 3 months. At each cycle, newly added qualified and accident data are automatically extracted, the thresholds are recalculated, and the baseline table is updated.

[0054] When implementing multimodal association analysis, a batch of samples is read from the sample library and categorized by data format into structured data parameter tables, semi-structured data JSON format operation logs, and unstructured data text logs. A pre-trained multimodal Transformer model is loaded, and the model is fine-tuned using 10,000 sample data from the current industry to adapt the model to the characteristics of the industry data. After fine-tuning, the accuracy is no less than 94%. The three types of data are input into the fine-tuned model, and the model calculates the association strength between the data through a self-attention mechanism and outputs an association matrix. If the association strength of a certain group of data exceeds 0.8 and meets the characteristics of an abnormal pattern, such as pressure fluctuations, response delays, and error codes, it is marked as an abnormal association pattern.

[0055] When implementing continuity calculation and weight determination, the Dynamic Time Warping (DTW) algorithm is invoked to compare the current batch data with 100 historical qualified batch data to calculate the similarity. If the similarity is less than 85%, it is marked as a breakpoint and its position is recorded. A two-level indicator system of hardware consumption and process impact is constructed, and five industry experts are invited to score the importance of the indicators from 1 to 10. The consistency test coefficient (CR) of the scoring matrix is ​​calculated. If the CR is less than 0.1, the weight coefficients are calculated: hardware consumption coefficient 0.4 and process impact coefficient 0.6. If the CR is not less than 0.1, experts are invited to re-score until the CR is less than 0.1. The equipment importance table is queried, and the accuracy weight of core equipment data is adjusted to 0.8, while that of non-core equipment is 0.3.

[0056] When calculating the quality assessment value and generating the report, the sample data is compared with the baseline threshold. The completeness score is calculated by dividing the number of complete data entries by the total number of data entries and multiplying by 100, and the accuracy score is calculated by dividing the number of accurate data entries by the total number of data entries and multiplying by 100. The single-dimensional scores are multiplied by the corresponding weight coefficients and summed to obtain the total data management quality assessment value of 100 points. The single-dimensional scores, details of abnormal correlation patterns, and list of breakpoint data are integrated to generate a PDF report, which is stored in the report library and pushed to the calculation module and the hardware and business collaborative scheduling mode.

[0057] In this embodiment, the step of calculating the data management judgment value specifically includes: S501. Introduce a fault causal correlation coefficient to replace the traditional correlation coefficient; mine precursor features in scarce fault data by activating inference algorithms, and determine the coefficient value by combining the causal inference model; among them, the coefficient of core precursor data is higher than the coefficient of indirect correlation data; scarce fault data comes from system log records and industry historical fault databases. S502. Optimize the calculation formula for data quality assessment results; the specific formula is: Data quality assessment result = Data management quality assessment value × Data proportion × (Actual data volume / Average data volume of similar data) × Fault causal correlation coefficient × Process stage correction coefficient; Data proportion is the proportion of each type of source data in the total data volume, and the process stage correction coefficient is set according to the differences in the categories of process stages; S503. Calculate the total amount of correction data; adopt a two-dimensional marking rule of amplitude + impact: for data that needs to be corrected in the analysis, if the correction amplitude exceeds the preset amplitude threshold or involves core process parameters, it will be automatically marked as a manual review item; after the manual review is completed, the correction results and review opinions will be updated to the industry-specific dynamic quality benchmark library for the purpose of optimizing the subsequent quality baseline; S504. Calculate the data management judgment value; the quality assessment results of various types of source data are weighted and summed according to the data proportion to obtain the final data management judgment value; S505. If the data management judgment value is within the preset reasonable range of the industry-specific management threshold, then the dual hardware channel secondary analysis is started; the main channel recalculates the judgment value using the current hardware resources, and the backup channel calls the redundant computing nodes of the central control system to recalculate the judgment value; if the similarity between the two results reaches the preset threshold, then the final judgment value is output; if it does not reach the threshold, then the re-integration analysis is triggered until the similarity between the two results reaches the threshold.

