A cluster-based continuous annealing furnace knowledge base dynamic learning and parameter control method
Patent Information
- Application Number
- CN202610794031.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-03
- Publication Date
- 2026-09-25
AI Technical Summary
上述方案，改善了传统知识库分档依赖人工规则、新优秀知识学习不足及跨品规检索能力弱的问题，实现了数据驱动的自适应分档、择优更新与智能推荐，可提升连退炉的控制精度和生产稳定性。
Smart Images
Figure CN122816110A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial process control technology, and in particular to a method for dynamic learning and parameter control of a knowledge base for continuous annealing furnaces based on clustering. Background Technology
[0002] In actual production, continuous annealing furnaces (LCF furnaces) need to rationally set key control parameters such as temperature in each furnace zone and internal tension based on strip specifications (including steel grade, thickness, etc.) and real-time operating conditions to ensure that core process indicators such as strip outlet temperature stably reach target values. However, LCF furnaces produce a wide variety of product specifications and experience frequent changes in operating conditions. Relying on manual experience to set parameters results in inconsistencies and difficulties in accumulating and passing on knowledge.
[0003] To address this, continuous annealing furnaces have adopted knowledge base methods in their control, using stored historical production experience to assist in setting control parameters. However, existing knowledge base methods still suffer from the following prominent problems: Product grading relies on manually defined rules to delineate fixed intervals, resulting in grading boundaries that do not match the actual data distribution, and the grading granularity cannot be automatically adjusted as production data accumulates; knowledge update strategies are crude, with simple replacements or fixed-weight averaging methods failing to distinguish control quality differences between different historical data, leading to the dilution or coverage of high-quality knowledge; continuous annealing furnace strip switching states and abnormal data lack effective filtering mechanisms, and direct involvement in knowledge updates significantly reduces the quality of the knowledge base; knowledge retrieval methods are simple, only able to recommend single stable strip states, lacking the ability to accurately recommend continuous annealing furnace strip switching states.
[0004] To address the aforementioned issues, patent CN 116485198 A discloses a knowledge-enhanced performance evaluation method, device, and storage device. It utilizes fuzzy C-means clustering to obtain the optimal combination of key parameters under multiple operating conditions and uses static data clustering to construct a performance evaluation model for the annealing process. Patent CN 121413414 A discloses a method for recommending an expert database for slurry circulation pump groups that integrates K-Means clustering and knowledge experience. It mines historical data using the K-Means clustering algorithm to construct knowledge experience rules, but lacks online learning of excellent knowledge experience. Patent CN 119758914 B discloses a method for optimizing grate cooler equipment parameters based on clustering algorithms. It trains clusters, creates core samples, updates the original sample set using a "replacement" strategy, retrains the model, and then recommends parameters. It is evident that existing methods lack online detection of sample quantity and adaptive clustering optimization, as well as online optimal updating of control parameters, failing to provide adaptive control for continuous annealing furnace operating condition switching.
[0005] Therefore, there is an urgent need for a knowledge base construction and control method for continuous annealing furnaces that can overcome the above-mentioned shortcomings, realize data-driven automatic classification of production conditions, online dynamic accumulation and optimal updating of knowledge, and intelligent recommendation of control parameters, thereby comprehensively improving the quality of continuous annealing furnace process control. Summary of the Invention
[0006] To address the aforementioned technical problems in existing technologies, this invention provides a clustering-based dynamic learning and parameter control method for a continuous annealing furnace knowledge base. This method effectively enables data-driven automatic classification of production conditions, online dynamic optimization of knowledge accumulation, and intelligent parameter recommendation across product specifications. This provides more reliable technical support for the intelligent and adaptive setting of continuous annealing furnace control parameters and the continuous optimization of the knowledge base. The technical solution is as follows: A clustering-based dynamic learning and parameter control method for a continuous annealing furnace knowledge base, the method comprising: S1. Establish a set of control parameters and construct a joint feature vector of production conditions using strip steel specification features and continuous annealing furnace condition detection features; S2. Using the joint feature vector of each sample of historical production data as input, clustering is used to calculate the cluster center and the number of clusters and initialize the corresponding records in the knowledge base; S3. For the current moment, use a sliding window to cache data, determine whether it meets the data quality requirements, find the nearest cluster center of the cached data, and execute a new cluster or a best-preference strategy to update the knowledge base. S4. For the current moment, synchronously determine whether the number of cluster samples exceeds the threshold, perform local clustering, and determine whether to accept the clustering results and update the knowledge base based on the Bayesian information criterion. S5. Determine the state of the strip in the furnace for each parameter in the current control parameter set and make a weighted recommendation of the control parameters.
