Intelligent mine management and control platform and method based on whole process optimization large model of mineral processing
By constructing a unified data warehouse and a set of interconnected models, the problem of process oscillation caused by the decoupling of prediction and decision-making in the mineral processing process was solved. This enabled real-time verification of set values and cross-process collaborative optimization, thereby improving the stability and prediction accuracy of the mineral processing process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN GOLD DESIGN INST
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-09
AI Technical Summary
In existing mineral processing, the decoupling between prediction and decision-making leads to process oscillations caused by trial-and-error of setpoints, making it difficult to achieve cross-process collaborative optimization.
A unified data warehouse is constructed to collect heterogeneous data from multiple sources and perform quality verification and spatiotemporal alignment. A set of linked models is formed by using twin prediction sub-models and optimization decision sub-models to generate future state trends and candidate set values, and real-time verification and adjustment are carried out through a closed-loop feedback mechanism.
Reduce process oscillations caused by setpoint probing, improve the consistency of trend prediction and decision-making, enhance setpoint attainability and cross-business item execution stability, and increase the credibility of multi-objective solutions.
Smart Images
Figure CN122175100A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial control technology, and in particular to an intelligent mine management and control platform and method based on a large-scale optimization model of the entire mineral processing process. Background Technology
[0002] The management and control of the mineral processing process is evolving from partial control to a data-driven approach across the entire process. A unified data warehouse and real-time stream processing aggregate multi-source heterogeneous data and complete quality verification, spatiotemporal alignment, and standardization to form mineral processing data that can be used for modeling. Twin prediction uses a long short-term memory network to output trends of multi-step process parameters, and optimization decisions are based on an actor-critic strategy to generate candidate setpoints. The central node is scheduled and operates in a coordinated manner, and distributed to business items such as crushing, grinding and classification, beneficiation and separation, and dewatering and concentration through a message queue.
[0003] Current practices mostly involve offline prediction combined with empirical thresholds or single-loop PID segmented control, decoupling setpoint generation from process dynamics. They lack a mechanism to pre-setpoints as boundary conditions within the same data closed loop for real-time verification and feedback, making it difficult to suppress process oscillations caused by setpoint probing and limiting the achievement of cross-process collaborative optimization. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides an intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process to solve the problem of setpoint testing and process oscillation caused by the decoupling of prediction and decision-making.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, this invention provides an intelligent mine management and control method based on a large-scale optimization model for the entire mineral processing process. This method includes: constructing a unified data warehouse; collecting multi-source heterogeneous data; performing quality verification and spatiotemporal alignment to generate mineral processing data; using the mineral processing data to train a twin prediction sub-model and an optimization decision sub-model to form a linked model set; running the linked model set at the central node of the intelligent mine management and control system to generate future state trends and candidate setpoints, and distributing them to various business items; each business item forming optimization instructions based on the candidate setpoints and preset optimization targets, and issuing them to the underlying control layer for execution; and the new mineral processing data generated during execution flowing back to the unified data warehouse.
[0007] As a preferred embodiment of the intelligent mine management and control method based on the large-scale optimization model of the entire mineral processing process described in this invention, the following features are provided: a distributed storage architecture and real-time data stream processing are adopted in the process of constructing a unified data warehouse; the multi-source heterogeneous data includes real-time sensor data of mine production equipment, environmental monitoring data, production plan data and historical operation record data; and the mineral processing data includes standardized data sequences.
[0008] As a preferred embodiment of the intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process described in this invention, the specific steps for forming the linkage model set are as follows: The twin prediction sub-model is trained using mineral processing data to optimize the predictive ability of mineral processing data; Based on the trained twin prediction sub-model, the ability to generate the optimized decision sub-model is trained using mineral processing data; The trained twin prediction sub-model and the optimization decision sub-model are coupled through a data interface to establish an inter-model feedback mechanism to form a linked model set, and the linked model set is then validated and its parameters are tuned.
[0009] As a preferred embodiment of the intelligent mine management and control method based on the large-scale optimization model of the entire mineral processing process described in this invention, the intelligent mine management and control central node includes a scheduling linkage model set running in the mine network center, and managing the distribution tasks of the data generated by the linkage model set to various business items.
[0010] As a preferred embodiment of the intelligent mine management and control method based on the large-scale optimization model of the entire mineral processing process described in this invention, the business items include ore crushing, grinding and classification, beneficiation and separation, and dewatering and concentration.
