Machine learning based cobalt-containing scrap recycling yield prediction method

Abstract

The present application belongs to the technical field of industrial data processing, and particularly relates to a cobalt-containing waste recycling and refining yield prediction method based on machine learning, comprising the following steps: obtaining historical smelting records and historical measured yield to construct a training sample set, so as to obtain a basic prediction model through training; obtaining tapping accounting data of a previous furnace cycle, performing quality balance calculation to obtain a conservation estimate value, the conservation estimate value representing the mass fraction of residual cobalt in the slag phase; calculating the difference between the historical measured yield of the previous furnace cycle and the corresponding model predicted yield to obtain a previous residual value; and combining the on-site initial parameters of the current furnace cycle with the conservation estimate value to obtain current input features. Through the dual mechanisms of physical conservation constraint and time series statistical compensation, the present application cuts off the cross-furnace cycle transmission chain of model prediction errors from the source, significantly improving the yield prediction accuracy and control stability of the multi-batch cycle smelting period.

Description

Technical Field

[0001] This invention belongs to the field of industrial data processing technology, specifically relating to a method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning. Background Technology

[0002] With the rapid expansion of the new energy and catalyst industries, the recycling and resource utilization of complex aluminum-based nickel-cobalt-molybdenum-vanadium waste has become a crucial link in the non-ferrous metal smelting field. Currently, the hydrometallurgical-pyrometallurgical co-process is widely used for the efficient extraction of such multi-metal waste. In the multi-heat internal circulation enrichment stage of pyrometallurgical electric arc furnace smelting, the feed material is usually a mixture of fresh hydrometallurgical solid products and metal-rich intermediate slag returned from the previous heat in a certain proportion. Due to the dynamic changes in the proportion and residual metal concentration of the returned slag during the circulation process, the actual total amount of cobalt fed into each heat exhibits significant fluctuations and uncertainties. In order to achieve stable operation of the smelting system and maximize the recovery of metal resources, there is an urgent need in industrial settings for a technical solution that can accurately predict the cobalt yield of the current heat before furnace start-up, thereby providing a quantitative basis for the precise setting of core process parameters such as reducing agent ratio and smelting temperature.

[0003] To address the aforementioned requirements for parameter prediction and setting, the industry currently typically employs traditional basic prediction techniques based on single direct testing or static data-driven approaches. These conventional methods rely primarily on random sampling and testing of various feed materials before furnace start-up, or on collecting simple operational data from historical batches to construct a basic regression prediction model. When dealing with highly variable materials like return slag, traditional solutions generally directly accept localized sampling and testing values ​​from the previous furnace's waste slag, or rely on an initial model to indirectly estimate the slag phase composition of previous furnaces. This value is then used as a fixed static feature input into the current furnace's prediction system to calculate the yield estimate. Operators then passively adjust on-site process parameters based on this single result.

[0004] However, the aforementioned traditional technical solutions suffer from insurmountable technical defects in the complex multi-heat internal circulation conditions of actual operations. First, due to the extremely non-uniform physical state of the slag phase produced by electric arc furnace smelting, direct sampling and analysis often carries significant measurement random noise, resulting in severe distortion of the underlying input features characterizing the composition of the returned slag. Second, if the composition of the returned slag is indirectly estimated using a basic model, the model prediction errors generated in previous heats will be directly carried into the input of the current heat, serving as new feature bases. In the continuous multi-batch cyclic enrichment process, this unverified nested input mechanism will inevitably lead to a chain-like accumulation or malignant divergence of prediction errors. Existing technologies cannot completely sever the cross-heat transmission path of model prediction errors at the physical source, nor do they possess the adaptive and comprehensive compensation capability for the systematic fluctuations of time-series residuals. Ultimately, this often leads to the complete failure of parameter control in the later stages of the cycle, severely limiting the improvement of overall smelting yield and production efficiency. Summary of the Invention

[0005] This invention provides a machine learning-based method for predicting the yield of cobalt-containing waste recycling and refining, in order to solve the technical problem of yield prediction divergence caused by indirect estimation and error accumulation in the process of multi-batch cyclic smelting.

