Heterogeneous data fusion management method for calcium-based material production line

Abstract

The invention discloses a calcium-based material production line heterogeneous data fusion management method, and belongs to the technical field of industrial automation control and information, and the method comprises the steps: obtaining real-time data streams and lag high-precision data generated by a plurality of heterogeneous data sources in a calcium-based material production line; establishing an edge side symbiosis unit for each heterogeneous data source, operating a dynamic behavior model to generate a prediction data stream, and generating an instantaneous deviation record through comparison; constructing a data influence topology network; in response to the lagging high-precision data, executing reverse tracing by using a topological network to determine a target unit and calculating a parameter compensation amount; distributing the compensation amount to perform distributed correction on the model parameters; and generating an updated prediction data stream by using the corrected model, and fusing the updated prediction data stream with the real-time data stream to generate fused data for production line control. According to the method, edge calculation and center coordination are adopted, dynamic modeling is combined, dynamic fusion of multi-source heterogeneous data can be achieved, and the real-time performance, reliability and self-adaptive capacity of production line control are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control and information technology, and in particular to a method for heterogeneous data fusion management of calcium-based material production lines. Background Technology

[0002] The production of calcium-based materials, such as the calcination process of cement clinker, is a typical continuous, nonlinear, and time-delayed industrial process. Its production line integrates numerous sensors and actuators, generating massive real-time data streams including temperature, pressure, flow rate, and composition. Simultaneously, the quality of the final product, such as the activity and strength of the clinker, relies on periodic offline laboratory analyses with significant time lags. These data, from different sources, with varying frequencies and levels of precision, collectively constitute the heterogeneous data environment of calcium-based material production lines. Effective management and utilization of this data are fundamental to achieving refined control and optimization of the production line.

[0003] Among related technologies, Chinese invention patent CN120687265B discloses a method and system for intelligent management and control of technology operations based on big data. The method includes: collecting heterogeneous data sources during technology operations and generating a standardized operation dataset through multimodal fusion processing; then performing dynamic feature grading on the dataset, extracting time-series features and spatial correlation features of key operation indicators, and constructing a multi-level operation status map based on feature importance weights; then matching the business rule base with the map, identifying abnormal nodes and resource conflict paths, and generating an optimized instruction set containing node repair priorities and conflict resolution strategies; then generating an executable management and control operation sequence based on this; finally collecting feedback data streams and updating the business rule base and feature importance weights through an incremental learning mechanism to form a closed-loop optimization link.

[0004] However, the aforementioned existing technical solutions have the following technical drawbacks: **Limitations of centralized processing:** Existing technologies may overly rely on centralized data processing, failing to effectively address the challenges of heterogeneous data sources, high concurrency, and high real-time requirements in calcium-based material production lines. This leads to data processing delays and affects the real-time performance of control. **Insufficient utilization of lagging high-precision data:** Existing technologies may lack an effective mechanism to deeply integrate and reverse-correct lagging, high-precision laboratory data with real-time data streams, failing to utilize high-precision data to dynamically improve the accuracy of real-time data and the predictive ability of the model. **Weak model adaptability:** Existing models may be static or have long update cycles, unable to quickly self-adjust and optimize parameters based on changes in production line conditions, equipment status fluctuations, and raw material characteristics, resulting in poor adaptability. **Data fusion accuracy and reliability issues:** When fusing multi-source heterogeneous data, existing technologies may not fully consider the causal and temporal relationships between data, and the fusion method may not be precise enough, leading to low reliability of the fused data provided to the control system, affecting the stability of production line control and product quality. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a heterogeneous data fusion management method for calcium-based material production lines. By employing edge computing and central coordination, combined with dynamic modeling, it enables dynamic fusion of multi-source heterogeneous data, thereby improving the real-time performance, reliability, and adaptability of production line control.

[0006] The above objectives can be achieved through the following approach: A heterogeneous data fusion management method for a calcium-based material production line includes: acquiring real-time data streams and delayed high-precision data from multiple heterogeneous data sources in the calcium-based material production line; establishing edge-side symbiotic units for each heterogeneous data source and running a dynamic behavior model to generate a predicted data stream, generating instantaneous deviation records through comparison; constructing a data influence topology network; responding to the delayed high-precision data, using the topology network to perform reverse tracing to determine the target unit and calculate parameter compensation amounts; distributing the compensation amounts to perform distributed correction on the model parameters; using the corrected model to generate updated predicted data streams and fusing them with the real-time data streams to generate fused data for production line control.

[0007] Optionally, generating instantaneous deviation records includes: configuring an edge computing node for each heterogeneous data source to perform local computation; storing a dynamic behavior model initialized based on the relationship between the heterogeneous data source and its key upstream process parameters in the edge computing node; driving the dynamic behavior model to continuously receive and process real-time data of the key upstream process parameters and output a predicted data stream; calculating the difference between the predicted data stream and the actual real-time data stream generated by the heterogeneous data source, and generating and storing instantaneous deviation records.

[0008] Optionally, the step of constructing a data influence topology network based on the causal and temporal influence relationships between heterogeneous data sources in the production line includes: defining a corresponding node for each heterogeneous data source and its edge-side symbiotic unit; establishing directed edges between the nodes with direct causal influence to form an initial topology; configuring an influence coefficient for quantifying the influence intensity and a delay attribute for characterizing the influence transmission time for each directed edge of the initial topology; and integrating the nodes, directed edges, influence coefficients, and delay attributes to generate a data influence topology network.

[0009] Optionally, the step of using the data influence topology network to perform reverse tracing, determining one or more target edge-side symbiotic units, and calculating the parameter compensation amount for the dynamic behavior model in the target edge-side symbiotic unit in conjunction with the instantaneous deviation record includes: comparing the lagging high-precision data with the predicted data stream value of the corresponding node in the data influence topology network at the same historical moment to generate an authoritative deviation value; starting from the corresponding node, searching backward along the data influence topology network to determine all relevant upstream nodes; allocating the authoritative deviation value according to the influence coefficient between the upstream node and the corresponding node to form a preliminary compensation requirement for each upstream node; obtaining the instantaneous deviation record of each upstream node at its corresponding historical moment, and coupling the preliminary compensation requirement with the instantaneous deviation record to generate the parameter compensation amount for the dynamic behavior model of each relevant upstream node.

[0010] Optionally, distributing the parameter compensation amount to the target edge-side co-occurrence unit and performing distributed correction on the parameters of the dynamic behavior model includes: encapsulating the parameter compensation amount for different target edge-side co-occurrence units into independent correction instruction packages; routing each correction instruction package to the edge computing node where its corresponding target edge-side co-occurrence unit is located; having the edge computing node parse the correction instruction package to obtain the parameter compensation amount, and using an incremental update method to superimpose and correct the internal parameters of the locally stored dynamic behavior model, thereby updating the local parameters of the dynamic behavior model.

