Hot rolling data material tracking anomaly analysis method and system based on link modeling

Abstract

The invention discloses a hot rolling data material tracking anomaly analysis method and system based on link modeling. The method comprises the steps that a cross-system timestamp link model is constructed, and nodes of all working procedures of hot rolling production are modeled as graph structure nodes; performing feature engineering processing on the time sequence data, the process parameters and the MES records of each material collected at each process node to form a unified feature vector; modeling the feature vectors by using a graph neural network model, and outputting an abnormal score of each material at each process node; performing anomaly judgment and aggregation based on the anomaly score and a preset threshold value, and identifying an abnormal material and an abnormal process; providing interpretable output to assist root cause tracing; and model closed-loop optimization and threshold adaptive adjustment are supported. According to the method, the problems of cross-system data islands, material tracking distortion and abnormality positioning difficulty in hot rolling production are solved, and the data consistency verification efficiency, the abnormality detection sensitivity and the root cause analysis capability are improved.

Description

Technical Field

[0001] This invention relates to the field of automation and manufacturing technology in the iron and steel metallurgy industry, and in particular to a method and system for analyzing material tracking anomalies in hot rolling data based on link modeling, which is used to realize cross-system data consistency verification, material tracking anomaly detection and root cause tracing during hot rolling production. Background Technology

[0002] Hot rolling production lines are a crucial link in the steel manufacturing process, encompassing multiple steps such as continuous casting, heating furnaces, roughing and finishing rolling, cooling, packaging, weighing, and warehousing. The process is lengthy and involves numerous pieces of equipment. During production, each step generates a large amount of real-time data on temperature, pressure, speed, and other parameters. This data is collected by PLC systems and other automated control systems, while the MES system generates material tracking data. However, due to issues such as dispersed data sources, inconsistent collection frequencies, differences in timestamp accuracy, and communication delays, information is highly fragmented, forming data silos that severely impact overall material tracking and intelligent management throughout the production process.

[0003] Existing material tracking methods primarily rely on manual operation or independent tracking modules, lacking effective real-time detection and early warning mechanisms for production data errors and sudden anomalies. Once data delays, omissions, or timing anomalies occur, traditional methods struggle to pinpoint the root cause in a timely manner, severely impacting production scheduling efficiency and product quality management. Therefore, there is an urgent need for a technical solution that can deploy multi-level detection equipment at key processes to achieve accurate material modeling across system environments, and perform real-time comparison and anomaly analysis through time series and location identifiers. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for anomaly analysis of hot-rolled data material tracking based on link modeling. By constructing a cross-system timestamp link model and using graph neural networks for anomaly scoring, the invention achieves automatic identification and interpretable tracing of material tracking anomalies.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for analyzing material tracking anomalies in hot-rolled data based on link modeling, comprising the following steps: Step S1: Construct a cross-system timestamp link model, and model each process node of the hot rolling production line as a graph structure node; Step S2: Perform feature engineering on the time-series data, process parameters, and MES records collected for each material at each process node to form a unified feature vector; Step S3: Use a graph neural network model to model the feature vectors and output the anomaly score of each material at each process node; Step S4: Based on the anomaly score and preset threshold, perform anomaly judgment and aggregation to identify abnormal materials and abnormal processes; Step S5: Provide interpretable output demonstrating the key features that led to the high anomaly score; Step S6: Perform closed-loop optimization and threshold adaptive adjustment on the model.

[0006] Optionally, in step S2, the feature engineering process includes: Extract time characteristics, process quantity characteristics, identification characteristics and / or counting characteristics, equipment health and network characteristics; The features are standardized and normalized to form a unified input space.

[0007] Optionally, in step S3, the graph neural network model is a GNN model built based on the material-process dependency relationship, which outputs the anomaly score of each material at each process node. The higher the score, the greater the probability of an anomaly.

[0008] Optionally, in step S4, the anomaly determination and aggregation includes: For a single process node, compare the anomaly score with a process-specific threshold to determine whether it is abnormal; For the entire material supply chain, weighted and aggregated abnormal scores for each process are used and compared with global thresholds to determine whether a material is abnormal and to mark the main abnormal processes.

