Blast furnace state monitoring method and device based on explainability enhanced neural network

Abstract

This invention discloses a blast furnace condition monitoring method and apparatus based on an interpretability-enhanced neural network, comprising: acquiring historical blast furnace parameter data and historical blast furnace condition indicator variable data; adding time delays to the historical blast furnace parameter data and selecting feature variables to obtain historical blast furnace feature variable data; clustering the historical blast furnace feature variable data to obtain several basic operating condition data of the blast furnace operating status; establishing a blast furnace condition indicator variable monitoring model based on the basic operating condition data, and training the blast furnace condition indicator variable monitoring model based on the historical blast furnace parameter data and historical blast furnace condition indicator variable data; inputting real-time blast furnace feature variable data into the trained blast furnace condition indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace condition indicator variables. This overcomes the problem of insufficient interpretability in existing data-driven blast furnace condition monitoring methods, making them difficult for industrial field operators to trust.

Description

Technical Field

[0001] This application relates to the field of blast furnace smelting process monitoring technology, and in particular to a blast furnace condition monitoring method and device based on interpretable enhanced neural networks. Background Technology

[0002] Blast furnace ironmaking is an indispensable and crucial link in modern steel production, and its operating status directly affects the quality of molten iron, energy consumption, and the safety of equipment operation. The internal reaction processes of a blast furnace are complex, encompassing multiphase transport involving solid, liquid, and gas phases, as well as intense physicochemical reactions at high temperatures. These processes are characterized by strong nonlinear coupling, high time-varying nature, and unobservable characteristics. Therefore, accurately monitoring and determining the operating status of the blast furnace is a key technical challenge for achieving intelligent ironmaking.

[0003] Currently, industrial site conditions mainly rely on operator experience or expert knowledge-based rule-based status judgment methods. However, these methods are limited by human subjectivity and the scope of knowledge coverage, making it difficult to meet the real-time and accuracy requirements under complex operating conditions. Although some studies have attempted to introduce physical mechanism models to describe blast furnace conditions, the highly nonlinear and difficult-to-model blast furnace system means that pure mechanistic models often face problems of insufficient accuracy and poor adaptability in practical applications.

[0004] In recent years, with the development of data acquisition and industrial Internet of Things (IoT) technologies, massive amounts of real-time operational data have made data-driven blast furnace condition monitoring possible. Neural networks, due to their powerful nonlinear modeling capabilities, are widely used for tasks such as condition identification, prediction, and optimization control, becoming an important tool in intelligent blast furnace research. However, traditional neural network models are mostly "black box" structures, lacking interpretability, and face significant obstacles to widespread adoption in industrial applications with high safety requirements. Operators find it difficult to understand the judgment criteria provided by the model, hindering the traceability of results and the tracing of faults, and also impeding the accumulation and dissemination of operational experience.

[0005] Therefore, there is an urgent need for a neural network model that combines high accuracy and high interpretability for monitoring the condition of blast furnaces. This model should retain the expressive power of deep learning while enhancing the interpretable correlation between the model's internal features and output, helping operators understand the basis of the model's judgments, thereby improving the model's reliability and practical value in industrial settings. Summary of the Invention

[0006] The purpose of this application is to provide a method and apparatus for monitoring the condition of a blast furnace based on an interpretability-enhanced neural network. This invention designs a multi-channel attention multilayer perceptron model with added physical constraints to monitor the blast furnace condition. The model enhances the interpretability of the neural network by introducing physical prior constraints and assigning physical meaning to the base operating condition weights of the intermediate layer units. This blast furnace condition monitoring method provides accurate monitoring results while also possessing high interpretability.

[0007] According to a first aspect of the embodiments of this application, a blast furnace condition monitoring method based on an interpretable enhanced neural network is provided, comprising: S1: Obtain historical blast furnace parameter data and historical blast furnace status indicator variable data; S2: Add time delay to the historical blast furnace parameter data and select feature variables to obtain historical blast furnace feature variable data; S3: Cluster the historical blast furnace characteristic variable data to obtain several basic operating condition data of blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition status indicator variable data; S4: Establish a blast furnace condition indicator variable monitoring model based on the aforementioned base operating condition data, and train the blast furnace condition indicator variable monitoring model based on historical blast furnace parameter data and historical blast furnace condition indicator variable data, wherein: The blast furnace condition indicator variable monitoring model is a multi-channel attention multilayer perceptron model with added physical prior constraints. Its structure consists of an input layer, a physical prior constraint layer, an attention mechanism layer, and an intermediate hidden layer. Softmax Layer, base condition weight output layer, base condition state indicator variable input layer, Matmul The model consists of a blast furnace state indicator variable output layer and a physical prior constraint layer. The constraint rules of the physical prior constraint layer are as follows: based on the Euclidean distance between the input historical blast furnace parameter data and the basic operating condition characteristic variable data, a predetermined number of target basic operating condition characteristic variable data are selected from the basic operating condition characteristic variable data; the model is trained based on the root mean square error between the output of the blast furnace state indicator variable output layer and the historical blast furnace state variable data. S5: Input the real-time blast furnace characteristic variable data into the trained blast furnace state indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace state indicator variables.