[0058] The activated inference causal chain model is based on Bayesian causal inference, with core mechanisms including a two-dimensional labeling mechanism and a dual-hardware channel secondary analysis mechanism. This model addresses the evaluation problem of scarce fault data; traditional correlation analysis is susceptible to spurious associations, while causal inference can uncover direct causal relationships. The model design is based on a Bayesian network, with nodes including fault type nodes, precursor feature nodes, hardware status nodes, and process stage nodes. During model construction, in the data preparation phase, 500 rare fault data points covering 10 types of rare faults (30-50 data points per type) are collected, and the fault type is labeled with its corresponding precursor features, hardware status, and process stage. In the network structure learning phase, a PC algorithm is used to learn the causal relationships between nodes from the data, constructing the network structure. Directed edges between nodes in a Bayesian network structure represent causal relationships. During the parameter learning phase, Bayesian estimation is used to calculate the conditional probability table between nodes, such as the probability of a sudden pressure surge fault occurring under conditions of increased response delay. In the causal strength calculation phase, the causal strength of each precursor feature to the fault type is calculated based on the conditional probability table, ranging from 0 to 1. After normalization, the fault causal correlation coefficient is obtained. The coefficient is set to 0.9 for core precursor features with a causal strength > 0.8, and to 0.3 for indirect correlation features with a causal strength < 0.5. When in use, the model is stored in a fault database, and the model inference and updates are implemented using Python's pgmpy library. For each newly added rare fault data, the conditional probability table and causal correlation coefficient are automatically updated to improve model adaptability.

[0059] The core design of the optimized calculation formula is to integrate multi-dimensional factors to comprehensively reflect data quality and fault risk. The formula is: Data Quality Assessment Result = Data Management Quality Assessment Value × Data Proportion × (Actual Data Volume / Average Data Volume of Similar Data) × Fault Causal Correlation Coefficient × Process Stage Correction Coefficient. Here, the data proportion reflects the distribution of data importance, with core data typically accounting for 60% and auxiliary data for 40%. The actual data volume / average data volume of similar data corrects for differences in data scale, avoiding assessment bias caused by insufficient data volume. The process stage correction coefficient adapts to the risk level of different stages: the adjustment stage has large parameter fluctuations and high risk, so the coefficient is set to 1.2; the stable stage has low risk, so the coefficient is set to 1.0; and the downtime stage has low data value, so the coefficient is set to 0.8. In use, the parameter values ​​are automatically extracted from the database, and the data proportion, actual data volume, and average data volume of similar data are calculated through SQL queries. The data management quality assessment value is read from the quality assessment report, and the correction coefficient is read from the process stage table. These are then substituted into the formula for calculation, and the result is rounded to two decimal places.

[0060] The dual-dimensional marking mechanism uses both magnitude and impact for screening, ensuring that severely distorted or core data is manually reviewed. The correction magnitude threshold is set at 30%. Based on practical experience in data quality calibration, a correction magnitude exceeding 30% indicates a large original error in the data, requiring manual confirmation. The list of core process parameters is developed by industry experts and includes parameters that directly affect production safety. During use, the data correction records from the quality assessment process are read and stored in the calibration log table. Each record is iterated to determine whether the correction magnitude exceeds 30% or whether it belongs to a core process parameter. If so, it is marked as a manually reviewed item, generating a review list containing the data ID, original value, correction value, and associated process unit, which is then pushed to the operations and maintenance terminal. After review by operations and maintenance personnel, the review result is written to the calibration log table as pass / reject. If rejected, the quality assessment is re-executed; if passed, the baseline threshold of the industry-specific dynamic quality benchmark library is updated.

[0061] The dual-hardware-channel secondary analysis mechanism is designed to improve the reliability of judgment values ​​and avoid misjudgments caused by single hardware failures. The main channel is the main CPU of the current server, and the backup channel is the CPU of the redundant computing node. Both channels are deployed with the same calculation program, parameter configuration, and data input. When in use, after the judgment value is generated, if it falls within the industry-specific management threshold ±5% range (e.g., threshold 20, range 19 to 21), parallel calculation of the dual channels is initiated. The main channel reads the raw data and parameters from the database and executes the judgment value calculation process. The backup channel synchronously obtains the same raw data and parameters through the industrial real-time bus and executes the calculation independently. After the calculation is completed, the judgment values ​​of the two channels are compared, and the difference percentage is calculated as |main channel value - backup channel value| / main channel value × 100%. If the difference percentage is ≤1%, the main channel value is output as the final judgment value. If the difference percentage is >1%, it is recalculated twice, and the average of the three results is taken as the final judgment value. If the difference in all three calculations is >1%, a hardware fault alarm is triggered, prompting maintenance personnel to check the server.

