A hot rolling secondary planning risk management and control method and system combining anomaly diagnosis and adaptive recommendation

Abstract

The application provides a hot rolling secondary plan risk management and control method and system combining anomaly diagnosis and adaptive recommendation, and belongs to the technical field of hot rolling production scheduling and intelligent management and control.The method comprises the following steps: collecting multiple-source production data in the execution process of a hot rolling secondary plan in real time; after the multiple-source production data are preprocessed, a fusion feature vector reflecting the state of a rolling process is extracted; based on the fusion feature vector, step-by-step diagnosis is carried out through a process rule rapid filtering layer, an integrated statistical anomaly detection layer and a sequence risk propagation evaluation layer in sequence, and the risk level and the anomaly type of a current rolling unit are output; according to the risk level and the anomaly type, a process parameter adjustment scheme for the anomaly type is generated through a meta-learning model in combination with historical production cases; the process parameter adjustment scheme is applied to a production process, and a historical case database and a basic learner used for parameter recommendation are updated based on the execution result, so that a continuously optimized risk management and control closed loop is formed.

Description

Technical Field

[0001] This invention relates to the field of hot rolling production scheduling and intelligent control technology, and more specifically, to a method and system for risk control of secondary planning in hot rolling that integrates anomaly diagnosis and adaptive recommendation. Background Technology

[0002] Hot-rolled strip steel production is one of the core processes in steel manufacturing. The hot-rolling secondary computer system (L2), as the production execution layer that connects the upper and lower levels, is responsible for decomposing the production orders of the tertiary planning system (L3) into specific rolling unit sequences and calculating the control setting parameters of each stand. The scientificity and reliability of its secondary planning directly determine the stability of the production line operation, the consistency of product quality, the service life of equipment, and the production cost.

[0003] Currently, there are many technical problems in the secondary planning and risk management of hot rolling production lines at home and abroad: First, risk identification relies on experience and lacks systematic diagnosis. It is difficult for manual identification of potential risks caused by multi-parameter coupling and equipment status accumulation, which can easily lead to sudden production anomalies. Second, parameter settings are fixed and lack adaptive optimization. Static mechanism models or empirical formulas cannot adapt to the dynamic changes in equipment status and operating conditions, resulting in fluctuations in product quality and increased energy consumption. Third, anomaly response is passive and lacks proactive prevention. The "post-event remediation" handling mode has the problem of adjustment lag, which can easily lead to production losses.

[0004] Existing research on intelligent hot rolling mainly focuses on single aspects, such as quality prediction based on machine learning, single-parameter forecasting based on data-driven methods, and process simulation based on digital twins. None of these have formed a complete closed loop from systematic diagnosis of planned risks to adaptive parameter recommendation. Moreover, most of them rely on a large amount of historical data and have poor adaptability to data-scarce scenarios such as new production lines and new product launches.

[0005] Therefore, there is an urgent need for a forward-looking intelligent management and control method that can systematically diagnose potential risks in secondary plans and adaptively recommend optimization parameters, so as to achieve a fundamental shift in risk management and control of hot rolling secondary plans from "experience-driven and passive response" to "data-driven and proactive prevention". Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for risk management of hot rolling secondary planning that integrates anomaly diagnosis and adaptive recommendation. This invention primarily utilizes a fusion architecture of multi-level anomaly diagnosis and meta-learning adaptive parameter recommendation. It diagnoses secondary planning risks through a three-level process of rule filtering, statistical detection, and sequence evaluation. Based on a meta-learning model, it achieves rapid parameter adaptation under small sample conditions, thereby transforming the process from "experience-driven, passive response" to "data-driven, proactive prevention." This enables closed-loop management of hot rolling secondary planning risks, encompassing pre-emptive prevention, in-process control, and post-event optimization, significantly improving production stability, product quality, and adaptability to new scenarios.

[0007] The technical means employed in this invention are as follows:

[0008] A risk management method for hot rolling secondary planning that integrates anomaly diagnosis and adaptive recommendation includes: S1. Collect multi-source production data during the execution of the hot rolling secondary plan, and after preprocessing the multi-source production data, extract the fusion feature vector reflecting the rolling process status. S2. Based on the fused feature vector, the process is sequentially diagnosed through the process rule fast filtering layer, the integrated statistical anomaly detection layer, and the sequence risk propagation assessment layer, and the risk level and anomaly type of the current rolling unit are output. S3. Based on the risk level and anomaly type, and combined with historical production cases, generate a process parameter adjustment plan for the anomaly type using a meta-learning model; S4. Apply the process parameter adjustment scheme to the production process, and update the historical case library and basic learner for parameter recommendation based on the execution results to form a continuously optimized risk management closed loop.

[0009] Further, step S1 includes: S11. Collect multi-source production data during the execution of the hot rolling secondary plan, including rolling process data, equipment status data and temperature trajectory data obtained through real-time data channels, as well as quality inspection parameters and plan and order data obtained through offline data channels; S12. Preprocess the collected multi-source production data, including data cleaning and data standardization; S13. Perform multi-scale feature extraction, extracting multi-dimensional original features from the preprocessed data, including basic specification features reflecting the product's geometric dimensions and material properties, rolling force energy features reflecting the load distribution and energy consumption characteristics of each stand, temperature field features reflecting the temperature control level of the rolling process, equipment health features reflecting the operating status of key equipment, and quality correlation features reflecting the historical quality control level. S14. Perform dimensionality reduction processing on the multi-dimensional original features to generate a low-dimensional fused feature vector. The dimensionality reduction processing adopts a kernel method or a nonlinear dimensionality reduction method to reduce the feature dimension while retaining the main information, thereby obtaining the fused feature vector.

[0010] Further, in step S11, the planned and order data includes steel grade, target thickness, target width, target temperature, and rolling sequence information (data sourced from the L2 / L3 system database interface); the rolling process data includes (covering F1 to F7) rolling force, rolling speed, reduction, and main motor operating parameters for each stand (collected via the L1 system real-time data bus); the temperature trajectory data includes roughing mill exit temperature, finishing mill inlet temperature, finishing mill exit temperature, and coiling temperature (collected using an infrared pyrometer array); the equipment status data includes the cumulative rolling mileage of the work rolls, the effective value of the vibration velocity of the support roll bearings, the main motor bearing temperature, and the hydraulic system pressure (data sourced from the equipment management system, vibration sensors, temperature sensors, and pressure sensors); the quality inspection parameters include finished product thickness, finished product width, crown, straightness, yield strength, and tensile strength (collected using a thickness gauge, profile gauge, straightness gauge, and tensile testing machine, collected after each coil is produced or during each test).

