Construction method and system of iron-making blast furnace target system based on deep clustering algorithm

Abstract

The application provides a kind of based on deep clustering algorithm's iron-making blast furnace to benchmark system construction method and system, comprising: step 1: based on production management system data, by statistical model establishes each blast furnace individual portrait, then carries out abnormal data removal and normalization processing, is abstracted as data point in high-dimensional space;Step 2: based on deep clustering algorithm, the data point in high-dimensional space is clustered, and the clustering result is obtained;Step 3: the clustering result is visualized;Step 4: the clustering result is monitored, and when the performance index is lower than the preset threshold, self-training is carried out, and the corresponding performance is improved.The present application solves the problem that the index weight is difficult to allocate due to the large number of indexes and the scattered attribute distribution in the blast furnace benchmarking process by combining supervised learning and unsupervised learning, and provides a quantitative standard for blast furnace benchmarking and evaluation.

Description

Technical Field

[0001] This invention relates to the field of benchmarking system construction technology for blast furnaces, specifically to a method and system for constructing a benchmarking system for blast furnaces based on a deep clustering algorithm. Background Technology

[0002] Currently, there are over 300 blast furnace steel plants nationwide, with more than 1,000 blast furnaces. Due to the different technological approaches employed by various blast furnaces, there is a lack of unified production evaluation standards. In big data production practices, current solutions simply aggregate indicators such as blast furnace output and energy consumption. However, these quantitative indicators cannot be standardized due to differences in products and production lines across steel mills, resulting in an existing evaluation system lacking scientific rigor and persuasiveness.

[0003] Patent document CN110069815A (application number: CN201910193841.1) discloses a method, system, and terminal device for constructing an indicator system, including: acquiring simulation data obtained by simulating a weapon system and preprocessing the simulation data; constructing a weighted hypernet model of the weapon system based on the preprocessed simulation data; and constructing an indicator system for evaluating the effectiveness of the weapon system based on the weighted hypernet model and a preset effectiveness evaluation target of the weapon system.

[0004] The existing technology has the following drawbacks: First, there is a lack of a large number of qualitative indicators for blast furnace evaluation; second, there is a lack of standardized or normalized quantitative means for indicators between different steel mills and blast furnaces; and third, in the face of the high-dimensional problems brought about by a large number of indicators, there is a lack of reasonable classification or clustering algorithms to construct blast furnace profiles and evaluate blast furnace technology.

[0005] Patent document CN109800215A (application number: CN201811598897.7) discloses a method, apparatus, computer storage medium, and terminal for benchmarking processing, including: constructing a first feature vector for each target table to be benchmarked; performing clustering operations on the constructed first feature vectors; merging target tables belonging to the same cluster based on the clustering results to obtain wide tables corresponding to each cluster; and selecting one of the obtained wide tables as the standard table for benchmarking processing based on the similarity information between the original table and each obtained wide table. This patent provides a clustering-based benchmarking system, but firstly, its processing object is too broad, failing to perform data preprocessing and necessary optimization operations for the specific scenario of blast furnace ironmaking; secondly, the clustering algorithm it uses is relatively simple, while blast furnace indicators are numerous and their attributes are widely distributed. Simple clustering algorithms only process them uniformly and cannot reasonably assign different weights to different indicators, leading to clustering results that seriously deviate from the business meaning. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for constructing a benchmarking system for ironmaking blast furnaces based on a deep clustering algorithm.

[0007] The method for constructing a benchmarking system for ironmaking blast furnaces based on a deep clustering algorithm provided by the present invention includes:

[0008] Step 1: Based on the production management system data, establish individual profiles for each blast furnace through statistical models, including operational indicators, emission indicators, molten iron cost indicators, and energy consumption indicators. Then, perform outlier removal and normalization on the above indicators, and abstract them into data points in a high-dimensional space.

[0009] Step 2: Based on the deep clustering algorithm, cluster the data points in the high-dimensional space to obtain the clustering results;

[0010] Step 3: Visualize the clustering results;

[0011] Step 4: Monitor the clustering results and perform self-training when the performance index is below the preset threshold to improve the corresponding performance.

[0012] Preferably, the abnormal data removal includes: removing data exceeding a preset range by modeling the time series of production data of a single blast furnace; removing data exceeding a preset range by modeling the data of all blast furnaces in the industry; the upper limit of the preset range is 75th quartile + 1.3 * interquartile range, and the lower limit of the preset range is 25th quartile - 1.3 * interquartile range.

