A hierarchical strategy approach for enhancing molten iron quality-related modeling and fault detection.

Abstract

This invention discloses a hierarchical strategy method for enhancing hot metal quality-related modeling and fault detection. The steps include: multiphase and multi-field mechanism modeling of the blast furnace ironmaking process, data-driven stationary feature extraction, hot metal quality (MIQ) modeling based on a time-matched generalized learning system, and the design of a fault diagnosis process. A mechanism- and data-driven collaborative strategy is designed to improve model transparency and MIQ prediction. By partitioning the furnace and applying mechanism-based features to capture material and thermal trends, and combining a novel stationary generalized feature learning system, interference caused by non-stationary process features is mitigated, and intrinsic information embedded in the blast furnace ironmaking process is mined. Subsequently, by combining stationary feature representations with mechanism features, process and quality variables are aligned with MIQ as the target. Validation with real data shows consistent process alignment, powerful feature extraction, and improved MIQ modeling, thereby achieving better fault detection.

Description

Technical Field

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

[0002] The Blast Furnace Ironmaking Process (BFIP) is considered an essential process, accounting for 70% of total energy consumption and significantly impacting subsequent steelmaking and rolling. For the BFIP product, molten iron, trace elements such as silicon, phosphorus, and sulfur play a crucial role in assessing the overall molten iron quality (MIQ). Unlike other industrial production processes, the blast furnace itself is a highly complex, large-scale system with intricate mechanistic models. Furthermore, assessing molten iron quality for MIQ testing involves a series of steps from discharge to sampling and subsequent laboratory analysis, taking approximately 30 minutes. Considering the approximately 30-minute BFIP cycle time, the inherent delay in MIQ assessment presents a significant challenge for field engineers aiming for real-time and precise adjustments. Additionally, BFIP is susceptible to various potential MIQ-related faults; neglecting timely monitoring of these issues can severely impact the entire production process, potentially escalating into serious production accidents in extreme cases. Furthermore, process variables in blast furnace ironmaking operations are affected by non-stationary factors, which introduces complexity and significantly hinders the rapid and reliable detection of quality-related faults. Therefore, establishing an effective monitoring system to promptly detect and address related faults is crucial to ensuring the smooth and safe operation of BFIP (Breakfast Factory Injection Process).

[0003] Currently, MIQ modeling and process monitoring mainly rely on data-driven techniques. Existing studies have used methods such as least squares support vector regression (LS-SVR) and graph neural networks (GNN) to predict MIQ. However, most of these methods only focus on a single MIQ parameter, neglecting to consider multiple parameters simultaneously. To overcome these limitations, Zhou et al. proposed a modeling method based on stochastic vector function linking networks (RVFLN), which considers multiple MIQ parameters and overcomes challenges related to 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) based on statistical analysis and improved kernel orthogonal projection to latent structure (IKOPLS) to achieve simultaneous multi-MIQ modeling and process monitoring. Although progress has been made, some challenges remain to be addressed.

[0004] The complex characteristics of BFIP introduce variability and fluctuations, posing significant challenges to accurately identifying potential intrinsic 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, leading to patterns that change over time.

[0005] While data-driven modeling approaches can provide accurate predictions, they often lack interpretability. The black-box nature of these models makes it challenging to discern the intrinsic relationship between process variables and MIQ. Interpretable models provide insights into how various factors influence MIQ, enabling operators and engineers to gain a clear understanding of the process. This greater transparency allows them to make informed decisions by analyzing the model's reasoning.

[0006] Considering 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; this delay can be caused by various factors in the measurement workflow, such as sampling transport time and the duration of laboratory testing.

