Method for optimizing a comminution-grinding-flotation synergy process based on metal liberation degree control

Abstract

The application relates to the technical field of mineral processing automation, and discloses a crushing-grinding-flotation synergic process optimization method based on metal dissociation degree regulation, which comprises the following steps: acquiring multi-dimensional sensor data in a crushing-grinding-flotation production process in real time; establishing a benchmark prediction model offline based on historical process data; in an online running stage, quantifying and determining the working condition drift state of a current process system online; when it is determined that significant working condition drift occurs, constructing a residual compensation model online; fusing the prediction result of the benchmark prediction model and the prediction result of the residual compensation model to generate a final process index prediction value; and based on the final process index prediction value, solving through an optimization algorithm to generate a synergic control set value. The application greatly reduces the calculation cost and data requirement of model maintenance, and significantly improves the response speed and adaptability of the system to working condition changes, thereby ensuring the long-term stability and reliability of the synergic optimization effect.

Description

Technical Field

[0001] This application relates to the field of mineral processing automation technology, specifically to an optimization method for a crushing-flotation synergistic process based on metal liberation degree control. Background Technology

[0002] In the field of mineral processing, crushing and flotation are two closely coupled and crucial processes. The crushing process aims to fully liberate valuable minerals from gangue minerals, and the particle size distribution of the product directly determines the recovery rate and concentrate grade of subsequent flotation units. To optimize the technical and economic indicators of the entire beneficiation process, the industry commonly adopts optimization control methods based on mathematical models. By adjusting the operating parameters of crushing and flotation units, coordinated production is achieved. These models typically utilize historical production data to establish complex nonlinear relationships between operating parameters and process indicators.

[0003] However, the mineral processing process is a time-varying system. Fluctuations in the properties of the raw ore, wear and tear on equipment, and changes in environmental factors can all cause the operating conditions of the process system to drift. After conducting an in-depth analysis of the existing technology, the applicant found that the prediction accuracy of the existing models in practical applications decreases significantly over time, and the optimization control effect is difficult to maintain in a stable manner over a long period of time. Summary of the Invention

[0004] This application provides a method for optimizing the synergistic crushing-flotation process based on metal dissociation degree control, comprising: real-time acquisition of multi-dimensional sensor data during the crushing-flotation production process, wherein the multi-dimensional sensor data includes process data of the crushing unit and process data of the flotation unit; based on historical process data, establishing offline a benchmark prediction model to characterize the nonlinear mapping relationship between the operating parameters of the crushing unit, the operating parameters of the flotation unit, and process indicators including metal dissociation degree; during the online operation phase, extracting real-time operating condition feature vectors from the real-time acquired multi-dimensional sensor data, and quantifying and determining the current process system's operating condition drift state online by calculating the statistical deviation between the real-time operating condition feature vectors and a pre-set benchmark operating condition feature space; when the drift state is determined... When a significant operating condition drift occurs, a small amount of real-time process data under the current operating condition is collected, and an online residual compensation model is constructed to predict the residual between the predicted value of the benchmark prediction model and the actual process index value under the current operating condition. The prediction results of the benchmark prediction model and the prediction results of the residual compensation model are fused to generate a final process index prediction value that reflects the current operating condition drift state. Based on the final process index prediction value, an optimization algorithm is used to solve and generate and output a set of collaborative control setpoints that optimize the expected mineral processing technical and economic indicators. The collaborative control setpoints include the setpoints of the crushing and grinding unit operating parameters and the flotation unit operating parameters to achieve dynamic collaborative control of the crushing and grinding process and the flotation process.

[0005] By adopting the above technical solution, a composite modeling framework of benchmark model plus online residual compensation is proposed. It combines a complex benchmark model that describes the long-term stability law of the system with a lightweight residual model that describes the short-term operating condition drift. This avoids the frequent and expensive retraining of the complex benchmark model. Only the residual model needs to be built online to achieve dynamic adaptation to operating condition drift. This not only greatly reduces the computational cost and data requirements of model maintenance, but also significantly improves the system's response speed and adaptability to changes in operating conditions, thereby ensuring the long-term stability and reliability of the collaborative optimization effect.

[0006] Optionally, determining that a significant operating condition drift has occurred specifically includes: quantifying the statistical deviation by calculating the Mahalanobis distance or other statistical distance between the real-time operating condition feature vector and the center point of the baseline operating condition feature space, and comparing the statistical deviation with a preset threshold. When the statistical deviation is greater than the preset threshold, a significant operating condition drift is determined to have occurred.