[0013] The normalization process includes: normalizing each field X, with the expression being:

[0014] Preferably, step 2 includes:

[0015] Step 2.1: Based on the neural network self-decoder model, it is divided into encoder and decoder. The encoder transforms the high-dimensional blast furnace image data into low-dimensional expression space data through the deep neural network. The decoder inversely transforms the low-dimensional expression space data into high-dimensional blast furnace image data. If the error between the fitted data after encoder and decoder and the original data is within the preset range, the expression space learning is completed.

[0016] Step 2.2: After the high-dimensional blast furnace profile data is processed by the encoder, clustering is performed based on the GMM model. For a single blast furnace data point, its final cluster affiliation is determined according to the probability distribution of belonging to different clusters.

[0017] Step 2.3: Based on the loss functions of supervised and unsupervised learning models, establish a unified loss function and optimization model, and find the optimal solution for parameters in the representation space learning and clustering models;

[0018] Step 2.4: Output the results, including the original image of each blast furnace, low-dimensional representation space data, and the cluster to which the cluster belongs.

[0019] Preferably, X represents the source data, and Z represents the low-dimensional space of the source data X, which is obtained through an encoder, i.e., Z = encode(X); This represents the reconstruction of source data X through an encoding-decoding process, i.e. For clustering in low-dimensional representation spaces, we use {u1, u2, ..., u...} j ,… represents the central kernel of the clustering algorithm, using As sample X i The expression space of Z i With the central core U j The similarity function, where i and j are sequence numbers;

[0020] The expression for the unified loss function is then:

[0021] Based on the loss function, the optimal solution for each parameter in the model is calculated using the Adam gradient descent algorithm.

[0022] Preferably, the visualization of clustering results includes: a list of indicator data, a comparative analysis chart of single indicators for multiple objects, a comparative heatmap of multiple indicators for multiple objects, and a big data analysis clustering chart of multiple indicators for multiple objects.

[0023] The blast furnace benchmarking system construction system based on deep clustering algorithm provided by the present invention includes:

[0024] Module M1: Based on production management system data, establish individual profiles for each blast furnace through statistical models, including operational indicators, emission indicators, molten iron cost indicators, and energy consumption indicators. Then, perform abnormal data removal and normalization on the above indicators, and abstract them into data points in a high-dimensional space.

[0025] Module M2: Based on deep clustering algorithm, it clusters data points in high-dimensional space to obtain clustering results;

[0026] Module M3: Visualizes the clustering results;

[0027] Module M4: Monitors clustering results and performs self-training when performance metrics fall below a preset threshold to improve performance.

[0028] Preferably, the abnormal data removal includes: removing data exceeding a preset range by modeling the time series of production data of a single blast furnace; removing data exceeding a preset range by modeling the data of all blast furnaces in the industry; the upper limit of the preset range is 75th quartile + 1.3 * interquartile range, and the lower limit of the preset range is 25th quartile - 1.3 * interquartile range.

[0029] The normalization process includes: normalizing each field X, with the expression being:

[0030] Preferably, the module M2 includes:

[0031] Module M2.1: Based on a neural network self-decoder model, it consists of an encoder and a decoder. The encoder transforms high-dimensional blast furnace image data into low-dimensional expression space data through a deep neural network. The decoder then inversely transforms the low-dimensional expression space data into high-dimensional blast furnace image data. If the error between the fitted data after encoder and decoder and the original data is within a preset range, then the expression space learning is complete.

[0032] Module M2.2: After the high-dimensional blast furnace profile data is processed by the encoder, it is clustered based on the GMM model. For a single blast furnace data point, its final cluster affiliation is determined according to the probability distribution of belonging to different clusters.

[0033] Module M2.3: Based on the loss functions of supervised and unsupervised learning models, establish a unified loss function and optimization model, and find the optimal solution of parameters in expression space learning and clustering models;

[0034] Module M2.4: Output results, including the original image of each blast furnace, low-dimensional representation space data, and the cluster to which the cluster belongs.