[0011] Further, step S2 includes: S21. The process rule rapid filtering layer takes the basic specification characteristics and some process characteristics of the rolling unit as input, and constructs a structured knowledge base rule system based on the hot rolling production process principle and practical experience. The rule system includes rolling force rules, temperature control rules, equipment status rules, planning and scheduling rules, and quality association rules. Each rule includes triggering conditions, risk level, and handling suggestions, and outputs a list of alarms triggered by the rules. S22. The integrated statistical anomaly detection layer takes the fused feature vector as input, and uses three models in parallel detection: adaptive deep isolated forest, dynamic memory support vector machine (OC-SVM), and interpretable variational autoencoder. It also uses Dempster-Shafer evidence theory to fuse the model results, determine whether it is a statistical anomaly, and infer the anomaly type by combining the alarm information of the process rule layer. S23. The sequence risk propagation assessment layer uses the fused feature vectors of multiple consecutive rolling units in the current plan and the abnormal scores of the integrated statistical anomaly detection layer as inputs to construct a graph convolutional network model to assess the sequence propagation risk of anomalies; the comprehensive risk value of the rolling unit is calculated by weighted summation, and the comprehensive risk value is divided into three risk levels: low, medium, and high.

[0012] Further, in step S23, the graph convolutional network model is constructed with each rolling unit as a node, and the node feature is the fused feature vector of that rolling unit; if two rolling units are adjacent in the production plan, an undirected edge is constructed between the corresponding nodes, and the edge weight represents the process coupling strength between the two units, which is calculated by an exponential function to ensure that the more similar the process characteristics of the units, the higher the coupling strength.

[0013] in, Representation unit and unit The difference in yield strength between them Representation unit and unit The width difference between them. This represents the standardized parameter of the strength difference. Indicates the standardized parameter of width difference ( , (Calculated based on historical production data) Furthermore, in the graph convolutional network model: The graph convolutional network propagation adopts a two-layer network structure, with each layer having the same output dimension. The ReLU function is used as the activation function. Through feature propagation in the two-layer network, each node obtains an embedding vector that incorporates information from its neighboring nodes. Sequence risk prediction inputs node embedding vectors into a fully connected layer and outputs the risk probability of that node in the sequence context.

[0014] Furthermore, the training of the graph convolutional network model uses historical planning data, labeled with whether a sequence-related fault actually occurred (such as a previous unit failure leading to a subsequent unit failure), and the model is optimized through supervised learning.

[0015] Further, step S3 includes: S31. Divide historical production data into multiple meta-tasks, each meta-task corresponding to a production condition; train a basic learner using a model-independent meta-learning algorithm to obtain general initialization parameters applicable to multiple conditions; the basic learner adopts a neural network structure, with the input being a fusion representation of condition features and abnormal information, and the output being the adjustment amount of process parameters; S32. Define the risk diagnosis results and operating condition characteristics of the current rolling unit as a new meta-task; retrieve several historical cases similar to the current operating condition from the historical case library to form a support set; the similarity is calculated by distance measurement or similarity measurement between feature vectors; S33. Based on the support set data, starting with the general initialization parameters, perform a few-step iterative optimization using the gradient descent method to obtain the basic learner parameters adapted to the current working condition; the number of iteration steps is dynamically adjusted according to the size of the support set. S34. Input the current operating condition characteristics and anomaly codes into the adapted basic learner to generate preliminary parameter adjustment amounts; combine process constraints to perform process constraint verification and rolling simulation verification, output the final recommended parameter vector that has passed verification, and simultaneously output the risk level reduction prediction, key process indicators, and reference historical case numbers.

[0016] Further, step S4 includes: S41. After the process parameter adjustment plan is confirmed by the operator, it is sent to the secondary process control system in real time via industrial Ethernet to replace the original control parameters of the current rolling unit and drive the production equipment to execute the rolling process according to the process parameter adjustment plan. S42. Collect the actual production data after the rolling unit executes the process parameter adjustment scheme, including rolling force load rate, final rolling temperature and quality inspection data; S43. Compare the actual production data with the preset success criteria to determine whether the execution effect meets expectations; the success criteria include that the rolling force load rate is within the safe range, the final rolling temperature deviation is within the allowable range, and the quality inspection data meets the product standards. S44. When the execution effect meets the success determination condition, the working condition characteristics, process parameters and production effect of the rolling unit are added to the historical case library in the form of a triplet; the historical case library is periodically deduplicated, the most representative cases are retained, and the total size of the case library is controlled within a preset upper limit. S45. Periodically use newly added successful cases from the historical case library to incrementally train the meta-learning model and update the general initialization parameters of the basic learner so that the basic learner can adapt to the slow changes in the production process.

[0017] This invention also provides a hot rolling secondary plan risk management system based on the above-mentioned integrated anomaly diagnosis and adaptive recommendation method, comprising: The data acquisition and preprocessing module is used to acquire multi-source data of the hot rolling production process, including rolling process parameters and equipment status parameters obtained through real-time data channels, as well as quality inspection parameters obtained through offline data channels; by preprocessing and feature engineering the multi-source data, a fusion feature vector reflecting the state of the rolling unit is generated.

[0018] The multi-dimensional anomaly diagnosis module is used to perform step-by-step diagnosis based on the fused feature vector, sequentially through the process rule fast filtering layer, the integrated statistical anomaly detection layer, and the sequence risk propagation assessment layer, and output the risk level and anomaly type of the current rolling unit. The meta-learning adaptive parameter recommendation module is used to generate process parameter adjustment schemes for the aforementioned anomaly type based on the risk level and anomaly type, combined with historical production cases, through a meta-learning model. The closed-loop control and decision output module is used to apply the process parameter adjustment scheme to the production process, and update the historical case library and basic learner for parameter recommendation based on the execution results, forming a continuously optimized risk control closed loop.

[0019] Compared with the prior art, the present invention has the following advantages: 1. This invention enables pre-diagnosis and proactive prevention of risks in the secondary planning of hot rolling. After application, unplanned downtime caused by improper secondary planning can be reduced by more than 30%. Through dynamic optimization of parameter recommendations, it effectively solves problems such as excessive rolling force, excessive temperature, and abnormal process coupling, and significantly improves the stability of key product quality.

[0020] 2. This invention can complete the full planning risk assessment and parameter optimization of dozens or even hundreds of rolling units within minutes. Compared with the hours of manual analysis by senior engineers, it allows sufficient time for pre-production adjustments and achieves true pre-control.

[0021] 3. This invention continuously transforms human experience, successful production cases, and production data into storable, reusable, and optimizable digital models and case libraries, effectively solving the problem of process knowledge transformation and inheritance in steel enterprises, and laying a solid foundation for the continuous improvement of the intelligent level of hot rolling production.

[0022] 4. The parameter recommendation mechanism based on meta-learning in this invention can quickly adapt to new working conditions with only a small number of samples, effectively solving the cold start problem in scenarios with scarce data, such as new production lines and new product launches, and has a wide range of applications.

[0023] 5. This invention realizes the automation and intelligence of risk management for secondary planning in hot rolling, reduces reliance on the personal experience of planners, lowers the error rate of manual operation, and further reduces energy consumption and equipment wear by optimizing process parameters, thereby reducing production costs.