[0007] By adopting the above technical solution, and by introducing statistical distances such as Mahalanobis distance to quantify the deviation between the real-time operating condition feature vector and the baseline operating condition feature space, an objective quantitative means of determining operating condition drift is provided. Compared with the passive triggering method that relies on the prediction error threshold, this method can detect changes in the internal state of the system earlier and more sensitively.

[0008] Optionally, the real-time operating condition feature vector is a low-dimensional feature vector with high information content extracted from the multi-dimensional sensor data through an autoencoder, principal component analysis, or slow feature analysis data dimensionality reduction or feature learning algorithm.

[0009] By adopting the above technical solution, before drift determination, the original high-dimensional sensor data is extracted using dimensionality reduction algorithms such as autoencoders. This effectively filters out noise and redundant information in the original data and extracts low-dimensional and high-information-content operating condition feature vectors that can characterize the core state of the system. This not only improves the accuracy and robustness of operating condition drift detection but also reduces the computational complexity of subsequent models.

[0010] Optionally, the benchmark prediction model is a Gaussian process regression model, a support vector regression model, or a deep neural network model.

[0011] By adopting the above technical solutions, the benchmark prediction model is limited to advanced models such as Gaussian process regression, support vector regression, or deep neural networks. These models have strong nonlinear fitting capabilities and can deeply explore and accurately characterize the complex physicochemical relationships between variables in the mineral processing process, thereby ensuring that the benchmark prediction model has high initial accuracy and good generalization performance.

[0012] Optionally, the residual compensation model is a lightweight regression model that is rapidly trained online using a small amount of real-time process data, and its structural complexity is lower than that of the benchmark prediction model.

[0013] By adopting the above technical solution, the residual compensation model is designed as a lightweight regression model, which can be quickly trained online using a small amount of real-time data. Its simple structure and low computational cost ensure that the compensation mechanism can be embedded in the real-time control loop, realizing the instant calibration of model predictions. This is an advantage that traditional model retraining methods cannot match.

[0014] Optionally, the input information of the residual compensation model includes the real-time operating condition feature vector, thereby establishing a direct correlation between the operating condition drift state and the predicted residual of the benchmark prediction model.

[0015] By adopting the above technical solution, the extracted real-time operating condition feature vector is used as the input of the residual compensation model, and a direct mapping relationship between the operating condition drift state and the prediction residual of the benchmark model is established. This enables the residual compensation model to accurately predict the corresponding error size according to the specific drift pattern and degree. Compared with the method of using only time series information to predict errors, it has stronger mechanism interpretability and higher prediction accuracy.

[0016] Optionally, the operating parameters of the grinding unit include at least one of the following: mill feed rate, mill speed, hydrocyclone inlet pressure, or mill ball replenishment rate; the operating parameters of the flotation unit include at least one of the following: flotation cell level, flotation cell aeration rate, flotation reagent type, flotation reagent addition location, or flotation reagent addition amount.

[0017] By adopting the above technical solution, the specific operating parameters that can be adjusted for the crushing and grinding unit and the flotation unit are clarified, so that the technical solution of this application is closely integrated with the actual production equipment and operation practice of the concentrator. The execution object of the collaborative optimization is clearly defined, ensuring that this application has high industrial applicability and operability.

[0018] Optionally, the dynamic coordinated control obtains the target metal dissociation distribution as an intermediate target for process optimization control, with the ultimate goal of achieving the optimal technical and economic indicators of the final mineral processing.

[0019] By adopting the above technical solution, the target metal dissociation distribution is used as an intermediate target for process optimization. Metal dissociation is the core physical bridge connecting the grinding effect and flotation performance. By introducing this intermediate target, this method decomposes the macroeconomic indicator optimization problem into a physical indicator control problem based on the process mechanism, making the optimization process more physically meaningful, the optimization results more stable and reliable, and avoiding the optimization algorithm from getting stuck in local optima or generating control commands that violate the process principle.

[0020] Optionally, the final beneficiation techno-economic indicators include a comprehensive evaluation function formed by at least one or a combination of concentrate recovery rate, concentrate grade, comprehensive energy consumption per unit of ore processing, or comprehensive reagent consumption per unit of ore processing.

