Method for predicting content of accessory product based on zinc oxide production

Abstract

The invention discloses an accessory product content prediction method and system based on zinc oxide production, and relates to the technical field of intelligent monitoring, and the method comprises the steps: collecting process parameters in real time through a sensor, constructing digital twin platform parallel operation mechanism simulation, and obtaining simulation characteristics; constructing a main random forest model, and calculating the resampling membership degree of the sample according to the fuzzy rule of the process parameters; when each node of the decision tree is split, recalculating a working condition label in real time according to the digital twin output of the current node; and based on digital twinborn analysis simulation features, preferentially selecting features having significant influence on the affiliated products during node segmentation, constructing a lightweight residual random forest for secondary correction, and outputting a final affiliated product content prediction value. According to the method, high-precision and low-delay prediction of the accessory products is realized. And the prediction stability and sensitivity under multiple working conditions, nonlinearity and abrupt change states are obviously improved. And the method has better identification capability in key working conditions and risk boundary sections.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to a method for predicting the content of ancillary products based on zinc oxide production. Background Art

[0002] With the rapid development of non-ferrous metallurgy and new materials industries, zinc oxide, as an important chemical raw material and functional material, has been widely used in many industrial fields such as electronics, ceramics, rubber, etc. The production process of zinc oxide mainly includes direct method, indirect method and wet method, etc. The process involves high-temperature gas-solid reaction, multiphase flow and complex energy and mass transfer. In the actual production process, due to the dynamic changes of parameters such as raw material composition, furnace temperature fluctuations, atmosphere environment, feed rate, etc., a variety of by-products (such as sulfides, carbonates, chlorides, etc.) are often generated. These by-products not only affect the product purity and recovery rate of zinc oxide, but also have a significant impact on the stable operation of subsequent process equipment, energy consumption control and environmental protection indicators. Therefore, how to achieve accurate prediction of the content of by-products in the zinc oxide production process is a key link in improving product quality, optimizing process parameters and realizing intelligent production.

[0003] Currently, mainstream byproduct monitoring and prediction technologies in the industry fall into two main categories: physical and chemical simulation based on mechanistic models, and data-driven statistical or machine learning methods. Mechanistic models rely on the principles of thermodynamics, kinetics, and fluid mechanics to infer the generation of various products by establishing mathematical models of the furnace reaction system. However, due to practical limitations such as parameter acquisition, model calibration, and high computational complexity, they struggle to adapt to the online, real-time prediction needs of production lines. Data-driven methods utilize large amounts of historical production data and laboratory analysis results to train soft sensors or regression models for content estimation. While this improves prediction efficiency to a certain extent, it faces challenges such as poor sample balance, insufficient extrapolation capabilities, and poor model interpretability. At the same time, with the development of emerging technologies such as digital twins, edge computing, and parallel simulation, a growing number of studies are attempting to integrate physical mechanisms with data-driven approaches to improve model generalization and prediction accuracy. However, most of the existing integration methods remain at the level of "step-by-step series connection" or "simple compensation", and have not yet achieved deep coupling of working condition partitioning and model training. It is difficult to capture complex behaviors such as nonlinearity, strong disturbances, and multiple working condition alternations in zinc oxide production, and the prediction stability and robustness under key boundary conditions are still obviously insufficient. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing technology for predicting the content of by-products in the zinc oxide production process generally has problems such as poor real-time performance, limited generalization ability, weak adaptability to complex working conditions, and insufficient early warning sensitivity in key sections.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a method for predicting the content of ancillary products in zinc oxide production, comprising:

[0007] By collecting process parameters in real time through sensors, a digital twin platform is built to run mechanism simulation in parallel to obtain simulation features;

[0008] Build a main random forest model and calculate the resampling membership of samples based on the fuzzy rules of process parameters. When each node of the decision tree splits, recalculate the working condition label in real time based on the digital twin output of the current node and guide the node splitting.

[0009] Based on the digital twin analysis and simulation features, the features that have a significant impact on the subsidiary products are prioritized when splitting the nodes;

[0010] The main random forest model is used to calculate the prediction residuals of the training samples, and a lightweight residual random forest is constructed for secondary correction to output the final prediction value of the ancillary product content.

[0011] As a preferred embodiment of the method for predicting the content of by-products based on zinc oxide production according to the present invention, the construction of a digital twin platform for parallel operation of mechanism simulation includes collecting data on furnace temperature, gas composition, feed rate, pressure, and humidity from field sensors to form field sensor characteristics;

[0012] The digital twin platform is divided into a data layer, a simulation layer, a fusion layer, and a presentation layer. The data layer is responsible for storing on-site sensor data and historical test results. The simulation layer includes an operation mechanism simulation module to perform furnace flow field calculations and multiphase chemical reaction kinetics calculations. The fusion layer is responsible for comparing and calibrating the simulation layer output with the real-time data from the data layer and outputting it to the upper-level scheduling or control logic. The presentation layer is responsible for providing a visual interface.

[0013] By running the mechanism simulation in parallel on the digital twin platform, the digital twin simulation characteristics are obtained; the simulation characteristics specifically include local temperature field, reaction rate field, and concentration field information at the gas-solid interface.