[0024] Based on the above reasons, this invention can be widely applied in fields such as hot rolling production scheduling and intelligent control. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a detailed workflow and data flow diagram of the multi-level anomaly diagnosis of the present invention.

[0028] Figure 3 The logic diagram of the adaptive parameter recommendation strategy of this invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0031] like Figure 1 As shown, this invention provides a method for risk management of hot rolling secondary planning that integrates anomaly diagnosis and adaptive recommendation, including: S1. Collect multi-source production data during the execution of the hot rolling secondary plan, and after preprocessing the multi-source production data, extract the fusion feature vector reflecting the rolling process status. S2. Based on the fused feature vector, the process is sequentially diagnosed through the process rule fast filtering layer, the integrated statistical anomaly detection layer, and the sequence risk propagation assessment layer, and the risk level and anomaly type of the current rolling unit are output. S3. Based on the risk level and anomaly type, and combined with historical production cases, generate a process parameter adjustment plan for the anomaly type using a meta-learning model; S4. Apply the process parameter adjustment scheme to the production process, and update the historical case library and basic learner for parameter recommendation based on the execution results to form a continuously optimized risk management closed loop.

[0032] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11. Collect multi-source production data during the execution of the hot rolling secondary plan, including rolling process data, equipment status data and temperature trajectory data obtained through real-time data channels, as well as quality inspection parameters and plan and order data obtained through offline data channels; S12. Preprocess the collected multi-source production data, including data cleaning and data standardization; In this embodiment, data cleaning includes missing value imputation and outlier removal, wherein missing value imputation employs... K KNN interpolation method k =5) Fill in a small number of missing values ​​(missing rate <5%) using linear interpolation based on time-series trends; fill in continuous missing values ​​(missing rate 5%-10%); outlier removal combined with improved method 3. The criteria and Local Outlier Factor (LOF) identify and eliminate outliers of single indicators (such as sudden changes caused by sensor failures) and localized dense outliers (such as short-term equipment fluctuations).

[0033] In this embodiment, data standardization is performed on features of different magnitudes (such as temperature, current, and component content) using Z-score standardization (suitable for normally distributed data, such as process temperature) and Min-Max standardization (suitable for non-normally distributed data, such as raw material composition) to normalize them and eliminate the influence of dimensions. The formula for score standardization is as follows:

[0034] in, The mean, Standard deviation; The formula for Min-Max standardization is as follows:

[0035] in, , These are the minimum and maximum values ​​of the feature, respectively. S13. Perform multi-scale feature extraction, extracting multi-dimensional original features from the preprocessed data, including basic specification features reflecting the product's geometric dimensions and material properties, rolling force energy features reflecting the load distribution and energy consumption characteristics of each stand, temperature field features reflecting the temperature control level of the rolling process, equipment health features reflecting the operating status of key equipment, and quality correlation features reflecting the historical quality control level. In this embodiment, a 65-dimensional feature vector is extracted from the cleaned standardized data, covering five major categories of features: basic specifications, rolling force, temperature field, equipment health, and quality-related features, as detailed below: There are five basic specifications: target finished product thickness, target finished product width, width-to-thickness ratio (calculated by the ratio of target width to target thickness), steel grade strength coefficient (maps the steel grade to the corresponding standard value of yield strength), and planned position coefficient (calculated by the ratio of the current rolling unit number to the total number of planned units).

[0036] Thirty-four rolling force characteristics are categorized into single-stand characteristics and overall characteristics. For each stand from F1 to F7, four characteristics are extracted: mean rolling force, standard deviation of rolling force, peak rolling force factor (calculated by dividing the difference between the peak and mean values ​​by the standard deviation), and rolling force load rate (calculated by the ratio of the mean to the rated maximum force of the stand). For the entire rolling process, six characteristics are extracted: estimated total rolling power (calculated by the sum of the products of the mean rolling force of each stand and the rolling speed), power fluctuation coefficient (calculated by the ratio of the standard deviation of total power to the mean total power), uniformity of force distribution (calculated by subtracting the standard deviation of the rolling force load rate of each stand from 1), stand number with maximum load, maximum load value, load gradient (calculated by the difference between the load rate of stand F7 and the load rate of stand F1), and number of load jumps (counting the number of times the rolling force load rate changes by more than 0.1). These characteristics comprehensively reflect the force consumption and distribution characteristics of the rolling process.

[0037] The temperature field features include 16 dimensions. For four key temperature measurement points—roughing mill exit, finishing mill inlet, finishing mill exit, and coiling—four features are extracted for each point: mean temperature, minimum temperature, standard deviation of temperature, and mean temperature deviation (calculated by the difference between the actual mean temperature and the target temperature). The system depicts the temperature change trajectory and control effect of the rolling process.

[0038] The equipment health characteristics comprise 17 dimensions, including single-stand characteristics and overall characteristics: For each stand from F1 to F7, two characteristics are extracted: the cumulative rolling mileage of the work rolls and the effective value of the vibration velocity of the support roll bearings; For the overall equipment status, three characteristics are extracted: the average bearing temperature of the main motor, the pressure fluctuation rate of the hydraulic system, and the comprehensive equipment health index (calculated by subtracting 0.3 times the ratio of the average vibration value to the threshold, 0.2 times the ratio of the temperature deviation to the threshold, and 0.5 times the ratio of the cumulative rolling mileage to the roll changing mileage from 1), to comprehensively assess the equipment's operating status and health level.

[0039] The quality-related characteristics include four dimensions: historical thickness control accuracy (calculated using the standard deviation of thickness from the most recent 5 rolls of the same specification), historical convexity control accuracy (calculated using the standard deviation of convexity from the most recent 5 rolls of the same specification), historical strength fluctuation (calculated using the standard deviation of yield strength from the most recent 5 rolls of the same specification), and overall quality score (based on thickness pass rate). 0.4+ convexity pass rate 0.3 Strength pass rate (Weighted sum calculation of 0.3), combined with historical quality data, reflects the quality control capability of the production process.

[0040] S14. Perform dimensionality reduction processing on the multi-dimensional original features to generate a low-dimensional fused feature vector. The dimensionality reduction processing adopts a kernel method or a nonlinear dimensionality reduction method to reduce the feature dimension while retaining the main information, thereby obtaining the fused feature vector.

[0041] In this embodiment, the kernel principal component analysis method is used for feature dimensionality reduction. A Gaussian kernel function is selected, the kernel parameter is set to 0.1, and n_components=30. The 76-dimensional standardized features are reduced to 30 dimensions, ensuring that more than 95% of the original information is retained after dimensionality reduction. This reduces the computational load of the model, improves the running efficiency, and avoids information loss that leads to a decline in model performance.