[0009] Optionally, process-specific thresholds and global thresholds are selected on the training set using quantiles or ROC curves of historical normal samples and are periodically re-estimated during system operation.

[0010] Optionally, the interpretability output in step S5 includes high-contribution features, timeline comparisons, and relevant sensor data to form anomaly evidence fragments.

[0011] Optionally, model training and closed-loop optimization in step S6 include: Collect and label historical normal and / or abnormal events for model training and validation; Deploy the model for online inference and save the inference results along with manually confirmed labels; The model is periodically retrained based on the new labeled data to improve detection accuracy.

[0012] Secondly, this invention provides a hot-rolled data material tracking anomaly analysis system based on link modeling, used to implement any of the above-mentioned hot-rolled data material tracking anomaly analysis methods based on link modeling, the system comprising: The link modeling module is configured to build a cross-system timestamp link model, modeling each process node of the hot rolling production line as a graph structure node. The feature engineering module is configured to perform feature engineering processing on the time-series data, process parameters and MES records collected for each material at each process node, forming a unified feature vector; The graph neural network analysis module is configured to model the feature vector using a graph neural network model and output the anomaly score for each material at each process node. The anomaly detection and aggregation module is configured to perform anomaly detection and aggregation based on anomaly scores and preset thresholds to identify abnormal materials and abnormal processes. The interpretability output module is configured to provide interpretable output that shows the key features that lead to high anomaly scores; The model optimization management module is configured to perform closed-loop optimization and adaptive threshold adjustment on the model.

[0013] In another aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described link-based modeling-based hot-rolled data material tracking anomaly analysis methods.

[0014] On another front, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements any of the above-mentioned methods for anomaly analysis of hot-rolled data material tracking based on link modeling.

[0015] The above technical solution has the following technical effects: This invention achieves accurate detection and intelligent diagnosis of anomalies in hot-rolled material tracking by constructing a cross-system timestamp link model and using graph neural networks to model and analyze multi-process features. Its beneficial effects are mainly reflected in: solving the problem of material tracking distortion caused by data silos, and improving the consistency of cross-system data through feature engineering and unified modeling; significantly enhancing the sensitivity and real-time performance of anomaly detection, using graph neural networks to capture complex dependencies between processes, and combining a multi-level threshold judgment mechanism to quickly identify abnormal materials and locate abnormal processes; providing interpretable traceability capabilities, intuitively displaying key features through anomaly evidence fragments to assist maintenance personnel in analyzing root causes; and possessing adaptive optimization capabilities, continuously improving detection accuracy and system robustness through closed-loop learning and dynamic threshold adjustment, ultimately effectively ensuring the continuous stability and quality control of hot-rolled production. Attached Figure Description

[0016] Figure 1 This is a flowchart of a hot-rolled data material tracking anomaly analysis method based on link modeling according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a hot-rolled data material tracking anomaly analysis system based on link modeling, according to another embodiment of the present invention. Figure 3 This is a schematic diagram of an electronic device structure according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0018] In practical applications, many steel mills have installed and equipped MES systems to achieve end-to-end material tracking and quality control from raw materials to finished products. However, real-time monitoring of the rolling mill status still requires high-frequency data sampling, resulting in a huge amount of data transmission between the field process control system (L2) and the MES system, increasing the difficulty of data synchronization and consistency assurance. Existing material tracking mainly relies on manual operation or independent tracking modules. The actual production environment is complex, with many unexpected situations, making it easy for tracking omissions and data loss to occur. At the same time, there is a lack of effective real-time detection and early warning mechanisms for production data errors and sudden anomalies. Once production data experiences delays, omissions, or timing anomalies, traditional methods struggle to pinpoint the root cause of the anomaly in a timely manner, seriously affecting production scheduling efficiency and product quality management.