[0007] In S1: The control parameters include strip exit temperature and tension. The set of control parameters is denoted as {y1, y2, ..., y...}. m}; The strip steel specification features include steel grade, quenching and tempering degree, annealing curve, strip thickness, and width. The character-type features (steel grade, quenching and tempering degree, annealing curve) within these features require one-hot encoding. The strip steel specification features need to be normalized, and single-row feature data can be used to construct strip steel specification feature sub-vectors. n1 represents the number of strip steel specification features.
[0008] The furnace condition detection features of the continuous annealing furnace include the measured temperature, measured tension, measured atmosphere values, and measured unit speed in each furnace zone. These furnace condition detection features need to be normalized, and a single row of feature data can be used to construct a furnace condition detection feature sub-vector. n2 represents the number of furnace condition detection features.
[0009] The joint feature vector x of production conditions is a one-dimensional ordered array formed by orderly splicing of single-row strip steel specification features and continuous annealing furnace condition detection feature data.
[0010]
[0011] The specific process in S2 is as follows: S21. Using the joint feature vector of each sample in the historical production data as input, use a clustering algorithm to determine the cluster centers. Calculation of the number of clusters K.
[0012] For example, the X-means clustering algorithm can be used to automatically determine the number of clusters based on the actual distribution density of the data; the more data there is, the denser the clusters. The joint feature vector of the samples is the joint feature vector corresponding to the historical data. The historical production data needs to exclude data with multiple coils of steel in the furnace.
[0013] S22. The knowledge base is initialized using the tuple (i, k) consisting of the control parameter i and the cluster number k as the unique index key. Each record in the knowledge base includes the mean vector of furnace condition detection features. Comprehensive score for quality control Cluster sample size . The calculation formula is:
[0014] In the formula, The weighting coefficients are satisfied. ; These are the control parameters corresponding to the k-th cluster. and target values of control parameters Mean absolute error, mean square error, and coefficient of determination.
[0015] Among them, the The mean of the feature vectors for each sample furnace condition in the corresponding cluster is detected; This represents the number of samples in the corresponding cluster.
[0016] The specific process in S3 is as follows: S31. Continuously cache the joint feature vector using a sliding window of time length T to obtain the cached joint feature two-dimensional matrix. .
[0017] Where T is a hyperparameter, determined through experiments or human experience; S32, Cache joint feature matrix Implement conditional checks; proceed to subsequent steps only if all conditions are met; otherwise, the knowledge base will not be updated. Conditions include: 1) Condition A, steady-state detection: Determine if the cache period falls under a non-steady-state condition; if so, skip the update. Non-steady-state conditions include multiple coils of strip steel simultaneously experiencing furnace and unit speed variations exceeding a threshold. .
[0018] 2) Condition B, anomaly detection: Outlier detection is performed on each control parameter sequence in the cached data (based on the 3σ criterion or interquartile range method). If the proportion of outlier data points in a certain sequence exceeds a preset threshold, the outlier detection is performed. If the data is invalid, the update will be skipped. in, Determined through experiments or human experience Determined through experiments or human experience.