[0011] Optionally, after updating the local parameters of the dynamic behavior model, the method further includes: the target edge-side symbiotic unit using the updated dynamic behavior model to generate an updated prediction data stream based on real-time input key upstream process parameter data; weighting and fusing the updated prediction data stream with the real-time data stream to output fused data; and providing the fused data to an advanced control algorithm for performing control optimization as its input parameters.

[0012] Optionally, the step of fusing the updated predicted data stream with the corresponding real-time data stream to generate fused data for production line control includes: performing a weighted average calculation on the updated predicted data stream and the corresponding real-time data stream to obtain a weighted average value; and performing data smoothing processing on the weighted average value to generate fused data for production line control.

[0013] Optionally, obtaining the instantaneous deviation records of each upstream node at its corresponding historical moment includes: sending a historical data query request carrying the timestamp information of the corresponding historical moment to the edge computing node corresponding to each upstream node; receiving the query request from each edge computing node based on the timestamp information and returning the value of the instantaneous deviation record; and collecting all the returned values ​​of the instantaneous deviation record to form an instantaneous deviation record set for the coupled computation.

[0014] Optionally, after generating the fused data for production line control, the method further includes: monitoring the changes in key performance indicators of the production line operation after control is implemented based on the fused data; calculating the adjustment amount of the influence coefficient in the data influence topology network based on the deviation between the changes in the key performance indicators and the preset performance target used to describe the expected control effect; and updating the influence coefficient in the data influence topology network using the adjustment amount.

[0015] Based on the same inventive concept, this invention also provides a heterogeneous data fusion management system for a calcium-based material production line. The system includes: a data acquisition module for acquiring real-time data streams and delayed high-precision data generated by multiple heterogeneous data sources in the calcium-based material production line; an edge-side symbiotic processing module for establishing an edge-side symbiotic unit for each heterogeneous data source, and using the edge-side symbiotic unit to run a dynamic behavior model describing the behavior of the heterogeneous data source, generating a predicted data stream, comparing the predicted data stream with the corresponding real-time data stream, and generating an instantaneous deviation record; a data influence topology network construction module for constructing a data influence topology network based on the causal and temporal influence relationships between the heterogeneous data sources in the production line; and reverse tracing. The system includes a parameter compensation calculation module, which, in response to receiving the delayed high-precision data, performs reverse tracing using the data to influence the topology network, identifies one or more target edge-side symbiotic units, and calculates the parameter compensation amount for the dynamic behavior model in the target edge-side symbiotic unit in conjunction with the instantaneous deviation record; a distributed parameter correction module, which distributes the parameter compensation amount to the target edge-side symbiotic units and performs distributed correction on the parameters of the dynamic behavior model; and a data fusion and output module, which uses the distributedly corrected dynamic behavior model to generate an updated predicted data stream, and fuses the updated predicted data stream with the corresponding real-time data stream to generate fused data for production line control.

[0016] Compared with the prior art, the present invention has the following advantages: This invention fundamentally resolves the contradiction between traditional fixed processes and fluctuating raw materials by constructing a raw material characteristic profile and deriving optimal process parameters in reverse, enabling the production process to proactively adapt to the characteristics of each batch of raw materials. This raw material-driven process mode breaks through the rigid bottleneck of the original control strategy and improves the adaptability of the production line and the stability of product quality when raw material composition and physical form change.

[0017] This invention incorporates energy consumption and quality indicators into the optimization objectives of process inversion and dynamically solves for the process parameter set that minimizes overall cost, thereby achieving optimal allocation of production resources. Compared to traditional control methods that rely on experience-based setpoints and make delayed adjustments after deviations occur, this invention can proactively avoid unnecessary energy consumption, reduce the overall production cost per unit product while ensuring product qualification rate, and improve overall economic efficiency.

[0018] The disturbance testing and model self-evolution closed-loop mechanism established in this invention transforms the control system from a static execution tool into an intelligent agent with autonomous learning and continuous optimization capabilities. By conducting safe and controlled testing on a real production line and continuously training the core model using feedback data, the system can continuously accumulate process knowledge, automatically correct model deviations, gradually reduce reliance on manual intervention and expert experience, and enhance the robustness and intelligence level of the production system.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a heterogeneous data fusion management method for a calcium-based material production line according to an embodiment of the present invention.

[0022] Figure 2 This is a distributed parameter correction diagram according to an embodiment of the present invention.

[0023] Figure 3 This is a closed-loop diagram of data fusion and performance feedback in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of a heterogeneous data fusion management system for a calcium-based material production line according to an embodiment of the present invention. Detailed Implementation

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

[0026] Reference Figure 1 One embodiment of the present invention proposes a heterogeneous data fusion management method for calcium-based material production lines. By using edge computing and central coordination, combined with dynamic modeling, it can realize the dynamic fusion of multi-source heterogeneous data, thereby improving the real-time performance, reliability and adaptability of production line control.

[0027] The method described in this embodiment specifically includes: S1. Acquire real-time data streams and delayed high-precision data generated by multiple heterogeneous data sources in the calcium-based material production line; Specifically, acquiring real-time data streams and delayed high-precision data from multiple heterogeneous data sources in a calcium-based material production line is fundamental to achieving data fusion. In calcium-based material production lines such as limestone calcination lines, heterogeneous data sources refer to various devices with different data formats, acquisition frequencies, or physical properties, including kiln temperature sensors, fuel weighing scales, raw material feeders, and offline laboratory analytical instruments. The acquisition process uses industrial Ethernet, RS485 bus, or OPC UA communication protocols to read production parameters in real time from programmable logic controllers (PLCs) or distributed control systems (DCS) at the production site. Real-time data streams refer to high-frequency online monitoring data, such as kiln tail temperature or fuel flow rate values ​​acquired every 100 milliseconds. These are characterized by high real-time performance, but their accuracy is significantly affected by ambient noise. Delayed high-precision data refers to compositional data obtained after physical sampling and detailed analysis in a laboratory, such as the clinker free calcium oxide content analyzed every two hours. While this type of data possesses extremely high authority and accuracy, there is a significant lag between sampling and obtaining results. In order to enable the high-precision lagging data to be logically compared with historical real-time data of the same physical meaning, the system attaches precise millisecond-level timestamp information to each data piece when acquiring data and stores it in a unified historical database.

[0028] S2. Establish an edge-side symbiotic unit for each heterogeneous data source, and use the edge-side symbiotic unit to run a dynamic behavior model to describe the behavior of the heterogeneous data source, generate a predicted data stream, compare the predicted data stream with the corresponding real-time data stream, and generate an instantaneous deviation record. Optionally, generating the instantaneous deviation record includes: Configure an edge computing node for performing local computations for each of the heterogeneous data sources; The edge computing node stores a dynamic behavior model initialized based on the relationship between the heterogeneous data source and its key upstream process parameters; The dynamic behavior model is driven to continuously receive and process real-time data of the key upstream process parameters and output a predictive data stream. Calculate the difference between the predicted data stream and the actual real-time data stream generated by the heterogeneous data source, and generate and store instantaneous deviation records.