[0021] By adopting the above technical solution, the final optimization target is set as a comprehensive evaluation function of technical and economic indicators such as concentrate recovery rate, grade, comprehensive energy consumption or reagent consumption. This gives the method a high degree of flexibility, allowing mineral processing enterprises to dynamically adjust the optimization target according to the strategic priorities of different periods such as market conditions, production plans and environmental protection requirements, so as to maximize the overall benefits of the enterprise.

[0022] Optionally, it also includes a model maintenance step: based on the cumulative degree of operating condition drift or a preset maintenance cycle, the benchmark prediction model is retrained offline using the process data accumulated during the online operation phase to update its model structure or model parameters, thereby ensuring its long-term prediction accuracy.

[0023] By adopting the above technical solution, offline maintenance steps for the baseline model are added, ensuring the long-term health and sustainable operation of the system. Although the residual compensation model can effectively cope with short-term operating condition drift, long-term cumulative system changes eventually need to be solidified into the baseline model. Regular model maintenance can prevent the baseline model from deviating excessively from the actual operating conditions, ensuring that the residual compensation model always works within a reasonable range, forming a complete closed loop that combines short-term and long-term adaptation. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the process optimization method for the combined crushing-flotation process based on the control of metal dissociation degree in this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be described in detail below. The specific embodiments described herein are merely illustrative and not intended to limit the scope of protection of this application. It should be understood that various modifications or variations can be made to this application by those skilled in the art without departing from the spirit and scope of this application, and such modifications and variations also fall within the scope of protection of this application.

[0026] like Figure 1 As shown in the embodiments of this application, a method for optimizing the crushing-flotation synergistic process based on metal liberation degree control is disclosed. This method can be applied to large-scale copper-molybdenum beneficiation plants. By constructing a two-layer prediction architecture that includes a benchmark prediction model containing long-term patterns and a residual compensation model for real-time correction, the method can achieve rapid perception and dynamic adaptation to the drift of the beneficiation process conditions, thereby ensuring the long-term effectiveness of synergistic optimization control. The method specifically includes the following steps.

[0027] S01: Real-time acquisition of multi-dimensional sensor data during the crushing-flotation production process.

[0028] Understandably, in a typical copper-molybdenum beneficiation production line, the collected multi-dimensional sensor data systematically covers the process data of the crushing and grinding unit and the flotation unit, so as to comprehensively reflect the operating status of the entire process.

[0029] Specifically, for the grinding unit, a series of process data needs to be continuously collected and recorded. This process data includes, but is not limited to: the raw ore feed rate into the first-stage ball mill, usually measured in tons per hour; ore hardness and ore grade, which characterize the physical properties of the ore, and these data can be obtained from an upstream online ore analyzer or through periodic manual sampling and testing; the amount of grinding media added to the mill cylinder, such as the weight or number of steel balls added, which is an important operating variable affecting grinding efficiency; the power consumption or current value of the mill's main motor, which indirectly reflects the load state inside the mill and the grinding condition of the material; the inlet slurry pressure of the hydrocyclone group, which is a key parameter determining the classification efficiency; the underflow circulation load of the hydrocyclone group; and the amount of process water added to the mill. For the flotation unit, a series of process data also need to be collected, which may include: the pulp level height inside each flotation cell, which directly affects the stability of the flotation froth layer and the froth removal efficiency; the amount of air introduced into the bottom of the flotation cell, which controls the degree of bubble dispersion and the probability of collision between mineral particles and bubbles; the movement speed or stability of the froth layer in the flotation cell, which can be obtained through machine vision-based intelligent sensors; the amount of various flotation reagents added, which are further subdivided into collectors used to enhance the hydrophobicity of the target mineral, such as the addition rate of ethyl xanthate or butyl xanthate; frothers used to generate stable froth, such as the addition rate of No. 2 oil; and modifiers used to adjust the pulp chemical environment to achieve selective flotation, such as the amount of lime added to suppress pyrite, which directly affects the pH value of the pulp. All these sensor data from different locations and of different types can undergo preprocessing after acquisition. This includes filtering out random noise using median filtering or Kalman filtering algorithms, identifying and removing outliers using the Laida criterion or isolated forest algorithm, and aligning all time series data to a unified timestamp using interpolation or resampling techniques. This forms a multidimensional data matrix that is dimensionally regular, time-synchronized, and can accurately reflect the true state of the process system at every moment.

[0030] S02: Based on historical process data, establish an offline benchmark prediction model to characterize the nonlinear mapping relationship between crushing and grinding unit operating parameters, flotation unit operating parameters and process indicators including metal dissociation degree.