[0062] The detailed implementation process is divided into three parts: fault causal correlation coefficient calculation, optimization formula calculation, and dual-dimensional labeling and dual hardware verification. When calculating the fault causal correlation coefficient, fault data with no more than 50 samples of a single type are extracted from the industry's historical fault database, i.e., scarce fault data. Each data point is labeled with fault type and precursor features, such as the precursor features of a sudden pressure surge fault being increased response delay and increased error code frequency. Using Bayesian network tools, a causal graph of precursor features and fault types is constructed using scarce fault data as training samples, and the causal strength of each precursor feature to the fault type is calculated. The causal strength is normalized to the interval between 0 and 1. The coefficient of core precursor feature with a causal strength exceeding 0.8 is set to 0.9, and the coefficient of indirect correlation feature with a causal strength below 0.5 is set to 0.3.

[0063] When implementing the optimization formula, the following steps are taken: First, extract the percentage of a certain type of data entries from the effective data table (divided by the total number of data entries and the actual data volume). Second, extract the average data volume of the same type of data from the historical database. Third, extract the data management quality assessment value from the quality assessment report. Fourth, extract the process stage correction coefficients from the process stage table (adjustment stage 1.2, stable stage 1.0, shutdown stage 0.8). Substitute these values ​​into the optimization formula: Data quality assessment result = Data management quality assessment value × Data percentage × Actual data volume divided by the average data volume of the same type of data × Fault causal correlation coefficient × Process stage correction coefficient. Calculate the quality assessment results for core data and auxiliary data separately. Finally, weighted summation of the single-type data quality assessment results according to the data percentage yields the data management judgment value.

[0064] When implementing dual-dimensional marking and dual-hardware verification, the data correction records during the quality assessment process are read. If the correction range of a certain data exceeds 30% or belongs to a core process parameter, it is marked as a manual review item, a review list is generated and pushed to the operation and maintenance terminal, and the operation and maintenance personnel update the results to the benchmark library after review. The main channel uses the current server CPU, and the backup channel uses the redundant computing node CPU. Both are loaded with the same calculation program and parameter configuration. The main and backup channels are started to calculate the judgment value in parallel. After the calculation is completed, the difference between the two is compared. If the difference does not exceed 1%, that is, the similarity is not less than 99%, the judgment value is output. If the difference exceeds 1%, it is recalculated twice, and the average of the three results is taken as the final judgment value.