[0019] S33, Confirm X Cluster of Belonging and minimum distance : calculate The mean of the joint feature vector of the corresponding production conditions ,calculate Distances to all existing cluster centers (e.g., Euclidean distances) :
[0020] Choose the cluster with the smallest distance as the current cluster:
[0021]
[0022] in, The function iterates through all possible... k Find The smallest one k Assign a value to ; S34. Execute the new cluster or preferred update strategy, following the logical branches below: 1) Scenario 1: The current production conditions belong to a new region, and a new cluster is being built: The current production conditions belong to a new area not yet covered by the knowledge base. Create a cluster around the center, and simultaneously create a new knowledge base record for each of the m control parameters. and initialize the corresponding This enables the autonomous expansion of the knowledge base; in, The preset maximum coverage radius is determined through experiments or human experience.
[0023] 2) Scenario Two: The current production conditions are assigned to an existing cluster, and a selective update is performed: For each control parameter i, update the knowledge base record according to the following rules. : Calculate the i-th control parameter based on the current cached data. New comprehensive score under the cluster Compared with the ratings already stored in the knowledge base Comparison: like (The old one is better), the knowledge base is not updated; like (New and Better), perform a weighted fusion update based on the score:
[0024]
[0025]
[0026] In the formula, This is the updated mean vector of furnace condition detection features in the knowledge base. The updated knowledge base control quality comprehensive score, This represents the updated sample size. To update the mean vector of furnace condition detection features in the previous knowledge base, This is the feature vector for detecting the furnace condition at the current moment. To update the overall quality score of the knowledge base before the update, The overall control quality score at the current moment. This represents the maximum coverage radius.
[0027] Ensure that the control parameters of each (i,k) record in the knowledge base always evolve towards better control effects, and at the same time retain the effective components of historical high-quality knowledge through weighted fusion.
[0028] The specific process in S4 is as follows: S41. For the current moment, synchronously determine whether the number of cluster samples exceeds [the limit]. If the threshold is exceeded, local clustering is performed within the clusters that exceed the threshold; if the threshold is not exceeded... Skip S4 and proceed to S5.
[0029] in, Determined through experiments or human experience.
[0030] S42. Determine whether to accept the clustering results and update the knowledge base based on the Bayesian Information Criterion (BIC). The determination process is as follows: 1) If the BIC criterion supports subdivision and accepts the results of local clustering, then the cluster is split into several subclusters. The centers and number of subclusters are reinitialized, and the knowledge base records corresponding to the subclusters continue the original clusters' records. Number of samples in sub-clusters Reinitialize.
[0031] 2) If the BIC criterion does not support subdivision and does not accept the results of local clustering, it means that the data within the cluster is highly homogeneous. The current cluster will be maintained, the grading will remain unchanged, and the cluster will not be updated.
[0032] The specific process in S5 is as follows: S51, Confirm The corresponding monitoring furnace area (e.g.) (This refers to the tension in the heating section, corresponding to the monitored furnace area being the heating section). The amount of strip steel in the furnace at the current moment is checked, and two scenarios are handled: 1) Scenario 1: The corresponding furnace area is a single coil of strip steel (steady state). Obtain the real-time specification information of the strip steel at the current moment and the furnace condition detection data, and construct the joint feature vector of the production conditions at the current moment. ,calculate Choose the cluster with the Euclidean distance from all cluster centers. of Recommended as a control parameter after inverse normalization; 2) Scenario Two: The corresponding furnace area contains multiple coils of strip steel (transitional state). The strip steel in this furnace area is denoted as... Each has a furnace length of The total length of the furnace area is For each coil of strip steel Each product specification information and furnace condition detection data is used to construct a corresponding joint feature vector of production conditions. Calculate the distance to the center of the current cluster and query the nearest cluster. control parameters ,Will Recommended as a control parameter after inverse normalization; Weighting is performed using the following formula:
[0033] S52. Summarize the cases in S51, output all recommended control parameters, and complete the control parameter optimization. After making a recommendation, the process continues to process the next control parameter until all m parameters are traversed, and the complete control parameters for the current moment are summarized for recommendation.
