An intelligent optimization control method and system for a beneficiation process based on a knowledge graph

Abstract

The application discloses a kind of based on knowledge graph's mineral processing flow intelligent optimization control method and system, it is related to mineral processing process control technical field, including, acquisition static attribute data and dynamic operation data, build causal knowledge graph, obtain the causal relationship and optimization target constraint between nodes in causal knowledge graph;Based on the causal relationship and optimization target constraint between nodes in causal knowledge graph, data is screened from historical database, and high-value historical data set is formed;According to high-value historical data set, construct reward function model, train strategy function, obtain reinforcement learning control strategy;Run reinforcement learning control strategy, generate control instruction, and the control instruction is issued to target equipment, obtain dynamic optimization control;Real-time monitoring is carried out to dynamic optimization control, when detecting that operating state appears abnormal deviation, based on causal knowledge graph, acquire counterfactual reasoning result.

Description

Technical Field

[0001] This invention relates to the field of mineral processing control technology, and in particular to an intelligent optimization control method and system for mineral processing based on knowledge graphs. Background Technology

[0002] Mineral processing optimization and control involves the coordinated operation of multiple stages, including ore composition, grinding, flotation / magnetic separation, thickening, and filtration. On-site data acquisition links are typically constructed using DCS / PLC and online analysis. Parameter tuning and strategy adjustment are then performed by combining mechanistic models, advanced process control, and data-driven optimization. Conventional methods include forming a knowledge base based on flow charts and laboratory information, using historical samples for empirical regression or heuristic optimization, and continuously revising setpoints based on online feedback. In recent years, there have also been explorations of using knowledge graphs for process knowledge management and reinforcement learning for continuous control, serving the comprehensive goals of improving recovery rate, ensuring grade, and energy economy.

[0003] Within the aforementioned conventional framework, two areas still require improvement: First, knowledge representation and control strategies often exist in parallel, lacking a unified representation based on causal relationships, making it difficult to directly map optimization objectives and constraints into auditable strategy updates. Second, anomaly handling often relies on experience thresholds and offline analysis, making it difficult to provide traceable inferences and closed-loop adaptation regarding "what will happen if a certain control quantity is adjusted." Existing mineral processing process control technologies mainly face the problems of knowledge-policy separation and insufficient consistency in anomaly handling. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a knowledge graph-based intelligent optimization control method for mineral processing to solve the problems of fragmented knowledge strategies and insufficient consistency in anomaly handling.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent optimization control method for mineral processing based on knowledge graphs, comprising,

[0008] Collect static attribute data and dynamic runtime data, construct a causal knowledge graph, and obtain the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints.

[0009] Based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, data is selected from the historical database to form a high-value historical dataset.

[0010] Based on high-value historical datasets, a reward function model is constructed, a policy function is trained, and a reinforcement learning control policy is obtained.

[0011] The reinforcement learning control strategy is run to generate control commands and send them to the target device to obtain dynamic optimization control.

[0012] Real-time monitoring of dynamic optimization control is performed. When an abnormal deviation in the operating status is detected, counterfactual reasoning is performed based on the causal knowledge graph to obtain the counterfactual reasoning results.

[0013] Based on the counterfactual reasoning results, the edge weights in the causal knowledge graph are modified, and the reinforcement learning control strategy is updated through incremental reinforcement learning.

[0014] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the steps of collecting static attribute data and dynamic operational data, constructing a causal knowledge graph, and obtaining the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints are as follows:

[0015] Static attribute data is collected through mineral samples and laboratory tests, while dynamic operational data is collected through DCS / PLC and online analysis.

[0016] Construct a causal knowledge graph in a graph database cluster, freeze the variable list and solidify it into a variable dictionary table;

[0017] The process flow diagram is converted into a set of candidate directed edges. Each candidate directed edge is labeled with a positive or negative sign and monotonicity. The dimensionless edge weight strength is calculated. Based on the positive or negative sign, monotonicity, and dimensionless edge weight strength of the candidate directed edges, the causal relationship between nodes is obtained.

[0018] In a causal knowledge graph, target nodes and constraint nodes are established. Variables are connected to target nodes or constraint nodes through association edges to obtain optimization target constraints.

[0019] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the step of selecting data from a historical database to form a high-value historical dataset based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints includes the following specific steps:

[0020] Based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, a uniform dimensionless utility score is calculated for each sample in the historical database.

[0021] Sort the samples by utility score from high to low, remove any samples that violate hard constraints, retain the same proportion of high-utility-score samples for each process segment to form a candidate sample set, and combine the selected samples to form a high-value historical dataset.

