A metal recovery rate prediction method and system based on Bayesian theory

Abstract

The application discloses a metal recovery rate prediction method based on Bayesian theory, converts the characteristics of a furnace into a standard form, and simultaneously carries out clustering and KD tree construction on the obtained furnace standardized characteristics; based on the KD tree index structure, after inputting new furnace characteristics, the furnace data of the current furnace is standardized first, and the nearest several nodes to the new furnace characteristics are matched based on the KD tree index structure; the KNN classification is carried out on the new furnace standardized characteristics in combination with the furnace characteristic classification result, the type with the largest proportion of the type quantity in the node is obtained as the classification data of the current furnace; and the Bayesian prediction estimation is carried out in combination with the current furnace classification data, the clustered furnace classification result and the furnace characteristic statistical information in the furnace characteristic database. The application utilizes data analysis and machine learning, combines historical furnace data and parameters, establishes an accurate model, and solves the problem that the existing technology cannot accurately calculate the metal recovery rate due to unstable tapping and remaining steel quantity.

Description

TECHNICAL FIELD

[0001] The present application relates to a metal recovery rate prediction method, in particular to a metal recovery rate prediction method and system based on Bayesian theory. BACKGROUND

[0002] In the electric furnace hot charging molten iron process system, the complex and fluctuating material conditions in the furnace are balanced in real time, and the metal recovery rate is accurately predicted, which can effectively shorten the electric furnace smelting cycle, reduce energy and material consumption, and achieve the purpose of energy saving and consumption reduction. In the actual electric furnace smelting process, there is a big difference in the amount of alloy added for different steel grades during production, and there is a certain amount of steel and slag left in the furnace, which leads to unstable tapping and steel left in the electric furnace, which brings great interference to the calculation of the metal recovery rate, and further affects the calculation of the actual tapping molten steel quantity and heat balance.

[0003] In order to eliminate the interference of abnormal data and improve the accuracy of metal recovery rate prediction, the current electric furnace smelting process mainly uses traditional experience and manual monitoring and operation. The experience method is to obtain a reasonable metal recovery rate constant through experiments, and to use a fixed metal recovery rate to support the electric furnace hot charging molten iron process system. In actual production, the factors such as molten iron scrap ratio and molten iron chemical elements affecting the metal recovery rate are constantly changing, so the use of constant value cannot accurately represent the metal recovery rate of each furnace, resulting in that the electric furnace hot charging molten iron process system cannot work better. Manual monitoring and operation is to calculate the input-output ratio of adjacent furnace times, and the calculation results are used to guide the operation of the subsequent electric furnace hot charging molten iron process system. This method still cannot adapt to the constantly changing and unstable factors in the electric furnace steelmaking process.

[0004] In view of the shortcomings of experience and manual monitoring and operation, some scholars begin to introduce advanced machine learning technology for steel smelting control process, such as using at least two machine learning-based methods to train a prediction sub-model, and then using Bayesian algorithm to couple multiple training results to determine the final prediction model, so as to improve the hit rate of system prediction, but the method of integrating multiple models undoubtedly increases the complexity of the system, and multiple training models may be affected by the same or similar training data and bias sources, resulting in correlation between models, and further increasing the risk of data overfitting.

[0005] At present, the reason for inaccurate metal recovery rate prediction is that the feature correlation between furnaces cannot be mined, resulting in that relevant furnace data cannot be used to guide the calculation of the metal recovery rate of the current furnace. SUMMARY

[0006] Invention purposes: The purpose of the present application is to solve the problems existing in the prior art, provide a metal recovery rate prediction method and system based on Bayesian theory, use advanced data analysis and machine learning technology, combine a large amount of historical furnace data and related parameters, and establish an accurate prediction model to assist operators in making scientific decisions. The present application is aimed at the actual problem that the unstable tapping and remaining steel quantity in the electric furnace smelting process leads to the difficulty in accurately calculating the metal recovery rate, and overcomes the shortcomings of unstable prediction and excessive resource consumption of the existing method.