[0007] Mechanistic modeling is based on fundamental theoretical knowledge, incorporating principles such as chemical reactions and mass / heat conservation, and understanding the evolutionary characteristics of processes. Early research by Ridgion et al. focused on materials and heat balance equations to understand the differences in chemical reactions occurring in different temperature zones within a blast furnace, thus gaining a deeper understanding of thermochemical properties. Subsequent work further enhanced BFIP modeling by introducing additional equations influencing factors such as heat transfer, gas-solid reactions, and component behavior. Furthermore, recent studies have limited mechanistic modeling to local furnace zones, neglecting a holistic understanding. However, it must be recognized that existing BFIP modeling cannot fully reflect the complexity of processes. In recent years, synergistic approaches combining mechanistic modeling and data-driven techniques have emerged. By combining the advantages of data-driven modeling with the insights of mechanistic models, co-driven models can leverage the accuracy of data-driven models while gaining perspectives from process understanding, thereby promoting improvements in tasks such as MIQ modeling and process monitoring. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a hierarchical strategy method for enhancing hot metal quality-related modeling and fault detection.

[0009] This invention proposes a Mechanism-Data Co-Driven Strategy (MDCDS) to address the challenges posed by the complexity of non-stationary characteristics, intricate mechanisms, and mismatch between MIQ and process samples. The BFIP mechanistic model here divides the blast furnace into different reaction zones, allowing for in-depth analysis of reactions at different depths. By combining major thermochemical reactions, conservation laws, and region-specific external interactions, the model achieves both interpretability and robustness, naturally deriving a large number of unmeasured mechanistic features. Furthermore, a Sequential Iterative Algorithm (SIA) is proposed to further refine the model parameters, iteratively improving model performance. A novel data-driven approach, the mechanistic model and a data-driven stationary generalized feature learning system (StaBFLS), is also proposed for estimating stationary feature representations (StaFR) within the autoencoder framework. A key issue in StaBFLS modeling is how to combine steady-state loss with reconstruction error loss. Here, we consider transforming the difference between the mean and variance of the Kullback-Leibler divergence for each period into a quadratic form by employing a second-order Taylor approximation (SOTA). Therefore, the extracted StaFR statistical characteristics remain consistent over time, and this approach is further reinforced by a theoretically derived optimal closed-form solution. This problem is addressed by generating lag samples and incorporating time-matching modules through the proposed Temporal Matching Generalized Learning System (TMBLS) model. This enables the identification of process variables closely related to MIQ and allows for the prediction of optimal MIQ values ​​that take time lags into account, resulting in more accurate MIQ modeling and fault detection.

[0010] The technical solution for achieving the objective of this invention is as follows:

[0011] A hierarchical strategy method for enhancing hot metal quality-related modeling and fault detection includes the following steps: multiphase and multifield mechanism modeling of blast furnace ironmaking process (BFIP), data-driven stationary feature extraction, hot metal quality (MIQ) modeling based on time-matched generalized learning system (TMBLS), and design of fault diagnosis process.

[0012] The method described improves model transparency and MIQ prediction through a mechanism and data-driven strategy (MDCDS). By partitioning the furnace and applying mechanism-based features to capture material and thermal trends, and combining a stationary generalized feature learning system (StaBFLS), it mitigates the interference caused by non-stationary process features and mines the intrinsic information embedded in the blast furnace ironmaking process. Subsequently, by combining stationary feature representations with mechanism features, the TMBLS system aligns process and quality variables with MIQ as the target, uses mechanism and data-driven features to build process monitoring statistics, and detects modeling biases.

[0013] The multiphase, 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 based on the current temperature T(t). This is achieved using the Arrhenius equation, which provides the reaction rate constant. The relationship with temperature change; assuming all reactions of each reactant are first-order reactions, then the range of each reaction in each region. The subscripts 1 to 7 represent the seven main reactions, which can be deduced as follows:

[0015] In the above formula, α is the apparent frequency factor, e represents the exponential function, and E a R is the Arrhenius activation energy constant. g The molar gas constant, arrive They represent Different reaction rate constants for 7 reactions This indicates the number of moles of different substances; different subscripts here indicate different substances.

[0016] Therefore, the change in the amount of substance after each reaction can be calculated. Different subscripts here represent different substances:

[0017]

[0018] Where L t The reaction constant is used; through complete reaction, the load and gas content changes of each layer are considered; for batching, the volume V after reaction is calculated. u,z (t), based on molar volume Calculate the volume change V a,z (t) and the proportions of each component In each region:

[0019] This represents the specific molar number of different substances j; then, the inter-band transport of the load is evaluated, including the molar loss of each substance from the current z region to regions below z+1. As shown below:

[0020] To accurately calculate the moles of each substance at subsequent time points Execute the following formula:

[0021]

[0022] The feed molar amount at the next moment is calculated as follows:

[0023]

[0024] Subsequently, regarding the gas behavior within the blast furnace, the bottom region receives gas of constant composition from the hot blast stove and transports it to the upper region. In the middle, while receiving gas from below and reacting, the gas is directly discharged from the top region.