[0008] According to a second aspect of the embodiments of this application, a blast furnace condition monitoring device based on an interpretable enhanced neural network is provided, characterized in that it includes: The data acquisition module is used to acquire historical blast furnace parameter data and historical blast furnace status indicator variable data; The feature variable selection module is used to add time delay to the historical blast furnace parameter data and select feature variables to obtain historical blast furnace feature variable data. The basic operating condition acquisition module is used to cluster the historical blast furnace characteristic variable data to obtain several basic operating condition data of the blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition status indicator variable data. The model building and training module is used to establish a blast furnace condition indicator variable monitoring model based on the base operating condition data, and to train the blast furnace condition indicator variable monitoring model based on historical blast furnace parameter data and historical blast furnace condition indicator variable data, wherein: The blast furnace condition indicator variable monitoring model is a multi-channel attention multilayer perceptron model with added physical prior constraints. Its structure consists of an input layer, a physical prior constraint layer, an attention mechanism layer, and an intermediate hidden layer. Softmax Layer, base condition weight output layer, base condition state indicator variable input layer, Matmul The model consists of a blast furnace state indicator variable output layer and a physical prior constraint layer. The constraint rules of the physical prior constraint layer are as follows: based on the Euclidean distance between the input historical blast furnace parameter data and the basic operating condition characteristic variable data, a predetermined number of target basic operating condition characteristic variable data are selected from the basic operating condition characteristic variable data; the model is trained based on the root mean square error between the output of the blast furnace state indicator variable output layer and the historical blast furnace state variable data. The monitoring output module is used to input real-time blast furnace characteristic variable data into the trained blast furnace state indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace state indicator variables.

[0009] According to a third aspect of the embodiments of this application, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0010] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0011] The technical solutions provided by the embodiments of this application may include the following beneficial effects: As illustrated in the above embodiments, this application proposes a blast furnace condition monitoring method based on enhanced interpretability. Its core lies in constructing a blast furnace condition indicator variable monitoring model that integrates physical prior constraints and neural networks. This method first extracts representative basic operating condition data from massive historical blast furnace data. During real-time monitoring, the model uses physical prior constraints to filter out the historical basic operating conditions most similar to the real-time blast furnace state, and then uses a neural network model to accurately calculate the contribution weight of each basic operating condition data point to the current state. This design makes the decision-making process of the blast furnace condition monitoring method based on the neural network model no longer a black box. The output weights clearly reveal which basic operating conditions, and to what extent, constitute the real-time blast furnace state, thus effectively overcoming the technical challenge of traditional data-driven models being difficult for on-site operators to trust due to insufficient interpretability. Furthermore, operators can trace the decision-making basis according to the weights, providing reliable support for analyzing the real-time blast furnace state, predicting trends, and precise control, significantly enhancing the model's practicality, reliability, and application value in industrial settings.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0014] Figure 1 This is a flowchart illustrating a blast furnace condition monitoring method based on an interpretable enhanced neural network according to an exemplary embodiment.

[0015] Figure 2 This is a bar chart showing the minimum redundancy maximum correlation score values ​​of each time-delay blast furnace parameter according to an exemplary embodiment.

[0016] Figure 3 This is a graph illustrating the monitoring results of silicon content in molten iron after applying the blast furnace condition monitoring method to blast furnace data, according to an exemplary embodiment.

[0017] Figure 4 This is a block diagram illustrating a blast furnace condition monitoring device based on decoupling and reconfiguration of operating status, according to an exemplary embodiment. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0019] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0020] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0021] Figure 1 This is a flowchart illustrating a blast furnace condition monitoring method based on an interpretable enhanced neural network according to an exemplary embodiment, such as... Figure 1 As shown, the method may include the following steps: In the specific implementation of S1, historical blast furnace parameter data and historical blast furnace status indicator variable data are obtained; Specifically, obtain historical blast furnace parameter data X={x1,x2,..,x m},in m The number of types of blast furnace parameters, each parameter includes n Sample data with a sampling interval of 1 hour; obtain historical blast furnace state indicator variable data Y={y1,y2,…,y o},in o The number of data types for blast furnace condition indicator variables; each historical blast furnace condition indicator variable also includes n Sample data with a sampling interval of 1 hour.

[0022] This step is fundamental to building any data-driven model. As a complex industrial process, the operating state of a blast furnace is determined by numerous operating parameters. Therefore, it is essential to acquire real historical data covering a long timeframe and containing rich variations in operating conditions as the basis for subsequent model training and testing.