[0042] In a preferred embodiment of the present invention, in step S11, the planned and order data includes steel grade, target thickness, target width, target temperature, and rolling sequence information (data sourced from the L2 / L3 system database interface); the rolling process data includes (covering F1 to F7) rolling force, rolling speed, reduction, and main motor operating parameters for each stand (collected via the L1 system real-time data bus); the temperature trajectory data includes roughing mill exit temperature, finishing mill inlet temperature, finishing mill exit temperature, and coiling temperature (collected using an infrared pyrometer array); the equipment status data includes the cumulative rolling mileage of the work rolls, the effective value of the vibration velocity of the support roll bearings, the main motor bearing temperature, and the hydraulic system pressure (data sourced from the equipment management system, vibration sensors, temperature sensors, and pressure sensors); and the quality inspection parameters include finished product thickness, finished product width, crown, flatness, yield strength, and tensile strength (collected using a thickness gauge, profile gauge, flatness gauge, and tensile testing machine, collected after each coil is produced or during each test).

[0043] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. The process rule rapid filtering layer takes the basic specification characteristics and some process characteristics of the rolling unit as input, and constructs a structured knowledge base rule system based on the hot rolling production process principle and practical experience. The rule system includes rolling force rules, temperature control rules, equipment status rules, planning and scheduling rules, and quality association rules. Each rule includes triggering conditions, risk level, and handling suggestions, and outputs a list of alarms triggered by the rules. S22. The integrated statistical anomaly detection layer takes a 30-dimensional fused feature vector as input, and uses three models in parallel detection: adaptive deep isolated forest, dynamic memory support vector machine (OC-SVM), and interpretable variational autoencoder. The model results are fused through Dempster-Shafer evidence theory to determine whether it is a statistical anomaly, and the anomaly type is inferred by combining the alarm information of the process rule layer. In this embodiment, the core improvement of the adaptive deep isolation forest lies in the introduction of a feature importance weighting mechanism. During the segmentation process of constructing the isolation trees, key features with high correlation to anomalies, such as load rate and temperature deviation, are given higher segmentation weights, thereby improving the model's sensitivity to important anomaly features. The model is trained using only historical normal data, i.e., 20,000 samples with no alarm records and excellent quality. The model parameters are set to 200 decision trees, 256 samples per tree, and a preset anomaly ratio of 0.05, outputting anomaly scores ranging from 0 to 1.

[0044] In this embodiment, the Dynamic Memory Support Vector Machine (OC-SVM) uses a sliding time window mechanism to update the training set, retaining only the production data from the most recent three months to ensure the model can adapt to the characteristics of slowly changing processes. The model selects the radial basis function as the kernel function, with the kernel parameter set to 0.1 and the anomaly proportion parameter set to 0.03. The output is converted into anomaly probabilities in the 0-1 interval using the Sigmoid function.

[0045] In this embodiment, the interpretable variational autoencoder (AE) network structure adopts a symmetrical design. The encoder part progressively compresses 30-dimensional input features to 5-dimensional latent variables, while the decoder part progressively reconstructs 30-dimensional features from the 5-dimensional latent variables. The latent variables follow a standard normal distribution. The model loss function consists of a weighted sum of reconstruction error and KL divergence, with a weight coefficient set to 0.5. This loss function is used for model training. Anomaly measurement uses a weighted sum of reconstruction error and KL divergence, with a weight of 0.6 for reconstruction error and 0.4 for KL divergence. The measurement result is then normalized to the 0-1 interval as the anomaly score. Simultaneously, by calculating the gradient of the input features with respect to the reconstruction error, the top three anomaly features with the largest contribution are identified, achieving anomaly localization.

[0046] In this embodiment, model fusion and decision-making adopt the Dempster-Shafer evidence theory, as detailed below: First, an identification framework is defined, containing two basic propositions: "normal" and "abnormal." Then, the outputs of the three models are treated as independent pieces of evidence. The basic probability allocation function for each piece of evidence is set as the probability of the "abnormal" proposition corresponding to the model's anomaly score, minus the probability of the "normal" proposition corresponding to the anomaly score (1). No probability allocation is set for uncertain propositions. Next, the synthesized basic probability allocation function is calculated using evidence synthesis rules to obtain the confidence function value and likelihood function value for the anomalous proposition. The overall anomaly score is calculated as a weighted sum of the confidence function value and the likelihood function value, with a weight of 0.7 for the confidence function value and 0.3 for the likelihood function value. Finally, a statistical anomaly judgment threshold of 0.75 is set. If the overall anomaly score exceeds this threshold, it is judged as a statistical anomaly. Anomaly type inference combines key anomaly features located by the variational autoencoder with alarm information triggered by the process rule layer to comprehensively determine the anomaly type, such as rolling force anomaly, temperature anomaly, equipment anomaly, etc.

[0047] S23. The sequence risk propagation assessment layer takes the 30-dimensional fusion feature vector of multiple consecutive rolling units in the current plan and the abnormal score of the integrated statistical anomaly detection layer as input to construct a graph convolutional network model to assess the sequence propagation risk of anomalies; the comprehensive risk value of the rolling unit is calculated by weighted summation, and the comprehensive risk value is divided into three risk levels: low, medium and high.

[0048] In this embodiment, the final comprehensive risk rating is calculated by weighted summation of the comprehensive risk value for each rolling unit: the weight of process rule alarms is 0.3, with no alarm scoring 0 points, a yellow warning scoring 0.5 points, and a red warning scoring 1 point; the weight of statistical anomaly scores is 0.5; and the weight of sequence risk probability is 0.2. Based on the comprehensive risk value, three risk levels are defined: low risk (comprehensive risk value less than 0.4), requiring no special handling and maintaining normal monitoring; medium risk (comprehensive risk value between 0.4 and 0.7), issuing warning information and suggesting optimization of process parameters; and high risk (comprehensive risk value greater than or equal to 0.7), issuing a strong warning and requiring adjustment of the production plan or process parameters. The output of this layer includes the risk level, comprehensive risk value, anomaly type, main anomaly characteristics, and rule alarm list for each rolling unit.

[0049] In a preferred embodiment of the present invention, in step S23, the graph convolutional network model is constructed with each rolling unit as a node, and the node feature is the fused feature vector of that rolling unit. If two rolling units are adjacent in the production plan, an undirected edge is constructed between the corresponding nodes, and the edge weight represents the process coupling strength between the two units, which is calculated by an exponential function to ensure that the more similar the process characteristics of the units, the higher the coupling strength.

[0050] in, Representation unit and unit The difference in yield strength between them Representation unit and unit The width difference between them. This represents the standardized parameter of the strength difference. Indicates the standardized parameter of width difference ( , (Calculated based on historical production data).

[0051] In a specific implementation, as a preferred embodiment of the present invention, the graph convolutional network model includes: The graph convolutional network propagation adopts a two-layer network structure, with each layer having the same output dimension of 16. The ReLU function is used as the activation function. Through feature propagation of the two layers, each node obtains a 16-dimensional embedding vector that incorporates information from its neighboring nodes. Sequence risk prediction inputs node embedding vectors into a fully connected layer and outputs the risk probability of that node in the sequence context.