[0006] This invention provides a machine learning-based method for predicting the yield of cobalt-containing waste recycling and refining, comprising the following steps: Historical smelting records and historical measured yields are used to construct a training sample set in order to train a basic prediction model. Obtain the tapping calculation data from the previous furnace, perform mass balance to obtain the conservation estimate, which represents the residual cobalt mass fraction in the slag phase; calculate the difference between the historical measured yield and the corresponding model predicted yield from the previous furnace to obtain the previous residual value. The current furnace batch's initial field parameters are combined with the conserved estimated values ​​to obtain the current input features; the current input features are then input into the basic prediction model to obtain the basic prediction values. The residual correlation is calculated based on the historical accumulated previous residual values. The residual correlation characterizes the error propagation law. The error correction amount is determined based on the residual correlation and the previous residual values. The base prediction value is superimposed with the error correction amount to obtain the final prediction value, which is then output to adjust the smelting parameters of the current furnace.

[0007] By acquiring historical records to build a model, the current field parameters are combined with the conservation estimates obtained from the mass balance as input features. Then, the error correction amount calculated based on the previous residual is superimposed, which effectively removes the dynamic prediction interference in the composition of the returned slag. The static high-dimensional features and dynamic time-series errors are decoupled, so that the final predicted value of the output has extremely high fidelity, providing a reliable quantitative decision-making basis for subsequent precise adjustment of smelting parameters such as reducing agent.

[0008] Furthermore, a mass balance is performed to obtain conservation estimates, including: Obtain the total amount of cobalt fed into the furnace, the total amount of cobalt recovered, and the total mass of the slag phase from the previous furnace batch; The difference between the total amount of cobalt fed into the furnace and the total amount of cobalt recovered is calculated to obtain the amount of residual cobalt in the slag phase; The ratio of residual cobalt in the slag phase to the total mass of the slag phase is calculated to obtain a conservation estimate.

[0009] By dividing the difference between the total cobalt amount fed into the furnace and the total cobalt amount recovered by the total mass of the slag phase, the conservation estimate is established directly using the physical results of weighing and balancing at the macroscopic level. This avoids the data dispersion noise caused by uneven slag phase sampling in the electric arc furnace smelting site, ensuring that the index truly reflects the concentration of residual elements inside the slag phase, and ensuring the objective physical accuracy of subsequent parameter inputs from the bottom up.

[0010] Furthermore, by combining the initial field parameters of the current furnace batch with the conserved estimated values, the current input characteristics are obtained, including: Obtain the raw material composition, furnace charge basicity, smelting temperature, proportioning coefficient, and batch number from the initial parameters obtained on site; The raw material composition, furnace charge basicity, smelting temperature, proportioning coefficient, batch number, and conservation estimate are combined to obtain the initial feature set; The initial feature set is normalized by subtracting the mean and dividing by the standard deviation using the mean and standard deviation of historical parameters, thus obtaining the current input features.

[0011] Furthermore, a training sample set is constructed by obtaining historical smelting records and historical measured yields, including: Extract the original test records for each historical furnace batch; Based on the furnace output data of subsequent furnaces from each historical furnace, the slag phase backtracking value is obtained by reverse calculation according to the law of conservation of mass. By replacing the historical slag components in the original test records with slag phase retrospective values, a training sample set is obtained, thereby breaking the error propagation chain at the model training end.

[0012] By extracting historical original records and using the slag phase backtracking value derived from subsequent furnace output data according to the law of conservation of mass, the historical slag composition is forcibly replaced. When constructing the underlying foundation of machine learning, the initial training samples are historically corrected and denoised in advance, which effectively eliminates the systematic fitting misalignment of the model itself in the initial stage and greatly accelerates the convergence rate of the algorithm's prediction accuracy.

[0013] Furthermore, the basic prediction model is trained, including: Extract the average value of all yield data in the training sample set as the initial prediction baseline; Calculate the fitting difference between the current model prediction and the historical measured yield in each iteration round; Construct a single decision tree with the objective of minimizing the fitting difference; The output of a single decision tree is multiplied by the learning step size and then added to the current model. This process is repeated iteratively until the preset total number of trees is reached, resulting in the basic prediction model.

[0014] Furthermore, the residual correlation is calculated based on the historically accumulated prior residual values, including: Select a preset number of previous residual values ​​from history and combine the previous residual values ​​from adjacent furnaces to construct residual data pairs; Arrange the first term of each residual data pair into the first column and the second term into the second column; Calculate the expected value of the product of corresponding elements in the first and second sequences; Calculate the product of the expected value of the first sequence and the expected value of the second sequence, and use it as the basic expected product; Calculate the product of the standard deviations of the first and second series, and use it as the joint standard deviation; Calculate the difference between the expected value and the product of the basic expected values, and divide the difference by the joint standard deviation to obtain the residual correlation.

