A method and system for identifying bottleneck processes in a flotation full flow

Abstract

The embodiment of the application provides a flotation full-process bottleneck process identification method and system. After operation data is acquired, the method can extract feasibility features, robustness features and sensitivity features from the operation data. Then, evaluation indexes such as a matching degree index and an adjustability index are calculated based on the extracted operation features. Then, a PI-PNN index qualitative prediction model based on a prior model is constructed according to the operation data and domain knowledge, and a bottleneck process variable or a bottleneck process combination is identified by calculating an index marginal contribution value of the flotation process based on the established prediction model. The method can effectively predict the qualitative influence of the process parameters on the process quality index, reduce the cost generated in the actual field invalid debugging process, make the improvement process have a basis, and reduce the blindness of the solution.

Description

Technical Field

[0001] This application relates to the field of mineral processing technology, and in particular to a method and system for identifying bottleneck processes in the entire flotation process. Background Technology

[0002] Flotation, or flotation for short, is a mineral processing technique that separates mineral particles based on differences in the physicochemical properties of their surfaces. The flotation process utilizes differences in the wettability of mineral surfaces; by adding flotation reagents, mineral particles selectively adhere to air bubbles, thus achieving mineral separation. Taking the flotation process of a copper mine as an example, after crushing and grinding, the copper ore particles reach a suitable flotation size. Then, collectors, frothers, and other flotation reagents are added to form a slurry with a specific concentration and pH value. This slurry is then fed into a flotation machine, where copper mineral particles are floated to the froth by air bubbles, yielding a rough concentrate. This rough concentrate is then refined 2-3 times to obtain a copper concentrate. The copper concentrate obtained from flotation is then concentrated using a thickener to remove excess water, and further dewatered using a filter to obtain a dry copper concentrate, improving copper recovery and grade.

[0003] Due to the accumulation of minerals and the influence of the storage environment, ores are prone to oxidation. Oxidized ores produce hydrophilic substances, thus reducing their floatability. For example, in the flotation process of copper ore, oxidized copper ore mainly consists of malachite, with small amounts of pseudomalachite, azurite, and chalcopyrite, while iron minerals mainly include magnetite and maghemite, and also contain a very small amount of native silver-bearing gold ore. In the direct flotation method for oxidized copper ore, the altered properties of the original ore and the increased composition of oxidized ore lead to the formation of more hydrophilic substances, further reducing the mineral's floatability.

[0004] To adapt to the impact of mineral oxidation processes on flotation, it is necessary to identify bottleneck factors that may change under different operating conditions or disturbances, and to develop modification plans based on these identified bottlenecks. However, due to the variable properties of oxidation feedstocks, the matching degree between actual production processes and operations, as well as the diagnostic difficulty of adjustment, are insufficient. Specifically, in actual industrial processes, the mineralogy characteristics of the raw ore, such as composition and particle size distribution, exhibit significant fluctuations and uncertainties. Coupled with differences in operators, this directly affects the stability of the flotation process and the effective control of process parameters. Furthermore, relying heavily on experience-based judgment or single-indicator evaluation lacks systematicity and objectivity, making it difficult to meet the needs of process diagnosis. This results in insufficient ability to identify process bottleneck variables, inadequate consideration of the dynamic correlation characteristics between variables, and an inability to accurately capture bottleneck factors that may change under different operating conditions or disturbances, thus affecting the accuracy and effectiveness of modification plans. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and system for identifying bottleneck processes in the entire flotation process, in order to solve the problem of insufficient accuracy and effectiveness of the modification scheme.

[0006] According to one aspect of this application, a method for identifying bottleneck processes in the entire flotation process is provided, the method comprising:

[0007] Acquire operational data and extract operational features from the operational data, the operational data including operational variables recorded throughout the entire flotation process; the operational features include feasibility features, robustness features, and sensitivity features;

[0008] Evaluation indicators are calculated based on the operational characteristics, including the calculation matching degree index and the adjustability index of the operational sample under ideal operational conditions.

[0009] A predictive model is established based on the operational data and process domain knowledge. The predictive model is a qualitative predictive model of indicators based on physical information assisted by a prior model and a product-based neural network. The predictive model includes a front-end network and a back-end network. The front-end network is used to fit key state parameters. The back-end network is used to perform quality indicator level classification.

[0010] Based on the prediction model, bottleneck process data is identified by quantitatively assessing the marginal contribution of process variables or process combinations in the indicator prediction. The bottleneck process data includes at least one of bottleneck process variables and bottleneck process combinations.

[0011] In some embodiments, extracting operational features from the operational data includes:

[0012] The operation subspace is divided according to the operation variables in the operation data, and the operation subspace includes an ideal operation space and a non-ideal operation space.

[0013] Calculate the ideal probability of the operation subspace;

[0014] The number of ideal spaces and the number of non-empty subspaces are counted, where the number of ideal spaces is the number of subspaces with an ideal probability of 1.

[0015] The feasibility characteristics are obtained by calculating the ratio of the number of ideal spaces to the number of non-empty subspaces.

[0016] In some embodiments, dividing the operation subspace according to the operation variables in the operation data includes:

[0017] Extract the operational variables and the process information associated with the operational variables from the operational data;

[0018] The operational variables are divided into multiple variable groups according to the process information, wherein the operational variables in the same variable group form a joint variable.

[0019] The joint space formed by the joint variables is divided into a predetermined number of subspaces;

[0020] Calculate the benefit value of the subspace;

[0021] If the benefit value reaches the preset benefit standard, the subspace is marked as an ideal operating space;

[0022] If the benefit value does not meet the preset benefit standard, the subspace is marked as a non-ideal operating space.

[0023] In some embodiments, extracting operational features from the operational data includes:

[0024] Calculate the information entropy of the operation subspace;

[0025] The number of uncertain subspaces is counted, where the uncertain subspace is the subspace whose information entropy is greater than or equal to a preset entropy threshold.

[0026] The robustness feature is obtained by calculating the ratio of the number of uncertain subspaces to the number of non-empty subspaces.

[0027] In some embodiments, extracting operational features from the operational data includes:

[0028] Obtain the flotation ore grade information corresponding to the operation data;

[0029] The Least Squares Support Vector Machine (LSSVRM) algorithm model is invoked, which is used to fit the nonlinear relationship function between the operational data and the flotation ore grade information.

[0030] Based on the KS sensitivity test, the conditional cumulative distribution parameters are obtained using the LSSVRM algorithm model, and the unconditional cumulative distribution parameters are obtained using the operational data.

[0031] Calculate a one-dimensional sensitivity value based on the conditional cumulative distribution parameter and the unconditional cumulative distribution parameter;

[0032] The sensitivity feature is obtained by calculating the sum of the one-dimensional sensitivity values ​​of the joint variables.

[0033] In some embodiments, calculating evaluation metrics based on the operational characteristics includes:

[0034] A score vector is generated based on the operational characteristics. The score vector includes a normal cycle score vector and an abnormal cycle score vector. The normal cycle score vector includes cycles in which the efficiency value meets production requirements. The abnormal cycle score vector includes cycles in which the efficiency value is lower than production requirements.

[0035] The evaluation index is calculated based on the score vector;

[0036] The matching threshold and the difficulty adjustment threshold are obtained. The matching threshold is a threshold obtained through experiments using a prior model. The difficulty adjustment threshold is a preset value.