[0065] Example 2 Based on Example 1, this application provides a data intelligent management system suitable for a central control system, which includes: The acquisition module is used to acquire source data, hardware resource status data, and industry process characteristic data, and to establish a three-dimensional mapping table of data acquisition, hardware status, and process nodes. The output of the acquisition module is connected to the determination module, generation module, and calculation module respectively, providing basic data for each module. The determination module, based on the three-layer collaborative model of hardware, data and business and the data output by the acquisition module, determines the dynamic and differentiated data management cycle information; the output end of the determination module connects to the generation module and the analysis module, synchronizing the cycle information to the data generation and analysis stages. The generation module generates multiple batches of data sample sets based on the dynamic differentiated data management cycle information output by the determination module through dynamic resource scheduling and adaptive compression, and has a hierarchical caching management function; the output end of the generation module is connected to the analysis module to provide labeled sample sets for the analysis stage. The analysis module, based on an industry-specific dynamic quality benchmark library and the sample set output by the generation module, uses a multimodal Transformer model to output data management quality assessment values ​​and a 3D quality assessment report. The output of the analysis module is connected to the calculation module and the hardware and business collaborative scheduling module, feeding back the assessment results to the judgment calculation and resource scheduling stages. The calculation module uses an activation inference causal chain algorithm, combined with the data management quality assessment value output by the analysis module and rare fault data, to calculate the data management judgment value, and has a dual hardware channel secondary analysis function; the output end of the calculation module is connected to the judgment processing module, and the output judgment value is used for storage / removal decision. The judgment and processing module determines whether the data management judgment value output by the calculation module is greater than the industry-specific management threshold, performs storage or rejection operations, and generates a quality optimization report containing process parameter adjustment suggestions; the output end of the judgment and processing module is connected to the core database, data acquisition nodes, and operation and maintenance terminals respectively to realize data storage and result feedback. The hardware and business collaborative scheduling module is used to receive the 3D quality assessment report output by the analysis module and the hardware resource optimization plan output by the judgment and processing module, so as to realize the cyclical adaptation of data management, hardware resource allocation and industry process requirements; data management corresponds to the determination module and the generation module, hardware resource allocation corresponds to the generation module, and industry process requirements correspond to the acquisition module. Each module interacts with other modules via an industrial-grade real-time bus. The bandwidth of the industrial-grade real-time bus is dynamically adjusted according to the amount of data transmitted and the process priority, with the transmission bandwidth of core process data being guaranteed to the highest priority. The hardware carrier of each module is an industrial-grade server, which has the ability to resist electromagnetic interference and operate in a wide temperature range.

[0066] In this embodiment, the acquisition module specifically includes: The multi-protocol acquisition unit has built-in parsers for multiple industrial protocols, supports time stamp synchronization with preset accuracy, and can access source data, hardware resource status data, and industry process characteristic data in parallel. The three-dimensional association unit establishes a three-dimensional association mapping table between data acquisition, hardware status, and process nodes, and automatically backs it up according to a preset cycle; the input end of the three-dimensional association unit is connected to the multi-protocol acquisition unit, and the output end is connected to the two-stage filtering unit. The two-stage filtering unit first filters overload data based on hardware thresholds, then filters invalid data based on process parameter ranges, and associates the valid data with a three-dimensional association mapping table. The outputs of the two-stage filtering units are respectively connected to the multi-dimensional information acquisition unit of the determination module and the dynamic resource allocation unit of the generation module, providing filtered valid data for subsequent modules. The process parameter monitoring interface, as a component of the data access node, supports direct access and analog-to-digital conversion of industrial sensor signals; the output of the process parameter monitoring interface connects to a multi-protocol acquisition unit, providing it with raw signals of industry process characteristic data.

[0067] In this embodiment, the determining module specifically includes: The multi-dimensional information acquisition unit collects information from the system operation status data output by the acquisition module, including storage structure, access frequency, process influence weight, time stamp information of key operating parameters, and hardware resource consumption values; the output of the multi-dimensional information acquisition unit is connected to the four-dimensional model construction unit. The four-dimensional model building unit, based on the data output by the multi-dimensional information acquisition unit, constructs a four-dimensional correlation model of key parameter index, operation scenario identification, hardware load and process stage on the basis of the three-layer collaborative model of hardware, data and business; the output end of the four-dimensional model building unit is connected to the cycle setting unit. The first cycle setting unit sets the first dynamic management cycle based on the four-dimensional correlation model; the output of the first cycle setting unit is connected to the cycle integration unit. The load prediction unit uses a fusion model of LSTM and attention mechanism to predict the hardware resource usage trend within a preset time period; the input of the load prediction unit is connected to the hardware resource status data of the acquisition module, and the output is connected to the periodic calibration unit. The periodic calibration unit calibrates the initial management cycle based on the trend results output by the load prediction unit; the output of the periodic calibration unit is connected to the periodic integration unit. The cycle integration unit summarizes each dynamic management cycle, generates dynamic differentiated data management cycle information, and synchronizes it to the generation module and the analysis module. The process safety verification unit is used to perform process safety level label association operations to ensure that the collection interval of core process data meets process safety requirements; the output of the process safety verification unit is connected to the cycle integration unit to add safety labels to the cycle information.