[0022] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the specific steps for constructing a reward function model based on a high-value historical dataset are as follows:

[0023] Based on high-value historical datasets and target nodes and constraint nodes in causal knowledge graphs, calculate the geometric mean of gains and the geometric mean of costs.

[0024] The logarithmic utility is obtained by taking the natural logarithm of the ratio of the geometric mean of the benefit to the geometric mean of the cost, and a reward function model is constructed to calculate the final immediate reward.

[0025] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the specific steps for training the policy function to obtain the reinforcement learning control policy are as follows:

[0026] The high-value historical dataset is divided into training and validation sets according to time sequence. The final immediate reward output by the reward function model is used as the reward signal for supervised early stopping and PPO of the behavior clone. The control policy is obtained by training the policy function through behavior clone hot start and small step fine-tuning based on PPO.

[0027] Perform mini-batch updates on the control policy to obtain a reinforcement learning control policy.

[0028] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the steps of running a reinforcement learning control strategy, generating control commands, and issuing the control commands to the target equipment to obtain dynamic optimization control are as follows:

[0029] Load the reinforcement learning control policy version and configuration, set a fixed control period, and in each control period, input the aligned data snapshot into the reinforcement learning control policy to obtain the original actions item by item in the variable list;

[0030] The system executes a safety envelope on the original action, generates control commands, and sends the control commands to the target device for dynamic optimization control through the industrial control interface. It also monitors any hard constraint triggers and abnormal events, and packages the trigger events and control commands into an exception buffer entry.

[0031] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the following steps are involved: real-time monitoring of dynamic optimization control; when an abnormal deviation in the operating state is detected, counterfactual reasoning is performed based on the causal knowledge graph to obtain the counterfactual reasoning result.

[0032] Take the latest control cycle closed-loop record from the abnormal buffer entries and lock it as the current working sample. Monitor and verify the current working sample and perform causal knowledge graph positioning.

[0033] Based on the current working sample, a unique intervention quantity is identified. Counterfactual reasoning is performed based on the causal knowledge graph. Hypothetical interventions are made on the unique intervention quantity to obtain the counterfactual reasoning results, which are then bound to the unique ID of the current working sample to generate an archive package.

[0034] As a preferred embodiment of the intelligent optimization control method for mineral processing based on knowledge graphs described in this invention, the step of modifying the edge weights in the causal knowledge graph based on counterfactual reasoning results includes the following specific steps:

[0035] Based on the counterfactual reasoning results, the edge weights in the causal knowledge graph are modified. When modifying the edge weights, only the edge pointing to the target quality index of the unique intervention quantity identified this time is updated.

[0036] By comparing the predictions with the actual measurements and adjusting the actual range, the local sensitivity is obtained. Aligned with a fixed process delay, the old edge weights and the local sensitivity are weighted and synthesized to obtain the new edge weights.

[0037] As a preferred embodiment of the knowledge graph-based intelligent optimization control method for mineral processing described in this invention, the step of updating the reinforcement learning control strategy through incremental reinforcement learning specifically includes the following steps:

[0038] Extract fields from the archive package for deduplication, cleaning, and alignment, generate incremental data slices, and update parameters for the reinforcement learning control strategy;

[0039] After the updated candidate policies are tested offline by relative change upper limit and hard constraints, they enter grayscale verification. The old reinforcement learning control policy and the updated reinforcement learning control policy are compared in parallel on the mirror data stream. If the hard conditions are met, they can be released; otherwise, they are rolled back and the reason is recorded. After release, the updated reinforcement learning control policy is marked as the latest reinforcement learning control policy.

[0040] Secondly, this invention provides a knowledge graph-based intelligent optimization control system for mineral processing, comprising,

[0041] The knowledge graph construction module collects static attribute data and dynamic runtime data to construct a causal knowledge graph, thereby obtaining the causal relationships between nodes in the causal knowledge graph and optimization objective constraints.

[0042] The data filtering module, based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, filters data from the historical database to form a high-value historical dataset.

[0043] The control policy module constructs a reward function model based on a high-value historical dataset, trains the policy function, and obtains a reinforcement learning control policy.

[0044] The control command module runs a reinforcement learning control strategy, generates control commands, and sends the control commands to the target device to obtain dynamic optimized control.

[0045] The counterfactual reasoning module monitors the dynamic optimization control in real time. When an abnormal deviation in the operating status is detected, it performs counterfactual reasoning based on the causal knowledge graph and obtains the counterfactual reasoning results.

[0046] The control strategy update module modifies the edge weights in the causal knowledge graph based on counterfactual reasoning results and updates the reinforcement learning control strategy through incremental reinforcement learning.