[0037] By comparing the matching degree index and the matching threshold, the matching status of operation and process is generated;

[0038] By comparing the adjustability index and the adjustment difficulty threshold, the relative operational difficulty of the adjustment operation and the original operation is generated.

[0039] In some embodiments, a predictive model is established based on the operational data and process domain knowledge, including:

[0040] Based on the operational data and prior knowledge of the process, a prior model of key state information for the entire process is established. The input data of the prior model are operational data and process variables, and the output data of the prior model are measurable key state parameters.

[0041] The relationship function between input data and key state information is extracted using the multivariate Taylor function fitting method through the prior model.

[0042] The relation function is set as the loss function of the front-end network of the prediction model, and the front-end network of the prediction model is trained based on the loss function;

[0043] The distortion parameters of the prediction in the front-end network of the prediction model are replaced with the predicted values ​​of the prior model.

[0044] The predicted values ​​are input into the backend network of the prediction model, so that the backend network trains the prediction model within a product-based neural network framework.

[0045] In some embodiments, a predictive model is established based on the operational data and process domain knowledge, including:

[0046] Obtain the SHAP value and SHAP interaction value of the prediction model input; the SHAP value is used to characterize the contribution of a single process variable to the prediction result; the SHAP interaction value is used to characterize the contribution of multiple process variables to the prediction result.

[0047] The bottleneck process is determined based on the SHAP value and SHAP interaction value.

[0048] Obtain the contribution parameters of different bottleneck processes to the final prediction result. The contribution parameters include the contribution value of a single process variable to the prediction result, and / or the joint contribution value of multiple process variables to the prediction result.

[0049] The subsequent modification targets are determined based on the contribution parameters. The subsequent modification targets include process variables whose SHAP values ​​are higher than a preset SHAP value threshold, and / or process combinations whose SHAP interaction values ​​are higher than a preset SHAP interaction value threshold.

[0050] In some embodiments, based on the prediction model, bottleneck variables are identified by quantitatively assessing the marginal contribution of process variables or process combinations in indicator prediction, including:

[0051] The interaction parameters of variables between process combinations are calculated based on the SHAP interaction value, and the interaction parameters are either synergistic interaction parameters or antagonistic interaction parameters.

[0052] Identify key bottleneck variables in the process flow based on the aforementioned operational parameters;

[0053] Calculate the priority of modification for the key bottleneck variables.

[0054] According to another aspect of this application, a system for identifying bottleneck processes in the entire flotation process is provided, the system comprising:

[0055] The feature extraction module is used to acquire operational data and extract operational features from the operational data, which includes operational variables recorded throughout the entire flotation process; the operational features include feasibility features, robustness features, and sensitivity features.

[0056] The evaluation index calculation module is used to calculate evaluation indexes based on the operational characteristics. The evaluation indexes include the calculation matching degree index and the adjustability index of the operational sample under ideal operational conditions.

[0057] The model building module is used to build a prediction model based on the operational data and process domain knowledge. The prediction model is a qualitative prediction model of indicators based on physical information assisted by a prior model and a product-based neural network. The prediction model includes a front-end network and a back-end network. The front-end network is used to fit key state parameters. The back-end network is used to perform quality indicator level classification.

[0058] The bottleneck identification module is used to identify bottleneck process data by quantitatively assessing the marginal contribution of process variables or process combinations in the index prediction based on the prediction model. The bottleneck process data includes at least one of bottleneck process variables and bottleneck process combinations.

[0059] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described method for identifying bottleneck processes in the entire flotation process.

[0060] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described method for identifying bottleneck processes in the entire flotation process.

[0061] By employing the above technical solutions, embodiments of this application provide a method and system for identifying bottleneck processes in the entire flotation process. After acquiring operational data, the method extracts feasibility, robustness, and sensitivity features from the operational data. Then, based on the extracted operational features, it calculates evaluation indicators such as matching degree and adjustability indicators. Next, based on the operational data and process domain knowledge, it constructs a PI-PNN qualitative prediction model for indicators assisted by a prior model. Based on the prediction model, it identifies bottleneck variables by quantitatively evaluating the marginal contribution of process variables in indicator prediction. This method integrates physical constraints into the neural network training process, combines empirical knowledge and data internal knowledge, and uses the calculated values ​​of the prior model to replace the predicted values ​​of distorted variables in the prediction model, thereby improving the classification and prediction accuracy of the backend network. Furthermore, it uses SHAP values ​​and SHAP interaction values ​​to analyze bottleneck processes, and the analysis results can provide a basis for determining the priority of process variable modifications. It can effectively predict the qualitative impact of process parameters on process quality indicators, reduce the costs incurred during ineffective on-site debugging, make the modification process based on evidence, and alleviate the blindness of decision-making.

[0062] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0063] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0064] Figure 1 A flotation process flow diagram provided for embodiments of this application;

[0065] Figure 2 This is a schematic diagram of the process flow diagram for identifying bottleneck processes in the entire flotation process provided in the embodiments of this application;

[0066] Figure 3 This is a schematic diagram of the marker subspace process provided in an embodiment of this application;

[0067] Figure 4 This is a schematic diagram of the calculation and evaluation index process provided in the embodiments of this application;

[0068] Figure 5 This is a schematic diagram of the process for generating rating results provided in an embodiment of this application;

[0069] Figure 6 This is a schematic diagram illustrating the process of establishing a prediction model provided in the embodiments of this application;

[0070] Figure 7 This is a schematic diagram of the structure of the flotation process bottleneck identification system provided in the embodiments of this application. Detailed Implementation

[0071] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0072] In this embodiment, flotation is short for flotation mineral processing, a type of mineral processing technology. Flotation is a mineral separation process that separates mineral particles based on differences in the physicochemical properties of mineral surfaces. Flotation utilizes the differences in wettability of mineral surfaces; that is, mineral particle surfaces are classified as hydrophilic or hydrophobic. Hydrophilic mineral particles are easily wetted by water, while hydrophobic mineral particles are not easily wetted. By adding flotation reagents, the hydrophobicity or hydrophilicity of the mineral surface can be altered, causing mineral particles to selectively adhere to air bubbles. Specifically, hydrophobic mineral particles adhere to air bubbles and rise to the surface of the slurry, forming a foam layer, while hydrophilic mineral particles remain in the slurry, thus achieving mineral separation.

[0073] like Figure 1 As shown, in some embodiments, the flotation process may include steps such as raw material processing, agitation and aeration, bubble mineralization, and mineralization-foam separation. Raw material processing can be achieved through operations such as grinding, slurry preparation, and reagent addition to process the raw ore into a slurry. Specifically, the ore is first ground to a certain particle size using grinding equipment such as ball mills or rod mills, allowing the valuable minerals to be liberated from other minerals or gangue minerals. The ground ore is then prepared into a slurry of appropriate concentration, ensuring uniform dispersion of reagents. Collectors, frothers, and depressants are added to enhance the difference in floatability between the surface of the valuable minerals and gangue minerals.

[0074] Agitation and aeration involve using the agitator and aerator in the flotation machine to stir and draw in air, generating a large number of appropriately sized and relatively stable bubbles. After reacting with the flotation agent, hydrophobic mineral particles adhere to the bubbles and gradually rise to the surface of the slurry, forming mineralized froth. The mineralized froth is then scraped off by rotating scrapers in the flotation machine to obtain froth concentrate, while the product remaining in the slurry and subsequently discharged is tailings.

