Mechanism data co-driven metal smelting process key process index sensing method

Abstract

The invention relates to the technical field of metal smelting, and particularly discloses a mechanism data co-driven metal smelting process key process index sensing method, which comprises the following steps of S01, carrying out mechanism modeling in a smelting process, and providing a mechanism model for key process indexes to predict a key process index preliminary interpolation; s02, key process index accumulative error compensation: after a key process index preliminary interpolation is obtained through a mechanism model, spatial feature extraction is carried out on collected operation variables by introducing a CNN model, accumulative errors output by the mechanism model are compensated, and a complete key process index historical data sequence is obtained; and S03, based on probabilistic prediction of the compensation sequence, carrying out time sequence modeling and bidirectional feature capture on the sequence subjected to error compensation, improving the prediction precision in combination with an attention mechanism, and realizing prediction of each confidence interval and single-value prediction by adopting a mixed quantile regression mode. According to the method, the problem that the accuracy and the reliability of process index sensing in the traditional metal smelting process are low is solved.

Description

Technical Field

[0001] This application relates to the technical field of metal smelting, and specifically discloses a method for sensing key process indicators in the metal smelting process driven by mechanism data co - driving. Background Art

[0002] With the continuous development of industrial manufacturing and materials science, the metal smelting process plays a crucial role in the performance of metal materials and production quality. However, there are complex heating and heat transfer mechanisms and extremely harsh sensor environments in this process. Due to external factors such as high temperature and strong corrosion, modern sensors are still difficult to work stably for a long time even after multiple technological iterations, and the measurement accuracy may also be significantly reduced due to various interferences. To avoid the rapid damage of sensors in high - temperature scenarios, only manual - held sensing tools can be used for limited sampling before, making it difficult to comprehensively and accurately monitor the dynamic changes of the smelting process. The limited manual sampling data is obviously insufficient to provide an effective reference for the overall physical field and also difficult to meet the requirements for real - time optimization control of key process indicators such as temperature;

[0003] The metal smelting process is a key link in modern industrial production, and the accurate sensing of its process indicators directly affects product quality, production efficiency, and energy consumption. Traditional methods for sensing process indicators mainly rely on mechanism models and data - driven models, each with its own advantages and disadvantages. In recent years, attempts to combine mechanism models with data - driven methods to improve the accuracy and reliability of process indicator sensing have become the mainstream trend. However, due to various problems brought about by the harsh and complex environment of industrial smelting, new solutions still need to be designed to further solve them;

[0004] Mechanism models are based on an in - depth understanding of physical, chemical, and thermodynamic principles, and establish mathematical models to describe various reactions and transport phenomena in the smelting process. For example, the reaction mechanism of antimony in the copper smelting process has been studied, the multiphase distribution differences of antimony in different copper smelting processes have been analyzed, and a multiphase equilibrium model for the oxygen - enriched bottom - blown copper smelting process has been established to explore the influence of raw material composition on the distribution behavior of antimony. In addition, based on the process reaction mechanism and multiphase equilibrium principle, a thermodynamic simulation model of the copper side - blown smelting process has been established using the chemical equilibrium constant method to simulate the distribution behavior of product composition and impurity elements, providing theoretical support for optimizing process parameters. These mechanism models provide a theoretical basis for understanding complex phenomena in the smelting process. However, due to the existence of various uncertain factors in actual production, relying solely on mechanism models may be difficult to comprehensively and accurately sense process indicators;

[0005] With the development of industrial informatization, a large amount of production data has been accumulated in the smelting process. Data-driven methods utilize this historical data and adopt machine learning and deep learning algorithms to model and predict process indicators. For example, a machine learning model was established, and based on the experimental data of the National Institute of Standards and Technology (NIST) in the United States, the molten pool size of future parts during metal additive manufacturing was predicted, achieving accurate prediction of molten pool dynamics. In addition, the development of data-driven soft sensing technology in the blast furnace ironmaking process was reviewed, and the modeling methods and engineering applications were discussed, providing new ideas for the perception of process indicators in the blast furnace ironmaking process. However, data-driven methods have high requirements for data quality and quantity, and the physical meaning of the model may not be clear enough, which may lead to insufficient generalization ability;

[0006] To make up for the deficiencies of a single method, researchers have begun to explore the integration of mechanism models and data-driven methods to improve the accuracy and robustness of process indicator perception. For example, a computational model for decarburization kinetics in the solid metal smelting process was studied. Combining a macroscopic kinetic model and a simplified calculation method, six sets of equations for the melting decarburization process were established, providing a theoretical basis for the optimization of the smelting process. In addition, a data-driven materials innovation infrastructure was proposed, emphasizing the application of data-intensive strategies and machine learning algorithms in materials science, providing a new perspective for the integration of mechanism models and data-driven methods. These integration methods effectively combine the physical constraints of mechanism models and the learning ability of data-driven models, improving the accuracy of process indicator perception.

[0007] However, due to the complex thermodynamic environment and poor sensor conditions in the metal smelting process, a considerable part of the process indicators are sparsely collected, with large measurement deviations, and manual sampling may also introduce potential data variations. Traditional single-value prediction may not meet the decision-making needs for the physical field in the metal smelting process, and it is unable to effectively predict abnormal process indicators to assist in fault diagnosis, which is not conducive to improving the liquid phase quality and process stability. The above work has not proposed an effective method to solve this problem;

[0008] In view of this, the present invention provides a method for perceiving key process indicators in the metal smelting process driven by mechanism and data to solve the above problems. Summary of the Invention

[0009] The purpose of the present invention is to solve the problem of low accuracy and reliability in the perception of process indicators in the traditional metal smelting process.