[0023] In this embodiment, 1440 blast furnace data points from two months within a certain year were selected from a database of 2650m³ blast furnaces in China. The sampling interval was 1 hour. The first 1140 data points were used as the training set for decoupling the blast furnace operating state and training the subsequent multi-channel attention multilayer perceptron model with added physical constraints. The last 300 data points were divided into a validation set and a test set in a 2:1 ratio to validate and test the blast furnace operating state reconstruction model. The blast furnace parameter data included total differential pressure (kPa), permeability index (m³ / s), and pressure difference (m³ / s). 6 ·kPa⁻¹·h⁻²), blower kinetic energy (kg·m 2 ·s -2 Hot air pressure (kPa), hot air temperature (°C), furnace top temperature (northeast, southwest, northwest, southeast), and blast humidity (g·m³). 3 Actual wind speed (m) 3 / s), furnace gas index (m 3 / (h·m 2 The data on blast furnace condition indicators include the following: resistance coefficient, theoretical combustion temperature (°C), cold air flow rate (m³ / h), cold air pressure (kPa), furnace top pressure (kPa), oxygen enrichment rate (%), oxygen enrichment flow rate (m³ / h), and pulverized coal injection rate (t). The blast furnace condition indicator variables include the silicon content (%) in molten iron, the H2 content (%) at the furnace top, the CO content (%) at the furnace top, the CO2 content (%) at the furnace top, and the molten iron temperature (°C).

[0024] In the specific implementation of S2, a time lag is added to the historical blast furnace parameter data and feature variables are selected to obtain historical blast furnace feature variable data; this step includes the following sub-steps: S21: Add time lag to each type of historical blast furnace parameter data, and calculate the mutual information between the lagged historical blast furnace parameter data and the silicon content data in the historical blast furnace molten iron. Specifically, the feature variable selection steps are as follows: (1) Add time lags of 0h, 1h, 2h, and 3h to each type of blast furnace parameter data, and calculate the mutual information between the lagged historical blast furnace parameter data and the historical silicon content data in molten iron. The formula for calculating the mutual information is:

[0025] Where X represents historical blast furnace parameter data, and y represents historical silicon content data in molten iron from the blast furnace.p ( x , y ) represents the joint probability distribution. p ( x ), p ( y ) represent the marginal probability distributions; S22: Select the historical blast furnace parameter data with the greatest mutual information with the silicon content data in the molten iron from all historical blast furnace parameter data, and use it as the first type of selected historical blast furnace characteristic variable data. Specifically, the historical blast furnace parameter data are sorted according to the mutual information value, and the parameter data with the largest mutual information with the silicon content data in the historical blast furnace molten iron is selected from all the historical blast furnace parameter data as the first type of selected blast furnace characteristic variable data. S23: For each type of historical blast furnace parameter data, calculate its minimum redundancy maximum correlation score, select the historical blast furnace parameter data with the highest score and add it to the selected historical blast furnace feature variable data, and remove it from the historical blast furnace parameter data. Repeat S23 until several types of historical blast furnace feature variable data are selected.

[0026] Specifically, for each type of blast furnace parameter data, its minimum redundancy maximum correlation score is calculated using the following formula:

[0027] in, For the selected historical feature variable dataset; select the historical blast furnace parameter data with the highest score and add it. Remove it from the historical blast furnace parameter data, and repeat step S23 until selected. q The historical blast furnace characteristic variable data obtained from the blast furnace characteristic variable data is denoted as follows: ,in q The number of blast furnace characteristic variables selected.

[0028] Blast furnace ironmaking is a physicochemical process with significant inertia and time lag. Changes in input parameters (such as adjusting pulverized coal injection) do not have an instantaneous impact on output states (such as molten iron temperature). Therefore, adding different time lags to the input parameters is to simulate and compensate for this physical delay effect at the data level. The original blast furnace parameters are numerous and interrelated, exhibiting information redundancy and noise interference. Directly using all parameters for modeling would increase model complexity, prolong training time, and potentially lead to overfitting. Advanced feature selection algorithms such as mRMR are employed to automatically and systematically identify the core parameter combinations that most decisively affect the blast furnace state. By considering time lags, the causal relationships of blast furnace operation are captured more realistically, significantly improving the model's prediction accuracy. By carefully selecting feature variables, the data dimensionality is effectively reduced, the computational burden is decreased, and the model can focus more on key influencing factors.

[0029] In this embodiment, historical blast furnace characteristic variable data are selected from the historical blast furnace parameter data in the training set. The minimum redundancy maximum correlation score between each historical blast furnace parameter data and the silicon content data in historical molten iron is as follows: Figure 2 As shown.

[0030] The top 10 blast furnace parameters with the highest minimum redundancy and maximum correlation scores were selected as blast furnace characteristic variables. The final selected blast furnace characteristic variables include: blast humidity (g·m³). 3 ), oxygen enrichment flow rate (m³ / h), cold air flow rate with time lag 3h (m³ / h), hot air temperature (°C), theoretical combustion temperature (°C), furnace top temperature southeast (°C), resistance coefficient with time lag 3h, furnace top temperature southwest (°C), total pressure difference (kPa), pulverized coal injection rate (t).