[0052] In a specific implementation, as a preferred embodiment of the present invention, the training of the graph convolutional network model uses historical planning data, labeled with whether a sequence-related fault actually occurs (such as a previous unit abnormality leading to a subsequent unit fault), and the model is optimized through supervised learning.

[0053] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. Divide historical production data into multiple meta-tasks, each meta-task corresponding to a production condition; train a basic learner using the Model Independent Meta-Learning (MAML) algorithm to obtain general initialization parameters applicable to multiple conditions; the basic learner adopts a neural network structure, with the input being a fusion representation of condition features and abnormal information, and the output being the adjustment amount of process parameters; In this embodiment, the basic learner adopts a lightweight neural network structure. The input layer is a 38-dimensional vector (30-dimensional fused feature vector and 8-dimensional anomaly encoding vector, the anomaly encoding vector being generated according to the anomaly type). Two hidden layers are set: the first layer is 64-dimensional and uses the ReLU activation function, and the second layer is 32-dimensional and uses the ReLU activation function. The output layer is 8-dimensional and uses the Tanh activation function. The Tanh function is used to limit the parameter adjustment range of the output layer to [-1, 1], and then the actual adjustment amount is mapped to the allowable adjustment range of each parameter. The 8-bit anomaly code represents: 1st bit for single-machine rolling force anomaly, 2nd bit for temperature field anomaly, 3rd bit for equipment health anomaly, 4th bit for schedule anomaly, 5th bit for historical quality correlation anomaly, 6th bit for total rolling power anomaly, 7th bit for sequence propagation risk anomaly, and 8th bit for other anomalies.

[0054] In this embodiment, the training process of the meta-learning model constructs 200 meta-tasks from historical production data. Each meta-task includes a support set and a query set. The support set consists of 50 successful production samples, and the query set consists of 2 samples. The Model-Independent Meta-Learning (MAML) algorithm is used for training, with the core objective of minimizing the average loss across all meta-tasks. The specific training logic is as follows: For each meta-task, the gradient of the loss function is calculated based on the support set data, and the parameters of the base learner are updated using gradient descent (with the inner loop learning rate set to 0.01). The updated parameters are used to calculate the loss function on the query set, and the average query set loss of all meta-tasks is used as the meta-loss function. The meta-loss function is optimized to obtain the general initialization parameters of the base learner, ensuring that these parameters have the ability to quickly adapt to new working conditions.

[0055] S32. Define the risk diagnosis results and operating condition characteristics of the current rolling unit as a new meta-task; retrieve several historical cases similar to the current operating condition from the historical case library to form a support set; the similarity is calculated by distance measurement or similarity measurement between feature vectors; In this embodiment, the K most similar historical cases to the current working condition are retrieved from the historical case library (K defaults to 10, but can be reduced to 3 if the working condition is entirely new), forming the support set for the new meta-task. Case similarity is calculated by comparing the 30-dimensional fused feature vector of the current working condition with the feature vectors of the historical cases using cosine similarity.

[0056] S33. Based on the support set data, starting with the general initialization parameters, perform a few-step iterative optimization using the gradient descent method to obtain the basic learner parameters adapted to the current working condition; the number of iteration steps is dynamically adjusted according to the size of the support set. In this embodiment, the gradient of the loss function of the base learner on the support set is calculated, and the parameters are updated by gradient descent (the learning rate of the inner loop is kept at 0.01). The number of adaptation steps is adjusted according to the size of the support set, usually 1-3 steps, to obtain model parameters adapted to the current working conditions.

[0057] S34. Input the current operating condition characteristics and anomaly codes into the adapted basic learner to generate preliminary parameter adjustment amounts; combine process constraints to perform process constraint verification and rolling simulation verification, output the final recommended parameter vector that has passed verification, and simultaneously output the risk level reduction prediction, key process indicators, and reference historical case numbers.

[0058] In this embodiment, the 30-dimensional fused feature vector of the current operating condition is concatenated with the 8-dimensional anomaly encoding vector and input into the base learner with adapted parameters to obtain preliminary values ​​for the 8-dimensional parameter adjustment. Combining the allowable adjustment range of each process parameter, the preliminary adjustment is mapped to the actual adjustment, and added to the currently used original parameters to obtain preliminary recommended parameters.

[0059] In a specific implementation, as a preferred embodiment of the present invention, step S4 includes: S41. After the process parameter adjustment plan is confirmed by the operator, it is sent to the secondary process control system in real time via industrial Ethernet to replace the original control parameters of the current rolling unit and drive the production equipment to execute the rolling process according to the process parameter adjustment plan. S42. Collect actual production data after the rolling unit executes the process parameter adjustment scheme, including rolling force load rate, final rolling temperature and quality inspection data; S43. Compare the actual production data with the preset success criteria to determine whether the execution effect meets expectations; the success criteria include that the rolling force load rate is within the safe range, the final rolling temperature deviation is within the allowable range, and the quality inspection data meets the product standards. S44. When the execution effect meets the success determination condition, the working condition characteristics, process parameters and production effect of the rolling unit are added to the historical case library in the form of a triplet; the historical case library is periodically deduplicated, the most representative cases are retained, and the total size of the case library is controlled within a preset upper limit. S45. Periodically use newly added successful cases from the historical case library to incrementally train the meta-learning model and update the general initialization parameters of the basic learner so that the basic learner can adapt to the slow changes in the production process.

[0060] This invention also provides a hot rolling secondary plan risk management system based on the above-mentioned integrated anomaly diagnosis and adaptive recommendation method, comprising: The data acquisition and preprocessing module is used to acquire multi-source production data in real time during the execution of the hot rolling secondary plan. After preprocessing the multi-source production data, a fusion feature vector reflecting the rolling process status is extracted. The multi-dimensional anomaly diagnosis module is used to perform step-by-step diagnosis based on the fused feature vector, sequentially through the process rule fast filtering layer, the integrated statistical anomaly detection layer, and the sequence risk propagation assessment layer, and output the risk level and anomaly type of the current rolling unit. The meta-learning adaptive parameter recommendation module is used to generate process parameter adjustment schemes for the aforementioned anomaly type based on the risk level and anomaly type, combined with historical production cases, through a meta-learning model. The closed-loop control and decision output module is used to apply the process parameter adjustment scheme to the production process, and update the historical case library and basic learner for parameter recommendation based on the execution results, forming a continuously optimized risk control closed loop.