[0047] The beneficial effects of this invention are as follows: by constructing a causal knowledge graph and establishing optimization target constraints, a unified causal expression of knowledge and manipulative variables is achieved. The edge weights in the causal knowledge graph are corrected and the reinforcement learning control strategy is updated through incremental reinforcement learning, realizing a closed-loop self-evolution between the knowledge layer and the strategy layer. This is used to resolve operational deviations online, ultimately achieving the beneficial effects of alleviating the problems of knowledge-policy fragmentation and insufficient consistency in anomaly handling, and improving indicators and security compliance. Attached Figure Description

[0048] 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.

[0049] Figure 1 This is a flowchart of a knowledge graph-based intelligent optimization control method for mineral processing.

[0050] Figure 2 This is a schematic diagram of an intelligent optimization control system for mineral processing based on knowledge graphs.

[0051] Figure 3 Flowchart for screening high-value historical datasets.

[0052] Figure 4 Flowchart for training reinforcement learning control strategies. Detailed Implementation

[0053] 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.

[0054] 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.

[0055] 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.

[0056] Reference Figures 1-4 This is one embodiment of the present invention, which provides a knowledge graph-based intelligent optimization control method for mineral processing, comprising the following steps:

[0057] S1. Collect static attribute data and dynamic operation data, construct a causal knowledge graph, and obtain the causal relationships between nodes in the causal knowledge graph and the optimization target constraints.

[0058] Static attribute data (such as batch particle size, mineral ratio, and grindability) are collected through mineral samples and laboratory tests, and then stored in the database according to batch identification.

[0059] By using DCS / PLC (Distributed Control / Programmable Logic Controller) and online analysis, dynamic operating data (such as control quantities, process quantities, result quantities, quality and energy consumption, etc.) are collected and aligned to a fixed sampling step size with a unified clock. The time when the change of the control quantity of the parent node is reflected in the result quantity of the child node is represented by a fixed time difference. The aligned parent node control quantity is used as the effective value and aligned to the moment when the result quantity of the child node changes according to the fixed time difference. The effective value is stored in the graph database as the hysteresis attribute of the parent node.

[0060] Construct a causal knowledge graph in a graph database cluster using a node-directed edge pattern, freeze the variable list and solidify it into a variable dictionary table, and group it according to static attributes, control quantities, process quantities, result quantities / quality indicators, and energy consumption / cost, etc.

[0061] Static properties refer to batch particle size distribution, mineral composition ratio, grindability index, mineral phase category, initial solids content of slurry, etc.

[0062] Controlled quantities refer to ore feed rate, water addition rate, reagent addition rate (such as collectors, frothers and inhibitors), aeration rate, stirring speed and valve / liquid level settings, etc.

[0063] Process parameters include grinding fineness, slurry density, pH, redox potential, gas holdup, foam layer velocity, and liquid level.

[0064] The quantitative / quality indicators of the results refer to the target concentrate grade, tailings grade, recovery rate, and yield, etc.

[0065] Energy consumption / cost refers to unit energy consumption, reagent cost, and pump / fan power, etc.

[0066] The process flow diagram is converted into a set of candidate directed edges. Only directed candidate edges are retained between variable pairs within the same process segment (such as grinding segment, flotation roughing segment, flotation cleaning segment, magnetic separation segment, thickening segment, and filtration segment) that have clear causal mechanisms (such as mass / energy conservation, fluid / mass transfer, chemical reaction or surface chemistry, and equipment structure / control loop logic, etc.). The rest are directly eliminated to avoid overfitting.

[0067] The process flow diagram comes from the controlled engineering archives within the plant, that is, the formal process drawing that has been approved by MOC (Change Management) and has a version number and effective date;

[0068] Based on the causal mechanism, each directed candidate edge is labeled with a positive or negative sign and global or segmental monotonicity. The labeling is used as a regularization prior for edge weight strength estimation to avoid edge weights that physically violate the rules.

[0069] Parent and child variables are determined based on the candidate directed edge set. The expected difference in the child variable caused by intervention at the upper or lower operation point of the parent variable is used as the effect size. The effect size is then normalized using the standard deviation of the child variable to obtain the dimensionless edge weight strength, expressed as:

[0070] ;

[0071] in, Indicates from parent variable Pointer to child variable The dimensionless edge weight strength, This represents the parent variable, which is the process quantity that is intervened or manipulated by external factors. This represents a sub-variable, which is either a result quantity or an intermediate process quantity. Represents the expectation operator. This indicates that external factors are interfering with the budget. This indicates the top operand of the parent variable. This indicates the lower operand of the parent variable. Indicates the variance of the sub-variables;

[0072] Based on the process flow diagram approved by MOC, variables within the same process segment are constructed into a directed acyclic structure according to the causal order of material / energy and control loop. Positive and negative signs and monotonicity labels are completed for each candidate directed edge, and the dimensionless edge weight is used as the strength measure. The source version and timestamp are written into the edge attributes to obtain the causal relationship between nodes.