[0014] As a preferred embodiment of the method for predicting the content of by-products based on zinc oxide production described in the present invention, the fuzzy rules include introducing fuzzy membership rules defined by expert experience when constructing the Bootstrap sample set of each decision tree, calculating the membership of each training sample in the high-risk and mutation areas, and dynamically adjusting the weight of the training sample drawn online to the tree node training set.

[0015] Based on production experience and historical data, define the ranges for furnace temperature and oxygen content;

[0016] For each sample, read its furnace temperature and oxygen content values, and calculate its membership under the two membership functions of furnace temperature and oxygen function respectively; according to the preset weight ratio, add the two memberships weightedly to obtain the original fuzzy weight of the sample; smooth the original fuzzy weights of all samples and normalize them proportionally in all samples to obtain the sampling probability of each sample.

[0017] As a preferred embodiment of the method for predicting the content of by-products based on zinc oxide production described in the present invention, the method includes recalculating the operating condition label based on the digital twin output of the current node, comprising: when constructing a training set for each decision tree, extracting a number of samples equal to the total sample size through weighted random sampling with replacement based on the sampling probability to obtain a sample subset of the decision tree; at the root node, collecting all subset samples, and calling the digital twin platform and expert-defined rules to label each sample with an operating condition; and preliminarily grouping the samples in the root node into multiple operating condition groups based on the labeling results;

[0018] After the root node samples are labeled, sampling probabilities are assigned to all features based on the feature sensitivity calculated by digital twin analysis.

[0019] As a preferred embodiment of the method for predicting the content of by-products based on zinc oxide production according to the present invention, the node segmentation is guided by extracting a feature subset from all features according to sampling probability for use in segmentation of the current node; analyzing the sample size and historical experience data of each working condition grouping to determine the working condition category to be prioritized for the current segmentation; sequentially trying different segmentation points for the selected candidate feature subsets, and calculating the error change before and after the segmentation on all samples of the current node;

[0020] Before training begins, set the minimum error drop threshold as the basis for division;

[0021] Select the feature and split value combination that can maximize the node purity, use only the latest annotated working condition information of the current node, and divide the current node into two child nodes, left and right, each carrying half of the samples; before splitting, evaluate the out-of-bag error level of the current node in the unsplit state; combine the out-of-bag errors of the two child nodes obtained after the current node is split into the overall error after splitting; compare the difference between the error before splitting and the overall error after splitting;

[0022] If the error reduction caused by the split is lower than the minimum error reduction threshold, the split is considered invalid, further splitting of the child node is abandoned, the current node is marked as a leaf node, and the average value of all training sample labels of the current node is used as the predicted output of the leaf node; for each child node that has not been stopped, the cycle of working condition labeling, feature extraction, optimal splitting, and convergence detection is repeated;

[0023] In the child node, if the working condition of the sample is different from that of the parent node, the child node will re-call the digital twin and expert rules to label the sample with a new working condition label;

[0024] When all leaf nodes trigger the error reduction stop judgment, the entire tree growth ends, and the output mean of each leaf node is recorded as the decision tree's prediction of the output.

[0025] As a preferred embodiment of the method for predicting the content of ancillary products based on zinc oxide production according to the present invention, the digital twin-based analysis and simulation feature includes using the digital twin to perform a small perturbation on each input feature, recording the difference in product content before and after the perturbation, and obtaining a sensitivity index corresponding to each feature;

[0026] The sensitivity values ​​of all features are normalized into weights, and features with high sensitivity have a higher probability in subsequent sampling; when each tree grows, a fixed-proportion subset is randomly extracted from all features according to the weight, and the best split is only found within the subset; the sensitivity of each feature is regularly re-evaluated in the digital twin environment, and the sensitivity weight of the feature is updated.

[0027] As a preferred embodiment of the method for predicting the content of ancillary products based on zinc oxide production according to the present invention, the lightweight residual random forest is constructed, which includes using the main random forest model to predict all training samples and recording the predicted value of each sample; calculating the residual for each sample to obtain a residual label set; the residual includes the true label minus the main model prediction value;

[0028] The original feature vectors and residual labels are used together to train a lightweight residual random forest model. During online prediction, the main random forest model is used to make a prediction on the new sample to obtain a preliminary estimate.

[0029] The same feature vector is input into the residual random forest model to obtain the correction value of the main model error; the preliminary estimate and the correction value are added to output the final prediction result.

[0030] As a preferred solution of the system for predicting the content of ancillary products based on zinc oxide production described in the present invention, it includes: a data acquisition module, a digital twin platform, an online reasoning module, and an online calibration module;

[0031] The data acquisition module is used to collect original process parameters from on-site sensors in real time, and obtain the content labels of regularly tested ancillary products from the database and transmit them to the digital twin platform;

[0032] The digital twin platform is used to call a parallel mechanism simulation engine to simulate the gas-solid two-phase flow, heat transfer and chemical reaction mechanism in the furnace, and output the internal physical quantities of temperature field, concentration field and reaction rate;

[0033] The online inference module is used to train the adaptive random forest model, obtain the latest fusion features in real time, and call the trained main forest and residual forest to complete the by-product content prediction;

[0034] The online calibration module is used to compare the simulation results with the field measured data and dynamically adjust the simulation parameters according to the errors.