[0075] Taking the flotation process of a copper mine as an example, after the copper ore particles are crushed and ground to a suitable size for flotation, they are processed with flotation reagents such as collectors and frothers to form a slurry with a specific concentration and pH value. The slurry is then fed into a flotation machine, where copper mineral particles are floated to the froth by bubbles, yielding a rough concentrate. This rough concentrate is then subjected to 2-3 cleaning processes to obtain a copper concentrate. The copper concentrate obtained from flotation is then concentrated using a thickener to remove excess water, and further dewatered using a filter to obtain a dry copper concentrate, thus improving the copper recovery rate and grade.

[0076] Due to the accumulation of minerals and the influence of the storage environment, ores are prone to oxidation. Oxidized ores produce hydrophilic substances, thus reducing their floatability. For example, in the flotation process of copper ore, oxidized copper ore mainly consists of malachite, with small amounts of pseudomalachite, azurite, and chalcopyrite, while iron minerals mainly include magnetite and maghemite, and also contain a very small amount of native silver-bearing gold ore. In the direct flotation method for oxidized copper ore, the altered properties of the original ore and the increased composition of oxidized ore lead to the formation of more hydrophilic substances, further reducing the mineral's floatability.

[0077] To adapt to the impact of mineral oxidation processes on flotation, it is necessary to identify bottleneck factors that may change under different operating conditions or disturbances, and to develop modification schemes based on these identified bottleneck factors. In some embodiments, fault diagnosis can be performed on operational data to identify bottleneck factors. For example, process diagnosis can rely on operator experience, providing an intuitive, interpretable, and easily implemented fault diagnosis method. However, in actual processes, the mineralogy characteristics of the raw ore, such as composition and particle size distribution, exhibit significant fluctuations and uncertainties. Furthermore, the varying skill levels of operators directly affect the stability of the flotation process and the effective control of process parameters. This results in a lack of systematic and quantitative diagnostic methods for experience-based process diagnosis, leading to insufficient diagnostic diagnoses regarding the matching degree between the actual production process and operation, and the difficulty of adjustment.

[0078] In some embodiments, process diagnosis can also be performed based on machine learning and deep neural network methods to identify bottleneck processes. This involves collecting operational data and inputting it into a pre-trained neural network model. The model extracts operational features and uses these features to predict revenue, thereby identifying bottleneck processes based on the revenue prediction results. While neural network-based process diagnosis offers high diagnostic efficiency, its prediction accuracy decreases when dealing with small industrial data samples, making it difficult to achieve the desired accuracy. Furthermore, deep neural networks have poor interpretability, making them unsuitable for process diagnosis in multi-source heterogeneous data environments. This results in insufficient ability to identify bottleneck variables, inadequate consideration of dynamic correlations between variables, and an inability to accurately capture bottleneck factors that may change under different operating conditions or disturbances, thus affecting the accuracy and effectiveness of modification solutions.

[0079] To address the issues of insufficient accuracy and effectiveness in modification schemes, some embodiments of this application provide a method for identifying bottleneck processes in the entire flotation process. This method can be applied to electronic devices with data processing capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control machines. In some embodiments of this application, an electronic device is used as the execution subject to describe the specific process of the method for identifying bottleneck processes in the entire flotation process. It should be understood that the method can also be applied to other types of execution subjects, which will not be shown individually in this application. Figure 2 As shown, the method includes:

[0080] S101. Acquire operation data and extract operation features from the operation data.

[0081] When identifying bottleneck processes in the entire flotation process, electronic equipment can first acquire operational data. This operational data includes operational variables recorded throughout the entire flotation process. Operational data broadly refers to all process data for the entire flotation process, encompassing the various variables input, generated, and output as each step of the flotation process is executed. For example, for the raw material processing step in the flotation process, operational data may include data such as the operating parameters of the grinding equipment, the flow rate of the grinding material, the number of cycles, the stirring speed, the aeration flow rate, the dosage of flotation reagents such as collectors, frothers, and depressants, and the scraper operating speed.

[0082] In order to acquire operational data, in some embodiments, the electronic device can generate a data acquisition request when it needs to acquire operational data and send the data acquisition request to a data source. The data source may include at least one of a data storage module and a sensor. After sending the data acquisition request to the data source, the data source can respond to the data acquisition request by feeding back operational data to the electronic device. The electronic device can then acquire the operational data by receiving the operational data fed back by the data source.

[0083] After acquiring operational data, the electronic device can extract features from the operational data to extract operational features, including feasibility features, robustness features, and sensitivity features. For feasibility features, the electronic device can reasonably divide the operational space and extract the probability distribution of ideal and non-ideal operational samples to calculate the feasibility features of each process state, which is used to measure whether the current operation is within the ideal operating range.

[0084] Therefore, in some embodiments, in order to extract feasibility features, the electronic device may first divide the operation subspace according to the operation variables in the operation data when extracting operation features from the operation data. The operation subspace includes an ideal operation space and a non-ideal operation space.

[0085] like Figure 3 As shown, when an electronic device divides an operation subspace based on operation variables in the operation data, it can extract the operation variables and the associated process information from the operation data. Then, according to the process information, the operation variables are divided into multiple variable groups, where operation variables within the same variable group form joint variables. By dividing the joint space formed by the joint variables into a preset number of subspaces and calculating the benefit value of each subspace, the operation subspace is divided based on whether the benefit value meets a preset benefit standard. If the benefit value meets the preset benefit standard, the subspace is marked as an ideal operation space; if the benefit value does not meet the preset benefit standard, the subspace is marked as a non-ideal operation space.

[0086] After dividing the operation subspace, the electronic device can calculate the ideal probability of the operation subspace and count the number of ideal subspaces and the number of non-empty subspaces. The number of ideal subspaces is the number of subspaces with an ideal probability of 1. The feasibility feature is obtained by calculating the ratio of the number of ideal subspaces to the number of non-empty subspaces.

[0087] Electronic devices can divide the N-dimensional joint operation variable space into 10 NFor each subspace, count the number of ideal and non-ideal operations. Ideal operations have better economic efficiency, while non-ideal operations have worse economic efficiency. Assuming a subspace has n ideal operations and n' non-ideal operations, the ideal probability can be calculated using the following formula:

[0088]

[0089] For example, when extracting feasibility features, all operational variables can be divided into several groups according to the process, and the operational variables within each group can be used as joint variables. Then, the joint space composed of the joint variables can be divided into 10... 3 Each subspace. Based on the operational data corresponding to the joint space, the mineral grade is calculated, and the expected value of the mineral is estimated based on the mineral grade. Simultaneously, the production costs incurred during actual production, including raw material costs, equipment costs, process costs, and labor costs, are obtained from the operational data corresponding to the joint space. The benefit value is obtained by calculating the difference between the expected value and the production output.

[0090] In some embodiments, to achieve the optimal benefit value, electronic devices can also establish an operational variable optimization model. This model determines the optimal (or near-optimal) operating values ​​under a given operating condition, thereby obtaining the optimal target grade value and calculating the benefit value. The operational variable optimization model can be used in fields such as control, engineering, or machine learning to adjust the performance of a system to achieve a desired goal, and is expressed as follows:

[0091]

[0092] Among them, g and g goal These represent the target variable and the target value. b, c, v air R and R represent variables that serve as constraints, which can be variables that affect the benefit value in the flotation process.

