Hierarchical strategy method for enhancing molten iron quality related modeling and fault detection

Abstract

The invention discloses a hierarchical strategy method for enhancing molten iron quality related modeling and fault detection. The hierarchical strategy method comprises the steps of blast furnace ironmaking process multiphase multi-field mechanism modeling, data-driven stationary feature extraction, time matching generalized learning system-based molten iron quality (MIQ) modeling and fault diagnosis process design. And a mechanism and data collaborative driving strategy is designed, so that the transparency and MIQ prediction of the model are improved. According to the method, material and thermal trends are captured by dividing a hearth and applying mechanism-based characteristics, and a novel stable generalized characteristic learning system is combined, so that interference caused by non-stable process characteristics is reduced, and internal information embedded in the blast furnace ironmaking process is excavated. Subsequently, process and quality variables are aligned with MIQ as a target by combining stationary feature representations with mechanism features. Through verification with real data, the model shows consistent process alignment, powerful feature extraction and improved MIQ modeling, so that better fault detection is realized.

Description

Technical Field

[0001] The present invention relates to a hierarchical strategy method for enhancing molten iron quality-related modeling and fault detection. Background Art

[0002] The Blast Furnace Ironmaking Process (BFIP) is positioned as an essential process, accounting for not only 70% of the total energy consumption but also having a significant impact on subsequent steelmaking and rolling. For the product of BFIP, namely molten iron, trace elements such as silicon, phosphorus, and sulfur play a crucial role in evaluating the overall Molten Iron Quality (MIQ). Different from other industrial production processes, the blast furnace itself is a highly complex large-scale system, and its production process mechanism model is also complex. At the same time, for MIQ testing, evaluating the quality of molten iron involves a series of steps from tapping to sampling and subsequent laboratory analysis, which takes about 30 minutes. Considering that the working cycle of BFIP is about 30 minutes, the inherent delay in MIQ evaluation poses a huge challenge to on-site engineers, with the goal of making real-time and precise adjustments. Meanwhile, BFIP is vulnerable to various potential failures related to MIQ, and ignoring the timely monitoring of these problems will seriously affect the entire production process and may escalate into serious production accidents in extreme cases. In addition, the process variables in blast furnace ironmaking operations are affected by non-stationary factors, which brings complexity and greatly hinders the rapid and reliable detection of quality-related faults. Therefore, it is crucial to establish an effective monitoring system to timely detect and handle related faults and ensure the smooth and safe operation of BFIP.

[0003] Currently, the modeling and process monitoring of MIQ mainly rely on data-driven technologies. Existing studies have used methods such as Least Squares Support Vector Regression (LS-SVR) and Graph Neural Network (GNN) to predict MIQ. However, most of these methods only focus on a single MIQ parameter and ignore the simultaneous consideration of multiple parameters. To overcome the limitations, Zhou et al. proposed a modeling method based on Random Vector Functional Link Network (RVFLN), which considers multiple MIQ parameters and overcomes challenges in aspects such as variable selection, time-varying coupling of process dynamics, and model robustness. To further advance this field, Wang et al. developed methods such as Kernel Partial Least Squares (KPLS) and Improved Kernel Orthogonal Projection to Latent Structures (IKOPLS) based on statistical analysis to achieve simultaneous multi-MIQ modeling and process monitoring. Although progress has been made, there are still some challenges to be solved.

[0004] The complex characteristics of BFIP introduce variations and fluctuations, posing significant challenges in accurately identifying potential underlying data features. This non-stationarity can be caused by changes in raw material composition, variations in process conditions, and other operating factors affecting the blast furnace, resulting in patterns that change over time.

[0005] Although data-driven modeling methods can provide accurate predictions, they often lack interpretability. The black-box nature of these models makes it challenging to discern the underlying relationships between process variables and MIQ. Interpretable models provide insights into how various factors affect MIQ, enabling operators and engineers to clearly understand the process. Greater transparency allows them to make informed decisions by analyzing the reasoning of the model.

[0006] Accounting for the time lag between process samples and MIQ measurements is crucial for accurate modeling and monitoring. Typically, there is a time lag between changes in MIQ and the corresponding process data, which can be caused by various factors in the measurement workflow, such as sampling transportation time and the duration of laboratory tests.