[0035] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for predicting the content of ancillary products based on zinc oxide production.

[0036] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for predicting the content of ancillary products based on zinc oxide production.

[0037] Beneficial effects of the present invention: The method for predicting the content of by-products in zinc oxide production provided by the present invention achieves high-precision, low-latency prediction of the content of by-products in the zinc oxide production process through the deep integration of the digital twin platform and the random forest model. The system can use the process parameters and simulation features collected in real time to adaptively adjust the sampling and modeling strategies, significantly improving the prediction stability and sensitivity under multiple working conditions, nonlinearity and mutation states. The introduction of fuzzy rules and feature sensitivity weights enables the model to have better recognition capabilities in key working conditions and risk boundary sections. The residual forest is used for secondary correction to further reduce errors and improve the robustness and reliability of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 This is an overall flow chart of a method for predicting the content of by-products based on zinc oxide production provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0040] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0041] Example 1, with reference to Figure 1 , as one embodiment of the present invention, provides a method for predicting the content of ancillary products based on zinc oxide production, comprising:

[0042] S1: Collect process parameters in real time through sensors, build a digital twin platform to run mechanism simulation in parallel, and obtain simulation features.

[0043] Furthermore, the construction of the digital twin platform and the parallel operation mechanism simulation include collecting the furnace temperature, gas composition, feed rate, pressure, and humidity data of the field sensors to form the field sensor characteristics. The formula of the original feature vector of the field sensor is expressed as:

[0044]

[0045] Among them, x1 represents the furnace temperature, x2 represents the CO concentration, and x m Represents the feed rate (kg / h). Digital twin simulation output feature vector:

[0046]

[0047] Where s1 represents the local temperature field, s2 represents the node reaction rate, and s k Represents the gas-solid interface concentration. Fusion eigenvectors are expressed as:

[0048]

[0049] Real label (content of by-product obtained by testing):

[0050]

[0051] After interpolating missing values ​​and removing 3σ anomalies for both x and s, Z-Score standardization is performed:

[0052]

[0053] Training set D = {(z (n) ,y (n) )|n=1,…,N}.

[0054] Field sensor characteristics include furnace temperature, concentration of various components in the reducing atmosphere (such as CO, O2, etc.), feed rate, furnace pressure, humidity and other process parameters that are easy to collect in real time on the production line.

[0055] For example, "furnace temperature" is expressed in numerical form as the instantaneous temperature value of the furnace on that day, and "CO concentration" is expressed as the percentage of carbon monoxide in the reducing atmosphere.

[0056] These parameters are automatically acquired through factory-installed temperature sensors, gas analyzers, material flow metering devices, etc., and pushed to the data acquisition system at fixed time intervals (such as every second or every minute).

[0057] Based on the digital twin platform, the physical and chemical processes in the furnace under current operating conditions are simulated to obtain some implicit contents that cannot be directly measured by on-site sensors but are crucial to the generation of ancillary products.

[0058] Typical outputs include: local temperature distribution at certain key locations inside the furnace; instantaneous reaction rate in the reaction area (such as the conversion rate of reactants per unit volume); and concentration field information at the gas-solid interface (such as the concentration distribution over a period of time on the contact surface between zinc oxide and gas).

[0059] These values ​​are quickly calculated by the simulation software based on real-time sensor input and output to the data platform within a predetermined clock (such as every minute or every ten minutes).

[0060] The digital twin platform is divided into data layer, simulation layer, fusion layer and display layer; the data layer is responsible for storing on-site sensor data and historical test results; the simulation layer includes an operation mechanism simulation module to perform furnace flow field calculations and multiphase chemical reaction kinetics calculations; the fusion layer is responsible for comparing and calibrating the simulation layer output with the real-time data of the data layer, and outputting it to the upper-level scheduling or control logic; the display layer is responsible for providing a visual interface.

[0061] By running the mechanism simulation in parallel on the digital twin platform, the digital twin simulation characteristics are obtained; the simulation characteristics specifically include local temperature field, reaction rate field, and concentration field information at the gas-solid interface.

[0062] It should be noted that for each training sample, the historical mean of the feature is subtracted and then divided by its historical standard deviation. After processing, all features are centered around zero and scaled with unit variance to ensure comparability across all dimensions within the model. Digital twin features are also normalized. Although the simulation output of a digital twin is inherently a continuous variable, its range of values ​​often differs from that of sensor data. Therefore, the same normalization principle is applied: the historical mean of the digital twin feature is subtracted and divided by its standard deviation to ensure that the scale of the entire fused feature vector remains consistent.

[0063] S2: Based on the digital twin analysis simulation features, the features with significant impact on the subsidiary products are given priority when splitting the nodes.

[0064] Furthermore, the digital twin platform was used to conduct local perturbation experiments on each dimension of the training set. That is, for a certain characteristic value (such as furnace temperature, temperature at a certain point or reaction rate, gas-solid interface concentration, etc.), the value of the characteristic was slightly increased and decreased in the digital twin model, and the changes in the byproduct content in the simulation output were observed.