[0034] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: The above solution improves upon the problems of traditional knowledge bases relying on manual rules for classification, insufficient learning of new and excellent knowledge, and weak cross-product specification retrieval capabilities. It achieves data-driven adaptive classification, optimal updating, and intelligent recommendation, which can improve the control accuracy and production stability of continuous annealing furnaces.
[0035] This method can not only effectively extract and reuse historical production experience, but also continuously evolve itself in continuous production, providing a data-driven intelligent parameter decision-making scheme with self-learning capabilities for the coupled control of multiple temperature zones in continuous annealing furnaces. Attached Figure Description
[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0037] Figure 1 This is a flowchart of a clustering-based dynamic learning and parameter control method for a continuous annealing furnace knowledge base, provided by an embodiment of the present invention. Figure 2 This is a curve showing the measured and target values of the strip outlet temperature in the continuous annealing furnace heating section during the online knowledge base update process in this embodiment of the invention. Figure 3 The comprehensive score change curve corresponding to the index key (1,1) of the knowledge base online update process in this embodiment of the invention; Figure 4 This is a curve showing the change of the recommended value of the measured temperature in zone 1 of the heating section during the online update process of the knowledge base in this embodiment of the invention. Detailed Implementation
[0038] The technical solution of the present invention will now be described with reference to the accompanying drawings.
[0039] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.
[0040] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.
[0041] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0042] This invention provides a clustering-based method for dynamic learning and parameter control of a continuous annealing furnace knowledge base. For example... Figure 1 The flowchart shown is for a cluster-based dynamic learning and parameter control method for a continuous annealing furnace knowledge base. This method may include the following steps: S1. Establish a set of control parameters and construct a joint feature vector of production conditions using strip steel specification features and continuous annealing furnace condition detection features; S2. Using the joint feature vector of each sample of historical production data as input, clustering is used to calculate the cluster center and the number of clusters and initialize the corresponding records in the knowledge base; S3. For the current moment, use a sliding window to cache data, determine whether it meets the data quality requirements, find the nearest cluster center of the cached data, and execute a new cluster or a best-preference strategy to update the knowledge base. S4. For the current moment, synchronously determine whether the number of cluster samples exceeds the threshold, perform local clustering, and determine whether to accept the clustering results and update the knowledge base based on the Bayesian information criterion. S5. Determine the state of the strip in the furnace for each parameter in the current control parameter set and make a weighted recommendation of the control parameters.
[0043] The following description, in conjunction with specific embodiments, illustrates this point.
[0044] In the specific implementation process, the following steps shall be followed: S1. Establish a set of control parameters and construct a joint feature vector of production conditions using strip steel specification features and continuous annealing furnace condition detection features.
[0045] The set of control parameters is denoted as {y1, y2, ..., y...}. m In this example, the control parameter set is: The meaning is the strip temperature (°C) at the outlet of the heating section of the continuous annealing furnace. The target temperature for the heating section of the continuous annealing furnace is 710℃.
[0046] The strip steel specification features should at least include steel grade, quenching and tempering degree, annealing curve, strip thickness, and width; the character-type features of the strip steel specification need to be one-hot encoded. The strip steel specification features need to be normalized, and feature sub-vectors can be constructed. n1 represents the number of strip steel specification features.
[0047] In this example, the categorical features include steel grade, quenching and tempering degree, and annealing curve; the numerical features include strip thickness (mm) and width (mm), n1=5. The character features of steel grade, quenching and tempering degree, and annealing curve are converted into numerical values using unique thermal coding, and normalized using the min-max method.
[0048] The furnace condition detection characteristics of a continuous annealing furnace can include the measured temperature of each furnace zone, the measured tension of each furnace zone, the measured atmosphere value inside the furnace, and the measured speed of the unit. These furnace condition detection characteristics need to be normalized, and a furnace condition detection feature sub-vector can be constructed. n2 represents the number of furnace condition detection features.