[0011] As a preferred embodiment of the intelligent mine management method based on a large-scale optimization model of the entire mineral processing process described in this invention, the specific steps for generating future state trends and candidate setpoints and distributing them to various business items are as follows: Real-time mineral processing data is input into the linkage model set at the central node of the intelligent mine management and control system; the linkage model set outputs the state trend prediction results of the future mineral processing process. Based on the state trend prediction results, multiple candidate setting value schemes are generated for each business item and distributed to the local execution terminal corresponding to each business item.
[0012] As a preferred embodiment of the intelligent mine management and control method based on the large-scale optimization model of the entire mineral processing process described in this invention, the candidate setting value is a set of adjustment values for the operation parameters of each business item; the preset optimization target is a quantitative indicator for redistributing the quantitative weight values of each business item.
[0013] As a preferred embodiment of the intelligent mine management and control method based on the large-scale optimization model of the entire mineral processing process described in this invention, the process of forming optimization instructions adopts a combination of multi-objective optimization algorithm and local constraints of business items to transform candidate set values into specific executable control instructions.
[0014] As a preferred embodiment of the intelligent mine management and control method based on the large-scale optimization model of the entire mineral processing process described in this invention, the step of returning the new mineral processing data generated during execution to the unified data warehouse includes: collecting the new mineral processing data generated after the bottom control layer executes the optimization instructions, transmitting and unifying the format of the new mineral processing data in real time, and updating it to the unified data warehouse.
[0015] Secondly, this invention provides an intelligent mine management and control platform based on a large-scale optimization model for the entire mineral processing process, including a data governance module, a model training module, a scheduling and distribution module, and an optimization execution module. The data governance module is used to construct a unified data warehouse, collect multi-source heterogeneous data, perform quality verification and spatiotemporal alignment, and generate mineral processing data. The model training module is used to train twin prediction sub-models and optimization decision sub-models using the mineral processing data, forming a linked model set. The scheduling and distribution module is used to run the linked model set at the central node of the intelligent mine management and control system, generate future state trends and candidate setpoints, and distribute them to various business items. The optimization execution module is used by each business item to form optimization instructions based on the candidate setpoints and preset optimization targets, and issue them to the underlying control layer for execution. The new mineral processing data generated during execution is fed back into the unified data warehouse.
[0016] The beneficial effects of this invention are as follows: by establishing a callback-driven closed-loop feedback within the application programming interface, the candidate setpoints generated by the optimization decision are immediately encapsulated as input boundary conditions for twin prediction, triggering a new round of state trend prediction. The prediction results are synchronously fed back for subsequent decision-making. The setpoint evaluation and process response verification are unified into a data closed loop, forming a pre-verification of setpoints, reducing invalid trials and process oscillations, improving the consistency between trend prediction and decision-making, enhancing the attainability of setpoints and the stability of cross-business item execution, and providing highly reliable input for multi-objective solutions. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of an intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process.
[0019] Figure 2 This is a module diagram of an intelligent mine management and control platform based on a large-scale optimization model of the entire mineral processing process.
[0020] Figure 3 A flowchart for message queue distribution.
[0021] Figure 4 A flowchart for data backflow updates. Detailed Implementation
[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process, including the following steps: S1: Build a unified data warehouse, collect heterogeneous data from multiple sources, perform quality verification and spatiotemporal alignment, and generate mineral processing data.
[0026] S1.1: Distributed storage architecture and real-time data stream processing are adopted in the process of building a unified data warehouse.
[0027] Specifically, the distributed storage architecture consists of a master node and multiple data nodes.
[0028] The master node is responsible for managing the logical structure and metadata of the file directory and maintaining the correspondence between data blocks and data nodes. Each data node acts as an independent physical server, responsible for storing data blocks and handling read and write requests from clients.
[0029] The data writing process begins with the client requesting the location information of the target data node from the master node.
[0030] The client divides the data to be written into fixed-size data blocks, and then, based on the list of data node addresses returned by the master node, pipelines these data blocks to multiple data nodes in the list. Each data block is written to multiple data nodes simultaneously, forming data replicas for redundant storage. When reading data, the client obtains the address of the data node containing the required data block from the master node, establishes a connection directly with the corresponding data node, and reads the data.
[0031] Example: The data block size is typically set to 64MB. Each data block creates 3 replicas by default and is distributed across data nodes on different racks. The master node monitors the cluster status by periodically receiving heartbeat signals from the data nodes.
[0032] Furthermore, real-time data stream processing works in conjunction with a distributed storage architecture, with the stream processing engine continuously receiving data streams from the data acquisition end; the data streams enter the stream processing engine in the form of a continuous sequence of events, and the stream processing engine processes the data records one by one or in micro-batch.