[0065] For each feature, we perform multiple similar small perturbation experiments, recording the differences in product content caused by these perturbations. By summarizing the absolute values ​​of these differences, we determine the "average sensitivity" of each feature. This step aims to determine how much the predicted product content fluctuates when a slight change in a feature occurs, thus serving as a basis for measuring the "importance" of that feature. Features with larger values ​​in the sensitivity assessment indicate a greater impact on product content; features with smaller values ​​are relatively less important.

[0066] Based on the average sensitivity of each feature, it is normalized and converted into a "sampling weight." This normalization process can be simply understood as first concentrating the sensitivity values ​​of all features into the same magnitude range, and then allocating a sampling probability proportionally to them. This way, features with high sensitivity are assigned a greater probability in subsequent random selections, while features with low sensitivity are assigned a smaller probability.

[0067] During each tree growth process in a random forest, a subset of features is randomly selected from all available features. The optimal split point is then found based on this subset. Traditionally, this involves uniformly random selection of all features; however, here we employ a "weighted random selection" approach.

[0068] Specifically, for a given decision tree to be generated, several features are randomly selected from the entire feature set according to the previously calculated sampling weight distribution. The total number of features selected accounts for a preset proportion of the total number of features (for example, 30% of the total number of features). Because features with higher sensitivity have a greater weight, they are naturally more likely to be selected and, accordingly, have more opportunities to participate in node splitting.

[0069] When splitting each node, the above-mentioned "weighted random selection" step is first performed to obtain a feature subset (usually containing several features that are strongly related to the product content).

[0070] Then, within this feature subset, different splitting schemes are tried, selecting the one that maximizes node purity. Because highly weighted features are more likely to be included in the subset, they more frequently become split candidates. Therefore, the entire tree prioritizes physical quantities that contribute significantly to the prediction of ancillary products.

[0071] Digital twin local sensitivity, for the i-th normalized feature Make small perturbations ±δ on the digital twin platform and simulate and predict the ancillary products Changes:

[0072]

[0073] where e i is the i-th unit vector, and δ can be taken as a relative perturbation of 5%.

[0074] Take the average or median of all samples to get the overall sensitivity of feature i

[0075]

[0076] Sensitivity weight, which defines the weighting factor w of the i-th feature i :

[0077]

[0078] Among them, ε represents the prevention of situations (such as for non-critical features).

[0079] Calculate the sampling probability of feature subsets, Normalize and define the probability P of the i-th feature being selected i :

[0080]

[0081] Each tree needs to be selected from {1,…,p} according to the distribution {P i Randomly extract [αp] features with replacement Among them, α∈(0,1) is a hyperparameter.

[0082] The digital twin-based analysis and simulation feature includes using the digital twin to make a small disturbance to each input feature, recording the difference in product content before and after the disturbance, and obtaining a sensitivity index corresponding to each feature.

[0083] The sensitivity values ​​of all features are normalized into weights, and features with high sensitivity have a higher probability in subsequent sampling; when each tree grows, a fixed-proportion subset is randomly extracted from all features according to the weight, and the best split is only found within the subset; the sensitivity of each feature is regularly re-evaluated in the digital twin environment, and the sensitivity weight of the feature is updated.

[0084] It should be noted that as production conditions change, the actual importance of certain features may drift. Therefore, sensitivity assessments should be performed regularly within the digital twin model (e.g., weekly or after each major adjustment) to obtain the latest feature sensitivity values. These updated sensitivity values ​​are converted into new sampling weights, allowing all subsequently generated decision trees to automatically adapt to the new feature distribution and impact relationships under the new operating conditions.

[0085] S3: Build the main random forest model and calculate the resampling membership of the samples based on the fuzzy rules of the process parameters; when each node of the decision tree splits, recalculate the working condition label in real time based on the digital twin output of the current node, and guide the node splitting.

[0086] Furthermore, the fuzzy rules include introducing fuzzy membership rules defined by expert experience when constructing the Bootstrap sample set of each decision tree, calculating the membership of each training sample in the high-risk and mutation areas, and dynamically adjusting the weight of the training sample being drawn into the tree node training set online.

[0087] First, based on the industry experts’ experience that “high temperature or oxygen-deficient environment will lead to a sharp increase in by-products” during zinc oxide production, membership functions were defined for furnace temperature and oxygen content respectively.

[0088] When the furnace temperature is lower than a certain threshold, it is considered to be the "low temperature zone"; when the furnace temperature is higher than another threshold, it belongs to the "high temperature zone"; the transition zone between them is considered to be the "intermediate temperature".

[0089] Similarly, for the oxygen content in the furnace, when the content is lower than a certain lower limit, it is considered as an "oxygen-deficient zone", when it is higher than a certain upper limit, it is considered as an "oxygen-rich zone", and the content in between is considered as "neutral oxygen content".

[0090] These two membership functions map the originally monotonously continuous temperature and oxygen content inputs into language variables such as "low / medium / high" and "lack / medium / rich", thereby reflecting the fuzzy judgment that "the higher the temperature and the lower the oxygen content, the greater the risk of by-products".

[0091] For each historical sample, first read its corresponding furnace temperature value and oxygen content value, and then use the membership function defined in the previous step to calculate the membership of the sample under the two unfavorable working conditions of "high temperature" and "lack of oxygen" respectively.