[0049] In this example, the real-time detection data includes the measured temperature (°C) of each heating section 1, the measured temperature (°C) of each heating section 2, and the measured speed of the unit (m / min), n2=3, and normalization is performed using the min-max method.
[0050] The joint feature vector x of production conditions is a one-dimensional ordered array formed by orderly splicing of single-row strip steel specification features and continuous annealing furnace condition detection feature data.
[0051]
[0052] S2. Using the joint feature vector of each sample of historical production data as input, clustering is used to calculate the cluster center and the number of clusters and initialize the corresponding records in the knowledge base.
[0053] S21, using the joint feature vector of each sample in the historical production data as input, uses a clustering algorithm to determine the cluster centers. Number of clusters K The calculation.
[0054] In this example, the X-means clustering algorithm is used, with the joint feature vector of each sample in the historical production data over a period of time as input. The number of clusters is 4, and the cluster centers are shown in the table below: Table 1
[0055] S22, the knowledge base is initialized using the tuple (i, k) consisting of the control parameter i and the cluster number k as the unique index key. Each record in the knowledge base includes the mean vector of furnace condition detection features. Comprehensive score for quality control Cluster sample size . The calculation formula is:
[0056] In the formula, The weighting coefficients are satisfied. ; These are the control parameters corresponding to the k-th cluster. and target values of control parameters The mean absolute error, mean square error, and coefficient of determination; This represents the mean of the feature vectors for furnace condition detection in the corresponding cluster. This represents the number of samples in the corresponding cluster. In this example, .
[0057] In this example, the knowledge base initialization records are shown in Table 2 below: Table 2
[0058] S3. For the current moment, use a sliding window to cache the data. After determining whether the data meets the quality requirements, find the nearest cluster center to the cached data and execute either a new cluster creation or a best-preference strategy to update the knowledge base. S31. Continuously cache the joint feature vector using a sliding window of time length T to obtain the cached joint feature two-dimensional matrix. .
[0059] Where T is a hyperparameter, determined through experiments or human experience; S32, Cache joint feature matrix Implement conditional checks; proceed to subsequent steps only if all conditions are met; otherwise, the knowledge base will not be updated. Conditions include: 1) Condition A, steady-state detection: Determine if the cache period is a non-steady state; if so, skip the update. The unsteady state includes multiple coils of strip steel simultaneously experiencing speed changes in the furnace and unit exceeding a threshold. ; 2) Condition B, anomaly detection: Outlier detection is performed on each control parameter sequence in the cached data (based on the 3σ criterion or interquartile range method). If the proportion of outlier data points in a certain sequence exceeds a preset threshold, the outlier detection is performed. If the data is invalid, the update will be skipped. in, Determined through experiments or human experience Determined through experiments or human experience.
[0060] S33, Confirm X Cluster of Belonging and minimum distance : calculate The mean of the joint feature vector of the corresponding production conditions ,calculate Distances to all existing cluster centers (e.g., Euclidean distances) :
[0061] Choose the cluster with the smallest distance as the current cluster:
[0062]
[0063] in, The function iterates through all possible... k Find The smallest one k Assign a value to ; S34. Execute the new cluster or preferred update strategy, following the logical branches below: 1) Scenario 1: The current production conditions belong to a new region, and a new cluster is being built: The current production conditions belong to a new area not yet covered by the knowledge base. Create a cluster around the center, and simultaneously create a new knowledge base record for each of the m control parameters. and initialize the corresponding This enables the autonomous expansion of the knowledge base; in, The preset maximum coverage radius is determined through experiments or human experience.