[0007] Technical scheme: A metal recovery rate prediction method based on Bayesian theory, comprising the following steps:

[0008] Step 1, Z-score feature standardization; first, obtain all the furnace features, the average value μ and the standard deviation σ of each attribute column furnace feature element in the furnace feature database, then convert all the furnace features into a standard form by Z-score standardization to eliminate the influence of dimension, and obtain the furnace standardized features;

[0009] Step 2, for the furnace standardized features obtained in step 1, K-means++ clustering and KD tree construction are simultaneously performed;

[0010] When K-means++ clustering is performed, all furnaces are first divided into D groups, and in actual application, the parameter D should be adjusted according to the final prediction result to achieve the best classification effect, and the type number of each furnace feature attribute is obtained. The furnace classification result is saved to the type attribute of the corresponding record in the furnace feature database; the statistical data required in Bayesian estimation can be conveniently obtained by means of database sql query;

[0011] KD tree construction, each node of the KD tree stores the furnace standardized features and the furnace classification result, the purpose is to find the nearest C nodes to the current furnace standardized feature data with the smallest time complexity, and the setting of the parameter C needs to be adjusted according to the specific prediction result in application;

[0012] Step 3, based on the KD tree index structure constructed in step 2, the new furnace features are input, the furnace data of the current furnace is standardized, and based on the KD tree index structure constructed in step 2, the nearest nodes to the new furnace features are matched, the KNN classification of the new furnace standardized features is performed combined with the furnace feature classification result obtained by K-means++, and the type with the largest proportion of type number in the node is obtained as the classification data of the current furnace;

[0013] Step 4, combining the current furnace classification data obtained in step 3, the K-means++ clustering furnace classification result obtained in step 2, and the furnace feature statistical information in the furnace feature database, Bayesian prediction estimation is performed.

[0014] Further, the specific process of Bayesian prediction estimation is as follows:

[0015] Step 4.1, the number of all furnace characteristics in the current database is obtained from the furnace characteristic database, and is recorded as n. The number of all records of the same type B as the current new furnace is counted, and is recorded as countB. Thus, the probability of the specific furnace type can be calculated, as shown in formula (2):

[0016]

[0017] Step 4.2, the number of records of the metal recovery rate A in the current database is obtained as countA. Thus, the probability of a specific metal recovery rate A is calculated, as shown in formula (3):

[0018]

[0019] Step 4.3, the number of records of the metal recovery rate A and the furnace type B is obtained as countAB. Thus, the probability of a specific furnace type B under a specific recovery rate A is calculated, as shown in formula (4):

[0020]

[0021] Step 4.4, the maximum metal recovery rate P(A|B) in the classification to which the current furnace belongs is predicted and estimated by the Bayesian method shown in formula (5), which is used as the final metal recovery rate prediction value.

[0022]

[0023] Further, the calculation formula of the Z-score feature standardization in step 1 is as follows:

[0024]

[0025] All furnace characteristics values are converted to -1-1 by formula (1). The standardized furnace characteristics convert different order of magnitude characteristics to the same order of magnitude, eliminate the influence of dimension, and ensure the comparability between data. The attributes in the furnace characteristic database include stoveId, ironWaterScrapWeightPro, ironWaterTem, cRate, mnRate, and standardTemperature.

[0026] Further, the specific method of automatically clustering the furnace characteristics by K-means++ clustering processing of the furnace standardized characteristics in step 2 is as follows:

[0027] Firstly, find K cluster centers from all the normalized features of the furnaces, randomly select a sample point from the normalized features of the furnaces as the first initial cluster center;

[0028] Next, calculate the shortest distance between each sample and the current existing cluster center;

[0029] Then, calculate the probability of each sample point being selected as the next cluster center, the calculation method of the sample point probability is: the shortest distance between the current sample point and the cluster center / the sum of the shortest distances between all sample points and the sample center;

[0030] Finally, select the sample point corresponding to the maximum probability value as the next cluster center, repeatedly repeat the above center finding method until K cluster centers are selected, number the K cluster centers with different numbers respectively, the same number indicates that the sample points belong to the same class; continue to cluster all samples into K classes, distribute the sample set to the nearest cluster center according to the minimum distance principle, and modify the sample number, update the cluster center using the sample mean of each cluster; repeatedly repeat the above clustering process until the cluster center no longer changes;

[0031] Through K-means++ clustering, the type number of each furnace feature can be finally obtained; and the classification result thereof is updated to the type attribute of the corresponding record in the database;

[0032] The distance between sample points in K-means++ is the Euclidean distance, and the distance formula is shown in formula (6):

[0033]

[0034] Wherein, X i represents the i-th dimensional feature of the sample point X, Y i represents the i-th dimensional feature of the sample point Y.