[0073] In the causal knowledge graph, target nodes and constraint nodes are established respectively. Business parameters such as the quality baseline, the upper limit of reagent cost, and the upper limit of unit energy consumption are fixed in the attributes of the target node. The value range of liquid level, pH, valve position and discharge are specified in the constraint node and hard or soft constraints and priorities are marked. Variables are connected to the corresponding nodes through the associated edges to form optimization target constraints.

[0074] Hard constraints refer to those that must not be exceeded; once exceeded, they will be immediately removed (e.g., physical upper or lower limits of valve positions, upper limits of environmental emissions, and key safety interlock points).

[0075] Soft constraints allow for small deviations, but impose penalties, which do not lead to sample deletion but rather lower utility scores;

[0076] Directed edges, edge weights, target nodes, and constraint nodes are written to the graph database cluster. Each update is tagged with a version number and timestamp, and the source number of the process flow diagram and the version of the variable list are retained to generate a read-only view.

[0077] S2. Based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, data is selected from the historical database to form a high-value historical dataset.

[0078] Read directed edges and edge weights, target nodes, constraint nodes, and historical data corresponding to the variable list in the historical database from the read-only view. Using production data from the past twelve consecutive months, and only using the process segments publicly available in the variable list, calculate a uniform dimensionless utility score for each sample in the historical database—that is, an aligned record for a specific process segment at a specific point in time under a unified clock. The expression is as follows:

[0079] ;

[0080] in, Indicates sample dimensionless utility fraction, Indicates the sample index. Indicates from parent variable Pointer to child variable The directed edge, Denotes the set of candidate directed edges. Representing an edge The dimensionless edge weight strength, Indicates sample In parent variable The observed values ​​on Indicates sample In child variables The observed values ​​on Represents the parent variable The mean over the time range, Represents the parent variable Standard deviation over the time range Subvariable Standard deviation over the time range This represents the constraint penalty constant, used to adjust the reduction in score due to constraint violation. Indicates the number of soft constraint terms. Indicates a constraint index. This indicates a violation of the degree function, which is automatically generated from the metadata of the constraint node. Indicates sample The variable vector contains all fields of the corresponding process segment in the variable list;

[0081] Sort the samples by dimensionless utility scores from high to low, remove any samples that violate hard constraints, and extract non-adjacent samples at fixed time intervals to reduce temporal redundancy in the same state. For each process segment, retain the same proportion of high-utility score samples to form a candidate sample set.

[0082] Fixed time interval refers to the minimum sampling interval in the sample extraction stage to reduce time sequence redundancy, and is a value that is fixed in advance in the configuration table;

[0083] The number of times the candidate sample set covers the first few strong edges is counted. If a sample in the candidate sample set contains both the effective value of the parent variable and the observed value of the child variable at the corresponding time after lag alignment, and the sample has been filtered by hard constraints, it is recorded as covering the strong edge once.

[0084] Strong edges refer to the top few directed edges selected from the candidate directed edge set, sorted by the absolute value of the edge weight from largest to smallest.

[0085] The selected samples are compiled into a high-value historical dataset and then stored in a read-only view.

[0086] S3. Based on high-value historical datasets, construct a reward function model, train the policy function, and obtain a reinforcement learning control policy.

[0087] Based on high-value historical datasets and target and constraint nodes in causal knowledge graphs, the geometric mean of rewards and costs is calculated. The natural logarithm of these geometric mean values ​​is then taken to transform the multiplicative relationship into an additive metric, yielding logarithmic utility. Finally, action smoothing and operational robustness terms are removed from the logarithmic utility to obtain the immediate reward, expressed as:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] in, Indicates time The geometric mean normalized value of the returns, Indicates time The grade of the concentrate, This represents the time mean of concentrate grade on a high-value historical dataset. Indicates time Recovery rate, This represents the time mean of the recovery rate on high-value historical datasets. Indicates time The geometric mean normalized value of the cost. Indicates time The cost of the medicine, This represents the time mean of drug costs on high-value historical datasets. Indicates time energy consumption per unit This represents the time mean of energy consumption per unit on a high-value historical dataset. Indicates time logarithmic utility Indicates time The final instant reward, Represents the motion smoothing constant. Indicates time The manipulator vector, This represents the vector of control inputs from the previous time step. Represents the robustness penalty constant. Indicates time The process quantity vector, This represents the mean vector of the process quantity vector on a high-value historical dataset.