[0093] The input to the operational variable optimization model considers fluctuations in ore properties. Uncertain parameters are sampled according to a certain distribution. After optimizing the operational variables, the optimal target value is obtained. Based on the optimal target value, different operator levels are classified. For example, operator level classification can include excellent operators, average operators, and mixed operators. The convergence of different operational levels can be used to evaluate operator level. For example, the mixed operational level is obtained by mixing excellent and average operators in a 3:2 ratio.

[0094] The operation is then classified as ideal or non-ideal based on whether the benefit value meets the target. The ratio of the number of subspaces with an ideal probability of 1 to the number of effective subspaces is used as the feasibility characteristic value. The number of effective subspaces is the number of non-empty subspaces in the operation. Based on the discrete distribution of the original data in three-dimensional space, the entire three-dimensional space can be divided into multiple subspaces. Each subspace simultaneously possesses a certain probability of ideal operation and a probability of non-ideal operation. The probability of ideal operation represents the expected degree of benefit from the operation in that subspace. The ideal probability of each subspace is calculated, and the number of subspaces with an ideal probability of 1 is counted, along with the number of non-empty subspaces. The feasibility characteristic is determined by calculating the ratio of the number of ideal subspaces to the number of non-empty subspaces. A larger ratio indicates better operational feasibility.

[0095] Regarding robustness, electronic devices can collect process operation data and their corresponding revenue data in each subspace, and convert each set of revenue data r i The data are treated as continuous observation variables. A Gaussian Mixture Model (GMM) is used to model the operational data, dividing all observations into ideal and non-ideal sets. Shannon entropy is then used to quantitatively analyze the distributional complexity of different operational states, constructing robustness characteristics that reflect the degree of system fluctuation.

[0096] Therefore, in some embodiments, in order to extract robustness features, the electronic device can calculate the information entropy of the operation subspace and count the number of uncertain subspaces when extracting operation features from the operation data, wherein the uncertain subspaces are subspaces whose information entropy is greater than or equal to a preset entropy threshold. The robustness features are then obtained by calculating the ratio of the number of uncertain subspaces to the number of non-empty subspaces.

[0097] For example, using the subspaces obtained when extracting feasibility features in the above embodiments, the information entropy of the effective subspace can be calculated. The effective subspace entropy is an indicator used to measure the amount of information in a subspace. For a high-dimensional dataset, by analyzing the correlation between variables, strongly coupled variables can be grouped into a joint variable group to find the effective subspace. In this subspace, the distribution of data can be described by a probability density function. The entropy of the effective subspace is the entropy of this probability density function. Therefore, for continuous random variables, the formula for calculating entropy is:

[0098] H(X) = -∫p(x)logp(x)dx;

[0099] Where H(X) is the entropy value; p(x) is the probability density function of the random variable X.

[0100] For example, assuming the ideal set and the non-ideal set follow distributions N1~(μ1,σ1) and N2~(μ2,σ2), and γ is the proportion of N1, then the observation function for each point can be expressed as:

[0101]

[0102] These parameters are estimated using the maximum likelihood estimation (MLE) method, i.e.:

[0103]

[0104] By maximizing the likelihood function, we can obtain the probability distributions of the ideal set and the non-ideal set.

[0105] For each subspace, the ideal probability P of that subspace is obtained through the GMM model. ideal Non-ideal probability P non-ideal Then, the Shannon information entropy formula is used to quantify the uncertainty of this subspace, that is:

[0106] H(x) = -p ideal ·log2p ideal -p non-ideal ·log2p non-ideal ;

[0107] The stability and predictability of a subspace can be assessed by calculating its entropy value H. A high entropy value indicates greater uncertainty in that subspace, suggesting potentially lower robustness of the system. The number of all valid subspaces (i.e., subspaces containing data) is counted, and the number of high-uncertainty subspaces with an information entropy greater than 0.7 is calculated. The ratio of the number of uncertain subspaces to the number of non-empty subspaces is used to represent the robustness of the system in different operating regions: a smaller ratio indicates higher robustness in these subspaces, while a larger ratio indicates poorer stability in these regions.

[0108] For sensitivity characteristics, the Kolmogorov-Smirnov test (KS test) can be used to analyze the significance of the distribution differences between key process variables and system state in electronic equipment, extract the degree of influence of each variable on the change of operating state, and form the operation sensitivity characteristics.

[0109] Therefore, in some embodiments, in order to extract sensitive features, when the electronic device extracts operational features from the operational data, it can call the Least Squares Support Vector Machine (LSSVRM) algorithm model after obtaining the flotation ore grade information corresponding to the operational data.

[0110] The LSSVRM algorithm model is a machine learning algorithm based on statistical learning theory, used for classification and regression tasks. It solves the optimization problem using the least squares method, transforming inequality constraints into equality constraints, thereby simplifying the solution process. Therefore, the LSSVRM algorithm model can be used to fit the nonlinear relationship function between the operational data and the flotation ore grade information.

[0111] After invoking the LSSVRM algorithm model, the electronic device can obtain the conditional cumulative distribution parameters using the LSSVRM algorithm model based on the KS sensitivity test, and obtain the unconditional cumulative distribution parameters using the operational data. Then, a one-dimensional sensitivity value is calculated based on the conditional and unconditional cumulative distribution parameters, and the sensitivity feature is obtained by summing the one-dimensional sensitivity values ​​of the joint variables.

[0112] For example, electronic devices can use the LSSVRM algorithm to fit the nonlinear relationship between operation and grade, and based on the KS sensitivity test, determine the sensitivity of grade changes to a certain operation variable by comparing the maximum absolute difference between the cumulative distribution functions (CDF) of two datasets. This involves using the LSSVRM model to obtain the conditional cumulative distribution. The unconditional cumulative distribution is obtained using the original data. The KS statistic represents the maximum difference between the two CDFs mentioned above. The KS sensitivity values ​​for each operand can then be calculated using the following formula:

[0113]

[0114] Among them, unconditional CDF To directly statistically analyze the model's raw output data (such as concentrate grade), we calculate its cumulative distribution function (CDF). While keeping other variables constant, let the single variable x i Following a uniform distribution, the corresponding output value is obtained through the LSSVRM prediction model, and the CDF is calculated.

[0115] After calculating the KS sensitivity value, the electronic device can calculate the sum of the one-dimensional sensitivity values ​​of the joint variables, that is, the KS statistic for each input variable can be calculated separately using the method described above. For example, the LSSVRM prediction model has a maximum relative error of 0.08 on the test set, indicating high accuracy; therefore, the KS statistic calculated from it has high reliability. To quantify the degree of influence of the variables considered comprehensively, critical values ​​for the KS statistic are given based on expert experience, allowing for further sensitivity calculations of the joint variables.

[0116]

[0117] KS0 is determined by experience in actual production.

[0118] S102. Calculate the evaluation index based on the operational characteristics.

[0119] After extracting operational features, the electronic device can calculate evaluation indicators based on these features. These evaluation indicators include a computational matching degree indicator and an adjustability indicator for the operational sample under ideal operational conditions. Based on the aforementioned three categories of operational features—feasibility, robustness, and sensitivity—the electronic device can construct a multi-dimensional evaluation indicator system and employ the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) to score the comprehensive performance of different operational samples, calculating the computational matching degree indicator G of the sample under ideal operational conditions. f With adjustability index G m Determine whether the process needs to be modified.

