A knowledge graph-based mineral processing optimization control method and system

Abstract

This invention discloses a mineral processing optimization control method and system based on knowledge graphs, relating to the field of mineral processing control technology. The method includes: collecting and preprocessing data from the entire mineral processing process to generate an initial multimodal dataset; extracting spatiotemporal process elements from the initial multimodal dataset to construct an initial mineral processing optimization knowledge graph; establishing a mineral processing process topology graph, incorporating the process, equipment, and material topologies into the mineral processing optimization knowledge graph to construct an evolutionary mineral processing optimization knowledge graph; aligning the initial multimodal dataset with timestamps using the evolutionary mineral processing optimization knowledge graph to obtain a time-aligned multimodal dataset; performing hierarchical temporal modeling and cross-modal association calculations on the time-aligned multimodal dataset to generate multimodal temporal association features; and inputting the multimodal temporal association features into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy.

Description

Technical Field

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

[0002] Mineral processing is a crucial process for the efficient utilization of mineral resources, typically including crushing, screening, grinding, flotation, dewatering, and tailings treatment. During production, multi-source data is generated, including ore raw material properties, equipment operation, reagent usage, and energy and environmental parameters. Existing research often utilizes sensor detection, spectral analysis, and industrial camera acquisition to obtain information such as ore particle size distribution, ore chemical composition ratios, reagent dosage, tailings concentration, and power consumption. These data are then combined with time series modeling, spatial layout modeling, and semantic parsing methods to achieve a unified representation of the multi-source data. These methods provide a foundation for process modeling and optimization, enabling researchers to describe and analyze the operational characteristics of the entire mineral processing process.

[0003] However, conventional methods still have limitations under complex process conditions. On the one hand, cross-process dynamic modeling mainly relies on time series methods, but when material transport delays and energy coupling exist, it is often difficult to accurately reveal the dynamic dependencies across stages, resulting in limited prediction accuracy for grinding cycle load rates or flotation reagent effects. On the other hand, cross-modal data fusion methods typically rely on spatial coordinate mapping and single synchronization rules, but when it is necessary to align ore images, spectral features, and numerical sequences under a unified time index, there is a lack of effective correlation mechanisms, thus limiting the in-depth characterization of multimodal dynamic evolution relationships. Summary of the Invention

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

[0005] Therefore, this invention provides a mineral processing optimization control method based on knowledge graphs, which solves the problems of difficulty in accurately modeling dynamic dependencies across processes and difficulty in effectively aligning cross-modal data.

[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 a mineral processing optimization control method based on knowledge graphs, comprising,

[0008] Collect data from the entire mineral processing process and preprocess it to generate an initial multimodal dataset;

[0009] Spatiotemporal process elements were extracted from the initial multimodal dataset to construct an initial mineral processing optimization knowledge graph;

[0010] Establish a mineral processing topology diagram, incorporate the process, equipment and material topology into the mineral processing optimization knowledge graph, and construct an evolutionary mineral processing optimization knowledge graph.

[0011] The initial multimodal dataset was timestamped using an evolutionary mineral processing optimization knowledge graph to obtain a time-axis aligned multimodal dataset.

[0012] Hierarchical temporal modeling and cross-modal correlation calculation are performed on time-axis aligned multimodal datasets to generate multimodal temporal correlation features;

[0013] Multimodal temporal correlation features are input into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy;

[0014] The optimal mineral processing control strategy is transformed into process control commands for execution. Execution feedback is collected and the evolutionary mineral processing optimization knowledge graph is updated online to generate the optimal mineral processing control strategy for the next cycle.

[0015] As a preferred embodiment of the knowledge graph-based mineral processing optimization control method of the present invention, the mineral processing data includes ore raw material attribute data, crushing and screening process data, grinding process data, flotation reagent process data, tailings treatment and discharge process data, and energy consumption and environmental parameter data.

[0016] As a preferred embodiment of the knowledge graph-based mineral processing optimization control method of the present invention, the preprocessing to generate an initial multimodal dataset includes unifying the time scale and normalizing the spatial coordinates of the entire mineral processing data to obtain a numerical sequence.

[0017] Perform acquisition error correction and index annotation on the ore particle morphology images and micro fracture images in the ore raw material attribute data to obtain image samples;

[0018] Noise smoothing, missing measurement completion, and wavelength range normalization were performed on the fluorescence spectrum and chemical composition spectrum of the pulp reagents in the flotation reagent process data to obtain spectral characteristics.

[0019] Numerical sequences, image samples, and spectral features are combined to form an initial multimodal dataset.

[0020] As a preferred embodiment of the knowledge graph-based mineral processing optimization control method of the present invention, wherein:

[0021] The construction of the initial mineral processing optimization knowledge graph includes obtaining time-dimensional process elements by performing time series analysis on the initial multimodal dataset;

[0022] The location information of the collection points in the initial multimodal dataset is mapped using a unified spatial index to obtain spatial dimension process elements;

[0023] Semantic parsing is performed on the initial multimodal dataset to obtain semantic dimension process elements;

[0024] By combining time-dimensional, spatial-dimensional, and semantic-dimensional process elements, a set of process entities is established.

[0025] Within the set of process entities, time evolution relationships are generated based on time-dimensional process elements, spatial connection relationships are generated based on spatial-dimensional process elements, and semantic dependency relationships are generated based on semantic-dimensional process elements. The time evolution relationships, spatial connection relationships, and semantic dependency relationships are recorded as multidimensional relationships of process entities.

[0026] By using the set of process entities as nodes and the multidimensional relationships between process entities as edges, a structured organization is constructed to create an initial mineral processing optimization knowledge graph.

[0027] As a preferred embodiment of the mineral processing optimization control method based on knowledge graph described in this invention, the construction of the evolutionary mineral processing optimization knowledge graph includes obtaining process nodes, equipment nodes, and material nodes based on the full-process data of mineral processing and the production process layout diagram, and generating a set of nodes for the mineral processing process topology diagram;

[0028] Based on the process sequence, material channel connection, and energy consumption correspondence, sequential constraint edges, material transport edges, and energy consumption correlation edges are established to generate the edge set of the mineral processing process topology graph.

[0029] Label and verify the node set and edge set to generate a list of data structures for the mineral processing topology graph;

[0030] The process nodes, equipment nodes, and material nodes in the data structure list are used as entity nodes, and the sequence constraint edges, logistics transmission edges, and energy consumption association edges are used as relation edges. These are then merged and updated with the initial mineral processing optimization knowledge graph to construct an evolutionary mineral processing optimization knowledge graph.

[0031] As a preferred embodiment of the knowledge graph-based mineral processing optimization control method of the present invention, the time-axis aligned multimodal dataset includes the generation of a time-axis aligned multimodal dataset by aligning the numerical sequences, image samples, and spectral features of the initial multimodal dataset with timestamps under the constraints of sequential constraint edges and logistics transmission edges through an evolutionary mineral processing optimization knowledge graph.

[0032] As a preferred embodiment of the knowledge graph-based mineral processing optimization control method of the present invention, the generation of multimodal time-series correlation features includes performing short-term and long-term modeling on numerical sequences to obtain short-term time-series latent vectors and long-term time-series latent vectors.