[0010] To achieve the above purpose, the present invention provides the following basic solution:

[0011] A method for perceiving key process indicators in the metal smelting process driven by mechanism and data, comprising the following steps:

[0012] S01: Mechanism Modeling of the Melting Process: Clean and align historical operation data, and then conduct mechanism modeling to provide a mechanism model for key process indicators and predict the initial interpolation of key process indicators;

[0013] S02: Cumulative Error Compensation of Key Process Indicators: After obtaining the initial interpolation of key process indicators through the mechanism model, introduce a CNN model to extract spatial features of the collected operating variables, compensate for the cumulative error output by the mechanism model, and obtain a complete historical data sequence of key process indicators;

[0014] S03: Probabilistic Prediction Based on the Compensated Sequence: Design a probabilistic prediction method using a tandem model architecture, conduct time series modeling and bidirectional feature capture on the sequence after error compensation, combine the attention mechanism to improve the prediction accuracy, and use the hybrid quantile regression method to achieve predictions for each confidence interval and single-value prediction, and finally provide a key indicator prediction result that can reflect uncertainty information.

[0015] Furthermore, in step S01, a mass conservation equation, a solid-liquid two-phase dynamic heat balance equation, and auxiliary equations were constructed, and the overall mechanism model was built based on this dynamic equation system. The specific steps are as follows:

[0016] B01: Through the thermal balance analysis and mass dynamic balance modeling of the three stages of solid-state heating, phase change, and liquid-state heating, the physical description of the entire melting process was realized and the mechanism prediction of key process indicators was completed;

[0017] B02: For the heat loss caused by intermittent operations such as furnace door opening and closing, an event-driven discrete-continuous hybrid model was proposed for accurate characterization;

[0018] B03: In terms of numerical solution, use an adaptive numerical solution framework to adopt the explicit Euler method to advance the mechanism model in time, and combine the event detection algorithm to capture the switching of operating states.

[0019] Furthermore, within the melting time, the model iteratively calculates the melting and temperature change processes of solid-liquid metal in a high-temperature environment with a time step of Δt. The model inputs include the initial solid-liquid mass and temperature, the mass flow rates of fuel and air, the furnace door opening state and opening duration, the solid-liquid contact area, and relevant physical property constants; the outputs are the liquid temperature, solid temperature, solid mass, and liquid mass at any moment. In each iteration, the model first determines whether the metal is in the meltable temperature range and calculates the melting rate, updates the solid-liquid two-phase mass distribution and contact area; then considers the heat dissipation situation of furnace door opening or closing and other heat losses, and finally comprehensively obtains the solid-liquid two-phase temperature at the current moment based on the fuel heat input, air heat input, and various heat losses; if the solid metal is completely melted in a certain step, it is updated to pure liquid metal in a timely manner and the process information is recorded.

[0020] Furthermore, in step S02, a Convolutional Neural Network (CNN) is used to implement cumulative error compensation. The CNN has a network structure with multi-channel inputs and can adapt to multi-modal input data. The multi-channel CNN includes a convolutional layer (Conv), a batch normalization layer (BN), a pooling layer (Pooling), and a fully connected layer, which respectively extract spatial features from input data of different dimensions. The depth feature vectors output by each channel will be mapped to their respective sub fully connected layers to obtain relatively independent high-level information expressions. Finally, the multi-channel information is fused in an additional fully connected layer. Through this fusion process, the CNN model can effectively integrate the key information from the real sampling values, the mechanism model interpolation, and the process variables, output the optimal error compensation value, and then obtain a complete historical data sequence of the key process indicators.

[0021] Furthermore, the serial model architecture is a CNN-BiLSTM-Attention serial architecture. Probabilistic prediction is achieved through the CNN-BiLSTM-Attention serial architecture. The serial model architecture includes a probabilistic prediction model of a deep fusion convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism (Attention).

[0022] Furthermore, the method for constructing the serial architecture is as follows:

[0023] First, multiple groups of convolutional kernels are used for parallel scanning to complete the reconstruction of the feature space. By means of depthwise separable convolution, complex local index change patterns are captured in the time dimension, potential correlations among multiple variables such as the temperatures inside and outside the furnace and the gas flow rate are extracted, and at the same time, noise interference is reduced and meaningful process features are retained.

[0024] Next, the extracted high-dimensional features are input into the BiLSTM network, which accurately depicts the dynamic evolution process of the heating and cooling stages from both the forward and backward directions respectively, so as to take into account both historical states and future trends.

[0025] Finally, to further highlight the decisive influence of key operation nodes on the prediction results, the model adaptively weights the BiLSTM output sequence through the attention mechanism, and sets different importance weights in both the time and feature dimensions.

[0026] Furthermore, to achieve probabilistic prediction, in quantile regression, the serial architecture will be trained separately for each quantile;

[0027] The training objective is to minimize the average loss function L of each quantile q , which is described as:

[0028]

[0029] When many quantiles need to be trained, there is a high computational burden; to alleviate this computational burden, this solution supplements the model output with a multi-dimensional output layer and uses the loss function as the average pinball loss L for all quantiles:

[0030]

[0031] T: The total time T (total number of samples in one round of training) corresponding to one round of training; q: The target quantile; The q-quantile estimated at time t (the t-th sample in one round of training); L q,t Represents the pinball loss for the q-quantile at time t; Q: The total number of quantiles.

[0032] Furthermore, to make up for the defects of the pinball loss function, the Huber norm is introduced into the loss function. The Huber norm can make the loss function differentiable everywhere with only a small approximation. Specifically, the Huber norm can be regarded as a combination of the L1 and L2 norms:

[0033]

[0034] where ε represents the threshold amplitude of the L1 and L2 norms. When the prediction error is below the threshold, the Huber norm is the L2 norm; when the prediction error is greater than the threshold, the Huber norm is the L1 norm. Then, substituting into the equation, using the Huber norm, the approximate pinball loss can be calculated as:

[0035]

[0036] Furthermore, to further balance the quantile regression accuracy and the central tendency constraint, a hybrid quantile loss method is adopted, which weights and fuses the pinball loss and the mean square error. Its loss training objective can be written as:

[0037]

[0038] where MSE is used to constrain the prediction from diverging too much; is used to retain the flexibility of quantile regression in interval prediction; α reg ∈[0,1] is used to balance the importance of the two; when α reg is large, it is more inclined to reduce the overall error, that is, the interval will relatively narrow; when α reg is small, it focuses more on quantile learning, that is, the interval is wider but the coverage is better.