[0120] TOPSIS is a multi-attribute decision analysis method used to select the optimal solution from multiple alternatives. TOPSIS evaluates the merits of each solution by calculating its distance from the ideal solution (optimal solution) and the negative ideal solution (worst solution). In some embodiments, when using TOPSIS to score the overall performance of different operational samples, the electronic device can first construct a decision matrix. By grouping m evaluation objects under n indicators into an original matrix, a decision matrix X can be constructed, i.e.:

[0121]

[0122] Where, x ij This represents the value of the i-th scheme on the j-th attribute.

[0123] To eliminate the influence of different attribute units, the decision matrix needs to be normalized after construction. This is done by standardizing the data by dividing each element by the square root of the sum of the squares of its column, resulting in the normalized matrix R:

[0124]

[0125] Where, r ij This represents the normalized result of the value of the i-th scheme on the j-th attribute, i.e.:

[0126]

[0127] After data normalization, electronic devices can assign corresponding weights based on the importance of each attribute. jAnd construct a weighted standardized decision matrix V, that is: v ij =w j ·r ij Next, determine the optimal and worst solutions. The optimal solution (ideal solution) is a vector composed of the maximum values ​​of each attribute, and the worst solution (negative ideal solution) is a vector composed of the minimum values ​​of each attribute. That is:

[0128] Optimal solution Z + ={max(v 1j ), max(v 2j ), ..., max(v mj )};

[0129] Worst solution Z - ={min(v 1j ), min(v 2j ), ..., min(v mj )};

[0130] Then, based on the determined optimal and worst solutions, calculate the cosine distance from each object to the optimal solution, and the cosine distance from each object to the worst solution. That is, for each object i, calculate its Euclidean distance to the optimal and worst solutions:

[0131]

[0132]

[0133] Then, based on the cosine distance between the proposed solution and the optimal and worst solutions, the relative proximity is calculated. Relative proximity S i This value is used to measure the merits of each solution. The closer the value is to 1, the closer the solution is to the ideal solution. The relative closeness can be calculated using the following formula:

[0134]

[0135] Therefore, based on the relative proximity S i Sort all the solutions and select the one with the highest relative similarity as the optimal solution.

[0136] like Figure 4 As shown, in some embodiments, in order to calculate evaluation indicators, the electronic device can generate a score vector based on the operational characteristics when calculating the evaluation indicators. The score vector includes a normal cycle score vector and an abnormal cycle score vector. The normal cycle score vector includes cycles where the benefit value meets production requirements; the abnormal cycle score vector includes cycles where the benefit value is lower than production requirements.

[0137] For example, electronic devices can divide all cycles in the operation process into normal cycles and abnormal cycles according to whether the output requirements are met, and obtain a scoring vector for each. and A normal cycle refers to a period where the revenue meets production requirements, and the corresponding score vector is assigned to the normal group; an abnormal cycle refers to a period where the revenue does not meet the requirements, and the score vector is assigned to the abnormal group.

[0138] The evaluation index is then calculated based on the score vector, and a matching threshold and an adjustment difficulty threshold are obtained. The matching threshold is a threshold obtained through experiments using a prior model; the adjustment difficulty threshold is a preset value. The matching degree index and the matching threshold are then compared to generate the matching state between the operation and the process. Furthermore, the relative operational difficulty between the adjusted operation and the original operation is generated by comparing the adjustability index and the adjustment difficulty threshold.

[0139] By defining a geometric eigenvalue (G) f As a parameter to measure the degree of matching between operations and processes, the similarity between normal and abnormal score distributions is evaluated. Nonlinear fitting is performed on the normal and abnormal periodic score distributions using the least squares method (LSM). After selecting an appropriate kernel function, the slope of the fitted curve is obtained, and the length of its horizontal projection is calculated. Geometric eigenvalues ​​are then calculated.

[0140]

[0141] Where Δk is the slope difference of the fitted curves; p is the ratio of the projections of the fitted curves onto the x-axis.

[0142] At the same time, the geometric mean (G) is used m This involves calculating the distribution trend of vectors to quantify the trend in operational adjustment difficulty. The geometric mean is calculated for each scoring vector, and the adjustability index is defined as follows:

[0143]

[0144] For example, such as Figure 5 As shown, after extracting operational features, electronic devices can classify normal cycles and abnormal cycles based on these features. Normal cycles, where the economic benefits meet production requirements, are then divided into score vectors. Score vectors of normal and abnormal periods It consists of all cycles where the economic benefit value is lower than the production requirements. Then, the matching degree index G is calculated. f With adjustability index G m And obtain the matching threshold G f0 and adjusting the difficulty threshold G m0 Wherein, the matching threshold G f0The difficulty threshold G for operation adjustment was obtained through experiments using a prior model. m0 It is 0.

[0145] According to the matching index G f Matching threshold G f0 Perform a match degree diagnosis. When G f >G f0 When G is in a certain position, it indicates that the operation and process are matched. Conversely, when G... f ≤G f0 This indicates a mismatch between the operation and the process.

[0146] Meanwhile, according to the adjustability index G m And adjusting the difficulty threshold G m0 Adjustability diagnosis of the operation. When G m When G ≥ 0, it indicates that the adjustment operation is easier than the original operation; conversely, when G ≥ 0, it indicates that the adjustment operation is easier than the original operation. m When the value is less than 0, it indicates that the adjustment operation is more difficult.

[0147] Based on the above judgment, if the operation and process have a low degree of matching and the operation is easier to adjust, then operation adjustment should be prioritized; otherwise, process modification should be carried out.

[0148] S103. Establish a prediction model based on the operational data and process domain knowledge.

[0149] After calculating the evaluation indicators, it is possible to determine whether the current flotation process needs to be modified based on the evaluation indicators. When the diagnosis result that process modification is required is obtained, a qualitative prediction model of the PI-PNN index assisted by prior model can be constructed based on the operating data and domain knowledge. That is, the electronic equipment can establish a prediction model based on the operating data and process domain knowledge.

[0150] Since only operational data exists in actual production, and process data is unavailable, it is necessary to combine operational data with process domain knowledge. The front end of the prediction model uses operational data and process knowledge to model key state variables under physical rule constraints, while the back end uses the predicted key state variables to perform qualitative classification prediction of process quality indicators. Furthermore, because complex industrial production processes are highly nonlinear and strongly coupled, directly fitting the correspondence between input and output indicators is difficult. Therefore, within the PINN framework, the PI-PNN network is used. The front end of the PI-PNN is used to estimate key state information in complex industrial production processes, thereby enabling the final indicator classification prediction using the PI-PNN back end network.

[0151] The prediction model is a qualitative prediction model for indicators based on prior model-assisted physical information and a product-based neural network (PI-PNN). The PI-PNN qualitative prediction model is a model that combines physical information (Physics-Informed) and a product-based neural network (PNN) for qualitative prediction. The prediction model includes a front-end network and a back-end network. The front-end network is used to fit key state parameters; the back-end network is used to perform quality indicator level classification.

[0152] The PI-PNN qualitative prediction model can incorporate physical information, embedding physical laws (such as partial differential equations) as prior knowledge into the model, allowing it to fit data while satisfying physical constraints during training. It can also process categorical features based on a product-based neural network architecture, converting features into low-dimensional vectors through embedding layers and capturing interactions between features through product operations.