[0031] Specifically, it is necessary to extract historical process data from the concentrator's historical process database over a sufficiently long period, such as the past one to two years. This data should cover as many known production conditions as possible, including processing raw ores from different origins and of different grades, experiencing different environmental conditions such as high summer temperatures and low winter temperatures, and recording equipment operation under different wear conditions. The extracted data should include all process data from the aforementioned crushing and flotation units, as well as corresponding key process indicators measured through manual sampling analysis or online monitoring instruments. These process indicators should especially include data on the liberation degree distribution of the target metal, such as copper minerals, at different particle size levels, the final copper concentrate recovery rate, copper concentrate grade, molybdenum concentrate recovery rate and grade, the comprehensive power consumption per unit of raw ore processed, and the comprehensive reagent consumption per unit of raw ore processed. After obtaining this high-quality historical process dataset, a benchmark prediction model can be constructed.

[0032] Understandably, the baseline prediction model can be a Gaussian process regression model, a support vector regression model, or a deep neural network model. In this embodiment, we can use a Gaussian process regression model. A Gaussian process regression model is a non-parametric probabilistic model based on Bayesian theory. It does not directly learn the fixed functional relationship between input and output, but instead defines a probability distribution over a function space composed of all possible functions. It can not only provide accurate point predictions of process indicators but also simultaneously give the uncertainty or confidence interval of the predicted value. When constructing a Gaussian process regression model, it is necessary to select a suitable kernel function, also known as a covariance function, to measure the similarity between any two input data points. In this application, the input data point is a multidimensional vector composed of the operating parameters of the grinding unit and the flotation unit; for example, a radial basis function kernel can be selected, which can capture smooth and complex nonlinear relationships. The training process of the benchmark prediction model essentially involves using historical data to optimize the hyperparameters of the radial basis function kernel by maximizing the marginal likelihood function. Once trained, the benchmark prediction model solidifies the operating patterns of the crushing-flotation system under historical conditions. It can predict the expected values ​​and probability distributions of key process indicators such as metal liberation distribution, concentrate recovery rate, and grade based on any given set of crushing and flotation operation parameter settings.

[0033] S03: During the online operation phase, real-time operating condition feature vectors are extracted from the multi-dimensional sensor data acquired in real time. By calculating the statistical deviation between the real-time operating condition feature vectors and the pre-set benchmark operating condition feature space, the current operating condition drift state of the process system is quantified and determined online.

[0034] Understandably, in actual production, fluctuations in the properties of the raw ore, such as changes in ore hardness, embedding characteristics, and gangue composition, are the most significant and frequent factors causing operational drift. To accurately capture this drift, directly using the aforementioned high-dimensional and noisy raw multidimensional sensor data is ineffective. Therefore, this application can employ a data dimensionality reduction or feature learning algorithm to extract a low-dimensional and highly informative feature vector from high-dimensional and noisy original multi-dimensional sensor data. This feature vector is the real-time operating condition feature vector. In this embodiment, we can use an autoencoder to perform this task. An autoencoder is an unsupervised deep learning model consisting of an encoder network and a decoder network connected in series. The encoder network is responsible for compressing the high-dimensional original multi-dimensional sensor data matrix into a low-dimensional latent variable vector, i.e., the operating condition feature vector. The decoder network attempts to reconstruct the original multi-dimensional sensor data from this low-dimensional operating condition feature vector. The goal of the autoencoder is to minimize the reconstruction error between the original input and the decoder output. In this way, the autoencoder is forced to learn how to extract the most essential and informative features from the original data and encode these features into the low-dimensional operating condition feature vector, while filtering out a large amount of noise and redundant information. In the baseline model establishment phase, we can train an autoencoder using historical process data and use it to process sensor data under all stable operating conditions within a historical period, obtaining a series of historical operating condition feature vectors. The distribution of these historical operating condition feature vectors in a low-dimensional space constitutes a pre-defined baseline operating condition feature space. This baseline operating condition feature space can be viewed as a digital portrait of the system's normal operating state, and its center point and covariance matrix can be accurately calculated and stored. After entering the online operation phase, every sampling period, such as every minute, the system inputs the latest real-time sensor data into the encoder part of this pre-trained autoencoder to generate real-time operating condition feature vectors.

