A sintering process real-time optimization method based on edge computing

Abstract

This invention discloses a real-time optimization method for sintering processes based on edge computing. It acquires raw multi-indicator time-series data collected in real-time from the edge of the sintering process. For sensor points at different locations in the sintering production line, sampling data and their corresponding process labels are recorded. Based on real-time prediction task load, edge node computing power, and network bandwidth monitoring results, it determines whether the current node has reached the elastic migration trigger condition. If so, subsequent task migration operations are executed. Based on the GPU / CPU load, storage space, and prediction task fluctuations of each node, the computing and storage quotas of the edge nodes are dynamically adjusted. This invention effectively improves the accuracy, real-time performance, and edge system adaptability of multi-indicator collaborative prediction in complex industrial processes by capturing the deep, time-varying coupling relationships between process indicators, constructing an elastic and dynamic resource scheduling and model adaptation mechanism, enhancing the dynamic graph modeling capability of the edge side, and introducing a model self-learning optimization mechanism.

Description

Technical Field

[0001] This invention relates to the field of industrial edge computing technology, and in particular to a real-time optimization method for sintering processes based on edge computing. Background Technology

[0002] Currently, the fields of industrial edge computing and sintering process optimization are developing rapidly, especially in large-scale continuous production scenarios such as steel and metallurgy. Real-time prediction and collaborative optimization of multiple process parameters based on edge intelligent analysis has become an important direction for the industry's digital transformation. Mainstream technical solutions typically involve deploying various types of industrial sensors (such as temperature, pressure, and gas composition sensors) at different stages of the production line. Edge nodes locally collect and pre-process this multi-source real-time data. After pre-processing, time-series-based prediction models (such as LSTM and GRU) or traditional physical mechanism-based models are used to achieve short-term prediction of key process indicators and early warning of process anomalies. At a higher level, some systems are attempting to incorporate concepts such as federated learning and distributed optimization control to improve information collaboration and overall response speed among multiple production units.

[0003] In recent years, Graph Neural Networks (GNNs), as an advanced tool for modeling complex structured data (such as social networks, power systems, and distributed Internet of Things), have been initially introduced into the field of industrial process modeling and analysis. Its technological trend is gradually shifting from predictive inference of single indicators or process segments to collaborative analysis and deep coupling modeling among multiple process parameters.

[0004] In terms of typical application scenarios, mainstream technologies in the fields of "industrial edge computing," "industrial intelligent prediction and optimization control," and "dynamic graph neural network modeling and collaborative analysis" are widely used in production industries such as steel sintering, continuous casting, intelligent logistics, and energy dispatching. These technologies are used for edge prediction of key process parameters, production line anomaly detection, analysis of the impact between indicators, and distributed process optimization. For example, in the steel sintering process, edge computing nodes collect multivariate parameters such as production line temperature, pressure, and gas content in real time, perform local preprocessing, and use methods such as multivariate regression and temporal neural networks based on historical data to predictively regulate downstream processes. Furthermore, in the context of the Industrial Internet of Things (IIoT), edge nodes focus on localized data processing, improving data security and ensuring timely response to high-frequency prediction needs. The gradual application of dynamic graph neural networks enables some cutting-edge systems to perform structured modeling of complex parameter relationships in the production process, improving the depth and breadth of process anomaly detection and indicator linkage analysis.

[0005] While existing technologies have made some progress in real-time prediction of multiple indicators and collaborative optimization at the edge, the following significant shortcomings and unmet needs still exist:

[0006] First, most existing edge prediction systems either model each indicator independently or perform multi-indicator linkage analysis using only simple static statistical correlations, failing to fully capture and utilize the deep, time-varying, and strongly coupled relationships between indicators in complex industrial processes such as sintering. This results in insufficient scientific rigor and robustness of the system's predictions under dynamic conditions such as process disturbances and increased indicator volatility, leading to a weak decision-making basis for collaborative optimization.

[0007] Secondly, for multi-indicator prediction tasks and collaborative allocation of edge computing resources, existing technologies generally rely on static models or manual parameter adjustments, lacking elastic scaling and dynamic adaptive scheduling mechanisms that integrate information from multiple sources such as real-time prediction task pressure, node computing power, and network bandwidth. Once edge nodes are handling high-load tasks or encounter resource bottlenecks, it is difficult to guarantee prediction response speed and service continuity, and the collaborative prediction and model migration capabilities across nodes are severely limited.

[0008] Furthermore, although some high-end systems have explored dynamic graph modeling, they are mostly implemented in a centralized cloud manner. Their capabilities for adaptive adjustment of dynamic graph structures driven by high-frequency local data on the edge and dynamic mining of mutual information between indicators are limited, and they cannot support the high requirements of industrial sites for real-time performance, data localization and system scalability.

[0009] Furthermore, existing systems rarely incorporate mechanisms such as model self-learning optimization and online transfer adaptation, making it impossible to automatically correct structural parameters based on historical prediction results and changing process scenarios. This hinders the evolution of prediction accuracy and the system's ability to adapt to high dynamics in industrial environments over long-term operation.

[0010] Based on the above situation, and the urgent technical needs of industrial sites in areas such as dynamic interactive capture of multiple indicators, edge collaborative prediction, flexible resource scheduling and adaptive migration, there is an urgent need for an innovative method that can comprehensively improve the accuracy, real-time performance and adaptability of multi-indicator collaborative prediction. Summary of the Invention

[0011] To address the aforementioned technical problems, this invention provides a real-time optimization method for the sintering process based on edge computing. The technical solution of this invention is implemented as follows: A real-time optimization method for the sintering process based on edge computing, comprising:

[0012] S1. Obtain the raw multi-index time series data collected in real time on the edge side of the sintering process. For the sensing points at different sections of the sintering production line, record the sampling data and their corresponding process labels respectively.

[0013] S2. Preprocess the multi-indicator time series data to generate a set of data stream features for the input of the adaptive graph neural network;

[0014] S3. Based on the process physical constraints of different work sections and historical statistical correlations, a dynamic graph structure is constructed with each key indicator in the multi-index time series data as nodes and the coupling relationship as edges, and the preprocessed data stream feature set is mapped to the dynamic graph structure.

[0015] S4. Input the dynamic graph structure into a lightweight graph neural network deployed at the edge, and output the predicted value distribution of each key indicator and its mutual information parameters within a predetermined future time window.

[0016] S5. Based on the mutual information parameters, dynamically evaluate the coupling strength between process indicators, and optimize the internal structure weights of the graph neural network accordingly.

[0017] S6. Based on the real-time predicted task load, edge node computing power and network bandwidth monitoring results, determine whether the current node has reached the elastic migration trigger condition. If it has, execute the subsequent task migration operation.

[0018] S7. For migration trigger results, a task adaptive migration protocol is used to distribute some or all graph structure fragments and prediction weights to nearby edge nodes or edge clouds, to complete the dynamic expansion or contraction of the prediction task and automatically synchronize the model state.

[0019] S8. Based on the GPU / CPU load, storage space, and predicted task fluctuations of each node, dynamically adjust the computing and storage quotas of edge nodes to jointly optimize the overall resource utilization and system prediction response speed.

[0020] S9. Embed a self-learning optimization mechanism for the entire multi-indicator collaborative prediction task and resource scheduling process. Automatically correct model transfer parameters and scheduling priorities based on historical prediction accuracy and resource allocation performance.

[0021] S10. When a significant change in process conditions is detected or the predicted anomaly rate increases, the model structure fine-tuning and parameter self-update process is automatically triggered.

[0022] Furthermore, in step S1, the raw multi-index time series data collected in real time on the edge side of the sintering process are obtained, and the sampling data and their corresponding process labels are recorded for the sensing points at different sections of the sintering production line.

[0023] Furthermore, in step S1, the specific steps are as follows:

[0024] For the various types of industrial-grade sensing points deployed in each section of the sintering production line, including temperature sensors, pressure sensors and gas component analyzers, data acquisition operations are performed to obtain the raw time-series signals of key process parameters such as temperature, pressure and gas content, forming a basic multi-index sensing signal stream.

[0025] Based on industrial Ethernet or customized acquisition channels, high-frequency real-time sampling of sensor signal streams is performed, and each original sampled data is uniquely identified by timestamp and node numbering mechanism to obtain multi-indicator original data packets with sampling time and spatial positioning.

[0026] The original data packets are mapped to their locations using the sintering process flow database, and the process physical section labels are automatically associated based on the section information to which the sensing points belong.

[0027] For raw data packets with multiple indicators and process physical section labels, embed the process real-time status code and operation team identifier, and write the operation context information into the raw time series data through adaptive data label mapping.

[0028] For the original time-series data of multiple indicators that have been completed and bound to the location and tags, automatic archiving is carried out in time periods through the high-speed caching mechanism of edge computing nodes, generating original time-series database entries with multiple process tags.

[0029] Furthermore, in step S2, the multi-index time series data is preprocessed to generate a set of data stream features for the input of the adaptive graph neural network.

[0030] Furthermore, in step S2, the specific steps are as follows:

[0031] Quality inspection is performed on the collected time series data of multiple indicators. All outlier data points that exceed the normal fluctuation range of process parameters are marked based on the statistical threshold method. Abnormal interference data caused by sensor failure, communication packet loss or extreme operating conditions are removed.

[0032] For the multi-index time series data after outlier removal, an interval scaling normalization algorithm is used to uniformly transform various process parameters to a preset dimensionless interval, resulting in normalized multi-index time series data.

[0033] For missing records in normalized multi-indicator time series data, a time series interpolation algorithm is applied to fill in the gaps by reasonable numerical inference, and to reconstruct a complete and continuous multi-indicator data stream.

[0034] For the complete multi-index data stream obtained after outlier removal, normalization and missing data interpolation, a sliding time window segmentation method is used to decompose the data stream into equal-length window segments according to the process attributes, and generate a window-level feature vector with time recursion for each window segment.

