Converter steelmaking endpoint carbon temperature prediction soft measurement method based on adaptive on-the-fly learning

Abstract

The application discloses a soft measurement method for converter steelmaking endpoint carbon temperature prediction based on adaptive instant learning, adopts a WGK measurement criterion with introduced label information to construct a WGKSOM clustering algorithm for guiding a clustering direction, improves clustering quality of the algorithm and reduces influence of data fluctuation of a class cluster on modeling; secondly, a membership of a to-be-measured sample is calculated by using a Gaussian posterior probability and a DRCA measurement strategy is constructed by introducing a dynamic factor, so that adaptive selection of the sample to form an algorithm learning set is realized and is used for local model training, and finally online prediction of the converter steelmaking endpoint carbon temperature is completed. Compared with other instant learning soft measurement methods, the application is more accurate in prediction of the converter steelmaking endpoint carbon temperature.

Description

Technical Field

[0001] This invention belongs to the field of complex industrial process modeling and soft measurement technology, specifically involving a soft measurement method for predicting the carbon temperature at the end point of converter steelmaking based on adaptive real-time learning. Background Technology

[0002] As a vital pillar of the national economy, the steel industry plays a crucial role in national industrial and socio-economic development. Converter steelmaking, due to its high efficiency and low cost, is one of the world's main steelmaking methods. Currently, converter steelmaking methods include traditional manual experience methods and secondary lance detection methods. With technological advancements, methods such as flame image analysis and spectral image analysis have also emerged. However, these methods suffer from drawbacks, including significant susceptibility to human error, high equipment usage and maintenance costs, difficulty in extracting key information from flame images, and susceptibility of spectral equipment to severe interference from high-temperature objects such as ladles. These problems pose significant challenges to converter steelmaking production and limit the widespread application of these technologies.

[0003] With the development of computer technology and the improvement of industrial intelligent detection levels in recent years, it has become possible to predict the final carbon temperature using soft sensing methods that establish mathematical models of auxiliary variables and the final carbon temperature during converter steelmaking. However, in actual industrial production, factors such as the varying quality of raw materials used in steelmaking, deviations in data acquisition equipment, and errors in manual recording often lead to significant fluctuations in the collected converter steelmaking process data, posing a significant challenge to soft sensing modeling. To address these issues, existing technologies include real-time learning soft sensing methods. These methods cluster historical data and then select samples from the cluster to which the sample to be tested belongs to form a local algorithm learning set, thereby training a local model to predict the final carbon temperature. These methods, to some extent, overcome the impact of data volatility on soft sensing models and improve the model's performance in predicting the final carbon temperature.

[0004] However, existing instant learning soft sensor modeling methods neglect the important value of label information, lack fault-tolerant mechanisms to avoid classification errors, and have problems in selecting samples from a single cluster. As a result, current instant learning soft sensor models are unable to achieve the expected results in predicting the final carbon temperature of converter steelmaking.

[0005] In summary, the current data on converter steelmaking processes exhibits high volatility and nonlinearity, resulting in low-quality learning sets for algorithms that perform real-time learning metrics. This weakens the model's predictive performance and leads to inaccurate predictions of the final carbon temperature in converter steelmaking.

[0006] Therefore, a prediction method is needed to improve the prediction performance of real-time learning soft measurement models in terms of the quality of local algorithm learning sets and the prediction performance of endpoint carbon temperature. Summary of the Invention

[0007] To address the aforementioned technical issues, this invention provides a soft measurement method for predicting the final carbon temperature in converter steelmaking based on adaptive real-time learning. It employs the WGK metric criterion with introduced label information to construct the WGKSOM clustering algorithm, which guides the clustering direction, improves the clustering quality, and reduces the impact of cluster data volatility on modeling. Secondly, it uses Gaussian posterior probability to calculate the membership degree of the samples to be tested and introduces a dynamic factor to construct a DRCA metric strategy, thereby enabling adaptive selection of samples to form the algorithm's learning set for local model training. Finally, it completes the online prediction of the final carbon temperature in converter steelmaking.

[0008] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: a soft measurement method for predicting the final carbon temperature in converter steelmaking based on adaptive real-time learning, characterized by comprising the following steps:

[0009] S1: Collect production process data during the actual production process of the steel plant;

[0010] S2: After preprocessing the above data and removing outliers, normalize all data so that the features of each dimension are at the same level of measurement.