[0039] Furthermore, it also includes performing simulation verification on the data obtained through the above method. The specific simulation verification operations are as follows:

[0040] Adopt R 2 as the performance index; for aspects such as the reliability, width, comprehensive quality of the prediction interval, and the accuracy of conditional quantiles, multiple evaluation indexes are selected for quantification. Among them, the prediction interval coverage probability (PICP) and average coverage error (ACE) are used to measure the reliability of interval prediction, the prediction interval normalized average width (PINAW) is used to evaluate the concentration degree of interval prediction, and the pinball loss (PL) is used to reflect the accuracy of the prediction model on conditional quantiles.

[0041] The principle and effect of this solution are as follows:

[0042] 1. Compared with the prior art, this solution proposes a key process index perception method driven by mechanism data co-driving. By introducing the mechanism model of the metal melting process and combining the learning ability of the convolutional neural network, the sparse data sampled manually is compensated for errors, improving the generalization ability and prediction accuracy of the model. For the complete data after compensation, this solution additionally combines the bidirectional long short-term neural network and the attention mechanism. Based on the tandem model architecture CNN-BiLSTM-Attetntion, a probabilistic prediction method is designed using the hybrid quantile regression method, giving richer prediction result information, which is beneficial to the further progress of control decision-making. Through the above solution, the accuracy and reliability of the metal melting process are further improved, providing guarantee for the stability and safety of the production process.

[0043] 2. Compared with the prior art, this solution proposes a key process index perception method for the metal melting process driven by mechanism data co-driving. First, a key index mechanism model with the dynamic heat balance of the metal melting process as the main body is established to preliminarily interpolate the key process index sequence with only a small amount of manually sampled data; secondly, a cumulative error compensation model is designed by integrating the characteristics of the melting process, and the mechanism prediction value is compensated for errors using the manually sampled data to obtain the key process index sequence that completes the final interpolation; finally, by designing a sequence probabilistic prediction model, rich perception information such as confidence intervals and single-value predictions can be given simultaneously, and the compensated sequence is used to guide the model to realize the effective perception of key process indexes, assisting in fault diagnosis and continuous process improvement.

[0044] 3. Compared with the prior art, the method proposed in this solution can effectively interpolate and construct the key index sequence using less manually sampled data, and obtain a perception result with higher accuracy, providing a new solution for improving the production efficiency of the metal melting process and reducing the manual labor intensity. Description of the Drawings

[0045] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0046] Figure 1 Fig. shows a schematic diagram of the framework of the key process index perception method for the metal melting process driven by mechanism data and co-driven data in the embodiments of the present application;

[0047] Figure 2 Fig. shows a schematic diagram of the CNN cumulative error compensation model in the key process index perception method for the metal melting process driven by mechanism data and co-driven data in the embodiments of the present application;

[0048] Figure 3 Fig. shows a schematic diagram of the architecture of the probabilistic prediction model in the key process index perception method for the metal melting process driven by mechanism data and co-driven data in the embodiments of the present application;

[0049] Figure 4 Fig. shows a schematic diagram of the probabilistic prediction effect display of each loss function in the key process index perception method for the metal melting process driven by mechanism data and co-driven data in the embodiments of the present application. Detailed implementation manners

[0050] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in combination with the accompanying drawings and preferred embodiments, detail the specific implementation manners, structures, features, and their effects of the present invention as follows.

[0051] Embodiments are as Figures 1-4 shown:

[0052] A key process index perception method for the metal melting process driven by mechanism data and co-driven data includes the following steps:

[0053] S01: Mechanism modeling of the melting process: Clean and align the historical operation data, and then perform mechanism modeling to provide a mechanism model for key process indexes to predict the preliminary interpolation of key process indexes;

[0054] Cleaning: For the non-label values in the historical operation data, remove the outliers and interpolate and supplement the missing values using simple KNN; Alignment: Align the time stamps of the collected values of each sensor for obtaining different historical operation data, unify the sampling frequency, and the unified method adopts linear interpolation upsampling and moving average downsampling;

[0055] S02: Cumulative Error Compensation for Key Process Indicators: After obtaining the preliminary interpolation of key process indicators through the mechanism model, a CNN model is introduced to extract spatial features from the collected operating variables, compensate for the cumulative error output by the mechanism model, and obtain a complete historical data sequence of key process indicators;

[0056] S03: Probabilistic Prediction Based on the Compensated Sequence: A probabilistic prediction method is designed using a tandem model architecture. Time series modeling and bidirectional feature capture are performed on the sequence after error compensation. The attention mechanism is combined to improve the prediction accuracy, and a hybrid quantile regression method is used to achieve predictions for each confidence interval and single-value prediction. Finally, a key indicator prediction result that can reflect uncertainty information is provided.

[0057] Regarding S01:

[0058] For the metal melting process, a mass conservation equation, a solid-liquid two-phase dynamic heat balance equation, and various auxiliary equations are constructed, and the overall mechanism model is built based on this dynamic equation system. First, through the heat balance analysis and mass dynamic balance modeling of the three stages of solid-state heating, phase change, and liquid-state heating, the physical description of the entire melting process is realized and the mechanism prediction of key process indicators is completed. For the heat loss caused by intermittent operations such as furnace door opening and closing, an event-driven discrete-continuous hybrid model is proposed for accurate characterization. In terms of numerical solution, an adaptive numerical solution framework is used to perform time marching on the mechanism model using the explicit Euler method, and an event detection algorithm is combined to capture the switching of operating states.