[0035] Preferably, X represents the source data, and Z represents the low-dimensional space of the source data X, which is obtained through an encoder, i.e., Z = encode(X); This represents the reconstruction of source data X through an encoding-decoding process, i.e. For clustering in low-dimensional representation spaces, we use {u1, u2, ..., u...} j ,… represents the central kernel of the clustering algorithm, using As sample X i The expression space of Z i With the central core U j The similarity function, where i and j are sequence numbers;

[0036] The expression for the unified loss function is then:

[0037] Based on the loss function, the optimal solution for each parameter in the model is calculated using the Adam gradient descent algorithm.

[0038] Preferably, the visualization of clustering results includes: a list of indicator data, a comparative analysis chart of single indicators for multiple objects, a comparative heatmap of multiple indicators for multiple objects, and a big data analysis clustering chart of multiple indicators for multiple objects.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) This invention solves the problem of difficulty in allocating indicator weights due to the large number of indicators and the dispersed distribution of attributes in the process of blast furnace benchmarking by combining supervised learning and unsupervised learning, and provides a quantitative standard for blast furnace benchmarking and evaluation.

[0041] (2) This invention uses deep learning algorithms and models to reduce human intervention and is suitable for scenarios that require rapid business deployment and iteration, cold start, etc. For example, when the business needs to increase or decrease the consideration of multiple production and economic benefit indicators, it can train itself to obtain the most reasonable clustering results without relying on the experience of a large number of front-line personnel to judge the weight of indicators. Attached Figure Description

[0042] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0043] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0044] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0045] Example:

[0046] This invention establishes a method for benchmarking and identifying differences among multiple blast furnaces and production sites. By introducing a deep learning clustering algorithm, based on indicators such as production and economic benefits, blast furnaces under different environments are clustered. This not only reveals the commonalities in furnace conditions and ironmaking capacity among blast furnace groups, individual furnaces, and individual furnaces within different groups, but also provides a means to benchmark and identify differences, evaluating high-quality blast furnaces among numerous blast furnaces and discovering the gaps between high-quality blast furnaces and other blast furnaces. This invention combines supervised and unsupervised learning, first abstracting the data of each blast furnace into points in a high-dimensional space, and then mapping it to a low-dimensional manifold for clustering. This solves the problem of difficulty in quantifying and allocating indicator weights due to the large number of indicators and the dispersed distribution of attributes in the blast furnace benchmarking process.

[0047] like Figure 1 Specifically, it includes the following steps:

[0048] Step 1: Based on production management system data, establish individual profiles for each blast furnace using statistical models. This includes operational indicators (furnace volume utilization coefficient, hearth cross-section utilization coefficient, gas utilization rate, shutdown rate, fuel ratio, coke ratio, coal ratio, oxygen enrichment rate, etc.), emission indicators (SO2 in hot blast stove flue gas, NOx in hot blast stove flue gas, SO2 in furnace top gas, dustfall, etc.), iron cost indicators (total cost, variable cost, fixed cost, raw material cost, fuel cost, energy and medium cost, depreciation per ton of iron, etc.), and energy consumption indicators (energy consumption in ironmaking processes, energy consumption in blast furnace processes, etc.). After outlier removal and normalization preprocessing, these indicators are abstracted into data points in a high-dimensional space.

[0049] Outlier Removal: Two algorithms are used for outlier removal. First, by modeling the time series data of a single blast furnace, data points greater than (75th quartile + 1.3 * interquartile range) or less than (25th quartile - 1.3 * interquartile range) are removed. Second, by modeling the data of all blast furnaces in the industry, data points greater than (75th quartile + 1.3 * interquartile range) or less than (25th quartile - 1.3 * interquartile range) are removed.

[0050] Normalization: Normalize each field (represented by X) using the following expression:

[0051]

[0052] Step 2: For the high-dimensional spatial data points mentioned above, clustering is performed based on the DeepCluster algorithm. This DeepCluster algorithm includes three parts: an expression space learning module, a low-dimensional spatial clustering module, and an optimal solution algorithm module. After this step, the system will output three results: the original image of each blast furnace, the low-dimensional expression space data, and the cluster to which it belongs.

[0053] Step 3: Visualization, including the following four methods: list display of indicator data; comparative analysis chart of single indicators for multiple objects; comparative heatmap display of multiple indicators for multiple objects; and clustering graph display of big data analysis for multiple indicators for multiple objects. The clustering graph display utilizes the t-SNE algorithm to map the clustering results to a 2D planar space.