[0033] Simultaneously, image-type samples and spectral features are converted into image feature vectors and spectral feature vectors, respectively;

[0034] Cross-modal correlation calculations are performed on short-term temporal latent vectors, long-term temporal latent vectors, image feature vectors, and spectral feature vectors using the delay correlation function, and the maximum value of the delay correlation function is taken as the multimodal temporal correlation feature.

[0035] As a preferred embodiment of the mineral processing optimization control method based on knowledge graph described in this invention, the optimal mineral processing control strategy includes matching multimodal temporal correlation features with process nodes, equipment nodes, and material nodes in the evolutionary mineral processing optimization knowledge graph to obtain grinding particle size control index, flotation reagent dosage control index, energy consumption efficiency control index, and tailings discharge stability control index.

[0036] Each control index is transformed into an optimization objective function, and a multi-objective optimization algorithm is used to solve it under the constraint set composed of sequential constraint edges, logistics transmission edges, and energy consumption correlation edges, so as to obtain the Pareto optimal solution set, which serves as the optimal control strategy for mineral processing.

[0037] As a preferred embodiment of the knowledge graph-based mineral processing optimization control method of the present invention, the step of converting the optimal mineral processing control strategy into process control instructions includes converting the optimal mineral processing control strategy into process control instructions through an industrial communication protocol and sending them to the process equipment for execution.

[0038] During the execution of process control commands, process execution feedback data such as equipment power load, reagent flow stability, concentrate grade fluctuation, and tailings discharge concentration are collected.

[0039] The process execution feedback data is input into the evolutionary beneficiation optimization knowledge graph to update the attribute values ​​of process nodes, equipment nodes, and material nodes. The process execution feedback data is then used as new optimization input to generate the optimal beneficiation control strategy for the next cycle.

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

[0041] The data acquisition module is used to collect data from the entire mineral processing process and preprocess it to generate an initial multimodal dataset.

[0042] The element extraction module is used to extract spatiotemporal process elements based on the initial multimodal dataset and construct an initial mineral processing optimization knowledge graph.

[0043] The topology construction module is used to build a mineral processing topology graph, incorporate the process equipment and material topology into the mineral processing optimization knowledge graph, and construct an evolutionary mineral processing optimization knowledge graph.

[0044] The time alignment module is used to align the timestamps of the initial multimodal dataset using the evolutionary mineral processing optimization knowledge graph, thereby obtaining a time-axis aligned multimodal dataset.

[0045] The temporal modeling module is used to perform hierarchical temporal modeling and cross-modal correlation calculation on time-axis aligned multimodal datasets, generating multimodal temporal correlation features;

[0046] The strategy reasoning module is used to input multimodal temporal correlation features into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy;

[0047] The instruction execution module is used to convert the optimal mineral processing control strategy into process control instructions for execution, collect execution feedback, update the evolutionary mineral processing optimization knowledge graph online, and generate the optimal mineral processing control strategy for the next cycle.

[0048] The beneficial effects of this invention are as follows: by aligning the initial multimodal dataset with timestamps using an evolutionary mineral processing optimization knowledge graph, a high-precision alignment effect of multimodal process information on the time axis is achieved; by performing hierarchical time series modeling and cross-modal correlation calculation on the time-axis aligned multimodal dataset, short-term fluctuation features and long-term dependency features of numerical sequences are extracted, and combined with image feature vectors and spectral feature vectors, cross-modal time series dependencies are calculated using a delay correlation function to obtain multimodal time series correlation features, ensuring a unified representation of process dynamics and material characteristics. Attached Figure Description

[0049] 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.

[0050] Figure 1 This is a flowchart of a knowledge graph-based mineral processing optimization control method.

[0051] Figure 2 This is a schematic diagram of a mineral processing optimization control system based on a knowledge graph.

[0052] Figure 3 A flowchart for hierarchical time series modeling and cross-modal correlation calculation.

[0053] Figure 4 A flowchart for reasoning the optimal control strategy for mineral processing. Detailed Implementation

[0054] 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.

[0055] 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.

[0056] 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.

[0057] Reference Figures 1-4 This is one embodiment of the present invention, which provides a mineral processing optimization control method based on knowledge graphs, comprising the following steps:

[0058] S1. Collect data from the entire mineral processing process and preprocess it to generate an initial multimodal dataset.

[0059] Furthermore, the data for the entire mineral processing process includes ore raw material attribute data, crushing and screening process data, grinding process data, flotation reagent process data, tailings treatment and discharge process data, and energy consumption and environmental parameter data.

[0060] The data includes, but is not limited to, ore raw material attribute data, ore particle size distribution, ore chemical composition ratio, ore hardness value, and moisture content; crushing and screening process data, but is not limited to, crusher feed particle size, crusher discharge particle size, screening efficiency, and equipment operating power; grinding process data, but is not limited to, mill feed particle size, mill discharge particle size, unit energy consumption, and circulating load rate; flotation reagent process data, but is not limited to, reagent type, reagent dosage, flotation time, froth layer thickness, and slurry concentration; tailings treatment and discharge process data, but is not limited to, tailings concentration, tailings particle size, and tailings water recovery rate; and energy consumption and environmental parameter data, but is not limited to, electricity consumption, equipment power load, and ambient temperature.

[0061] The data for the entire mineral processing process includes, but is not limited to, real-time acquisition using laser particle size analyzers, X-ray fluorescence spectrometers, industrial cameras, online moisture sensors, pipeline flow sensors, power metering devices, and temperature sensors.

[0062] Furthermore, preprocessing includes unifying the time scale, standardizing the spatial coordinates, correcting acquisition errors, and organizing the format and structure of the data from the entire mineral processing process.

[0063] In the time scale unification, the data on ore raw material properties, crushing and screening process data, grinding process data, flotation reagent process data, tailings treatment and discharge process data, and energy consumption and environmental parameters are processed in a unified time scale. The production line sampling control clock (e.g., PLC sampling clock) is used as a unified time reference. All kinds of collected records are aligned to a unified time index and formed standardized numerical fields arranged in chronological order.

[0064] In spatial coordinate normalization, the production process layout diagram (such as engineering drawings showing the physical layout and logical relationship of the process flow of a mineral processing plant) is used as a reference. The collection points of crushing, grinding, flotation, tailings and energy consumption are located in a unified spatial coordinate system on the process layout diagram and bound to the time index to obtain a numerical sequence with spatiotemporal anchoring characteristics.

[0065] In the acquisition error correction, numerical deviations caused by acquisition fluctuations in ore raw material attribute data, flotation reagent process data, and energy consumption and environmental parameter data are smoothed and corrected. At the same time, missing measurements caused by acquisition interruption or sensor malfunction in crushing and screening process data, grinding process data, and tailings treatment and discharge process data are interpolated and completed, thereby correcting acquisition errors and restoring the continuity and rationality of numerical records. Furthermore, acquisition error correction and unified index labeling are performed on ore particle morphology images and microcrack images in ore raw material attribute data, and the acquisition time and spatial location are merged into a unified coordinate system to obtain image samples corresponding to numerical sequences. Noise smoothing, missing measurement value completion, and wavelength range normalization are performed on the fluorescence spectrum and chemical composition spectrum of slurry reagent action acquired in flotation reagent process data. After unifying the time index and spatial coordinates, spectral features corresponding to process links are obtained.

[0066] In the process of formatting and structuring, the data of the entire mineral processing process are processed to unify field naming, standardize measurement units, align time indexes and normalize spatial indexes, thereby forming an initial multimodal dataset that is time-corresponding, spatially normalized, error-corrected and consistent in field structure.