[0035] Further, the step 4 first determines the newly input furnace features, and the newly input furnace features are subjected to Z-score standardization processing; then searches K nearest sample points according to the index structure of the KD tree obtained in step 2, and the specific search method is:

[0036] 1) Create a large root heap to store the distance information of the K nearest nodes;

[0037] 2) Find the leaf node containing the target point x in the KD tree index structure: start from the root node, recursively search the KD tree downward: if the coordinate of the target point x in the current dimension is less than the coordinate of the split point, move to the left child node, otherwise move to the right child node, until the leaf node is reached; this leaf node is the "current nearest point" and is added to the large root heap;

[0038] 3) The method for adding to the max-heap is to recursively backtrack upwards, performing the following operations on each node: If the instance point stored in this node is closer to the target point than the current top node of the max-heap, then if the number of elements in the max-heap is less than K, add it directly; if it is greater than K, delete the first element and add the new element to the max-heap. Update the distance of the "current nearest point" to the distance between the first element of the max-heap and the target point. Simultaneously, check if there is a closer point in the region corresponding to the other child node of the child node's parent node. Specifically:

[0039] Check if the region corresponding to the other child node intersects with a circle or hypersphere centered at the target point and with the distance between the target point and the "current nearest point" as its radius: if they intersect, there may be a point closer to the target point in the region corresponding to the other child node, move to the other child node, and then continue the nearest neighbor search recursively; if they do not intersect, backtrack upwards.

[0040] 4) When the search ends at the root node, the nodes stored in the max-heap are the K nodes closest to the target node.

[0041] This invention also discloses a metal recovery rate prediction system based on Bayesian theory, characterized by comprising a Z-score feature normalization module, a K-means++ clustering module, a KD-tree construction module, a KNN classification module, and a Bayesian prediction module.

[0042] The Z-score feature standardization module is used to eliminate the influence of units on the classification results. Statistical analysis is performed on all data attributes in the furnace feature database table to obtain the mean and variance of each attribute column in the feature database, denoted as μ and σ, respectively. Based on the mean and variance of each attribute column, Z-score standardization is performed on the corresponding attribute value of each record. Z-score standardization is shown in formula (1):

[0043]

[0044] Z-score standardization converts all furnace feature values ​​to a range of -1 to 1. Standardized furnace features transform features of different magnitudes into the same magnitude, eliminating the influence of dimensions and ensuring data comparability. Specific attributes in the furnace feature database table include `stoveId`, `ironWaterScrapWeightPro`, `ironWaterTem`, `cRate`, `mnRate`, and `standardTemperature`.

[0045] The K-means++ clustering module performs cluster analysis on the standardized furnace features. Feature samples with high similarity are clustered into the same cluster to obtain furnace features of different clusters, i.e., to generate classification results for different furnace features. These results are then saved to the `type` attribute column of the `stove` table in the furnace feature database. Specific attributes included in the `stoveId`, `ironWaterWeight`, `ironWaterTem`, `ironScrapWeight`, `steelTappingWeight`, `steelTappingTem`, `powerCompression`, and `type`.

[0046] The KD-tree construction module addresses the KNN classification algorithm's need to find the k nearest neighbors of a given furnace feature. To optimize search efficiency, this method constructs a KD-tree from the standardized furnace data, serving as the index structure for storing training data samples. The KD-tree structure is as follows: Figure 2 As shown.

[0047] The KNN classification module is based on the index structure of the KD tree. When the user inputs a new furnace feature, the KD tree quickly matches the few nodes closest to the new furnace feature. Then, the majority voting method is used, combined with the furnace feature classification results obtained by K-means++, to obtain the type with the largest proportion of types in the nodes as the classification result of the new furnace.

[0048] The Bayesian prediction and estimation module combines the current furnace classification data obtained in step 3, the furnace classification results of K-means++ clustering obtained in step 2, and the furnace feature statistics in the furnace feature database to perform Bayesian prediction and estimation, and obtain the final metal recovery rate.