[0092] Based on the expert-defined ratio, the "high temperature membership" and "hypoxia membership" are weighted and added together to obtain the sample's fuzzy weight. The larger the weight, the closer the sample is to the hazardous conditions of "high temperature and / or hypoxia," and thus should be prioritized during resampling.

[0093] Based on production experience and historical data, define the ranges for furnace temperature and oxygen content.

[0094] According to expert experience, two fuzzy membership functions are constructed for temperature x1 and oxygen content x3 (x3 represents O2 concentration):

[0095]

[0096] Where T min ,T max are the threshold intervals of “low temperature → high temperature”, O min ,O max They are respectively the threshold range of "O2 low → high".

[0097] Define the comprehensive fuzzy weight. Expert experience suggests that "at high temperatures and (high CO / low O2) the amount of by-products increases dramatically." Therefore, the fuzzy weight gain coefficient for sample n is constructed:

[0098]

[0099] Among them, ω T ,ω O >0 is the weight constant assigned by the expert. Indicates the membership of the nth sample as “at high temperature”. Indicates the degree of “being in hypoxia”.

[0100] If α (n) The higher it is, the closer the sample is to the “critical boundary condition” and a higher probability is required during sampling.

[0101] Calculate the final resampling probability for all standardized samples First calculate the original weight α (n) .

[0102] To avoid extreme values, the following smoothing normalization can be performed:

[0103]

[0104] Among them, λ controls the degree of “weight prominence”.

[0105] Finally, when Bootstrap sampling, the probability α′ is selected from D (n) Take N samples (repeated samples are allowed) to form the training subset of the cost tree.

[0106] For each sample, read its furnace temperature and oxygen content values, and calculate its membership under the two membership functions of furnace temperature and oxygen function respectively; according to the preset weight ratio, add the two memberships weightedly to obtain the original fuzzy weight of the sample; smooth the original fuzzy weights of all samples and normalize them proportionally in all samples to obtain the sampling probability of each sample.

[0107] The recalculation of the operating condition label based on the digital twin output of the current node includes: when constructing the training set of each decision tree, based on the sampling probability, weighted random sampling with replacement is performed to extract the same number of samples as the total sample size to obtain a sample subset of the decision tree; at the root node, all subset samples are collected, and the digital twin platform and expert-defined rules are called to label the operating condition of each sample; based on the labeling results, the samples in the root node are preliminarily classified into multiple operating condition groups.

[0108] After the root node samples are labeled, sampling probabilities are assigned to all features based on the feature sensitivity calculated by digital twin analysis.

[0109] The guidance for node splitting includes extracting a feature subset from all features according to sampling probability for use in splitting the current node; analyzing the sample size and historical experience data of each working condition grouping to determine the working condition category that is prioritized for the current splitting; for the selected candidate feature subset, trying different splitting points in turn, and calculating the error change before and after the split on all samples of the current node.

[0110] Before training begins, set the minimum error drop threshold as the basis for division.

[0111] Select the feature and split value combination that can maximize the node purity, only use the latest annotated working condition information of the current node, and divide the current node into two child nodes on the left and right, each carrying half of the samples; before splitting, evaluate the out-of-bag error level of the current node in the unsplit state; merge the out-of-bag errors of the two child nodes obtained after the current node is split into the overall error after splitting; compare the difference between the error before splitting and the overall error after splitting.

[0112] If the error reduction brought by the split is lower than the minimum error reduction threshold, it is judged as an invalid split, and further splitting of the child node is abandoned. The current node is marked as a leaf node, and the average value of all training sample labels of the current node is used as the predicted output of the leaf node; for each child node that has not been stopped, the cycle of working condition labeling, feature extraction, optimal splitting, and convergence detection is repeated.

[0113] In the child node, if the working condition of the sample is different from that of the parent node, the child node will re-call the digital twin and expert rules to label the sample with a new working condition label.

[0114] When all leaf nodes trigger the error reduction stop judgment, the entire tree growth ends, and the output mean of each leaf node is recorded as the decision tree's prediction of the output.

[0115] The working condition classification function C(z) defines a differentiable "working condition partition function":

[0116] r=C(z),r∈{1,2,…,R}

[0117] Where r represents the operating condition category number (such as "normal", "high temperature and partial oxygen", "low temperature and partial oxygen deficiency", etc.).

[0118] C(·) can be determined by the digital twin simulation output and expert rules. The formula for C(z) is:

[0119]

[0120] where τ T ,τ O ,τ CO are the thresholds obtained by digital twin simulation.

[0121] Dynamic coupled tree growth process For the growth process of the t-th tree, D is no longer divided into r categories at one time in advance; instead, when each node splits, it is divided into r categories in real time based on the current node sample. The digital twin output is used to recalculate the working condition labels and guide the segmentation accordingly.

[0122] Generated using Bootstrap resampling

[0123] Initialize the Root node so that the current node sample set

[0124] For any non-leaf node, "real-time partitioning" is performed on each sample in the node. Call Get the latest working condition number r of the sample (n) .

[0125] The samples in the node are divided into

[0126] Node split candidate features The extraction is done by randomly sampling from {1,…,p} according to the sensitivity weighting (as shown in the following section). Features, composition

[0127] Prioritize the working conditions and observe the Sample size The r with the largest sample size is given priority for secondary segmentation; if they are similar, the next segmentation step is determined by comprehensively considering the "significance of the working condition" and the degree of failure of the previous working condition segmentation.