[0019] In actual production, continuous casting machines, heating furnaces, rolling mills, bundling, weighing, and warehousing processes are connected in series, each controlled by different automated control programs. This leads to significant problems in cross-system collaboration: Firstly, due to differences in data acquisition frequency, timestamp accuracy, and communication latency across systems, accurate material identification across systems is difficult, and data links are prone to interruption. Secondly, MES systems rely heavily on process data to estimate material flow status; when primary instruments or PLC data are abnormal, the system struggles to detect abnormal material flow or missed detections in a timely manner. Furthermore, the lack of a real-time cross-system comparison mechanism based on timestamps and location identifiers prevents primary detection data from effectively verifying the correctness of the MES system, further impacting production traceability and anomaly location efficiency. Therefore, how to deploy multi-level detection equipment at key processes to achieve accurate material modeling in a cross-system environment, and how to perform real-time comparison and anomaly analysis through time series and location identifiers, is a crucial technical problem that needs to be solved in hot-rolled digital production. This is of great significance for improving production process traceability, scheduling efficiency, and product quality management.

[0020] To improve the sensitivity and robustness of cross-system link anomaly detection, this invention upgrades the timestamp link comparison method from a simple threshold judgment to a model-based anomaly scoring mechanism. This mechanism takes the time-series data, process parameters, and MES records collected for each material at each process stage as input. The reconstructed model, trained through this process, outputs an anomaly score for each material at each process node. This score is then compared with the threshold of the corresponding process to determine abnormal materials and abnormal processes.

[0021] The following is combined Figure 1 The flowchart below provides a detailed explanation of the method of the present invention.

[0022] A first aspect of the present invention provides a method for anomaly analysis of hot-rolled data material tracking based on link modeling, comprising the following steps: Step S1: Construct a cross-system timestamp link model, and model each process node of the hot rolling production line as a graph structure node; Step S2: Perform feature engineering on the time-series data, process parameters, and MES records collected for each material at each process node to form a unified feature vector; Step S3: Use a graph neural network model to model the feature vectors and output the anomaly score of each material at each process node; Step S4: Based on the anomaly score and preset threshold, perform anomaly judgment and aggregation to identify abnormal materials and abnormal processes; Step S5: Provide interpretable output demonstrating the key features that led to the high anomaly score; Step S6: Perform closed-loop optimization and threshold adaptive adjustment on the model.

[0023] The above embodiments illustrate the core process of the method of the present invention. Exemplarily, in practical applications, in step S1, data is collected from the control systems and sensors of processes such as the Manufacturing Execution System (MES) of the hot rolling production line, continuous casting machine, heating furnace, roughing mill, finishing mill, cooling line, bundling machine, and weighing device. A time-series graph model is constructed using the material number as a clue, with processes as nodes and material flow as edges. Next, in step S2, feature extraction and standardization are performed on the data of each material (e.g., steel billet) at each process. A feature vector is constructed for each PLC / detection event (first-level signal data) and MES record (business-level record of the manufacturing system). Where i represents the material (or billet) identifier, and m represents the process node number. The feature vector may include, but is not limited to: Time characteristics: Device trigger timestamp MES records time Time difference between the two Duration (enter / exit duration); Process characteristics: real-time parameters and their statistics (mean, variance, skewness) such as temperature (average / maximum / minimum), speed, thickness, and weight. Identification and / or counting characteristics: material number, batch, serial number, pass count (continuous casting count → rolling count); Equipment health and network characteristics: PLC heartbeat delay, data packet loss rate, and local clock skew estimation. These characteristics are standardized and normalized at the edge gateway or middleware to form a unified input space. Subsequently, in step S3, the feature vector is input into a trained graph neural network model, which outputs the anomaly score vector of the material in each process. For example, a graph neural network (GNN) model is used to... Modeling is performed to establish material-process dependencies. The model provides anomaly scores for each input. The score is defined as follows: .in The higher the score of the trained model, the more likely the material is to be abnormal in that process. In step S4, anomaly determination and aggregation are performed based on the anomaly score and a preset threshold. If the score of a specific process exceeds its threshold, it is marked as an abnormal process, and the material is marked as an abnormal material. Specifically, for a single process, by comparison... Process-specific thresholds To determine: if If so, material i is determined to be abnormal in process m.

[0024] For end-to-end evaluation of materials, weighted summation can be used for aggregation:

[0025] in Let m be the weight of process m. The abnormal score of material i in process m. The overall anomaly score for material i across the entire supply chain.