[0049] Furthermore, the Bayesian prediction estimation module predicts the final metal recovery rate using the following method:

[0050] Suppose P(A) represents the probability of a certain metal recovery rate A occurring, P(B) represents the probability of a specific furnace type B occurring, and P(B|A) represents the probability of a specific furnace type B under a certain metal recovery rate A. Then, P(A|B) represents the probability of a certain metal recovery rate A under a specific furnace type B.

[0051] After the furnace feature data is classified by KNN, the furnace type can be determined. Under this specific furnace type, for all possible metal recovery rates within a selected reasonable range, the probability of each possible metal recovery rate is obtained by looking up the furnace table. The metal recovery rate that maximizes P(A|B) is then used as the prediction result.

[0052] Beneficial effects: This invention establishes a classification prediction model by combining a large amount of historical furnace data, and performs material balance calculations in real time for the complex and fluctuating material conditions of electric arc furnace hot charging molten iron. It effectively solves the practical problem of unstable steel output in the electric arc furnace hot charging molten iron process system, avoids the impact of abnormal steelmaking data on subsequent steelmaking results, improves the accuracy of steel yield prediction, shortens the electric arc furnace smelting cycle, reduces energy and material consumption, and achieves the ultimate energy-saving effect.

[0053] This invention is scalable. Its purpose is to address the problem of inaccurate metal recovery rate prediction in intelligent auxiliary decision-making systems for electric arc furnace (EAF) hot charging molten iron processes. Therefore, the extracted features are specific to EAF hot charging molten iron processes. If it is necessary to predict the metal recovery rate of other steelmaking methods, reliable features can be obtained again from actual production data, allowing the invention to be applied to other scenarios. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the data signal relationships between the modules in this invention;

[0055] Figure 2 This is a schematic diagram of some standardized features in a KD tree according to one embodiment of the present invention;

[0056] Figure 3 This is a flowchart of Bayesian estimation in one embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0058] This invention presents a metal recovery rate prediction method based on Bayesian theory. It performs feature clustering analysis on furnace characteristics, grouping furnaces with similar characteristics into one category and constructing a KD-tree index structure to quickly classify the current furnace data features. Then, using Bayesian theory, it estimates the metal recovery rate with the highest probability in its respective category. This invention avoids the operational costs and random errors associated with traditional empirical methods. Without excessively increasing system complexity and resource consumption, it enables the electric arc furnace hot charging hot iron process system to quickly and automatically predict metal recovery rates and obtain accurate furnace energy and material consumption figures. Compared to previous methods, it achieves better energy conservation and consumption reduction.

[0059] like Figure 1 As shown, the metal recovery rate prediction method based on Bayesian theory in this embodiment includes the following steps:

[0060] Step 1: Z-score feature standardization; First, obtain the mean μ and standard deviation σ of all furnace features in the furnace feature database and the furnace feature elements of each attribute column. Then, convert all furnace features into standard form through Z-score standardization to eliminate the influence of dimensions and obtain standardized furnace features.

[0061] Step 2: For the standardized furnace features obtained in Step 1, perform K-means++ clustering and construct a KD tree simultaneously;

[0062] When performing K-means++ clustering, all furnaces are first divided into D groups. In practical applications, the parameter D should be adjusted according to the final prediction results to achieve the best classification effect. The type number of the attribute to which each furnace feature belongs is obtained, and the furnace classification results are saved to the type attribute of the corresponding record in the furnace feature database. The statistical data required for Bayesian estimation can be easily obtained by using SQL queries in the database.

[0063] Construct a KD tree, where each node stores the standardized features of the furnace and the furnace classification result. The goal is to find the C nodes closest to the standardized feature data of the furnace with the least time complexity. The parameter C needs to be adjusted according to the specific prediction results in the application.

[0064] Step 3: Based on the KD tree index structure constructed in Step 2, after inputting new furnace features, first standardize the furnace data of this furnace, and based on the KD tree index structure constructed in Step 2, match the nodes closest to the new furnace features. Combine the furnace feature classification results obtained by K-means++, perform KNN classification on the new standardized furnace features, and obtain the type with the largest proportion of type quantity in the node as the classification data of the current furnace.