[0033] It should be noted that the stream processing engine's processing includes data parsing and rule validation. Data parsing converts the raw byte stream into structured data records (e.g., parsing field boundaries, data types, and timestamps according to the communication protocol message format, splitting the continuous byte stream into data records containing acquisition time, device identifier, and measurement value fields). Rule validation filters and labels the data records according to business rules, including constraints on the range of measurement values, data field integrity requirements, and time order consistency conditions. Data records that pass the rule validation are immediately written to data nodes in the distributed storage architecture and persistently stored.
[0034] S1.2: Multi-source heterogeneous data includes real-time sensor data from mining production equipment, environmental monitoring data, production plan data, and historical operation record data; mineral processing data includes standardized data sequences.
[0035] Specifically, the structured data records output from real-time data stream processing serve as an initial set of multi-source heterogeneous data. Subsequently, quality checks are performed on this initial set of multi-source heterogeneous data. The quality checks include integrity checks to verify that there are no missing data, accuracy checks to verify that the data values are reasonable through predefined range rules, and consistency checks to ensure that the data logic is consistent.
[0036] It should be noted that the predefined range rules specifically represent the upper and lower limits of the values corresponding to each data point, forming an allowed value range. The accuracy verification process compares the collected data values with the corresponding allowed value range. If the data value is within the range, it is considered reasonable; if it exceeds the range, it is marked as abnormal. For example, for the data "crusher speed," based on equipment specifications and historical operation records, the upper limit of the predefined range rule is set to 500 revolutions per minute, and the lower limit is set to 100 revolutions per minute.
[0037] Performing spatiotemporal alignment aligns the quality-verified data in both time and space.
[0038] Among them, time alignment unifies data with different sampling rates to a common timestamp through linear interpolation, and spatial alignment associates device data with a unified geographic reference system through coordinate mapping, ultimately outputting a set of well-aligned data in both time and spatial dimensions.
[0039] The spatiotemporally aligned regular dataset is used as input, and data standardization is performed to generate mineral processing data. The data standardization process uses the min-max normalization method.
[0040] It should be noted that the min-max normalization method scales each value of a data field to the range of zero to one based on the historical maximum and minimum values of each data field in the normalized dataset, thereby generating a standardized data sequence in a uniform format.
[0041] S2: Use mineral processing data to train twin prediction sub-models and optimization decision sub-models respectively, forming a set of linked models.
[0042] S2.1: Use mineral processing data to train the twin prediction sub-model to optimize the prediction capability of mineral processing data.
[0043] Specifically, mineral processing data is used as the training dataset to train the twin prediction sub-model.
[0044] The mineral processing data contains a standardized data sequence of historical time series. The twin prediction sub-model uses a long short-term memory network architecture to process the time series data. The long short-term memory network consists of an input gate, a forget gate, and an output gate, which is used to capture time dependencies.
[0045] The training process divides the mineral processing data into a training set and a validation set in chronological order.
[0046] The training set is used for model parameter learning, and the validation set is used to monitor overfitting. The training objective of the Siamese prediction sub-model is to minimize the difference between the predicted output and the true value. The loss function is the mean squared error loss. The gradient is calculated using the backpropagation algorithm, and the network weights are updated using a stochastic gradient descent optimizer. Training iterations are performed until the loss function stabilizes on the validation set.
[0047] The mean squared error loss function is expressed as follows: ; In the formula, The loss value is a quantitative indicator representing the difference between the predicted result and the actual value. For the sample size, For the true value, For predicted values, For sample index, Indicates the first The actual observed value corresponding to each sample Indicates the twin prediction sub-model for the first The predicted value for each sample, Indicates the calculation of the first The difference between the true value and the predicted value of a sample.
[0048] It should be noted that training iterations until the loss function stabilizes on the validation set means that the training process continues until, over multiple consecutive training epochs, the loss function value on the validation set no longer exhibits a continuously decreasing trend, and its fluctuation becomes relatively flat. For example, when the fluctuation range of the validation set loss value over ten consecutive training epochs is less than a preset small threshold, it is considered to have reached a stable state. The preset small threshold is set before training begins based on an empirical proportion of the initial validation set loss value; an example value range is 0.001 to 0.05.
[0049] S2.2: Based on the trained twin prediction sub-model, use mineral processing data to train and optimize the generation capability of the decision sub-model.
[0050] Specifically, based on the trained twin prediction sub-model, the optimization decision sub-model is trained.