[0059] The following analyzes typical mechanism equations, and flexible corrections and transformations can be made according to special circumstances in actual applications;

[0060] Mass conservation equation:

[0061] Among them, is the change rate of the solid-phase metal mass with respect to time, is the change rate of the liquid-phase metal mass with respect to time, is the melting rate of the solid metal.

[0062] Solid-liquid dynamic heat balance equation:

[0063]

[0064] Among them, m s is the solid-phase metal mass, c s (T s ) is the specific heat capacity of the solid-phase metal, ΔT s is the change amount of the solid-phase metal temperature per unit time, h sf is the heat transfer coefficient at the solid-liquid interface, A sf is the solid-liquid contact area, Tl is the liquid-phase metal temperature, T s is the solid-phase metal temperature, and L is the latent heat of fusion of the metal.

[0065] m l C l ΔT l = Q fuel + Q air - Q loss - h sf A sf (t)(T l - T s );

[0066] Wherein, m l is the mass of the liquid-phase metal, C l is the specific heat capacity of the liquid-phase metal, ΔT l is the change amount of the liquid metal temperature per unit time, Q fuel is the heat input by fuel combustion, Q air is the heat brought in by air, Q res is the initial heat of the remaining liquid metal, Q loss is the total heat loss of the furnace body.

[0067] Auxiliary equation:

[0068] For dealing with the problem of the latent heat term, aiming at the characteristics of solving dynamic heat balance, the liquid fraction method is adopted to approximately calculate the liquid phase ratio, so as to calculate the two-phase mass, and the melting rate can be calculated according to the values at adjacent moments. For the melting process, that is, when T1 < T s < T2, according to relevant literature, the liquid phase ratio of the metal can be approximately expressed as:

[0069]

[0070] Where f L (t) is the liquid phase ratio of the metal at time t, and T1 and T2 are the solidus temperature and liquidus temperature of the metal respectively.

[0071] The melting rate at time t + 1 can be deduced as:

[0072]

[0073] The contact area is affected by complex factors and it is difficult to accurately estimate in some industrial environments. In the case of being unable to measure, according to relevant literature, it is assumed that the melting is reduced in proportion to the melt, then the mass change during the melting process can be described by the following formula. If there are other requirements for the mass evolution process, etc., refer to the empirical formulas of each smelting process for estimation or dynamic measurement;

[0074]

[0075] The heat input mainly considers the calorific value of fuel combustion Q fuel and the heat brought in by air Q air as the typical heat input part. The specific analysis is as follows:

[0076]

[0077] Among them, is the mass flow rate of natural gas, Q dw is the lower calorific value of natural gas, and η is the combustion efficiency.

[0078]

[0079] Among them, is the air flow rate, c air is the specific heat capacity of air, T air is the air preheating temperature, T amb is the ambient temperature.

[0080] Q res =m r0 C l T r0 ;

[0081] Among them, m r0 is the initial mass of the liquid metal, T r0 is the initial temperature of the remaining liquid metal.

[0082] The heat loss Q loss mainly considers the heat dissipation of the furnace body Q furnace the heat dissipation of the flue gas escaping Q somke and the heat dissipation of the furnace door opening escaping Q door . The specific analysis is as follows:

[0083] Q loss =Q furnace +Q somke +Q door ;

[0084]

[0085] Q door =h door A door (T l -T amb )f open (t);

[0086] Among them, A wall is the inner surface area of the furnace wall, k ins is the thermal conductivity of the furnace wall insulation material, d ins is the thickness of the furnace wall, ∈ is the radiation emissivity of the furnace wall surface, σ is the Stefan-Boltzmann constant, h dooris the comprehensive heat transfer coefficient of the furnace door (including radiation and convection), A door is the opening area of the furnace door, f open (t) is the opening time of the furnace door;

[0087] By calculating through multi - physical equations, the prediction results of the corresponding mechanism model can be effectively calculated. It has high flexibility. The melting rate formula, contact area, etc. can be updated or modified according to the actual metal and the needs of the smelting process. Heat newly introduced due to equipment, etc. can also be incorporated into equations such as heat loss. For unmeasured heat, etc., it can be replaced or simplified in the form of a percentage according to industry experience formulas to complete mechanism interpolation. And the relevant errors caused by simplification are processed in the cumulative error compensation;

[0088] Taking the metal temperature in the smelting process as a typical case, a mechanism model is built to form a prediction framework. This mechanism model mainly calculates the melting and temperature change process of solid - liquid metal in a high - temperature environment with a time step of Δt during smelting. The model inputs include the initial solid - liquid mass and temperature, the mass flow rates of fuel and air, the opening state and opening duration of the furnace door, the solid - liquid contact area, and relevant physical property constants, etc.; the outputs are the liquid temperature, solid temperature, solid mass, and liquid mass at any moment. In each iteration step, the model first determines whether the metal is in the meltable temperature range and calculates the melting rate, updating the mass distribution and contact area of the solid - liquid two - phase; then it considers the heat dissipation situation when the furnace door is opened or closed and other heat losses (such as the furnace body, flue gas, etc.). Finally, based on the fuel heat input, air heat input, and various heat losses, the solid - liquid two - phase temperature at the current moment is obtained. If the solid metal is completely melted (or the temperature exceeds the target melting upper limit) in a certain step, it is updated to pure liquid metal in a timely manner and the process information is recorded. Through this cyclic iteration, the model can simulate the dynamic changes of metal temperature and melting state over time under different operating conditions;

[0089] The advantages of this model are its strong flexibility and robustness: on the one hand, the known or measurable parameter terms in the heat balance equation can be appropriately increased or decreased according to actual production data, so as to overall compensate and eliminate errors without changing the main physical mechanisms; on the other hand, it considers the synergistic effects of multiple factors such as fuel, air, furnace door opening state, heat dissipation loss, etc., and can more accurately reflect the real operation situation in the industrial smelting process;

[0090] In S02:

[0091] In the online estimation process of the mechanism model, errors often accumulate over time, resulting in deviations in subsequent prediction results. To solve this problem, this solution uses a convolutional neural network (CNN) to achieve cumulative error compensation. The model architecture is as Figure 2As shown, the CNN has a network structure with multi-channel inputs, which can adapt to multi-modal input data, learn the complex relationships between variables from multiple levels, and then capture more targeted information.