[0067] The initial multimodal dataset includes numerical sequences, image samples, and spectral features. The numerical sequences include numerical fields from all mineral processing data. The image samples include ore particle morphology images and microcrack images collected from ore raw material attribute data. The spectral features include fluorescence spectra and chemical composition spectra of slurry reagents collected from flotation reagent process data.

[0068] S2. Extract spatiotemporal process elements based on the initial multimodal dataset and construct an initial mineral processing optimization knowledge graph.

[0069] Furthermore, spatiotemporal technological elements include time-dimensional technological elements, spatial-dimensional technological elements, and semantic-dimensional technological elements.

[0070] By conducting time series analysis on ore raw material attribute data, crushing and screening process data, grinding process data, flotation reagent process data, tailings treatment and discharge process data, and energy consumption and environmental parameter data, time-dimensional process elements are obtained, including ore particle size distribution variation trend, crushing and screening efficiency fluctuation pattern, grinding cycle load rate dynamic change, flotation reagent action time series, tailings concentration change curve, and energy consumption parameter fluctuation characteristics.

[0071] The spatial dimension process elements are obtained by extracting relationships based on the location information of the collection points under a unified spatial index. These relationships include the spatial connection between the crushing and screening process data collection points and the grinding process data collection points, the material transfer relationship between the grinding process data collection points and the flotation reagent process data collection points, the flow direction relationship between the flotation reagent process data collection points and the tailings discharge process data collection points, and the spatial proximity relationship between the energy consumption and environmental parameter collection points and each process data collection point.

[0072] Semantic dimension process elements are obtained by performing semantic parsing on ore raw material attribute data, flotation reagent process data, and tailings treatment and discharge process data. Semantic parsing refers to the logical analysis and summarization of key process data fields in the entire mineral processing process, such as the proportion of ore chemical composition, reagent type and dosage, the combination relationship between ore properties and reagent effects, and tailings concentration and water recovery rate, to transform the original numerical fields or recorded results into tags with clear process semantics, including ore type tags, reagent effect tags, ore reaction trend tags, and tailings residual property tags; among which, the ore type tag is rooted in... The ore types are classified and labeled according to the proportion of chemical components in the ore raw material attribute data; the reagent action labels are labeled according to the reagent types and dosages in the flotation reagent process data, indicating the functional mechanisms of the reagents, including collection, foaming, or adjustment effects; the ore reaction trend labels are labeled according to the combination relationship between the proportion of chemical components in the ore and the types of reagents, indicating the possible reaction trends of minerals and reagents, such as sulfidation or oxidation trends; and the tailings residual properties labels are labeled according to the tailings concentration, tailings particle size, and tailings water recovery rate in the tailings treatment and discharge process data, indicating the potential residual components in the tailings and their environmental impact properties.

[0073] Furthermore, a set of process entities is established based on spatiotemporal process elements, including ore raw material entities, crushing and screening process entities, grinding process entities, flotation reagent process entities, tailings discharge process entities, and energy consumption and environmental entities. Each process entity is derived from the extraction results of process elements in the time dimension, spatial dimension, and semantic dimension, respectively. For example, the ore raw material entity corresponds to the ore particle size distribution change trend, ore type label, and ore reaction trend label; the crushing and screening process entity corresponds to the crushing and screening efficiency fluctuation pattern and the spatial connection relationship between it and the grinding process data collection point; the grinding process entity corresponds to the dynamic change of grinding cycle load rate and the material transfer relationship between it and the flotation reagent process entity. By clarifying the composition of process entities, the initial beneficiation optimization knowledge graph has full-chain process coverage.

[0074] Based on the set of process entities, multidimensional relationships are established between them, including temporal evolution relationships, spatial connectivity relationships, and semantic dependencies. Temporal evolution relationships are established through time-series correlation analysis of the numerical sequences corresponding to the process entities. When the change of one process indicator has a significant or lagged correlation with another process indicator over time, a temporal evolution relationship is established between the corresponding process entities. For example, under a unified time retrieval, the correlation between the trend of ore particle size distribution change and the fluctuation pattern of crushing and screening efficiency is calculated, and a temporal evolution relationship is established. Spatial connectivity relationships are established through the production process layout diagram and a unified spatial index. When two process data collection points are located in material channels, pipelines, or physically adjacent positions in the process layout diagram, a spatial connectivity relationship is established between the corresponding process entities. Connection relationships are established, for example, spatial connections are established between data collection points in the grinding process and data collection points in the flotation reagent process; semantic dependencies are established through semantic parsing results. When there is logical adaptation or coupling between classification labels, semantic dependencies are established between corresponding process entities. For example, when the ore type label is sulfide ore and the reagent action label is collector, a semantic dependency is established between the ore raw material entity and the reagent entity. When there is an environmental impact correspondence between the tailings residual property label and the energy consumption and environment entity's record results, a semantic dependency is established between the tailings discharge process entity and the energy consumption and environment entity. By establishing multidimensional relationships, it is ensured that the initial mineral processing optimization knowledge graph not only contains single data fields, but also fully presents the logical connections and coupling mechanisms between process links.

[0075] The process entity set and multidimensional relationships are structured and organized to form an initial mineral processing optimization knowledge graph. Specifically, ore raw material entities, crushing and screening process entities, grinding process entities, flotation reagent process entities, tailings discharge process entities, and energy consumption and environmental entities are used as nodes, and temporal evolution relationships, spatial connection relationships, and semantic dependencies are used as edges. A directed graph structure is constructed for the initial mineral processing optimization knowledge graph using a unified field naming, unified time index, and unified spatial index. The constructed initial mineral processing optimization knowledge graph realizes the node-based expression of process entities and the change representation of process multidimensional relationships in terms of structure, ensuring that the initial mineral processing optimization knowledge graph has the data organization capability to cover the entire mineral processing process and the ability to represent process multidimensional relationships.

[0076] S3. Establish a mineral processing topology diagram, incorporate the process equipment and material topology into the mineral processing optimization knowledge graph, and construct an evolutionary mineral processing optimization knowledge graph.

[0077] Furthermore, based on the entire mineral processing data and the production process layout diagram, a set of nodes for the mineral processing process topology diagram is constructed, including process nodes, equipment nodes, and material nodes.

[0078] The process nodes are obtained based on the crushing process, grinding process, flotation process, dewatering process and tailings treatment process as divided in the production process layout diagram.

[0079] Equipment nodes are obtained based on the physical equipment corresponding to the collection points of each process in the entire mineral processing data. For example, in the crushing process node, the collection point corresponds to the crusher and the screening machine; in the grinding process node, the collection point corresponds to the ball mill and the hydrocyclone; in the flotation process node, the collection point corresponds to the flotation machine and the reagent dosing device; in the dewatering process node, the collection point corresponds to the dewatering screen and the thickener; and in the tailings treatment process node, the collection point corresponds to the tailings conveying pump and the tailings dam facility.

[0080] Material nodes are obtained based on the characteristics of ore raw materials and products in the entire mineral processing data. For example, the ore type and the proportion of ore chemical composition contained in the ore raw material attribute data are used as ore raw material nodes; the slurry particle size distribution and slurry flow rate formed in the crushing and screening process data and the grinding process data are used as slurry intermediate product nodes; the concentrate output after the action of flotation reagents in the process data is used as concentrate nodes; and the tailings concentration, tailings particle size and tailings water recovery rate in the tailings treatment and discharge process data are used as tailings nodes.