[0035] The feature vectors of the multi-index data stream, after being segmented by a window, are associated with the corresponding process labels, and a set of data stream features for the input of the adaptive graph neural network is generated based on the data mapping rules.

[0036] Furthermore, in step S3, based on the process physical constraints of different work sections and historical statistical correlations, a dynamic graph structure is constructed with each key indicator in the multi-index time series data as nodes and coupling relationships as edges, and the preprocessed data stream feature set is mapped to the dynamic graph structure.

[0037] Furthermore, in step S3, the specific steps are as follows:

[0038] Based on the process physical constraint model, the index partitioning and attribution determination is performed on the preprocessed multi-index data stream characteristics to obtain the physical distribution labels of each key index in different sections of the sintering production line.

[0039] For process indicators that have been assigned to different regions, the historical process database and historical data collection logs at the edge are used to perform correlation coefficient matrix calculation and Granger causality analysis to extract statistical correlation parameters among key indicators.

[0040] By combining the process physical constraint model and historical statistical correlation parameters, node sets are generated for all key indicators, and the initial coupling relationship edge weights are labeled for indicator pairs with physical or statistical correlations to form a basic static graph structure topology.

[0041] Based on the newly acquired data stream feature set, a sliding time window mechanism and a dynamic edge weight assignment algorithm are applied to update the edge weight strength in the static graph structure in real time.

[0042] Map the node attributes of the data stream feature set covered by the current sliding time window based on the dynamic graph structure described above.

[0043] For the dynamically adjusted dynamic graph structure, a distributed consensus protocol based on edge nodes is used to verify the consistency of the dynamic graph structure of multiple nodes and synchronize the dynamic graph structure to all edge nodes that need to participate in the collaborative prediction of multiple indicators.

[0044] Furthermore, in step S4, the dynamic graph structure is input into a lightweight graph neural network deployed at the edge, and the predicted value distribution and mutual information parameters of each key indicator within a predetermined future time window are output.

[0045] Furthermore, in step S4, the specific steps are as follows:

[0046] Using the dynamic graph structure constructed in step S3 as the input object, the lightweight graph neural network inference instance already deployed locally on the edge node is called to complete the dynamic graph input data format adaptation and model initialization.

[0047] Based on the input dynamic graph structure, the GNN information aggregation algorithm is used to perform graph signal transmission and feature embedding operations on each key indicator node and its adjacent relationship edges, generating the feature embedding vector of each indicator node in the current time series to enhance the correlation of nodes.

[0048] Based on node feature embedding vectors, a lightweight multi-task prediction sub-network is driven. Based on the sintering process prediction window parameters, parallel multi-index distribution prediction calculations are performed on each key index node to obtain the multi-index prediction value distribution results.

[0049] Based on the distribution results of the predicted values ​​of the multiple indicators, and combining node feature embedding and dynamic graph structure, a quantitative analysis is performed on the information gain and joint distribution relationship between all key indicator node pairs, and the mutual information parameter matrix results between each indicator pair are output.

[0050] The distribution results of the multi-indicator predicted values ​​and the results of the mutual information parameter matrix are packaged into a standardized structured output.

[0051] Furthermore, in step S5, the coupling strength between process indicators is dynamically evaluated based on the mutual information parameters.

[0052] Furthermore, in step S5, the specific steps are as follows:

[0053] The mutual information parameters output by the preorder graph neural network collaborative prediction module are captured and structured in real time to obtain the mutual information distribution matrix between different key process indicators.

[0054] Based on the mutual information distribution matrix, an information theory-based coupling coefficient calculation algorithm is applied to dynamically quantify the coupling strength between each pair of key process index nodes and output the process index coupling strength matrix.

[0055] Using the coupling strength matrix of process indicators as input, a weighted adaptive update strategy is used to dynamically optimize and adjust the node weights and edge connection weights inside the graph structure of the graph neural network.

[0056] After completing the weight adjustment, the lightweight model at the edge is retrained iteratively based on the updated graph neural network structure to perform rapid incremental learning on the multi-indicator feature set collected in the recent time window.

[0057] The results of structural weight optimization are continuously and dynamically monitored. By comparing them with real-time prediction accuracy and historical performance indicators, the effectiveness of the current graph neural network optimization adjustment is evaluated. The key weight parameters and related data on the coupling strength of indicators during the optimization process are cached in the edge strategy library.

[0058] Furthermore, in step S6, based on the real-time predicted task load, edge node computing power, and network bandwidth monitoring results, it is determined whether the current node has reached the elastic migration trigger condition. If it has, the subsequent task migration operation is executed.

[0059] Furthermore, in step S6, the specific steps are as follows:

[0060] Real-time scheduling is performed on the prediction task queue within the edge node. Based on the task queue length, task response latency, and prediction task concurrency industrial scheduling parameters, the prediction task load statistics of the edge node are dynamically obtained, and prediction task load monitoring indicators are generated.

[0061] Based on the predicted task load monitoring index, and combining the current computing power utilization rate and memory storage occupancy rate of edge nodes, a multivariate resource evaluation algorithm is used to calculate the comprehensive computing power status vector of edge nodes.

[0062] Based on the comprehensive computing power state vector of edge nodes, the network bandwidth between the local node and neighboring nodes is continuously monitored. An adaptive bandwidth detection algorithm is used to periodically calculate the effective data throughput between nodes and generate a cross-node network bandwidth availability evaluation index.

[0063] Based on the predictive task load monitoring index, the comprehensive computing power state vector and the network bandwidth availability assessment index, combined with multi-condition decision logic, the elastic migration triggering algorithm is applied to determine whether the edge node has reached the migration criterion and generate an elastic migration triggering signal.

[0064] Based on the generated elastic migration trigger signal, the elastic migration task decision results are output in real time, including the type of multi-indicator collaborative prediction subtask to be migrated, the prediction data stream to be migrated, the current model parameter snapshot, and the migration target node information, which are used to drive the adaptive migration protocol of subsequent tasks.

[0065] Furthermore, in step S7, based on the migration triggering result, a task adaptive migration protocol is used to distribute some or all graph structure fragments and prediction weights to nearby edge nodes or edge clouds, thereby completing the dynamic expansion or contraction of the prediction task and automatically synchronizing the model state.

[0066] Furthermore, in step S7, the specific steps are as follows:

[0067] The real-time resource monitoring data of edge nodes is matched and analyzed with dynamic graph structure fragments. Based on the node CPU / GPU load, storage quota and dynamic graph structure fragment characteristics, the dynamic graph structure fragments that need to be migrated and their corresponding graph neural network prediction weights are determined.

[0068] Based on the selected dynamic graph structure fragments and their corresponding graph neural network predicted weights, fragment transfer requests are generated through a task adaptive migration protocol, and the computing power status and bandwidth conditions of the target's neighboring edge nodes or edge clouds are automatically detected.

[0069] A distributed task scheduling algorithm is used to schedule and allocate the generated shard transfer requests and target edge node or edge cloud capability parameters, and to allocate dynamic graph structure shards and graph neural network prediction weights to the optimal target node.

[0070] The scheduled dynamic graph structure fragments and graph neural network prediction weights are migrated and transmitted through an efficient network protocol, and the graph neural network state parameters during the migration process are synchronized using a model incremental synchronization mechanism.

[0071] After completing dynamic graph structure sharding and graph neural network prediction weight deployment at the target edge node or edge cloud, perform a model prediction engine state consistency check and automatically correct the model state after migration using an end-to-end task signature mechanism.

[0072] Furthermore, in step S8, based on the GPU / CPU load, storage space, and predicted task fluctuations of each node, the computing and storage quotas of the edge nodes are dynamically adjusted to jointly optimize the overall resource utilization and the system's predicted response speed.

[0073] Furthermore, in step S8, the specific steps are as follows:

[0074] The current GPU load, CPU load, and storage space usage of edge nodes are monitored at high frequency, and resource status acquisition algorithms are used to obtain resource utilization indicators in real time.

[0075] Based on the obtained resource utilization rate indicators, combined with the historical scheduling of the predicted tasks and the fluctuation characteristics of the current predicted tasks, the time series autocorrelation analysis method is used to generate the predicted task load estimation parameters.

[0076] The resource utilization rate index and the predicted task load parameters are jointly calculated using an elastic bound constraint model to determine the current task pressure status and idle potential of each edge node, and output elastic resource adjustment suggestions.

[0077] Based on the elastic resource adjustment recommendations, a dynamic resource quota allocation algorithm is adopted to adaptively adjust the GPU / CPU computing quota and storage resource quota of edge nodes in stages, generating the adjusted computing quota distribution and storage quota distribution.

[0078] The adjusted computing and storage quota distributions of each edge node are synchronized with neighboring edge nodes and the edge cloud through a resource scheduling communication protocol. A distributed consensus algorithm is used to optimize the joint resource utilization and form a globally optimal edge-side resource configuration state.

[0079] Based on the adjusted global resource configuration status, the edge multi-indicator collaborative prediction load redistribution action is triggered, and some prediction subtasks are dynamically transferred between nodes through the task migration protocol.

[0080] Furthermore, in step S9, a self-learning optimization mechanism is embedded in the entire multi-indicator collaborative prediction task and resource scheduling process. The model migration parameters and scheduling priorities are automatically corrected based on historical prediction accuracy and resource allocation performance.

[0081] Furthermore, in step S9, the specific steps are as follows:

[0082] Acquire historical multi-indicator collaborative prediction task prediction accuracy data and resource allocation performance indicators of edge nodes to form the input basis for the system's self-learning optimization mechanism;

[0083] Based on the currently collected prediction accuracy data, a multi-objective evaluation algorithm is used to calculate the performance deviation of each prediction task, so as to quantify the difference between the model prediction effect and the historical best index and generate a dynamic evaluation vector.

[0084] By combining the resource allocation performance indicators of each node, the adaptability of the model transfer parameters is estimated using the Bayesian update mechanism, and a set of corrected parameter suggestions is output.