[0092] In an actual production environment, the compensation model needs to make full use of multi-source data to accurately capture the characteristics of error accumulation at different stages. On the one hand, the model receives the true values at the sampling time points to obtain observable information closely related to the actual working conditions; on the other hand, the interpolation outputs generated by the mechanism model during the corresponding period are also incorporated to help the model understand the possible prediction biases under the prior physical mechanism. In addition, to more comprehensively describe the dynamic relationships in the smelting process, each process variable is also integrated into the input layer, so as to ensure that the network can perform data fusion in a higher dimension and finally generate more accurate error compensation values.

[0093] In the multi-channel CNN for realizing the effective integration and feature learning of the above multi-source data, each channel, with the cooperation of basic modules such as the convolutional layer (Conv), batch normalization layer (BN), pooling layer (Pooling), and fully connected layer (FC), respectively extracts features from input data of different dimensions (such as key index sequences and process variables). The sliding of the convolutional kernel in the local area can identify the latent local correlations in industrial process data, and the BN layer and Pooling layer help to stabilize the network training and extract more representative intermediate features. Thereafter, the deep feature vectors output by each channel will be mapped into their respective sub fully connected layers to obtain relatively independent high-level information expressions, and finally, the multi-channel information is fused in an additional fully connected layer. Through this fusion process, the model can effectively integrate the key information from the true sampling values, mechanism model interpolation, and process variables, and output the optimal error compensation value.

[0094] The multi-channel CNN compensation model adopted in this solution, by efficiently capturing local correlations and multi-variable information, can deeply describe and correct the deficiencies in the cumulative error of the interpolation output. When the error compensation value generated by it is superimposed on the preliminary interpolation of the mechanism model, a more accurate final interpolation label can be obtained, significantly improving the understanding of the key process indicators in the smelting process. This solution not only reduces the deviation caused by error accumulation but also provides more robust data support for subsequent prediction and decision-making, thus realizing the refined and intelligent management of the industrial smelting process.

[0095] After using the interpolation method to supplement the missing values to obtain a complete data set, to solve the long-range dependence problem, a CNN-BiLSTM-Attention cascade architecture is adopted to achieve probabilistic prediction.

[0096] For the characteristics of strong non - linearity, time - lag, and multi - variable coupling existing in most of the key index sequences during the metal smelting process, this solution adopts a probability prediction model that deeply integrates a convolutional neural network (CNN), a bidirectional long short - term memory network (BiLSTM), and an attention mechanism (Attention). Through the feature space reconstruction and information focusing mechanism, this architecture effectively solves the problems of insufficient sensitivity of traditional single models to local features and incomplete modeling of temporal dependencies. Its overall architecture is as shown in Figure 3 shown;

[0097] After interpolating and complementing the smelting index sequence and compensating for the cumulative error to obtain a complete data set, this solution realizes a unified framework for probabilistic prediction of key process indicators during the metal smelting process through a cascaded model, that is, by combining a convolutional neural network (CNN), a bidirectional long short - term memory network (BiLSTM), and an attention mechanism (Attention).

[0098] First, multiple groups of convolutional kernels are used for parallel scanning to complete the reconstruction of the feature space. By means of depth - separable convolution, complex local index change patterns are captured in the time dimension, potential correlations among multi - variables such as the temperature inside and outside the furnace and gas flow are extracted, while noise interference is reduced and meaningful process features are retained. Then, the extracted high - dimensional features are input into the BiLSTM network, which accurately depicts the dynamic evolution process of the heating and cooling stages from both the forward and backward directions respectively, thus taking into account both historical states and future trends. Compared with traditional unidirectional temporal modeling methods, the reverse - propagation features captured by the bidirectional structure can more comprehensively reveal the potential global dependencies during the index change process and effectively alleviate the problem of temporal blind spots caused by insufficient utilization of future information.

[0099] Finally, to further highlight the decisive influence of key operation nodes on the prediction results, the model adaptively weights the BiLSTM output sequence through the attention mechanism, setting different importance weights in both the time and feature dimensions. By enhancing the sensitivity to multi - variable coupling relationships, long - range dependencies, and short - term mutations, this module provides a more interpretable decision - making basis for accurate prediction and anomaly detection;

[0100] In S03:

[0101] The above architecture overall constructs a three - level information processing flow of "local feature extraction - temporal dependence modeling - global information focusing", and uses a subsequent fully - connected layer to achieve multi - dimensional output. Finally, the upper and lower bounds of key process indicators in the future period are presented in the form of interval prediction combined with single - value prediction. This hierarchical structure helps to effectively complement between capturing local mutations and modeling long - range dependencies, and improves the model's ability to identify key process nodes in the global information focusing stage through a multi - dimensional attention mechanism. With the confidence ranges provided by interval prediction, more sufficient decision - making information can be obtained for process optimization and anomaly detection. To achieve probabilistic prediction, in traditional quantile regression, the model is trained separately for each quantile. The training objective is to minimize the average loss function \(L\) of each quantile q , which is described as:

[0102]

[0103] When many quantiles need to be trained, there will be a high computational burden. To reduce this computational burden, this solution supplements a multi - dimensional output layer in the model output, and uses the average pinball loss \(L\) of all quantiles as the loss function:

[0104] denoted as the quantile loss;

[0105] \(T\): the total number of time steps \(T\) (total number of samples in one round of training) corresponding to one round of training; \(q\): the target quantile; the \(q\) - th quantile estimated at time \(t\) (the \(t\) - th sample in one round of training); \(L\) q,t represents the pinball loss for the \(q\) - th quantile at time \(t\); \(Q\): the total number of quantiles (this solution uses the median quantile, i.e., the 0.5 quantile as the single - value output, and quantiles such as 0.025 and 0.975 can form the upper and lower intervals of the index prediction with a 95% confidence level, etc., providing a more comprehensive index perception for control decisions).