[0065] Step 4: Combining the current furnace classification data obtained in Step 3, the furnace classification results of K-means++ clustering obtained in Step 2, and the furnace feature statistics in the furnace feature database, perform Bayesian prediction and estimation to obtain the metal recovery rate.

[0066] Example:

[0067] This embodiment predicts the metal recovery rate in the hot charging process of an electric arc furnace, using C++ as the programming language and MySQL as the database platform. Specific attributes in the furnace characteristic database table include `stoveId`, `ironWaterScrapWeightPro`, `ironWaterTem`, `cRate`, `mnRate`, and `standardTemperature`. Specific attributes in the furnace table include: `stoveId`, `ironWaterWeight`, `ironWaterTem`, `ironScrapWeight`, `steelTappingWeight`, `steelTappingTem`, `powerCompression`, and `type`.

[0068] Table 1

[0069]

[0070]

[0071]

[0072] The data in the furnace feature database table is shown in Table 1. The data in the furnace feature database table is used to illustrate this embodiment. In application, data needs to be provided according to the actual situation of the current furnace.

[0073] The furnace feature data were standardized and then subjected to K-means++ clustering. The clustering results are shown in Table 2. Furnace features with the same type attribute belong to the same class. In this embodiment, the parameter D is set to 5.

[0074] Table 2

[0075]

[0076]

[0077] Since the furnace features are 6-dimensional data and cannot be directly displayed, this embodiment lists the KD tree structure of two dimensions of the furnace features (e.g., Figure 2 As shown in the figure, it is used to demonstrate how furnace features are combined with the KD tree.

[0078] like Figure 3 As shown, the method for estimating the metal recovery rate using Bayesian prediction in this embodiment is as follows:

[0079] Step 4.1: Obtain the number of all furnace features in the current database from the furnace feature database, denoted as n. Count the number of all records that belong to the same type B as the new furnace, denoted as countB. From this, the probability of this specific furnace type can be calculated, as shown in formula (2):

[0080]

[0081] Step 4.2: Obtain the number of records in the current database with a metal recovery rate of A as countA, and then calculate the probability of a specific metal recovery rate A, as shown in formula (3):

[0082]

[0083] Step 4.3: Obtain the number of records with metal recovery rate A and furnace type B, denoted as countAB. Calculate the probability of a specific furnace type B under a specific recovery rate A, as shown in formula (4):

[0084]

[0085] Step 4.4: Use the Bayesian method shown in formula (5) to predict the metal recovery rate P(A|B) with the highest probability in the current furnace category, and use it as the final predicted value of metal recovery rate;

[0086]

[0087] Specifically, first, the number of records in the current database is obtained using an SQL statement, denoted as n. Simultaneously, the number of records in the database corresponding to the new furnace type is obtained, denoted as countB. Let the metal recovery rate be 0.86, and the possible range of the metal recovery rate is determined to be 0.7 to 1.0. Therefore, starting from 0.7 and ending at 1.0, the loop iterates through all possible metal recovery rates with a step size of 0.02. The number of records in the current database with a metal recovery rate equal to the current metal recovery rate is obtained using an SQL statement, denoted as countA. If countA is 0, the probability of the current metal recovery rate is considered 0, and the current iteration is skipped. Finally, the number of records in the current database with a metal recovery rate equal to the current metal recovery rate and of the type of the new furnace is obtained using an SQL statement, denoted as countAB. Finally, the probability of the current metal recovery rate is obtained by using Bayes' formula P(A|B)=P(A)×P(B|A) / P(B). By comparing the values ​​of P(A|B) under different metal recovery rates, the metal recovery rate with the largest P(A|B) is selected as the prediction result. If P(A|B) is always 0, then the final estimated result is 0.86, which is a relatively reasonable constant for the metal recovery rate of the furnace.

[0088] As can be seen from the above embodiments, this invention utilizes Bayesian principles to analyze experimental data and automatically predict the metal recovery rate for each furnace smelting operation. This allows the electric arc furnace hot charging molten iron process system to calculate more accurate energy and material consumption, reduce resource waste caused by human error, and further promote the popularization and development of the electric arc furnace hot charging molten iron process.