[0026] like If so, the material is determined to be abnormal, and the main abnormal process is marked as follows. ,in A global threshold is set for determining anomalies across the entire connection. To make material i's abnormal score The largest process number m. A specific threshold for each process. The threshold can be set by selecting the quantiles (e.g., 95th percentile) of historical normal samples or by using ROC curves on the training set to ensure that the detection rate and false alarm rate meet preset requirements. As another possible implementation, a self-learning strategy is employed during operation: that is, the threshold is periodically (or window-based) re-estimated for the most recent normal samples, thereby achieving dynamic threshold setting. A report is generated in step S5, highlighting the key features leading to high anomaly scores. To assist maintenance and control personnel in diagnosing anomalies in material tracking, interpretable output is required, demonstrating the key features leading to high anomaly scores. For example, an "anomaly proof fragment" is automatically generated, containing high-contribution features, timeline comparison (PLC vs MES), and relevant sensor data to assist in tracing the root cause; finally, step S6 stores the current judgment result along with a manual confirmation label in the database for iterative model updates. For example, the model training process typically includes: collecting and labeling historical normal and / or abnormal events, training / validating the model; after training, the model is deployed to the edge or middle platform for online inference. By saving the inference results and manually confirmed labels, a closed loop is formed for periodic retraining, which helps to improve the detection accuracy of this application.

[0027] This embodiment fundamentally solves the technical problem of "material tracking distortion caused by data silos" by constructing a cross-system link model. Traditional methods suffer from inconsistent data collection frequencies and timestamp accuracy across systems, leading to difficulties in accurately matching material identifiers and frequent data link interruptions. This method integrates fragmented information into a unified graph structure, providing a data foundation for subsequent analysis. Furthermore, by using model-based anomaly scoring and multi-level threshold determination, it replaces the simple threshold method, enabling sensitive detection of severe local anomalies in specific processes and accurate tracing back to specific materials and processes. This achieves automatic identification and interpretable traceability of abnormal material flow or missed detection phenomena, significantly improving the overall material tracking capability of the production process.

[0028] In some embodiments, the feature engineering process in step S2 includes: Extract time characteristics, process quantity characteristics, identification characteristics and / or counting characteristics, equipment health and network characteristics; The features are standardized and normalized to form a unified input space.

[0029] The above embodiments disclose in detail the feature engineering process of step S2. For example, for the same steel billet, the system extracts the following features at the heating furnace process node: time features (such as equipment trigger timestamp, manufacturing execution system recording time, the time difference between the two, and furnace dwell time); process quantity features (such as average, maximum, and minimum temperature values, speed, thickness, weight, and other parameters and their statistics); identification features (such as material number, batch number, serial number, and pass count); and equipment health and network features (such as control system heartbeat delay, data packet loss rate, and local clock deviation estimation). Then, all numerical features are standardized and normalized to form the final unified feature vector.

[0030] This embodiment directly addresses the technical challenge of "heterogeneous cross-system data, making direct comparison difficult" through systematic feature engineering. Due to differences in data acquisition frequency, timestamp accuracy, and communication latency across systems, simple timestamp comparisons generate numerous false alarms. This method, however, introduces multi-dimensional features to construct an information-rich feature profile, enabling subsequent models to distinguish between "genuine business anomalies" and "system noise" from "data differences." This unified feature processing at the edge gateway or data platform is a crucial step in transforming dispersed and highly fragmented raw data into effective information for intelligent models to identify anomalies, providing a solid data foundation for achieving high-precision anomaly detection.

[0031] In some embodiments, in step S3, the graph neural network model is a GNN model built based on the material-process dependency relationship, which outputs the anomaly score of each material at each process node. The higher the score, the greater the probability of an anomaly.

[0032] For example, in this embodiment, the graph neural network model is a model built based on material-process dependencies. The input to the model is graph data containing material-process relationships, where the features of each node (such as the "finishing mill") are the feature vectors constructed by the aforementioned method. The model aggregates the feature information of neighboring nodes (such as the preceding process "roughing mill" and the following process "cooling line") through graph neural network layers, learning the normal patterns of nodes in their context. The model outputs an anomaly score for each material at each process node; a higher score indicates that the material is more likely to be an anomaly in that process.