[0015] By extracting the preceding residuals of two adjacent furnaces to construct sequences, calculating the expected product and difference of the two sequences and dividing by the joint standard deviation, and through rigorous Pearson statistical analysis, the originally discrete single fluctuation sequence is mapped into an autocorrelation time series matrix structure, which can sensitively quantify and capture the hidden same-direction accumulation or opposite-direction oscillation patterns between errors of adjacent furnaces, providing solid mathematical theoretical support for the implementation of intervention strategies.

[0016] Furthermore, based on the residual correlation and the preceding residual value, the error correction amount is determined, including: If the absolute value of the residual correlation is greater than the correlation threshold, it indicates that there is systematic error propagation in the same or opposite direction. The initial correction is obtained by multiplying the preceding residual value by the residual correlation. The absolute value of the initial correction is limited to the upper limit of the preset proportion of the absolute value of the preceding residual, and the original calculation sign is retained to obtain the error correction amount.

[0017] Furthermore, based on the residual correlation and the preceding residual value, the error correction amount is determined, including: If the absolute value of the residual correlation is not greater than the correlation threshold, it indicates that the preceding residual value is random noise without any systematic transmission pattern. The value 0 is defined as the error correction amount.

[0018] Furthermore, after obtaining and outputting the final predicted value, the process also includes: Obtain the new measured yield and the new calculation residual after the current output calculation; The current input features, the final predicted value, the new measured yield, and the new accounting residual are combined into an incremental sample. In response to the absolute value of the newly calculated residual for a single furnace being greater than a preset multiple of the average absolute value of the historical residuals, incremental samples are added to the training sample set for rolling retraining to obtain an updated prediction model.

[0019] By setting a hard trigger condition that the absolute value deviates from the historical average by a preset multiple, new accounting data that meets the fluctuation anomaly are combined into incremental samples for rolling retraining. Under the premise of filtering out conventional process disturbances, the system achieves agile capture of the essential variation law of the furnace material and rapid self-healing evolution of the model weights, ensuring the long-term effectiveness and adaptability of the production algorithm base from the perspective of the entire life cycle.

[0020] Furthermore, to adjust the smelting parameters for the current furnace cycle, including: Calculate the expected difference between the preset target yield and the final predicted value; If the expected difference is greater than the set warning value, it indicates that the amount of reducing agent added is insufficient; Increase the preset percentage of reducing agent ratio as the adjusted smelting parameters.

[0021] The beneficial effects are as follows: This invention introduces macroscopic quality balancing methods, using the total output data to inversely calculate the conservation estimate of slag phase residue, thus replacing direct laboratory values ​​to construct the input characteristics of the current furnace batch. This cuts off the historical error propagation chain caused by indirect prediction at the data input end. Based on this, the scheme performs autocorrelation calculations on the residual historical sequence of previous fluctuations, extracting the residual correlation degree characterizing the error propagation law, thereby adaptively targeting error superposition or silent suppression of the predicted values ​​output by the basic prediction model. This dual constraint mechanism based on physical conservation correction and temporal statistical compensation successfully overcomes the engineering bottleneck of systematic divergence in parameter prediction caused by local disturbances in complex multi-batch return slag recycling smelting scenarios, thereby significantly improving the accuracy and robustness of single-batch furnace prediction and the long-term stable cyclic control capability. Attached Figure Description

[0022] Figure 1 This is a flowchart of a machine learning-based method for predicting the yield of cobalt-containing waste recycling and refining.

[0023] Figure 2 The graph shows the evolution of the absolute error in cobalt yield prediction with each furnace cycle. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] This technical solution is applicable to the scenario of cobalt yield prediction and parameter control in the multi-heat internal circulation enrichment stage of a complex resource recovery process for aluminum-based nickel-cobalt-molybdenum-vanadium waste using both wet and pyrometallurgical methods. In multi-heat cyclic production in an electric arc furnace smelting workshop, the feed material for each heat is typically a mixture of fresh wet solid products and metal-rich intermediate slag returned from the previous heat in a specific ratio. Since the residual cobalt content in the returned slag dynamically changes with multiple cycles, on-site operators find it difficult to directly and accurately grasp the actual cobalt content of each heat, resulting in a lack of quantitative scientific basis for setting process parameters such as reducing agent ratio and smelting temperature. This invention, by outputting a high-precision cobalt yield prediction value before each heat is started, can effectively support the on-site optimization of various process parameters of the furnace charge in advance, thereby avoiding heat efficiency losses due to redundancy or insufficiency of parameters such as reducing agent.