Claims

1. A method for predicting metal recovery rate based on Bayesian theory, characterized in that, Includes the following steps: Step 1: Z-score feature standardization; first, obtain the average value of all furnace features in the furnace feature database and the furnace feature elements in each attribute column. and standard deviation Then, Z-score standardization is used to convert all furnace features into a standard form to obtain the standardized furnace features. Step 2: For the standardized furnace features obtained in Step 1, perform K-means++ clustering and construct a KD tree simultaneously; When performing K-means++ clustering, all furnace features are first divided into D groups, where D is the grouping parameter. The type number of the attribute to which each furnace feature belongs is obtained, and the furnace classification result represented by the type number is saved to the type attribute of the corresponding record in the furnace feature database. Construct a KD tree where each node stores the standardized features of the furnace and the furnace classification result. The goal is to find the C nodes closest to the standardized feature data of the furnace with the least time complexity. Step 3: Based on the KD-tree index structure constructed in Step 2, after inputting new furnace features, first standardize the furnace data, and then, based on the KD-tree index structure constructed in Step 2, match the nodes closest to the new furnace features. Combine the furnace feature classification results obtained from K-means++, perform KNN classification on the new standardized furnace features, and obtain the type with the largest proportion of types in the nodes as the classification data for the current furnace. The specific method is as follows: First, the features of the newly input furnace are determined, and then Z-score normalization is performed on these features. Next, the K nearest sample points are searched based on the index structure of the resulting KD-tree. The specific search method is as follows: 1) Create a max-heap to store information about the K nearest nodes; 2) Find the leaf node containing the target point x in the KD tree index structure: Starting from the root node, recursively search the KD tree downwards. If the coordinates of the target point x in the current dimension are less than the coordinates of the split point, move to the left child node; otherwise, move to the right child node, until the leaf node is reached. Treat this leaf node as the "current nearest node" and add it to the max-heap; 3) The method for adding to the max-heap is to recursively backtrack upwards, performing the following operations on each node: If the instance point stored in the node is closer to the target point than the current top node of the max-heap, then if the number of elements in the max-heap is less than K, add it directly; if it is greater than K, delete the first element and add the new element to the max-heap; update the distance of the "current nearest point" to the distance between the first element of the max-heap and the target point, and at the same time check if there is a closer point in the region corresponding to the other child node of the parent node of the child node. The specific steps are as follows: Check if the region corresponding to the other child node intersects with a circle or hypersphere centered at the target point and with the distance between the target point and the "current nearest point" as its radius: if they intersect, there may be a point closer to the target point in the region corresponding to the other child node, move to the other child node, and then continue to recursively perform the nearest neighbor search; if they do not intersect, backtrack upwards. 4) When the search ends at the root node, the nodes stored in the max-heap are the K nodes closest to the target node. Step 4: Combine the current furnace classification data obtained in Step 3, the furnace classification results of K-means++ clustering obtained in Step 2, and the furnace feature statistics in the furnace feature database to perform Bayesian prediction estimation.

2. The metal recovery rate prediction method based on Bayesian theory according to claim 1, characterized in that, Step 4, the specific process of Bayesian prediction estimation, is as follows: Step 4.1: Obtain the number of all furnace features in the current database from the furnace feature database, denoted as n. Count the number of all records that belong to the same type B as the new furnace, denoted as countB. From this, the probability of the new furnace belonging to type B can be calculated, as shown in formula (2): (2) Step 4.2: Obtain the number of records in the current database with a metal recovery rate of A as countA, and then calculate the probability of a specific metal recovery rate A, as shown in formula (3): (3) Step 4.3: Obtain the number of records with metal recovery rate A and furnace type B, denoted as countAB. Calculate the probability of a specific furnace type B under a specific recovery rate A, as shown in formula (4): (4) Step 4.4: Use the Bayesian method shown in formula (5) to predict the metal recovery rate P(A|B) with the highest probability in the current furnace category, and use it as the final predicted value of metal recovery rate; (5)。 3. The metal recovery rate prediction method based on Bayesian theory according to claim 1, characterized in that, The formula for calculating the Z-score feature standardization in step 1 is as follows: (1) Formula (1) converts the values ​​of all furnace features to between -1 and 1. The attributes in the furnace feature database include stoveId, ironWaterScrapWeightPro, ironWaterTem, cRate, mnRate, and standardTemperature.