[0033] This embodiment addresses the technical problem of traditional isolated detection failing to capture complex inter-process dependencies and propagation anomalies by employing a graph neural network model based on material-process dependencies. In the long-process hot rolling production, each stage is interconnected, and minor deviations in previous stages can be amplified in subsequent stages. Analyzing each stage independently makes it difficult to identify such root causes. The graph neural network model, through its message passing mechanism, can extract node and connection features, thereby capturing complex dependencies between multiple stages and identifying anomalies by reconstructing errors or anomaly scores. This method learns the embedded representation of the graph while preserving node attributes and structural information, thus more accurately reflecting the degree of anomalies in the hot rolling chain.

[0034] In some embodiments, the anomaly determination and aggregation in step S4 includes: For a single process node, compare the anomaly score with a process-specific threshold to determine whether it is abnormal; For the entire material supply chain, weighted and aggregated abnormal scores for each process are used and compared with global thresholds to determine whether a material is abnormal and to mark the main abnormal processes.

[0035] In this embodiment, specifically for a certain material, by obtaining the abnormal scores of each process, a process-level judgment is first performed: each score is compared with a specific threshold for the process; if the score of a process exceeds its threshold, the process is judged to be abnormal. Next, a material-level judgment is performed: the abnormal scores of each process are weighted and aggregated to obtain the comprehensive abnormal score of the material, which is compared with a global threshold; if it exceeds the threshold, the material is judged to be an abnormal material, and the process with the highest score is marked as the main abnormal process.

[0036] This embodiment effectively solves the technical problems of "how to balance local sudden anomalies with overall material quality" and "the lack of an effective real-time detection mechanism for production data corruption" by adopting a dual judgment and aggregation strategy of "process level + material level". A minor anomaly in a single process may not affect the overall material quality, but the cumulative minor anomalies of multiple processes can be fatal. This method ensures rapid response to severe local problems through process-specific thresholds, while simultaneously assessing the overall degree of material anomaly through weighted aggregation and global thresholds, preventing missed detections. This hierarchical judgment logic enables the system to accurately locate fault points and make a comprehensive judgment on the overall health of material flow, improving the automatic identification capability of abnormal materials and abnormal processes.

[0037] In some embodiments, process-specific thresholds and global thresholds are selected on the training set using quantiles or ROC curves of historical normal samples and are periodically re-estimated during system operation.

[0038] In this embodiment, the process of managing thresholds is specified. During the initialization phase, initial thresholds are set using historical normal samples (e.g., material-process records confirmed to be normal within the past three months). For a given process, its process-specific threshold can be selected using the 95th percentile of the abnormal scores of these historical normal samples for that process. The global threshold is also selected based on the aggregate score distribution of historical normal materials. Furthermore, a self-learning strategy is employed during operation: the thresholds are periodically (e.g., weekly) re-estimated using recently generated normal samples, enabling adaptive adjustment of the thresholds.

[0039] The above embodiments address the inherent limitation of static thresholds in adapting to dynamic production environments by using a threshold management strategy based on historical normal sample quantiles and periodically adaptive adjustments. Changes in production conditions, equipment status, and material types lead to natural variations in data distribution, and fixed thresholds gradually result in either increased false alarm rates or decreased detection rates. This method defines the "boundary of the normal range" using the quantiles of normal samples and selects appropriate quantiles based on business requirements (such as detection rate and false alarm rate). Periodic re-estimation allows this boundary to dynamically shift with the "new normal" of production, ensuring the timeliness and scientific rigor of anomaly detection standards and maintaining high system detection performance.

[0040] In some embodiments, the interpretability output in step S5 includes high-contribution features, timeline comparisons, and relevant sensor data to form an anomaly proof fragment.

[0041] The above embodiment demonstrates that in execution step S5, an interpretable output can be generated for anomalous materials. The report is presented in the form of an "anomaly proof fragment," clearly identifying the key features that led to the high anomaly score. These key features may include: high-contribution features (i.e., features that have the greatest impact on the anomaly score), timeline comparisons (such as comparing the difference between the control system trigger timestamp and the manufacturing execution system recorded timestamp for the material in a specific process), and relevant sensor data (such as real-time parameters such as temperature and pressure in the process).