[0026] An embodiment of the cobalt-containing waste recycling and refining yield prediction method based on machine learning provided by this invention: like Figure 1 As shown, the machine learning-based method for predicting the yield of cobalt-containing waste recycling includes the following steps: S101. Obtain historical smelting records and historical measured yields to construct a training sample set in order to train the basic prediction model.

[0027] In one embodiment, to break the chain of data error propagation at the source of model training and ensure high fidelity of input features, original test records from each historical furnace are extracted. Based on the output data of subsequent furnaces from each historical furnace, a reverse calculation is performed according to the law of conservation of mass to obtain slag phase retrospective values. These slag phase retrospective values ​​are then used to replace the historical slag components in the original test records to obtain a training sample set, thereby breaking the error propagation chain at the model training end. It should be noted that in conventional industrial settings, directly sampling and testing the slag phase is often limited by uneven sampling and carries uncontrollable measurement random noise. Introducing macroscopic material mass balance from the subsequent furnace to infer the objective physical data of the previous furnace can significantly improve the physical authenticity of the sample data. For example, if the historical slag composition value is 5.2% in the original test records of the first batch of historical furnaces, and the slag phase retrospective value obtained by reverse calculation based on the total weight of the furnace output and the weight of the recovered material from the second batch of historical furnaces is 6.0%, then when constructing the training sample set, 6.0% is forcibly used to replace the original 5.2%. Preferably, the minimum value of the total number of samples in the training sample set is data from 30 historical furnaces, to ensure that the subsequent machine learning algorithm has at least feature generalization ability.

[0028] After constructing a highly reliable training sample set, in order to establish a prediction carrier capable of mapping the nonlinear laws of complex pyrometallurgical smelting, the average value of all yield data within the training sample set is extracted as the initial prediction benchmark. The fitting difference between the current model's predicted value and the historical measured yield is calculated in each iteration. A single decision tree is constructed with the goal of minimizing this fitting difference. The output of each decision tree is multiplied by the learning step size and accumulated into the current model, and this process is iterated until the preset total number of trees is reached, resulting in the basic prediction model. Understandably, the above tree construction and iteration process uses the standard gradient boosting regression tree construction algorithm. When constructing a single decision tree, the evaluated target uses the standard mean squared error loss function, eliminating the need to construct non-standard complex networks, which fully suits the training characteristics of small-sample tabular data in industrial applications.

[0029] The training process involves several hyperparameters that determine the model structure. All hyperparameters are optimized from historical data using a grid search algorithm based on five-fold cross-validation. Specifically, exemplary reference values ​​for the following hyperparameters are provided: an exemplary reference value of 0.05 for the learning step size, an exemplary reference value of 100 for the preset total number of trees, an exemplary reference value of 4 layers for the maximum depth of a single decision tree, and an exemplary reference value of 3 samples for the minimum number of leaf nodes in a single decision tree.

[0030] By reconstructing and correcting historical test inputs through strict physical conservation logic, and in conjunction with the iterative fitting mechanism of the standard integrated tree model, the systematic characterization bias carried by the historical data itself is effectively removed. From the perspective of engineering practice, this provides a high-precision and robust algorithmic foundation for online yield prediction of the current furnace.

[0031] S102, obtain the furnace output calculation data of the previous furnace, perform mass balance to obtain the conservation estimate value, the conservation estimate value characterizes the residual cobalt mass fraction in the slag phase; calculate the difference between the historical measured yield of the previous furnace and the corresponding model predicted yield to obtain the previous residual value.

[0032] In one embodiment, to obtain objective benchmark data unaffected by prior prediction errors and to accurately quantify the actual deviation of the current prediction model, it is necessary to acquire production data from the previous furnace run using physical weighing and on-site testing methods. This involves acquiring the total cobalt input, total cobalt recovery, and total slag mass of the previous furnace run; calculating the difference between the total cobalt input and total cobalt recovery to obtain the residual cobalt content in the slag phase; and calculating the ratio of the residual cobalt content to the total slag mass to obtain a conservation estimate. To strictly adhere to the macroscopic law of conservation of mass to characterize the true cobalt retention level within the slag phase and to avoid sampling bias caused by direct slag phase testing, the conservation estimate satisfies the following relationship:

[0033] In the formula, This represents the conservative estimate. This indicates the total amount of cobalt fed into the furnace. This indicates the total amount of cobalt recovered. This represents the total mass of the slag phase. When obtaining the above variables, the measurement accuracy of the weighing equipment used to weigh the alloy phase and slag phase is preferably better than 0.5% of the total mass. The corresponding instrument for component analysis is preferably an inductively coupled plasma atomic emission spectrometer (ICP-AES), with a component detection accuracy preferably better than 0.1%. The above numerical ranges are determined based on conventional industrial verification procedures and cost-benefit balance points. Furthermore, when the absolute value of the difference between the conserved estimated value and the measured value of the slag phase composition obtained from direct analysis of the same batch in the previous heat exceeds 1.5%, a forced material balance verification is triggered. This preset trigger threshold of 1.5% is calibrated based on the 95th percentile data of the difference between the two values ​​in historical abnormal heats.