[0029] Specifically, a dedicated edge computing node is configured for each heterogeneous data source, providing localized, low-latency computing capabilities. This edge computing node is an industrial-grade computing device, typically equipped with a 4- to 8-core processor and 8 to 16 GB of memory. It is deployed in a cabinet less than 5 meters away from the corresponding data source, such as a thermocouple used to monitor kiln temperature or an online analyzer for analyzing raw material composition, achieving millisecond-level data interaction via industrial Ethernet. The edge-side symbiotic unit is the collection of this edge computing node and its internal software model, forming a digital symbiotic relationship with the physical device.

[0030] Deploy and initialize a dynamic behavioral model within this edge computing node. This model is a mathematical process description, which can be a mechanistic state-space equation or a Long Short-Term Memory (LSTM) network model trained on historical data. The core of model initialization is establishing a dynamic relationship between the data source itself and one or more key upstream process parameters. For example, for the kiln tail temperature data source in limestone calcination, key upstream process parameters might include fuel supply rate and primary air fan speed. During initialization, the model's initial weights or coefficient matrix are determined using an offline system identification algorithm based on production data from the past month, ensuring the model possesses a basic understanding of the process upon startup.

[0031] After deployment initialization, to drive continuous real-time predictions by the dynamic behavior model, edge computing nodes subscribe to and receive real-time data streams of the key upstream process parameters at a frequency of 10 Hz via the OPC UA protocol. Whenever a new upstream parameter data packet is received, or according to a fixed execution cycle of 100 milliseconds, the dynamic behavior model performs a forward computation, outputting a predicted data value. These continuously outputted predicted values ​​constitute the prediction data stream. This prediction data stream is synchronized in both value and timestamp with the data stream to be generated by the actual sensors.

[0032] To quantify the accuracy of model predictions in real time and provide a basis for subsequent model adjustments, edge computing nodes simultaneously collect the real-time data stream generated by the data source itself and subtract this real-time value from the predicted data stream value generated by the model at the same timestamp to generate an instantaneous deviation record. This calculation process can be expressed by the following formula: , in, Representative at The instantaneous deviation value is calculated and recorded at each moment; Representative at The actual process parameter values ​​are collected directly from the data source, such as a PLC or DCS system, at all times. This is the dynamic behavior model in the same The predicted value output at time step. Because... and Both refer to the same physical quantity, such as Celsius or Pascal; they have the same dimensions and can be directly subtracted. The generated instantaneous deviation record... Along with the corresponding timestamp, it is stored in a local circular buffer on the edge computing node. This buffer is typically set to store data from the most recent 24 hours, serving as direct data input for subsequent reverse tracing and parameter compensation calculations.

[0033] For example, in the limestone calcination stage of a calcium-based material production line, an industrial edge computing node with a quad-core processor and 8GB of memory is configured for the kiln tail temperature sensor, a heterogeneous data source. This node establishes a millisecond-level connection with the sensor via industrial Ethernet to form an edge-side symbiotic unit. A dynamic behavior model is stored and initialized within this edge computing node, defining the dynamic relationship between the kiln tail temperature and its key upstream process parameter, namely the fuel supply rate. During actual production, the edge computing node continuously receives real-time data on the fuel supply rate at a frequency of 10 Hz via the OPC UA protocol. When the system receives real-time input of the fuel supply rate, it drives the dynamic behavior model to perform forward calculations and output a predicted data stream. In a specific... At that time, assume the model outputs the predicted kiln tail temperature as follows: The temperature was measured in degrees Celsius, while the actual kiln tail temperature collected directly from the PLC system was [value missing]. Degrees Celsius. At this point, using the formula, we obtain... Degrees Celsius. The system then uses the generated instantaneous deviation value. Along with the current timestamp information, it is stored in a local circular buffer on the edge computing node to complete the generation and storage of instantaneous deviation records.

[0034] S3. Construct a data influence topology network based on the causal and temporal influence relationships among heterogeneous data sources in the production line; Optionally, constructing a data influence topology network based on the causal and temporal influence relationships among heterogeneous data sources in the production line includes: Define a corresponding node for each heterogeneous data source and its edge-side symbiotic unit; Directed edges are established between the nodes that have a direct causal relationship to form an initial topology. Configure an influence coefficient for quantifying the influence intensity and a delay attribute for characterizing the influence propagation time for each directed edge of the initial topology; By integrating the nodes, directed edges, influence coefficients, and delay attributes, a data-affected topology network is generated.

[0035] Specifically, each independent heterogeneous data source on the production line, such as a thermocouple monitoring the temperature of a rotary kiln head or a metering scale controlling the amount of raw material fed, along with its symbiotic units bound to the edge side, is abstracted as a basic unit in a network, namely a node. Each node has a unique identifier in the network and is associated with the physical attributes and real-time status of its corresponding data source.

[0036] Directed edges are established based on direct causal relationships to form the initial network topology. This step relies heavily on a deep understanding of the calcium-based material production process and an expert knowledge base. When a change in an upstream process parameter, such as the pulverized coal feed rate, directly and predictably causes a change in a downstream process parameter, such as the calcination temperature in the kiln, a directed edge is established between the node representing the pulverized coal feed rate and the node representing the calcination temperature. The direction of the edge clearly indicates the path of influence transmission from cause to effect, for example, from the pulverized coal feed node to the calcination temperature node.

[0037] Configure key quantifiable attributes for each established directed edge, namely the influence coefficient and the delay attribute. The influence coefficient is a dimensionless parameter that quantifies the strength of influence, representing the relative contribution of changes in upstream nodes to downstream nodes. This coefficient is derived through statistical analysis of a large amount of historical production data. Its calculation can be expressed by the following formula: , in, Represents the upstream node to downstream nodes The influence coefficient. It is achieved by calculating nodes aligned in time according to the delay attribute. Time series data and nodes The value is obtained by considering the covariance or correlation strength of time series data. It is a node The sum of the values ​​of all upstream input nodes. This normalization calculation ensures that all values ​​point to the same node. The sum of all upstream influence coefficients is 1, providing a mathematical basis for the proportional allocation of subsequent deviations. The delay attribute characterizes the time required for an influence to propagate from an upstream node to a downstream node. It is determined by calculating the cross-correlation function of the historical time series of two nodes; the time offset corresponding to the peak of the function is adopted as the delay attribute of the directed edge. For example, the delay attribute of the effect of a change in the opening of a control valve on the pressure at the end of a pipeline might be 5 to 10 seconds.