[0061] Example Taking a 2300mm wide strip steel hot continuous rolling production line of a large steel enterprise, covering more than 150 steel grades including low-carbon steel, low-alloy steel, high-strength steel, and automotive steel, with widths ranging from 1000-2300mm, and equipped with a three-level computer system of L1 (basic automation), L2 (process automation), and L3 (production planning), this paper takes a rolling unit (12th in the planning sequence, upstream of a Q235B steel grade with a thickness of 10mm and a width of 1700mm) of Q355B steel grade as an example to demonstrate the diagnostic process and results in detail. The detailed implementation method is as follows: Before applying the secondary planning risk management method of this invention, two high-performance servers, one primary and one backup, need to be deployed to deploy the algorithm model and database, and an industrial Ethernet switch to ensure real-time data transmission.

[0062] Step 1, Data Collection: Production data from the past three years of this production line were collected, and 25,000 normal samples with no abnormalities and excellent quality were selected (for training the isolated forest model), 8,000 fault samples (covering scenarios such as abnormal rolling force, excessive temperature, and equipment vibration), and 120,000 complete {feature vector, process parameter, production effect} triple data were used for meta-learning training and case library construction. After data cleaning, the effective sample retention rate was ≥96%.

[0063] Process rule library refinement: Based on the process specifications of this production line and the experience of senior engineers, five core refinement rules have been formed, as shown in Table 1.

[0064] Table 1 Detailed Process Rules

[0065] Data cleaning: Deploying the KNN interpolation algorithm ( k =5) Handle scattered missing values ​​with a missing rate of <5%, use linear interpolation to handle continuous missing values ​​of 5%-10%, and trigger a sensor fault alarm when the missing rate is >10%; use the improved 3σ criterion (threshold coefficient adjusted to 2.5) to identify single index outliers, and use the LOF algorithm (neighborhood size set to 20) to identify local dense outliers. Logs are automatically recorded after outlier removal.

[0066] Data standardization: Z-score standardization is used for normally distributed data such as temperature and current, and Min-Max standardization is used for non-normally distributed data such as raw material composition. The standardization process is executed in real time.

[0067] Feature extraction and dimensionality reduction: 65 original features were extracted according to the technical solution, and principal component analysis (Gaussian kernel function, kernel parameter 0.1) was used to reduce the dimensionality to 30 dimensions, ensuring that the information retention rate after dimensionality reduction is ≥95%.

[0068] Step 2, Multi-dimensional Anomaly Diagnosis: The first layer of process rule rapid filtering: The refined process rule library is deployed to the anomaly diagnosis server. The rule engine is used to achieve real-time matching. The input is basic specification features (steel strength coefficient, width-to-thickness ratio) and some process features (current equipment vibration speed, strength difference, width difference). The output is a list of rule alarms (including anomaly type, risk level, and confidence level).

[0069] The input data is shown in Table 2: Table 2 Input Data

[0070] Rule matching process: Matching R_F04 (F5 stand rolling force load rate > 0.95): Current value 0.96, triggering a red alarm; Matching R_T03 (|final rolling temperature - target value| = 35℃ > 25℃): Triggers a red alert; Matching R_P03 (strength difference 280MPa>200MPa): The current unit steel grade strength coefficient is 355MPa, the upstream Q235B steel grade strength coefficient is 235MPa, the difference is 280MPa, triggering a yellow warning; Matching R_P01 (width jump rate = 100mm ÷ 1700mm ≈ 0.059 < 0.20): Not triggered; None of the other rules met the trigger conditions.

[0071] The output results are shown in Table 3: Table 3 Output Results

[0072] Second layer integrated statistical anomaly detection layer: Input data: 30-dimensional fully fused feature vector; The parallel detection results of the three models are as follows: Adaptive Deep Isolation Forest: Anomaly score 0.82 (>0.75, considered anomaly), core sensitive features are "F5 rolling force load rate, final rolling temperature deviation, and steel strength coefficient"; Dynamic memory OC-SVM: anomaly probability 0.85 (>0.75, considered anomaly), outputs the probability value after Sigmoid transformation; Interpretable variational autoencoder: reconstruction error, KL divergence weighted sum 0.81 (>0.75, judged as abnormal), the first 3 abnormal features are located as "F5 rolling force load rate, final rolling temperature, and strength difference between adjacent units"; Model fusion (Dempster-Shafer evidence theory): Evidence 1 (Isolated Forest): Anomaly probability 0.82, normal probability 0.18; Evidence 2 (OC-SVM): Anomaly probability 0.85, normal probability 0.15; Evidence 3 (Variational Autoencoder): Anomaly probability 0.81, normal probability 0.19; Composite calculation: Trust function value = (0.82 + 0.85 + 0.81) / 3 = 0.83, Likelihood function value = max(0.82, 0.85, 0.81) = 0.85 + 0.01 = 0.86, Overall anomaly score = 0.83 × 0.7 + 0.86 × 0.3 = 0.84 (> 0.75, judged as statistically abnormal); Output results: Statistical anomaly (overall score 0.84), the core anomaly features are "F5 rolling force exceeds limit and final rolling temperature is seriously exceeded", and the suspected anomaly type is "rolling force-temperature coupling anomaly".

[0073] Third-level sequence risk propagation assessment layer: Input data: 30-dimensional feature vectors of the current unit and 9 units before and after it, totaling 19 rolling units; second-layer comprehensive anomaly score of 0.84; Graph structure construction: Node: 19 rolling units, the current unit is node 10, the node features include steel strength coefficient of 355MPa and width-to-thickness ratio of 225; Edge weight: Calculation of the process coupling strength between the current element (node ​​10) and the upstream node 9 (Q235B, steel strength coefficient 235MPa, width-to-thickness ratio 170): =355-235=120MPa =1800-1700=100mm, substitute into the formula ( Standardized parameter 0.0018, (Standardized parameter 0.0009) Edge weight = exp(-0.0018×120)×exp(-0.0009×100) = exp(-0.216)×exp(-0.09) 0.806 × 0.914 = 0.737, this value represents strong coupling, and this value is used as a feature in the graph convolutional network; Graph convolutional network output: Risk probability of the current unit sequence = 0.78; The overall risk rating is as follows: Process rule alarm score: 2 red alarms (1 point each) × 0.3 + 1 yellow warning (0.5 points) × 0.3 = (2 × 1 + 1 × 0.5) × 0.3 = 0.75; The outlier score was 0.84 × 0.5 = 0.42. Sequence risk probability: 0.78 × 0.2 = 0.156; Overall risk value = 0.75 + 0.42 + 0.156 = 1.326 (≥0.7, judged as high risk); The final output is as follows: Risk level: High risk; Overall risk value: 1.326; Core anomaly type: F5 rolling force exceeds limit, final rolling temperature is seriously exceeded, and strength transition is hard; Key anomaly characteristics: steel grade strength coefficient 355MPa, width-to-thickness ratio 225, F5 rolling force load rate 0.96, final rolling temperature deviation 35℃, strength difference between adjacent units 280MPa; Trigger rule list: R_F04 (red), R_T03 (red), R_P03 (yellow); Sequence transmission risk: High (probability 0.78); Total number of sequence evaluation units: 19 (current unit, previous 9, next 9); Step 3, Meta-learning adaptive parameter recommendation: Core input data for the model: Abnormal diagnosis result: High risk, core abnormality type "F5 rolling force exceeds limit, final rolling temperature seriously exceeds standard, and strength hard transition".