[0034] Understandably, the conservation estimate is positively correlated with the total cobalt input and negatively correlated with the total recovered cobalt and the total mass of the slag phase. Physically, this means that, given a constant total amount of effective elements input, the more effective components extracted into the alloy phase during smelting, or the larger the final volume of waste slag, the lower the concentration of residual effective elements carried in the unit waste slag. By using this macroscopic mass balance to replace single, localized sampling and testing, the physical carrier of indirect estimation errors in the model is completely eliminated from the input end, establishing an absolute physical truth level correction basis for subsequent steps.

[0035] When determining the preceding residual, the historical measured yield of the previous batch is directly subtracted from the model-predicted yield of the current batch, and the difference is recorded. If the result is positive, it indicates that the model has systematically underestimated the output of the previous batch; if it is negative, it indicates that the prediction has been overestimated.

[0036] By combining on-site weighing, testing and other physical measurement methods to perform global quality balance, and simultaneously and accurately calculating the numerical prediction residuals of the previous production cycle node, a solid data support is provided for the purification of model input feature sources and dynamic statistical compensation of errors in the next production cycle. This effectively prevents the prediction error from unidirectionally diverging and accumulating along multiple batch cycle sequences.

[0037] S103, combine the initial field parameters of the current furnace batch with the conserved estimated values ​​to obtain the current input features; input the current input features into the basic prediction model to obtain the basic prediction values.

[0038] In one embodiment, in order to transform the physical state and process settings of the current production batch into a standard data structure that can be recognized by machine learning algorithms, and to achieve effective dimensionality reduction and feature alignment of multidimensional heterogeneous parameters, the raw material composition, furnace charge basicity, smelting temperature, proportioning coefficient, and batch number in the initial parameters are obtained. The raw material composition, furnace charge basicity, smelting temperature, proportioning coefficient, batch number, and conservation estimates are merged to obtain an initial feature set. The initial feature set is normalized by subtracting the mean and dividing by the standard deviation using the mean and standard deviation of historical parameters to obtain the current input features.

[0039] It should be noted that the raw material composition specifically includes the mass fractions of cobalt, nickel, molybdenum, and vanadium in the furnace feed. The content of these substances is obtained through online analysis by pre-testing equipment. The basicity of the furnace charge is defined as the ratio of the mass of basic flux to the mass of acidic components. The proportioning coefficient characterizes the percentage of the reducing agent mass in the total mass of the furnace feed. The batch number is assigned sequentially according to the actual process position of the current furnace in this cycle. When constructing the initial feature set, this initial feature set is represented in data form as a multi-dimensional column vector composed of concatenated data. For the aforementioned multi-dimensional heterogeneous parameters with different physical dimensions and scaling scales, the standard Z-Score data normalization algorithm is used as the normalization method in this scheme. This method maps each feature dimension to a dimensionless space of standard normal distribution based on conventional data scaling and translation standard equations. This algorithm is a well-known standard data preprocessing technique in this field, and it can eliminate dimensional interference between various features without specific mathematical derivation.

[0040] For example, the preferred numerical parameter configuration for the initial parameters on site is as follows: the mass fraction of cobalt in the raw material is 2.5%, nickel is 1.8%, molybdenum is 0.5%, and vanadium is 0.3%; the basicity of the furnace charge is 1.2; the smelting temperature is 1450 degrees Celsius; the proportioning coefficient is 8.0%; and the batch number is 2. Furthermore, the hyperparameters required for the normalization process, namely the mean and standard deviation of the historical parameters of each feature, are calculated in the preceding steps based on the conventional statistical distribution characteristics of the same-dimensional feature fields of the full training samples, and are independently and persistently stored and retrieved by the database.

[0041] After completing the morphological transformation from high-dimensional vectors to low-dimensional standard space and obtaining the current input features, these features are input into the pre-constructed basic prediction model. Through the node threshold determination and branch convergence mechanism within the tree network model, the nonlinear combination logic of multidimensional complex feature variables is ultimately collapsed into a single-dimensional output result, thereby outputting an objective basic prediction value that is not contaminated by historical prediction errors.