[0025] Assume the furnace provides a constant gas composition:

[0026] To rigorously simulate the gas behavior of each layer, the molar changes of each layer were calculated based on the principle of mass balance. Different subscripts 'g' here represent different gas components:

[0027]

[0028] 2.2) The thermal equilibrium modeling of BFIP includes six main aspects, namely the heat of reaction. Heat loss ΔH from furnace wall w,z (t), the heat obtained by the furnace ΔH d,z (t), heat loss ΔH from iron / slag p,z (t), A represents the slag discharge volume. Fe (t) represents the hot metal discharge rate, and the thermal changes ΔH of rising gas and falling furnace charge. g,z (t), ΔH b,z (t), respectively, are represented as:

[0029]

[0030] ΔH w,z (t)=-L t ·k Tw ·AW z ·(T z (t)-T e (10)

[0031]

[0032]

[0033] in Indicates the heat of reaction at standard temperature; All represent the molar mean of substances with different subscripts; k Tw Indicates the furnace wall heat loss coefficient; AW z T represents the furnace wall area; e This is expressed as the temperature of the blast furnace outer wall; H v Indicates the rate of thermal change; Molar mass of air; Ths (t) represents the temperature of the hot air; Indicates CaSiO3 concentration; A Fe (t) represents the Fe concentration; integrating the above 6 parts, the total heat H in each region is... z (t) is represented as:

[0034]

[0035] Calculate the moles in each region and average specific heat capacity Determine the temperature change ΔT z (t) is:

[0036]

[0037] Ultimately, the temperatures of each region at subsequent times were identified as follows:

[0038] T z (t+1)=T z (t)+ΔT z (t) (17);

[0039] This coupled thermochemical evolution continued until the mechanistic performance error (MPE) relative to the output model stopped decreasing significantly after 10 iterations.

[0040] The data-driven stationary feature extraction includes the following steps:

[0041] 3.1) For StaBFLS, when the provided process data X is used as input, the relevant feature nodes and enhancement nodes are obtained.

[0042]

[0043] Where, ξ q ξ l For activation function, As weight, For bias terms;

[0044] Then, with the help of linear connections, and The combination is:

[0045]

[0046] When the extracted features are divided into 8 time periods, their mean and variance (μ) e ,Σ e ) with population mean and variance Matching;

[0047] To quantify the differences, the Kullback-Leibler divergence was used and a loss function was defined. as follows:

[0048]

[0049] Where W represents the parameter matrix to be optimized; the optimization problem is formulated as a quadratic objective:

[0050]

[0051] By adding an unconstrained logarithmic term, equation (20) is rewritten as:

[0052]

[0053] When equation (23) is in the optimal parameter matrix When the minimum value is reached, the first derivative disappears, and optimization can be performed using the second derivative; the notation [·] can be used. d,f As the (d,f)th element of the matrix, ([·] d ,[·] f ) is used as the (d,f)th row vector. The second derivative is derived as follows:

[0054]

[0055] Use the following derivation:

[0056] The derivation of the similarity formula (25);

[0057] Assumption satisfy Then it can be deduced that The relationship is:

[0058] According to the Cauchy-Schwarz inequality and from Received The lower bound of equation (26) holds:

[0059] Finally, by substituting the lower bound, we obtain the complete quadratic objective, which is the logarithmic term in equation (23);

[0060] 3.2) The extracted data-driven features can minimize the prediction under supervised supervision. The difference between the target X and the original target X, and the self-supervised loss with regularization. Represented as:

[0061]

[0062] A complete optimization objective is the total loss. Represented as:

[0063]

[0064] Where the weights λ1, λ2 ∈ (0, 1); for Taking the partial derivatives, we obtain the following results:

[0065]

[0066] By setting =0, Then the optimal parameter matrix It can be exported as:

[0067]

[0068] Subsequently, StaFR was approved:

[0069] The TMBLS-based MIQ modeling includes the following steps:

[0070] Hysteresis features and current features Combined items The matrix representation is as follows:

[0071] It has a time matching width of s;

[0072] This matching uses a judgment criterion based on Euclidean distance to optimally align the current MIQ with relevant lag information:

[0073]

[0074] A matching threshold λ is introduced to filter out unimportant connections;

[0075] By utilizing all candidate features, an information fusion connector is constructed to form a comprehensive feature, namely:

[0076]

[0077] Subsequently, during training, fusion features were utilized. Replace X to generate merge maps and enhanced nodes. and Enter the prediction task, through The ability to predict MIQ is achieved as follows:

[0078]

[0079] In the formula λ is the projection matrix of TMBLS, and λ3 is the regularization weight parameter to avoid numerical errors;

[0080] The prediction deviation is calculated as follows:

[0081] In the theoretical case where λ3 is zero, the inverse problem is transformed into a least squares problem, yielding the original pseudo-inverse:

[0082] At the other end of the spectrum, when λ3 approaches infinity, the solution of equation (36) is subject to a maximum constraint and thus approaches zero, which yields:

[0083]

[0084] The explanation of the regularization parameter λ1 function in equation (29) is similar to that of λ3.

[0085] The design of the fault diagnosis process includes the following steps:

[0086] 5.1) Construct three Ts 2 Monitoring statistics, i.e. and These are the prediction biases for mechanistic characteristics, StaFR, and MIQ:

[0087] Mechanism characteristics:

[0088] Data-driven StaFR:

[0089] MIQ prediction bias: Where k m k d and k y They represent The dimension of ε; each statistic is determined by its dimension and χ². 2 The significance level ζ chosen by the distribution is used to determine a single threshold, using:

[0090]

[0091] The final fault detection logic is as follows:

[0092]

[0093] The symbol “∩” represents the logical operator “AND”.

[0094] The beneficial effects of this invention are:

[0095] This invention addresses the limitations of data-driven models in BFIP by proposing a hierarchical MDCDS strategy. By partitioning the blast furnace according to mass and thermal balance principles, the model utilizes mechanistic features to capture material and temperature trends. To mitigate non-stationary disturbances and extract intrinsic information, data-driven StaBFLS is employed for StaFR extraction. Based on this, with MIQ as the target, StaFR is combined with mechanistic features, utilizing a time-series matching module to form a comprehensive TMBLS model. In summary, by bridging the gap between data-driven methods and mechanistic understanding, the proposed MDCDS method opens up a new avenue for collaboratively driven modeling and monitoring techniques in complex BFIPs. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of the multiphase and multifield mechanism modeling of the blast furnace ironmaking process of the present invention.

[0097] Figure 2 This is a schematic diagram of the hierarchical strategy method for enhancing molten iron quality-related modeling and fault detection according to the present invention.

[0098] Figure 3 The figure shows the test results of the mechanism model for the blast furnace ironmaking process of this invention;

[0099] The components include: (a) temperature variations in different regions during SIA; (b) variations in the composition of the top feed, including C, Fe2O3, CaCO3, and SiO2; and (c) demonstrations using the training set. (d) The predicted results of the mechanistic model; the predicted maximum temperature and the actual maximum temperature and the deviation between them.

[0100] Figure 4 This is a schematic diagram of the data-driven stationary feature extraction method of the present invention;

[0101] The components are: (a) the four-dimensional features of BLS-AE; (b) the four-dimensional features of the proposed stationary generalized feature learning system; (c) the probability density plot between features of BLS-AE; and (d) the probability density plot between features of the proposed stationary generalized feature learning system.

[0102] Figure 5 This is a comparison chart showing the predictive performance of the present invention for the quality of molten iron in the blast furnace ironmaking process.

[0103] The components are: (a) KPLS; (b) IKOPLS; (c) BLS; (d) MDCDS-D; (e) MDCDS-M; (f) MDCDS (the present invention).