[0011] S3: In the offline training part, the WGKSOM algorithm is used to cluster the historical process dataset to obtain a cluster dataset that can represent different modalities;

[0012] S3.1: Randomly initialize the weight vector w of each neuron node in the competition layer. k (k = 1, 2, ..., n), where n is the number of neurons in the competition layer, and each neuron has a weight w with the same h dimensions as the input data. kh =[w k1 ,w k2 ,…,w kd ];

[0013] S3.2: In the t-th iteration, a training sample x(t) is randomly selected from the training set as input. The similarity between x(t) and the weight vectors of all competing layer neurons is calculated. The neuron node most similar to x(t) is the winning node, denoted as the Best Match Neuron (BMU) of the current training sample. The calculation formula is as follows:

[0014]

[0015] Where, x i (t)∈X is the input training sample vector, n is the number of neuron nodes, and the calculation of the WGK metric is shown in the formula;

[0016] WGK(x i ,w k )=exp[-L(x i ,w k ) / 2δ 2 ]

[0017]

[0018] Where, x ih Represents the input vector x i The h-th dimension feature, w kh w represents the weight of the h-th dimension of the weight vector of the k-th competing layer neuron. gh Let WGK(x) be the weight of the h-th dimension of the global weighting coefficients. i w k The larger the value, the better the input sample matches the neurons in the competing layer;

[0019] When using the WGK metric, the first step is to determine the global weighting coefficient w in the WGK metric. gh This weighting factor enhances the features in the converter process data that are more correlated with the output carbon temperature results;

[0020] S3.3: The winning neuron node and its neighboring node set are denoted as N. BMU (by the domain function h) iBMU (Judgment), and then, using the training samples as a reference, for N BMU The weights of neurons within the range are updated using the following formula:

[0021]

[0022] Where t represents the current iteration number, L(t) is the current learning rate, and h iBMU Let (t) be the neighborhood function, and let the learning rate L(t) and the neighborhood function h be... iBMU Both L(t) and h(t) decrease monotonically with time. iBMU The calculation of (t) is shown in the following formula:

[0023]

[0024] Where L0∈[0,1] is the initial learning rate, and T is the total number of training iterations of the WGKSOM algorithm.

[0025] h iBMU (t) is the neighborhood function around the winning neuron BMU at time t, representing the distance from neuron i to the winning neuron BMU, as shown in the following formula:

[0026]

[0027]

[0028] Among them, Pos i Pos represents the coordinates of neighboring neuron i. BMU Let r(t) represent the coordinates of the winning neuron, r(t) be the neighborhood radius, r(t) decreases monotonically during training, and r0 be the initial neighborhood radius.

[0029] S3.4: When the termination condition is met, the clustering ends and the clustering results are obtained. Cluster centers of various types of clusters Where C represents the number of clusters, X k This represents the cluster sample set of the k-th cluster;

[0030] S4: In the online prediction part, for a new query sample, the DRCA algorithm is used to adaptively select samples from multiple clusters to form a local algorithm learning set;

[0031] Based on the size of the local algorithm learning set, determine the size of the sample subset measured from each type of cluster;

[0032] N k (x q ,X k )=N×P k ,k∈(1,C)

[0033] Where N is the sample x to be tested q The corresponding local algorithm learning set size, P k Represents x q The Gaussian posterior probability value of the k-th class of sample sets;

[0034] Based on the correlation component analysis algorithm, the local algorithm learning set is formed by measuring the sample with the smallest value among various clusters and the test sample.

[0035]

[0036]

[0037] Where, x q For the sample to be tested, x i For cluster samples, For the matrix obtained by training under each cluster, m j Let N be the mean of the samples in the j-th category subset. k This represents the number of samples within that cluster.

[0038] S5: Input the local algorithm learning set obtained in step 4 into the limit learning set for local model training, and finally complete the prediction of the endpoint carbon content / temperature of the current test sample.

[0039] The beneficial effects of this invention are:

[0040] This invention obtains a local algorithm learning set that better matches the test samples through adaptive sample selection, and compared with other instant learning soft measurement methods, it is more accurate in predicting the final carbon temperature of converter steelmaking. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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.

[0042] Figure 1 This is a schematic diagram illustrating the implementation process of the method of the present invention;

[0043] Figure 2 This is a schematic diagram of the algorithm structure of the method of the present invention;

[0044] Figure 3 This is a comparison of the results of the present invention and existing technologies in predicting the final temperature of converter steelmaking;

[0045] Figure 4 This is a comparison of the results of the present invention and existing technologies in predicting the carbon content at the end point of converter steelmaking. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0047] Example

[0048] Step 1: Collect production process data recorded by sensors during the actual production process in the steel plant;

[0049] Step 2: Preprocess the process data. First, remove outlier data, and then normalize all data to ensure that the features of each dimension are at the same level of measurement.