[0064] 2) Scenario Two: The current production conditions are assigned to an existing cluster, and a selective update is performed: For each control parameter i, update the knowledge base record according to the following rules. : Calculate the i-th control parameter based on the current cached data. New comprehensive score under the cluster Compared with the ratings already stored in the knowledge base Comparison: like (The old one is better), the knowledge base is not updated; like (New and Better), perform a weighted fusion update based on the score:
[0065]
[0066]
[0067] In the formula, This is the updated mean vector of furnace condition detection features in the knowledge base. The updated knowledge base control quality comprehensive score, This represents the updated sample size. To update the mean vector of furnace condition detection features in the previous knowledge base, This is the feature vector for detecting the furnace condition at the current moment. To update the overall quality score of the knowledge base before the update, The overall control quality score at the current moment. This represents the maximum coverage radius.
[0068] In this example, T is set to 40, representing 40 seconds.
[0069] In this example, For a given moment, the joint feature matrix of the decommissioned cache is such that conditions A and B are both satisfied, and the knowledge base is updated.
[0070] In this example, , The values are 0.07, 0.32, 1.52, and 1.44 respectively. The cluster k with the smallest distance is numbered 1. , .
[0071] In this example, , , The (older, better) knowledge base is not being updated.
[0072] S4. For the current moment, synchronously determine whether the number of cluster samples exceeds the threshold, perform local clustering, and determine whether to accept the clustering results and update the knowledge base based on the Bayesian information criterion.
[0073] In this example, since no knowledge base update occurs in S3, the number of samples in each cluster does not increase; therefore, the number of clusters does not exceed [a certain limit]. The tiers remain unchanged, and the clusters are not updated.
[0074] S5. Determine the state of the strip in the furnace for each parameter in the current control parameter set and make a weighted recommendation of the control parameters.
[0075] S51, Confirm The corresponding monitoring furnace area (e.g.) (This refers to the tension in the heating section, corresponding to the monitored furnace area being the heating section). The amount of strip steel in the furnace at the current moment is checked, and two scenarios are handled: 1) Scenario 1: The corresponding furnace area is a single coil of strip steel (steady state). Obtain the real-time specification information of the strip steel at the current moment and the furnace condition detection data, and construct the joint feature vector of the production conditions at the current moment. ,calculate Choose the cluster with the Euclidean distance from all cluster centers. of Recommended as a control parameter after inverse normalization; 2) Scenario Two: The corresponding furnace area contains multiple coils of strip steel (transitional state). The strip steel in this furnace area is denoted as... Each has a furnace length of The total length of the furnace area is For each coil of strip steel Each product specification information and furnace condition detection data is used to construct a corresponding joint feature vector of production conditions. The query retrieves the nearest cluster. control parameters The weighting is performed using the following formula:
[0076] Will Recommended as a control parameter after inverse normalization; S52. Summarize and output all recommended control parameters to complete the control parameter recommendation process. After making a recommendation, the process continues to process the next control parameter until all m parameters are traversed, and the complete control parameters for the current moment are summarized for recommendation.
[0077] In this example, the current production conditions are for single-coil strip steel (steady state), and the joint eigenvector at the current moment... The Euclidean distances between this cluster and the middle of all clusters are 0.07, 0.39, 1.53, and 1.44, respectively. Therefore, the nearest cluster is k=1, and it is... After inverse normalization, the recommended values for the control parameters are [1080, 1090, 120].
[0078] In this example, regarding control parameters The recommended values for the furnace condition parameters (measured temperature of heating section 1 hf1_fcetmp, measured temperature of heating section 2 hf2_fcetmp, and measured unit speed) are 1080℃, 1090℃, and 120m / min, respectively.
[0079] Experimental results: The dynamic learning and parameter control method for a knowledge base based on clustering provided in this embodiment of the invention is applied to a data sampling system for a continuous annealing furnace production line. Figure 2 After initializing the knowledge base, the control parameters for the online update process of the knowledge base (including the S3 dynamic self-learning and S4 adaptive grading processes) are as follows: (Measured strip temperature at the outlet of the heating section of the continuous annealing furnace) and The curve showing the change of the target temperature of the strip at the outlet of the heating section of the continuous annealing furnace (hf_exit_target_tmp) over a certain period of time. Figure 3 The online update process knowledge base index key (1,1) records the corresponding evolution process of the comprehensive quality score. Figure 4 The curve showing the change of the recommended parameter hf1_fcetmp in the knowledge base during online updates.