[0104] Figure 6This is a comparison chart of the root mean square error of the model for predicting the quality (Si, P, S) of molten iron in the blast furnace ironmaking process according to the present invention.

[0105] Figure 7 This invention is directed towards the blast furnace ironmaking process. Fault detection diagram of the case;

[0106] The components are: (a) KPLS; (b) IKOPLS; (c) BLS; (d) MDCDS-D; (e) MDCDS-M; (f1-f3) MDCDS (of the present invention). Detailed Implementation

[0107] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0108] like Figure 1 The diagram shown illustrates the multiphase, multi-field mechanism modeling process of the blast furnace ironmaking process according to the present invention, including the following steps:

[0109] 2.1) Establish a mass balance model for the blast furnace ironmaking process, and determine the kinetic coefficients of different reactions in each region based on the current temperature T(t). This is achieved using the Arrhenius equation, which provides the reaction rate constant. The relationship with temperature change. Assuming all reactions for each reactant are first-order reactions, the range of each reaction in each region. This can be deduced as:

[0110] In the above formula, The molar amounts of different substances are given below. Therefore, the change in the amount of each substance after the reaction can be calculated.

[0111]

[0112] Where L t This is the reaction constant. Achieving a full reaction while fully considering the load and gas content variations in each layer is crucial.

[0113] For the ingredients, calculate the volume V after the reaction. u,z (t), based on molar volume The volume change V can be calculated. a,z (t) and the proportions of each component In each region:

[0114]

[0115] The assessment of inter-band transport of loads includes materials from the current z-region to regions below z+1. The molar loss is as follows:

[0116] To accurately calculate the moles of each substance at subsequent time points The following formula can be executed:

[0117]

[0118] Therefore, the feed molar quantity at the next moment is calculated as follows:

[0119]

[0120] Subsequently, regarding the gas behavior within the blast furnace, the bottom region receives gas of constant composition from the hot blast stove and transports it to the upper region. In the middle, while receiving gas from below and reacting, the gas is discharged directly from the top region.

[0121] Here it is assumed that the furnace provides a constant gas composition:

[0122] To rigorously simulate the gas behavior of each layer, the molar changes of each layer were calculated based on the principle of mass balance.

[0123]

[0124] 2.2) The thermal equilibrium modeling of BFIP includes six main aspects, namely the heat of reaction. Heat loss ΔH from furnace wall w,z (t), the heat obtained by the furnace ΔH d,z (t), heat loss ΔH from iron / slag p,z (t), A represents the slag discharge volume. Fe (t) represents the hot metal discharge rate, and the thermal changes ΔH of rising gas and falling furnace charge. g,z (t), ΔH b,z (t), respectively, are represented as:

[0125]

[0126]

[0127] Integrating the above six parts, the total heat H in each region is... z (t) can be expressed as:

[0128]

[0129] Calculate the moles in each region and average specific heat capacity The temperature change ΔT can be determined. z (t) is:

[0130]

[0131] Finally, the temperatures of each region at subsequent times can be identified as follows:

[0132] T z (t+1)=T z (t)+ΔT z (t) (17).

[0133] This coupled thermochemical evolution continued until the mechanistic performance error (MPE) relative to the output model stopped decreasing significantly after 10 iterations.

[0134] like Figure 2 , Figure 3 , Figure 4 The figure shows a comparative case study of data-driven stationary feature extraction for the blast furnace ironmaking process according to the present invention, which includes the following steps:

[0135] 3.1) For StaBFLS, when the provided process data X is used as input, relevant feature nodes and enhancement nodes can be obtained.

[0136]

[0137] Where, ξ q ξ l For activation function, As weight, This is a bias term.

[0138] Then, with the help of linear connections, and It can be combined as:

[0139]

[0140] When the extracted features are divided into 8 time periods, their mean and variance (μ) e ,Σ e ) with population mean and variance Matching.

[0141] To quantify the differences, the Kullback-Leibler divergence was used and a loss function was defined. as follows:

[0142]

[0143] Where W represents the parameter matrix that needs to be optimized.