[0007] Mechanistic modeling is based on fundamental theoretical knowledge, combining principles such as chemical reactions, mass / heat conservation, and understanding the evolution characteristics of the process. Early studies by researchers such as Ridgion focused on material and heat balance equations to understand the differences in chemical reactions occurring in different temperature zones within the blast furnace, thereby gaining in-depth understanding of thermochemical properties. Subsequent work further enhanced BFIP modeling by introducing additional equations for influencing factors such as heat transfer, gas-solid reactions, and component behavior. Additionally, recent research has limited the scope of mechanistic modeling to local furnace areas, overlooking the overall understanding. However, it must be recognized that existing BFIP modeling does not fully reflect the complexity of the process. In recent years, a collaborative approach that combines mechanistic modeling and data-driven techniques has emerged. By combining the advantages of data-driven modeling with mechanistic model insights, co-driven models can leverage the accuracy of data-driven patterns while obtaining perspectives from process understanding, thereby promoting improvements in tasks such as MIQ modeling and process monitoring. Summary of the Invention

[0008] To overcome the deficiencies of the prior art, the present invention provides a hierarchical strategy method for enhancing molten iron quality-related modeling and fault detection.

[0009] The present invention proposes a mechanism-data co-driven strategy (MDCDS) to address the challenges posed by the complexity of non-stationary characteristics, complex mechanisms, and the mismatch between MIQ and process samples. The BFIP mechanism model here divides the blast furnace into different reaction zones, allowing in-depth analysis of reactions at different heights. By combining the main thermochemical reactions, conservation laws, and specific external interactions in each zone, the model is both interpretable and robust, and a large number of unmeasured mechanism features can be naturally derived. In addition, a sequential iterative algorithm (SIA) is proposed to further refine the model parameters and iteratively improve the model's performance. Meanwhile, a new data-driven method, the mechanism model and data-driven stationary generalized feature learning system (StaBFLS), is proposed to estimate the stationary feature representation (StaFR) within the autoencoder framework. How to combine the steady-state loss and the reconstruction error loss is a key issue in StaBFLS modeling. Here, it is considered to convert the difference between the mean and variance of each epoch based on the Kullback-Leibler divergence into a quadratic form by adopting the second-order Taylor approximation (SOTA). Therefore, the statistical features of the extracted StaFR over time remain consistent, and this method is also consolidated by the theoretically derived optimal closed-form solution. The proposed time-matching generalized learning system (TMBLS) model can generate lag samples and incorporate a time-matching module to address this problem, which enables the identification of process variables closely related to MIQ and allows the prediction of the optimal MIQ value considering time lags to obtain more accurate MIQ modeling and fault detection.

[0010] The technical solution for the present invention to achieve its objectives is as follows:

[0011] A hierarchical strategy method for enhancing molten iron quality-related modeling and fault detection, the steps including: multi-phase multi-field mechanism modeling of the blast furnace ironmaking process (BFIP), data-driven stationary feature extraction, molten iron quality (MIQ) modeling based on the time-matching generalized learning system (TMBLS), and the design of the fault diagnosis process;

[0012] The method described improves the model transparency and MIQ prediction through a mechanism and data co-driven strategy (MDCDS); by dividing the furnace hearth and applying mechanism-based features to capture material and heat trends, and combining with the stationary generalized feature learning system (StaBFLS), it reduces the interference caused by non-stationary process features and mines the intrinsic information embedded in the blast furnace ironmaking process; subsequently, by combining the stationary feature representation with the mechanism features, the TMBLS system aligns the process and quality variables with MIQ as the target, establishes process monitoring statistics using mechanism and data-driven features, and detects modeling deviations.