[0035] Understandably, to quantify the degree of drift, this application can quantify statistical deviation by calculating the Mahalanobis distance or other statistical distances between the real-time operating condition feature vector and the center point of the baseline operating condition feature space. Mahalanobis distance is an advanced statistical distance that considers the correlation between different dimensions of the data. Unlike simple Euclidean distance, it introduces the covariance matrix of the baseline operating condition feature space distribution, performing scale-independent and coordinate axis decoupling processing on the data. Therefore, it can more accurately measure the degree of anomalous deviation of new data points from the data distribution. The larger the calculated Mahalanobis distance value, the more severe the deviation of the current real-time operating condition feature vector from the baseline operating condition feature space, i.e., the more significant the operating condition drift. Finally, the calculated Mahalanobis distance value is compared with a preset threshold. If the Mahalanobis distance exceeds the threshold, the system determines that a significant operating condition drift has occurred. This threshold can be scientifically set based on the statistical distribution of Mahalanobis distance in historical data. For example, the 95% or 99% quantile of the Mahalanobis distance distribution under historical normal operating conditions can be used as the threshold, thereby ensuring the statistical significance and reliability of the drift determination. This feature space distance-based judgment method, compared to the traditional passive method that relies on model prediction error exceeding the limit, can detect changes in the internal state of the system earlier and more proactively, thus gaining valuable time for subsequent compensation measures.

[0036] S04: When a significant operating condition drift is determined to occur, a small amount of real-time process data under the current operating condition is collected, and an online residual compensation model is constructed to predict the residual between the predicted value of the benchmark prediction model under the current operating condition and the actual process index value.

[0037] Understandably, once the drift signal is triggered, the system will initiate a brief data acquisition window, for example, over the next 30 minutes to an hour, collecting several to dozens of complete process data samples. Each sample includes the crushing and flotation operating parameters at that moment, the actual values ​​of key process indicators obtained through online instruments or rapid manual analysis, such as the actual metal liberation degree or concentrate grade, and the predicted values ​​given by the benchmark prediction model based on the same operating parameters. For each sample, the system calculates the corresponding residual value, which is equal to the actual process indicator value of that sample minus the predicted value of the benchmark prediction model. This residual value represents the prediction deviation of the benchmark prediction model under the current drift condition. After collecting a sufficient number of samples containing residual values, the system can construct a residual compensation model online.

[0038] Specifically, the residual compensation model is a lightweight regression model that is rapidly trained online using a small amount of real-time process data, with a structural complexity far lower than the baseline prediction model. For example, it can be a simple linear regression model, a K-nearest neighbor regression model, or a shallow neural network containing only a single hidden layer and a small number of neurons. This lightweight design ensures that its training process can be completed within seconds or tens of seconds, meeting the needs of online real-time applications. Furthermore, the input information of the residual compensation model includes real-time operating condition feature vectors; specifically, the input of the residual compensation model is a real-time operating condition feature vector that characterizes the current operating condition drift state, while its output is a prediction of the calculated residual value. In this way, the residual compensation model directly establishes an explicit mapping relationship between the operating condition feature vector and the baseline model prediction error caused by this pattern, i.e., the residual. This means that the residual compensation model learns to intelligently predict the magnitude and direction of the error that the baseline model will produce based on the direction and distance of the current operating condition deviating from the baseline state. Compared to methods that merely attempt to predict the error evolution trend from a time series perspective, it has stronger mechanistic interpretability and higher prediction accuracy.

[0039] S05: The prediction results of the baseline prediction model and the prediction results of the residual compensation model are fused to generate a final process index prediction value that can reflect the current operating condition drift state.

[0040] Understandably, in each new control cycle, the system first inputs a set of operating parameters to be evaluated into the baseline prediction model to obtain preliminary predicted values ​​of the process indicators, i.e., the baseline predicted values. Simultaneously, the system inputs the real-time operating condition feature vector extracted by the autoencoder at the current moment into the newly trained online residual compensation model to obtain the predicted value of the baseline model error under the current operating condition, i.e., the predicted residual. Finally, the final predicted value of the process indicators is obtained by a simple algebraic addition of the baseline predicted value and the predicted residual; that is, the final predicted value of the process indicators equals the baseline predicted value plus the predicted residual. This fused predicted value retains the profound understanding of the long-term operating laws of the system contained in the baseline model, and through the residual compensation term, dynamically corrects the systematic deviations introduced by the drift of the current operating conditions in real time.