[0128] Find the best split point for each candidate feature and its possible segmentation threshold θ, calculate the segmentation pair The mean square error (MSE) gain is selected from all (j,θ) combinations. * ,θ * )=argmaxΔMSE(j,θ).

[0129] Press the node sample and Divided into Left and Right child nodes

[0130] It should be noted that an adaptive resampling method, which incorporates expert experience fuzzy rules on top of the traditional random forest static bootstrap method, allows samples from critical boundary conditions to be selected into the training subset with a higher probability. This significantly improves the model's sensitivity to sudden changes in product content under the critical conditions of "high temperature and hypoxia" and prioritizes capturing these sample features that are critical to prediction accuracy during training.

[0131] S4: Use the main random forest model to calculate the prediction residuals of the training samples, construct a lightweight residual random forest for secondary correction, and output the final prediction value of the ancillary product content.

[0132] Furthermore, the construction of the lightweight residual random forest includes using the main random forest model to predict all training samples and record the predicted value of each sample; calculating the residual for each sample to obtain a residual label set; the residual includes the true label minus the main model prediction value.

[0133] The standardized feature vector of each sample in the training set is fed into the trained primary random forest model. Based on the growth structure of each tree, the model sends the sample to the corresponding leaf node of each tree and aggregates the average values ​​of these leaf nodes to obtain a preliminary prediction for the sample. At this point, the primary random forest has learned most of the nonlinear relationships using the features and labels of the training set. However, due to factors such as model complexity and sample noise, some errors may still exist.

[0134] For each sample in the training set, the difference between the sample's true label value (i.e., the amount of the byproduct obtained by analysis) and the initial prediction from the primary random forest is compared. This difference is called the "residual" or "error." These residuals reflect information not captured by the primary model and are often associated with more subtle patterns or noise. Collecting the residuals for each sample forms a new set of "error labels."

[0135] The main RF predicts that for each trained tree t, any new sample Starting from the root node, the samples are divided into dynamic partitions and dropped to a leaf node.

[0136] The output of the tree is the average of the leaf nodes L t .

[0137] The prediction formula for the entire main forest (T trees in total) is expressed as:

[0138]

[0139] Residual forest training, for all training samples First, the main RF gives a prediction Then calculate the residual:

[0140]

[0141] use Construct a lightweight residual random forest RFres, the number of trees is usually T res <<T.

[0142] Residual forests can also adopt the three major mechanisms of "dynamic coupling", "fuzzy weighted sampling" and "convergence-aware pruning", but due to their smaller scale, they only need to ensure speed and a certain degree of accuracy in local error correction.

[0143] The final comprehensive output is for the new sample First get the main RF prediction Then we get from the residual forest

[0144] The final output formula is expressed as:

[0145]

[0146] The original feature vector and the residual label are used together to train a lightweight residual random forest model; during online prediction, the main random forest model is used to predict the new sample once to obtain a preliminary estimate.

[0147] The same feature vector is input into the residual random forest model to obtain the correction value of the main model error; the preliminary estimate and the correction value are added to output the final prediction result.

[0148] It should be noted that the original normalized feature vectors, along with the residual labels just calculated, are then fed into another random forest model for training. Because the labels are now "residual values," the newly trained random forest can specifically learn the error components that the primary model failed to fit. To ensure a more efficient and faster residual model, the residual forest is typically configured with a smaller number of trees and a smaller depth, making it more lightweight and reducing computational overhead while ensuring effective correction.

[0149] Example 2 is an embodiment of the present invention, which provides a method for predicting the content of by-products based on zinc oxide production. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0150] First of all, this embodiment takes the actual production line of a zinc oxide production enterprise as the background, and conducts a systematic test on the real-time prediction problem of the content of ancillary products (such as ZnS, ZnCO3, etc.). The experiment first collects key process parameters (such as furnace temperature, gas composition, feed rate, pressure, humidity, etc.) in seconds through a multi-point intelligent sensor network, and all data are automatically imported into the factory data center. A digital twin platform is established simultaneously, which includes a data layer, a simulation layer, a fusion layer and a display layer. The data layer stores the process data and historical test results collected by the sensor in real time. The simulation layer integrates CFD flow field simulation and multiphase chemical kinetics model, and can output simulation features such as local temperature field, reaction rate field, gas-solid interface concentration online. The fusion layer compares and calibrates the output of the simulation layer with the real data, and the display layer provides an intuitive monitoring and decision-making interface for production line operators and technical management personnel.

[0151] During the model training phase, all process parameters and simulation features are first normalized and subjected to sensitivity analysis. By performing perturbation simulations on a single variable, the impact of each feature on changes in byproduct content is quantified, resulting in a sensitivity index. Highly sensitive features are then assigned higher sampling weights, serving as the basis for subsequent random forest feature sampling probabilities. During the sample resampling phase, the probability of sampling high-risk areas for each sample is dynamically calculated based on the fuzzy membership intervals for furnace temperature and oxygen content defined by expert rules, combined with production experience and historical data. This creates a fuzzy weight distribution, further enhancing the modeling capabilities for sudden operating conditions.