[0093] High-value historical datasets are divided into training and validation sets according to time sequence. Hard and soft constraints are read from constraint nodes and solidified into online inspection rules during training.

[0094] The execution behavior cloning hot start and PPO-based small-step fine-tuning training policy function are trained by taking state data snapshots aligned with fixed process delays in high-value historical datasets as inputs and the historical executed control instructions corresponding to the state data snapshot timestamps as supervision targets, so that the control policy output approximates the actionable actions. On the validation set, the action error quantile and the average reward are used as stopping conditions. When the stopping conditions are met, the control policy is frozen as the initial control policy.

[0095] Fixed process delay refers to the shortest time required for an adjustment of a parent variable to have an observable impact on the corresponding child variable or target quality indicator after the adjustment of the parent variable is transmitted in the material, reaction, or control chain.

[0096] Action error quantile refers to predicting actions one by one for the validation set state using the current strategy, mapping the predicted actions to the same physical dimension as the historical actions, calculating the Euclidean norm of the action error at each time step, that is, the Euclidean norm of the error between the predicted action and the historical action, and normalizing it with the physical range of each actuator, and taking the quantile value of the action error on the entire validation set.

[0097] The average reward refers to the calculation of the final instantaneous reward for each time point, and the arithmetic average of the final instantaneous rewards in the validation set.

[0098] The trajectory is constructed on the offline playback sample using the initial control strategy. The cumulative return is obtained by gradually accumulating the immediate reward at each time step by step according to the discount. The final immediate reward at the current time step and the subsequent income are discounted to the current time step to form the current equivalent value income. The current equivalent value income is compared with the value at the current time step to obtain the deviation. The deviation is aggregated by slowing down and smoothing over time to obtain the advantage estimate.

[0099] The value at the current moment refers to the expected discounted return that can be obtained from the current moment until the end of the trajectory;

[0100] Offline playback samples refer to a continuous set of records sorted by time and aligned with lag.

[0101] Constructing a trajectory refers to generating predicted actions for each state at each time step using the current control strategy, and forming a sequence of states, predicted actions, immediate rewards, and the next state together with the measured indicators at each time step, i.e., a trajectory.

[0102] Based on the return and advantage estimates, multiple rounds of small-batch updates are performed to obtain the reinforcement learning control policy. By pruning the target and setting the relative entropy for the control policy changes in each round, the update range is ensured to be controllable. After each round of updates is completed, a hard constraint trigger check is run. If any out-of-bounds is found, the update in this round is discarded and the control policy of the previous version is rolled back.

[0103] S4. Run the reinforcement learning control strategy, generate control commands, and send the control commands to the target device to obtain dynamic optimization control;

[0104] Load the reinforcement learning control strategy version and configuration, set a fixed control cycle, and in each control cycle, read the dynamic running data according to the DCS / PLC and online analysis, align it, and obtain a state data snapshot;

[0105] The state data snapshot is input into the reinforcement learning control strategy to obtain the original actions (such as ore feed rate, reagent dosage, aeration rate, and valve or liquid level setting) corresponding to each item in the variable list.

[0106] A state data snapshot refers to an aligned snapshot of data sent to the reinforcement learning control policy in each control cycle, with fields derived from the variable list;

[0107] The system executes a safety envelope, trimming each original action to within the physical upper and lower limits of the equipment to obtain trimmed action values. It then applies a variation limit to the trimmed action values ​​based on adjacent control cycles to prevent abrupt changes and performs group constraint checks (e.g., total airflow and total chemical dosage not exceeding the upper limit). This yields a draft control command to be issued. The draft control command is then checked against soft and hard constraints using soft and hard constraint rules. If it passes both checks, the draft control command is confirmed as a control command. If it fails the soft constraint check, it is marked as a soft constraint violation, but the control command issuance is not blocked. If it fails the hard constraint check, the hard constraint takes precedence. If it fails the hard constraint check, no control command is generated, and it is marked as an abnormal event. The system then rolls back to the previously confirmed control command and sends the control command to the target equipment via the industrial control interface. The system reads and verifies the equipment feedback. If the command feedback deviation exceeds the allowable tolerance, the system immediately rolls back to the previously confirmed control command.

[0108] Command feedback deviation refers to the difference between the equipment feedback value and the confirmed control command in the current control cycle after the clock is unified and aligned according to a fixed process delay.