[0041] S06: Based on the predicted final process indicators, an optimization algorithm is used to generate and output a set of collaborative control setpoints that can optimize the expected mineral processing technical and economic indicators. The collaborative control setpoints include the setpoints of the crushing and grinding unit operating parameters and the flotation unit operating parameters, so as to realize the dynamic collaborative control of the crushing and grinding process and the flotation process.

[0042] It is understandable that the adjustable collaborative control setpoints are specific, executable engineering parameters. Among these, the setpoints for the grinding unit's operating parameters may include at least one of the following: target value for the mill feed rate, setpoint for the mill rotation speed, target value for the hydrocyclone inlet pressure, or planned value for the amount of steel balls added to the mill. The setpoints for the flotation unit's operating parameters may include at least one of the following: target value for the liquid level in each flotation cell, setpoint for the aeration rate in each flotation cell, selection of flotation reagent type and ratio scheme, setting of flotation reagent addition points at different locations, or target value for the specific reagent addition amount. The goal of the optimization algorithm is to find the optimal combination of these adjustable parameters so that the comprehensive evaluation function reaches its optimum.

[0043] Understandably, the final technical and economic indicators for mineral processing can be a comprehensive evaluation function that takes into account multiple sub-objectives such as concentrate recovery rate, concentrate grade, comprehensive energy consumption per unit of ore processing, or comprehensive reagent consumption per unit of ore processing. For example, this comprehensive evaluation function can be defined as: the comprehensive indicator equals weighting coefficient 1 multiplied by concentrate recovery rate plus weighting coefficient 2 multiplied by concentrate grade minus weighting coefficient 3 multiplied by comprehensive energy consumption minus weighting coefficient 4 multiplied by comprehensive reagent consumption; these weighting coefficients can be flexibly adjusted by the plant management based on current market prices, production tasks, and environmental protection requirements.

[0044] To make the optimization process more stable and consistent with the process mechanism, this application uses the target metal liberation distribution as an intermediate objective for process optimization control. Metal liberation directly determines the theoretical upper limit of flotation recovery; therefore, the optimization process is designed as a hierarchical search. Optimization algorithms, such as genetic algorithms or particle swarm optimization algorithms, utilize a fusion prediction model to evaluate the final technical and economic indicators of each candidate combination of operating parameters when searching for the optimal solution. During the iteration process, the optimization algorithm first explores the crushing and grinding operating parameter combinations that can achieve a certain ideal metal liberation distribution. Then, based on this, it further searches for a matching flotation operating parameter combination that maximizes the final economic indicators. By introducing metal liberation as an intermediate physical objective, the optimization process is anchored on a track consistent with the scientific principles of mineral processing, effectively avoiding the algorithm getting trapped in purely mathematical local optima or generating control commands that violate common production sense. After several iterations, the optimization algorithm converges and outputs a set of optimal cooperative control setpoints. These set values ​​are then sent to the underlying process control system of the concentrator, such as a distributed control system, which automatically adjusts various actuators, such as the frequency converter of the feed conveyor belt, the frequency converter of the pump, the valve opening and the speed of the dosing pump, thereby completing a closed-loop, dynamic, and coordinated control of the crushing and flotation process.

[0045] Furthermore, while online residual compensation models can effectively address short-term, rapid operational drift, long-term, slow, and cumulative system changes, such as continuous wear and tear of key equipment components, performance degradation of sensors, or permanent changes in mineral resources, ultimately need to be reflected and solidified in the baseline prediction model. To ensure the long-term effectiveness and sustainability of the entire system, this method includes offline maintenance steps for the baseline prediction model. The triggering conditions for these maintenance steps can be dual. Specifically, it can be a fixed maintenance cycle, such as once every six months or once a year; or it can be based on the cumulative degree of operational drift, for example, when the system detects that the residual compensation model is frequently triggered, or when the magnitude of residual compensation remains at a high level, the maintenance process is automatically triggered.

[0046] At the start of the maintenance process, technicians collect all high-quality process data accumulated during the online operation phase since the last baseline model update. This data naturally includes the system's operating records under various drift conditions. Using this updated and more comprehensive dataset, the baseline prediction model is retrained offline. Retraining can be done by training a completely new model from scratch or by fine-tuning the parameters of the existing model. The updated baseline prediction model will be able to more accurately reflect the inherent patterns of the system in its current stage. Simultaneously, the baseline feature space used for determining operating condition drift also needs to be recalculated and redefined using the updated data. By combining this periodic offline maintenance with online real-time compensation, this application constructs a dual-timescale adaptive optimization system that can respond quickly to instantaneous changes and fundamentally correct long-term evolution, thereby ensuring the excellent performance and high robustness of the collaborative optimization method throughout its entire lifecycle.