[0085] By combining the dynamic evaluation vector and the set of corrected suggestion parameters, a reinforcement learning optimization algorithm is used to reorder the scheduling priorities of the predicted tasks to obtain a priority allocation decision table.

[0086] Based on the priority allocation decision table, the task migration parameters and scheduling priority parameters are dynamically adjusted and synchronized to the edge node scheduling management module to achieve adaptive real-time optimization of the scheduling scheme and output the iteratively corrected system scheduling parameters.

[0087] We continuously monitor the model migration behavior and prediction accuracy changes of each edge node after self-learning optimization, and use a sliding window statistical detection algorithm to evaluate the effectiveness of the optimization scheme in real time to determine whether to trigger subsequent structural adjustments or further parameter fine-tuning.

[0088] Furthermore, in step S10, when a significant change in process conditions is detected or the predicted anomaly rate increases, the model structure fine-tuning and parameter self-update process is automatically triggered.

[0089] Furthermore, in step S10, the specific steps are as follows:

[0090] For the prediction results output based on edge deployment, continuously monitor the prediction error sequence of key process indicators and process condition data, identify the input conditions of prediction anomaly rate increase or sudden condition events through anomaly detection algorithm, and obtain online model adaptive trigger signals.

[0091] Based on the obtained model adaptive triggering signal, structural sensitivity analysis is performed on the currently operational multi-index coupled dynamic graph neural network structure. The node importance evaluation algorithm and graph structure reconstruction mechanism are applied to determine whether the weight configuration of nodes and coupling edges needs to be adjusted, and a structural optimization scheme for the current process changes is generated.

[0092] Based on the output-based structural optimization scheme, the dynamic graph neural network structure is optimized in real time using a structural fine-tuning mechanism, and the modeling method of multi-index coupling relationship is adjusted to generate a structurally optimized dynamic graph neural network model.

[0093] The generated optimized dynamic graph neural network model is input with historical process index data sequences and the latest collected data. An incremental parameter self-updating algorithm is used to perform immediate correction on the model parameters to obtain the dynamic graph neural network weight set after parameter self-updating.

[0094] By using the obtained parameter-updated dynamic graph neural network weight set, the process multi-indicator data streams collected in real time at the edge are re-predicted, the prediction performance improvement effect is automatically evaluated, and the evaluation results are fed back to the model for continuous optimization.

[0095] Beneficial effects

[0096] This invention addresses the problems of insufficient modeling of dynamic interaction relationships among multiple indicators in existing sintering processes, limited collaborative prediction capabilities, and limited edge elastic scalability. It proposes a method for edge multi-indicator coupling prediction and adaptive dynamic migration based on a cross-time-window graph neural network, which has the following beneficial effects:

[0097] (1) The core of this invention is the introduction of a dynamic multi-index graph structure, which uses key process parameters such as temperature, pressure, and gas content as nodes, and dynamically assigns weights to the edges between nodes based on physical constraints, historical correlations, and time-series causal relationships. Through a sliding window and dynamic graph mechanism, the coupling and decoupling characteristics of various indicators over time during the process can be dynamically captured. Compared with traditional single-index or simple multi-index regression models, this invention can significantly improve the ability to perceive the coordinated change trend of multiple indicators. Experimental results show that under multi-factor perturbation conditions, it can reduce the mean square error of prediction for each key parameter, effectively solving the problem that existing technologies are insufficient in characterizing the deep causal and coordinated relationships of multiple parameters.

[0098] (2) This invention deploys a lightweight graph neural network inference engine at edge nodes, fully utilizing the dynamic graph structure to efficiently aggregate information about the nodes themselves and their neighborhoods, forming associative embedding vectors. This not only enhances the model's spatial-temporal modeling capabilities but also promotes the full extraction and utilization of mutual information among various process indicators. While maintaining low model inference latency (less than 100ms per batch), it significantly improves the accuracy and response rate of multi-indicator joint prediction. The node embedding vectors can be used for subsequent process diagnosis and optimization decisions, enhancing the system's interpretability and industrial adaptability.

[0099] (3) The invention achieves real-time adaptation of the model to changes in process conditions by synchronously outputting multi-index mutual information parameters and dynamically adjusting the edge weights and node weights within the GNN. It automatically optimizes network connections and feature fusion methods, significantly enhancing the robustness of the system under complex environments such as load mutations, operating condition drift, and abnormal fluctuations. Compared to static structural models, it can track the evolution of coupling modes in real time, effectively avoiding performance degradation caused by process transformation or extreme disturbances.

[0100] (4) This invention incorporates a historical task self-learning optimization mechanism and a prediction anomaly adaptive reconstruction strategy. Based on actual prediction performance and resource scheduling effectiveness, it can automatically correct model migration parameters, predict load, and iteratively optimize the priorities of each stage. When prediction deviations or sudden changes in operating conditions are detected, the system can proactively trigger model structure fine-tuning and parameter incremental self-updates, avoiding human intervention and ensuring the system continuously adapts to dynamic changes in production conditions, achieving long-term evolution and autonomous management of prediction performance. This is particularly important in industrial settings where unattended operation and self-healing capabilities are required. Attached Figure Description

[0101] Figure 1 This is a flowchart of a real-time optimization method for the sintering process based on edge computing according to the present invention. Detailed Implementation

[0102] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0103] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0104] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0105] Please see Figure 1 As shown, a real-time optimization method for the sintering process based on edge computing includes:

[0106] S1. Obtain the raw multi-index time series data collected in real time on the edge side of the sintering process. The multi-index includes key process parameters such as temperature, pressure, and gas content. For the sensing points at different sections of the sintering production line, record the sampling data and their corresponding process labels respectively.

[0107] S2. Preprocess the acquired multi-indicator time series data, including outlier removal, normalization, and missing data interpolation, to generate a set of data stream features for the input of the adaptive graph neural network.

[0108] S3. Based on the process physical constraints and historical statistical correlation of different work sections, a dynamic graph structure is constructed with each key indicator in the multi-indicator time series data as a node and the coupling relationship as an edge. The preprocessed data flow features are mapped to this dynamic graph structure to realize dynamic multi-indicator distribution modeling.

[0109] S4. Input the dynamic graph structure into a lightweight graph neural network deployed at the edge, perform multi-indicator collaborative prediction, and output the predicted value distribution and mutual information parameters of each key indicator within a predetermined future time window.

[0110] S5. Based on the mutual information parameters output by S4, dynamically evaluate the coupling strength between process indicators, and optimize the weights of the internal structure of the graph neural network accordingly to achieve adaptive adjustment of feature relationships in response to actual process changes.

[0111] S6. Based on the real-time predicted task load, edge node computing power and network bandwidth monitoring results, determine whether the current node has reached the elastic migration trigger condition. If it has, execute the subsequent task migration operation.

[0112] S7. For migration trigger results, a task adaptive migration protocol is used to distribute some or all graph structure fragments and prediction weights to nearby edge nodes or edge clouds, to complete the dynamic expansion or contraction of the prediction task and automatically synchronize the model state.

[0113] S8. Implement a flexible resource scheduling strategy. Based on the GPU / CPU load, storage space, and predicted task fluctuations of each node, dynamically adjust the computing and storage quotas of edge nodes to jointly optimize the overall resource utilization and system prediction response speed.

[0114] S9. Embed a self-learning optimization mechanism into the entire multi-indicator collaborative prediction task and resource scheduling process. By using historical prediction accuracy and resource allocation performance, the model migration parameters and scheduling priorities are automatically corrected to improve the system's robustness and adaptability.

[0115] S10. When a significant change in process conditions or an increase in the prediction anomaly rate is detected, the model structure fine-tuning and parameter self-update process is automatically triggered to ensure that the multi-index coupled prediction model continuously adapts to the actual operating environment of sintering production.

[0116] S1 involves acquiring raw multi-indicator time-series data collected in real-time from the edge of the sintering process. These multi-indicators include key process parameters such as temperature, pressure, and gas content. For sensor points at different locations within the sintering production line, sampling data and their corresponding process labels are recorded, specifically including:

[0117] S1.1 Perform data acquisition operations on various types of industrial-grade sensing points (including temperature sensors, pressure sensors and gas composition analyzers) deployed in each section of the sintering production line to obtain the original time-series signals of key process parameters and form a basic multi-index sensing signal stream.

[0118] The data collection targets are various industrial-grade sensing points deployed at different times and locations in the sintering production line, mainly including temperature sensors, pressure sensors, and gas component analyzers, specifically involving multiple types of intelligent sensing units such as platinum resistance thermometers, diffused silicon pressure transmitters, and infrared laser gas analysis equipment.

[0119] A distributed data acquisition and control method is adopted (configuration parameters cover sensor type, sensitivity level and sampling interval) to realize local signal acquisition and preliminary steady-state control of various types of industrial sensors;

[0120] Furthermore, through a multi-channel analog / digital signal conditioning module (including filtering amplification, anti-interference circuit and AD conversion unit), the original electrical signal output from the sensing point is subjected to noise filtering, amplification and standardization processing to obtain a digitized original signal sequence of process parameters;

[0121] Furthermore, an event-driven data acquisition algorithm (with parameters set to adaptively switch between high-frequency polling and mutation detection modes) is adopted to acquire the digital sampling stream after signal conditioning in batches at different time periods, and to bind or label the sensor source point information according to the preset work section rules.

[0122] Furthermore, by utilizing the time-series caching module of the local high-reliability edge acquisition terminal, the conditioned and labeled multi-type sensor signal streams are cached in priority queues and summarized into stable basic multi-index sensor signal streams to facilitate subsequent data sampling identification, archiving and multi-modal feature integration.

[0123] Through the above-mentioned distributed acquisition, signal conditioning and timing caching process, the original timing information of key process parameters in each section is transformed into a structured, complete integrated multi-index sensing signal stream, so as to realize high-quality, low-latency, localized data acquisition and sensing foundation guarantee for complex process parameters in the sintering site.