[0031] In the specific implementation of S3, the historical blast furnace characteristic variable data are clustered to obtain several basic operating condition data of the blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition state indicator variable data; this step includes the following sub-steps: S31: Apply the mean-shift clustering algorithm to the historical blast furnace characteristic variable data to obtain several cluster centers; Specifically, the mean-shift clustering algorithm has the following steps: (1) Randomly select a sample from the historical blast furnace characteristic variable dataset as the center point c, and construct a radius of c. r Sliding window; (2) Calculate the average distance from other sample points within the sliding window to the center point, and take this average distance as the offset M. The calculation formula is as follows:

[0032] in, w The number of sample points in the sliding window. It is the set of all sample points in the sliding window; (3) Update the position of the center point according to the offset M. The calculation formula is as follows:

[0033] in, l The number of iterations; (4) Repeat steps (1) to (3) until the position of the center point no longer changes; (5) If the distance between the center point of the current class and the center point of another class is less than the set threshold, the two classes will be merged into one class; otherwise, the number of center points will be increased by 1. (6) Repeat steps (1) to (5) until all sample points are traversed to obtain the cluster center set of the historical blast furnace characteristic variable dataset.

[0034] S32: Calculate the Euclidean distance between each data point and each cluster center in the historical blast furnace characteristic variable data; S33: Select the historical blast furnace characteristic variable data with the smallest Euclidean distance from each cluster center as the blast furnace basic operating condition characteristic variable data, and save the historical blast furnace state indicator variable data at the corresponding time for each basic operating condition to obtain the basic operating condition state indicator variable data.

[0035] This is one of the core steps in achieving model interpretability. Its purpose is to simplify the continuously changing blast furnace state space, which is difficult to model directly, into a set of finite, relatively stable, and physically meaningful basic operating conditions. Each basic operating condition can be understood as a typical blast furnace operating mode. This transforms a complex continuous state monitoring problem into a relatively simple problem of identifying and combining discrete basic operating conditions. This greatly simplifies the subsequent modeling process and provides operators with an understandable state baseline, allowing the final monitoring results to be presented as combinations of different basic operating conditions representing the current state, thus laying the foundation for model interpretability.

[0036] In this embodiment, the mean-shift clustering algorithm is applied to the training set data, with the cluster radius parameter... r With the dimension set to 0.015, the final number of base condition data obtained after decoupling is 39.

[0037] In the specific implementation of S4, a blast furnace condition indicator variable monitoring model is established based on the base operating condition data, and the blast furnace condition indicator variable monitoring model is trained based on historical blast furnace parameter data and historical blast furnace condition indicator variable data. Specifically, the relationship between the number of basic operating conditions required for the blast furnace remodeling process and the number of monitored blast furnace condition indicator variables is analyzed; assuming a total of uFor each blast furnace operating condition, the corresponding blast furnace state indicator variables are: ,in type The category of blast furnace state indicator variables includes a series of state indicators that can indicate the operating conditions inside the blast furnace, such as silicon content in molten iron, molten iron temperature, CO content in top gas, and CO2 content in top gas. The blast furnace state reconstruction process can then be represented as:

[0038] in d The amount of base condition data used in the refactoring process. This represents the weight corresponding to the i-th base condition. It's important to note that not all base conditions are used in the reconstructing process; the number of base conditions used is related to the number of monitored blast furnace condition indicator variables. A detailed explanation follows: Without inequality constraints and ensuring the equations are independent, the existence, uniqueness, and multiplicity of solutions to a system of equations are directly related to the number of variables and the number of equations. In the problem described above, the number of variables represents the number of base cases used in the reconstruction process, and the number of equations is the same as the number of monitored variables. When the number of equations equals the number of variables, this is a typical full-rank case, and the system has a unique solution. When the number of equations is less than the number of variables, the system is an underdetermined system of equations. In this case, the solution space has more degrees of freedom, and the system has infinitely many solutions. When the number of equations is greater than the number of variables, the system is an overdetermined system of equations. In this case, the system usually has no solution.

[0039] When the system of equations includes inequality constraints, the solution will be affected by these constraints. However, in either case, adding inequality constraints only reduces the solution space. Therefore, to ensure the feasibility of the reconstruction process, the constructed system of equations should be at least a full-rank system or an underdetermined system. Furthermore, to ensure the safety and reliability of the production process, industrial sites generally require minimizing uncertainties in the system. Even with inequality constraints, an underdetermined system may still have multiple solutions. Constructing such a blast furnace state reconstruction equation system might lead to non-uniqueness of the corresponding basic operating condition combinations, meaning it's impossible to accurately determine which basic operating conditions constitute the current blast furnace state. Although this doesn't affect the monitoring results of the blast furnace state, it makes the physical meaning of the entire reconstruction process unclear. This reduces the model's interpretability and hinders subsequent operator decision-making and control system stability in practical applications. Based on the above analysis, establishing a full-rank system of equations to represent the blast furnace state reconstruction process is the ideal approach. l When a blast furnace condition indicator variable is used as a monitoring indicator, the equation system contains... l +1 equation, therefore choosel +1 base operating conditions to reconstruct the blast furnace state.