[0042] This embodiment directly addresses the pain point of traditional methods' difficulty in timely locating the root cause of anomalies, leading to low diagnostic efficiency for operations and maintenance personnel, by providing "anomaly proof fragments" containing multi-dimensional evidence. Simply knowing "where the anomaly is" is insufficient for rapid problem-solving; the key is understanding "why the anomaly was identified." This method makes the model's decision-making logic transparent and concrete by displaying key features, timeline comparisons, and relevant sensor data, assisting operations and control personnel in diagnosing anomalies in material tracking. Logically, this constructs a bridge from "anomaly detection" to "root cause analysis," which is of great significance for improving the traceability of the production process, scheduling efficiency, and product quality management.

[0043] In some embodiments, model training and closed-loop optimization in step S6 include: Collect and label historical normal and / or abnormal events for model training and validation; Deploy the model for online inference and save the inference results along with manually confirmed labels; The model is periodically retrained based on the new labeled data to improve detection accuracy.

[0044] In this embodiment, model training and closed-loop optimization in step S6 are implemented. First, during the initial training phase, historical normal and abnormal events are collected and labeled to train and validate the model. After training, the model is deployed to an edge server or data platform for online inference. The system saves the online inference results and combines them with manual confirmation labels from on-site maintenance personnel to form a dataset with new labels. Based on this continuously expanding dataset, a retraining process is periodically (e.g., monthly) to retrain and optimize the model, thereby improving its detection accuracy and adaptability.

[0045] The above embodiments successfully address the dual challenges of "performance degradation due to model staticization" and "insufficient initial labeled data" by constructing a closed-loop optimization process of "data collection and annotation - training and verification - online inference - manual confirmation - periodic retraining." A model trained solely on historical data is insufficient to cover new anomaly patterns that emerge during production. This method transforms operational feedback into valuable labeled data and enables the model to continuously learn the latest normal and anomaly patterns through periodic retraining, much like an experienced engineer. Logically, this achieves a self-improving, increasingly intelligent system that continuously enhances the sensitivity of anomaly detection and root cause analysis capabilities.

[0046] refer to Figure 2 To facilitate understanding of the second aspect of the invention, a hot-rolled data material tracking anomaly analysis system based on link modeling is provided, for implementing any of the above-described link modeling-based hot-rolled data material tracking anomaly analysis methods, the system comprising: The link modeling module is configured to build a cross-system timestamp link model, modeling each process node of the hot rolling production line as a graph structure node. The feature engineering module is configured to perform feature engineering processing on the time-series data, process parameters and MES records collected for each material at each process node, forming a unified feature vector; The graph neural network analysis module is configured to model the feature vector using a graph neural network model and output the anomaly score for each material at each process node. The anomaly detection and aggregation module is configured to perform anomaly detection and aggregation based on anomaly scores and preset thresholds to identify abnormal materials and abnormal processes. The interpretability output module is configured to provide interpretable output that shows the key features that lead to high anomaly scores; The model optimization management module is configured to perform closed-loop optimization and adaptive threshold adjustment on the model.

[0047] The above embodiments concretize the corresponding method claims into a physical system. In this embodiment, the link modeling module is responsible for constructing a cross-system timestamp link model, modeling each process node of the hot rolling production line as a graph structure node. The feature engineering module is responsible for performing feature engineering processing on the time-series data, process parameters, and manufacturing execution system records collected for each material at each process node, forming a unified feature vector. The graph neural network analysis module is responsible for modeling the feature vector using a graph neural network model, outputting the anomaly score for each material at each process node. The anomaly judgment and aggregation module is responsible for anomaly judgment and aggregation based on the anomaly score and a preset threshold, identifying abnormal materials and abnormal processes. The interpretability output module is responsible for providing interpretable output, displaying the key features that lead to high anomaly scores. The model optimization management module is responsible for closed-loop optimization of the model and adaptive threshold adjustment.

[0048] This embodiment addresses the challenge of "difficult engineering implementation of complex algorithms" at the system architecture level by concretizing each step in the method claims into a highly cohesive, loosely coupled software module. The modular design ensures clear and defined data flow, computation flow, and responsibility boundaries. For example, changes in feature engineering do not affect the structure of the graph neural network model, and threshold adjustments only require configuration updates without code modifications. The inherent logic of this architecture ensures the maintainability, scalability, and stability of the system. It is not merely a simple implementation of the method, but a systematic guarantee for its efficient and reliable operation, transforming innovative algorithms into stable, value-creating industrial software systems, significantly improving the efficiency of data consistency verification in the hot rolling production process.