[0081] Construct the edge set of the mineral processing topology graph, including sequence constraint edges, material transport edges, and energy consumption related edges.

[0082] The sequential constraint edges are established based on the process sequence in the production process layout diagram. They are connected in the order of crushing process nodes, grinding process nodes, flotation process nodes, dewatering process nodes, and tailings treatment process nodes, and are checked one by one against the process flow relationship on the production process layout diagram.

[0083] The logistics transmission edge is established based on the material channels and pipeline connections in the production process layout diagram. Simultaneously, a time sequence consistency check is performed under a unified time index. If the change in crusher discharge particle size occurs first, followed by the change in mill feed particle size, and the crusher's operating power record and the ball mill's operating power record exist continuously in the same time period, the logistics transmission edge from the crushing process node to the grinding process node is confirmed. If the change in mill discharge particle size occurs first, followed by the reagent addition record in the flotation reagent process data, and the ball mill's operating power record exists continuously in the current time period, the logistics transmission edge from the grinding process node to the flotation process node is confirmed. If, after the flotation process is completed, continuous dewatering equipment operating power records appear in the dewatering process, and the tailings concentration record in the tailings treatment and discharge process data continues to appear in subsequent time periods, the logistics transmission edge from the dewatering process node to the tailings treatment process node is confirmed.

[0084] Energy consumption association edges are established based on the correspondence between energy consumption environmental parameter data and equipment operating power. Under a unified spatial index, the energy consumption environmental parameter data collection points are located to specific equipment nodes. Under a unified time index, the power consumption records in the energy consumption environmental parameter data and the equipment operating power records are compared to ensure that they exist simultaneously in the same time period, thereby determining the energy consumption association edges from the entity corresponding to the energy consumption environmental parameter collection point to the equipment node.

[0085] Furthermore, after obtaining the node set and edge set of the mineral processing process topology diagram, a unified spatial index is used to label the spatial location and process level of process nodes, equipment nodes, and material nodes. A unified time index is used to label the time interval and occurrence period of sequential constraint edges, logistics transmission edges, and energy consumption related edges. Subsequently, a consistency check is performed based on the production process layout diagram. This includes checking the sequential relationship of sequential constraint edges, the spatial continuity of logistics transmission edges, and the equipment correspondence of energy consumption related edges. After the check is completed, a data structure list of the mineral processing process topology diagram is generated. The node part of the data structure list records the node name, node category, spatial location, and process level. The edge part of the data structure list records the source node name, destination node name, relationship category, spatial location description, and time interval.

[0086] After completing the data structure list, process nodes, equipment nodes, and material nodes are first recorded as entity nodes in the entity node set, retaining attribute fields such as process name, equipment identifier, and material characteristic value. Then, sequence constraint edges, logistics transmission edges, and energy consumption association edges are recorded as relation edges in the relation edge set, specifying the source node, destination node, and corresponding relation type. After establishing the entity node set and relation edge set, to ensure the temporal and spatial consistency of the overall topology, a unified time index is used to label the occurrence time of each entity node and relation edge, and a unified spatial index is used to label the corresponding physical location of each entity node and relation edge. After labeling with time and spatial indices, the entity node set and relation edge set can be compared and merged one by one with the existing entities and relations in the initial mineral processing optimization knowledge graph. When a new process or equipment appears, an entity node is added; when an existing entity node or relation is updated, the corresponding attributes are modified. Finally, the topology is fully incorporated into the initial mineral processing optimization knowledge graph, forming an evolutionary mineral processing optimization knowledge graph covering processes, equipment, and materials.

[0087] S4. The initial multimodal dataset is timestamped by optimizing the knowledge graph through evolutionary mineral processing to obtain a time-axis aligned multimodal dataset.

[0088] Furthermore, in the evolutionary mineral processing optimization knowledge graph, the sequence constraint edge records the sequence of processes, and the material transfer edge records the direction of material transfer. Based on the process nodes connected by both the sequence constraint edge and the material transfer edge, a pairing relationship between upstream process nodes and downstream process nodes is established. Each pairing relationship between upstream process nodes and downstream process nodes covers a production period, which is jointly determined by the time range recorded by the sequence constraint edge and the time range recorded by the material transfer edge.

[0089] In the numerical sequences of upstream and downstream process nodes, the fields most sensitive to changes in the boundary of process events are selected as benchmark fields. The selection of benchmark fields is based on the response time and response magnitude of the field to process events. For example, in the grinding process node, the hydrocyclone inlet pressure field has the highest time sensitivity to the start and end of grinding events, so the hydrocyclone inlet pressure field is selected as the benchmark field; in the flotation process node, the reagent dosage field has the highest time sensitivity to the start and end of flotation events, so the reagent dosage field is selected as the benchmark field.

[0090] Based on the selected benchmark field, an upstream numerical sequence is formed in the upstream process node, and a downstream numerical sequence is formed in the downstream process node. The upstream numerical sequence is used to represent the critical process quantity of the upstream process node on the unified time index, and the downstream numerical sequence is used to represent the critical process quantity of the downstream process node on the unified time index.

[0091] After establishing the upstream and downstream numerical sequences, candidate delay values ​​are set within the covered production period. The candidate delay values ​​are determined based on the delay range defined by the sequence constraint edges and the material transport edges. The example value range is 5 to 30 seconds, and the value is determined based on the actual time characteristics of material transport between adjacent processes. On the production line, the transport process of material from the crushing process to the grinding process and from the grinding process to the flotation process usually takes several seconds to tens of seconds. Therefore, setting the delay range to 5 to 30 seconds can cover common transport durations. Under the unified time index, the candidate delay values ​​are converted into a discrete set of candidate time displacements. When the unified time index step size is set to 1 second, the example value range of the candidate time displacement set is 5 to 30 steps. The candidate time displacement set consists of all discrete step sizes within the delay range. Each candidate time displacement is used to represent the alignment offset between the upstream and downstream numerical sequences on the unified time index.

[0092] In the candidate time shift set, the correlation maximization criterion is adopted. The time shift in which the correlation between upstream and downstream numerical sequences reaches its maximum value within the production period is defined as the optimal time shift, expressed as:

[0093] ;

[0094] in, This represents the optimal time displacement. Indicates the candidate time shift. Indicates the upstream numerical sequence at the time index The baseline field on Indicates the downstream numerical sequence at the time index The baseline field on Indicates the displacement at the candidate time. Below, the Pearson correlation coefficient between the upstream and downstream numerical sequences ranges from 0 to 1. The value is based on the fact that the physical meaning of the benchmark fields in the mineral processing process all show a positive correlation. For example, the hydrocyclone inlet pressure and the mill output particle size, and the reagent addition amount and the flotation froth layer thickness all show a monotonically positive correlation. Therefore, the Pearson correlation coefficient only retains the non-negative interval, where 0 indicates no correlation and 1 indicates a completely positive correlation.

[0095] Obtain the optimal time displacement Subsequently, the downstream numerical sequences were shifted as a whole. After translation processing, the upstream and downstream numerical sequences achieve point-to-point correspondence on a unified time index covering the production period, thereby completing the timestamp alignment of the numerical sequences.