[0040] The blast furnace condition indicator variable monitoring model is a multi-channel attention multilayer perceptron model with added physical prior constraints. Its structure consists of an input layer, a physical prior constraint layer, an attention mechanism layer, and an intermediate hidden layer. Softmax Layer, base condition weight output layer, base condition state indicator variable input layer, Matmul The model consists of a blast furnace state indicator variable output layer and a physical prior constraint layer. The constraint rules of the physical prior constraint layer are as follows: based on the Euclidean distance between the input historical blast furnace parameter data and the basic operating condition characteristic variable data, a predetermined number of target basic operating condition data are selected from the basic operating condition data; the model is trained based on the root mean square error between the output of the blast furnace state indicator variable output layer and the historical blast furnace state indicator variable data.

[0041] The specific steps of each layer of the blast furnace condition indicator variable monitoring model are as follows: (1) Input layer: The input layer's role is to receive external data, which is typically in the form of feature vectors. The form of representation, in which d It is the dimension of the features. The input layer will Pass it directly to the first hidden layer without any additional transformation:

[0042] (2) Physical prior constraint layer: In industrial systems, historical data and experience show that the distributions of characteristic variables under similar operating conditions tend to be close. By selecting geometrically similar baseline operating conditions to reconstruct the blast furnace operating state, the dynamic changes in operating conditions can be captured more accurately. According to the aforementioned embodiments, when simultaneously using... l When using state indicator variables as monitoring indicators, the number of ideal base conditions required for the blast furnace state reconstruction process is: l +1. At this point, the blast furnace reconstruction process will be constrained to the state that best conforms to objective physical laws, improving reconstruction accuracy while ensuring the clarity of the physical meaning of the reconstruction process. k The formula for calculating the Euclidean distance between the feature vectors of each basic operating condition and the blast furnace operating state is as follows:

[0043] in For the first k The first basic working condition i There are 1 feature variables. The one with the smallest distance is the first... lOne base operating condition will be selected for blast furnace operation status reconstruction. Only the subsequent channels corresponding to the selected base operating conditions will be opened, and real-time data will be transmitted to the base operating condition weight calculation network layer through these channels.

[0044] (3) Attention mechanism layer: Different attention modules are trained for different baseline conditions, enabling each attention module to allocate feature variable weights based on the data characteristics of the corresponding baseline condition. The calculation steps for the attention module are as follows: The attention module input consists of three main parts: a query vector Q representing the target to be focused on, a key vector K for matching the query, and a value vector V corresponding to the actual information of each key. For a given input sequence... These three parts are obtained through linear transformation:

[0045]

[0046]

[0047] in, , ， It is a learnable weight matrix. Similarity is calculated by taking the dot product of the query vector Q and the key vector K to quantify their correlation; the formula is as follows:

[0048] in, This is the dimension of the key vector. (Using...) Scaling is applied to avoid excessively large dot product values, thereby ensuring gradient stability.

[0049] use Softmax The function converts similarity into normalized attention weights:

[0050] The final attention output is generated by weighting and summing the value vector V using attention weights.

[0051] (4) Intermediate hidden layer: The hidden layer is the core component of a multilayer perceptron, containing one or more hidden layers. The output of the hidden layer is calculated by the following steps:

[0052] in The output of the hidden layer, lIndicates the index of the current layer. l =1,2,…, L ; For the first l The weight matrix of the layer represents the weight matrix of the first layer. l The connection strength between neurons in one layer and neurons in the layer above; It is the first l Layer bias vector; This is the activation function. In this invention, the number of neurons in the last hidden layer is set to the number of base cases, and it is assigned the physical meaning of the base case weights.

[0053] (5) Softmax layer: To ensure that the sum of the weights of all base cases is 1, the outputs of the neurons in the last hidden layer of the multilayer perceptron network corresponding to all base cases selected for the reconstruction process are processed... Softmax The calculation formula for the layer is as follows:

[0054] in w i For the first i The weights corresponding to each basic working condition i ∈C, where C is the set of base case indices selected for the reconstruction process. The weight output for base cases not selected for the reconstruction process is 0.

[0055] (6) Base condition weight output layer: The base condition weight output layer is used to output the calculation results of the base condition weight. Softmax The output of the base case weight output layer serves as the input to the base case weight output layer, which does not perform any additional transformations on the input.

[0056] (7) Input layer of basic working condition indicator variables: The role of the input layer for the basic operating condition indicator variables is to receive the data of the basic operating condition indicator variables. ,Will It is directly passed to subsequent network layers without any additional transformation.