[0124] For example, a large sintering plant deploys 100 industrial-grade temperature sensors (PT100 platinum resistance thermometers, accuracy ±0.2℃, sampling interval 0.1s), 50 pressure sensors (diffused silicon type, range 0~10MPa, sampling accuracy ±0.01MPa), and 30 infrared gas composition analyzers (CO / CO2 detection limit 0.01%, sampling period 0.2s) in the calcination, cooling, and discharge zones. A supporting edge-based field acquisition module prioritizes sensor sensitivity, and the signal conditioning board uses a third-order low-pass filter and a 24-bit Σ-Δ A / D converter to perform real-time analog-to-digital sampling. An event-driven capture algorithm switches between batch acquisition modes based on a mutation detection threshold of 3σ. After local redundant caching, a structured multi-indicator raw signal stream data packet containing various indicators such as temperature, pressure, and gas content is generated every 5 seconds. Tests showed that during 24 hours of continuous operation, the packet loss rate was less than 0.01%, and the signal-to-noise ratio was greater than 40dB, ensuring high reliability and adaptability for subsequent spatial positioning, multi-label annotation, and data archiving operations. This step significantly improved the accuracy of raw parameter capture and the foundation for collaborative modeling of multiple indicators in complex production scenarios.

[0125] S1.2. Based on industrial Ethernet or a customized acquisition channel, perform high-frequency real-time sampling of the sensor signal stream, and use a timestamp and node numbering mechanism to uniquely identify each original sampled data to obtain a multi-indicator original data packet with sampling time and spatial positioning.

[0126] For the various types of industrial-grade sensor signal streams buffered by step S1.1, the input data includes digital raw signal sequences of key process parameters such as temperature, pressure, and gas content. Each signal has a preliminary sensor point identifier and a local time cache index.

[0127] Employing industrial Ethernet communication protocols or customized high-speed data acquisition channels, with parameter settings including data packet bandwidth ≥100Mbps and transmission delay ≤10ms, it enables high-frequency real-time acquisition of various sensor signals and real-time acquisition of sampling signal packets;

[0128] Furthermore, through the high-precision clock synchronization mechanism built into the edge nodes, using the IEEE 1588PTP time synchronization protocol, the clock drift error is controlled to ≤1ms, and a unique timestamp T is automatically assigned to each sampled data stream. i (i is the data sequence number), to ensure standardized time positioning of sensor data;

[0129] Furthermore, through a node numbering allocation mechanism (using a fixed numbering strategy or dynamic DHCP assignment, node number N) j This enables a unique spatial or device mapping between each piece of raw sampling data and a physical sensor node;

[0130] The algorithm employs a multi-index data structure encapsulation method, which encapsulates each sampled data point according to its sampling timestamp T. i With node number N j The raw data package consists of the following multi-indicator data:

[0131] D i,j ={V temp V press V gas ,T i N j}

[0132] Where V temp V press V gas These represent quantitative signal sample values ​​for temperature, pressure, and gas concentration, respectively.

[0133] Furthermore, data consistency verification algorithms (such as CRC-32 and hash value verification) are used to verify the integrity of data packets, eliminate damaged or inconsistent signal segments, and retain only complete multi-index original data packet entries with valid spatiotemporal identifiers.

[0134] Through the combined processing of high-frequency real-time sampling, spatiotemporal unique identification, and integrity verification, the complex, distributed, multi-source asynchronous process raw signal stream is transformed into a structured multi-index raw data packet with spatiotemporal positioning characteristics, providing highly reliable raw data input for subsequent location mapping, process label annotation, and archiving management.

[0135] For example, a platinum resistance temperature sensor deployed in the sintering zone of the sintering plant is subjected to high-frequency signal extraction via industrial Ethernet at 0.1s intervals, and the signal is precisely synchronized via IEEE 1588 at the edge to assign a sampling timestamp.

[0136] T 1001 =2024-06-01, 12:00:05.100. Node numbers are fixedly assigned as N by the data acquisition and control system. 37 This indicates the 37th temperature sensing unit. When the sampled value V... temp=1123.4, ℃, V press =1.026 MPa

[0137] V gas =0.143%, after CO is read in real time, the data is structured as follows:

[0138] D 1001,37 ={1123.4,1.026,0.143,2024-06-01,12:00:05.100,37};

[0139] After successful CRC-32 verification, the data packet is incorporated into the original multi-indicator dataset, reliably completing the spatiotemporal positioning of the indicators at that sampling moment. For network events that interrupt sampling operations within 24 hours, a mechanism using independent node numbers and timestamps ensures complete data archiving, reducing the packet loss rate to below 0.01%, achieving seamless traceability and real-time recording of process big data. This processing effectively completes the high-frequency real-time identification encapsulation and input of the original multi-indicator sampling stream, laying a solid foundation for subsequent multi-label spatial annotation and process semantic binding.

[0140] S1.3. The original data packets are mapped to their locations using the sintering process flow database. Based on the section information of the sensing point, the physical section labels of the process are automatically associated to realize the spatial semantic annotation of the original data.

[0141] S1.4. For multi-index raw data packets with process physical section labels, embed the process real-time status code and operation team identifier. Through adaptive data label mapping, write the operation context information into the raw time series data to achieve dynamic coupling between data and process scenario.

[0142] S1.5. For the original time series data of multiple indicators with completed location and label binding, the high-speed caching mechanism of edge computing nodes is used to implement automatic archiving in different time periods, generating original time series database entries with multiple process labels to meet the input requirements of subsequent preprocessing, graph structure mapping and collaborative modeling.

[0143] Step S2 involves preprocessing the acquired multi-indicator time series data, including outlier removal, normalization, and missing data interpolation, to generate a data stream feature set suitable for the input of the graph neural network. Specifically, this includes:

[0144] S2.1. Perform quality inspection on the collected multi-indicator time series data, and mark all outlier data points that exceed the normal fluctuation range of process parameters based on the statistical threshold method, so as to remove abnormal interference data caused by sensor failure, communication packet loss or extreme working conditions, and achieve high-quality raw dataset screening.

[0145] S2.2 For the multi-index time series data after outlier removal, the interval scaling normalization algorithm is used to uniformly transform various process parameters (such as temperature, pressure, and gas content) to the preset dimensionless interval, so that the value range of different indicators can be standardized, so that the graph neural network can process and fuse the features of multiple heterogeneous process parameters simultaneously.

[0146] S2.3 For missing records in normalized multi-indicator time series data, apply time series interpolation algorithms (such as linear interpolation or sliding window mean interpolation) to fill in the gaps with reasonable numerical inferences, so as to reconstruct a complete and continuous multi-indicator data stream and prevent the continuity and prediction accuracy of graph neural network modeling from being affected by incomplete samples.

[0147] S2.4. For the complete multi-index data stream obtained after outlier removal, normalization and missing data interpolation, the sliding time window segmentation method is adopted to decompose the data stream into equal-length window segments according to the process attributes. A window-level feature vector with time recursion is generated for each window segment so that the dynamic graph structure model can utilize time correlation.

[0148] S2.5. Associate the feature vectors of the multi-indicator data stream after window segmentation with the corresponding process labels, and generate a set of data stream features for the input of the adaptive graph neural network based on the data mapping rules, so as to provide standardized and structured input features for subsequent dynamic graph structure construction and multi-indicator coupling prediction tasks.

[0149] S3, based on the process physical constraints and historical statistical correlations of different work sections, constructs a dynamic graph structure with key indicators in the multi-indicator time series data as nodes and coupling relationships as edges. The preprocessed data flow features are mapped to this dynamic graph structure to achieve dynamic multi-indicator distribution modeling. Specifically, this includes:

[0150] S3.1. Based on the process physical constraint model, determine the index partitioning and attribution of the preprocessed multi-index data stream features to obtain the physical distribution labels of each key index in different sections of the sintering production line.

[0151] S3.2 For process indicators that have been assigned to different partitions, use the historical process database and historical data collection logs from the edge to perform correlation coefficient matrix calculation and Granger causal analysis to extract statistical correlation parameters among key indicators.

[0152] Using the multi-index data stream characteristics determined by the process physical constraint partitioning as input, for the time series data sequences of each key process index in each physical segment, the cointegration test and Pearson correlation coefficient calculation method (parameters include time window length setting and historical reference period selection) are used to characterize the underlying correlation features between major process parameters such as temperature, pressure, and gas content.

[0153] Furthermore, the existence of long-term equilibrium relationships among key indicator sequences is determined by the cointegration test method. The Johansen cointegration test algorithm is introduced to extract cointegration vector sets that can be used to model the coordinated change trend among indicators, thereby obtaining long-term stable and related process parameter pairs.

[0154] The Pearson correlation coefficient was used to quantify the linear correlation strength between the main indicator pairs according to the following formula:

[0155]

[0156] Where, x i With y i Let represent the measured values ​​of two different process parameters at time i. and Its mean, n is the number of sampling points within the time window, and r xy The correlation coefficient is used to construct the statistical linear correlation matrix between pairs of indicators.

[0157] Furthermore, the Granger causal analysis method was applied, and based on the historical process database and locally stored multi-index segment samples, a multi-step lag VAR (vector autoregression) model was used to examine the temporal causal driving relationship between the key indicators.

[0158] Specifically, the Granger causality test statistic is calculated using the following formula:

[0159]

[0160] Among them, RSS r RSS represents the sum of squared residuals of a constrained model. ur Let m be the sum of squared residuals of the unconstrained model, m be the lag order, and n be the number of samples. This enables the determination of the significance of the time-series driving effects between each pair of process parameters.

[0161] The cointegration coefficient matrix, correlation coefficient matrix and Granger causality significance test results obtained above are merged and archived to automatically generate a set of statistical correlation parameters between each pair of key indicators, including linear correlation strength, long-term synergistic trend weight and time-series causality indicators.