[0013] The multi-phase multi-field mechanism modeling includes the following steps:

[0014] 2.1) Establish a mass balance model for the blast furnace ironmaking process, and determine the kinetic coefficients of different reactions in each region according to the current temperature T(t). This is achieved by using the Arrhenius formula, which provides the reaction rate constant. The relationship between the reaction rate constant and temperature change; assuming that all reactions of each reactant are first-order reactions, the range of each reaction within each region where the subscripts from 1 to 7 represent seven main reactions, can be deduced as follows:

[0015]

[0016] In the above formula, α is the apparent frequency factor, e represents the exponential function, E a is the Arrhenius activation energy constant, R g is the molar gas constant, to respectively represent the different reaction rate constants under 7 reactions, represents the number of moles of different substances, and different subscripts here represent different substances;

[0017] Therefore, the change in the amount of substance of each substance after the reaction is calculated where different subscripts here represent different substances:

[0018]

[0019] where L t is the reaction constant; through sufficient reaction, consider the load and gas content changes in each layer; for the burden, calculate the volume V u,z (t) after the reaction, and according to the molar volume obtain the volume change V a,z (t) and the proportion of each component in each region as:

[0020]

[0021] represents the specific number of moles of different substance j; then evaluate the inter-zone transportation of the load, including the molar loss of each substance from the current z zone to the regions below z + 1 as follows:

[0022]

[0023] To accurately calculate the number of moles of each substance at subsequent moments execute the following formula:

[0024]

[0025] The feed molar amount at the next moment is calculated as follows:

[0026]

[0027] Subsequently, for the gas behavior in the blast furnace, the bottom region receives the gas with a constant composition from the hot blast stove and conveys it to the upper region. In the middle, while receiving the gas from below and reacting, the gas is directly discharged from the top region;

[0028] Assume that the hot blast stove provides a constant gas composition of:

[0029]

[0030] In order to strictly simulate the gas behavior of each layer, according to the principle of mass balance, the molar changes of each layer are calculated Here, different subscripts g represent different gas components:

[0031]

[0032] 2.2) The thermal equilibrium modeling of BFIP includes 6 main aspects, namely the reaction heat The heat loss through the furnace wall ΔH w,z (t), the heat obtained by the hot blast stove ΔH d,z (t), the heat loss in the iron / slag ΔH p,z (t), is the slag discharge amount, A Fe (t) is the hot metal discharge amount, the heat changes of the ascending gas and the descending burden ΔH g,z (t), ΔH b,z (t), which are respectively expressed as:

[0033]

[0034] ΔH w,z (t) = -L t ·k Tw ·AW z ·(T z (t) - T e ) (10)

[0035]

[0036] Where represents the reaction heat at the standard temperature; are all expressed as the molar averages of substances with different subscripts; k Tw represents the furnace wall heat loss coefficient; AW z represents the furnace wall area; T eDenoted as the outer wall temperature of the blast furnace; H v Denotes the heat change rate; Molar mass of air; T hs (t) is the hot air temperature; Denotes CaSio 3 Concentration; A Fe (t) denotes the Fe concentration; Integrating the above 6 parts, the total heat H in each region z (t) is expressed as:

[0037]

[0038] Calculate the moles of each region And the average specific heat capacity Determine the temperature change ΔT z (t) is:

[0039]

[0040] Finally, the temperature of each region at subsequent moments is identified as:

[0041] T z (t + 1) = T z (t) + ΔT z (t)(17); This coupled thermochemical evolution continues until the mechanism performance error (MPE) with respect to the output model stops decreasing significantly in 10 iterations.

[0042] The data-driven stationary feature extraction described above includes the following steps:

[0043] 3.1) For StaBFLS, when the provided process data X is used as input, the relevant feature nodes and enhancement nodes are obtained

[0044]

[0045] where ξ q 、ξ l are activation functions, is the weight, is the bias term;

[0046] Then, by means of linear connection, and are combined as:

[0047]

[0048] When the extracted features are divided into 8 time periods, their mean and variance (μ e , Σ e ) and the overall mean and variance (Σe , ) match;

[0049] To quantify the difference, the Kullback-Leibler divergence is utilized and a loss function is defined as follows:

[0050]

[0051] where W represents the parameter matrix to be optimized; the optimization problem is formulated as a quadratic objective:

[0052]

[0053] By adding an unconstrained logarithmic term, equation (20) is rewritten as:

[0054]

[0055] When equation (23) reaches its minimum value at the optimal parameter matrix , the first derivative vanishes and the optimization can be carried out using the second derivative; using the notation [·] d,f as the (d,f)-th element of the matrix, ([·] d ,[·] f ) as the (d,f)-th row vector, Θ e = logdetW T ∑ e W, the second derivative is derived as:

[0056]

[0057] Using the following derivation:

[0058]

[0059] The derivation of is similar to equation (25);