[0057] (8) Matmul layer: Matmul The function of the layer is to perform matrix multiplication, linearly combining the base condition weight vector output by the previous layer with the corresponding base condition state indicator variable data matrix. The calculation formula is shown below:

[0058] (9) Output layer of blast furnace state indicator variables: The blast furnace condition indicator variable output layer is used to output the monitoring results of blast furnace condition indicator variables. Matmul The output of the layer serves as the input to the blast furnace state indicator variable output layer, and the blast furnace state indicator variable output layer does not perform any additional transformations on the input.

[0059] The model employs a physical prior constraint layer to ensure a high degree of consistency in the internal logic of the model during two independent training sessions. This significantly improves the model's interpretability and reproducibility, resolving the "black box" and "randomness" problems of traditional neural networks. This results in stable and reliable output interpretation, crucial for gaining operator trust in industrial applications. Furthermore, the model preprocesses the input data using a parallel attention module, selecting and weighting the most important features for each base condition. These focused features are then fed into the intermediate hidden layers of the neural network, leveraging its powerful nonlinear modeling capabilities to ensure the accuracy of the calculated base condition weights and the model's monitoring results.

[0060] In the specific implementation of S5, real-time blast furnace characteristic variable data are input into the trained blast furnace state indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace state indicator variables.

[0061] Specifically, in this embodiment, the structure of the multilayer sensor is 20-64-39, and the monitored blast furnace status indicators are silicon content in molten iron, H2 content at the top of the furnace, CO content at the top of the furnace, CO2 content at the top of the furnace, and molten iron temperature.

[0062] The overall performance of the blast furnace condition monitoring model is determined by two factors: the model's accuracy in monitoring all blast furnace condition indicator variables and its interpretability. Regarding the overall monitoring accuracy, the sum of the RMSE (Recovery Mean Squared Error) of all indicator monitoring results represents the model's overall accuracy. As for interpretability, a major reason why neural networks are difficult to interpret is their poor repeatability. That is, for the same set of data, the models trained twice have low similarity, which is visually manifested as significant differences in the output distribution of intermediate layer neurons in two repeated experiments. This negatively impacts the tracing and interpretation of the model's monitoring results. The model proposed in this invention assigns physical meaning to the weights of the base operating conditions to the intermediate layer neurons, inferring the monitoring results of the blast furnace condition indicator variables at the current moment by weighted combination of the values ​​of the blast furnace condition indicator variables corresponding to the base operating conditions. In the real physical world, the base operating conditions from which the blast furnace state is reconstructed should be uniquely determined. However, if the results obtained by the intermediate layer neurons are different in each training session, the reconstruction process becomes difficult to interpret. Therefore, the difference in the intermediate layer neurons under multiple training sessions is used as a verification indicator of the model's interpretability. The formula for calculating the design difference index is shown below:

[0063] in, W i For the first i The base condition weight matrix obtained from this experiment It is the average matrix of all weight matrices. Let f be the Frobenius norm of the matrix. Diff The smaller the value, the better the repeatability of the experiment. The final overall performance evaluation index of the constructed model is shown in the following formula:

[0064] in This is the overall monitoring accuracy index of the normalized model. This is the normalized model interpretability index. P The smaller the value, the better the overall performance of the model.

[0065] In this embodiment, the blast furnace condition monitoring model was trained 10 times and tested on a test set. To compare the superiority of the proposed model in terms of interpretability, a multilayer perceptron model without physical constraints was used for comparison. The evaluation metrics for the monitoring results of the two methods are shown in the table below.

[0066]

[0067] Compared to the multilayer perceptron model without physical constraints, the blast furnace condition monitoring model proposed in this invention has a slight decrease of 1.33% in overall monitoring accuracy, but a significant improvement of 51.93% in repeatability. Figure 3 This paper presents the monitoring results of the blast furnace condition monitoring model proposed in this invention on the blast furnace condition indicator variables. From top to bottom, the monitoring results are for silicon content in molten iron, H2 content at the furnace top, CO content at the furnace top, CO2 content at the furnace top, and molten iron temperature. The dashed lines represent the actual values ​​of the blast furnace condition indicator variables, and the solid lines represent the estimated values ​​of silicon content in the molten iron. Figure 3 As shown, among all the monitoring results of blast furnace condition indicator variables, the solid and dashed lines fit ideally, indicating that the monitoring results are accurate. These results demonstrate that the blast furnace condition monitoring model proposed in this invention not only ensures the accuracy of the monitoring results but also possesses higher interpretability and greater potential for industrial application.

[0068] Corresponding to the aforementioned embodiments of the blast furnace condition monitoring method based on interpretability-enhanced neural networks, this application also provides embodiments of a blast furnace condition monitoring device based on interpretability-enhanced neural networks.