[0153] When building a prediction model, electronic devices can perform front-end BPNN modeling and back-end PNN classification separately. The front-end network is built based on the BP neural network (Back Propagation Neural Network) structure. By fitting the nonlinear relationship between key state parameters of some measurable processes and the input, mathematical rules are incorporated into the network's convergence loss function through backpropagation. The loss function, as a representation of gradient information, can be expressed as:

[0154]

[0155] The front-end network can achieve convergence using the gradient descent algorithm, which includes inputs of M operational variables and N process variables and outputs of L key state parameters. Regarding the selection of key state parameters, the influence of operations and processes on state parameters can be fully reflected in the mechanism, which is a general cognitive rule in reality.

[0156] The backend network is built on a Probabilistic Neural Network (PNN) structure, fitting the relationship between output indicators and key state parameters of the measurable process. The quality level classification prediction is achieved by inputting the state parameters output from the front-end BPNN into the back-end PNN. The PNN uses Bayesian decision-making as its core, combining the advantages of Radial Basis Function Networks (RBFN) and Probability Density Estimation (PDE). It constructs probability density functions for each category using a Gaussian kernel function, calculates and normalizes the conditional probability of the test sample under each category, and finally outputs the quality level label with the highest probability. To ensure classification performance, the PNN uses the Parzen window function as the activation function in the last layer, ensuring the number of neurons matches the number of output categories, thereby improving the model's generalization ability and prediction accuracy.

[0157] Therefore, as Figure 6 As shown, in some embodiments, when an electronic device executes a predictive model based on operational data and process domain knowledge, it can establish a priori model of key state information for the entire process based on the operational data and prior process knowledge. The input data of the priori model consists of operational data and process variables, and the output data consists of measurable key state parameters.

[0158] The PI-PNN model, composed of a front-end BPNN and a back-end PNN, takes the operational and process variables of the process flow as input and outputs process state parameters and quality level prediction results. Combining actual operational data collected on-site with prior knowledge of the process (such as control mechanisms and empirical formulas), a model reflecting the actual process state is established. This model should take operational and process variables as input and key output indicators as outputs, such as quality indicators, energy consumption indicators, and economic benefits. The purpose of this model is to provide reference and guidance for subsequent processes, forming a "knowledge base" or "experience benchmark" for process operation.

[0159] After establishing a priori model of key state information throughout the entire process, the electronic device then uses a multivariate Taylor function fitting method to extract the relationship function between the input and output data through the prior model. Multivariate Taylor function fitting is a method of approximating a multivariate function by expanding it using a Taylor series around a certain point. A Taylor series is an infinite series of derivatives of an infinitely differentiable function at a point. The multivariate Taylor function fitting method is used to extract the input-output functional relationship of the model.

[0160] The relationship function is then set as the loss function of the front-end network of the prediction model, and the front-end network of the prediction model is trained based on the loss function. Therefore, electronic devices can use a multivariate Taylor function fitting method to extract the functional relationship between input and output data from a prior model. This functional relationship is then used as the loss function to train the front-end part of the PI-PNN.

[0161] While training the front-end PI-PNN network, the electronic device can also use the predicted values ​​of the prior model to replace the distortion parameters predicted by the front-end network of the prediction model. Simultaneously, the predicted values ​​are input into the back-end network of the prediction model, allowing the back-end network to train the prediction model within a product-based neural network framework. In other words, the electronic device can use the predicted values ​​of the prior model to replace the distortion parameters predicted by the PI-PNN front-end network. The prediction results with replaced distortion parameters are then input into the PI-PNN back-end network, which trains the final classification prediction model within the PNN framework.

[0162] For example, electronic devices replace the "distortion parameters" predicted by the PI-PNN front-end with more accurate parameter values ​​from the prior model, ensuring that the model parameters do not deviate from physical reality. These adjusted parameters are then fed into the PI-PNN back-end network. Ultimately, within the entire PNN framework, a model with enhanced classification and prediction capabilities is trained to identify whether a process is operating in a non-ideal state.

[0163] After training the front-end and back-end networks of the PI-PNN according to the methods provided in the above embodiments, the electronic device can apply the trained PI-PNN model to perform qualitative prediction of indicators. Therefore, in some embodiments, when the electronic device executes the prediction model based on the operational data and process domain knowledge, it can obtain the SHAP value and SHAP interaction value input to the prediction model.

[0164] The Shapley Additive Explanations (SHAP) value, based on the Shapley value, measures the contribution of each feature to the model's prediction. It fairly distributes the contribution of each feature to the model's prediction by considering all possible subsets of features, thus characterizing the degree of contribution of a single process variable to the prediction result. Electronic equipment can perform SHAP value and SHAP interaction value analysis on the trained prediction model to evaluate the marginal impact of each input variable (i.e., each process variable) on the output result and the synergistic effects between variables. Based on the SHAP values, the key variables with the greatest impact on the prediction result are identified as bottleneck process variables. If the joint SHAP interaction value of two or more process variables is high, it indicates a combined bottleneck and they should be considered as a whole. These high-contribution variables are the focus of subsequent process modifications.

[0165] For example, the formula for calculating the SHAP value of input variable X is:

[0166]

[0167] in, It is the model output mean. It is the SHAP marginal contribution value of the k-th variable. The formula for calculation is:

[0168]

[0169] The SHAP interaction value is used to measure the contribution of the synergistic effect between features to the model prediction, that is, to characterize the degree of contribution of multiple process variables to the prediction result. The SHAP interaction value can be decomposed into main effect value and interaction effect value. The main effect value represents the contribution of each feature individually to the model prediction result. The interaction effect value represents the contribution of the synergistic effect between features to the model prediction result. Therefore, the formula for calculating the SHAP interaction value is:

[0170]

[0171] After obtaining the SHAP value and SHAP interaction value, the electronic device determines the bottleneck process based on these values. Then, it obtains contribution parameters of different bottleneck processes to the final prediction result and determines subsequent modification targets based on these parameters. The contribution parameters include the contribution value of a single process variable to the prediction result, and / or the combined contribution value of multiple process variables. The subsequent modification targets include process variables whose SHAP values ​​are higher than a preset SHAP value threshold, and / or process combinations whose SHAP interaction values ​​are higher than a preset SHAP interaction value threshold.

[0172] In other words, electronic devices can determine the bottleneck process based on the SHAP value and SHAP interaction value input from the prediction model. Then, the contribution value or joint contribution value of different bottleneck processes to the final prediction result is obtained, and the process variables or process combinations with higher SHAP values ​​are selected as the targets for subsequent modification.

[0173] After acquiring the process dataset, the electronic device, following the prediction model establishment method provided in the above embodiments, repeatedly establishes a prior model for the complex industrial process, trains the front-end PI-PNN network, trains the final classification prediction model in the back-end network within the PNN framework, and determines the bottleneck process based on the SHAP value and SHAP interaction value input to the prediction model, until the prior model can no longer assist the PI-PNN model in prediction, reaching the convergence state of the algorithm. It is evident that, in order to accurately identify bottleneck process variables throughout the flotation process, the electronic device can identify bottleneck process variables and bottleneck process combinations based on SHAP values ​​and SHAP interaction values. SHAP values ​​are based on game theory, enabling the analysis of the individual contribution value of each variable and the joint contribution value of multiple variable interactions based on a limited amount of prediction data. Variables (or combinations) with higher contribution values ​​are considered process bottleneck variables (or combinations).