[0074] Current operating conditions: 30-dimensional fused feature vector (core indicators: F5 rolling force load rate 0.96, final rolling temperature 835℃, width 1800mm, steel grade Q355B, etc.); From 120,000 triplet data points, the top 10 cases with the cosine similarity to the current operating condition were retrieved, with a matching degree of 88%-92%, and were determined to be normal operating conditions.

[0075] The implementation process is as follows: Working condition matching and support set construction: Select the top 10 cases with the highest cosine similarity, and collect the {feature vector, process parameters, production effect} data of these cases to construct a support set (10 samples). Extract the core information of the support set. In this embodiment, the case with the highest similarity is used as an example: steel grade Q355B, thickness 8mm, width 1750mm, F5 rolling force load rate 0.93 (red alert), finishing rolling temperature 820℃ (yellow warning). The recommended parameters are a 3% reduction in F5 reduction rate and a 0.4m / s reduction in finishing rolling speed. After execution, the anomaly is resolved, and the quality is qualified.

[0076] Meta-learning for rapid adaptation: Using general parameters trained offline, the base learner employs a lightweight neural network structure. The input is a 38-dimensional vector (30-dimensional fused feature vector + 8-dimensional anomaly encoding vector, corresponding to rolling force, temperature, planned coding anomalies, and sequence risk propagation anomalies). The 30-dimensional features include standardized values ​​of the steel grade strength coefficient (355 MPa) and width-to-thickness ratio (225). The anomaly encoding is [1,1,0,1,0,0,1,0]. Two hidden layers are used: a 64-dimensional layer using ReLU activation and a 32-dimensional layer using ReLU activation. The output layer is 8-dimensional and uses Tanh activation. The Tanh function in the output layer limits parameter adjustments to [-1,1], and subsequently maps these adjustments to the actual values ​​based on the allowable range. The MSE loss is calculated based on the support set data, and the base learner parameters are updated for each gradient descent step. After adaptation, the query set loss value is 0.03 < 0.05, indicating successful adaptation.

[0077] Preliminary recommended parameters generated: Input the current working condition 38-dimensional vector (30-dimensional features + 8-dimensional anomaly encoding) into the adapted model, and output the 8-dimensional parameter adjustment amount: [-0.02, -0.01, 0, -0.03, -0.05, -0.02, 0.01, -0.08]; mapped to the actual adjustment amount (based on the initial parameters: F1-F7 reduction rate is 45% / 40% / 35% / 30% / 25% / 20% / 10%, finishing rolling speed is 5.2m / s). F1 compression ratio: 45% - 45% × 2% = 44.1%; F2 reduction rate: 40% - 40% × 1% = 39.6%; F3 compression ratio: 35.0% (no adjustment); F4 compression rate: 30% - 30% × 3% = 29.1%; F5 reduction rate: 25% - 25% × 5% = 23.75% (core adjustment, specifically reducing F5 load); F6 compression ratio: 20% - 20% × 2% = 19.6%; F7 compression rate: 10% + 10% × 1% = 10.1%; Finishing rolling speed: 5.2m / s - 5.2m / s×8% = 4.784m / s (reducing the speed controls the final rolling temperature); Preliminary recommended parameters: [44.1%, 39.6%, 35.0%, 29.1%, 23.75%, 19.6%, 10.1%, 4.784 m / s].

[0078] Constraint verification and simulation validation: Constraint verification: All parameters are within the process constraints, and the verification is passed; Rolling simulation verification: After inputting the recommended parameters, the simulation results show that the F5 rolling force load rate is reduced to 0.82, the final rolling temperature is reduced to 802℃, the steel grade strength coefficient has good adaptability, and there are no new anomalies.

[0079] The final output is as follows: Recommended 8-dimensional parameter vector: F1: 44.1%, F2: 39.6%, F3: 35.0%, F4: 28.5%, F5: 23.75%, F6: 20.0%, F7: 10.1%; Finishing speed: 4.784 m / s; Risk level: High risk (1.326) → Low risk (predicted comprehensive risk value 0.38); Key process parameters affecting the rolling force: F5 rolling force load rate (target 0.82), final rolling temperature (target 802℃), and thickness accuracy (target ±0.01mm). Reference historical case numbers: C20230512-08 (92% similarity), C20230603-15 (90% similarity); The adjustment was based on the following: reducing the F5 reduction rate by 5% to resolve the excessive rolling force; reducing the finishing rolling speed by 0.416 m / s to control the final rolling temperature; and fine-tuning the F4 / F6 reduction rates to adapt to the differences in the strength coefficients of the steel grades and alleviate the hard transition in strength.

[0080] The actual rolling results are as follows: Key indicators: F5 rolling force load rate 0.81, final rolling temperature 801℃, steel grade strength coefficient matching is normal, and the plate shape control corresponding to the width-to-thickness ratio of 225 meets the standard; Quality data: Thickness 7.99mm (deviation 0.01mm), crown 8μm, yield strength 358MPa (standard deviation 12MPa), all meet the standards; Finally, the case triplet (containing the key features of steel grade strength coefficient 355MPa and width-to-thickness ratio 225) was included in the historical case library (number C20240118-22) for subsequent incremental training of the meta-learner.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for risk management of hot rolling secondary planning that integrates anomaly diagnosis and adaptive recommendation, characterized in that, include: S1. Collect multi-source production data during the execution of the hot rolling secondary plan, and after preprocessing the multi-source production data, extract the fusion feature vector reflecting the rolling process status. S2. Based on the fused feature vector, a step-by-step diagnosis is performed sequentially through a process rule fast filtering layer, an integrated statistical anomaly detection layer, and a sequence risk propagation assessment layer, outputting the risk level and anomaly type of the current rolling unit, including: S21. The process rule rapid filtering layer takes the basic specification characteristics and some process characteristics of the rolling unit as input, and constructs a structured knowledge base rule system based on the hot rolling production process principle and practical experience. The structured knowledge base rule system includes rolling force rules, temperature control rules, equipment status rules, planning and scheduling rules, and quality association rules. Each rule includes triggering conditions, risk level, and handling suggestions, and outputs a list of alarms triggered by the rules. S22. The integrated statistical anomaly detection layer takes the fused feature vector as input and uses three models in parallel detection: adaptive deep isolated forest, dynamic memory support vector machine, and interpretable variational autoencoder. The model results are fused through Dempster-Shafer evidence theory to determine whether it is a statistical anomaly and to infer the anomaly type by combining the alarm information of the process rule layer. S23. The sequence risk propagation assessment layer uses the fused feature vectors of multiple consecutive rolling units in the current plan and the anomaly scores of the integrated statistical anomaly detection layer as inputs to construct a graph convolutional network model to assess the sequence propagation risk of anomalies; the comprehensive risk value of the rolling unit is calculated by weighted summation, and the comprehensive risk value is divided into three risk levels: low, medium, and high. S3. Based on the aforementioned risk level and anomaly type, and in conjunction with historical production cases, generate process parameter adjustment schemes for the anomaly type using a meta-learning model, including: S31. Divide historical production data into multiple meta-tasks, each meta-task corresponding to a production condition; train the basic learner using a model-independent meta-learning algorithm to obtain general initialization parameters applicable to multiple conditions; the basic learner adopts a neural network structure, with the input being a fusion representation of condition features and abnormal information, and the output being the adjustment amount of process parameters; S32. Define the risk diagnosis results and operating condition characteristics of the current rolling unit as a new meta-task; retrieve several historical cases similar to the current operating condition from the historical case library to form a support set; calculate the similarity through the distance measure or similarity measure between feature vectors; S33. Based on the support set data, starting with the general initialization parameters, a few steps of iterative optimization are performed using the gradient descent method to obtain the basic learner parameters that are suitable for the current working conditions; the number of iteration steps is dynamically adjusted according to the size of the support set. S34. Input the current working condition characteristics and anomaly codes into the adapted basic learner to generate preliminary parameter adjustment amounts; combine process constraints to perform process constraint verification and rolling simulation verification, output the final recommended parameter vector that has passed the verification, and simultaneously output the risk level reduction prediction, key process indicators, and reference historical case numbers. S4. Apply the process parameter adjustment plan to the production process, and update the historical case library and basic learner used for parameter recommendation based on the execution results to form a continuously optimized risk management closed loop.