[0038] All defined nodes, directed edges established based on causal relationships, and influence coefficients and delay attributes precisely calculated for each edge are integrated into a complete data influence topology network, generated and solidified in the memory or graph database of a central server. This network digitally and completely replicates the inherent logic of the production line process flow, existing as a benchmark model for dynamic correction and querying.

[0039] For example, in the clinker quality control stage of a calcium-based material production line, the thermocouple monitoring the rotary kiln head temperature, the metering scale controlling the raw material feed rate, and the frequency converter controlling the pulverized coal feed rate, along with their respective symbiotic units bound to the edge side, are first defined as nodes, and assigned unique identifiers to associate their physical attributes. Based on the calcium-based material production process, since changes in the pulverized coal feed rate directly cause changes in the calcination temperature inside the kiln, the system establishes a directed edge from upstream to downstream between the pulverized coal feed node and the calcination temperature node, forming an initial topology. Next, a quantization attribute is configured for this directed edge. By calculating the cross-correlation function of the historical data time series, the time offset corresponding to the function peak is determined to be 10 seconds and set as a delay attribute. For the calculation of the influence coefficient, it is assumed that the target node is the calcination temperature node. It has an upstream input node that is a pulverized coal feeding node. After time series alignment, the covariance strength between the pulverized coal feeding node and the temperature node was calculated. for At this point, the node is retrieved. The sum of the values ​​of all upstream input nodes Assuming that within this observation period The value is The influence coefficient of the pulverized coal feeding node on the temperature node is obtained by using the formula and substituting the data. Similarly, the influence coefficient of the primary air volume node can be calculated as follows: Ensure that the sum of the influence coefficients pointing to the same node is 1. Finally, the system integrates and solidifies these nodes, directed edges, influence coefficients, and delay attributes in the graph database, generating a complete data influence topology network, thereby reproducing the inherent logic of the production line process flow.

[0040] S4. In response to receiving the high-precision delayed data, the topology network is affected by the data to perform reverse tracing, one or more target edge-side co-occurring units are identified, and the parameter compensation amount for the dynamic behavior model in the target edge-side co-occurring unit is calculated in combination with the instantaneous deviation record. Optionally, the step of using the data to influence the topology network to perform reverse tracing, identifying one or more target edge-side co-occurring units, and calculating the parameter compensation amount for the dynamic behavior model in the target edge-side co-occurring units in conjunction with the instantaneous deviation record includes: The lagging high-precision data is compared with the predicted data stream values ​​of the corresponding nodes in the data-affected topology network at the same historical moment to generate an authoritative bias value. Starting from the corresponding node, search backwards along the data influence topology network to determine all relevant upstream nodes; Based on the influence coefficient between the upstream node and the corresponding node, the authority deviation value is distributed to form the initial compensation requirement for each upstream node. The instantaneous deviation records of each upstream node at its corresponding historical moment are obtained, and the preliminary compensation requirement is coupled with the instantaneous deviation records to generate the parameter compensation amount for the dynamic behavior model of each relevant upstream node.

[0041] Specifically, this is triggered by the generation of an authority bias value. When the system receives high-precision data with a lag, such as a report from a laboratory that sampled and analyzed clinker two hours ago, showing a free calcium oxide content of 1.5%, the system immediately retrieves the corresponding node from the historical database—the clinker quality node—and the predicted data stream value generated by its dynamic behavior model at the same time two hours prior, for example, a predicted value of 1.2%. The system generates an authority bias value by comparing these two values. This calculation can be expressed by the following formula: , in, This is the authority deviation value; It is a value that lags behind high-precision data; Data affects the corresponding nodes in the topology network. The predicted data stream values ​​at the same historical moment. Both are percentages representing the same physical quantity, have the same dimensions, and can be directly subtracted.

[0042] Once the authoritative deviation value is obtained, reverse tracing is immediately initiated. Starting with the clinker quality node, the system traverses the data influence topology network in the opposite direction of all directed edges pointing to that node. For example, it might trace back to nodes such as rotary kiln calcination temperature, raw meal composition control, and pulverized coal feed rate. This search process identifies the set of all relevant upstream nodes that influence the final clinker quality along the causal chain.

[0043] The system allocates the authority deviation value based on the influence coefficient. For each traceable direct upstream node, the system proportionally distributes the authority deviation value according to the influence coefficient configured on the directed edge connecting it to the clinker quality node, forming the initial compensation requirement for each upstream node. For example, if the influence coefficient of the calcination temperature node is 0.6, the initial compensation requirement allocated to that node is: , in, It is for upstream nodes The initial compensation needs; This is the previously calculated authoritative deviation value; It is the data that affects the storage of upstream nodes in the topology network. To the target node The influence coefficient is a dimensionless normalized value.

[0044] The coupled computation generates the final parameter compensation amount. The system then distributes the initial compensation amount to each upstream node that has been assigned it. The system queries the instantaneous deviation record at the corresponding historical moment, i.e., the time point after considering the time delay attribute. This instantaneous deviation record reflects the accuracy of the upstream node's own model's prediction at that time. The system couples the initial compensation requirement with this instantaneous deviation record to calculate the true parameter compensation amount for the node's dynamic behavior model. Its core logic is expressed by the following formula: , in, It is the final generation used to correct upstream nodes. The parameter compensation vector of the dynamic behavior model. It is a gain matrix whose function is to convert the calculated error signal into an adjustment amount for specific parameters inside the model. This is the preliminary compensation requirement calculated in the previous step. The upstream node obtained from the query The instantaneous deviation record value at the corresponding historical moment. It is a weighting factor, ranging from 0 to 1, used to adjust the degree of confidence in instantaneous bias. The significance of this coupling calculation is that if the bias of the upstream model has already been captured by the instantaneous bias record at that time, the amount of compensation should be reduced accordingly; otherwise, it indicates that the bias is a new problem unknown to the model, requiring a larger parameter correction.

[0045] For example, in the clinker quality monitoring process of a calcium-based material production line, when the system receives a high-precision, albeit lagging, data from laboratory analysis indicating a clinker free calcium oxide content of 1.5%, the system immediately retrieves the predicted data stream value of the clinker quality node in the data-affected topology network at the same historical moment two hours prior, assuming this predicted value is 1.2%. Using the formula... By performing comparison calculations and substituting the data, we obtain... Then, starting from the clinker quality node, a reverse search is performed along the topology network to determine the relevant upstream node as the rotary kiln calcination temperature node. Based on the preset influence coefficient between this upstream node and the clinker quality node, an influence coefficient is assumed. It is 0.6, using the formula Calculate the preliminary compensation requirement and obtain The system then queries the edge computing node to which the calcination temperature node belongs for the instantaneous deviation record value recorded at the corresponding historical moment. Assuming this value... Set the gain matrix to 0.1. The weighting factor is 1.2. The value is 0.5, calculated using the coupling formula. The parameter compensation amount is generated, and the calculation process is as follows: The calculated parameter compensation amount is ultimately encapsulated and distributed to the target edge-side symbiotic units, achieving accurate correction of model parameters based on hysteresis authoritative data.