[0096] Furthermore, after completing the timestamp alignment of the numerical sequence, the timestamp alignment of the image samples is performed. By using the sequential constraint edges in the evolutionary mineral processing optimization knowledge graph, the acquisition time of each frame of ore particle morphology image and micro-fracture image is matched with a unified time index. Specifically, to ensure alignment accuracy, the sampling position is located within the production period covered by the unified time index based on the trigger timestamp of the image acquisition device. Then, an interpolation method is used to map the image acquisition time that is not completely matched to the nearest unified time index point, so that each ore particle morphology image and micro-fracture image has a unique anchor position on the unified time index, thus completing the timestamp alignment of the image samples.

[0097] After image-based samples are aligned, spectral feature timestamp alignment is performed. Based on the corresponding logistics transport edges in the evolutionary mineral processing optimization knowledge graph, a correspondence between spectral acquisition times and a unified time index is established. To ensure the continuity of spectral features on the unified time index, the spectral sampling time is first discretized, and then the optimal time shift is used. The result is a global shift of the spectral sampling time, which ensures that the spectral acquisition time maintains a point-to-point correspondence with the numerical sequence and image samples, thereby completing the timestamp alignment of spectral features.

[0098] After completing the synchronous alignment of numerical sequences, image samples, and spectral features under a unified time index, a time-axis aligned multimodal dataset is generated.

[0099] S5. Perform hierarchical temporal modeling and cross-modal association calculation on the time-axis aligned multimodal dataset to generate multimodal temporal association features.

[0100] Furthermore, in a time-axis aligned multimodal dataset, short-term modeling of numerical sequences is performed to identify short-term dynamic fluctuations in numerical sequences under a unified time index, and the arrival time of process event boundaries is predicted. The specific steps are as follows:

[0101] Long Short-Term Memory (LSTM) networks are used to model numerical sequences in the short term. A fixed-length short window is constructed under a unified time index. The length of the short window is determined by the average duration of historical process disturbances, such as the average duration of rapid fluctuations in hydrocyclone inlet pressure. Each short window consists of several consecutive unified time index points and their corresponding numerical sequence fields. Before entering the LSTM, the numerical sequences within the short window are detrended to eliminate slow drift, and then standardized to unify the dimensions. The preprocessed short window is then input into the LSTM. The LSTM uses an input gate to control the entry of new values, a forget gate to control the discarding of historical values, and an output gate to control the output of hidden states, thereby extracting high-frequency dynamic features within the short window.

[0102] In the short-term modeling training process, the process event boundary indicator field of the next unified time index point after the end of the short-term window is used as the prediction target. The process event boundary indicator field is a numerical sequence field that can mark the start and end of a process event. The training objective function is set as the short-term modeling mean square error between the predicted output and the actual observed value of the process event boundary indicator field, which is used to measure the accuracy of short-term modeling in identifying process event boundaries.

[0103] To ensure that the training samples can cover all the changing features within the short window, the short batch size is typically set between 16 and 64. This setting is based on the fact that a short window usually contains only a few dozen uniform time index points. If the short batch size exceeds the number of uniform time index points, the training samples cannot completely cover the entire short window, and the transition segment of the process event boundary will be weakened, making it difficult to identify the process event boundary features. Keeping the short batch size between 16 and 64 ensures that each training batch completely covers the data changes within the short window and enhances the sensitivity to the response to process event boundaries.

[0104] To suppress training instability caused by drastic fluctuations within the short-term window, the short-term learning rate is typically set between 1×10^-4 and 5×10^-4. This setting is based on the fact that numerical sequences within the short-term window often exhibit strong instantaneous fluctuations. If the short-term learning rate is set too high, gradient updates will be too drastic, and the training process is prone to oscillations or divergence. Controlling the short-term learning rate between 1×10^-4 and 5×10^-4 allows the input gate weight matrix, forget gate weight matrix, output gate weight matrix, and corresponding bias parameters within the Long Short-Term Memory network to converge smoothly during iterative updates, while ensuring training speed and avoiding the loss of high-frequency dynamic features.

[0105] Short-term training uses the time-expanded backpropagation algorithm to iteratively update all weight matrices and bias parameters within the Long Short-Term Memory network. The optimization method is Adam. During short-term modeling training, an early stopping criterion for the validation set is set. When the mean square error of short-term modeling no longer decreases in several consecutive iterations, the update is stopped to avoid overfitting of the LSTM weight parameters due to excessive iteration.

[0106] After short-term modeling training is completed, LSTM outputs short-term time-series latent vectors under a unified time index. These latent vectors are used to characterize short-term dynamic features such as sudden increases in power consumption, rapid changes in hydrocyclone inlet pressure, and instantaneous adjustments in reagent addition.

[0107] Short-term time-series latent vectors are concatenated according to a unified time index to form a long-term input sequence. By performing long-term modeling on the numerical sequence, long-term dependencies across processes and stages are identified under the unified time index. For example, the impact of changes in grinding cycle load rate on flotation reagent addition and concentrate grade after a certain delay. The specific steps are as follows:

[0108] A Transformer model based on a self-attention mechanism is used to model numerical sequences over the long term. A long-term window of fixed length is constructed under a unified time index. The length of the long-term window is determined based on the average duration of cross-process lag effects. For example, the average lag time of the impact of grinding cycle load rate on concentrate grade is used as a reference value. The long-term window is fed into the Transformer model as input. The Transformer model calculates the correlation between all unified time index points within the long-term window through a multi-head attention mechanism, thereby establishing cross-process correlations between changes in cycle load rate and reagent addition adjustment, and between energy consumption accumulation and tailings concentration changes.

[0109] In the long-term modeling training process, process efficiency index fields (such as concentrate grade and tailings concentration fields) at several future unified time index points are used as prediction targets. The training objective function is set as the long-term modeling mean square error between the predicted output and the observed values ​​of the process efficiency index fields, used to measure the accuracy of long-term modeling in predicting cross-process dependency effects. The long-term batch size example ranges from 8 to 32, set based on the fact that the long-term window contains a large number of unified time index points. If the long-term batch size is too large, the memory usage of long-term modeling training will increase and lead to the loss of dependency features. Choosing a range of 8 to 32 can balance computational overhead and dependency feature learning effect. The long-term learning rate example ranges from 5×10^-5 to 2×10^-4, set based on... To address the long span of cross-process dependencies, excessively high long-term learning rates can cause significant fluctuations in the attention weight matrix and feedforward layer weight matrix within the Transformer model during iterative updates, leading to unstable convergence. Controlling the long-term learning rate between 5×10^-5 and 2×10^-4 ensures stable updates to the Transformer model's weight matrix and bias parameters, guaranteeing gradual convergence of cross-process dependent features. Regularization strategies include a dropout rate, with examples ranging from 0.1 to 0.3, designed to prevent excessive concentration of attention weights at a single, uniform time index. The long-term modeling training process employs an early stopping criterion on the validation set; long-term modeling training is stopped when the long-term mean square error no longer decreases within several consecutive iterations.

[0110] After long-term modeling training is completed, the Transformer model outputs a long-term time-series latent vector at each unified time index point. The long-term time-series latent vector can characterize the long-term effects of the slow increase of the cyclic load rate, the periodic adjustment of the reagent addition amount, and the cumulative energy consumption on the concentrate grade.