[0051] The training of the optimized decision sub-model is based on mineral processing data, and the dynamic environment of the mineral processing process is simulated using a twin prediction sub-model. The optimized decision sub-model generates decision actions in the simulated environment using the policy gradient method. The policy gradient method is based on an actor-commentator architecture, where the actor network outputs the probability distribution of actions, and the commentator network evaluates the value of the state.
[0052] The actor-critic architecture works as follows: at each decision moment, the current state vector of the simulated environment, constructed by the twin prediction sub-model, is read as input. The actor network performs forward propagation on the state vector to obtain the probability distribution of each candidate action in the action space. From the action probability distribution, a specific action is selected as the current candidate decision action according to the random sampling or greedy sampling rule and applied to the simulated environment to update the environment state. At the same time, the state sequence, action sequence, and the instantaneous feedback signal from the simulated environment are recorded.
[0053] The critic network performs forward inference with the current state vector as input and calculates the state value estimate and advantage function by combining the immediate reward signal corresponding to the current state. The advantage function represents the additional benefit measure of taking the current decision action in the current state relative to the average policy, and can be calculated as the difference between the action value function and the state value function.
[0054] The policy gradient is obtained by weighting the action probability distribution output by the actor network using the advantage function to update the actor network parameters. At the same time, the critic network parameters are updated with the temporal difference error as the objective. Through multiple rounds of alternating updates, the decision actions generated in the simulation environment gradually converge around the evaluation index, thus completing the decision action learning process of the optimized decision sub-model under the policy gradient method.
[0055] The training process uses the predicted output of the twin prediction sub-model as the input state of the optimization decision sub-model. The output action of the optimization decision sub-model corresponds to the candidate setpoint. The reward function is defined according to the preset optimization objective, and calculates a weighted combination of improved ore beneficiation recovery rate, reduced energy consumption, and reduced environmental pollution. Training collects experience trajectories through simulated environmental interactions, calculates the policy gradient using the advantage function, and updates the network parameters to maximize the expected cumulative reward.
[0056] The optimization objectives are determined before training by setting target value ranges and corresponding weights for three indicators: ore beneficiation recovery rate, unit energy consumption, and environmental emissions. For example, the target range for ore beneficiation recovery rate is set to 90%–95%, the target range for unit energy consumption is set to 2.0–2.5 kWh / t, and the target range for environmental emissions is set to a dimensionless emission index not exceeding 0.8. During training, these target value ranges and weights are consistently used as parameters in the reward function calculation. The evaluation index is the expected cumulative reward calculated under the constraints of the reward function, which weights the gains from improved ore beneficiation recovery rate, reduced unit energy consumption, and reduced environmental emissions into a single scalar.
[0057] S2.3: Couple the trained twin prediction sub-model and the optimization decision sub-model through a data interface, establish an inter-model feedback mechanism to form a linked model set, and verify and tune the linked model set.
[0058] Specifically, the trained twin prediction sub-model and the optimization decision sub-model are coupled through a data interface, which is implemented as an application programming interface (API). The API allows data flow between the twin prediction sub-model and the optimization decision sub-model and agrees on a unified feature representation format and state encoding format, enabling the trained twin prediction sub-model and the optimization decision sub-model to interact in the same state space and establishing an inter-model feedback mechanism. The API connects the three stages of candidate setting value generation, state prediction and decision update in a single call chain, avoiding the need to call the two types of sub-models through independent processes.
[0059] The feedback mechanism uses the candidate setpoints generated by the optimized decision-making sub-model as the input boundary conditions of the twin prediction sub-model. The twin prediction sub-model re-predicts the state trend and feeds the prediction results back to the optimized decision-making sub-model for decision adjustment. The twin prediction sub-model and the optimized decision-making sub-model, which operate in a coupled manner within the feedback mechanism, together constitute a linkage model set. In each decision cycle, the linkage model set is used to perform at least one virtual rolling prediction on the candidate setpoints. The candidate setpoints are retained for subsequent optimization processes only when the prediction results meet the preset safety constraints and benefit constraints. This ensures that the candidate setpoints undergo pre-screening by the full-process dynamic response provided by the twin prediction sub-model before actual issuance.
[0060] Specifically, in the feedback mechanism, the optimization decision sub-model generates multiple candidate setpoint combinations based on the unified data warehouse and the working conditions of each business item. The application programming interface encapsulates each candidate setpoint combination as an input boundary condition and passes it to the twin prediction sub-model.