[0060] Assume satisfies then it is derived that the relationship of is:

[0061]

[0062] According to the Cauchy-Schwarz inequality and the obtained from the lower bound of equation (26) holds:

[0063]

[0064] Finally, by substituting the lower bound, the complete quadratic objective is obtained, i.e., substituting the logarithmic term in Equation (23);

[0065] 3.2) The extracted data-driven features can minimize the prediction under supervision and the difference between the original objective X, the self-supervised loss with regularization terms is expressed as:

[0066]

[0067] A complete optimization objective, i.e., the total loss is expressed as:

[0068]

[0069] where the weights λ 1 , λ 2 ∈(0, 1); Taking the partial derivative with respect to gives the following result:

[0070]

[0071] By setting to 0, the optimal parameter matrix can be derived as:

[0072]

[0073] After that, StaFR is obtained by:

[0074] The TMBLS-based MIQ modeling described above includes the following steps:

[0075] The matrix of lag features and current features and the combination term is represented as:

[0076]

[0077] with a time-matching width s;

[0078] This matching uses a Euclidean distance-based judgment criterion to optimally align the current MIQ with the relevant lag information:

[0079]

[0080] where a matching threshold λ is introduced to filter out unimportant connections;

[0081] Using all candidate features, an information fusion connector is constructed to form a comprehensive feature, i.e.:

[0082]

[0083] Subsequently, during the training process, the fused features are used to replace X, generating a fused mapping and enhanced nodes and enter the prediction task, and the prediction ability for MIQ is achieved through as:

[0084]

[0085] where is the projection matrix of TMBLS, and λ 3 is the regularization weight parameter to avoid numerical errors;

[0086] The prediction deviation is calculated as:

[0087]

[0088] In the theoretical case where λ 3 is zero, the inverse problem is transformed into a least-squares problem, and the original pseudo-inverse is obtained:

[0089]

[0090] At the other end of the spectrum, when λ 3 tends to infinity, the solution of Equation (36) is extremely constrained and thus tends to zero, and we can obtain:

[0091]

[0092] For the regularization parameter λ 1 in Equation (29), the explanation of the function is the same as that of λ 3 .

[0093] The design of the fault diagnosis process described above includes the following steps:

[0094] 5.1) Construct three T 2 monitoring statistics, namely and which are respectively for mechanism features, StaFR, and MIQ prediction deviation:

[0095] Mechanism features:

[0096] Data-driven StaFR:

[0097] MIQ prediction deviation: where k m , k d and k y respectively represent and the dimension of ε; each statistic determines a single threshold according to its dimension and the significance level ζ selected by the χ 2 distribution, using:

[0098]

[0099] The final fault detection logic is as follows:

[0100]

[0101] where "∩" represents the logical operator "and".

[0102] Advantages of the present invention:

[0103] The present invention solves the limitations of the data-driven model in BFIP by proposing a hierarchical MDCDS strategy. By partitioning the blast furnace according to the principles of mass and heat balance, the model here utilizes mechanism characteristics to capture the trends of materials and temperature. To mitigate non-stationary interference and extract inherent information, data-driven StaBFLS is used here for StaFR extraction. On this basis, with MIQ as the goal, StaFR is combined with mechanism characteristics, and a comprehensive TMBLS model is formed using the time series matching module. Generally speaking, by bridging the gap between data-driven methods and mechanism understanding, the proposed MDCDS method opens up a new approach to the co-driven modeling and monitoring technology for complex BFIP. Brief Description of the Drawings

[0104] Figure 1 is a schematic diagram of the multi-phase and multi-field mechanism modeling of the blast furnace ironmaking process of the present invention.

[0105] Figure 2 is a schematic diagram of the hierarchical strategy method for enhancing the modeling and fault detection related to hot metal quality of the present invention.

[0106] Figure 3 is a test result diagram of the mechanism model for the blast furnace ironmaking process of the present invention;

[0107] where each part: (a) temperature changes in each region during SIA; (b) changes in the top feed composition, including C, Fe 2 O 3 , CaCO 3 and SiO 2 ; (c) showing the prediction results of the mechanism model using the training set ; (d) the predicted highest temperature and the actual highest temperature and the deviation between them.