[0080] The three sets of curves above reveal significant dynamic characteristics of continuous learning and parameter evolution within the knowledge base. Around 60 seconds, the control quality of the system significantly improves as the measured temperature hf_exit_tmp converges towards the target temperature hf_exit_target_tmp. At this point, the online evaluation mechanism of the knowledge base detects the improved control effect, and the comprehensive score Q of the corresponding cluster increases synchronously, triggering a selective update of the knowledge base records. After the update, the control parameter vector associated with that cluster is optimized, leading to an adaptive adjustment of the recommended parameter hf1_fcetmp. Similarly, between 340 and 360 seconds, the control deviation decreases, and the knowledge base again captures this positive change, manifested as a further increase in the Q value, subsequently completing a weighted update of the parameters within the cluster and a refresh of the recommended values. This demonstrates that during the update process, the knowledge base can intelligently detect the quality of control-related parameters and continuously update and optimize, intelligently recommending parameters.
[0081] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for dynamic learning and parameter control of a knowledge base for continuous annealing furnaces based on clustering, characterized in that, The method includes: S1. Establish a set of control parameters and construct a joint feature vector of production conditions using strip steel specification features and continuous annealing furnace condition detection features; S2. Using the joint feature vector of each sample of historical production data as input, clustering is used to calculate the cluster center and the number of clusters and initialize the corresponding records in the knowledge base; S3. For the current moment, use a sliding window to cache data, determine whether it meets the data quality requirements, find the nearest cluster center of the cached data, and execute a new cluster or a best-preference strategy to update the knowledge base. S4. For the current moment, synchronously determine whether the number of cluster samples exceeds the threshold, perform local clustering, and determine whether to accept the clustering results and update the knowledge base based on the Bayesian information criterion. S5. Determine the state of the strip in the furnace for each parameter in the current control parameter set and make a weighted recommendation of the control parameters.
2. The method for dynamic learning and parameter control of a knowledge base for continuous annealing furnace based on clustering as described in claim 1, characterized in that, In S1: The control parameters include strip exit temperature and tension. The set of control parameters is denoted as {y1, y2, ..., y...}. m }; The strip steel specification features include steel grade, quenching and tempering degree, annealing curve, strip thickness and width; character-type features in the strip steel specification features need to be individually heat-coded. The furnace condition detection characteristics of the continuous annealing furnace include the measured temperature of each furnace zone, the measured tension of each furnace zone, the measured value of the atmosphere inside the furnace, and the measured speed of the unit. The joint feature vector of production conditions is a one-dimensional ordered array formed by orderly splicing of single-row strip steel specification features and continuous annealing furnace condition detection feature data.
3. The method for dynamic learning and parameter control of a knowledge base for continuous annealing furnace based on clustering as described in claim 1, characterized in that, S2 specifically includes: S21. Using the joint feature vector of each sample in the historical production data as input, use a clustering algorithm to determine the cluster centers. Number of clusters K Calculation; S22. The knowledge base is initialized using the tuple (i, k) consisting of the control parameter i and the cluster number k as the unique index key. Each record in the knowledge base includes the mean vector of furnace condition detection features. Comprehensive score for quality control Cluster sample size .
4. The method for dynamic learning and parameter control of a knowledge base for continuous annealing furnace based on clustering according to claim 3, characterized in that, The clustering algorithm is the X-means algorithm.
5. The method for dynamic learning and parameter control of a knowledge base for a continuous annealing furnace based on clustering as described in claim 3, characterized in that, The overall control quality score The calculation formula is as follows: In the formula, The weighting coefficients are satisfied. ; These are the control parameters corresponding to the k-th cluster. and target values of control parameters Mean absolute error, mean square error, and coefficient of determination.