[0061] The twin prediction sub-model performs multi-step rolling state trend prediction on candidate setpoint combinations and outputs the change trajectories of output indicators, quality indicators, energy consumption indicators and safety constraint indicators. The optimization decision sub-model constructs a unified evaluation index based on the state trend prediction results to sort and prune candidate setpoint combinations. The unified evaluation index maps the prediction deviation and revenue performance of multiple business items and multiple objectives into a single scalar, which is used to eliminate candidate setpoint combinations that violate safety constraints or cause process fluctuations to exceed limits.
[0062] Within a continuous decision-making cycle, the multi-objective optimization weights and hyperparameter configurations of the linkage model set are adaptively adjusted based on the prediction deviation, enabling the twin prediction sub-model and the optimization decision sub-model to converge collaboratively within the same data loop. Furthermore, a unified evaluation index is used to simultaneously drive the selection of candidate setpoints and the updating of hyperparameters, forming an integrated linkage calibration process. Additionally, a virtual rolling prediction is used to select setpoints and coordinate the improvement of prediction accuracy and decision benefits throughout the entire process.
[0063] The linked model set was validated using an independent test set to evaluate its overall performance. Performance metrics included prediction accuracy and decision return rate. The hyperparameter tuning process employed Bayesian optimization to search the hyperparameter space and optimize the generalization ability of the linked model set. During the validation phase, prediction accuracy and decision return rate were combined into a unified evaluation metric, which was used as the objective function in Bayesian optimization. This ensured that the hyperparameters of the twin prediction sub-model and the optimized decision sub-model converged synchronously with joint performance as a constraint. The optimal hyperparameter configuration obtained by Bayesian optimization simultaneously constrained the collaborative performance of the twin prediction sub-model and the optimized decision sub-model, rather than tuning each individual model independently.
[0064] It should be noted that the inter-model feedback mechanism is established by creating closed-loop control logic; The closed-loop control logic is implemented within the application programming interface (API). It listens for output events from the optimization decision sub-model by setting callback functions. When a candidate setpoint is detected, the callback function is automatically triggered, encapsulating the candidate setpoint into a data packet conforming to the input format of the twin prediction sub-model. This data packet is then passed to the twin prediction sub-model as input boundary conditions. The twin prediction sub-model re-predicts the state trend based on the input boundary conditions. The prediction result is then fed back to the state input port of the optimization decision sub-model through another predetermined channel of the API to adjust subsequent decisions. The closed-loop control logic maintains an association cache between candidate setpoints and corresponding state trends within the API, ensuring a one-to-one correspondence between the decision states within the optimization decision sub-model and the output states of the twin prediction sub-model. Based on this association cache, it maintains the mapping relationship between candidate setpoints, state trends, and unified evaluation indicators, providing a consistent data foundation for subsequent linkage calibration.
[0065] A preferred approach is to establish a closed-loop feedback between the twin prediction sub-model and the optimization decision sub-model using an application programming interface (API). Callback functions are used to encapsulate candidate setpoints as boundary conditions in real time to drive a new round of state trend prediction, and the prediction results are fed back for the next round of decision-making. During the closed-loop feedback process, candidate setpoints are limited to fall within the feasible region verified by the twin prediction sub-model, and stationarity constraints are set for state trend changes in continuous decision-making cycles. Simultaneously, independent test sets and Bayesian optimization are used for joint verification and parameter tuning. A synergistic gain between prediction performance and decision benefits is achieved through Bayesian search of the joint evaluation index. Thus, an adaptive evolutionary link is formed within a unified data closed loop, from candidate setpoint generation, virtual rolling prediction, feasible region and stationarity constraint verification to decision adjustment and linkage calibration.
[0066] Compared to the decoupled control method that trains prediction models in an offline environment and performs segmented control on the control side based on preset rules or experience thresholds, this invention uses closed-loop feedback to move the impact of setpoints on process dynamics to the pre-decision verification stage, reducing the number of trials and the amplitude of process oscillations. Under the premise of meeting safety constraints, it synergistically improves the prediction accuracy and decision benefits of the entire process and enhances the generalization ability.
[0067] S3: Run the linkage model set at the central node of intelligent mine management and control, generate future state trends and candidate set values, and distribute them to various business items.
[0068] S3.1: The intelligent mine management and control hub node includes the scheduling linkage model set running in the mine network center, and manages the distribution tasks of the data generated by the linkage model set to various business items.
[0069] Furthermore, the intelligent mine management and control hub node is deployed in the mine network center. The linkage model set is loaded onto the computing node for execution through a task scheduler (such as the job scheduler in a distributed computing framework). The task scheduler dynamically allocates central processing unit and memory resources according to the utilization rate of computing resources and monitors the running status of the linkage model set to ensure real-time performance.