[0152] During training, each decision tree adopts a dynamic working condition partitioning mechanism, that is, when each node is split, the digital twin simulation and expert rules are called in real time to regroup the current node samples, and the splitting features and splitting thresholds are determined according to the changes in working conditions. Node splitting gives priority to samples with a large number of samples or samples at the boundary of working conditions to ensure that the model has good generalization capabilities under multiple working conditions. The out-of-bag error is evaluated before and after each node split. If the error caused by the split is less than the preset threshold, the split is stopped. After the training of the entire main model is completed, all training samples are predicted and the residuals are calculated. The lightweight residual random forest model is trained based on the original features and residuals to achieve automatic correction of the main model error. Finally, during online real-time inference, the system first uses the main model to make a prediction, and then uses the residual model to correct it, and outputs the final high-precision prediction value of the byproduct content.

[0153] After feature sensitivity analysis and weight assignment guided by the digital twin, the main random forest model has significantly outperformed traditional data-driven methods in predicting the content of byproducts in each production batch. For example, for batches such as T04 and T07, where the operating temperature is relatively high and the fuzzy sampling weight is large, the main model predictions are highly consistent with the measured values. For areas sensitive to process fluctuations, the dynamic adjustment of the fuzzy sampling weights significantly increases the number of samples under high-risk working conditions, effectively improving the model's fitting ability in key areas. When abnormal fluctuations occur in the feed rate or oxygen content, such as in batch T06, the model dynamically adjusts the working condition labels by calling the digital twin simulation output in real time, achieving timely adaptation and segmentation decisions for complex working conditions, and avoiding the accuracy loss caused by data averaging in conventional models.

[0154] Further analysis of the residual correction link shows that in all batches, the prediction residuals of the main model are generally controlled within ±0.04%, and the lightweight residual model effectively compensates for the minor errors of the main model under small local sample sizes or abnormal working conditions. The final output of the prediction result of the content of the subsidiary product has an average absolute error as low as within 0.01%, which is much better than the industry's conventional level of 0.05%. Overall, the innovations in the embodiment are reflected in: First, the quantitative weighting of feature sensitivity guided by digital twins greatly improves the model's adaptability and interpretability to complex and multi-working conditions; second, the dynamic resampling of high-risk area samples based on fuzzy rules enables the model to show extremely strong capture capabilities at process mutation points, overcoming the insensitivity to boundary conditions under traditional uniform sampling; third, combined with the secondary correction mechanism of the residual random forest, targeted compensation for the weaknesses of the main model is achieved, significantly reducing the overall prediction error.

[0155] Example 3 is an embodiment of the present invention, which provides a system for predicting the content of ancillary products based on zinc oxide production, including a data acquisition module, a digital twin platform, an online reasoning module, and an online calibration module.

[0156] The data acquisition module is used to collect original process parameters from on-site sensors in real time, and obtain the content labels of regularly tested ancillary products from the database and transmit them to the digital twin platform.

[0157] The digital twin platform is used to call the parallel mechanism simulation engine to simulate the gas-solid two-phase flow, heat transfer and chemical reaction mechanism in the furnace, and output the internal physical quantities of temperature field, concentration field and reaction rate.

[0158] The online inference module is used to train the adaptive random forest model, obtain the latest fusion features in real time, and call the trained main forest and residual forest to complete the by-product content prediction.

[0159] The online calibration module is used to compare the simulation results with the field measured data and dynamically adjust the simulation parameters according to the errors.

[0160] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0161] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0162] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0163] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logical functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc. It should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to be limiting. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications should be encompassed by the claims of the present invention.

[0164] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting the content of ancillary products based on zinc oxide production, characterized in that: include: By collecting process parameters in real time through sensors, a digital twin platform is built to run mechanism simulation in parallel to obtain simulation features; Based on the digital twin analysis and simulation features, the features that have a significant impact on the subsidiary products are prioritized when splitting the nodes; The main random forest model is constructed to calculate the resampled membership of samples according to the fuzzy rules of process parameters; When each node in the decision tree is split, the working condition label is recalculated in real time based on the digital twin output of the current node, and the node splitting is guided; The main random forest model is used to calculate the prediction residuals of the training samples, and a lightweight residual random forest is constructed for secondary correction to output the final prediction value of the ancillary product content.

2. The method for predicting the content of ancillary products based on zinc oxide production according to claim 1, wherein: The construction of the digital twin platform for parallel operation of the mechanism simulation includes collecting furnace temperature, gas composition, feed rate, pressure, and humidity data from field sensors to form field sensor characteristics; The digital twin platform is divided into data layer, simulation layer, fusion layer and display layer; The data layer is responsible for storing on-site sensor data and historical test results. The simulation layer includes an operation mechanism simulation module to perform furnace flow field calculations and multiphase chemical reaction kinetics calculations. The fusion layer is responsible for comparing and calibrating the simulation layer output with the real-time data from the data layer and outputting it to the upper-level scheduling or control logic. The presentation layer is responsible for providing a visual interface. Run mechanism simulations in parallel on the digital twin platform to obtain digital twin simulation features; The simulation features specifically include local temperature field, reaction rate field, and concentration field information at the gas-solid interface.