[0174] S104. Based on the prediction model, bottleneck process data are identified by quantitatively evaluating the marginal contribution of process variables or process combinations in the index prediction.

[0175] After establishing the prediction model, electronic devices can incorporate SHAP value analysis based on the prediction model to quantitatively evaluate the marginal contribution of each process variable in the indicator prediction. That is, bottleneck process data can be identified by quantitatively evaluating the marginal contribution of process variables or process combinations in the indicator prediction. Bottleneck process data includes at least one of bottleneck process variables and bottleneck process combinations.

[0176] In some embodiments, when an electronic device identifies bottleneck variables based on the prediction model by quantitatively assessing the marginal contribution of process variables in indicator prediction, it can first calculate the interaction parameters of variables between process combinations based on the SHAP interaction value. These interaction parameters can be synergistic or antagonistic parameters. Based on these interaction parameters, key bottleneck variables in the process flow are then identified. The modification priority of these key bottleneck variables is then calculated.

[0177] In identifying bottleneck variables, electronic equipment incorporates SHAP value analysis based on the predictive model to quantitatively assess the marginal contribution of each process variable in the indicator prediction. Simultaneously, it further analyzes SHAP interaction values ​​to identify the impact of synergistic effects between variables on process production, identify key bottleneck variables and their combinations in the process, and determine the priority of modifications.

[0178] For example, when the combination order is 2, 3, and 4, the number of combinations of process variables are 36, 84, and 136, respectively. Based on the bottleneck variable analysis results, the determined modification priorities are shown in the table below:

[0179] quantity combination Joint contribution Prioritization of modifications 1 Vcell1 0.3947 1 3 Vcell1+Vcell2+Vcell3 0.3143 2 4 Vcell1+Vcell2+Vcell3+dia2 0.3191 3 4 Vcell1+Vcell2+Vcell3+dia3 0.3184 4 4 Vcell1+Vcell2+Vcell3+DImp2 0.3179 5 4 Vcell1+Vcell2+Vcell3+Dimp3 0.3145 6 4 Vcell1+Vcell2+Vcell3+dia1 0.3031 7 4 Vcell1+Vcell2+Vcell3+Dimp1 0.2944 8

[0180] Where Vcell represents the volume of the flotation cell; DImp represents the impeller diameter; dia represents the diameter of the air inlet; and the suffix number represents the flotation stage, i.e., suffix number 1 represents the roughing stage, suffix number 2 represents the scavenging stage, and suffix number 3 represents the cleaning stage.

[0181] After each iteration, the data or parameters used by the model are updated until the prior model no longer provides significant improvement to the PI-PNN prediction results, i.e., the algorithm converges. At this point, the model has stably identified the process bottleneck, forming a reliable basis for bottleneck diagnosis and modification.

[0182] By applying the technical solutions of the above embodiments, the method for identifying bottleneck processes in the entire flotation process provided in the above embodiments can quantify the necessity of operational adjustments by constructing a comprehensive operational evaluation system. When there is no need for operational adjustments, a qualitative prediction model integrating process and operation indicators can be established for bottleneck process identification.

[0183] The method described above can effectively determine whether substandard production efficiency is caused by differences in operator skill levels through a comprehensive operational evaluation system, avoiding the blindness, randomness, and self-interest inherent in human judgment due to the decision-maker's style. Starting from the underlying characteristics of operations, the method analyzes the joint characteristics and one-dimensional sensitivity characteristics of operational variables to comprehensively and scientifically determine the necessity of improvement, i.e., whether the substandard efficiency is caused by improper operation.

[0184] The method quantifies abstract concepts, such as operational feasibility, operational robustness, operational sensitivity, degree of matching between operation and process, and difficulty of operational adjustment, and expresses distribution characteristics numerically. It fully mines the feature information contained in the original operational data, making the method easy to implement and highly interpretable and generalizable.

[0185] The method can also identify bottleneck process parameters as key process parameters to be modified, establish a PI-PNN qualitative prediction model, model the flotation process parameters based on the PINN framework, integrate physical constraints into the neural network training process, combine empirical knowledge and internal data knowledge, and use the calculated values ​​of the prior model to replace the predicted values ​​of the distorted variables of the prediction model, thereby ensuring the classification prediction accuracy of the back-end network.

[0186] Furthermore, a predictive model built based on operational data and process domain knowledge objectively analyzes bottleneck processes using SHAP values ​​and SHAP interaction values. The analysis results include joint contribution and synergistic antagonistic effect analysis, providing a basis for determining the priority of process variable modifications. This method can effectively predict the qualitative impact of process parameters on process quality indicators, reduce costs incurred during ineffective on-site debugging, ensure that the modification process is based on evidence, and avoid blind decision-making.

[0187] In some embodiments, as a specific implementation of the flotation process bottleneck identification method described in the above embodiments, some embodiments of this application also provide a flotation process bottleneck identification system, such as... Figure 7 As shown, the system includes:

[0188] The feature extraction module is used to acquire operational data and extract operational features from the operational data, which includes operational variables recorded throughout the entire flotation process; the operational features include feasibility features, robustness features, and sensitivity features.

[0189] The evaluation index calculation module is used to calculate evaluation indexes based on the operational characteristics. The evaluation indexes include the calculation matching degree index and the adjustability index of the operational sample under ideal operational conditions.

[0190] The model building module is used to build a prediction model based on the operational data and process domain knowledge. The prediction model is a qualitative prediction model of indicators based on physical information assisted by a prior model and a product-based neural network. The prediction model includes a front-end network and a back-end network. The front-end network is used to fit key state parameters. The back-end network is used to perform quality indicator level classification.

[0191] The bottleneck identification module is used to identify bottleneck process data by quantitatively assessing the marginal contribution of process variables or process combinations in the index prediction based on the prediction model. The bottleneck process data includes at least one of bottleneck process variables and bottleneck process combinations.

[0192] By applying the technical solutions of the above embodiments, the above embodiments provide a flotation full-process bottleneck process identification system. After acquiring operational data, the feature extraction module can extract feasibility features, robustness features, and sensitivity features from the operational data. The evaluation index calculation module then calculates evaluation indices such as matching degree indices and adjustability indices based on the extracted operational features. The model building module then constructs a PI-PNN index qualitative prediction model based on prior model assistance according to the operational data and process domain knowledge. Furthermore, the bottleneck identification module identifies bottleneck variables by quantitatively evaluating the marginal contribution of process variables or process combinations in index prediction based on the prediction model. The system can integrate physical constraints into the neural network training process, combining empirical knowledge and data internal knowledge, using the calculated values ​​of the prior model to replace the predicted values ​​of distorted variables in the prediction model, thereby improving the classification prediction accuracy of the backend network. Moreover, the system uses SHAP values ​​and SHAP interaction values ​​to analyze bottleneck processes, and the analysis results can provide a basis for determining the priority of process variable modification. It can effectively predict the qualitative impact of process parameters on process quality indicators, reduce the costs incurred during ineffective on-site debugging, make the modification process based on evidence, and alleviate the blindness of decision-making.

[0193] It should be noted that other corresponding descriptions of the functional units involved in the flotation process bottleneck identification system provided in the embodiments of this application can be found in the corresponding descriptions in the flotation process bottleneck identification method provided in the above embodiments, and will not be repeated here.