[0069] Figure 4 This is a block diagram illustrating a blast furnace condition monitoring device based on an interpretable enhanced neural network, according to an exemplary embodiment. (Refer to...) Figure 4 The device includes: Data acquisition module 1 is used to acquire historical blast furnace parameter data and historical blast furnace status indicator variable data; Feature variable selection module 2 is used to add time delay to the historical blast furnace parameter data and select feature variables to obtain historical blast furnace feature variable data; The basic operating condition acquisition module 3 is used to cluster the historical blast furnace characteristic variable data to obtain several basic operating condition data of the blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition status indicator variable data. Model building and training module 4 is used to establish a blast furnace condition indicator variable monitoring model based on the base operating condition data, and to train the blast furnace condition indicator variable monitoring model based on historical blast furnace parameter data and historical blast furnace condition indicator variable data, wherein: The blast furnace condition indicator variable monitoring model is a multi-channel attention multilayer perceptron model with added physical prior constraints. Its structure consists of an input layer, a physical prior constraint layer, an attention mechanism layer, and an intermediate hidden layer. Softmax Layer, base condition weight output layer, base condition state indicator variable input layer, MatmulThe model consists of a blast furnace state indicator variable output layer and a physical prior constraint layer. The constraint rules of the physical prior constraint layer are as follows: based on the Euclidean distance between the input historical blast furnace parameter data and the basic operating condition characteristic variable data, a predetermined number of target basic operating condition characteristic variable data are selected from the basic operating condition characteristic variable data; the model is trained based on the root mean square error between the output of the blast furnace state indicator variable output layer and the historical blast furnace state variable data. The monitoring output module 5 is used to input real-time blast furnace characteristic variable data into the trained blast furnace state indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace state indicator variables.

[0070] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0071] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0072] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the blast furnace condition monitoring method based on the interpretable enhanced neural network described above.

[0073] Accordingly, this application also provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the blast furnace condition monitoring method based on an interpretability-enhanced neural network as described above.

[0074] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0075] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A blast furnace condition monitoring method based on interpretable augmented neural networks, characterized in that, include: S1: Obtain historical blast furnace parameter data and historical blast furnace status indicator variable data; S2: Add time delay to the historical blast furnace parameter data and select feature variables to obtain historical blast furnace feature variable data; S3: Cluster the historical blast furnace characteristic variable data to obtain several basic operating condition data of blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition status indicator variable data; S4: Establish a blast furnace condition indicator variable monitoring model based on the aforementioned base operating condition data, and train the blast furnace condition indicator variable monitoring model based on historical blast furnace parameter data and historical blast furnace condition indicator variable data, wherein: The blast furnace condition indicator variable monitoring model is a multi-channel attention multilayer perceptron model with added physical prior constraints. Its structure consists of an input layer, a physical prior constraint layer, an attention mechanism layer, and an intermediate hidden layer. Softmax Layer, base condition weight output layer, base condition state indicator variable input layer, Matmul The model consists of a blast furnace state indicator variable output layer and a physical prior constraint layer. The constraint rules of the physical prior constraint layer are as follows: based on the Euclidean distance between the input historical blast furnace parameter data and the basic operating condition characteristic variable data, a predetermined number of target basic operating condition characteristic variable data are selected from the basic operating condition characteristic variable data; the model is trained based on the root mean square error between the output of the blast furnace state indicator variable output layer and the historical blast furnace state variable data. S5: Input the real-time blast furnace characteristic variable data into the trained blast furnace state indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace state indicator variables; Specifically, clustering the historical blast furnace characteristic variable data yields several basic operating condition data points for the blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition state indicator variable data, including: S31: Apply the mean-shift clustering algorithm to the historical blast furnace characteristic variable data to obtain several cluster centers; S32: Calculate the Euclidean distance between each data point and each cluster center in the historical blast furnace characteristic variable data; S33: Select the historical blast furnace characteristic variable data with the smallest Euclidean distance from each cluster center as the blast furnace basic operating condition characteristic variable data, and save the historical blast furnace state indicator variable data at the corresponding time for each basic operating condition to obtain the basic operating condition state indicator variable data. The training of the blast furnace condition indicator variable monitoring model based on historical blast furnace data includes: S41: Input historical blast furnace parameter data into the input layer of the blast furnace condition indicator variable monitoring model, and impose constraints on the subsequent base conditions by the physical prior constraint layer. The constraint rule is: among all base condition characteristic variable data, only the data with the closest Euclidean distance between the base condition characteristic variable data and the input historical blast furnace parameter data is retained. The characteristic variable data of each basic working condition are used for subsequent calculations, among which The number of output variables representing the blast furnace state in the model; S42: Select the... The basic operating condition characteristic variable data are concatenated with the input historical blast furnace parameter data, and the concatenated matrix is ​​input into the corresponding attention mechanism module, passing through the intermediate hidden layer and... Softmax The layer is normalized, and the weight of each base condition is obtained from the base condition weight output layer, where the weight of the base condition that is not retained is 0. S43: Input the base condition weights and the base condition state indicator variable data together. Matmul The layers are multiplied and summed to obtain the monitoring output of the blast furnace state indicator variables from the output layer of the blast furnace state indicator variables; S45: Determine the loss function based on the root mean square error between the monitoring output and the corresponding historical blast furnace condition indicator variable data, and train the model based on the loss function to obtain the trained blast furnace condition indicator variable monitoring model.