[0068] In this embodiment, the generation module specifically includes: The dynamic resource allocation unit allocates processing cores, memory quotas, and I / O bandwidth based on the hardware resource status data and process priority of the acquisition module through a dynamic resource scheduling algorithm; the output of the dynamic resource allocation unit is connected to the intelligent compression unit and the adaptive verification unit. The intelligent compression unit selects the appropriate compression algorithm based on the data type and initiates adaptive compression processing; the input of the intelligent compression unit is connected to the dynamic resource allocation unit, and the output is connected to the hierarchical caching unit. The adaptive verification unit adjusts the hardware load verification frequency according to the process complexity, triggering the dynamic resource allocation unit to reallocate resources; the input of the adaptive verification unit is connected to the intelligent compression unit, and the output is connected to the hierarchical caching unit. The hierarchical caching unit temporarily stores core process-related sample data in a high-speed memory area and auxiliary data in an SSD cache area; the output of the hierarchical caching unit is connected to the sample integration unit. The sample integration unit integrates data from various stages to generate multiple batches of data sample sets, and adds resource consumption tags, process association tags, and quality grade tags to the sample sets; the input end of the sample integration unit is connected to the hierarchical caching unit, and the output end is connected to the multimodal analysis unit of the analysis module to provide labeled samples for the analysis stage; The IO rate monitoring unit is used to monitor the storage IO rate in real time. When the rate is lower than a preset threshold, it triggers the intelligent compression unit to start compression processing. The output of the IO rate monitoring unit is connected to the intelligent compression unit to provide the rate basis for compression triggering.

[0069] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A data intelligent management method suitable for central control systems, characterized in that, The steps of this method are as follows: Acquire control parameters from the central control system, covering multiple sets of source data, hardware resource status data, and industry process characteristic data; Based on a three-layer collaborative model of hardware, data, and business, and combined with the acquired control parameters, dynamic and differentiated data management cycle information is determined; the model is generated through joint training of historical industry fault data and hardware load data. Based on the dynamic differentiated data management cycle information, a data sample set is generated. The hardware processing capacity is adapted to the current and preset time period through a hardware resource pre-scheduling algorithm. The input of the scheduling algorithm includes the current hardware resource status data and industry process characteristic data. Based on an industry-specific dynamic quality benchmark library, the data sample set is integrated and analyzed to obtain data management quality assessment values; the industry-specific dynamic quality benchmark library has built-in industry process templates and is updated in real time with the system's running data through incremental learning; The activation inference causal chain algorithm is used to calculate the data management judgment value by combining the data management quality assessment value and scarce fault data; the activation inference causal chain algorithm extracts fault precursor features and models causal relationships. The system determines whether the data management judgment value is greater than the preset industry-specific management threshold. If it is, the data is stored in the core database of the central control system in a hardware optimization format, and a hardware resource optimization plan is generated. Otherwise, invalid data is removed and a quality optimization report is generated.

2. The method according to claim 1, characterized in that, The specific steps for obtaining data include: The system synchronously acquires source data, hardware resource status data, and industry process characteristic data through the acquisition interface. The hardware resource status data includes processor utilization, memory usage, and storage I / O rate; the industry process characteristic data consists of parameters specific to the corresponding field; each set of source data includes system operating status, user operation behavior, equipment interaction response, and system log record data. Establish a three-dimensional mapping table relating data acquisition, hardware status, and process nodes, record the processor load value, response latency value, and associated process unit corresponding to each type of source data, and back them up regularly; A two-stage filtering mechanism is used to remove abnormal data. The first stage filters data whose processor utilization or response latency exceeds a preset reasonable range during data acquisition. The second stage filters invalid data that is not caused by hardware by verifying the range of process parameters. The effective data after two levels of filtering is associated with a three-dimensional association mapping table, which serves as the basic data for the three-layer collaborative model of hardware, data, and business.