[0106] Compared with other non - parametric methods, this quantile regression method is more flexible and can provide various forms of prediction results according to requirements; however, since it uses the pinball loss function in constructing the parameter training optimization model for the quantile prediction model, and this function is non - continuous, it leads to difficult parameter training and a large amount of calculation. Since the pinball loss is non - differentiable everywhere, therefore, the Huber norm is introduced into the loss function, and with only a small approximation, the loss function can be made differentiable everywhere. The Huber norm can be regarded as a combination of the L1 and L2 norms:

[0107]

[0108] where \(\varepsilon\) represents the threshold amplitude of the L1 and L2 norms. When the prediction error When below the threshold, the Huber norm is the L2 norm; when the prediction error is greater than the threshold, the Huber norm is the L1 norm. Then, substituting into the equation and using the Huber norm, the approximate pinball loss can be calculated as:

[0109]

[0110] Denoted as the norm quantile loss (HuberQuantileLoss). However, in the actual industrial process applications, it is found that the process error distribution is relatively complex, and using the norm method is prone to problems such as too narrow prediction intervals or inflexible characterization of quantiles. To further balance the accuracy of quantile regression and the central tendency constraint, this solution adopts a method of mixed quantile loss (MixedQuantileLoss), which weights and fuses the pinball loss (Pinball) and the mean squared error (MSE). Its loss training objective can be written as:

[0111]

[0112] Among them, MSE replaces the role of the norm quantile loss to constrain the prediction from diverging too much; is used to retain the flexibility of quantile regression in interval prediction; α reg ∈[0,1] is used to balance the importance of the two. When α reg is larger, it is more inclined to reduce the overall error, that is, the interval will relatively narrow; when α reg is smaller, it emphasizes more on quantile learning, that is, the interval is wider but the coverage is better.

[0113] Through this mixed form, on the one hand, it can avoid problems such as numerical instability and slow convergence caused by pure pinball loss, and on the other hand, it can also suppress the disadvantages of too narrow or inflexible intervals brought by

[0114] To obtain the quantification of single-value prediction accuracy on the basis of interval prediction, this solution adopts the median quantile, that is, the 0.5 quantile as the single-value output;

[0115] Verification of effectiveness:

[0116] This method has been verified by simulation in a typical smelting process of a certain metal. To ensure the representativeness of the experiment, the key process index of liquid phase temperature in the metal smelting process was selected, and its thermodynamic model was analyzed and matched with the aforementioned mechanism model. For some parameters that have not been mastered or missing data, approximate calculations were carried out based on historical experience and industry standards. Finally, the preliminary data completion was completed through the mechanism model. In addition, to further reduce the impact of residual error on model prediction, cumulative error compensation was carried out after mechanism interpolation, and the final sequence obtained after compensation was regarded as the real and complete measurement data. Subsequent experiments and performance indicators were carried out based on this data.

[0117] This patent is based on the cascaded model CNN-BiLSTM-Attention, and two simplified models CNN-BiLSTM and CNN-Attention are used for comparative experiments. Through this kind of comparison, the influence of multi-module fusion on prediction performance can be intuitively evaluated. At the same time, to verify the advantage of the hybrid loss L mixed in quantile prediction, the same model version with a designed non-norm loss function is adopted to investigate the performance differences under the same network structure but different loss function settings.

[0118] The prediction scheme of this patent is to improve the accuracy and stability of probability prediction. For the measurement of the prediction ability of single-value prediction, R 2 is used as the performance index; for aspects such as the reliability, width, comprehensive quality of the prediction interval, and the accuracy of conditional quantiles, a variety of evaluation indicators are selected for quantification. Among them, the prediction interval coverage probability (PICP) and average coverage error (ACE) are used to measure the reliability of interval prediction, the normalized average width of the prediction interval (PINAW) is used to evaluate the concentration degree of interval prediction, and the pinball loss (PL) is used to reflect the accuracy of the prediction model on conditional quantiles. By comprehensively considering these indicators, the practicability of the prediction interval and the effective degree of uncertainty quantification of this patent can be comprehensively judged.

[0119] R 2 can be expressed by the sum of squared residuals and the total sum of squares:

[0120]

[0121] where N is the total number of samples, y i is the true value of the sample target variable, is the model prediction value of the sample target variable. R 2 is optimally 1 in value, that is is always 0, the predicted value is exactly the same as the actual value, and the closer it is to 1, the better the single-value prediction performance.

[0122] MAPE measures the accuracy of the prediction model and can be expressed as:

[0123]

[0124] The closer the MAPE is to 0, the better the effect.

[0125] ACE quantifies the reliability by calculating the absolute value of the difference between the prediction interval coverage probability (PICP) and the prediction interval nominal confidence level (PINC):

[0126] ACE = |PICP - PINC|;

[0127] where it is assumed that and correspond to the lower bound and the upper bound of the prediction interval of the i-th sample at the confidence level of α respectively, then PICP can be calculated by the following formula:

[0128]

[0129] It can be found that the closer ACE is to zero, the higher the reliability of the model prediction interval is reflected.

[0130] PINAW reflects the width of the prediction interval and can quantitatively show the concentration degree of the interval prediction distribution. Since when the model reliability is the same, the narrower the prediction interval, the better the prediction effect and the greater the guiding significance for decision-making. Its definition is as follows:

[0131]

[0132] where R is the difference between the maximum value and the minimum value in the observed values. The smaller the value of PINAW, the smaller the width of the prediction interval, the more concentrated the prediction distribution, and the better the prediction effect is reflected.