[0109] The allowable tolerance refers to the conversion of the comprehensive accuracy, hysteresis, and resolution given in the instrument and actuator specifications into an equipment-level baseline tolerance, the conversion of feedback fluctuations statistically analyzed using high-value historical data during stable periods under historical stable operating conditions into a data-level baseline tolerance, and the conversion of mechanical dead zone or minimum fine-tunable step size into a mechanism-level baseline tolerance. Finally, the maximum baseline tolerance among the equipment-level baseline tolerance, data-level baseline tolerance, and mechanism-level baseline tolerance is taken as the allowable tolerance.

[0110] Monitor any hard constraint triggers and abnormal events, package the triggering event along with the current state and control instructions into an abnormal buffer entry, and write the remaining normal control cycle closed-loop records into the strategy replay library according to the control cycle.

[0111] S5. Real-time monitoring of dynamic optimization control; when an abnormal deviation in the operating status is detected, counterfactual reasoning is performed based on the causal knowledge graph to obtain the counterfactual reasoning results.

[0112] Take the latest control cycle closed-loop record from the abnormal buffer entries and lock it as the current working sample. Monitor and verify the current working sample. If any hard constraint is triggered, the final reward is continuously negative within a fixed window, or the deviation between the equipment feedback and the instruction feedback of the issued control command exceeds the allowable tolerance, it is judged as an abnormality and the causal knowledge graph positioning is performed.

[0113] Using the target quality index node corresponding to the current working sample as the key, select the candidate manipulation quantity from the direct parent node of the target quality index node, and select the first one according to the joint ranking of edge weight strength and current deviation magnitude to determine the unique intervention quantity.

[0114] The target quality indicator node refers to reading the process segment and the corresponding output flow from the current working sample. In the causal knowledge graph, the set of quality KPI nodes connected to the output flow is found along the association edge. The KPI node with the highest priority is selected. If there are ties, the KPI node with a higher association strength with the output flow of this process segment is selected.

[0115] Counterfactual reasoning is based on causal knowledge graphs. Starting from the current value of the only intervention quantity, it checks in both directions in a fixed step size to predict the target quality index at the sample timestamp and fixed process delay. It checks whether all constraint nodes are satisfied and compares them with the median value of the fixed window forward from the current working sample timestamp to see if it has returned to normal or improved. It stops when the minimum satisfaction step size is found.

[0116] Generate intervention recommendations (including unique intervention amount, adjustment range, expected improvement, expected effective time, and process quantities that need to be closely monitored, etc.) and send them back to the execution safety envelope. If blocked, mark it as not executed and record the reason.

[0117] Within the expected effective time window, read the actual quality indicators, generate a prediction-actual comparison, and bind the unique intervention quantity, whether it is executed, the adjustment range, the prediction and actual measurement, the causal knowledge graph version and edge weight snapshot, the status and control instruction context, and the constraint check results to the unique ID of the current working sample and archive them into an archive package.

[0118] S6. Based on the counterfactual reasoning results, the edge weights in the causal knowledge graph are corrected, and the reinforcement learning control strategy is updated through incremental reinforcement learning.

[0119] Using the causal knowledge graph version and reinforcement learning control policy version recorded in the archive as anchors, operations are performed on the snapshot, and all changes generate a new version number and are bound to the current evidence;

[0120] This evidence refers to the full set of contextual data that can prove the rationality of this edge weight or reinforcement learning control strategy update;

[0121] When revising edge weights, only the edges involved in the current evidence are updated, that is, the edges that point to the target quality index by the only intervention quantity identified in this case. The local sensitivity is obtained by comparing the predicted and measured values ​​with the actual adjustment range. The old edge weights and the local sensitivity are weighted and synthesized according to the fixed process delay to obtain the new edge weights. If the direction of the new edge weights conflicts with the monotonicity of the label, the update is truncated to the fixed minimum range and written into the new causal knowledge graph version.

[0122] The status, control commands, device feedback, final immediate rewards and constraint checks in the archive package are merged and deduplicated, obvious anomalies are removed, and then aligned to form incremental data slices.

[0123] In offline playback or simulation environments, incremental data slices are used to make very small parameter updates to the reinforcement learning control policy, and candidate policies detected offline through relative change upper limits and hard constraints are entered into grayscale verification.

[0124] The upper limit of relative change refers to the limit that the change in the output distribution of the reinforcement learning control policy relative to the old reinforcement learning control policy after each round of update shall not exceed a fixed upper limit;

[0125] Offline hard constraint detection refers to checking each hard constraint for triggering. If a hard constraint is triggered, the update is rolled back and the upper limit of relative change is reduced. The update is rejected.