[0108] Figure 4 is a schematic diagram of the data-driven stationary feature extraction of the present invention;

[0109] Each part: (a) Four-dimensional features of BLS-AE; (b) Four-dimensional features of the proposed stationary generalized feature learning system; (c) Probability density map between BLS-AE features; (d) Probability density map between features of the proposed stationary generalized feature learning system.

[0110] Figure 5 This is a comparison chart of the prediction performance of the present invention for the molten iron quality in the blast furnace ironmaking process;

[0111] Each part: (a) KPLS; (b) IKOPLS; (c) BLS; (d) MDCDS-D; (e) MDCDS-M; (f) MDCDS (the present invention). Figure 6 This is a comparison chart of the root mean square error of the model for the prediction of molten iron quality in the blast furnace ironmaking process of the present invention;

[0112] Each part: (a) Box plot of RMSE predicted by MIQ; (b) Average value and standard deviation recorded by the comparison model.

[0113] Figure 7 This is for the blast furnace ironmaking process of the present invention Fault detection chart of the case;

[0114] Each part: (a) KPLS; (b) IKOPLS; (c) BLS; (d) MDCDS-D; (e) MDCDS-M; (f1-f3) MDCDS (of the present invention). Specific implementation manner

[0115] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0116] As Figure 1 shown, it is a process diagram of multi-phase and multi-field mechanism modeling for the blast furnace ironmaking process of the present invention, including the following steps:

[0117] 2.1) Establish a mass balance model for the blast furnace ironmaking process, and determine the kinetic coefficients of different reactions in each region according to the current temperature T(t) This is achieved by using the Arrhenius formula, which provides the reaction rate constant The relationship with temperature change. Assuming that all reactions of each reactant are first-order reactions, the range of each reaction in each region Can be deduced as:

[0118]

[0119] And in the above formula, Indicates the moles of different substances, which will be given below. Therefore, the change in the amount of substance after the reaction of each substance can be calculated

[0120]

[0121] where L t is the reaction constant. It is crucial to fully consider the load and gas content changes in each layer through sufficient reaction.

[0122] For the batching, by calculating the volume V u,z (t) after reaction, based on the molar volume the volume change v a,z (t) and the proportion of each component in each region can be obtained as follows:

[0123]

[0124] Evaluate the inter-zone transport of the load, including the molar loss of each substance from the current z-region to the region below z + 1 as follows:

[0125]

[0126] To accurately calculate the molar of each substance at subsequent moments, the following formula can be executed:

[0127]

[0128] Therefore, the feed molar amount at the next moment is calculated as:

[0129]

[0130] Subsequently, for the gas behavior in the blast furnace, the bottom region receives gas with a constant composition from the hot blast stove and transports it to the upper region. In the middle, while receiving the gas from below and reacting, the gas is directly discharged from the top region.

[0131] Here, it is assumed that the hot stove provides a constant gas composition of:

[0132]

[0133] To strictly simulate the gas behavior in each layer, based on the principle of mass balance, calculate the molar change in each layer

[0134]

[0135] 2.2) The thermal equilibrium modeling of BFIP includes 6 main aspects, namely the reaction heat the heat loss through the furnace wall ΔH w,z (t), the heat obtained by the hot stove ΔHd,z (t), the heat loss ΔH in iron / slag p,z (t), is the slag discharge amount, A Fe (t) is the hot metal discharge amount, and the heat changes ΔH of the ascending gas and descending burden g,z (t), ΔH b,z (t) are respectively expressed as:

[0136]

[0137]

[0138] Integrating the above 6 parts, the total heat H in each region z (t) can be expressed as:

[0139]

[0140] Calculate the moles in each region and the average specific heat capacity to determine the temperature change ΔT z (t) is:

[0141]

[0142] Finally, the temperatures of each region at subsequent moments can be identified as:

[0143] T z (t + 1) = T z (t) + ΔT z (t)(17).

[0144] This coupled thermochemical evolution continues until the mechanism performance error (MPE) with respect to the output model stops decreasing significantly in 10 iterations.

[0145] As Figure 2 , Figure 3 , Figure 4 shown, it is a case study comparison of data-driven stationary feature extraction for the blast furnace ironmaking process of the present invention, including the following steps:

[0146] 3.1) For StaBFLS, when the provided process data X is used as the input, relevant feature nodes and enhancement nodes can be obtained

[0147]

[0148] where ξ q , ξ l are activation functions, is the weight, is the bias term.