[0046] Optionally, obtaining the instantaneous deviation records of each upstream node at its corresponding historical moment includes: Send historical data query requests carrying timestamp information of the corresponding historical time to the edge computing nodes corresponding to each upstream node; Receive query requests from each edge computing node based on the timestamp information and return the value of the instantaneous deviation record; The values ​​of all returned instantaneous deviation records are collected to form an instantaneous deviation record set for the coupled calculation.

[0047] Specifically, the central system initiates a targeted and time-precise data query command. Once the reverse tracing process identifies the upstream node requiring correction, the central management system sends a historical data query request via Industrial Ethernet to the edge computing nodes corresponding to each relevant upstream node. This request is a standardized data packet whose key payload is the timestamp information of the corresponding historical moment, calculated based on the latency attributes in the data-affected network topology. For example, if a request is made for an event occurring at an upstream node with a 90-second causal delay from the high-precision data production time, the timestamp carried in the request would be the high-precision data production timestamp minus 90 seconds. This request is directly routed to the server endpoint of the target edge computing node via IP-based communication protocols such as HTTP, RESTful, or API calls.

[0048] The edge computing nodes respond to query commands and return accurate historical data. Each edge computing node runs a lightweight service program that continuously listens for query requests from the central management system. Upon receiving a request, the program parses the timestamp information. Using this timestamp, it performs a precise index query on its locally stored time-series database of instantaneous deviation records. The goal of the query is to find the value of the instantaneous deviation record that exactly matches or is closest in time to the requested timestamp. Once found, the edge computing node encapsulates this value in a response data packet and returns it to the central management system via the original communication link. The entire query and return process is typically required to be completed within 100 milliseconds to ensure the efficiency of the overall correction process.

[0049] All scattered historical data are aggregated to prepare a complete input dataset for the next stage of coupled computation. After issuing multiple query requests, the central management system asynchronously waits for and receives return data from various edge computing nodes. The system sets a reasonable timeout, such as 2 seconds, to handle possible network latency or node unresponsiveness. Whenever a returned instantaneous deviation record value is received, the system associates and stores it with the corresponding upstream node identifier. When all expected responses have arrived or the timeout period has expired, the system aggregates all collected values ​​to form a structured instantaneous deviation record set, such as a mapping table with node identifiers as keys and instantaneous deviation values ​​as values. This set comprehensively reflects the model prediction performance of each relevant upstream link at critical causal moments and serves as a key input directly for the subsequent coupled computation stage of calculating parameter compensation.

[0050] For example, in the reverse tracing process of a calcium-based material production line, the central management system initiates a targeted data query command to perform coupled calculations of parameter compensation. Once the reverse tracing determines that the rotary kiln calcination temperature node, an upstream node, needs correction, the central management system sends a historical data query request carrying the corresponding historical time timestamp information to the edge computing node corresponding to that node. Assuming the production time of the lagging high-precision data is 14:00:00, based on the 90-second delay attribute defined in the data influence topology network, the system calculates the corresponding historical time timestamp as 13:58:30 and sends it as a payload to the target edge computing node via industrial Ethernet. The service program of the edge computing node parses the timestamp and performs an index query in its local circular buffer's instantaneous deviation record time series database. The node finds an instantaneous deviation record value matching 13:58:30 with 0.15 degrees Celsius, and then encapsulates this value in a response data packet and returns it to the central management system. The central management system asynchronously receives and aggregates all returned values, associates the upstream node identifier with the instantaneous deviation record value of 0.15, and stores them to form a set of instantaneous deviation records for coupled calculation.

[0051] S5. Distribute the parameter compensation amount to the target edge-side co-occurrence unit to perform distributed correction on the parameters of the dynamic behavior model; Optionally, distributing the parameter compensation amount to the target edge-side co-occurrence unit and performing distributed correction on the parameters of the dynamic behavior model includes: The parameter compensation amounts for different target edge-side co-occurring units are encapsulated into independent correction instruction packages; Each of the correction instruction packets is routed to the edge computing node where its corresponding target edge-side co-occurrence unit is located; The edge computing node parses the correction instruction packet to obtain the parameter compensation amount, and uses an incremental update method to superimpose and correct the internal parameters of the locally stored dynamic behavior model, thereby updating the local parameters of the dynamic behavior model.

[0052] Specifically, the calculated parameter compensation values ​​are converted into independent standard commands that can be transmitted over the network. After calculating the parameter compensation values ​​for one or more target edge-side co-occurring units, the central management system does not immediately issue the commands. Instead, it generates an independent correction command package for each target unit. This command package is typically encapsulated in a lightweight data exchange format such as JSON. Its internal structure includes a unique identifier for the target edge-side co-occurring unit for command addressing; a timestamp to ensure the timeliness of the command; and a parameter compensation value vector as the core payload. For example, a command packet sent to a kiln head temperature prediction model has a payload that is a set of specific adjustment values ​​for the model's internal parameters.

[0053] Through industrial IoT communication mechanisms, each correction instruction packet is delivered accurately to the designated edge computing node. The central management system publishes the encapsulated correction instruction packet to a message middleware, such as a message broker server using the MQTT protocol. Upon startup, each edge computing node subscribes to a dedicated topic from the message broker server based on the unique identifier of the edge-side symbiotic unit it hosts. When the central management system publishes a correction instruction packet to a specific topic, only the target edge computing node that has subscribed to that topic can receive the instruction. This publish-subscribe model achieves loosely coupled, one-to-one precise instruction delivery, with network communication latency typically controlled between 50 and 500 milliseconds.

[0054] Localized and incremental updates of model parameters are performed at the edge, avoiding the complex process of model retraining. After receiving and parsing the correction instruction packet, the edge computing node extracts the parameter compensation amount. Subsequently, it does not discard the original dynamic behavior model, but instead uses incremental updates to perform cumulative adjustments on the model's internal parameters stored in local memory or files. This process follows the following update rules: , in, It is a vector of internal parameters of the updated dynamic behavior model; It is a vector of the model's internal parameters stored locally before correction, such as a set of coefficients for an autoregressive model that describes the dynamics of a system. It is the parameter compensation vector parsed from the correction instruction packet. This formula is a vector addition, requiring... and Having the same dimensions and physical correspondence ensures the accuracy of the correction. This superimposed correction method is a lightweight online learning approach that only fine-tunes the original parameters, resulting in minimal computational overhead. It is particularly suitable for execution on resource-constrained edge computing nodes, thereby quickly updating the local parameters of the dynamic behavior model to more accurately reflect the latest operational characteristics of the production line. Distributed parameter correction, such as... Figure 2 As shown.