[0111] Furthermore, image samples and spectral features are converted into vectors that are aligned and correlated with numerical sequences. The specific steps are as follows:

[0112] Neural networks (CNNs) are used to vectorize image samples and spectral features;

[0113] Under a unified temporal index, superpixel segmentation is first performed on the ore particle morphology image and microcrack image to form several image region nodes. The feature fields of the image region nodes include the image region gray mean, image region gray variance, image region edge density, image region principal direction gradient statistics, and image region area. Then, based on the unified spatial index, the adjacency relationship of the image regions is established to obtain the image region edges. The weight of the image region edges is jointly calculated based on the proportion of the contact boundary length of the image region and the image region feature similarity. The image region graph is composed of the image region nodes and the image region edges.

[0114] The image region map is input into a graph neural network for feature propagation and aggregation. The number of layers in the graph neural network ranges from 2 to 4. The reason for this is that when the number of layers is less than 2, feature propagation is insufficient and cannot cover the spatial relationship between image regions. When the number of layers exceeds 4, oversmoothing is likely to occur, causing feature differences to disappear. The number of hidden dimensions ranges from 64 to 256. The reason for this is that when the hidden dimension is less than 64, it is difficult to fully express the complex features such as gray-level statistics, edge density, and geometric information of the image region. When the hidden dimension exceeds 256, it is likely to cause excessive computational overhead and overfitting.

[0115] The training objectives for the image region map include two parts. First, the readout results of the image region map must be consistent with the ore particle size distribution statistics of the ore raw material attribute data in step S1, ensuring that the vector representation reflects macroscopic particle size characteristics. Second, the changes in image feature vectors under adjacent unified time indices must remain smooth to ensure temporal continuity and reduce noise fluctuations. The training method employs the Adam optimization method. The example image batch size ranges from 4 to 16, set because gradient estimation is unstable when the batch size is less than 4, while exceeding 16 makes it difficult to guarantee complete training due to memory overhead and limitations on the number of image region nodes. The example image learning rate ranges from 1×10^-4 to 5×10^-4, set because convergence is too slow when the learning rate is below 1×10^-4, while the short-term gradient update amplitude is too large when the learning rate is above 5×10^-4, easily leading to training oscillations. After the image region map training is completed, image feature vectors are output under a unified time index. These image feature vectors are used to characterize the ore particle morphology and microcrack distribution characteristics.

[0116] Under a unified time index, the fluorescence spectrum and chemical composition spectrum of the slurry reagent are discretized at equal intervals according to wavelength ranges to form spectral band nodes. The characteristic fields of the spectral band nodes include spectral band intensity, spectral band first derivative, and spectral band second derivative. Spectral band edges are established based on wavelength adjacency. At the same time, cross-band edges are established when there is a characteristic peak neighborhood to enhance the dependence of spectral peak intervals. The weight of the spectral band edges is calculated jointly based on the wavelength spacing and spectral intensity difference. The spectral band map is composed of spectral band nodes and spectral band edges.

[0117] The spectral band map is input into a graph neural network for feature propagation and aggregation. The number of layers in the graph neural network ranges from 2 to 4. The reason for this is that when the number of layers is less than 2, the correlation propagation between spectral band nodes is insufficient, making it difficult to capture cross-band dependencies. When the number of layers exceeds 4, oversmoothing effects are likely to occur, which weakens the differences in spectral peak regions. The number of hidden dimensions ranges from 64 to 256. The reason for this is that when the hidden dimension is less than 64, it is difficult to fully express the nonlinear combination of spectral intensity and derivative features. When the hidden dimension exceeds 256, it will bring redundant parameters and the risk of overfitting.

[0118] The training objectives for the spectral band plots include two parts. First, the readout results of the spectral band plots must be consistent with the proportion of ore chemical composition in the ore raw material attribute data of step S1, ensuring that the feature vector can accurately represent the chemical composition. Second, the readout results of the spectral band plots must be consistent with the reagent type records in the flotation reagent process data of step S1, ensuring that the spectral features can distinguish different reagent action modes. The training method uses the Adam optimization method, with the sample value range for the spectral batch size being 8 to 32. This value is set based on the gradient estimation when the spectral batch size is less than 8. Excessive bias can lead to unstable training, while a bias exceeding 32 results in an excessively large batch size, making it difficult to maintain fine-grained variations in spectral features within a small range. The example value range for the spectral learning rate is 5×10^-5 to 2×10^-4. This is based on the fact that training convergence is too slow when the learning rate is below 5×10^-5, while gradient updates are too drastic when the learning rate is above 2×10^-4, which can easily cause oscillations in the spectral peak region. After the spectral band map is trained, a spectral feature vector is output under a unified time index. The spectral feature vector is used to characterize the drug's action characteristics and the distribution characteristics of its chemical components.

[0119] It should be noted that the short-term latent vectors, long-term latent vectors, image feature vectors, and spectral feature vectors of numerical sequences are mapped to a unified model through linear projection operations. 3D space.

[0120] After completing the unified dimension mapping, cross-modal correlation calculations are performed on the short-term time-series latent vectors, long-term time-series latent vectors, image feature vectors, and spectral feature vectors of numerical sequences using the delay correlation function, as shown below:

[0121] ;

[0122] in, Indicates time index The candidate delay vectors of different modal vectors (short-term latent vectors of numerical sequences, long-term latent vectors of numerical sequences, image feature vectors, and spectral feature vectors) Cross-modal delay correlation values ​​under the given conditions Indicates the candidate delay amount. Represents the vector dimension. Short-term time-series latent vectors representing numerical sequences In time index The dimensional components, Long-term time-series latent vectors representing numerical sequences In time index The dimensional components, Represents the image feature vector at the time index The dimensional components, Represents the spectral eigenvector at the time index The Dimensional components.

[0123] Within the range of candidate delay values, the maximum value of the delay correlation function is taken as the multimodal temporal correlation feature, expressed as:

[0124] ;

[0125] in, Represents multimodal temporal correlation features. This indicates the search range for candidate delay values.

[0126] S6. Input the multimodal temporal correlation features into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy.

[0127] Furthermore, based on the relationships between process nodes, equipment nodes, and material nodes in the evolutionary mineral processing optimization knowledge graph, the multimodal temporal correlation features are mapped to mineral processing control indicators, including grinding particle size control indicators, flotation reagent dosage control indicators, energy consumption efficiency control indicators, and tailings discharge stability control indicators.

[0128] During the mapping process, multimodal time-series correlation features are matched with the historical records of process nodes to determine the target range of each process indicator under a unified time index. Specifically, multimodal time-series correlation features are matched and mapped with the historical records of grinding process nodes and their subordinate equipment nodes (including ball mills and hydrocyclones) to obtain the dynamic range of hydrocyclone inlet pressure and mill output particle size, forming grinding particle size control indicators; multimodal time-series correlation features are matched and mapped with the historical records of flotation process nodes and reagent dosing devices to obtain the dynamic range of reagent type and reagent dosage, forming flotation reagent dosage control indicators; multimodal time-series correlation features are matched and mapped with energy consumption and environmental parameter data nodes to obtain the correspondence between power consumption and equipment power load, forming energy efficiency control indicators; and multimodal time-series correlation features are matched and mapped with the historical records of tailings treatment process nodes and tailings conveying pump nodes to obtain the dynamic range of tailings concentration and tailings water recovery rate, forming tailings discharge stability control indicators.