3. The method according to claim 1, characterized in that, The steps to determine the dynamic differential data management cycle information are as follows: The data storage structure, access frequency, and process impact weight are obtained from the system operation status data; the process impact weight is set according to the category differences between core process-related data and auxiliary data. Extract the timestamp information and hardware resource consumption values ​​of key operating parameters from the system's operating status data; the hardware resource consumption values ​​include memory usage, storage bandwidth utilization, and CPU computation time during the data writing process. Based on the process impact weight and hardware resource consumption value, a four-dimensional correlation model is constructed on the basis of the three-layer collaboration model of hardware, data and business, including key parameter index, operation scenario identification, hardware load and process stage; and the first dynamic management cycle corresponding to the system operation status data is set based on the four-dimensional correlation model. A fusion model of LSTM and attention mechanism is adopted to predict the hardware resource usage trend within a preset time period; and the initial management cycle corresponding to user operation behavior, device interaction response and system log data is dynamically calibrated according to the load prediction results. The first dynamic management cycle and other management cycles after calibration are summarized to generate dynamic differentiated data management cycle information; this cycle information includes data acquisition interval, analysis frequency and storage update frequency, and is associated with the corresponding process safety level label; Synchronize with hardware resource pre-scheduling algorithms and industry-specific dynamic quality benchmark libraries.

4. The method according to claim 1, characterized in that, The steps to generate a data sample set include: Based on the process priority determined by hardware resource status data and industry process characteristic data, a dynamic resource scheduling algorithm is used to allocate dedicated processing cores, memory quotas and IO bandwidth to each data management task. Among them, the allocation priority of core process data is higher than that of auxiliary data. The system monitors the storage I / O rate in real time, and initiates adaptive compression when the storage I / O rate falls below a preset threshold. Structured data is encoded using an improved dictionary encoding, while unstructured data is compressed using a streaming compression algorithm. An adaptive verification node is set in the data acquisition sequence, and the verification frequency is dynamically adjusted according to the process complexity. If the verification finds that the hardware load exceeds the preset threshold, the acquisition interval is extended and a dynamic resource scheduling algorithm is triggered to reallocate resources. The core process-related sample data is temporarily stored in the high-speed memory area, and the auxiliary data is temporarily stored in the SSD cache area; Integrate various types of data after resource allocation, compression, verification, and caching, and generate a data sample set according to the dynamic and differentiated data management cycle information; the sample set carries resource consumption tags, process association tags, and quality grade tags.

5. The method according to claim 1, characterized in that, The integration and analysis steps are as follows: The quality baseline is retrieved from the industry-specific dynamic quality benchmark library and matches the industry process characteristic data. The quality baseline includes quantitative indicators such as completeness, accuracy, and process relevance, and the baseline threshold is updated in combination with industry accident cases. The data sample set is distinguished by data type to obtain structured, semi-structured, and unstructured data results; A multimodal Transformer model was used to perform cross-type correlation analysis on the three types of data results to identify hidden abnormal correlation patterns. A dynamic time warping algorithm is introduced to calculate the continuity evaluation value of data information; if the similarity between the current batch data and the historical qualified batch data is lower than the preset similarity threshold, it is marked as breakpoint data and the breakpoint location and associated process operation are recorded; The weights of each quality dimension were determined using the analytic hierarchy process (AHP). The weights are calculated by weighting the hardware consumption coefficient and the process impact coefficient; among them, the accuracy of data related to core equipment has a higher weight than that of non-core equipment. By combining the quality baseline, correlation analysis results, continuity assessment value, and weighting, a data management quality assessment value is calculated, and a three-dimensional quality assessment report is generated.

6. The method according to claim 1, characterized in that, The steps for calculating data management judgment values ​​include: By activating inference algorithms to mine precursor features in scarce fault data, and combining them with causal inference models to determine coefficient values; among them, the coefficient of core precursor data is higher than that of indirect correlation data; scarce fault data comes from system log records and industry historical fault databases. The formula for calculating the optimized data quality assessment result is: Data quality assessment result = Data management quality assessment value × Data proportion × (Actual data volume / Average data volume of similar data) × Fault causal correlation coefficient × Process stage correction coefficient, where the data proportion is the proportion of each type of source data in the total data volume, and the process stage correction coefficient is set according to the differences in the categories of process stages. The total amount of calibration data is counted. If the calibration range exceeds the preset range threshold or involves core process parameters, it is automatically marked as a manual review item. After the manual review is completed, the calibration results and review comments are updated to the industry-specific dynamic quality benchmark library. The quality assessment results of various source data are weighted and summed according to their data proportions to obtain the final data management judgment value. If the data management judgment value is within the preset reasonable range of the industry-specific management threshold, a secondary analysis through dual hardware channels is initiated. The main channel recalculates the judgment value using the current hardware resources, while the backup channel calls the redundant computing nodes of the central control system to recalculate the judgment value. If the similarity between the two results reaches the preset threshold, the final judgment value is output. Otherwise, a re-integration analysis is triggered until the similarity between the two results reaches the threshold.