2. The hot rolling secondary planning risk management method integrating anomaly diagnosis and adaptive recommendation according to claim 1, characterized in that, Step S1 includes: S11. Collect multi-source production data during the execution of the hot rolling secondary plan, including rolling process data, equipment status data and temperature trajectory data obtained through real-time data channels, as well as quality inspection parameters and plan and order data obtained through offline data channels; S12. Preprocess the collected multi-source production data, including data cleaning and data standardization; S13. Perform multi-scale feature extraction, extracting multi-dimensional original features from the preprocessed data, including basic specification features reflecting the product's geometric dimensions and material properties, rolling force energy features reflecting the load distribution and energy consumption characteristics of each stand, temperature field features reflecting the temperature control level of the rolling process, equipment health features reflecting the operating status of key equipment, and quality correlation features reflecting the historical quality control level. S14. Perform dimensionality reduction processing on the multi-dimensional original features to generate a low-dimensional fused feature vector. The dimensionality reduction processing adopts a kernel method or a nonlinear dimensionality reduction method to reduce the feature dimension while retaining the main information, thereby obtaining the fused feature vector.

3. The hot rolling secondary planning risk management method integrating anomaly diagnosis and adaptive recommendation according to claim 2, characterized in that, In step S11, the planning and order data includes steel grade, target thickness, target width, target temperature, and rolling sequence information; the rolling process data includes rolling force, rolling speed, reduction, and main motor operating parameters for each stand; the temperature trajectory data includes roughing mill exit temperature, finishing mill inlet temperature, finishing mill exit temperature, and coiling temperature; the equipment status data includes the cumulative rolling mileage of the work rolls, the effective value of the vibration velocity of the support roll bearings, the main motor bearing temperature, and the hydraulic system pressure; and the quality inspection parameters include finished product thickness, finished product width, crown, flatness, yield strength, and tensile strength.

4. The hot rolling secondary planning risk management method integrating anomaly diagnosis and adaptive recommendation according to claim 1, characterized in that, In step S23, the graph convolutional network model is constructed with each rolling unit as a node, and the node feature is the fusion feature vector of the rolling unit. If two rolling units are adjacent in the production plan, an undirected edge is constructed between the corresponding nodes. The edge weight represents the process coupling strength between the two units, which is calculated by an exponential function to ensure that the more similar the process characteristics of the units, the higher the coupling strength.

5. The hot rolling secondary planning risk management method integrating anomaly diagnosis and adaptive recommendation according to claim 1, characterized in that, In the graph convolutional network model: The graph convolutional network propagation adopts a two-layer network structure, with each layer having the same output dimension. The ReLU function is used as the activation function. Through feature propagation in the two-layer network, each node obtains an embedding vector that incorporates information from its neighboring nodes. Sequence risk prediction inputs node embedding vectors into a fully connected layer and outputs the risk probability of that node in the sequence context.

6. The hot rolling secondary planning risk management method integrating anomaly diagnosis and adaptive recommendation according to claim 1, characterized in that, The graph convolutional network model is trained using historical planning data, labeled with whether or not a sequence-related fault actually occurred, and the model is optimized through supervised learning.

7. The hot rolling secondary planning risk management method integrating anomaly diagnosis and adaptive recommendation according to claim 1, characterized in that, Step S4 includes: S41. After the process parameter adjustment plan is confirmed by the operator, it is sent to the secondary process control system in real time via industrial Ethernet to replace the original control parameters of the current rolling unit and drive the production equipment to execute the rolling process according to the process parameter adjustment plan. S42. Collect the actual production data after the rolling unit executes the process parameter adjustment scheme, including rolling force load rate, final rolling temperature and quality inspection data; S43. Compare the actual production data with the preset success criteria to determine whether the execution effect meets expectations; the success criteria include that the rolling force load rate is within the safe range, the final rolling temperature deviation is within the allowable range, and the quality inspection data meets the product standards. S44. When the execution effect meets the success determination condition, the working condition characteristics, process parameters and production effect of the rolling unit are added to the historical case library in the form of a triplet; the historical case library is periodically deduplicated, the most representative cases are retained, and the total size of the case library is controlled within a preset upper limit. S45. Periodically use newly added successful cases from the historical case library to incrementally train the meta-learning model and update the general initialization parameters of the basic learner so that the basic learner can adapt to the slow changes in the production process.

8. A hot rolling secondary-level plan risk management system based on the hot rolling secondary-level plan risk management method of any one of claims 1-7, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source production data in real time during the execution of the hot rolling secondary plan. After preprocessing the multi-source production data, a fusion feature vector reflecting the rolling process status is extracted. The multi-dimensional anomaly diagnosis module is used to perform step-by-step diagnosis based on the fused feature vector, sequentially through the process rule fast filtering layer, the integrated statistical anomaly detection layer, and the sequence risk propagation assessment layer, and output the risk level and anomaly type of the current rolling unit. The meta-learning adaptive parameter recommendation module is used to generate process parameter adjustment schemes for the aforementioned anomaly type based on the risk level and anomaly type, combined with historical production cases, through a meta-learning model. The closed-loop control and decision output module is used to apply the process parameter adjustment scheme to the production process, and update the historical case library and basic learner for parameter recommendation based on the execution results, forming a continuously optimized risk control closed loop.

Images