[0047] In summary, this application, by combining a stable benchmark prediction model and a flexible online residual compensation model, successfully solves the common problem in existing technologies where optimized models struggle to adapt to drifting industrial process conditions. Those skilled in the art should recognize that the above embodiments are merely examples, and any equivalent substitutions or improvements made in accordance with the spirit of this application should be included within the scope of protection of this application.

[0048] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0049] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0050] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0051] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0052] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0053] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for optimizing a comminution-flotation synergistic process based on metal liberation degree control, characterized by, The method comprises the following steps: real-time acquisition of multi-dimensional sensor data in a grinding-flotation production process, the multi-dimensional sensor data comprising process data of a grinding unit and process data of a flotation unit; offline establishment of a benchmark prediction model for representing a nonlinear mapping relationship between grinding unit operation parameters, flotation unit operation parameters and process indicators including metal dissociation degree based on historical process data; in an online running stage, extraction of a real-time working condition feature vector from the real-time acquired multi-dimensional sensor data, online quantification and determination of a working condition drift state of a current process system by calculating a statistical deviation between the real-time working condition feature vector and a preset benchmark working condition feature space; when a significant working condition drift is determined, a small amount of real-time process data under the current working condition is collected, and a residual compensation model for predicting a residual between a predicted value of the benchmark prediction model under the current working condition and an actual process indicator value is constructed online; fusion of a prediction result of the benchmark prediction model and a prediction result of the residual compensation model to generate a final process indicator prediction value capable of reflecting the current working condition drift state; based on the final process indicator prediction value, a set of coordinated control set values capable of optimizing expected beneficiation technical and economic indicators are generated and output by solving through an optimization algorithm, the coordinated control set values including set values of grinding unit operation parameters and set values of flotation unit operation parameters to realize dynamic coordinated regulation and control of the grinding process and the flotation process.

2. The method of claim 1, wherein, The determination of a significant working condition drift specifically comprises: quantification of the statistical deviation by calculating a Mahalanobis distance between the real-time working condition feature vector and a center point of the benchmark working condition feature space, and comparison of the statistical deviation with a preset threshold value, when the statistical deviation is greater than the preset threshold value, a significant working condition drift is determined.

3. The method of claim 2, wherein, The real-time working condition feature vector is a low-dimensional and high-information-content feature vector extracted from the multi-dimensional sensor data through a self-encoder, principal component analysis or slow feature analysis data dimensionality reduction algorithm.

4. The method of claim 1, wherein, The benchmark prediction model is a Gaussian process regression model, a support vector regression model or a deep neural network model.

5. The method of claim 1, wherein, The residual compensation model is a lightweight regression model trained online quickly using a small amount of real-time process data, and has a lower structural complexity than the benchmark prediction model.

6. The method of claim 5, wherein, The input information of the residual compensation model includes the real-time working condition feature vector, thereby establishing a direct association between the working condition drift state and the prediction residual of the benchmark prediction model.

7. The method of claim 1, wherein, The grinding unit operation parameters include at least one of a mill feed amount, a mill rotation speed, a cyclone inlet pressure or a mill steel ball supplement amount; and the flotation unit operation parameters include at least one of a flotation tank liquid level, a flotation tank aeration amount, a flotation reagent type, a flotation reagent addition position or a flotation reagent addition amount.

8. The method of claim 1, wherein, The dynamic coordinated regulation and control takes a target metal dissociation degree distribution as an intermediate target of process optimization control, and takes optimization of final beneficiation technical and economic indicators as a final target.

9. The method of claim 8, wherein, The final beneficiation technical and economic index includes at least one of concentrate recovery rate, concentrate grade, comprehensive energy consumption per unit of ore processing or comprehensive reagent consumption per unit of ore processing or a combination thereof to form a comprehensive evaluation function.

10. The method of claim 1, wherein, The model maintenance step is also included: according to the accumulated degree of working condition drift or the preset maintenance period, the offline retraining is performed on the benchmark prediction model by using the accumulated process data in the online running stage to update the model structure or model parameters, so as to ensure the long-term prediction accuracy.

Images