[0055] For example, in the distributed correction process of clinker quality in a calcium-based material production line, the central management system will calculate the parameter compensation vector based on the rotary kiln head temperature prediction model. The correction instruction is encapsulated into a separate correction instruction package. This package is in JSON format and contains the unique identifier of the target edge-side co-occurrence unit, a timestamp, and the specific parameter compensation value. The system then routes this correction instruction package to the corresponding edge computing node via the MQTT protocol. The edge computing node parses the correction instruction package to obtain the parameter compensation value; assuming the parsed parameter compensation value vector... The value is [0.05, -0.02]. The node uses an incremental update method to perform cumulative corrections on the internal parameters of the locally stored dynamic behavior model. If the model's internal parameter vector before correction is [0.05, -0.02]... Given the range [0.80, 0.45], calculate according to the formula above. Substituting the specific values ​​yields... =[0.80+0.05,0.45+(-0.02)]=[0.85,0.43]. Through this calculation process, the edge computing node successfully updated the local parameters of the dynamic behavior model, enabling it to more accurately reflect the latest operating characteristics of the production line.

[0056] Optionally, after updating the local parameters of the dynamic behavior model, the method further includes: The target edge-side symbiotic unit utilizes the updated dynamic behavior model to generate an updated predictive data stream based on real-time input key upstream process parameter data; The updated predicted data stream and the real-time data stream are weighted and fused to output fused data; The fused data is provided as input parameters to an advanced control algorithm used to perform control optimization.

[0057] Specifically, by utilizing the dynamically updated behavior model with updated parameters, prediction sequences with higher accuracy are generated in real time. The parameters of the local dynamic behavior model are then set at the target edge symbiotic unit. After the update, the model immediately enters a new working state. It will continue to receive real-time input of key upstream process parameter data, such as pulverized coal feed rate or primary air volume data, at a fixed interval of, for example, every 2 seconds, and utilize the updated model. The calculations generate an updated predictive data stream. This new predictive data stream, because the model parameters have been calibrated with high-precision, lagged data, can more accurately predict future trends in downstream process variables compared to the original version.

[0058] This system organically combines the model's forward-looking predictions with real-time sensor measurements to output a fused data stream that is both smooth and accurate. To overcome potential random noise, spikes, or transient drift in the underlying real-time sensor data stream, and to compensate for the inherent slight lag and uncertainty in model predictions, the system does not directly use the updated prediction data stream. Instead, it employs a weighted fusion algorithm to merge the updated prediction data stream with the real-time data stream from this heterogeneous data source. , in, Is The fused data values ​​generated at any given time; It is the value of the updated predicted data stream generated at time t using the updated dynamic behavior model; yes The value of the real-time data stream acquired by the time sensor; It is a weighting coefficient, with a value ranging from 0 to 1, and is usually set between 0.3 and 0.7. The value of determines whether the fusion result leans more towards model predictions or real-time measurements; its setting can be based on the model's historical prediction confidence or an assessment of the sensor's noise level. Because and Since the dimensions are the same, this weighted average calculation is physically valid, and the result... It includes both the model's dynamic prediction information and is anchored to the current actual measurement value.

[0059] High-quality fused data is supplied to advanced control systems to improve the accuracy and optimization of production line control. The calculated fused data... As a purified and optimized estimate of process variables, this is periodically pushed to advanced control algorithm modules, such as Model Predictive Control (MPC), used to perform control optimization. This is in contrast to directly using noisy real-time data streams. Data integration It provides MPC with smoother and more accurate current state input, enabling MPC to make more stable and forward-looking control decisions, such as more finely adjusting the coal feed rate at the kiln head or the frequency of the grate cooler fan, thereby effectively improving the stability of clinker quality, reducing energy consumption, and ultimately achieving optimized operation of the production line.

[0060] For example, in the advanced control stage of a calcium-based material production line, after the target edge-side symbiotic unit updates the parameters of its local dynamic behavior model, it immediately uses this model to generate an updated predictive data stream based on real-time input key upstream process parameters. Assuming the system continuously receives pulverized coal feed rate data at a 2-second interval and performs forward calculations using the updated model parameters, the updated predictive data stream value generated at the current moment... The temperature is 860 degrees Celsius. At this time, the real-time data stream value of the kiln tail temperature collected by the sensor... The temperature is 854 degrees Celsius. To balance the model's forward-looking predictions with the sensor's actual measurements, the system sets weighting coefficients. The value is 0.6. The calculation is performed according to the weighted fusion formula, and the specific value is substituted to obtain... =0.6×860+(1-0.6)×854=516+341.6=857.6 degrees Celsius. This calculation yields the fused data. It overcomes random noise in real-time data and is then periodically pushed to advanced control algorithms that perform control optimization as input parameters for model predictive control algorithms, used to finely adjust the coal feed rate at the kiln head to achieve optimized operation of the production line.

[0061] S6. Generate an updated prediction data stream using the dynamic behavior model after the distributed correction, and fuse the updated prediction data stream with the corresponding real-time data stream to generate fused data for production line control.

[0062] Optionally, fusing the updated predicted data stream with the corresponding real-time data stream to generate fused data for production line control includes: The updated predicted data stream and the corresponding real-time data stream are weighted and averaged to obtain a weighted average value. The weighted average value is smoothed to generate fused data for production line control.

[0063] Specifically, the model-driven predictions are initially integrated with real-time measurements from the physical world. This step involves performing a weighted average calculation, combining the updated prediction data stream generated by the parameter-corrected dynamic behavior model with the real-time data stream directly collected from the heterogeneous data source to obtain a preliminary weighted average. This calculation aims to balance the predictive power of the model with the immediacy of the measurements.

[0064] The weighted average data sequence is filtered to eliminate potential short-term fluctuations and random noise, ensuring that the final fused data output has a sufficiently high signal-to-noise ratio and stability, thereby avoiding unnecessary disturbances to the downstream control system. Using the weighted average sequence obtained in the previous step as input, a data smoothing algorithm, such as the Exponential Moving Average (EMA), is applied to generate the final fused data used for production line control. The final output fused data sequence, as a smooth, continuous, and high-quality signal that accurately reflects the dynamic trend of the process, is directly supplied to advanced control algorithms.