[0070] It should be noted that a mining network center refers to a central node that carries the main data flow and control flow of the mining production process and centrally deploys the core switching equipment and computing and storage resources of the industrial control network. For example, it is a data center used by mining enterprises to run production control and data processing services.
[0071] S3.2: Input real-time mineral processing data into the linkage model set at the intelligent mine management and control center node; the linkage model set outputs the state trend prediction results of the future mineral processing process.
[0072] Specifically, real-time mineral processing data is input into the linkage model set, and the real-time mineral processing data is obtained from the unified data warehouse in real time through the data interface.
[0073] The linked model set outputs the state trend prediction results of the future mineral processing process, which are generated by the twin prediction sub-models in the linked model set.
[0074] The twin prediction sub-model processes input data based on a long short-term memory network and outputs predicted values of process parameters for multiple future time steps, expressed as: ; In the formula, The output of the twin prediction sub-model represents the prediction of the future. Predicted process parameters for the trend of mineral processing status at each time step. This indicates the time range covered by the prediction, i.e., the next time step. To the future time step A continuous period of time, This represents the forward computation function of a Long Short-Term Memory (LSTM) network. For the current moment Real-time mineral processing data input into the twin prediction sub-model, This represents the set of all weight parameters of a Long Short-Term Memory (LSTM) network model.
[0075] It should be noted that the forward computation function of the Long Short-Term Memory (LSTM) network involves the collaborative updating of the input gate, forget gate, output gate, and cell state, specifically including: The forget gate controls the degree to which information in each dimension of the cell state from the previous time step is retained. The output of the forget gate is a vector between zero and one.
[0076] The calculation for the forget gate is represented as follows: ; in, Indicates the forget gate at time step The output vector, It is the sigmoid activation function. Here is the weight matrix for the forget gate. This is the hidden state from the previous time step. Input for the current time step. For the bias term of the forget gate, As an identifier, it is used to identify whether the variable belongs to the forget gate.
[0077] The input gate calculates gate weights using a sigmoid function for nonlinear mapping based on the input features of the current time step and the hidden state of the previous time step. These gate weights are applied to the candidate information vector to determine the new information content to be written into the cell state. The cell state is updated based on the outputs of the forget gate and the input gate, while the output gate determines the hidden state output of the current time step. By repeating the gate calculation and state update sequentially step by step, a complete predicted output sequence is obtained. This predicted output sequence is the result of future... The sequence formed by arranging the predicted values of process parameters for each time step in chronological order.
[0078] It should be noted that the input features of the current time step refer to the feature vector composed of process parameters and state variables collected and organized from the mineral processing process under the current time index. For example, it includes a set of normalized values such as slurry concentration, feed rate, equipment speed, current and particle size index at the current moment.
[0079] S3.4: Based on the state trend prediction results, generate multiple candidate setting value schemes for each business item and distribute them to the local execution terminal corresponding to each business item.
[0080] Specifically, based on the predicted output sequence, multiple candidate setpoint schemes are generated for each business item. The generation process is executed by the optimization decision sub-model in the linkage model set. The optimization decision sub-model is based on the actor-critic architecture. The actor network outputs the action probability distribution according to the predicted values of the current process parameters, and generates multiple sets of different candidate setpoint schemes through sampling. The candidate setpoint schemes are generated separately for the ore crushing business item, the grinding and classification business item, the beneficiation and separation business item, and the dewatering and concentration business item.
[0081] Candidate settings for the ore crushing business include crusher speed and discharge port size; candidate settings for the grinding and classification business include mill feed rate and classifier overflow concentration; candidate settings for the beneficiation and separation business include reagent addition amount and flotation machine liquid level; and candidate settings for the dewatering and thickening business include thickener rake speed and underflow discharge frequency.
[0082] For example, in the dewatering and thickening business, the thickener rake speed can be understood as the rotation speed that drives the rake arm inside the thickener to rotate (such as setting it to 0.2 to 0.4 rpm to control the mud layer thickness and flocculation and sedimentation effect), and the underflow discharge frequency can be understood as the number of times the underflow discharge pump is periodically turned on (such as setting it to discharge once every 5 minutes to adjust the underflow concentration and material level stability).
[0083] Data is distributed to the local execution terminals corresponding to each business item through message queues. The message queues create independent topics according to the business item type. The intelligent mine management and control hub node publishes the corresponding candidate setting value schemes to the corresponding topics, and the local execution terminals of each business item subscribe to the corresponding topics to receive data.