[0049] In some embodiments, such as Figure 3 This application also provides an electronic device, including a processor 301, a memory 302 and a bus 303, and a computer program stored in the memory. When the processor executes the program, it implements any of the above-described link-based modeling-based hot-rolled data material tracking anomaly analysis methods.

[0050] Furthermore, as an executable solution, the hot-rolled data material tracking anomaly analysis system based on link modeling can be a computer unit, which can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer unit may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described computer unit structure is merely an example and does not constitute a limitation on the computer unit. It may include more or fewer components, or combine certain components, or use different components. For example, the computer unit may also include input / output devices, network access devices, buses, etc., which are not limited in this embodiment of the invention.

[0051] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the computer unit, connecting various parts of the entire computer unit via various interfaces and lines.

[0052] In some embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods described above in the embodiments of the present invention.

[0053] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer unit by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0054] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for hot rolling data material tracking anomaly analysis based on link modeling, characterized in that, The method comprises the following steps: Step S1: constructing a cross-system timestamp link model, modeling each process node of the hot rolling production line as a graph structure node; Step S2: performing feature engineering processing on the time series data, process parameters and MES records collected at each process node of each material, to form a unified feature vector; Step S3: modeling the feature vector using a graph neural network model, and outputting an abnormal score of each material at each process node; Step S4: performing abnormality determination and aggregation based on the abnormal score and a preset threshold, to identify abnormal materials and abnormal processes; Step S5: providing an explainability output to show key features that cause high abnormal scores; Step S6: performing closed-loop optimization and threshold adaptive adjustment on the model.

2. The method of claim 1, wherein, In step S2, the feature engineering processing comprises: extracting time features, process quantity features, identification features and / or count features, device health and network features; performing standardization and normalization processing on the features to form a unified input space.

3. The method of claim 1, wherein, In step S3, the graph neural network model is a GNN model constructed based on material-process dependency, which outputs an abnormal score of each material at each process node, and a higher score indicates a higher abnormality possibility.

4. The method of claim 1, wherein, In step S4, the abnormality determination and aggregation comprises: for a single process node, comparing the abnormal score with a process-specific threshold to determine whether it is abnormal; for a material full link, using a weighted sum to aggregate the abnormal scores of each process, and comparing with a global threshold to determine whether it is an abnormal material and to label the main abnormal process.

5. The method of claim 4, wherein, The process-specific threshold and the global threshold are selected on the training set through the quantile or ROC curve of historical normal samples, and are periodically re-estimated during system operation.

6. The method of claim 1, wherein, The explainability output in step S5 comprises high-contribution features, timeline comparison and related sensor data, forming an abnormality proof segment.

7. The method of claim 1, wherein, The model training and closed-loop optimization in step S6 comprise: collecting and labeling historical normal and / or abnormal events for model training and verification; deploying the model for online inference, saving the inference results and manually confirming the labels; periodically retraining the model based on new label data to improve detection accuracy.

8. A link modeling based hot rolling data material tracking anomaly analysis system for implementing the method of any one of claims 1 to 7, characterized by, The system comprises: a link modeling module configured to construct a cross-system timestamp link model, modeling each process node of the hot rolling production line as a graph structure node; a feature engineering module configured to perform feature engineering processing on the time series data, process parameters and MES records collected at each process node of each material, to form a unified feature vector; a graph neural network analysis module configured to model the feature vector using a graph neural network model, and output an abnormal score of each material at each process node; an abnormality determination and aggregation module configured to perform abnormality determination and aggregation based on the abnormal score and a preset threshold, to identify abnormal materials and abnormal processes; an explainability output module configured to provide an explainability output to show key features that cause high abnormal scores; a model optimization management module configured to perform closed-loop optimization and threshold adaptive adjustment on the model.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the method of any one of claims 1-7 when executing the program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the method of any one of claims 1 to 7.

Images