[0129] The target range refers to the allowable range of values ​​for process control indicators determined in the evolutionary mineral processing optimization knowledge graph, combining historical records of process nodes, production experience, and process specifications.

[0130] After completing the mapping of process indicators, the dynamic change trend of multimodal time-series correlation features is compared with the target range of each process indicator, and process control prediction values ​​are generated within a number of unified time index points in the future. When the change trend of multimodal time-series correlation features falls into the target range, the process control prediction value is recorded as the value within the target range. When the change trend deviates from the target range, the process control prediction value is equal to the value of the deviation and the boundary of the target range, thus reflecting the possible numerical levels of grinding particle size, flotation reagent dosage, energy efficiency and tailings discharge stability within a given production cycle.

[0131] Furthermore, the grinding particle size control index, flotation reagent dosage control index, energy consumption efficiency control index, and tailings discharge stability control index are transformed into optimization objective functions, and the optimal mineral processing control strategy is generated under a multi-objective optimization algorithm.

[0132] Based on the matching results of grinding process nodes and equipment nodes (ball mill and hydrocyclone), the process control prediction value of mill output particle size is formed. Taking the target range of grinding particle size control index as the benchmark, the deviation between the process control prediction value of mill output particle size and the boundary of the target range of grinding particle size control index is used as the optimization objective function value of grinding particle size control index.

[0133] Based on the matching results of the flotation process nodes and reagent dosing device nodes, a process control prediction value for the amount of flotation reagent added is formed. Taking the target range of the flotation reagent dosage control index as a benchmark, the deviation between the process control prediction value of the amount of flotation reagent added and the boundary of the target range of the flotation reagent dosage control index is used as the optimization objective function value of the flotation reagent dosage control index.

[0134] Based on the matching results of energy consumption environmental parameter data nodes and equipment node power load, a process control prediction value for power consumption per unit processing volume is formed. Using the target range of energy efficiency control index as a benchmark, the deviation between the process control prediction value for power consumption per unit processing volume and the boundary of the target range of energy efficiency control index is used as the optimization objective function value of energy efficiency control index.

[0135] Based on the matching results of tailings treatment process nodes and tailings conveying pump nodes, process control prediction values ​​for tailings concentration and tailings water recovery rate are generated. Taking the target range of tailings discharge stability control index as the benchmark, the deviation between the process control prediction values ​​of tailings concentration and tailings water recovery rate and the boundary of the target range of tailings discharge stability control index is used as the optimization objective function value of tailings discharge stability control index.

[0136] If the predicted value of the process control falls within the target range of the corresponding control indicator, the optimization objective function value of the control indicator is zero; if the predicted value of the process control deviates from the target range of the control indicator, the optimization objective function value of the control indicator is equal to the deviation between the predicted value of the process control and the boundary of the target range of the control indicator. The larger the deviation, the larger the optimization objective function value of the control indicator, indicating that the deviation of the process control indicator is higher and the control effect is worse.

[0137] Simultaneously, the sequential constraint edges in the evolutionary mineral processing optimization knowledge graph are used as time constraints, the logistics transmission edges are used as material flow constraints, and the energy consumption correlation edges are used as energy consumption constraints, forming a constraint set. The constraint set ensures that the generated control strategy must conform to the existing process logic and physical constraints in the production process layout diagram. For example, the crushing process must be completed before the grinding process, the adjustment of flotation reagents must take into account the material conveying delay time, and the equipment power load must be kept within the safe operating range.

[0138] After clarifying the objective function and constraint set of the control indicators, a multi-objective optimization algorithm is used to solve the problem. Specifically, during the optimization process, the objective function values ​​of the four control indicators are calculated one by one by searching the solution space that satisfies the constraint set. The solution set that is not simultaneously superior to any other solution in all objective functions is selected to form the Pareto optimal solution set. Each solution in the Pareto optimal solution set represents the optimal combination of grinding particle size, flotation reagent dosage, energy efficiency, and tailings discharge stability under different trade-offs.

[0139] Using the Pareto optimal solution set as the optimal control strategy for mineral processing, each optimal control strategy has a clear numerical description under a unified time index and a unified spatial index, and can be directly converted into process control instructions.

[0140] S7. Transform the optimal mineral processing control strategy into process control instructions for execution, collect execution feedback, update the evolutionary mineral processing optimization knowledge graph online, and generate the optimal mineral processing control strategy for the next cycle.

[0141] Furthermore, the optimal control strategy for mineral processing is transformed into process control commands through industrial communication protocols (such as OPC UA protocol) and distributed to process equipment such as crushers, ball mills, hydrocyclones, flotation machines, reagent pumps, dewatering screens, thickeners, and tailings conveying pumps. Among them, the grinding particle size control index corresponds to the adjustment commands for ball mill speed and hydrocyclone valve opening, the flotation reagent dosage control index corresponds to the setting command for reagent pump flow rate, the energy efficiency control index corresponds to the adjustment commands for crusher and mill power, and the tailings discharge stability control index corresponds to the adjustment commands for tailings pump flow rate and dewatering screen speed.

[0142] During the execution of process control commands, feedback data is collected in real time, including equipment power load, reagent flow stability, concentrate grade fluctuations, and tailings discharge concentration. The feedback data is input into the evolutionary beneficiation optimization knowledge graph under a unified time index and a unified spatial index. By updating the attribute values ​​of process nodes, equipment nodes, and material nodes, the attributes and relationships of the evolutionary beneficiation optimization knowledge graph are dynamically evolved. This allows the optimal beneficiation control strategy to combine the latest feedback to form new optimization inputs in the next production cycle, thereby continuously outputting an adaptive optimal beneficiation control strategy.

[0143] This embodiment also provides a mineral processing optimization control system based on a knowledge graph, including:

[0144] The data acquisition module is used to collect data from the entire mineral processing process and preprocess it to generate an initial multimodal dataset.

[0145] The element extraction module is used to extract spatiotemporal process elements based on the initial multimodal dataset and construct an initial mineral processing optimization knowledge graph.

[0146] The topology construction module is used to build a mineral processing topology graph, incorporate the process equipment and material topology into the mineral processing optimization knowledge graph, and construct an evolutionary mineral processing optimization knowledge graph.

[0147] The time alignment module is used to align the timestamps of the initial multimodal dataset using the evolutionary mineral processing optimization knowledge graph, thereby obtaining a time-axis aligned multimodal dataset.

[0148] The temporal modeling module is used to perform hierarchical temporal modeling and cross-modal correlation calculation on time-axis aligned multimodal datasets, generating multimodal temporal correlation features;

[0149] The strategy reasoning module is used to input multimodal temporal correlation features into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy;

[0150] The instruction execution module is used to convert the optimal mineral processing control strategy into process control instructions for execution, collect execution feedback, update the evolutionary mineral processing optimization knowledge graph online, and generate the optimal mineral processing control strategy for the next cycle.

[0151] In summary, this invention achieves high-precision alignment of multimodal process information on the time axis by: aligning the initial multimodal dataset with timestamps using an evolutionary mineral processing optimization knowledge graph; and by performing hierarchical temporal modeling and cross-modal correlation calculations on the time-aligned multimodal dataset, extracting short-term fluctuation features and long-term dependency features of numerical sequences, and combining image feature vectors and spectral feature vectors, using a delay correlation function to calculate cross-modal temporal dependencies, thereby obtaining multimodal temporal correlation features and ensuring a unified representation of process dynamics and material characteristics.