[0084] S4: Each business item generates optimization instructions based on candidate set values and preset optimization targets, and sends them to the underlying control layer for execution; the new ore beneficiation data generated by the execution is fed back to the unified data warehouse.
[0085] S4.1: The candidate setting value is the set of adjustment values for the operation parameters of each business item; the preset optimization target is the quantitative indicator that redistributes the quantitative weight values of each business item.
[0086] Specifically, the operating parameters for each business item include the crusher speed and discharge port size for the ore crushing business item, the mill feed rate and classifier overflow concentration for the grinding and classification business item, the reagent addition amount and flotation liquid level for the beneficiation and separation business item, and the thickener rake speed and underflow discharge frequency for the dewatering and thickening business item.
[0087] The reallocation process calculates the weight adjustment amount based on the historical performance data and real-time operating status of each business item. The historical performance data is obtained from a unified data warehouse, and the real-time operating status is obtained from sensor data streams. The weight adjustment amount is normalized to make the weights of mineral processing recovery rate, energy consumption and environmental pollution equal to one. For example, the weight of mineral processing recovery rate is increased to 0.5, the weight of energy consumption is reduced to 0.3, and the weight of environmental pollution is adjusted to 0.2.
[0088] S4.2: In the process of generating optimization instructions, a multi-objective optimization algorithm is combined with local constraints of business items to transform candidate setpoints into specific executable control instructions.
[0089] Specifically, the multi-objective optimization algorithm uses a weighted sum method, which combines the weight values corresponding to each objective in the preset optimization objectives with the normalized evaluation values of each objective in a weighted linear combination to construct a comprehensive objective function; the local constraints of the business items include the upper and lower limits of equipment operation and the safety threshold, and the local constraints of the business items limit the range of operating parameters in the form of inequalities.
[0090] Furthermore, the optimization process uses candidate setpoints as decision variables, solves for the maximum value of the objective function under the local constraints of the business item, generates the optimal combination of operating parameters, and transforms the optimal combination of operating parameters into specific executable control instructions. The control instructions include parameter setpoints and execution timestamps.
[0091] The control commands are sent to the lower-level control layer via the industrial Ethernet protocol and transmitted to the programmable logic controller (PLC). The PLC then drives the actuators to adjust the operating parameters.
[0092] A superior approach uses a multi-objective weighted sum method to transform candidate setpoints for four business items—crushing, grinding and classifying, beneficiation and separation, and dewatering and concentration—into executable control commands within local constraints such as equipment safety and operational limits. These commands are then distributed via message queues and processed in a unified manner to form a continuous closed loop. Compared to a decoupled method that uses single-loop PID tuning for each equipment and local parameter adjustments based on a single KPI, this invention simultaneously addresses recovery rate, energy consumption, and environmental indicators under given weights through joint solution. This reduces cross-process interactions and parameter tuning workload, improving the success rate of control command execution and overall process performance.
[0093] S4.3: The process of feeding new mineral processing data back to the unified data warehouse includes collecting new mineral processing data generated after the underlying control layer executes optimization instructions, transmitting and unifying the format of the new mineral processing data in real time, and updating it to the unified data warehouse.
[0094] Specifically, during the data acquisition process, the data stream is read in real time through the sensor interface. The data stream contains equipment operating parameters and process indicators.
[0095] The collected data stream is used as the source of new mineral processing data, and is transmitted in real time and processed in a unified format. Real-time transmission uses a message queue to transmit data asynchronously; the unified format processing includes data parsing, timestamp alignment, unit conversion and range standardization steps.
[0096] Furthermore, data parsing converts the received raw signals into structured data records; the parsed data records are timestamped and synchronized to a unified time base; the time-synchronized data records undergo unit conversion, converting the values to standard International Units (SI); and finally, the data undergoes range standardization, scaling the data values to the range of zero to one.
[0097] After all the formatting processes are completed, the new ore beneficiation data is appended to the data nodes of the distributed storage architecture through the data writing interface, thereby updating the unified data warehouse.
[0098] This embodiment also provides an intelligent mine management and control platform based on a large-scale optimization model for the entire mineral processing process, including: a data governance module, a model training module, a scheduling and distribution module, and an optimization execution module; the data governance module is used to build a unified data warehouse, collect multi-source heterogeneous data, perform quality verification and spatiotemporal alignment, and generate mineral processing data; the model training module is used to train twin prediction sub-models and optimization decision sub-models using mineral processing data to form a linkage model set; the scheduling and distribution module is used to run the linkage model set at the central node of the intelligent mine management and control system, generate future state trends and candidate setpoints, and distribute them to various business items; the optimization execution module is used for each business item to form optimization instructions based on candidate setpoints and preset optimization targets, and send them to the underlying control layer for execution; the new mineral processing data generated by the execution is fed back to the unified data warehouse.