[0050] Step 3: Use the weighted Gaussian kernel self-organizing map algorithm to cluster the historical process dataset, so that the data within the cluster are close and the data between the clusters are far apart, to obtain a cluster dataset that can represent different modalities;

[0051] Step 3.1: Randomly initialize the weight vector w of each neuron node in the competition layer. k (k = 1, 2, ..., n), where n is the number of neurons in the competition layer, and each neuron has a weight w with the same h dimensions as the input data. kh =[w k1 ,w k2 ,…,w kd ];

[0052] Step 3.2: In the t-th iteration, randomly select a training sample x(t) from the training set as input, calculate its similarity to the weight vectors of all competing layer neurons, and the neuron node most similar to it is the winning node, denoted as the Best Match Neuron (BMU) of the current training sample. The calculation formula is as follows:

[0053]

[0054] Where, x i (t)∈X is the input training sample vector, n is the number of neuron nodes, and the calculation of the WGK metric is shown in equation (2).

[0055] WGK(x i ,w k )=exp[-L(x i ,w k ) / 2δ 2 (2)

[0056]

[0057] Where, x ih Represents the input vector x i The h-th dimension feature, w kh w represents the weight of the h-th dimension of the weight vector of the k-th competing layer neuron. gh Let WGK(x) be the weight of the h-th dimension of the global weighting coefficients. i w k The larger the value, the better the input sample matches the neurons in the competing layer.

[0058] When using the WGK metric, the first step is to determine the global weighting coefficient w in the WGK metric. gh This weighting factor enhances the features in the converter process data that are more correlated with the output carbon temperature, and is defined as follows:

[0059] Definition 1: Pearson Correlation Coefficient Weighting Criterion

[0060] 1. Calculate the Pearson correlation coefficient (PCCs) between the input features of each dimension in the historical database and the output results. Then, the h-th dimension input feature x h The calculation of the output y is shown in equation (4):

[0061]

[0062] Wherein, cov(x) h (,y) represents the h-th dimension feature x of the input sample set X. h The covariance of the output y, where m represents the feature dimension of the input sample set, and σ * This indicates the calculation of the standard deviation of the variable.

[0063] 2. Based on equation (4), calculate the correlation between each input dimension and the endpoint carbon temperature, and then further calculate to obtain the weighted weight w. g ={w g1 w g2 w gm}

[0064]

[0065] Step 3.3: Let N be the set of the winning neuron node and its neighboring nodes. BMU (by the domain function h) iBMU (Judgment), and then, using the training samples as a reference, for N BMU The weights of neurons within the range are updated using the following formula:

[0066]

[0067] Where t represents the current iteration number, L(t) is the current learning rate, and h iBMU Let (t) be the neighborhood function, and let the learning rate L(t) and the neighborhood function h be... iBMU Both L(t) and h(t) decrease monotonically with time. iBMU The calculation of (t) is shown in formulas (7) and (8):

[0068]

[0069] Where L0∈[0,1] is the initial learning rate, and T is the total number of training iterations of the WGKSOM algorithm.

[0070] h iBMU (t) is the neighborhood function around the winning neuron BMU at time t, representing the distance from neuron i to the winning neuron BMU, as shown in the following formula:

[0071]

[0072]

[0073] Among them, Pos i Pos represents the coordinates of neighboring neuron i. BMU Let r(t) represent the coordinates of the winning neuron, r(t) be the neighborhood radius, r(t) decreases monotonically during training, and r0 be the initial neighborhood radius.

[0074] Step 3.4: When the termination condition is met, the clustering ends, and the clustering results are obtained. Cluster centers of various types of clusters Where C represents the number of clusters, X k Let represent the cluster sample set of the k-th cluster.

[0075] Step 4: For the sample data to be tested, based on the proposed dynamic correlation component analysis method, samples are adaptively selected from multiple clusters to form a local algorithm learning set;

[0076] Step 4.1: Calculate the Gaussian posterior probability of the test sample belonging to the same cluster, using the following formula:

[0077]

[0078] Among them, P k Represents the sample to be tested, x q The Gaussian posterior probability of the k-th cluster, where h represents the feature dimension of the input sample, u k Let ∑ represent the mean of the k-th cluster. k Let represent the covariance matrix of the k-th cluster.