[0149] Then, by means of linear connection, and can be combined as:

[0150]

[0151] When the extracted features are divided into 8 time periods, their mean and variance (μ e , Σ e ) match the population mean and variance (Σ e , ).

[0152] To quantify the difference, the Kullback-Leibler divergence is utilized and the loss function is defined as follows:

[0153]

[0154] where W represents the parameter matrix to be optimized.

[0155] The optimization problem is reformulated as a quadratic objective:

[0156]

[0157] By adding an unconstrained logarithmic term, Eq. (20) can be rewritten as:

[0158]

[0159] When Eq. (23) reaches the minimum value (close to 0) at the optimal parameter matrix , its first derivative vanishes, and the optimization can be performed using SOTA relying only on the second derivative.

[0160] Using the symbol [·] d,f as the (d, f)-th element of the matrix, ([·] d , [·] f ) as the (d, f)-th row vector, Θ e = logdetW T ∑ e W, the second derivative can be derived as:

[0161]

[0162] Using the following derivation:

[0163]

[0164] The derivation of

[0165] Suppose satisfies then it is deduced that The relationship of:

[0166]

[0167] According to the Cauchy - Schwarz inequality and from obtained The lower bound of Equation (26) holds:

[0168]

[0169] Finally, by substituting the lower bound mentioned here, the complete quadratic objective can be obtained, that is, substituting the logarithmic term in Equation (23).

[0170] 3.2) The data - driven features extracted should be able to minimize the difference between the prediction under supervision and the original objective X, and the self - supervised loss with a regularization term can be expressed as:

[0171]

[0172] A complete optimization objective, that is, the total loss can be expressed as:

[0173]

[0174] where the weights λ 1 、λ 2 ∈(0, 1).

[0175] For taking the partial derivative, the following results can be obtained:

[0176]

[0177] By setting to 0, then the optimal parameter matrix can be derived as:

[0178]

[0179] After that, StaFR can be obtained through:

[0180]

[0181] Such as Figure 5 、 Figure 6As shown, it is a case study comparison of MIQ modeling based on TMBLS for the blast furnace ironmaking process of the present invention, including the following steps:

[0182] The matrix representation of the lag features and the current features and the combined terms is:

[0183]

[0184] with a time matching width s. This matching uses a judgment criterion based on the Euclidean distance to optimally align the current MIQ with the relevant lag information:

[0185]

[0186] where a matching threshold λ is introduced to filter out unimportant connections.

[0187] Using all candidate features, an information fusion connector is constructed to form a comprehensive feature, that is:

[0188]

[0189] Subsequently, during the training process, the fused feature is used to replace X to generate the fused mapping and enhanced nodes and

[0190] Entering the prediction task, the prediction ability for MIQ is achieved through as:

[0191]

[0192] where is the projection matrix of TMBLS, and λ 3 is the regularization weight parameter to avoid numerical errors.

[0193] The prediction deviation can be calculated as:

[0194]

[0195] In the theoretical case where λ 3 is zero, the inverse problem is transformed into a least squares problem to obtain the original pseudo-inverse:

[0196]

[0197] At the other end of the spectrum, when λ 3 tends to infinity, the solution is extremely constrained and tends to zero:

[0198]

[0199] For the regularization parameter λ in Equation (29) 1 The description of the function is similar to that of λ 3 .

[0200] As Figure 7 shown, it is a case study comparison of the design of the fault diagnosis process for the blast furnace ironmaking process of the present invention, including the following steps:

[0201] Three T 2 monitoring statistics can be constructed, which are for mechanism characteristics, StaFR, and MIQ prediction deviation respectively:

[0202] Mechanism characteristics:

[0203] Data-driven StaFR:

[0204] MIQ prediction deviation: where k m , k d and k y represent respectively and the dimensions of ε.

[0205] Each statistic determines a single threshold according to its dimension and the significance level ζ selected according to the χ 2 distribution, and can be expressed as:

[0206]

[0207] The final fault detection logic is as follows:

[0208]

[0209] where "∩" represents the logical operator "and".