6. The method for dynamic learning and parameter control of a knowledge base for continuous annealing furnace based on clustering according to claim 1, characterized in that, Specifically, S3 includes: S31. Continuously cache the joint feature vectors using a sliding window of time length T to obtain the joint feature two-dimensional matrix. ; S32, to Implement conditional judgments; allow subsequent steps to be executed only if all conditions are met; otherwise, the knowledge base will not be updated. S33, Confirm X Cluster of Belonging and minimum distance ; S34. Execute the new cluster or the optimal update strategy.
7. The method for dynamic learning and parameter control of a knowledge base for continuous annealing furnace based on clustering as described in claim 6, characterized in that, The conditions in S32 include: 1) Condition A, steady-state detection: Determine whether the cache period is a non-steady state. If it is, skip the update; otherwise, condition A is met. 2) Condition B, anomaly detection: Outlier detection is performed on each control parameter sequence in the cached data. If the proportion of outlier data points in a certain sequence exceeds a preset threshold... If the current data is invalid, the update will be skipped; otherwise, condition B is met.
8. The method for dynamic learning and parameter control of a knowledge base for continuous annealing furnace based on clustering according to claim 6, characterized in that, The specific execution process in S34 is as follows: 1) Scenario 1: The current production conditions belong to a new region, so new clusters and knowledge base records are being created; 2) Scenario Two: The current production conditions are assigned to an existing cluster, and a selective update is performed: For each control parameter i, update the knowledge base record according to the following rules. : Calculate the i-th control parameter based on the current cached data. New comprehensive score under the cluster Compared with the ratings already stored in the knowledge base Comparison: like The knowledge base is not updated; like Perform a weighted fusion update based on the score: In the formula, This is the updated mean vector of furnace condition detection features in the knowledge base. The updated knowledge base control quality comprehensive score, This represents the updated sample size. To update the mean vector of furnace condition detection features in the previous knowledge base, This is the feature vector for detecting the furnace condition at the current moment. To update the overall quality score of the knowledge base before the update, The overall control quality score at the current moment. This represents the maximum coverage radius.
9. The method for dynamic learning and parameter control of a knowledge base for a continuous annealing furnace based on clustering as described in claim 1, characterized in that, S4 specifically includes: S41. For the current moment, synchronously determine whether the number of cluster samples exceeds the threshold. Local clustering is performed within clusters that exceed a threshold; S42. Based on the Bayesian Information Criterion (BIC), determine whether to accept the clustering results and update the knowledge base. The determination process is as follows: 1) If the BIC criterion supports subdivision and accepts the results of local clustering, then the cluster is split into several subclusters. The centers and number of subclusters are reinitialized, and the knowledge base records corresponding to the subclusters continue the original clusters' records. Number of samples in sub-clusters Reinitialize; 2) If the BIC criterion does not support subdivision, the current cluster is maintained and the cluster is not updated.
10. The method for dynamic learning and parameter control of a knowledge base for a continuous annealing furnace based on clustering according to claim 1, characterized in that, Specifically, S5 includes: S51, Confirm For the corresponding monitored furnace area, check the quantity of strip steel in the furnace at the current moment, and handle it in two ways: 1) Scenario 1: The corresponding furnace area is for single coil strip steel. Construct the joint feature vector of the production conditions at the current moment. ,calculate The Euclidean distances to all cluster centers yield the nearest cluster. The number was found in the knowledge base records. The denormalized value is recommended as the control parameter. 2) Scenario Two: Multiple coils of strip steel exist in the corresponding furnace area Each has a furnace length of The total length of the furnace area is For each coil of strip steel , respectively construct Calculate the distance to the center of the current cluster and find the nearest cluster. And the knowledge base records ,Will Recommended as a control parameter after inverse normalization; The The calculation is performed using the following formula: S52. Summarize the situations in S51 to form complete control parameters for the current moment and make recommendations.