[0054] Step 4: System Self-Learning. This includes clustering result monitoring and self-training modules. The system considers scenarios with dynamic data changes, periodically monitoring clustering model performance metrics such as density and splitting. When performance metrics significantly decrease (based on system-set thresholds), the self-training module is activated to retrain the model and improve its performance.

[0055] Step 2 includes the following steps:

[0056] Step 2.1: Representation Space Learning. The core of this part is the neural network self-decoder model, which consists of an encoder and a decoder. The encoder transforms the high-dimensional blast furnace image data into low-dimensional representation space data using a deep neural network, and the decoder inversely transforms the low-dimensional representation space data back into high-dimensional blast furnace image data. After training, if the error between the fitted data obtained through the encoder and decoder and the original data is sufficiently small, the representation space learning is complete.

[0057] Step 2.2: Low-dimensional spatial clustering. The core of this part is the Gaussian Mixture Model (GMM) clustering model. The results of the high-dimensional blast furnace profile data after the encoder in Step 2.1 are used to cluster the data based on the GMM model. For a single blast furnace data point, its final cluster affiliation is determined according to the probability distribution of belonging to different clusters.

[0058] Step 2.3: Optimal Solution Algorithm Module. This part comprehensively considers the loss functions of the supervised learning model in Step 2.1 and the unsupervised learning model in Step 2.2, establishes a unified loss function and an optimal model, and finds the optimal solution for parameters in the representation space learning and clustering model to optimize the models in Step 2.1 and Step 2.2.

[0059] Unified loss function and optimization model:

[0060] In this invention, it is necessary to establish a unified loss function that combines the loss functions of representation space learning and clustering algorithms. Hereinafter, X represents the source data, and Z represents the low-dimensional space of the source data X, which is obtained through an encoder, i.e., Z = encode(X); This represents the reconstruction of source data X through an encoding-decoding process, i.e. For clustering in low-dimensional representation spaces, we use {u1, u2, ..., u...} j ,…} represents the central kernel of the clustering algorithm, using

[0061]

[0062] As sample X i The expression space of Z i With the central core U j The similarity function.

[0063] Therefore, the formula for the unified loss function is:

[0064]

[0065] Based on the loss function, the optimal solution for each parameter in the model is calculated using the Adam gradient descent algorithm.

[0066] Step 2.4: Output results, including the original image of each blast furnace, low-dimensional representation space data (encoder results in Step 2.1), and the cluster to which the cluster belongs (clustering results in Step 2.2).

[0067] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0068] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for constructing a benchmarking system for blast furnaces based on deep clustering algorithm, characterized in that, include: Step 1: Based on the production management system data, establish individual profiles for each blast furnace through statistical models, including operational indicators, emission indicators, molten iron cost indicators, and energy consumption indicators. Then, perform outlier removal and normalization on the above indicators, and abstract them into data points in a high-dimensional space. Step 2: Based on the deep clustering algorithm, cluster the data points in the high-dimensional space to obtain the clustering results; Step 3: Visualize the clustering results; Step 4: Monitor the clustering results and perform self-training when the performance index is below the preset threshold to improve the corresponding performance; Step 2 includes: Step 2.1: Based on the neural network self-decoder model, it is divided into encoder and decoder. The encoder transforms the high-dimensional blast furnace image data into low-dimensional expression space data through the deep neural network. The decoder inversely transforms the low-dimensional expression space data into high-dimensional blast furnace image data. If the error between the fitted data after encoder and decoder and the original data is within the preset range, the expression space learning is completed. Step 2.2: After the high-dimensional blast furnace profile data is processed by the encoder, clustering is performed based on the GMM model. For a single blast furnace data point, its final cluster affiliation is determined according to the probability distribution of belonging to different clusters. Step 2.3: Based on the loss functions of supervised and unsupervised learning models, establish a unified loss function and optimization model, and find the optimal solution for parameters in the representation space learning and clustering models; Step 2.4: Output the results, including the original image of each blast furnace, low-dimensional representation space data, and the cluster to which the cluster belongs.