[0210] The solution here involves a two-layer architecture. The bottom layer contains independent data-driven models and mechanism models, while the top layer contains data and mechanism characteristics to drive the collaborative model. By describing the BFIP operation mechanism and domain knowledge, rich information is provided. On the other hand, the data-driven model explores non-linear and non-stationary data to extract potential StaFR. These two models coordinate and cooperate to align the information obtained with the model objectives through time matching, and finally obtain a more accurate MIQ model.

[0211] The present invention proposes a mechanism and data co-driven strategy (MDCDS) to enhance model transparency and hot metal quality (MIQ) prediction. First, by partitioning the furnace body and applying mechanism-based material and thermal trend features, together with a new stationary generalized feature learning system (StaBFLS), the interference caused by non-stationary process features is mitigated, and the inherent information embedded in the BFIP is mined. Subsequently, by combining the stationary feature representation with the mechanism features, the proposed time-matching generalized learning system (TMBLS) aligns the process variables and quality variables with the MIQ as the target. This integration allows the establishment of process monitoring statistics using mechanism and data-driven characteristics, as well as the detection of modeling deviations. Through the verification of actual BFIP data, the proposed MDCDS model demonstrates consistent process alignment, robust feature extraction, and improved MIQ modeling, resulting in better fault detection.

[0212] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.

Claims

1. A hierarchical strategy method for enhancing hot metal quality related modeling and fault detection, characterized in that the steps include: Multiphase and multifield mechanism modeling of blast furnace ironmaking process (BFIP), data-driven stationary feature extraction, molten iron quality (MIQ) modeling based on time-matching generalized learning system (TMBLS), and design of fault diagnosis process; The described approach improves model transparency and MIQ predictions through a mechanism and data collaboratively driven strategy (MDCDS); By dividing the furnace and applying mechanism-based features to capture material and thermal trends, combined with the stationary generalized feature learning system (StaBFLS), the interference caused by non-stationary process features is mitigated and the intrinsic information embedded in the blast furnace ironmaking process is mined; subsequently, by combining the stationary feature representation with the mechanism features, the TMBLS system aligns process and quality variables with MIQ as the goal, establishes process monitoring statistics using mechanism and data-driven features, and detects modeling deviations.

2. The method according to claim 1, characterized in that The multi-phase multi-field mechanism modeling includes the following steps: 2.1) Establish a mass balance model for the blast furnace ironmaking process and determine the kinetic coefficients of different reactions in each area based on the current temperature T(t) This is accomplished by using the Arrhenius formula, which provides the reaction rate constant and temperature change; Assuming that all reactions of each reactant are first-order reactions, the range of each reaction in each region The subscripts 1 to 7 represent seven main reactions, which can be deduced as follows: In the above formula, α is the apparent frequency factor, e represents the exponential function, and E a is the Arrhenius activation energy constant, R g is the molar gas constant, arrive Respectively Different reaction rate constants under 7 reactions, Indicates the number of moles of different substances, where different subscripts represent different substances; Therefore, calculate the change in the amount of each substance after the reaction Here different subscripts represent different substances: Where L t is the reaction constant; through sufficient reaction, the load and gas content of each layer are considered; for the batching, the volume V after reaction is calculated u,z (t), according to the molar volume Find the volume change V a,z (t) and the ratio of each component In each area: represents the specific number of moles of different substances j; then the inter-zone transport of the load is evaluated, including the molar loss of each substance from the current z zone to the zone below z+1 As shown below: In order to accurately calculate the molar V of each substance at a later time j z (t+1), execute the following formula: The molar amount of feed at the next moment is calculated as: Then, for the gas behavior in the blast furnace, the bottom zone receives the gas of constant composition from the hot blast stove and conveys it to the upper zone, while in the middle, the gas is received from below and reacts while the gas is discharged directly from the top zone; Assume that the hot furnace provides a constant gas composition of: In order to strictly simulate the gas behavior of each layer, the molar changes of each layer are calculated according to the mass balance principle. Here different subscripts g represent different gas compositions: 2.2) The thermal balance modeling of BFIP includes six main aspects, namely, reaction heat Heat loss from furnace wall ΔH w,z (t), the heat gained by the hot furnace is ΔH d,z (t), heat loss in iron / slag ΔH p,z (t), is the slag discharge, A Fe (t) is the discharge of molten iron, the heat change of rising gas and descending charge ΔH g,z (t), ΔH b,z (t), respectively expressed as: ΔH w,z (t)=-L t ·k Tw ·AW z ·(T z (t)-T e ) (10) in It represents the heat of reaction at standard temperature; All are expressed as the molar mean of substances with different subscripts; k Tw Indicates the furnace wall heat loss coefficient; AW z Indicates the furnace wall area; T e Expressed as the blast furnace outer wall temperature; H v represents the rate of thermal change; The molar mass of air; T hs (t) is the hot air temperature; Indicates CaSio3 concentration; A Fe (t) represents the Fe concentration; integrating the above 6 parts, the total heat H in each area z (t) is expressed as: Calculate the moles per area and average specific heat Determine the temperature change ΔT z (t) is: Finally, the temperature of each area at the subsequent time is identified as: T z (t+1)=T z (t)+ΔT z (t)(17); This coupled thermochemical evolution continues until the mechanistic performance error (MPE) relative to the output model stops decreasing significantly within 10 iterations.