[0162] Through the above parameterization process, the multi-indicator data of the process after attribution determination are effectively transformed into formal statistical correlation parameters, which provide criteria for subsequent dynamic graph node generation and coupling relationship edge weight labeling, and realize the structured extraction of coupling patterns between multi-source heterogeneous process data.

[0163] For example, in the cooling section of a sintering machine in a steel sintering plant, the monitoring period is 30 days. Three key indicators—temperature, pressure, and CO concentration—are selected, with a data collection period of 1 minute. A total of 43,200 time-series samples are assembled from the historical process database and local data collection logs. The Johansen cointegration test is performed on the 30-day samples, using a cointegration rank of 1. A significant long-term synergistic relationship between temperature and CO concentration is detected, and the cointegration vector [1, -0.85] is extracted accordingly. Pearson correlation coefficients are calculated pairwise for each of the three indicators, with a window length of n = 43,200. The results are: temperature-pressure r = 0.72, temperature-CO concentration r = 0.89, and pressure-CO concentration r = 0.58, all reaching a significance level of 0.01. A VAR model with a lag order of m = 4 is used, and the Granger causality test is employed to determine the bidirectional time-series influence of temperature and CO concentration. The F-statistic F0 is calculated. xy =13.27, indicating a significant Granger causal relationship between CO concentration and temperature sequence. Finally, the automatic archiving of this batch of temperature-CO concentration cointegration weights (0.85), Pearson correlation coefficients (0.89), and Granger causal significance (p<0.01) provides a structured statistical basis for subsequent construction of dynamic graph node sets and edge weight allocation for this segment, achieving high-precision, dynamic adaptive modeling of multi-index segment coupling relationships.

[0164] S3.3. Combining the process physical constraint model and historical statistical correlation parameters, generate node sets for all key indicators, and label the initial coupling relationship edge weights for indicator pairs with physical or statistical correlations to form a basic static graph structure topology.

[0165] Using the multi-indicator data stream features and parameterization results of attribution determination and statistical correlation parameter extraction as input objects, the node generation and initial edge weight assignment of the basic dynamic graph structure are carried out for the dataset.

[0166] A node generation method (input parameter: a set of multiple indicators determined by process physical constraint partitioning) is adopted to define the nodes of key indicators such as temperature, pressure, and gas content in each process physical section, forming an initial node set covering the entire sintering process.

[0167] By using edge generation and edge weight labeling methods (input parameters: process physical constraint model, cointegration vector group, Pearson correlation coefficient matrix, Granger causality analysis parameters), the coupling existence of index pairs with physical or statistical correlation is determined, and the correlation edges between index nodes are generated sequentially.

[0168] Furthermore, using a weight initialization algorithm, for index pairs with direct physical constraints (such as upstream and downstream control relationships or spatial coupling states) and statistical cointegration, high-strength correlation, or significant Granger causality, the weight strength is assigned according to the following weight calculation formula:

[0169]

[0170] in, Let C be the initial edge weight between nodes i and j. ij For the weights of the cointegrated vector, r ij Let I be the Pearson correlation coefficient. ij α represents the Granger causal significance coefficient, and α, β, γ are optimization ratio factors set by business scenarios, process sensitivity, and model prior experience.

[0171] Based on the edge weight assignment results, the edge weight matrix is ​​initialized for all key index node pairs that can be determined to be coupled and related, and the node-edge structure is registered to form the basic static graph structure topology.

[0172] A structural legality check and archiving mechanism is adopted to verify the connectivity and non-redundancy of the generated node set and initial edge weight matrix, ensuring that the graph structure adapts to subsequent dynamic graph adaptive adjustments;

[0173] Through the above-mentioned whole-chain processing method of node generation, edge generation and edge weight assignment driven by physical / statistical correlation, and structural legality archiving, the multi-indicator data of the process is transformed into the basic static graph structure topology, realizing the formal structural modeling of the coupling relationship of multiple indicators within the sintering process, and providing structured prior input for real-time adjustment of dynamic graph and collaborative prediction of multiple indicators.

[0174] For example, in the cooling section T1-T3 of a steel sintering plant, three types of sensors are used: temperature (node ​​1), pressure (node ​​2), and CO concentration (node ​​3). The cointegration weights are C_12 = 0.80, C_13 = 0.85, and C_23 = 0.65, respectively. The Pearson correlation coefficients are r_12 = 0.72, r_13 = 0.89, and r_23 = 0.58, respectively. The Granger causality coefficients are I_12 = 0.67, I_13 = 1.00, and I_23 = 0.55, respectively. The scaling factors are α = 0.4, β = 0.4, and γ = 0.2. The initial edge weights are calculated pairwise for nodes 1-3:

[0175]

[0176] Three nodes (nodes 1-3) and three directed edges (each edge labeled with initial weights of 0.742, 0.896, and 0.602) are generated to construct the basic structure of a static graph. After archiving and verifying the connectivity of the structure, the initial graph topology is stored in the database. In actual deployment, this basic structure serves as the input for subsequent dynamic graph edge weight adjustment and collaborative prediction algorithms, enabling refined coupled modeling of highly correlated segments with multiple indicators and real-time data-driven structure adaptation.

[0177] S3.4 Based on the newly acquired data stream feature set, a sliding time window mechanism and a dynamic edge weight assignment algorithm are applied to update the edge weight strength in the static graph structure in real time, thereby realizing the time-series adaptive adjustment of the dynamic graph structure.

[0178] S3.5. Map the data stream feature set covered by the sliding time window at the current moment to the node attributes based on the above dynamic graph structure to realize the one-to-one mapping and relationship constraint between the multi-indicator time series features and the nodes of the dynamic graph.

[0179] The feature set of multi-indicator data streams (including parameters such as temperature, pressure, and gas composition) dynamically collected within the coverage area of ​​the sliding time window is used as the input object for node attribute mapping;

[0180] A node attribute mapping method (parameters include node set, node spatial distribution label, and time window level multi-index feature vector) is adopted to assign a time-series feature vector corresponding to its physical process segment, sensing type and sampling time to each dynamic graph node, thereby realizing the structured injection of node attributes.

[0181] Furthermore, through a one-to-one mapping rule for multiple indicators (based on the forced matching logic of node encoding, process section labels and feature vector identifiers), it is ensured that all time-series features of multiple indicators within the sliding window can be accurately mapped to their corresponding nodes, thereby realizing the dynamic binding of multiple indicator data and node structure.

[0182] Furthermore, by using a relation constraint embedding algorithm (parameters including edge weight matrix, node attribute vector, and coupling correlation threshold), the coupling relationship parameters between feature vectors are mapped to attribute constraints between nodes, ensuring that the physical and statistical dependencies between nodes can be strictly reflected in the graph structure during dynamic attribute updates.

[0183] Furthermore, a node attribute temporal coverage verification algorithm (parameters include window length, time alignment step size and node data coverage threshold) is adopted to automatically verify the integrity and consistency of attribute mapping of each node within the current sliding time window, identify uncovered, abnormal or delayed node attributes, and achieve reliable assurance of node attribute mapping.

[0184] Through the above node attribute mapping, feature relationship constraint management and verification mechanism, the multi-indicator feature data covered by the sliding window is efficiently and securely embedded into the dynamic graph structure, realizing a one-to-one correspondence between the multi-indicator time series and the nodes and coupling edges of the dynamic graph and linking them with dynamic constraints, providing a strictly structured input for subsequent distributed GNN collaborative prediction;

[0185] For example, temperature, pressure, and oxygen content sensors are installed at stations T1, T2, and T3 of the sintering machine ignition section in a steel plant. The sliding window width is 60 minutes, the sampling period is 1 minute, and each node has 60 feature samples. During the node attribute mapping stage, taking temperature node 1 (T1), pressure node 2 (T2), and oxygen node 3 (T3) as examples, a 60-dimensional sliding window feature vector is injected into each node sequentially. Based on the process section label and feature timestamp, node 1 is mapped as T1-temperature-time series, node 2 as T2-pressure-time series, and node 3 as T3-oxygen concentration-time series. The relational constraint embedding algorithm, based on a preset coupling correlation threshold of 0.65, maps the coupling features between nodes 1 and 3 (such as the correlation between temperature and oxygen concentration) as constraint edges between node attributes, with a dynamic edge weight of 0.88. The node attribute time series coverage verification threshold is set to 95%, and edge database completion is automatically initiated for nodes with missing samples, achieving a 100% attribute mapping rate for each main node within the 60-minute window. The final output is a dynamic graph structure input with 60-dimensional node-level features, complete relational constraints, and consistency verification results, providing full and structured data support for lightweight GNN operations;

[0186] S3.6 For the dynamically adjusted dynamic graph structure, based on the distributed consensus protocol of edge nodes, verify the consistency of the dynamic graph structure of multiple nodes, and synchronize the dynamic graph structure to all edge nodes that need to participate in the collaborative prediction of multiple indicators.

[0187] Step S4 involves inputting the dynamic graph structure into a lightweight graph neural network deployed at the edge, performing multi-indicator collaborative prediction, and outputting the predicted value distribution and mutual information parameters of each key indicator within a predetermined future time window. Specifically, this includes:

[0188] S4.1. Using the dynamic graph structure constructed in step S3 as the input object, call the lightweight graph neural network (GNN) inference instance that has been deployed locally on the edge node to complete the dynamic graph input data format adaptation and model initialization, so as to provide a structured feature data stream for the subsequent multi-indicator collaborative prediction process.

[0189] The dynamic graph structure obtained by the S3 step is used as the input object. The inputs include multiple types of key process index nodes (such as temperature, pressure, gas composition, etc.) in different sections of the sintering production line and the coupling edge weights between each node after dynamic assignment based on process physical constraints and historical statistical correlation parameters.