[0133] PL can better reflect the accuracy of the conditional quantiles predicted by the model. When calculating using the two quantiles corresponding to each confidence interval, it can be expressed as:

[0134]

[0135] The smaller PL is, the higher the prediction accuracy is.

[0136] To compare the simulation experiment effects of the model and the comparative model, this patent shows the performance indicators of single-value prediction and interval prediction, as shown in Table 1 and Table 2 respectively. It can be found that the series model can better integrate the ability advantages of each part of the model and has achieved good effects in all aspects, verifying the effectiveness of this solution.

[0137] Table 1 Comparison of single-value prediction effects of each model

[0138]

[0139] Table 2 Comparison of interval prediction effects of each model

[0140]

[0141] Table 3 compares the above three quantile losses to verify the effectiveness of the theory. As can be seen from Table 3, there are significant differences in the three indicators of ACE, PINAW, and PL for different loss functions. Although the norm quantile loss can provide a relatively more compact interval, the ACE value is relatively large, that is, the coverage gap is relatively large; while the pure quantile loss shows good coverage, but often leads to an overly wide interval. In contrast, the hybrid quantile loss through the reasonable integration of the quantile loss and the norm term enables α reg to obtain a suitable value, which can not only ensure a small coverage error but also effectively reduce the interval width (for example, better PINAW values are obtained at three confidence levels of 85%, 90%, and 95%), thus achieving a more balanced comprehensive performance between coverage and interval compactness.

[0142] Comparison of the interval prediction effects of each loss function in Table 3

[0143]

[0144] Figure 4 More vividly demonstrates the effectiveness of this method. The comparison results shown in the figure indicate that if α reg is selected as 0.8, as shown in Figure 4 (b), the dependence on the norm error term in the hybrid quantile loss increases significantly and almost approaches Figure 4 (d) pure norm quantile regression; at this time, although the prediction interval is often more compact, the coverage rate will also decrease accordingly, and the robustness to extreme disturbances in the working conditions decreases. When α reg takes a smaller value, such as taking 0.3 as shown in Figure 4 (a), the model pays more attention to the quantile error. Although the interval is likely to become wider, it still shrinks to a certain extent compared with Figure 4 (c), enabling the coverage rate to be fully guaranteed. This adjustment process is conducive to further improving the interval prediction effect. In actual process requirements, in order to balance the monitoring ability of abnormal disturbances and the accuracy requirements for single-value prediction, the experiment finally selects α reg = 0.3, a relatively small value. It can be found that on the one hand, the sensitivity of the quantile regression to the tail distribution is fully utilized to ensure that the prediction interval has a reasonable coverage rate; on the other hand, it also relies on a certain proportion of norm error constraints, making the single-value prediction accuracy higher than that of the pure quantile regression, and further reducing the interval width during this process, taking into account both robustness and prediction practicality.

[0145] As shown in Figure 4 :

[0146] Figure 4 The medium curve and interval distribution also reflect this compromise effect. Within the range allowed by process requirements, it can cover more potential fluctuation situations, which is more valuable for anomaly detection and real-time monitoring.

[0147] Verification in the simulation environment of a typical metal smelting process shows that the CNN-BiLSTM-Attention cascade structure is superior to the comparison models in terms of prediction accuracy and interval coverage reliability, and has better comprehensive performance in aspects such as prediction interval width and single-value prediction accuracy. From various quantitative indicators (R 2 , MAPE, as well as ACE, PINAW, and PL), it can be seen that after constructing a three-level information processing flow of "local feature extraction - temporal dependence modeling - global information focusing", the model can not only accurately capture the evolution characteristics of temperature or other key indicators during the heating and cooling processes, but also provide a more robust upper and lower bound range at the interval prediction level, bringing more sufficient decision-making information for anomaly detection and process optimization.

[0148] In summary, the experimental results fully demonstrate that this probabilistic prediction scheme can effectively address complex thermodynamic nonlinearity and multivariable coupling problems in the metal smelting process, showing higher robustness and prediction accuracy in multiple indicators, indicating that this method has good feasibility and practical value in actual industrial production and subsequent extended applications.

[0149] The above is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or equivalent changes and modifications within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any brief modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for perceiving key process indicators in the metal smelting process driven by mechanism data, characterized in that, It includes the following steps: S01: Mechanism modeling of the smelting process: Clean and align historical operation data, and then conduct mechanism modeling to provide a mechanism model for key process indicators to predict the initial interpolation of key process indicators; S02: Cumulative error compensation for key process indicators: After obtaining the initial interpolation of key process indicators through the mechanism model, introduce a CNN model to extract spatial features of the collected operating variables, compensate for the cumulative error output by the mechanism model, and obtain a complete historical data sequence of key process indicators; S03: Probabilistic prediction based on the compensated sequence: Design a probabilistic prediction method using a tandem model architecture, conduct time series modeling on the sequence after error compensation, use convolution to extract features and bidirectional feature capture, combine the attention mechanism to improve prediction accuracy, and use a hybrid quantile regression method to achieve predictions for each confidence interval and single-value prediction, and finally provide a key indicator prediction result that can reflect uncertainty information.

2. The key process index perception method for the metal melting process driven by mechanism data according to claim 1, characterized in that In step S01, a mass conservation equation, a solid-liquid two-phase dynamic heat balance equation, and auxiliary equations are constructed, and the overall construction of the mechanism model is completed with this dynamic equation system. The specific steps are as follows: B01: Through the thermal balance analysis and mass dynamic balance modeling of the three stages of solid-state heating, phase change, and liquid-state heating, the physical description of the entire smelting process is realized and the mechanism prediction of key process indicators is completed; B02: A discrete-continuous hybrid model driven by events is proposed for accurate characterization of the heat loss caused by intermittent operations such as furnace door opening and closing; B03: In terms of numerical solution, use an adaptive numerical solution framework to adopt the explicit Euler method to advance the mechanism model in time, and combine an event detection algorithm to capture the switching of operating states.