[0042] By physically mapping and dimensionlessly standardizing the multi-dimensional heterogeneous on-site operating parameters with the conserved parameters after quality balance and debiasing, a high-fidelity, pure, and uncontaminated panoramic operating condition feature combination was successfully constructed. With the help of the nonlinear analytical base of the integrated regression model, the current production baseline expected output was quickly generated, providing a high-precision decision-making basis for subsequent prediction residual superposition and process optimization control from the engineering execution level.

[0043] S104. Calculate the residual correlation based on the historically accumulated preceding residual values. The residual correlation characterizes the error propagation pattern. Determine the error correction amount based on the residual correlation and the preceding residual values.

[0044] In one embodiment, to accurately capture the hidden transmission characteristics of prediction errors over time during multi-furnace cyclic production, it is necessary to perform autocorrelation analysis on the statistical dimension using a historical data matrix to extract regular indicators reflecting fluctuations before and after. A predetermined number of previous residual values ​​are selected from the historical data, and the previous residual values ​​of adjacent furnaces are combined to construct residual data pairs. The first term of each residual data pair is arranged as a first column, and the second term as a second column. The expected value of the product of corresponding elements in the first and second columns is calculated. The product of the expected value of the first and second columns is calculated as the basic expected product. The product of the standard deviations of the first and second columns is calculated as the joint standard deviation. The difference between the expected value and the basic expected product is calculated, and the difference is divided by the joint standard deviation to obtain the residual correlation.

[0045] It should be noted that the above calculation process essentially transforms a discrete single-wave sequence into a structured space with corresponding relationships between different time periods, thereby fully exposing the potential chain-like propagation of errors. Therefore, the residual correlation satisfies the following relationship:

[0046] In the formula, Indicates the degree of residual correlation. Indicates the expected value. Let represent the fundamental expected product, where Let be the expected value of the first sequence. Let be the expected value of the second sequence. This represents the joint standard deviation, where Let be the standard deviation of the first sequence. Let be the standard deviation of the second sequence. In engineering control practice, to balance the confidence requirements of statistical data with the need for rapid response to new operating conditions in industrial settings, the historical preset quantity is set to 6. Understandably, the residual correlation is positively correlated with the difference between the expected value and the product of the basic expected values, and negatively correlated with the joint standard deviation. This physical relationship indicates that when the residual changes in two adjacent batches are highly consistent, the absolute value of the correlation approaches the extreme value, suggesting that the prediction error in actual production exhibits a systematic unidirectional divergence or alternating oscillation, requiring the introduction of a targeted correction mechanism.

[0047] In another embodiment, for operating conditions with significant autocorrelation characteristics, when the absolute value of the residual correlation is greater than the correlation threshold, it indicates that there is systematic error propagation in the same or opposite direction; the product of the preceding residual value and the residual correlation is calculated to obtain the initial correction amount; the absolute value of the initial correction amount is limited to a preset upper limit of the absolute value of the preceding residual value, and the original calculation sign is retained to obtain the error correction amount.

[0048] In order to adaptively and dynamically feedforward compensate the benchmark prediction signal and quantify the actual proportion of the previous residual affecting the current furnace, the initial correction amount satisfies the following relationship:

[0049] In the formula, This represents the initial correction amount. Represents the preceding residual value. This represents the residual correlation. Understandably, the initial correction amount is positively correlated with both the preceding residual value and the residual correlation. The engineering implication is that the larger the historically accumulated residual base or the stronger its systematic chain correlation, the greater the required compensation intervention. Simultaneously, to avoid excessive intervention leading to control logic divergence under extreme boundary conditions, hard constraint boundaries must be established. For example, the correlation threshold is set at 0.3, a value strictly calibrated based on the lower bound of the confidence interval of the statistical Pearson correlation coefficient with a sample size of 6; the preset upper limit of the proportion is set at 80%, designed to forcibly interrupt abnormal correction signals exceeding the historical maximum stable amplitude.

[0050] In another embodiment, for error fluctuation scenarios exhibiting completely random walk characteristics, if the absolute value of the residual correlation is not greater than the correlation threshold, it indicates that the preceding residual value is random noise without any systematic transmission pattern; the value 0 is determined as the error correction amount. Understandably, when the prediction error manifests as chaotic white noise, any forcibly superimposed correction signal will introduce additional nonlinear perturbations. Directly setting the correction channel to zero and blocking it allows the overall prediction logic to smoothly degenerate to the original output state of the physical underlying model. This exit mechanism ensures the robustness of the algorithm under stable melting conditions.