[0152] 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 mineral processing optimization control method based on knowledge graphs, characterized in that: include, Collect data from the entire mineral processing process and preprocess it to generate an initial multimodal dataset; Spatiotemporal process elements were extracted from the initial multimodal dataset to construct an initial mineral processing optimization knowledge graph; Establish a mineral processing topology diagram, incorporate the process, equipment and material topology into the mineral processing optimization knowledge graph, and construct an evolutionary mineral processing optimization knowledge graph. The initial multimodal dataset was timestamped using an evolutionary mineral processing optimization knowledge graph to obtain a time-axis aligned multimodal dataset. Hierarchical temporal modeling and cross-modal correlation calculation are performed on time-axis aligned multimodal datasets to generate multimodal temporal correlation features; Multimodal temporal correlation features are input into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy; The optimal mineral processing control strategy is transformed into process control commands for execution. Execution feedback is collected and the evolutionary mineral processing optimization knowledge graph is updated online to generate the optimal mineral processing control strategy for the next cycle. The construction of the evolutionary mineral processing optimization knowledge graph includes obtaining process nodes, equipment nodes, and material nodes based on the mineral processing full-process data and production process layout diagram, and generating a set of nodes for the mineral processing process topology diagram; Based on the process sequence, material channel connection, and energy consumption correspondence, sequential constraint edges, material transport edges, and energy consumption correlation edges are established to generate the edge set of the mineral processing process topology graph. Label and verify the node set and edge set to generate a list of data structures for the mineral processing topology graph; The process nodes, equipment nodes, and material nodes in the data structure list are used as entity nodes, and the sequence constraint edges, logistics transmission edges, and energy consumption association edges are used as relation edges. These are then merged and updated with the initial mineral processing optimization knowledge graph to construct an evolutionary mineral processing optimization knowledge graph. The generation of multimodal temporal correlation features includes performing short-term and long-term modeling on numerical sequences to obtain short-term and long-term temporal latent vectors. Simultaneously, image-type samples and spectral features are converted into image feature vectors and spectral feature vectors, respectively; Cross-modal correlation calculations are performed on short-term temporal latent vectors, long-term temporal latent vectors, image feature vectors, and spectral feature vectors using the delay correlation function, and the maximum value of the delay correlation function is taken as the multimodal temporal correlation feature.

2. The mineral processing optimization control method based on knowledge graph as described in claim 1, characterized in that: The complete mineral processing data includes ore raw material attribute data, crushing and screening process data, grinding process data, flotation reagent process data, tailings treatment and discharge process data, and energy consumption and environmental parameter data.

3. The mineral processing optimization control method based on knowledge graph as described in claim 2, characterized in that: The preprocessing to generate the initial multimodal dataset includes unifying the time scale and normalizing the spatial coordinates of the entire mineral processing data to obtain a numerical sequence. Perform acquisition error correction and index annotation on the ore particle morphology images and micro fracture images in the ore raw material attribute data to obtain image samples; Noise smoothing, missing measurement completion, and wavelength range normalization were performed on the fluorescence spectrum and chemical composition spectrum of the pulp reagents in the flotation reagent process data to obtain spectral characteristics. Numerical sequences, image samples, and spectral features are combined to form an initial multimodal dataset.

4. The mineral processing optimization control method based on knowledge graph as described in claim 3, characterized in that: The construction of the initial mineral processing optimization knowledge graph includes obtaining time-dimensional process elements by performing time series analysis on the initial multimodal dataset; The location information of the collection points in the initial multimodal dataset is mapped using a unified spatial index to obtain spatial dimension process elements; Semantic parsing is performed on the initial multimodal dataset to obtain semantic dimension process elements; By combining time-dimensional, spatial-dimensional, and semantic-dimensional process elements, a set of process entities is established. Within the set of process entities, time evolution relationships are generated based on time-dimensional process elements, spatial connection relationships are generated based on spatial-dimensional process elements, and semantic dependency relationships are generated based on semantic-dimensional process elements. The time evolution relationships, spatial connection relationships, and semantic dependency relationships are recorded as multidimensional relationships of process entities. By using the set of process entities as nodes and the multidimensional relationships between process entities as edges, a structured organization is constructed to create an initial mineral processing optimization knowledge graph.

5. The mineral processing optimization control method based on knowledge graph as described in claim 4, characterized in that: The time-axis aligned multimodal dataset is generated by aligning the numerical sequences, image samples, and spectral features of the initial multimodal dataset with timestamps through an evolutionary mineral processing optimization knowledge graph under the constraints of sequential constraint edges and logistics transmission edges.

6. The mineral processing optimization control method based on knowledge graph as described in claim 5, characterized in that: The optimal control strategy for mineral processing includes matching multimodal temporal correlation features with process nodes, equipment nodes, and material nodes in the evolutionary mineral processing optimization knowledge graph to obtain grinding particle size control index, flotation reagent dosage control index, energy consumption efficiency control index, and tailings discharge stability control index. Each control index is transformed into an optimization objective function, and a multi-objective optimization algorithm is used to solve it under the constraint set composed of sequential constraint edges, logistics transmission edges, and energy consumption correlation edges, so as to obtain the Pareto optimal solution set, which serves as the optimal control strategy for mineral processing.

7. The mineral processing optimization control method based on knowledge graph as described in claim 6, characterized in that: The process of converting the optimal mineral processing control strategy into process control instructions includes converting the optimal mineral processing control strategy into process control instructions through an industrial communication protocol and sending them to the process equipment for execution. During the execution of process control commands, process execution feedback data such as equipment power load, reagent flow stability, concentrate grade fluctuation, and tailings discharge concentration are collected. The process execution feedback data is input into the evolutionary beneficiation optimization knowledge graph to update the attribute values ​​of process nodes, equipment nodes, and material nodes. The process execution feedback data is then used as new optimization input to generate the optimal beneficiation control strategy for the next cycle.

8. A mineral processing optimization control system based on a knowledge graph, based on the mineral processing optimization control method based on any one of claims 1 to 7, characterized in that: include, The data acquisition module is used to collect data from the entire mineral processing process and preprocess it to generate an initial multimodal dataset. The element extraction module is used to extract spatiotemporal process elements based on the initial multimodal dataset and construct an initial mineral processing optimization knowledge graph. The topology construction module is used to build a mineral processing topology graph, incorporate the process equipment and material topology into the mineral processing optimization knowledge graph, and construct an evolutionary mineral processing optimization knowledge graph. The time alignment module is used to align the timestamps of the initial multimodal dataset using the evolutionary mineral processing optimization knowledge graph, thereby obtaining a time-axis aligned multimodal dataset. The temporal modeling module is used to perform hierarchical temporal modeling and cross-modal correlation calculation on time-axis aligned multimodal datasets, generating multimodal temporal correlation features; The strategy reasoning module is used to input multimodal temporal correlation features into the evolutionary mineral processing optimization knowledge graph for reasoning to obtain the optimal mineral processing control strategy; The instruction execution module is used to convert the optimal mineral processing control strategy into process control instructions for execution, collect execution feedback, update the evolutionary mineral processing optimization knowledge graph online, and generate the optimal mineral processing control strategy for the next cycle.

Images