[0126] On the mirrored data stream, the old reinforcement learning control strategy and the updated reinforcement learning control strategy are compared in parallel for several control cycles. The moving average of the final immediate reward and the number of hard and soft constraints are calculated. The strategy can be released if the final immediate reward does not decrease, the hard constraints are triggered zero times, and the updated reinforcement learning control strategy does not trigger the soft constraints more frequently than the old reinforcement learning control strategy. Otherwise, it is rolled back and the reason is recorded. After release, the updated reinforcement learning control strategy is marked as the latest reinforcement learning control strategy, and the old reinforcement learning control strategy version is downgraded to a rollbackable version, which will automatically take over the latest reinforcement learning control strategy in the next control cycle.

[0127] This embodiment also provides a knowledge graph-based intelligent optimization control system for mineral processing, including:

[0128] The knowledge graph construction module collects static attribute data and dynamic runtime data to construct a causal knowledge graph, thereby obtaining the causal relationships between nodes in the causal knowledge graph and optimization objective constraints.

[0129] The data filtering module, based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, filters data from the historical database to form a high-value historical dataset.

[0130] The control policy module constructs a reward function model based on a high-value historical dataset, trains the policy function, and obtains a reinforcement learning control policy.

[0131] The control command module runs a reinforcement learning control strategy, generates control commands, and sends the control commands to the target device to obtain dynamic optimized control.

[0132] The counterfactual reasoning module monitors the dynamic optimization control in real time. When an abnormal deviation in the operating status is detected, it performs counterfactual reasoning based on the causal knowledge graph and obtains the counterfactual reasoning results.

[0133] The control strategy update module modifies the edge weights in the causal knowledge graph based on counterfactual reasoning results and updates the reinforcement learning control strategy through incremental reinforcement learning.

[0134] This embodiment also provides a computer device applicable to the intelligent optimization control method for mineral processing based on knowledge graphs, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent optimization control method for mineral processing based on knowledge graphs as proposed in the above embodiment.

[0135] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0136] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the intelligent optimization control method for mineral processing based on knowledge graphs as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0137] In summary, this invention achieves a unified causal expression of knowledge and manipulative variables by constructing a causal knowledge graph and establishing optimization objective constraints, modifying the edge weights in the causal knowledge graph and updating the reinforcement learning control strategy through incremental reinforcement learning, realizing a closed-loop self-evolution between the knowledge layer and the strategy layer, and using it to resolve operational deviations online. Ultimately, it achieves the beneficial effects of alleviating the problems of knowledge-policy fragmentation and insufficient consistency in anomaly handling, while improving indicators and security compliance.

[0138] 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 knowledge graph-based intelligent optimization control method for mineral processing, characterized in that: include, Collect static attribute data and dynamic runtime data, construct a causal knowledge graph, and obtain the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints. Based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, data is selected from the historical database to form a high-value historical dataset. Based on high-value historical datasets, a reward function model is constructed, a policy function is trained, and a reinforcement learning control policy is obtained. The reinforcement learning control strategy is run to generate control commands and send them to the target device to obtain dynamic optimization control. Real-time monitoring of dynamic optimization control is performed. When an abnormal deviation in the operating status is detected, counterfactual reasoning is performed based on the causal knowledge graph to obtain the counterfactual reasoning results. Based on the counterfactual reasoning results, the edge weights in the causal knowledge graph are modified, and the reinforcement learning control strategy is updated through incremental reinforcement learning. The training policy function is used to obtain a reinforcement learning control policy. The specific steps are as follows: the high-value historical dataset is divided into a training set and a validation set according to time sequence; the final immediate reward output by the reward function model is used as the reward signal for supervised early stopping and PPO of the behavior clone; and the control policy is obtained by training the policy function through behavior clone hot start and small-step fine-tuning based on PPO. Perform mini-batch updates on the control policy to obtain a reinforcement learning control policy; The dynamic optimization control is monitored in real time. When an abnormal deviation in the operating state is detected, counterfactual reasoning is performed based on the causal knowledge graph to obtain the counterfactual reasoning result. The specific steps are as follows: take the latest control cycle closed-loop record from the abnormal buffer entry and lock it as the current working sample; monitor and verify the current working sample; and perform causal knowledge graph positioning. Based on the current working sample, a unique intervention quantity is identified. Counterfactual reasoning is performed based on the causal knowledge graph. Hypothetical interventions are made on the unique intervention quantity to obtain the counterfactual reasoning results, which are then bound to the unique ID of the current working sample to generate an archive package. The specific steps for correcting the edge weights in the causal knowledge graph based on counterfactual reasoning results are as follows: When correcting the edge weights, only the edge pointing to the target quality indicator of the unique intervention quantity identified this time is updated. By comparing the predictions with the actual measurements and adjusting the actual range, the local sensitivity is obtained. Aligned with a fixed process delay, the old edge weights and the local sensitivity are weighted and synthesized to obtain the new edge weights. The specific steps for updating the reinforcement learning control strategy through incremental reinforcement learning are as follows: extracting archived data fields for deduplication, cleaning and alignment, generating incremental data slices, and updating the parameters of the reinforcement learning control strategy. After the updated candidate policies are tested offline by relative change upper limit and hard constraints, they enter grayscale verification. The old reinforcement learning control policy and the updated reinforcement learning control policy are compared in parallel on the mirror data stream. If the hard conditions are met, they can be released; otherwise, they are rolled back and the reason is recorded. After release, the updated reinforcement learning control policy is marked as the latest reinforcement learning control policy.