2. The method for constructing a benchmarking system for blast furnaces based on deep clustering algorithm according to claim 1, characterized in that, The abnormal data removal includes: removing data exceeding a preset range by modeling the time series of production data of a single blast furnace; removing data exceeding a preset range by modeling the data of all blast furnaces in the industry; the upper limit of the preset range is 75th quartile + 1.3 * interquartile range, and the lower limit of the preset range is 25th quartile - 1.3 * interquartile range. The normalization process includes: normalizing each field X, with the expression being: .

3. The method for constructing a benchmarking system for ironmaking blast furnaces based on deep clustering algorithm according to claim 1, characterized in that, Let X represent the source data, and Z represent the low-dimensional space of the source data X, which is obtained through an encoder, i.e. ;use This represents the reconstruction of source data X through an encoding-decoding process, i.e. For clustering in low-dimensional representation spaces, use The central kernel of the clustering algorithm is represented by... As a sample Expression space expression With the central core The similarity function, where i and j are sequence numbers; The expression for the unified loss function is then: ; Based on the loss function, the optimal solution for each parameter in the model is calculated using the Adam gradient descent algorithm.

4. The method for constructing a benchmarking system for blast furnaces based on deep clustering algorithm according to claim 1, characterized in that, The visualization of clustering results includes: a list of indicator data, a comparative analysis chart of single indicators for multiple objects, a comparative heatmap of multiple indicators for multiple objects, and a big data analysis clustering chart of multiple indicators for multiple objects.

5. A system for constructing a benchmarking system for ironmaking blast furnaces based on a deep clustering algorithm, characterized in that, include: Module M1: Based on production management system data, establish individual profiles for each blast furnace through statistical models, including operational indicators, emission indicators, molten iron cost indicators, and energy consumption indicators. Then, perform abnormal data removal and normalization on the above indicators, and abstract them into data points in a high-dimensional space. Module M2: Based on deep clustering algorithm, it clusters data points in high-dimensional space to obtain clustering results; Module M3: Visualizes the clustering results; Module M4: Monitors clustering results and performs self-training when performance metrics fall below a preset threshold to improve performance. The module M2 includes: Module M2.1: Based on a neural network self-decoder model, it consists of an encoder and a decoder. The encoder transforms high-dimensional blast furnace image data into low-dimensional expression space data through a deep neural network. The decoder then inversely transforms the low-dimensional expression space data into high-dimensional blast furnace image data. If the error between the fitted data after encoder and decoder and the original data is within a preset range, then the expression space learning is complete. Module M2.2: After the high-dimensional blast furnace profile data is processed by the encoder, it is clustered based on the GMM model. For a single blast furnace data point, its final cluster affiliation is determined according to the probability distribution of belonging to different clusters. Module M2.3: Based on the loss functions of supervised and unsupervised learning models, establish a unified loss function and optimization model, and find the optimal solution of parameters in expression space learning and clustering models; Module M2.4: Output results, including the original image of each blast furnace, low-dimensional representation space data, and the cluster to which the cluster belongs.

6. The system for constructing a benchmarking system for ironmaking blast furnaces based on deep clustering algorithm according to claim 5, characterized in that, The abnormal data removal includes: removing data exceeding a preset range by modeling the time series of production data of a single blast furnace; removing data exceeding a preset range by modeling the data of all blast furnaces in the industry; the upper limit of the preset range is 75th quartile + 1.3 * interquartile range, and the lower limit of the preset range is 25th quartile - 1.3 * interquartile range. The normalization process includes: normalizing each field X, with the expression being: .

7. The system for constructing a benchmarking system for blast furnaces based on deep clustering algorithm according to claim 5, characterized in that, Let X represent the source data, and Z represent the low-dimensional space of the source data X, which is obtained through an encoder, i.e. ;use This represents the reconstruction of source data X through an encoding-decoding process, i.e. For clustering in low-dimensional representation spaces, use The central kernel of the clustering algorithm is represented by... As a sample Expression space expression With the central core The similarity function, where i and j are sequence numbers; The expression for the unified loss function is then: ; Based on the loss function, the optimal solution for each parameter in the model is calculated using the Adam gradient descent algorithm.

8. The system for constructing a benchmarking system for blast furnaces based on deep clustering algorithm according to claim 5, characterized in that, The visualization of clustering results includes: a list of indicator data, a comparative analysis chart of single indicators for multiple objects, a comparative heatmap of multiple indicators for multiple objects, and a big data analysis clustering chart of multiple indicators for multiple objects.

Images