[0051] By using standard statistical tools, the underlying coupling mechanism of error propagation between multiple production batches was deeply quantified and refined. Combined with a control strategy of dynamically assigning adaptive amplification compensation or silent exit under hard threshold conditions, the global prediction distortion problem caused by small biases within the model in complex long-cycle physical cycles was fundamentally mitigated.

[0052] S105, the basic predicted value is superimposed with the error correction amount to obtain the final predicted value and output it, so as to adjust the smelting parameters of the current furnace.

[0053] In one embodiment, to form a closed-loop system logic from single prediction deduction to continuous iterative optimization, the new measured yield and new calculation residual after the current furnace output are obtained; the current input features, final predicted value, new measured yield, and new calculation residual are combined into incremental samples; in response to the absolute value of the new calculation residual of a single furnace being greater than a preset multiple of the average absolute value of historical residuals, the incremental samples are added to the training sample set for rolling retraining to obtain an updated prediction model. It should be noted that the above mechanism constructs a data flywheel for the self-evolving weights of the model's underlying layers. By setting a trigger threshold for residual mutations, the algorithm can both quickly respond to and capture essential regular mutations in the composition of new batches of materials and effectively filter reasonable physical fluctuations under normal smelting conditions. For example, the preferred reference value for the preset multiple of the average absolute value of historical residuals is 2.0, which is calibrated based on the boundary extreme value of the normal distribution of errors from the entire historical sample set. When the retraining condition is triggered, the old data is updated using the incremental samples containing the latest real physical response mapping, thereby continuously strengthening the long-term fitting ability of the algorithm on a unified conservation accounting benchmark.

[0054] In another embodiment, in order to effectively convert the preceding multidimensional computational data into control actions of real physical entities in the industrial field to adjust the smelting parameters of the current furnace, the following steps are taken: calculating the expected difference between the preset target yield and the final predicted value; in response to the expected difference being greater than the set warning value, indicating that the amount of reducing agent added is insufficient; and increasing the reducing agent ratio by a preset percentage as the adjusted smelting parameters.

[0055] In order to clearly quantify the hard trigger threshold for process parameter adjustment, the expected difference is to satisfy the following relationship:

[0056] In the formula, Indicates the expected difference. Indicates the preset target yield. This represents the final predicted value.

[0057] Understandably, the expected difference is positively correlated with the preset target yield and negatively correlated with the final predicted value. The engineering and physical significance of this formula is that when the predicted output, obtained through rigorous calculation, is far lower than the process design requirements, it indirectly confirms that the current furnace reducing atmosphere concentration cannot break the chemical bond constraints of the waste metal ions. Immediate intervention and compensation must be made by forcibly adding reducing media such as carbonaceous auxiliaries. For example, combining actual acceptance standards in industrial settings, the preset target yield is set at 90.0%, the warning value at 3.0%, and the preset percentage at 0.5%. When the calculated expected difference reaches 4.0%, the on-site operation instruction to strictly increase the reducing agent ratio by 0.5% is executed, thereby blocking the inefficient smelting process.

[0058] By superimposing the basic expectations of static high-dimensional feature mapping with the residual compensation of dynamic temporal patterns, and simultaneously establishing an instant physical feedback channel for production materials and auxiliary materials and a self-healing evolution channel for the underlying core algorithm at the end, the efficient, stable, precise, controllable and continuous convergence of complex industrial smelting processes in multiple furnaces is guaranteed from the entire life cycle dimension of engineering control.

[0059] like Figure 2 The figure shows a dynamic comparison of the absolute error in cobalt yield prediction using the method of this invention and existing methods over 20 consecutive furnace cycles. The horizontal axis represents the furnace number, and the vertical axis represents the absolute error in prediction. As can be seen from the figure, the existing technology, due to its reliance on local sampling and lack of error feedback, exhibits a significant fluctuating and accumulating characteristic in its prediction error as the number of cycles increases. In contrast, the method of this invention, benefiting from real-time compensation of the correlation between the conserved estimate and the residual, keeps the error at a low level throughout, and the average prediction accuracy is significantly improved compared to the existing technology, verifying the stability of this scheme under long-cycle operating conditions.