3. A key process index perception method for the metal smelting process co-driven by mechanism data according to claim 2, characterized in that, During the smelting time, the model iteratively calculates the melting and temperature change processes of solid-liquid metal in a high-temperature environment with a time step of Δt. The model inputs include the initial solid-liquid mass and temperature, the mass flow rates of fuel and air, the furnace door opening state and opening duration, the solid-liquid contact area, and relevant physical property constants; the outputs are the liquid temperature, solid temperature, solid mass, and liquid mass at any moment. In each iteration, the model first determines whether the metal is in the meltable temperature range and calculates the melting rate, and updates the solid-liquid two-phase mass distribution and contact area; Subsequently, consider the heat dissipation situation of furnace door opening or closing and other heat losses, and finally obtain the solid-liquid two-phase temperature at the current moment based on the comprehensive fuel heat input, air heat input, and various heat losses; If the solid metal is completely melted in a certain step, it is updated to pure liquid metal in time and the process information is recorded.

4. A method for perceiving key process indicators in the metal smelting process driven by mechanism data according to claim 3, characterized in that, In step S02, a Convolutional Neural Network (CNN) is used to achieve cumulative error compensation. The CNN has a network structure with multi-channel inputs and can adapt to multi-modal input data. The multi-channel CNN includes a convolutional layer (Conv), a batch normalization layer (BN), a pooling layer (Pooling), and a fully connected layer (FC). It extracts spatial features from input data of different dimensions respectively. The depth feature vectors output by each channel will be mapped to their respective sub-fully connected layers to obtain relatively independent high-level information expressions. Finally, the fusion of multi-channel information is achieved in an additional fully connected layer. Through this fusion process, the CNN model can effectively integrate the key information from the real sampling values, the interpolation of the mechanism model, and the process variables, output the optimal error compensation value, and then obtain a complete historical data sequence of the key process indicators.

5. A method for perceiving key process indicators in the metal smelting process driven by mechanism data according to claim 4, characterized in that The tandem model architecture is the CNN-BiLSTM-Attention tandem architecture. Probabilistic prediction is achieved through the CNN-BiLSTM-Attention tandem architecture. The tandem model architecture includes a probabilistic prediction model of a deep fusion convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism (Attention).

6. The key process index perception method for the metal melting process driven by mechanism data according to claim 5, characterized in that The method for constructing the tandem architecture is as follows: First, multiple groups of convolutional kernels are used for parallel scanning to complete the reconstruction of the feature space. With the help of depthwise separable convolution, complex local index change patterns are captured in the time dimension, and the potential correlations among multiple variables such as the temperature inside and outside the furnace and the gas flow rate are extracted. At the same time, noise interference is reduced and meaningful process features are retained. Next, the extracted high-dimensional features are input into the BiLSTM network, which accurately depicts the dynamic evolution process of the heating and cooling stages from both the forward and backward directions respectively, thus taking into account both historical states and future trends. Finally, to further highlight the decisive influence of key operation nodes on the prediction results, the model adaptively weights the output sequence of the BiLSTM through the attention mechanism, setting different importance weights in both the time and feature dimensions.

7. A method for perceiving key process indicators in the metal smelting process driven by mechanism data according to claim 6, characterized in that To achieve probabilistic prediction, in quantile regression, the tandem architecture will be trained separately for each quantile. The training objective is to minimize the average loss function \(L\) of each quantile, q which is described as: When many quantiles need to be trained, there will be a high computational burden. To reduce this computational burden, a multi-dimensional output layer is added to the model output, and the loss function is taken as the average pinball loss L of all quantiles: T: The total time T corresponding to one round of training (the total number of samples in one round of training); q: The target quantile; The q - quantile estimated at time t (at the t - th sample in one round of training); L q,t Denotes the pinball loss for the q - quantile at time t; Q: The total number of quantiles.

8. A key process index perception method for the metal smelting process driven by mechanism data according to claim 7, characterized in that To make up for the deficiencies of the pinball loss function, the Huber norm is introduced into the loss function. The Huber norm only requires a small approximation to make the loss function differentiable everywhere. Specifically, the Huber norm can be regarded as a combination of the L1 and L2 norms: where ε represents the threshold amplitude of the L1 and L2 norms. When the prediction error is below the threshold, the Huber norm is the L2 norm; when the prediction error is greater than the threshold, the Huber norm is the L1 norm, and then is substituted into the equation, and using the Huber norm, the approximate pinball loss can be calculated as:

9. The key process index perception method for the metal smelting process driven by mechanism data according to claim 8, characterized in that, To further balance the accuracy of quantile regression and the central tendency constraint, a hybrid quantile loss method is adopted, which weights and fuses the pinball loss and the mean squared error. Its loss training objective can be written as: Among them, MSE is used to constrain the prediction from diverging excessively; is used to retain the flexibility of quantile regression in interval prediction; α reg ∈[0,1] is used to balance the importance of the two; when α reg is larger, it is more inclined to reduce the overall error, that is, the interval will relatively narrow; when α reg is smaller, it pays more attention to quantile learning, that is, the interval is wider but the coverage is better.

10. A method for perceiving key process indicators in the metal smelting process driven by mechanism data according to claim 9, characterized in that, It also includes performing simulation verification on the data obtained through the above method. The specific operation of the simulation verification is as follows: Adopt R 2 and MAPE as single-value prediction performance indicators; for interval prediction, in terms of the reliability, width, comprehensive quality of the prediction interval, and the accuracy of conditional quantiles, a variety of evaluation indicators are selected for quantification. Among them, the prediction interval coverage probability (PICP) and the average coverage error (ACE) are used to measure the reliability of interval prediction, the prediction interval normalized average width (PINAW) is used to evaluate the concentration degree of interval prediction, and the pinball loss (PL) is used to reflect the accuracy of the prediction model on conditional quantiles.

Images