3. The method according to claim 2, characterized in that The data-driven stationary feature extraction includes the following steps: 3.1) For StaBFLS, when the process data X is provided as input, the relevant feature nodes and enhanced nodes are obtained Among them, ξ q , l is the activation function, is the weight, is the bias term; Then with the help of linear connection, and The combination is: When the extracted features are divided into 8 time periods, their mean and variance (μ e ,Σ e ) and the population mean and variance Match; To quantify the difference, we use the Kullback-Leibler divergence and define the loss function as follows: Where W represents the parameter matrix to be optimized; the optimization problem is formulated as a quadratic objective: By adding the unconstrained logarithmic term, equation (20) can be rewritten as: When equation (23) is in the optimal parameter matrix When the minimum is reached, the first-order derivative vanishes and the optimization can be performed using the second-order derivative; using the symbol [·] d,f , as the (d,f)th element of the matrix, ([·] d ,[·] f ) as the (d,f)th row vector, Θ e =logdetW T ∑ e W, The second-order derivative is derived as: Use the following derivation: The derivation of is similar to formula (25); Assumptions satisfy Then it is deduced The relationship is: According to the Cauchy-Schwarz inequality and from What you get The lower bound of formula (26) holds: Finally, by substituting the lower bound, we get the complete quadratic objective, which is the logarithmic term in (23); 3.2) The extracted data-driven features can minimize the supervised prediction The difference between the original target X and the self-supervised loss with regularization It is expressed as: A complete optimization objective, namely the total loss It is expressed as: The weights λ1, λ2∈(0,1); Taking partial derivatives, we get the following results: By setting is 0, The optimal parameter matrix Can be exported as: After that, StaFR was approved:

4. The method according to claim 3, characterized in that The MIQ modeling based on TMBLS includes the following steps: The lagged feature and the current feature With combination item The matrix representation of is: With time matching width s; The matching uses a criterion based on Euclidean distance to optimally align the current MIQ with the relevant lag information: The matching threshold λ is introduced to filter out unimportant connections; Using all candidate features, an information fusion connector is constructed to form a comprehensive feature, namely: Then, the fused features are used during training Replace X, generate fusion map and enhancement node and Enter the prediction task through The prediction capability of MIQ is achieved as follows: In the formula is the projection matrix of TMBLS, λ3 is the regularization weight parameter to avoid numerical errors; The prediction bias is calculated as: In the theoretical case where λ3 is zero, the inverse problem is transformed into a least squares problem, and the original pseudo-inverse is obtained: At the other end of the spectrum, when λ3 tends to infinity, the solution of equation (36) is greatly constrained and tends to zero, and we can obtain: The explanation of the regularization parameter λ1 function in equation (29) is similar to that of λ3.

5. The method according to claim 4, characterized in that The design of the fault diagnosis process includes the following steps: 5.1) Constructing three Ts 2 Monitoring statistics, i.e. and They are for mechanism characteristics, StaFR and MIQ prediction deviations: Mechanism characteristics: Data drives StaFR: MIQ prediction bias: where k m , k d and k y Respectively and ε; each statistic is calculated according to its dimension and χ 2 The selected significance level ζ of the distribution is used to determine a single threshold, using: The final fault detection logic is as follows: Where "∩" represents the logical operator "and".

Images