3. The method for predicting the content of ancillary products based on zinc oxide production according to claim 2, wherein: The fuzzy rules include introducing fuzzy membership rules defined by expert experience when constructing the Bootstrap sample set of each decision tree, calculating the membership of each training sample in the high-risk and mutation areas, and dynamically adjusting the weight of the training sample drawn online to the tree node training set. Based on production experience and historical data, define the ranges for furnace temperature and oxygen content; For each sample, read its furnace temperature and oxygen content values, and calculate its membership under the two membership functions of furnace temperature and oxygen respectively; according to the preset weight ratio, add the two memberships weightedly to obtain the original fuzzy weight of the sample; The original fuzzy weights of all samples are smoothed and normalized proportionally among all samples to obtain the sampling probability of each sample.

4. The method for predicting the content of ancillary products based on zinc oxide production according to claim 3, wherein: The digital twin-based analysis of simulation features includes using the digital twin to make a small perturbation to each input feature, recording the difference in product content before and after the perturbation, and obtaining a sensitivity index corresponding to each feature; The sensitivity values ​​of all features are normalized into weights, and features with high sensitivity have a higher probability in subsequent sampling; when each tree grows, a fixed-proportion subset is randomly extracted from all features according to the weight, and the best split is only found within the subset; the sensitivity of each feature is regularly re-evaluated in the digital twin environment, and the sensitivity weight of the feature is updated.

5. The method for predicting the content of ancillary products based on zinc oxide production according to claim 4, wherein: Recalculating the operating condition label based on the digital twin output of the current node includes: when constructing the training set of each decision tree, based on the sampling probability, weighted random sampling with replacement is performed to extract the same number of samples as the total sample size to obtain a sample subset of the decision tree; at the root node, all subset samples are collected, and the digital twin platform and expert-defined rules are called to label the operating condition of each sample; According to the labeling results, the samples in the root node are preliminarily classified into multiple working condition groups; After the root node samples are labeled, sampling probabilities are assigned to all features based on the feature sensitivity calculated by digital twin analysis.

6. The method for predicting the content of ancillary products based on zinc oxide production according to claim 5, wherein: The node splitting guidance includes extracting a feature subset from all features according to sampling probability for use in the splitting of the current node; analyzing the sample size and historical experience data of each working condition grouping to determine the working condition category that should be prioritized for the current splitting; for the selected candidate feature subset, sequentially trying different splitting points, and calculating the error change before and after the splitting on all samples of the current node; Before training begins, set the minimum error drop threshold as the basis for division; Select the feature and split value combination that can maximize the purity of the node. Only the latest annotated working condition information of the current node is used. The current node is divided into two child nodes, left and right, each carrying half of the samples. Before splitting, the out-of-bag error level of the current node in the unsplit state is evaluated. The out-of-bag errors of the two child nodes obtained after the current node is split are combined into the overall error after splitting. Compare the difference between the error before segmentation and the overall error after segmentation; If the error reduction caused by the split is lower than the minimum error reduction threshold, the split is considered invalid, further splitting of the child node is abandoned, the current node is marked as a leaf node, and the average value of all training sample labels of the current node is used as the predicted output of the leaf node; for each child node that has not been stopped, the cycle of working condition labeling, feature extraction, optimal splitting, and convergence detection is repeated; In the child node, if the working condition of the sample is different from that of the parent node, the child node will re-call the digital twin and expert rules to label the sample with a new working condition label; When all leaf nodes trigger the error reduction stop judgment, the entire tree growth ends, and the output mean of each leaf node is recorded as the decision tree's prediction of the output.

7. The method for predicting the content of ancillary products based on zinc oxide production according to claim 6, wherein: The lightweight residual random forest construction includes using the main random forest model to predict all training samples and recording the predicted value of each sample; Calculate the residual for each sample to obtain a residual label set; the residual includes the true label minus the main model prediction value; The original feature vectors and residual labels are used together to train a lightweight residual random forest model. During online prediction, the main random forest model is used to make a prediction on the new sample to obtain a preliminary estimate. The same feature vector is input into the residual random forest model to obtain the correction value of the main model error; the preliminary estimate and the correction value are added to output the final prediction result.

8. A system using the method for predicting the content of by-products based on zinc oxide production according to any one of claims 1 to 7, characterized in that: Includes data acquisition module, digital twin platform, online reasoning module, and online calibration module; The data acquisition module is used to collect original process parameters from on-site sensors in real time, and obtain the content labels of regularly tested ancillary products from the database and transmit them to the digital twin platform; The digital twin platform is used to call a parallel mechanism simulation engine to simulate the gas-solid two-phase flow, heat transfer and chemical reaction mechanism in the furnace, and output the internal physical quantities of temperature field, concentration field and reaction rate; The online inference module is used to train the adaptive random forest model, obtain the latest fusion features in real time, and call the trained main forest and residual forest to complete the by-product content prediction; The online calibration module is used to compare the simulation results with the field measured data and dynamically adjust the simulation parameters according to the errors.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for predicting the content of by-products based on zinc oxide production according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting the content of by-products based on zinc oxide production according to any one of claims 1 to 7 are implemented.

Images