[0190] A dynamic graph input data format adaptation method (parameters include: multi-dimensional feature vectors of node attributes, adjacency matrix of edge relationships, edge weight strength values, and sliding time window labels) is adopted to automatically map the original process multi-indicator time-series data and their spatial attribution labels to the standardized input format required by the lightweight GNN inference engine. The node attribute structure is arranged as follows: Where N represents the number of indicator nodes, F is the feature dimension to which each node belongs, and the adjacency matrix... Each element in the middle is determined by the edge weight w ij Dynamic settings;

[0191] Furthermore, by using model input verification and graph structure integrity checking algorithms (parameters include multi-index node coverage, edge connectivity threshold, and consistency identifier), the validity and consistency of dynamic graph structure input are verified. Error prompts and repair suggestions are automatically given for data format defects, abnormal node / edge mapping, or insufficient temporal coverage, ensuring the structural integrity of the initialization of inference instances.

[0192] Furthermore, through the local lightweight GNN inference engine scheduling mechanism on the edge node (which relies on containerized deployment at the edge, automatic task queue allocation, chip resource identification, and model loading parameters), the GNN model weights θ allocated to this node are invoked. GNN The system includes a structure description file, which initializes the inference task for the verified dynamic graph structure input, and determines the preconditions for operation such as model parameter loading, graph structure locking, and inference window step size setting.

[0193] Furthermore, a lightweight GNN inference data stream construction mechanism is used (parameters include: node feature stream X). t Structured Tag Stream A t Sliding time window step T w (Definition of the inference target set) After the data stream and structure stream inputs are standardized, the dynamic graph structure of the current process multi-index and the description of the specified prediction target are integrated into a structured and reusable data stream input set, providing an efficient parallel computing data layer for subsequent GNN graph signal processing and collaborative multi-index prediction;

[0194] Through the above-mentioned technology chain, we can achieve seamless integration and standardized format reorganization of the dynamically adjusted multi-indicator distribution relationship structure, meet the requirements of edge-side GNN inference engine for input feature integrity, node attribute diversity and relationship structure rationality, and lay a solid data and structural foundation for multi-indicator collaborative prediction calculation.

[0195] For example, in the ignition and cooling sections of a sintering machine in a steel metallurgical plant, three types of nodes—temperature, pressure, and CO concentration—are deployed respectively. Node features are captured using a 60-minute sliding window with a 1-minute sampling period; therefore, each node is injected with a 60-dimensional temporal feature vector, and the number of nodes N=3. Based on the edge weights obtained in the preceding steps, A in the adjacency matrix A... 12 =0.742, A 13 =0.896, A 23 =0.602. The node set X consists of three sets of 60-dimensional feature vectors: temperature T1, pressure T2, and CO concentration C1. After input formatting, it is automatically adapted to... Data input consistency verification was performed, ensuring 100% coverage of nodes and edges, no structural redundancy, and that the input passed verification. The edge GNN inference container instance is configured with 2 CPU cores, 256MB of memory, and model weights θ. GNN Loading time less than 1 second, input data window width T w The value was set to 60. Finally, the parameterized dynamic graph structure data stream was successfully input into the local GNN inference instance, and subsequent information aggregation and feature embedding calculations began, significantly improving the immediacy of edge prediction and the model's adaptability under the drive of multi-index coupling relationships;

[0196] S4.2 Based on the input dynamic graph structure, the GNN information aggregation algorithm is used to perform graph signal transmission and feature embedding operations on each key indicator node and its adjacent relationship edges, generating the correlation enhancement node feature embedding vector of each indicator node in the current time series, laying an interpretable foundation for subsequent parallel prediction operations.

[0197] The dynamic graph structure input after processing in step S4.1 (including the multidimensional feature vectors of nodes, adjacency matrix and edge weight strength) is used as the processing object in this step;

[0198] The GNN information aggregation algorithm is used (parameter: node feature set). Graph structure adjacency matrix With a preset aggregation layer number L), feature iterative propagation and aggregation of each key process indicator node and its adjacent relationship edges are realized;

[0199] Through message passing, for each layer l = 1, 2, ... L, based on the characteristics of the current node... Its neighboring nodes The following information aggregation formula is used for recursion:

[0200]

[0201] in, For the feature embedding of node i at layer l, wij Let W be the edge weight between nodes i and j. (l) Let b be the linear transformation weight matrix for this layer. (l) For bias, Let σ be the set of neighbors of node i, and let σ be the activation function.

[0202] Furthermore, through multi-layer superimposed feature aggregation iteration (usually L=2 to 4 layers), the feature itself and neighborhood information are gradually integrated on the graph structure, so that the feature embedding vector of each process index node contains the process state information and coupling relationship weight of its spatial neighboring nodes.

[0203] By introducing relation-aware normalization (such as edge weighted normalization), edge information is corrected in the following ways:

[0204]

[0205] and use w in the alternative information aggregation formula ij This ensures that the feature aggregation bias caused by the difference in correlation coefficients of different indicators can be effectively corrected, thereby improving the scientific nature of feature fusion among multiple indicators.

[0206] By performing node-level feature normalization and relation normalization, the final correlation-enhancing feature embedding vector for each key indicator node is generated. Provides structured inputs for parallel multi-indicator forecasting;

[0207] Through the above information aggregation and normalization embedding processing, the node features and their adjacency coupling information in the dynamic graph structure are fully injected into the embedding space, thereby improving the feature expressiveness and the interpretability of the relation structure, and providing a high-quality input foundation for subsequent distributed and collaborative multi-index prediction calculations.

[0208] For example, in the dynamic diagram structure of the cooling section of a steel plant, the three key process nodes correspond to temperature, pressure, and CO concentration, respectively. The input feature of each node is a 60-dimensional sampling window vector, and the edge weight in the adjacency matrix A is A. 12 =0.742, A 13 =0.896, A 23 =0.602. A two-layer feature aggregation using a GNN is employed, with the first layer weights... Second floor The activation function is ReLU. After edge-weighted normalization, the following embedding propagation is performed continuously for each node: The first layer sums the neighborhood features of each node with weights, performs a linear transformation, and adds activation, outputting a 32-dimensional intermediate embedding; the second layer further aggregates industry relationships and outputs a 16-dimensional final node embedding vector. In actual deployment, the final embedding vector for the temperature node is... This method integrates the feature information and coupling edge weights of the local node, pressure node, and CO concentration node. Through visualization analysis of node embedding vectors, the collaborative relationship between nodes is effectively characterized. When subsequently input into a multi-task prediction network, it achieves an improvement of approximately 12.5% ​​in prediction mean square accuracy compared to the baseline method, and significantly improves the interpretability of feature interactions between nodes.

[0209] S4.3. Using the node feature embedding vectors mentioned above, a lightweight multi-task prediction sub-network is driven. Based on the sintering process prediction window parameters, parallel multi-index distribution prediction calculations are performed on each key index node (including temperature index, pressure index, gas content index, etc.) to obtain the data structure of the predicted value distribution of each index within the future predetermined time window.

[0210] S4.4. Based on the distribution results of the predicted values ​​of multiple indicators obtained in step S4.3, and combining node feature embedding and dynamic graph structure, the mutual information calculation module is used to employ the edge mutual information estimation algorithm to quantitatively analyze the information gain and joint distribution relationship between all key indicator node pairs, and output the mutual information parameter matrix results between each indicator pair to reveal the strength of the time-varying coupling relationship between indicators.

[0211] The multi-index predicted value distribution results output from step S4.3 and the final feature embedding vector of each node are used as input objects. The objects involved include the future temporal distribution of each key process index node in the sintering production process, the dynamic graph structure coupling information between nodes, and the node-level feature aggregation representation.

[0212] Marginal mutual information estimation algorithm is used (parameter: predicted value distribution results) Node feature embedding {h i}, adjacency matrix A of the dynamic graph, time window width T w This enables joint distribution modeling of prediction results for each pair of key indicator nodes;

[0213] Furthermore, the probability density function kernel estimation method (parameter: predicted value sequence) is used. With an adaptive bandwidth σ for the window width, a continuous probability distribution model is performed on the predicted value sequence of each node and the joint predicted value sequence of node pairs, and the marginal probability and joint probability of each node and node pair are calculated.

[0214] Furthermore, using the mutual information calculation formula, the information gain parameter MI between the prediction results of all indicator node pairs (i,j) is calculated. ij The formula for calculating mutual information is as follows:

[0215]

[0216] in, and Let be the marginal probability densities of the predicted values ​​for nodes i and j, respectively. This represents the corresponding joint probability density;

[0217] Furthermore, by recursively applying a sliding time window, the most recent T is... w Within a given time period, all nodes perform dynamic evolution calculations of mutual information to form a complete mutual information parameter matrix within the current window. Where N is the number of key process indicator nodes, each MI ij Reflects the strength of coupling between node pairs in the current time series;

[0218] Furthermore, by integrating node feature embedding information with dynamic graph structure, the MI matrix is ​​normalized through parameter normalization and outlier removal algorithms. Invalid mutual information terms with confidence levels below the threshold are removed, and a high-confidence mutual information parameter matrix is ​​output in a structured form, providing quantitative input for subsequent coupling strength evaluation and adaptive adjustment of model weights.

[0219] By using the above-mentioned edge mutual information estimation and dynamic matrix archiving processing method, the distribution of multi-indicator prediction values ​​and the embedding results of node features are transformed into a mutual information parameter matrix that reflects the coupling strength between key process indicators in real time, thereby achieving accurate quantification of the time-varying coupling strength between indicators and significantly improving the scientific nature of collaborative prediction and the system's adaptive capability.