2. The intelligent optimization control method for mineral processing based on knowledge graphs as described in claim 1, characterized in that: The process of collecting static attribute data and dynamic operational data to construct a causal knowledge graph, and obtaining the causal relationships between nodes in the causal knowledge graph and optimization objective constraints, involves the following specific steps: Static attribute data is collected through mineral samples and laboratory tests, while dynamic operational data is collected through DCS / PLC and online analysis. Construct a causal knowledge graph in a graph database cluster, freeze the variable list and solidify it into a variable dictionary table; The process flow diagram is converted into a set of candidate directed edges. Each candidate directed edge is labeled with a positive or negative sign and monotonicity. The dimensionless edge weight strength is calculated. Based on the positive or negative sign, monotonicity, and dimensionless edge weight strength of the candidate directed edges, the causal relationship between nodes is obtained. In a causal knowledge graph, target nodes and constraint nodes are established. Variables are connected to target nodes or constraint nodes through association edges to obtain optimization target constraints.

3. The intelligent optimization control method for mineral processing based on knowledge graphs as described in claim 2, characterized in that: The process involves selecting data from historical databases to create a high-value historical dataset based on the causal relationships between nodes in a causal knowledge graph and optimization objective constraints. The specific steps are as follows: Based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, a uniform dimensionless utility score is calculated for each sample in the historical database. Sort the samples by utility score from high to low, remove any samples that violate hard constraints, retain the same proportion of high-utility-score samples for each process segment to form a candidate sample set, and combine the selected samples to form a high-value historical dataset.

4. The intelligent optimization control method for mineral processing based on knowledge graphs as described in claim 3, characterized in that: The specific steps for constructing a reward function model based on high-value historical datasets are as follows: Based on high-value historical datasets and target nodes and constraint nodes in causal knowledge graphs, calculate the geometric mean of gains and the geometric mean of costs. The logarithmic utility is obtained by taking the natural logarithm of the ratio of the geometric mean of the benefit to the geometric mean of the cost, and a reward function model is constructed to calculate the final immediate reward.

5. The intelligent optimization control method for mineral processing based on knowledge graphs as described in claim 4, characterized in that: The process involves implementing a reinforcement learning control strategy, generating control commands, and sending these commands to the target device to achieve dynamic optimized control. The specific steps are as follows: Load the reinforcement learning control policy version and configuration, set a fixed control period, and in each control period, input the aligned data snapshot into the reinforcement learning control policy to obtain the original actions item by item in the variable list; The system executes a safety envelope on the original action, generates control commands, and sends the control commands to the target device for dynamic optimization control through the industrial control interface. It also monitors any hard constraint triggers and abnormal events, and packages the trigger events and control commands into an exception buffer entry.

6. A knowledge graph-based intelligent optimization control system for mineral processing, based on the knowledge graph-based intelligent optimization control method for mineral processing as described in any one of claims 1 to 5, characterized in that: include, The knowledge graph construction module collects static attribute data and dynamic runtime data to construct a causal knowledge graph, thereby obtaining the causal relationships between nodes in the causal knowledge graph and optimization objective constraints. The data filtering module, based on the causal relationships between nodes in the causal knowledge graph and the optimization objective constraints, filters data from the historical database to form a high-value historical dataset. The control policy module constructs a reward function model based on a high-value historical dataset, trains the policy function, and obtains a reinforcement learning control policy. The control command module runs a reinforcement learning control strategy, generates control commands, and sends the control commands to the target device to obtain dynamic optimized control. The counterfactual reasoning module monitors the dynamic optimization control in real time. When an abnormal deviation in the operating status is detected, it performs counterfactual reasoning based on the causal knowledge graph and obtains the counterfactual reasoning results. The control strategy update module modifies the edge weights in the causal knowledge graph based on counterfactual reasoning results and updates the reinforcement learning control strategy through incremental reinforcement learning.

Images