4. The method for predicting metal recovery rate based on Bayesian theory according to claim 1, characterized in that, The specific method for automatically clustering furnace features by performing K-means++ clustering on the standardized furnace features in step 2 is as follows: First, find K cluster centers from all the standardized features of the smelting furnace, and randomly select a sample point from the standardized features of the smelting furnace as the first initial cluster. Next, calculate the shortest distance between each sample and the existing cluster centers; Then, the probability of each sample point being selected as the next cluster center is calculated. The probability of a sample point is calculated as: the shortest distance of the current sample point from the cluster center / the sum of the shortest distances of all sample points from the cluster center. Finally, select the sample point corresponding to the highest probability value as the next cluster center, and repeat the above method of finding centers until K cluster centers are selected. Number these K cluster centers with different numbers, and the same number indicates that the sample points belong to the same class. Continue to cluster all samples into K classes, and assign the sample set to the nearest cluster center according to the minimum distance principle, and modify the sample number. Update the cluster center using the sample mean of each cluster. Repeat the above clustering process until the cluster centers no longer change. K-means++ clustering can ultimately yield the type number of each furnace feature; and the classification result is then updated to the type attribute of the corresponding record in the database. In K-means++, the distance between sample points is Euclidean distance, and the distance formula is shown in formula (6): (6) in, This represents the i-th dimension of the feature of sample point X. Let Y represent the i-th dimension of the feature of sample point Y.

5. A system for implementing the Bayesian theory-based metal recovery rate prediction method according to any one of claims 1 to 4, characterized in that, It includes a Z-score feature normalization module, a K-means++ clustering module, a KD-tree construction module, a KNN classification module, and a Bayesian prediction module. Statistical analysis was performed on all data attributes in the furnace characteristic database, and the mean and variance of each attribute column were obtained as follows: and The Z-score feature standardization module performs Z-score standardization on the corresponding attribute values ​​of each record based on the mean and variance of each attribute column, as shown in formula (1): (1) All furnace standardization features have values ​​between -1 and 1. The specific attributes in the furnace feature database table include stallId, ironWaterScrapWeightPro, ironWaterTem, cRate, mnRate, and standardTemperature. The K-means++ clustering module performs cluster analysis on the standardized features of the furnace. Feature samples with high similarity are clustered into the same cluster to obtain furnace features of different clusters, i.e., to generate classification results of different furnace features, and save them to the type attribute column of the furnace table stove in the furnace feature database. The specific attributes of the furnace table include: stoveId, ironWaterWeight, ironWaterTem, ironScrapWeight, steelTappingWeight, steelTappingTem, powerCompression, and type. The KD tree construction module constructs the standardized features of the furnace into a KD tree, which is used as the index structure when storing training data samples. The KNN classification module is based on the index structure of the KD tree. When the user inputs a new furnace feature, the KD tree quickly matches the few nodes closest to the new furnace feature. Then, the majority voting method is used, combined with the furnace feature classification results obtained by K-means++, to obtain the type with the largest proportion of types in the nodes as the classification result of the new furnace. The Bayesian prediction and estimation module combines the current furnace classification data obtained in step 3, the furnace classification results of K-means++ clustering obtained in step 2, and the furnace feature statistics in the furnace feature database to perform Bayesian prediction and estimation, and obtain the final metal recovery rate.

6. The system for predicting metal recovery rate based on Bayesian theory according to claim 5, characterized in that, The method used by the Bayesian prediction estimation module to predict the final metal recovery rate is as follows: Suppose P(A) represents the probability of a certain metal recovery rate A occurring, P(B) represents the probability of a specific furnace type B occurring, and P(B|A) represents the probability of a specific furnace type B under a certain metal recovery rate A. Then, P(A|B) represents the probability of a certain metal recovery rate A under a specific furnace type B. After the furnace feature data is classified by KNN, the furnace type can be determined. Under this specific furnace type, for all possible metal recovery rates within a selected reasonable range, the probability of each possible metal recovery rate is obtained by looking up the furnace table. The metal recovery rate that maximizes P(A|B) is then used as the prediction result.

Images