[0194] This application also provides a computer device, specifically a personal computer, server, network device, etc. The computer device includes a bus, processor, memory, and communication interface, and may also include input / output interfaces and a display device. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores location information. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0195] Those skilled in the art will understand that the structure of the computer device described above is only a partial structure related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. A specific computer device may include more or fewer components, or combine certain components, or have different component arrangements.

[0196] In one embodiment, a computer-readable storage medium is also provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0197] In one embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0198] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0199] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0200] Any references to memory, database, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc.

[0201] Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0202] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0203] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0204] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying bottleneck processes in the entire flotation process, characterized in that, The method includes: Acquire operational data and extract operational features from the operational data, the operational data including operational variables recorded throughout the entire flotation process; the operational features include feasibility features, robustness features, and sensitivity features; Evaluation indicators are calculated based on the operational characteristics, including the calculation matching degree index and the adjustability index of the operational sample under ideal operational conditions. A predictive model is established based on the operational data and process domain knowledge. This predictive model is a qualitative predictive model of indicators based on prior model-assisted physical information and a product-based neural network. The predictive model includes a front-end network and a back-end network. The front-end network is used to fit key state parameters; the back-end network is used to perform quality indicator level classification. Establishing the predictive model based on the operational data and process domain knowledge includes: establishing a prior model of key state information for the entire process based on the operational data and prior process knowledge. The input data of the prior model are operational data and process variables, and the output data of the prior model are measurable key state parameters. Establishing the predictive model based on the operational data and process domain knowledge includes: obtaining the SHAP value and SHAP interaction value of the predictive model input; the SHAP value is used to characterize the contribution of a single process variable to the prediction result; the SHAP interaction value is used to characterize the contribution of multiple process variables to the prediction result; and the SHAP value and SHAP interaction value are used to characterize the contribution of multiple process variables to the prediction result. The process involves: identifying bottleneck processes through cross-value comparison; obtaining contribution parameters of different bottleneck processes to the final prediction result, including the contribution value of a single process variable to the prediction result, and / or the joint contribution value of multiple process variables to the prediction result; determining subsequent modification targets based on the contribution parameters, including process variables with SHAP values ​​higher than a preset SHAP value threshold, and / or process combinations with SHAP interaction values ​​higher than a preset SHAP interaction value threshold; extracting the relationship function between input and output data using a multivariate Taylor function fitting method through the prior model; setting the relationship function as the loss function of the front-end network of the prediction model, and training the front-end network of the prediction model based on the loss function; replacing the distortion parameters predicted by the front-end network of the prediction model with the predicted values ​​of the prior model; and inputting the predicted values ​​into the back-end network of the prediction model to train the prediction model within a product-based neural network framework. Based on the prediction model, bottleneck process data is identified by quantitatively evaluating the marginal contribution of process variables or process combinations in indicator prediction. The bottleneck process data includes at least one of bottleneck process variables and bottleneck process combinations. Identifying bottleneck variables by quantitatively evaluating the marginal contribution of process variables or process combinations in indicator prediction based on the prediction model includes: calculating the interaction parameters between variables in the process combination based on the SHAP interaction value, where the interaction parameters are synergistic or antagonistic interaction parameters; identifying key bottleneck variables in the process flow based on the interaction parameters; and calculating the modification priority of the key bottleneck variables.

2. The method according to claim 1, characterized in that, Extracting operational features from the operational data includes: The operation subspace is divided according to the operation variables in the operation data, and the operation subspace includes an ideal operation space and a non-ideal operation space. Calculate the ideal probability of the operation subspace; The number of ideal spaces and the number of non-empty subspaces are counted, where the number of ideal spaces is the number of subspaces with an ideal probability of 1. The feasibility characteristics are obtained by calculating the ratio of the number of ideal spaces to the number of non-empty subspaces.

3. The method according to claim 2, characterized in that, The operation subspace is divided according to the operation variables in the operation data, including: Extract the operational variables and the process information associated with the operational variables from the operational data; The operational variables are divided into multiple variable groups according to the process information, wherein the operational variables in the same variable group form a joint variable. The joint space formed by the joint variables is divided into a predetermined number of subspaces; Calculate the benefit value of the subspace; If the benefit value reaches the preset benefit standard, the subspace is marked as an ideal operating space; If the benefit value does not meet the preset benefit standard, the subspace is marked as a non-ideal operating space.

4. The method according to claim 2, characterized in that, Extracting operational features from the operational data includes: Calculate the information entropy of the operation subspace; The number of uncertain subspaces is counted, where the uncertain subspace is the subspace whose information entropy is greater than or equal to a preset entropy threshold. The robustness feature is obtained by calculating the ratio of the number of uncertain subspaces to the number of non-empty subspaces.

5. The method according to claim 3, characterized in that, Extracting operational features from the operational data includes: Obtain the flotation ore grade information corresponding to the operation data; The Least Squares Support Vector Machine (LSSVRM) algorithm model is invoked, which is used to fit the nonlinear relationship function between the operational data and the flotation ore grade information. Based on the KS sensitivity test, the conditional cumulative distribution parameters are obtained using the LSSVRM algorithm model, and the unconditional cumulative distribution parameters are obtained using the operational data. Calculate a one-dimensional sensitivity value based on the conditional cumulative distribution parameter and the unconditional cumulative distribution parameter; The sensitivity feature is obtained by calculating the sum of the one-dimensional sensitivity values ​​of the joint variables.

6. The method according to claim 1, characterized in that, The evaluation index is calculated based on the operational characteristics, including: A score vector is generated based on the operational characteristics. The score vector includes a normal cycle score vector and an abnormal cycle score vector. The normal cycle score vector includes cycles in which the efficiency value meets production requirements. The abnormal cycle score vector includes cycles in which the efficiency value is lower than production requirements. The evaluation index is calculated based on the score vector; The matching threshold and the difficulty adjustment threshold are obtained. The matching threshold is a threshold obtained through experiments using a prior model. The difficulty adjustment threshold is a preset value. By comparing the matching degree index and the matching threshold, the matching status of operation and process is generated; By comparing the adjustability index and the adjustment difficulty threshold, the relative operational difficulty of the adjustment operation and the original operation is generated.

7. A system for identifying bottleneck processes in the entire flotation process, characterized in that, The system is applied to the method for identifying bottleneck processes in the entire flotation process according to any one of claims 1-6; the system comprises: The feature extraction module is used to acquire operational data and extract operational features from the operational data, which includes operational variables recorded throughout the entire flotation process; the operational features include feasibility features, robustness features, and sensitivity features. The evaluation index calculation module is used to calculate evaluation indexes based on the operational characteristics. The evaluation indexes include the calculation matching degree index and the adjustability index of the operational sample under ideal operational conditions. The model building module is used to build a prediction model based on the operational data and process domain knowledge. The prediction model is a qualitative prediction model of indicators based on physical information assisted by a prior model and a product-based neural network. The prediction model includes a front-end network and a back-end network. The front-end network is used to fit key state parameters. The back-end network is used to perform quality indicator level classification. The bottleneck identification module is used to identify bottleneck process data by quantitatively assessing the marginal contribution of process variables or process combinations in the index prediction based on the prediction model. The bottleneck process data includes at least one of bottleneck process variables and bottleneck process combinations.

Images