2. The method according to claim 1, characterized in that, By adding time delays to the historical blast furnace parameter data and selecting characteristic variables, historical blast furnace characteristic variable data is obtained, including: S21: Add time lag to each type of historical blast furnace parameter data, and calculate the mutual information between the lagged historical blast furnace parameter data and the silicon content data in the historical blast furnace molten iron. S22: Select the historical blast furnace parameter data with the greatest mutual information with the silicon content data in the molten iron from all historical blast furnace parameter data, and use it as the first type of selected historical blast furnace characteristic variable data. S23: For each type of historical blast furnace parameter data, calculate its minimum redundancy maximum correlation score, select the historical blast furnace parameter data with the highest score and add it to the selected historical blast furnace feature variable data, and remove it from the historical blast furnace parameter data. Repeat S23 until several types of historical blast furnace feature variable data are selected.

3. A blast furnace condition monitoring device based on an interpretable enhanced neural network, characterized in that, include: The data acquisition module is used to acquire historical blast furnace parameter data and historical blast furnace status indicator variable data; The feature variable selection module is used to add time delay to the historical blast furnace parameter data and select feature variables to obtain historical blast furnace feature variable data. The basic operating condition acquisition module is used to cluster the historical blast furnace characteristic variable data to obtain several basic operating condition data of the blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition status indicator variable data. The model building and training module is used to establish a blast furnace condition indicator variable monitoring model based on the base operating condition data, and to train the blast furnace condition indicator variable monitoring model based on historical blast furnace parameter data and historical blast furnace condition indicator variable data, wherein: The blast furnace condition indicator variable monitoring model is a multi-channel attention multilayer perceptron model with added physical prior constraints. Its structure consists of an input layer, a physical prior constraint layer, an attention mechanism layer, and an intermediate hidden layer. Softmax Layer, base condition weight output layer, base condition state indicator variable input layer, Matmul The model consists of a blast furnace state indicator variable output layer and a physical prior constraint layer. The constraint rules of the physical prior constraint layer are as follows: based on the Euclidean distance between the input historical blast furnace parameter data and the basic operating condition characteristic variable data, a predetermined number of target basic operating condition characteristic variable data are selected from the basic operating condition characteristic variable data; the model is trained based on the root mean square error between the output of the blast furnace state indicator variable output layer and the historical blast furnace state variable data. The monitoring output module is used to input real-time blast furnace characteristic variable data into the trained blast furnace state indicator variable monitoring model to obtain the monitoring values ​​of all monitored blast furnace state indicator variables. Specifically, clustering the historical blast furnace characteristic variable data yields several basic operating condition data points for the blast furnace operating status, including basic operating condition characteristic variable data and basic operating condition state indicator variable data, including: S31: Apply the mean-shift clustering algorithm to the historical blast furnace characteristic variable data to obtain several cluster centers; S32: Calculate the Euclidean distance between each data point and each cluster center in the historical blast furnace characteristic variable data; S33: Select the historical blast furnace characteristic variable data with the smallest Euclidean distance from each cluster center as the blast furnace basic operating condition characteristic variable data, and save the historical blast furnace state indicator variable data at the corresponding time for each basic operating condition to obtain the basic operating condition state indicator variable data. The training of the blast furnace condition indicator variable monitoring model based on historical blast furnace data includes: S41: Input historical blast furnace parameter data into the input layer of the blast furnace condition indicator variable monitoring model, and impose constraints on the subsequent base conditions by the physical prior constraint layer. The constraint rule is: among all base condition characteristic variable data, only the data with the closest Euclidean distance between the base condition characteristic variable data and the input historical blast furnace parameter data is retained. The characteristic variable data of each basic working condition are used for subsequent calculations, among which The number of output variables representing the blast furnace state in the model; S42: Select the... The basic operating condition characteristic variable data are concatenated with the input historical blast furnace parameter data, and the concatenated matrix is ​​input into the corresponding attention mechanism module, passing through the intermediate hidden layer and... Softmax The layer is normalized, and the weight of each base condition is obtained from the base condition weight output layer, where the weight of the base condition that is not retained is 0. S43: Input the base condition weights and the base condition state indicator variable data together. Matmul The layers are multiplied and summed to obtain the monitoring output of the blast furnace state indicator variables from the output layer of the blast furnace state indicator variables; S45: Determine the loss function based on the root mean square error between the monitoring output and the corresponding historical blast furnace condition indicator variable data, and train the model based on the loss function to obtain the trained blast furnace condition indicator variable monitoring model.

4. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-2.

5. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-2.

Images