[0144] The optimization problem is reformulated as a quadratic objective:

[0145]

[0146] By adding an unconstrained logarithmic term, Eq.(20) can be rewritten as:

[0147]

[0148] When equation (23) is in the optimal parameter matrix When the value reaches its minimum (close to 0), its first derivative disappears, and optimization can be performed using state-of-the-art methods that rely solely on the second derivative.

[0149] Using the symbol [·] d,f As the (d,f)th element of the matrix, ([·] d ,[·] f Θ is the (d,f)th row vector. e =logdetW T ∑ e W, The second derivative can be derived as follows:

[0150]

[0151] Use the following derivation:

[0152]

[0153] The derivation is similar to that of equation (25).

[0154] Assumption satisfy Then it can be deduced that The relationship is:

[0155]

[0156] According to the Cauchy-Schwarz inequality and from Received The lower bound of equation (26) holds:

[0157]

[0158] Finally, by substituting the lower bound mentioned here, we can obtain the complete quadratic objective, which is the logarithmic term in equation (23).

[0159] 3.2) The extracted data-driven features should be able to minimize the prediction under supervised supervision. The difference between the target X and the original target X, and the self-supervised loss with regularization. It can be represented as:

[0160]

[0161] A complete optimization objective is the total loss. It can be represented as:

[0162]

[0163] The weights λ1 and λ2 are ∈ (0, 1).

[0164] right Taking the partial derivatives, we can obtain the following results:

[0165]

[0166] By setting =0, Then the optimal parameter matrix It can be exported as:

[0167]

[0168] StaFR can then be approved:

[0169]

[0170] like Figure 5 , Figure 6 The figure shows a comparative case study of MIQ modeling based on TMBLS for blast furnace ironmaking processes according to the present invention, which includes the following steps:

[0171] Hysteresis features and current features Combined items The matrix representation is as follows:

[0172] It has a time matching width s. This matching employs a judgment criterion based on Euclidean distance to optimally align the current MIQ with relevant lag information:

[0173]

[0174] A matching threshold λ is introduced to filter out unimportant connections.

[0175] By utilizing all candidate features, an information fusion connector is constructed to form a comprehensive feature, namely:

[0176]

[0177] Subsequently, during training, fusion features were utilized. Replace X to generate merge maps and enhanced nodes. and

[0178] Enter the prediction task, through The ability to predict MIQ is achieved as follows:

[0179]

[0180] In the formula λ is the projection matrix of TMBLS, and λ3 is the regularization weight parameter to avoid numerical errors.

[0181] Prediction bias can be calculated as follows:

[0182]

[0183] In the theoretical case where λ3 is zero, the inverse problem is transformed into a least squares problem, yielding the original pseudo-inverse:

[0184]

[0185] At the other end of the spectrum, as λ3 approaches infinity, the solution is subject to a maximal constraint and approaches zero:

[0186]

[0187] The description of the regularization parameter λ1 function in equation (29) is similar to that of λ3.

[0188] like Figure 7 The diagram shows a comparative design case study of the fault diagnosis process for the blast furnace ironmaking process according to the present invention, which includes the following steps:

[0189] Three Ts can be constructed 2 The monitoring statistics are for mechanistic characteristics, StaFR, and MIQ prediction bias, respectively:

[0190] Mechanism characteristics:

[0191] Data-driven StaFR:

[0192] MIQ prediction bias:

[0193] Where k m k d and k y They represent The dimension of ε.

[0194] Each statistic is determined by its dimension and χ².2 The significance level ζ chosen by the distribution is used to determine a single threshold, which can be achieved using:

[0195]

[0196] The final fault detection logic is as follows:

[0197]

[0198] The symbol “∩” represents the logical operator “AND”.

[0199] The proposed solution involves a two-tier architecture. The bottom layer contains independent data-driven and mechanistic models, while the top layer contains data and mechanistic characteristics to drive a collaborative model. Rich information is provided by describing the operational mechanisms of BFIP and domain knowledge. Meanwhile, the data-driven model explores nonlinear and non-stationary data to extract potential StaFRs. These two models coordinate and collaborate, aligning the acquired information with the model objectives through time matching, ultimately resulting in a more accurate MIQ model.