[0065] For example, in the heterogeneous data fusion generation stage of the calcium-based material production line, the system uses a parameter-corrected dynamic behavior model to generate an updated predicted data stream and fuses it with the real-time data stream collected by sensors. First, the model-driven prediction information is initially integrated with the real-time measurement information from the physical world. Assuming that at a specific sampling moment, the model output updated predicted data stream value after distributed correction is 910 degrees Celsius, while the corresponding real-time data stream value directly collected from the heterogeneous data source is 900 degrees Celsius, the system performs a weighted average calculation on these two values, setting the weights of model prediction and real-time measurement to be equal. The calculated weighted average is 905 degrees Celsius. Subsequently, to eliminate potential short-term fluctuations and random noise in the data sequence, the system uses the obtained weighted average sequence as input and applies the exponential moving average method to perform data smoothing. Assuming the smoothing coefficient is set to 0.8 and the fused data value at the previous moment is 902 degrees Celsius, the fused data value used for production line control at the current moment is calculated to be 904.4 degrees Celsius according to the data smoothing logic. The final output fused data sequence, as a smooth and continuous high-quality signal, is directly supplied to advanced control algorithms, thereby avoiding unnecessary disturbances to downstream control systems.

[0066] Optionally, after generating the fused data for production line control, the process further includes: Monitor the changes in key performance indicators of the production line operation after control is implemented based on the fused data; Based on the deviation between the changes in the key performance indicators and the preset performance targets used to describe the expected control effect, the adjustment amount of the influence coefficient in the data influence topology network is calculated. The adjustment amount is used to update the influence coefficients of the data in the topology network.

[0067] Specifically, the data fusion and performance feedback closed loop, such as Figure 3 As shown, the actual operational effectiveness of the production line after applying integrated data for control is quantitatively evaluated. The central management system continuously and periodically monitors changes in one or more predefined key performance indicators (KPIs) for the production line. These KPIs are macroscopic variables characterizing the final production efficiency, such as standard coal consumption per ton of clinker, 28-day compressive strength of clinker, or kiln system operating rate. The system obtains the actual values ​​of these indicators from the Manufacturing Execution System (MES) or Laboratory Information Management System (LIMS), and the monitoring cycle may be 8 hours per shift or 24 hours per day.

[0068] The actual production results are compared with the expected targets, and the resulting macroscopic performance deviation is transformed into fine-tuning instructions for the specific connection weights in the data-influenced topology network. The actual values ​​of monitored key performance indicators are compared with a preset performance target describing the expected control effect to calculate the performance deviation. Then, based on this performance deviation, the adjustment amount for one or more influence coefficients in the data-influenced topology network is calculated. This calculation process is not a direct allocation, but rather based on a reinforcement learning or gradient descent approach. That is, if the parameter compensation allocation over a past period was based on the influence coefficient of a certain path, and the final KPI did not meet expectations, it indicates that the influence coefficient may have been overestimated or underestimated. , in, It is for upstream nodes to downstream nodes The adjustment amount of the influence coefficient on the directed edge; It is a learning rate, used to control the step size of network parameter updates, and is usually a decimal between 0.01 and 0.1; These are the actual values ​​of key performance indicators obtained from the production line; It is the preset performance target value; The performance deviation is made dimensionless, making it a relative error rate; This is the influence coefficient value of the current data on this path in the topology network. This formula indicates that when actual performance does not meet the target, i.e. Compare If the error is poor, the system will adjust the influence coefficient of the path that plays a role in error sharing in the opposite direction.

[0069] The calculated adjustment amounts are applied to the data-driven topology network, enabling the network model to learn and evolve iteratively. The central management system utilizes these calculated adjustment amounts. The system updates the corresponding influence coefficients in the topology network based on the data stored in the system. , in, This is the updated influence coefficient value. This is its original value. After the update, the system will also perform a normalization operation, which ensures that the sum of the influence coefficients of all upstream paths pointing to the same downstream node remains at a constant value, such as 1, thus guaranteeing the self-consistency of the deviation allocation logic. Through this closed-loop feedback and correction mechanism, the data influence topology network can continuously learn and adapt to the slow changes in the production line over time, such as equipment wear and tear, and fluctuations in raw material composition, making its description of the causal relationships within the production line increasingly accurate.

[0070] For example, in the self-learning evolution phase after data fusion control is completed on the calcium-based material production line, the central management system first quantitatively evaluates the actual operational effectiveness of the production line after applying the fused data for control. The system obtains the actual value of the key performance indicator, standard coal consumption per ton of clinker, from the production execution system at a monitoring frequency of 8 hours per shift. The target value is 110 kg of standard coal per ton. Simultaneously, the system sets preset performance target values ​​to describe the expected control effect. The standard yield is 105 kg of standard coal per ton. Next, the system calculates fine-tuning instructions for the specific connection weights in the data topology network based on the deviation between the actual production results and the expected target, and sets the learning rate. The value is 0.05, and the influence coefficient of the current data on the coal powder feeding node to the temperature node in the topology network is obtained. The value is 0.8. Calculations are performed using the formula, and the value is substituted to obtain... Finally, the central management system uses the calculated adjustment amount to update the impact coefficients stored in the system, calculating the updated impact coefficient values ​​according to the formula. =0.8+(-0.0019)=0.7981. The system then performs a normalization operation on each upstream path of the downstream node to ensure that the sum of the influence coefficients remains at 1, thereby realizing the self-learning and iterative evolution of the topology network in response to wear and tear of production line equipment or fluctuations in raw material composition.

[0071] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides a heterogeneous data fusion management system for a calcium-based material production line, the system comprising: The data acquisition module is used to acquire real-time data streams and delayed high-precision data generated by multiple heterogeneous data sources in the calcium-based material production line; The edge-side symbiotic processing module is used to establish an edge-side symbiotic unit for each heterogeneous data source, and use the edge-side symbiotic unit to run a dynamic behavior model describing the behavior of the heterogeneous data source, generate a predicted data stream, compare the predicted data stream with the corresponding real-time data stream, and generate an instantaneous deviation record. The data influence topology network construction module is used to construct a data influence topology network based on the causal and temporal influence relationships between various heterogeneous data sources in the production line. The reverse tracing and parameter compensation calculation module is used to respond to the received high-precision delayed data, use the data to influence the topology network to perform reverse tracing, determine one or more target edge-side co-occurring units, and calculate the parameter compensation amount for the dynamic behavior model in the target edge-side co-occurring unit in combination with the instantaneous deviation record; A distributed parameter correction module is used to distribute the parameter compensation amount to the target edge-side co-occurrence unit and perform distributed correction on the parameters of the dynamic behavior model. The data fusion and output module is used to generate an updated predictive data stream using the dynamic behavior model after the distributed correction, and to fuse the updated predictive data stream with the corresponding real-time data stream to generate fused data for production line control.