[0099] In summary, this invention establishes a callback-driven closed-loop feedback within the application programming interface, instantly encapsulating candidate setpoints generated by optimization decisions into input boundary conditions for twin predictions, triggering a new round of state trend predictions, and synchronously feeding back the prediction results for subsequent decisions. Setpoint evaluation and process response verification are unified into a data closed loop, forming a pre-setpoint verification, reducing invalid trials and process oscillations, improving the consistency between trend prediction and decision-making, enhancing setpoint attainability and cross-business item execution stability, and providing highly reliable input for multi-objective solutions.
[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A smart mine management and control method based on a large-scale optimization model of the entire mineral processing process, characterized by: include, Build a unified data warehouse, collect heterogeneous data from multiple sources, perform quality verification and spatiotemporal alignment, and generate mineral processing data; Using mineral processing data, twin prediction sub-models and optimization decision sub-models are trained separately to form a set of linked models; The linkage model set is run at the central node of the intelligent mine management and control system to generate future state trends and candidate set values, and then distributed to various business items. Each business item generates optimization instructions based on candidate settings and preset optimization targets, which are then sent to the underlying control layer for execution; the new ore beneficiation data generated during execution is fed back into the unified data warehouse.
2. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 1, characterized in that: The construction of a unified data warehouse adopts a distributed storage architecture and real-time data stream processing; the multi-source heterogeneous data includes real-time sensor data from mining production equipment, environmental monitoring data, production plan data, and historical operation record data; the ore beneficiation data includes standardized data sequences.
3. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 2, characterized in that: The specific steps for forming the linkage model set are as follows: The twin prediction sub-model is trained using mineral processing data to optimize the predictive ability of mineral processing data; Based on the trained twin prediction sub-model, the ability to generate the optimized decision sub-model is trained using mineral processing data; The trained twin prediction sub-model and the optimization decision sub-model are coupled through a data interface to establish an inter-model feedback mechanism to form a linked model set, and the linked model set is then validated and its parameters are tuned.
4. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 3, characterized in that: The intelligent mine management and control hub node includes a scheduling linkage model set running in the mine network center, and manages the distribution tasks of the data generated by the linkage model set to various business items.
5. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 4, characterized in that: The business items include ore crushing, grinding and classifying, beneficiation and separation, and dewatering and concentration.
6. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 5, characterized in that: The specific steps for generating future state trends and candidate set values, and distributing them to each business item, are as follows: Real-time mineral processing data is input into the linkage model set at the central node of the intelligent mine management and control system; the linkage model set outputs the state trend prediction results of the future mineral processing process. Based on the state trend prediction results, multiple candidate setting value schemes are generated for each business item and distributed to the local execution terminal corresponding to each business item.
7. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 6, characterized in that: The candidate setting value is a set of adjustment values for the operation parameters of each business item; the preset optimization target is a quantitative indicator for redistributing the quantitative weight values of each business item.
8. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 7, characterized in that: The process of forming optimization instructions uses a combination of multi-objective optimization algorithms and local constraints of business items to transform candidate set values into specific executable control instructions.
9. The intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in claim 8, characterized in that: The process of feeding new mineral processing data back to the unified data warehouse includes collecting new mineral processing data generated after the underlying control layer executes optimization instructions, transmitting and unifying the format of the new mineral processing data in real time, and updating it to the unified data warehouse.
10. An intelligent mine management and control platform based on a large-scale optimization model of the entire mineral processing process, and based on the intelligent mine management and control method based on a large-scale optimization model of the entire mineral processing process as described in any one of claims 1 to 9, characterized in that: It includes a data governance module, a model training module, a scheduling and distribution module, and an optimization and execution module; The data governance module is used to build a unified data warehouse, collect multi-source heterogeneous data, perform quality verification and spatiotemporal alignment, and generate mineral processing data. The model training module is used to train the twin prediction sub-model and the optimization decision sub-model respectively using mineral processing data to form a linked model set; The scheduling and distribution module is used to run the linkage model set at the intelligent mine management and control center node, generate future state trends and candidate set values, and distribute them to each business item. The optimization execution module is used to generate optimization instructions for each business item based on candidate set values and preset optimization targets, and send them to the underlying control layer for execution; the new ore beneficiation data generated by the execution is fed back to the unified data warehouse.