[0079] Step 4.2: Determine the size of the adaptive local algorithm learning set corresponding to the current test sample, which is calculated as follows:

[0080] N(x q ,X)=Dynum(|Pcc(x q ,x i )|≥e),x i ∈X (11)

[0081] Where, x i ∈X represents a sample in the historical database, and e is a dynamic factor used to adaptively adjust the size of the local algorithm learning set. The larger e is, the smaller the local algorithm learning set is, and vice versa. Pcc(x) q x i ) indicates the calculation of x q and x iThe Pearson correlation coefficient is used to calculate the sample size of the test sample and the historical database sample results that satisfy the dynamic factor condition, which is then used as the size of the learning set for the adaptive algorithm of the test sample.

[0082] Step 4.3: Based on the size of the local algorithm learning set in Step 4.2, determine the size of the sample subset measured from each type of cluster;

[0083] N k (x q ,X k )=N×P k ,k∈(1,C) (12)

[0084] Where N is the sample x to be tested q The corresponding local algorithm learning set size (calculated in step 4.2), P k Represents x q The Gaussian posterior probability value of the k-th class of sample sets.

[0085] Step 4.4: Based on the correlation component analysis algorithm, measure the sample with the smallest value from each cluster to form a local algorithm learning set;

[0086]

[0087]

[0088] Where, x q For the sample to be tested, x i For cluster samples, For the matrix obtained by training under each cluster, m j Let N be the mean of the samples in the j-th category subset. k This represents the number of samples within this cluster.

[0089] Step 5: Input the local algorithm learning set obtained in Step 4 into the extreme learning set to train the local model, and finally complete the prediction of the endpoint carbon content / temperature of the current test sample.

[0090] In summary, this patent obtains a local algorithm learning set that better matches the sample to be tested through adaptive sample selection, and compared with other instant learning soft measurement methods, it is more accurate in predicting the final carbon temperature of converter steelmaking.

[0091] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0092] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A soft measurement method for predicting the final carbon temperature in converter steelmaking based on adaptive real-time learning, characterized in that, Specifically, the steps include the following: S1: Collect production process data recorded by sensors during the actual production process in the steel plant; S2: Preprocessing the above data: After removing outlier data, normalize all data; S3: Offline Training: The dataset processed in S2 is clustered using the WGKSOM algorithm to obtain cluster datasets representing different modalities; the WGKSOM algorithm is constructed using the WGK metric that incorporates label information; S3 specifically includes the following steps: S3.1: Randomly initialize the weight vector of each neuron node in the competition layer. ,in This represents the number of neurons in the competitive layer. S3.2: In the first In each iteration, a training sample is randomly selected from the training set. As input, the similarity between the node and the weight vectors of all competing layer neurons is calculated. The neuron node with the highest similarity is the winning node and is denoted as the best matching unit in the current training sample. The calculation formula is as follows: The formula for calculating the measurement criteria is as follows: in, Represents the input vector No. Dimensional features, Indicates the first The weight vector of the nth competitive layer neuron node The weights of the dimensions, The global weighting coefficient is the first The weights of the dimensions, The larger the value, the better the input sample matches the neurons in the competing layer; Sure Calculate the Pearson correlation coefficient between the input features of the historical samples and the output results, and obtain the first... 3D input features With output The calculation formula is used to calculate the correlation between each input dimension and the endpoint carbon temperature, resulting in the following calculation. ; S3.3: The winning neuron node and its neighboring node set are denoted as... Using the training samples as a reference, for Update the weights of neurons within the specified range; S3.4: When the termination condition is met, the clustering ends, and the clustering results and the cluster centers of each type of cluster are obtained; S4: Online Prediction Part: For new query samples, a dynamic correlation component analysis algorithm is used to adaptively select samples from multiple clusters to form a local algorithm learning set. This includes the following steps: S4.1: Calculate the Gaussian posterior probability of the test sample belonging to the same cluster class; S4.2: Determine the size of the adaptive local algorithm learning set corresponding to the current test sample, which is calculated as follows: in, Representing samples in the historical database, This is a dynamic factor used to adaptively adjust the size of the local algorithm learning set. The larger the value, the smaller the local learning set of the algorithm; conversely, the smaller the value, the larger the local learning set of the algorithm. Indicates calculation and Pearson correlation coefficient, Function calculation The sample size is used as the size of the learning set for the adaptive algorithm of the test sample; S4.3: Based on the Gaussian posterior probabilities of the test sample and each type of cluster in S4.1 and the size of the adaptive local algorithm learning set in S4.2, determine the size of the sample subset measured from each type of cluster. S4.4: Based on the correlation component analysis algorithm, the sample with the smallest distance to the test sample from various clusters is used to form a local algorithm learning set; S5: Input the local algorithm learning set from S4 into the extreme learning machine to train the local model; S6. Use the local model trained in S5 to predict the endpoint carbon temperature of the current test sample.

Images