[0220] For example, in the sintering and cooling section of a steel plant, 60-dimensional predicted value distributions are obtained for the temperature node (T), pressure node (P), and CO concentration node (C) within a 60-minute prediction window. Using the Gaussian kernel density estimation method with a bandwidth σ = 5.0, the marginal probabilities and joint probabilities of the predicted value sequences for the three node pairs (T,P), (T,C), and (P,C) are calculated. The mutual information calculation formula is then applied to obtain the MI. TP =0.37, MI TC =0.62, MI PC =0.29. The MI matrix is ​​normalized using a confidence threshold of 0.25, low-confidence terms are removed, and a 3×3 mutual information parameter normalization matrix is ​​output. This matrix is ​​passed as a structured quantitative indicator to the subsequent coupling strength assessment and GNN weight adaptive adjustment module, realizing traceable modeling of dynamic coupling strength among multiple indicators and optimization of actual control effect;

[0221] S4.5. Package the obtained multi-index prediction value distribution results and mutual information parameter matrix into a standardized structured output to support the dynamic coupling strength evaluation and GNN internal weight adaptive adjustment in the subsequent S5 step, and at the same time provide real-time input for real-time digital twin, optimization control and dynamic elastic migration of the sintering production process.

[0222] S5, based on the mutual information parameters output by S4, dynamically evaluates the coupling strength between process indicators and optimizes the weights of the internal structure of the graph neural network accordingly, achieving adaptive adjustment of feature relationships in response to actual process changes. Specifically, this includes:

[0223] S5.1. Real-time capture and structured processing of the mutual information parameters output by the pre-graph neural network collaborative prediction module are performed to obtain the mutual information distribution matrix between different key process indicators, providing basic data support for subsequent quantitative evaluation of coupling strength.

[0224] S5.2 Based on the mutual information distribution matrix, the coupling coefficient calculation algorithm based on information theory is applied to dynamically quantify the coupling strength between each pair of key process index nodes and output the process index coupling strength matrix to accurately depict the evolution of the collaborative relationship between multiple indicators.

[0225] S5.3. Using the above-mentioned process index coupling strength matrix as input, the weighted adaptive update strategy is used to dynamically optimize and adjust the node weights and edge connection weights inside the graph structure of the graph neural network, so as to realize the adaptive reconstruction of the importance and relationship of index nodes and improve the sensitivity of the model to actual process changes.

[0226] S5.4 After completing the weight adjustment, based on the updated graph neural network structure, retrain the lightweight model at the edge end and perform rapid incremental learning on the multi-indicator feature set collected in the recent time window to achieve synchronous adaptive updating of structural weight optimization and model prediction capability.

[0227] S5.5 Continuously monitor the structural weight optimization results dynamically. By comparing them with real-time prediction accuracy and historical performance indicators, evaluate the effectiveness of the current graph neural network optimization adjustment. Also, cache the key weight parameters and indicator coupling strength related data in the optimization process to the edge strategy library to provide a mechanism for rapid model rollback and parameter backup in case of subsequent anomalies or significant process changes.

[0228] S6, based on real-time predicted task load, edge node computing power, and network bandwidth monitoring results, determines whether the current node has met the elastic migration trigger condition. If it has, subsequent task migration operations are executed. Specifically, this includes:

[0229] S6.1. Perform real-time scheduling on the prediction task queue within the edge node. Based on industrial scheduling parameters such as task queue length, task response latency, and prediction task concurrency, dynamically obtain the prediction task load statistics of the edge node and generate prediction task load monitoring indicators for subsequent resource status perception.

[0230] S6.2. For the load monitoring indicators of the prediction task, combine the current computing power utilization rate of the edge node (such as GPU / CPU real-time utilization rate, floating-point operation power consumption per unit time) and the storage occupancy rate such as memory, and use a multivariate resource evaluation algorithm to calculate the comprehensive computing power status vector of the edge node as the input for edge node computing power bottleneck analysis.

[0231] S6.3 Based on the comprehensive computing power state vector of edge nodes, the network bandwidth between the local node and neighboring nodes is continuously monitored. An adaptive bandwidth detection algorithm is used to periodically calculate the effective data throughput between nodes and generate cross-node network bandwidth availability evaluation index, providing a basis for bandwidth constraint judgment for elastic migration.

[0232] S6.4. For the prediction task load monitoring index, comprehensive computing power state vector and network bandwidth availability evaluation index, combine the multi-condition decision logic and apply the elastic migration trigger algorithm to determine whether the edge node currently meets the following migration criteria: such as a sharp increase in prediction queue pressure, a sustained high computing power utilization rate, or sufficient network bandwidth, and output an elastic migration trigger signal.

[0233] S6.5 Once the elastic migration trigger signal is generated, the elastic migration task decision results are output in real time, including the type of multi-indicator collaborative prediction subtask to be migrated, the prediction data stream to be migrated, the current model parameter snapshot, and the migration target node information. This is used to drive the adaptive migration protocol of subsequent tasks, realize the collaborative expansion or contraction between multiple edge nodes, and ensure the real-time execution performance of key indicator prediction tasks.

[0234] S7, based on the migration trigger result, employs a task-adaptive migration protocol to distribute some or all graph structure fragments and prediction weights to neighboring edge nodes or edge clouds, completing the dynamic expansion or contraction of the prediction task and automatically synchronizing the model state. Specifically, this includes:

[0235] S7.1. Match and analyze the real-time resource monitoring data of edge nodes with dynamic graph structure fragments. Based on the node CPU / GPU load, storage quota and dynamic graph structure fragment characteristics, determine the dynamic graph structure fragments that need to be migrated and their corresponding graph neural network prediction weights in order to screen the objects to be migrated.

[0236] S7.2 Based on the selected dynamic graph structure sharding and its corresponding graph neural network predicted weights, a sharding transfer request is generated through a task adaptive migration protocol. The computing power status and bandwidth conditions of the target's neighboring edge nodes or edge cloud are automatically detected to achieve dynamic perception of the collaborative node's capabilities and provide basic data for subsequent efficient migration.

[0237] S7.3. A distributed task scheduling algorithm is adopted to schedule and allocate the generated sharding transfer requests and target edge node or edge cloud capability parameters, and allocate dynamic graph structure sharding and graph neural network prediction weights to the optimal target node to ensure the continuity and load balancing of key prediction services during the migration process.

[0238] S7.4 The scheduled dynamic graph structure fragments and graph neural network prediction weights are migrated and transmitted through an efficient network protocol. The graph neural network state parameters during the migration process are synchronized using a model incremental synchronization mechanism to ensure that the target node can seamlessly connect to the prediction and inference task and reduce the response delay caused by model cold start.

[0239] S7.5 After completing the dynamic graph structure sharding and graph neural network prediction weight deployment on the target edge node or edge cloud, perform a model prediction engine state consistency check. Use the end-to-end task signature mechanism to automatically correct the model state after migration, ensuring the consistency of prediction services and data integrity among all nodes of the system before and after migration.

[0240] S8 implements an elastic resource scheduling strategy, dynamically adjusting the computing and storage quotas of edge nodes based on the GPU / CPU load, storage space, and predicted task fluctuations of each node, jointly optimizing overall resource utilization and system prediction response speed. Specifically, this includes:

[0241] S8.1. High-frequency sampling and monitoring of the current GPU load, CPU load and storage space usage of edge nodes are performed. Resource status acquisition algorithm is adopted to obtain resource utilization indicators in real time, providing underlying monitoring data for subsequent resource allocation decisions.

[0242] S8.2 Based on the obtained resource utilization rate indicators, combined with the historical and current fluctuation characteristics of the predicted task scheduling, the time series autocorrelation analysis method is used to generate the predicted task load estimation parameters, and input the predicted load trend information for dynamic quota adjustment decisions.

[0243] S8.3. The resource utilization rate index and the predicted task load estimation parameters are jointly calculated using the elastic boundary constraint model to determine the current task pressure status and idle potential of each edge node, and output elastic resource adjustment suggestion parameters.

[0244] S8.4 Based on the elastic resource adjustment recommendations, a dynamic resource quota allocation algorithm is adopted to adaptively adjust the GPU / CPU computing quota and storage resource quota of edge nodes in stages, generating the adjusted computing quota distribution and storage quota distribution to ensure the resource priority and real-time processing of key prediction tasks.

[0245] S8.5. The adjusted computing quota distribution and storage quota distribution of each edge node are synchronized with neighboring edge nodes and edge cloud through the resource scheduling communication protocol. The distributed consensus algorithm is used to optimize the joint resource utilization and form the globally optimal edge side resource configuration state.

[0246] S8.6 Based on the adjusted global resource configuration status, trigger the edge multi-indicator collaborative prediction load redistribution action, and dynamically transfer some prediction subtasks between nodes through the task migration protocol to achieve efficient matching of active tasks and available resources, thereby improving the overall prediction response speed and robustness of the system.

[0247] S9 embeds a self-learning optimization mechanism into the entire multi-indicator collaborative prediction task and resource scheduling process. Through historical prediction accuracy and resource allocation performance, it automatically corrects model transfer parameters and scheduling priorities to improve system robustness and adaptability. Specifically, this includes:

[0248] S9.1 Obtain the prediction accuracy data and resource allocation performance indicators of historical multi-indicator collaborative prediction tasks of edge nodes to form the input basis for the system's self-learning optimization mechanism;

[0249] S9.2 Based on the currently collected prediction accuracy data, a multi-objective evaluation algorithm is used to calculate the performance deviation of each prediction task, so as to quantify the difference between the model prediction effect and the historical best index and generate a dynamic evaluation vector.

[0250] S9.3. Combining the resource allocation performance indicators of each node, the adaptability of the model migration parameters is estimated using the Bayesian update mechanism, and a set of corrected parameter suggestions is output to improve the efficiency of model migration to adapt to new loads.

[0251] S9.4. Combining the dynamic evaluation vector and the set of corrected suggestion parameters, a reinforcement learning optimization algorithm is used to reorder the scheduling priorities of the predicted tasks to obtain a priority allocation decision table, thereby improving the global optimality of the system scheduling.

[0252] S9.5 Based on the priority allocation decision table, dynamically adjust and synchronize task migration parameters and scheduling priority parameters to the edge node scheduling management module to achieve adaptive real-time optimization of the scheduling scheme and output the iteratively corrected system scheduling parameters.