[0200] This invention proposes a Mechanism- and Data-Driven Strategy (MDCDS) to enhance model transparency and molten iron quality (MIQ) prediction. First, by partitioning the furnace body and applying mechanistic material and thermal trend features, coupled with a novel Stationary Generalized Feature Learning System (StaBFLS), interference from non-stationary process features is mitigated, and inherent information embedded in the molten iron in process inlet (BFIP) is mined. Subsequently, by combining stationary feature representations with mechanistic features, the proposed Time-Matched Generalized Learning System (TMBLS) aligns process and quality variables with MIQ as the objective. This integration allows for the establishment of process monitoring statistics using mechanistic and data-driven characteristics, as well as the detection of modeling biases. Validation on actual BFIP data demonstrates consistent process alignment, robust feature extraction, and improved MIQ modeling, resulting in better fault detection.

[0201] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A hierarchical strategy approach for enhancing the molten iron quality related modeling and fault detection characterized by the steps of The application relates to a multi-phase multi-field mechanism modeling method for a blast furnace iron-making process, a stationary feature extraction method driven by data, a molten iron quality modeling method based on a time-matching generalized learning system, and a fault diagnosis process design method. The method improves the transparency of the model and the prediction of the molten iron quality through a mechanism and data collaborative driving strategy; By dividing the furnace and applying mechanism-based features to capture material and heat trends, and combining a stationary generalized feature learning system, the disturbance caused by non-stationary process features is reduced, and the inherent information embedded in the blast furnace iron-making process is mined; Then, by combining the stationary feature representation with the mechanism features, the time-matching generalized learning system aligns the process and quality variables with the molten iron quality as the target, uses mechanism and data-driven features to establish process monitoring statistics, and detects modeling deviations. The multi-phase multi-field mechanism modeling method comprises the following steps:

2. The method of claim 1, wherein, The molar quantity of the next time feeding is calculated as follows: 2.1) Establishing a mass balance model for the blast furnace ironmaking process, determining the kinetic coefficients of different reactions in each zone according to the current temperature T(t) This is achieved by using the Arrhenius formula, which provides the reaction rate constant as a function of temperature change; assuming that all reactions of each reactant are first order reactions, the range of each reaction in each zone where the subscripts 1 to 7 represent the seven main reactions, which can be derived as: In the above equation, a is an apparent frequency factor, e represents an exponential function, E a is an Arrhenius activation energy constant, R g is a molar gas constant, to respectively represent different reaction rate constants under 7 different reactions, represent the number of moles of different substances, where different subscripts represent different substances; Thus, the change in the amount of substance of each substance after the reaction is calculated Here different subscripts indicate different substances: where L t is the reaction constant; by a sufficient reaction, considering the load and gas content change of each layer; for the ingredients, by calculating the volume V u,z (t) after the reaction according to the molar volume The volume change V a,z (t) and the proportion of each component In each region: represents the specific number of moles of different species j; then the interzonal transport of the load is evaluated, including the loss of moles of each species from the current z zone to the z+1 zone below as follows: To accurately calculate the molar amount of each substance at the subsequent time step The following equation is executed: Then, for the gas behavior in the blast furnace, the bottom region receives the gas with constant composition from the hot blast stove and delivers it to the upper region, in the middle, the gas is directly discharged from the top region while receiving the gas below and reacting; It is assumed that the hot stove provides a constant gas composition as follows: Finally, the temperature of each region at the next time is identified as follows: In order to strictly simulate the gas behavior of each layer, the molar change of each layer is calculated according to the mass balance principle Here different subscripts g represent different gas components: 2.2) Heat balance modeling of the blast furnace ironmaking process includes six main aspects, namely, reaction heat Heat loss ΔH from furnace wall w,z (t), the heat obtained by the furnace ΔH d,z (t), the heat lost in the iron / slag ΔH p,z (t), A represents the slag discharge volume. Fe (t) represents the hot metal discharge rate, and the thermal changes ΔH of rising gas and falling furnace charge. g,z (t), ΔH b,z (t), respectively, are represented as: wherein represents the standard temperature heat of reaction; all represent the molar average of the different subscripted species; k Tw represents the furnace wall heat loss coefficient; AW z represents the furnace wall area; T e represents the blast furnace outer wall temperature; H v represents the rate of heat change; the molar mass of air; T hs (t) is the hot air temperature; represents the CaSio3concentration; A Fe (t) represents the Fe concentration; the total heat H z (t) is represented as: The molar heat capacity of each region is calculated and the average specific heat capacity The temperature change ΔT is determined z (t) is: This coupled thermo-chemical evolution continues until the mechanism performance error relative to the output model stops decreasing significantly in 10 iterations. T z (t+1) = T z (t) + ΔT z (t) (17); Where W represents the parameter matrix to be optimized; the optimization problem is expressed as a quadratic objective:

3. The method of claim 2, wherein, The data-driven stationary feature extraction includes the following steps: 3.1) for StaBFLS, when the provided process data X is taken as input, the relevant feature nodes and augmented nodes are obtained wherein, ξ q , ξ l is an activation function, is a weight, is a bias term; Then by means of linear connection, and combine to: When the extracted features are divided into 8 time segments, their mean and variance (μ e ,Σ e ) match those of the population mean and variance ; To quantify the difference, the Kullback-Leibler divergence is utilized and a loss function is defined as follows: L(θ) = -log p(θ) By adding an unconstrained logarithmic term, equation (20) is rewritten as: The following derivation is used: When equation (23) reaches a minimum at the optimal parameter matrix , the first derivative vanishes and optimization can proceed using the second derivative; with the notation [·] d,f , as the (d,f)-th element of the matrix, ([·] d ,[·] f ) as the (d,f)-th row vector, Θ e = log det W T ∑ e W, The second derivative is derived as: Finally, by substituting the lower bound, the complete quadratic objective is obtained, that is, the logarithmic term in equation (23) is substituted; the derivation of which is analogous to equation (25); Assume satisfies then the relationship is derived as: ​ From the Cauchy-Schwarz inequality and the fact that we get the lower bound of formula (26) holds: The molten iron quality modeling method based on the time-matching generalized learning system comprises the following steps: 3.2) The extracted data-driven features are able to minimize the difference between the predictions under supervision and the original target X, with a self-supervised loss with a regularization term representing: representing: one complete optimization objective, i.e. total loss is represented as: where the weights λ1, λ2∈(0, 1); for Taking the partial derivative, we obtain the following result: By setting 0, then the optimal parameter matrix can be derived as Afterwards, the data-driven StaFR is obtained by:

4. The method of claim 3, wherein, With a time-matching width s; combining the lagging features and the current features with the combination terms is represented by a matrix An optimal alignment of the current molten iron quality and the related lag information is performed by using a judgment criterion based on the Euclidean distance: Where a matching threshold lambda is introduced to filter unimportant connections; By using all candidate features, an information fusion connector is constructed to form a comprehensive feature, that is, The prediction deviation is calculated as follows: Subsequently, the fused features are utilized in the training process Replacing X, generating fused mapping and augmented nodes and Into the prediction task, by The prediction ability for the quality of molten iron is realized as: In the formula is the projection matrix of the time-matching generalized learning system, and λ3is a regularization weight parameter for avoiding numerical errors. In the theoretical case where lambda3 is zero, the inverse problem is converted into a least square problem, and the original pseudo-inverse is obtained: At the other end of the spectrum, when lambda3 tends to infinity, the solution of equation (36) is greatly constrained and tends to zero, and the following equation is obtained: The explanation of the regularization parameter lambda1 function in equation (29) is the same as that of lambda3. The fault diagnosis process design method comprises the following steps:

5. The method of claim 4, wherein, The final fault detection logic is as follows: 5.1) Constructing three T 2 Monitoring statistics, i.e. and are for mechanism features, data-driven StaFR and hot metal quality prediction bias, respectively: Mechanism features: Data-driven StaFR: Prediction deviation of hot metal quality: where k m , k d , and k y denote the dimension of the vector of and ε, respectively; each statistic determines a single threshold value according to its dimension and the chosen significance level ζ of the χ 2 distribution, with: Where "∩" represents the logical operator "and". ​

Images