[0072] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0073] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for heterogeneous data fusion management in a calcium-based material production line, characterized in that, The method includes: Acquire real-time data streams and delayed high-precision data from multiple heterogeneous data sources in the calcium-based material production line; An edge-side symbiotic unit is established for each heterogeneous data source, and a dynamic behavior model describing the behavior of the heterogeneous data source is run using the edge-side symbiotic unit to generate a predicted data stream. The predicted data stream is then compared with the corresponding real-time data stream to generate an instantaneous deviation record. Based on the causal and temporal influence relationships among various heterogeneous data sources in the production line, a data influence topology network is constructed. In response to receiving the high-precision delayed data, the system uses the data to influence the topology network to perform reverse tracing, identifies one or more target edge-side co-occurring units, and calculates the parameter compensation amount for the dynamic behavior model in the target edge-side co-occurring unit in combination with the instantaneous deviation record. The parameter compensation amount is distributed to the target edge-side co-occurrence unit to perform distributed correction on the parameters of the dynamic behavior model; An updated predictive data stream is generated using the distributed corrected dynamic behavior model, and the updated predictive data stream is fused with the corresponding real-time data stream to generate fused data for production line control.

2. The method for heterogeneous data fusion management of a calcium-based material production line according to claim 1, characterized in that, The generation of instantaneous deviation records includes: Configure an edge computing node for performing local computations for each of the heterogeneous data sources; The edge computing node stores a dynamic behavior model initialized based on the relationship between the heterogeneous data source and its key upstream process parameters; The dynamic behavior model is driven to continuously receive and process real-time data of the key upstream process parameters and output a predictive data stream. Calculate the difference between the predicted data stream and the actual real-time data stream generated by the heterogeneous data source, and generate and store instantaneous deviation records.

3. The heterogeneous data fusion management method for a calcium-based material production line according to claim 1, characterized in that, The construction of the data influence topology network based on the causal and temporal influence relationships among various heterogeneous data sources in the production line includes: Define a corresponding node for each heterogeneous data source and its edge-side symbiotic unit; Directed edges are established between the nodes that have a direct causal relationship to form an initial topology. Configure an influence coefficient for quantifying the influence intensity and a delay attribute for characterizing the influence propagation time for each directed edge of the initial topology; By integrating the nodes, directed edges, influence coefficients, and delay attributes, a data-affected topology network is generated.

4. The heterogeneous data fusion management method for a calcium-based material production line according to claim 1, characterized in that, The process of using the data to influence the topology network to perform reverse tracing, identifying one or more target edge-side co-occurring units, and calculating the parameter compensation amount for the dynamic behavior model in the target edge-side co-occurring units in conjunction with the instantaneous deviation records includes: The lagging high-precision data is compared with the predicted data stream values ​​of the corresponding nodes in the data-affected topology network at the same historical moment to generate an authoritative bias value. Starting from the corresponding node, search backwards along the data influence topology network to determine all relevant upstream nodes; Based on the influence coefficient between the upstream node and the corresponding node, the authority deviation value is distributed to form the initial compensation requirement for each upstream node. The instantaneous deviation records of each upstream node at its corresponding historical moment are obtained, and the preliminary compensation requirement is coupled with the instantaneous deviation records to generate the parameter compensation amount for the dynamic behavior model of each relevant upstream node.

5. The method for heterogeneous data fusion management of a calcium-based material production line according to claim 1, characterized in that, The step of distributing the parameter compensation amount to the target edge-side co-occurrence unit and performing distributed correction on the parameters of the dynamic behavior model includes: The parameter compensation amounts for different target edge-side co-occurring units are encapsulated into independent correction instruction packages; Each of the correction instruction packets is routed to the edge computing node where its corresponding target edge-side co-occurrence unit is located; The edge computing node parses the correction instruction packet to obtain the parameter compensation amount, and uses an incremental update method to superimpose and correct the internal parameters of the locally stored dynamic behavior model, thereby updating the local parameters of the dynamic behavior model.

6. The heterogeneous data fusion management method for a calcium-based material production line according to claim 5, characterized in that, The process of updating the local parameters of the dynamic behavior model further includes: The target edge-side symbiotic unit utilizes the updated dynamic behavior model to generate an updated predictive data stream based on real-time input key upstream process parameter data; The updated predicted data stream and the real-time data stream are weighted and fused to output fused data; The fused data is provided as input parameters to an advanced control algorithm used to perform control optimization.

7. The heterogeneous data fusion management method for a calcium-based material production line according to claim 1, characterized in that, The step of fusing the updated predicted data stream with the corresponding real-time data stream to generate fused data for production line control includes: The updated predicted data stream and the corresponding real-time data stream are weighted and averaged to obtain a weighted average value. The weighted average value is smoothed to generate fused data for production line control.

8. The method for heterogeneous data fusion management of a calcium-based material production line according to claim 4, characterized in that, The acquisition of the instantaneous deviation records of each upstream node at its corresponding historical moment includes: Send historical data query requests carrying timestamp information of the corresponding historical time to the edge computing nodes corresponding to each upstream node; Receive query requests from each edge computing node based on the timestamp information and return the value of the instantaneous deviation record; The values ​​of all returned instantaneous deviation records are collected to form an instantaneous deviation record set for the coupled calculation.

9. The method for heterogeneous data fusion management of a calcium-based material production line according to claim 1, characterized in that, The process of generating fused data for production line control also includes: Monitor the changes in key performance indicators of the production line operation after control is implemented based on the fused data; Based on the deviation between the changes in the key performance indicators and the preset performance targets used to describe the expected control effect, the adjustment amount of the influence coefficient in the data influence topology network is calculated. The adjustment amount is used to update the influence coefficients of the data in the topology network.

10. A heterogeneous data fusion management system for a calcium-based material production line, applied to the heterogeneous data fusion management method for a calcium-based material production line as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire real-time data streams and delayed high-precision data generated by multiple heterogeneous data sources in the calcium-based material production line; The edge-side symbiotic processing module is used to establish an edge-side symbiotic unit for each heterogeneous data source, and use the edge-side symbiotic unit to run a dynamic behavior model describing the behavior of the heterogeneous data source, generate a predicted data stream, compare the predicted data stream with the corresponding real-time data stream, and generate an instantaneous deviation record. The data influence topology network construction module is used to construct a data influence topology network based on the causal and temporal influence relationships between various heterogeneous data sources in the production line. The reverse tracing and parameter compensation calculation module is used to respond to the received high-precision delayed data, use the data to influence the topology network to perform reverse tracing, determine one or more target edge-side co-occurring units, and calculate the parameter compensation amount for the dynamic behavior model in the target edge-side co-occurring unit in combination with the instantaneous deviation record; A distributed parameter correction module is used to distribute the parameter compensation amount to the target edge-side co-occurrence unit and perform distributed correction on the parameters of the dynamic behavior model. The data fusion and output module is used to generate an updated predictive data stream using the dynamic behavior model after the distributed correction, and to fuse the updated predictive data stream with the corresponding real-time data stream to generate fused data for production line control.

Images