[0253] S9.6 Continuously monitor the model migration behavior and prediction accuracy changes of each edge node after self-learning optimization, and use a sliding window statistical detection algorithm to evaluate the effectiveness of the optimization scheme in real time to determine whether to trigger subsequent structural adjustments or further parameter fine-tuning.

[0254] S10, upon detecting a significant change in process conditions or an increase in the prediction anomaly rate, automatically triggers a model structure fine-tuning and parameter self-update process to ensure that the multi-index coupled prediction model continuously adapts to the actual operating environment of sintering production. Specifically, this includes:

[0255] S10.1 For the prediction results output based on edge deployment, continuously monitor the prediction error sequence of key process indicators and process condition data, and identify the input conditions for the increase in prediction anomaly rate or sudden change in operating conditions through anomaly detection algorithms (such as sliding window anomaly detection or sensitivity analysis) to obtain online model adaptive trigger signals.

[0256] S10.2 Based on the model adaptive trigger signal obtained in S10.1, perform structural sensitivity analysis on the currently operational multi-index coupled dynamic graph neural network structure, apply the node importance evaluation algorithm and graph structure reconstruction mechanism to determine whether it is necessary to adjust the weight configuration of nodes and coupling edges, so as to generate a structural optimization scheme for the current process changes.

[0257] S10.3. For the structural optimization scheme output by S10.2, use the structural fine-tuning mechanism (such as node relationship rearrangement and edge weight self-tuning) to optimize the dynamic graph neural network structure in real time, adjust the multi-index coupling relationship modeling method, and generate a dynamic graph neural network model with optimized structure.

[0258] S10.4. The dynamic graph neural network model with optimized structure generated in S10.3 is input with historical process index data sequence and the latest collected data. An incremental parameter self-updating algorithm (such as online gradient descent or Bayesian posterior update) is used to perform immediate correction on the model parameters to obtain the dynamic graph neural network weight set after parameter self-updating.

[0259] S10.5. Using the dynamic graph neural network weight set obtained from S10.4 after parameter self-updating, the subsequent real-time acquisition of multi-indicator process data streams at the edge is re-predicted, the prediction performance improvement effect is automatically evaluated, and the evaluation results are fed back to the anomaly detection and adaptive strategy module to form a closed loop of model performance-process conditions-model self-adjustment, continuously ensuring the adaptability and prediction robustness of the multi-indicator coupled prediction model to the actual operating environment of sintering production.

[0260] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0261] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A real-time optimization method for sintering processes based on edge computing, characterized in that, Includes the following steps: S1. Obtain the raw multi-index time series data collected in real time on the edge side of the sintering process. For the sensing points at different sections of the sintering production line, record the sampling data and their corresponding process labels respectively. S2. Preprocess the multi-indicator time series data to generate a set of data stream features for the input of the adaptive graph neural network; S3. Based on the process physical constraints and historical statistical correlations of different work sections, a dynamic graph structure is constructed with key indicators in the multi-index time series data as nodes and coupling relationships as edges. The preprocessed data stream feature set is mapped to the dynamic graph structure, specifically: Based on the process physical constraint model, the index partitioning and attribution determination is performed on the preprocessed multi-index data stream characteristics to obtain the physical distribution labels of each key index in different sections of the sintering production line. For process indicators that have been assigned to different regions, the historical process database and historical data collection logs at the edge are used to perform correlation coefficient matrix calculation and Granger causality analysis to extract statistical correlation parameters among key indicators. By combining the process physical constraint model and historical statistical correlation parameters, node sets are generated for all key indicators, and the initial coupling relationship edge weights are labeled for indicator pairs with physical or statistical correlations to form a basic static graph structure topology. Based on the newly acquired data stream feature set, a sliding time window mechanism and a dynamic edge weight assignment algorithm are applied to update the edge weight strength in the static graph structure in real time. Map the node attributes of the data stream feature set covered by the current sliding time window based on the dynamic graph structure described above. For the dynamically adjusted dynamic graph structure, a distributed consensus protocol based on edge nodes is used to verify the consistency of the dynamic graph structure of multiple nodes and synchronize the dynamic graph structure to all edge nodes that need to participate in multi-indicator collaborative prediction. S4. Input the dynamic graph structure into a lightweight graph neural network deployed at the edge, and output the predicted value distribution and mutual information parameters of each key indicator within a predetermined future time window, specifically: Using a dynamic graph structure as input, the system calls a lightweight graph neural network inference instance already deployed locally on the edge nodes to complete the dynamic graph input data format adaptation and model initialization. Based on the input dynamic graph structure, a graph neural network information aggregation algorithm is used to perform graph signal transmission and feature embedding operations on each key indicator node and its adjacent relationship edges, generating the feature embedding vector of each indicator node in the current time series to enhance the correlation of the nodes. Based on the node feature embedding vector, a lightweight multi-task prediction sub-network is driven. Based on the sintering process prediction window parameters, parallel multi-index distribution prediction calculations are performed on each key index node to obtain the multi-index prediction value distribution results. Based on the distribution results of the predicted values ​​of the multiple indicators, and combining node feature embedding and dynamic graph structure, a quantitative analysis is performed on the information gain and joint distribution relationship between all key indicator node pairs, and the mutual information parameter matrix results between each indicator pair are output. The distribution results of the multi-indicator predicted values ​​and the results of the mutual information parameter matrix are packaged into a standardized structured output; S5. Based on the mutual information parameters, dynamically evaluate the coupling strength between process indicators, and optimize the internal structure weights of the graph neural network accordingly. S6. Based on real-time predicted task load, edge node computing power and network bandwidth monitoring results, determine whether the current node has reached the elastic migration trigger condition. If it has, execute the subsequent task migration operation. S7. For migration trigger results, a task adaptive migration protocol is used to distribute part or all of the graph structure fragments and prediction weights to nearby edge nodes or edge clouds, so as to complete the dynamic expansion or contraction of the prediction task and automatically synchronize the model state. S8. Based on the GPU / CPU load, storage space, and predicted task fluctuations of each node, dynamically adjust the computing and storage quotas of edge nodes to jointly optimize the overall resource utilization and system prediction response speed.

2. The real-time optimization method for the sintering process based on edge computing according to claim 1, characterized in that, Following step S8, the following is also included: S9. Embed a self-learning optimization mechanism for the entire multi-indicator collaborative prediction task and resource scheduling process. Automatically correct model transfer parameters and scheduling priorities based on historical prediction accuracy and resource allocation performance. S10. When a significant change in process conditions is detected or the predicted anomaly rate increases, the model structure fine-tuning and parameter self-update process is automatically triggered.

3. The real-time optimization method for the sintering process based on edge computing according to claim 1, characterized in that, Step S1 specifically includes: Data acquisition operations are performed on various types of industrial-grade sensor points deployed in each section of the sintering production line to obtain raw timing signals and form a basic multi-index sensor signal stream. Based on industrial Ethernet or customized acquisition channels, high-frequency real-time sampling of sensor signal streams is performed, and each original sampled data is uniquely identified by timestamp and node numbering mechanism to obtain multi-indicator original data packets with sampling time and spatial positioning. The original data packets are mapped to their locations using the sintering process flow database, and the process physical section labels are automatically associated based on the section information to which the sensing points belong. For raw data packets with multiple indicators and process physical section labels, embed the process real-time status code and operation team identifier, and write the operation context information into the raw time series data through adaptive data label mapping. For the original time-series data of multiple indicators that have been completed and bound to the location and tags, automatic archiving is carried out in time periods through the high-speed caching mechanism of edge computing nodes, generating original time-series database entries with multiple process tags.

4. The real-time optimization method for sintering process based on edge computing according to claim 3, characterized in that, The original timing signals include key process parameters such as temperature, pressure, and gas content.

5. The method according to claim 3, characterized in that, Employing industrial Ethernet or a custom acquisition channel, parameter settings include packet bandwidth ≥100Mbps, transmission delay ≤10ms, and a built-in high-precision clock synchronization mechanism at the edge nodes, using the IEEE 1588 PTP time synchronization protocol, with clock drift error controlled to ≤1ms.

6. The real-time optimization method for sintering process based on edge computing according to claim 1, characterized in that, Step S5 specifically includes: The mutual information parameters output by the preorder graph neural network collaborative prediction module are captured and structured in real time to obtain the mutual information distribution matrix between different key process indicators. Based on the mutual information distribution matrix, an information theory-based coupling coefficient calculation algorithm is applied to dynamically quantify the coupling strength between each pair of key process index nodes and output the process index coupling strength matrix. Using the coupling strength matrix of the process index as input, a weighted adaptive update strategy is used to dynamically optimize and adjust the node weights and edge connection weights inside the graph structure of the graph neural network. After completing the weight adjustment, the lightweight model at the edge is retrained iteratively based on the updated graph neural network structure to perform rapid incremental learning on the multi-indicator feature set collected in the recent time window. The results of structural weight optimization are continuously and dynamically monitored. By comparing them with real-time prediction accuracy and historical performance indicators, the effectiveness of the current graph neural network optimization adjustment is evaluated. The key weight parameters and related data on the coupling strength of indicators during the optimization process are cached in the edge strategy library.

7. The real-time optimization method for sintering process based on edge computing according to claim 2, characterized in that, The self-learning optimization mechanism calculates the deviation of the predicted task performance through a multi-objective evaluation algorithm, corrects the model transfer parameters using a Bayesian update mechanism, and reorders the task scheduling priority using a reinforcement learning optimization algorithm.

8. The real-time optimization method for sintering process based on edge computing according to claim 2, characterized in that, The model structure fine-tuning and parameter self-updating process includes: using an anomaly detection algorithm to identify changes in process conditions or predict events that increase the anomaly rate; generating an optimization scheme through node importance assessment and graph structure reconstruction; and using an incremental parameter self-updating algorithm to correct the model parameters.

Images