[0060] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning, characterized in that, Includes the following steps: Historical smelting records and historical measured yields are used to construct a training sample set in order to train a basic prediction model. Obtain the tapping calculation data from the previous furnace, perform mass balance to obtain the conservation estimate, which represents the residual cobalt mass fraction in the slag phase; calculate the difference between the historical measured yield and the corresponding model predicted yield from the previous furnace to obtain the previous residual value. The current furnace batch's initial field parameters are combined with the conserved estimated values ​​to obtain the current input features; the current input features are then input into the basic prediction model to obtain the basic prediction values. The residual correlation is calculated based on the historical accumulated previous residual values. The residual correlation characterizes the error propagation law. The error correction amount is determined based on the residual correlation and the previous residual values. The base prediction value is superimposed with the error correction amount to obtain the final prediction value, which is then output to adjust the smelting parameters of the current furnace.

2. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, The conservation estimates are obtained by performing mass balance calculations, including: Obtain the total amount of cobalt fed into the furnace, the total amount of cobalt recovered, and the total mass of the slag phase from the previous furnace batch; The difference between the total amount of cobalt fed into the furnace and the total amount of cobalt recovered is calculated to obtain the amount of residual cobalt in the slag phase; The ratio of residual cobalt in the slag phase to the total mass of the slag phase is calculated to obtain a conservation estimate.

3. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, By combining the initial field parameters of the current furnace batch with the conserved estimated values, the current input characteristics are obtained, including: Obtain the raw material composition, furnace charge basicity, smelting temperature, proportioning coefficient, and batch number from the initial parameters obtained on site; The raw material composition, furnace charge basicity, smelting temperature, proportioning coefficient, batch number, and conservation estimate are combined to obtain the initial feature set; The initial feature set is normalized by subtracting the mean and dividing by the standard deviation using the mean and standard deviation of historical parameters, thus obtaining the current input features.

4. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, A training sample set was constructed by obtaining historical smelting records and historical measured yields, including: Extract the original test records for each historical furnace batch; Based on the furnace output data of subsequent furnaces from each historical furnace, the slag phase backtracking value is obtained by reverse calculation according to the law of conservation of mass. By replacing the historical slag components in the original test records with slag phase retrospective values, a training sample set is obtained, thereby breaking the error propagation chain at the model training end.

5. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, The basic prediction model is obtained through training, including: Extract the average value of all yield data in the training sample set as the initial prediction baseline; Calculate the fitting difference between the current model prediction and the historical measured yield in each iteration round; Construct a single decision tree with the objective of minimizing the fitting difference; The output of a single decision tree is multiplied by the learning step size and then added to the current model. This process is repeated iteratively until the preset total number of trees is reached, resulting in the basic prediction model.

6. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, The residual correlation is calculated based on the historical accumulated prior residual values, including: Select a preset number of previous residual values ​​from history and combine the previous residual values ​​from adjacent furnaces to construct residual data pairs; Arrange the first term of each residual data pair into the first column and the second term into the second column; Calculate the expected value of the product of corresponding elements in the first and second sequences; Calculate the product of the expected value of the first sequence and the expected value of the second sequence, and use it as the basic expected product; Calculate the product of the standard deviations of the first and second series, and use it as the joint standard deviation; Calculate the difference between the expected value and the product of the basic expected values, and divide the difference by the joint standard deviation to obtain the residual correlation.

7. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, Based on the residual correlation and the preceding residual values, the error correction amount is determined, including: If the absolute value of the residual correlation is greater than the correlation threshold, it indicates that there is systematic error propagation in the same or opposite direction. The initial correction is obtained by multiplying the preceding residual value by the residual correlation. The absolute value of the initial correction is limited to the upper limit of the preset proportion of the absolute value of the preceding residual, and the original calculation sign is retained to obtain the error correction amount.

8. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, Based on the residual correlation and the preceding residual values, the error correction amount is determined, including: If the absolute value of the residual correlation is not greater than the correlation threshold, it indicates that the preceding residual value is random noise without any systematic transmission pattern. The value 0 is defined as the error correction amount.

9. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, After obtaining and outputting the final predicted value, the following is also included: Obtain the new measured yield and the new calculation residual after the current output calculation; The current input features, the final predicted value, the new measured yield, and the new accounting residual are combined into an incremental sample. In response to the absolute value of the newly calculated residual for a single furnace being greater than a preset multiple of the average absolute value of the historical residuals, incremental samples are added to the training sample set for rolling retraining to obtain an updated prediction model.

10. The method for predicting the yield of cobalt-containing waste recycling and refining based on machine learning according to claim 1, characterized in that, To adjust the smelting parameters for the current furnace, including: Calculate the expected difference between the preset target yield and the final predicted value; If the expected difference is greater than the set warning value, it indicates that the amount of reducing agent added is insufficient; Increase the preset percentage of reducing agent ratio as the adjusted smelting parameters.

Images