7. A data intelligent management system suitable for central control systems, characterized in that, The management system includes: The acquisition module is used to acquire source data, hardware resource status data, and industry process characteristic data, and to establish a three-dimensional correlation mapping table from data acquisition and hardware status to process nodes. The determination module is used to determine the dynamic and differentiated data management cycle information based on the three-layer collaborative model of hardware, data and business and the collected data; The generation module generates a set of data samples through dynamic resource scheduling and adaptive compression based on the output dynamic differentiated data management cycle information, and also has a hierarchical cache management function; The analysis module, based on an industry-specific dynamic quality benchmark library and sample set, uses a multimodal Transformer model to output data management quality assessment values ​​and three-dimensional quality assessment reports. The calculation module uses the activation inference causal chain algorithm, combined with data management quality assessment value and scarce fault data, to calculate the data management judgment value, and outputs the judgment value for storage / removal decision; The judgment and processing module is used to determine whether the data management judgment value is greater than the industry-specific management threshold, perform storage or rejection operations, and generate a quality optimization report containing process parameter adjustment suggestions.

8. The system according to claim 7, characterized in that, The data acquisition module specifically includes: Multi-protocol acquisition unit with built-in timestamp synchronization is used to access source data, hardware resource status data and industry process characteristic data; The three-dimensional association unit establishes a three-dimensional association mapping table between data acquisition, hardware status and process nodes and automatically backs it up according to a preset cycle. The two-stage filtering unit first filters overload data based on hardware thresholds, then filters invalid data based on process parameter ranges, and associates valid data with a three-dimensional association mapping table. The process parameter monitoring interface, as a component of the data access node, can directly access and convert industrial sensor signals into analog and digital signals.

9. The system according to claim 7, characterized in that, The module specifically includes: The multi-dimensional information acquisition unit is used to collect information such as storage structure, access frequency, process influence weight, time stamp information of key operating parameters, and hardware resource consumption value from system operation status data. The four-dimensional model building unit, based on the data output by the multi-dimensional information acquisition unit, builds a four-dimensional correlation model of key parameter index, running scenario identification, hardware load, and process stage on the basis of the three-layer collaborative model of hardware, data and business. The first cycle setting unit sets the first dynamic management cycle based on the four-dimensional correlation model; The load prediction unit uses a fusion model of LSTM and attention mechanism to predict the hardware resource usage trend within a preset time period; The periodic calibration unit calibrates the initial management cycle based on the trend results output by the load prediction unit. The cycle integration unit is used to summarize each dynamic management cycle and generate dynamic differentiated data management cycle information; The process safety verification unit is used to perform process safety level label association operations, control the process data collection interval to meet process safety requirements, and add safety labels to periodic information.

10. The system according to claim 7, characterized in that, The generation module specifically includes: The dynamic resource allocation unit allocates processing cores, memory quotas, and I / O bandwidth based on hardware resource status data and process priorities through a dynamic resource scheduling algorithm. The intelligent compression unit selects the appropriate compression algorithm based on the data type and initiates adaptive compression processing. The adaptive verification unit adjusts the hardware load verification frequency according to the process complexity, triggering the dynamic resource allocation unit to reallocate resources. The tiered caching unit temporarily stores core process-related sample data in a high-speed memory area and auxiliary data in an SSD cache area. The sample integration unit integrates data from each stage to generate a data sample set, and adds resource consumption tags, process association tags, and quality grade tags to the sample set; The IO rate monitoring unit is used to monitor the storage IO rate in real time. When the rate is lower than a preset threshold, the intelligent compression unit is triggered to start compression processing.

Images