An AI-driven method for fully automated collaborative control of a mineral processing plant

Abstract

This invention discloses an artificial intelligence-driven fully automated collaborative control method for a mineral processing plant, relating to the field of mine automation and intelligent control technology. The method includes the following steps: S1: Constructing a material flow time-delay map of the entire mineral processing plant process to quantify the material transmission delay time between each process, from raw ore input, crushing, grinding, classification, flotation, concentration, filtration, tailings discharge to concentrate output; S2: Real-time collection of process data for each process using online detection instruments and sensors distributed throughout the process; dynamic time-axis alignment of the process data for each process based on the material flow time-delay map. By establishing a material flow time-delay map of the entire mineral processing plant process, the sampled data of each process are dynamically aligned according to the actual arrival time of the material, effectively eliminating the problem of data spatiotemporal misalignment caused by material transmission delays.

Description

Technical Field

[0001] This invention relates to the field of mine automation and intelligent control technology, and in particular to an artificial intelligence-driven method for full-process automated collaborative control of a mineral processing plant. Background Technology

[0002] Mineral processing plants are a core component of mining production. Their production process typically encompasses multiple steps, including raw ore input, crushing, grinding, classification, flotation, concentration, filtration, and tailings discharge. These steps involve strong material flow coupling and dynamic time delays. The mineral processing process is characterized by multiple sequential steps, significant material flow delays, frequent operational disturbances, and mutually constraining control objectives, posing a significant challenge to achieving coordinated and optimized control across the entire process.

[0003] In existing mineral processing plant process control technologies, distributed control systems (DCS) or programmable logic controllers (PLCs) are widely used to automate various processes. However, existing technologies have the following shortcomings:

[0004] Significant material transfer delays exist between different processes in a mineral processing plant. After the raw ore enters the crusher, it undergoes multiple stages of processing, including crushing, grinding, and classification, before the changes in the raw ore's properties affect the quality of the flotation concentrate. This material flow lag causes misalignment of sampling data across different processes on the timeline. Furthermore, the control systems of each process in existing mineral processing plants are relatively independent, lacking a global collaborative optimization mechanism, making it difficult for the overall technical and economic indicators of the plant to reach optimal levels. Additionally, the lack of simulation verification mechanisms in existing mineral processing plants leads to the system model easily outputting unreasonable control setpoints, resulting in production fluctuations. Summary of the Invention

[0005] The embodiments of the present invention provide an artificial intelligence-driven fully automated collaborative control method for a mineral processing plant, which aims to solve problems such as misalignment of sampling data in each process on the time axis, lack of a global collaborative optimization mechanism leading to difficulty in achieving the best overall technical and economic indicators for the mineral processing plant, and lack of a simulation verification mechanism leading to the system model easily outputting unreasonable control setpoints and causing production fluctuations.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0007] An AI-driven method for fully automated collaborative control of a mineral processing plant includes the following steps:

[0008] S1: Construct a material flow time delay map of the entire process of the mineral processing plant to quantify the material transfer delay time between each process from raw ore input, crushing, grinding, classification, flotation, concentration, filtration, tailings discharge to concentrate output;

[0009] S2: By using online detection instruments and sensors distributed in each process, process data of each process is collected in real time. Based on the material flow time delay map, the process data of each process is dynamically aligned on the time axis to obtain the standard timestamp after the data of each process is aligned.

[0010] S3: Construct a full-process state vector based on the standard timestamps after aligning the data of each process.

[0011] S4: Construct a time-delay perception fusion attention time-series model, train the time-delay perception fusion attention time-series model, input the full-process state vector into the trained time-delay perception fusion attention time-series model for inference calculation, and output the optimal control setpoint for each process.

[0012] S5: Perform simulation verification on the optimal control setpoints for each process to obtain the simulation verification results;

[0013] S6: When the simulation verification result shows an anomaly, block the issuance of the optimal control setting value for each process, and regenerate a new optimized control setting value from the time delay perception fusion attention timing model.

[0014] S7: When the simulation verification results show that everything is normal, closed-loop feedback control is completed based on the optimal control setpoints of each process to achieve coordinated and optimized operation of the entire mineral processing plant.

[0015] Furthermore, in S1, the material flow time delay map includes: material transport path for each process, process node number, single process time delay duration, inter-process coupling time delay, and total process cumulative time delay, with the specific formula as follows:

[0016]

[0017] in: The total material transport time delay for the entire mineral processing process; This represents the total number of mineral processing steps. This is the process sequence number, with a value range of 1 ≤ ≤ ; For the first Material handling delays in each process step; This refers to the number of material transfer nodes between adjacent processes. For the first Material transfer interval time delay at each transfer node.

[0018] Furthermore, in S2, the dynamic time axis alignment of the process data for each process step specifically involves: relying on the inherent time delay parameters of each process step in the material flow time delay diagram, and using the quantified raw ore input time as the global time reference. Combining the formula:

[0019]

[0020] Calculation yields the first Standard time coordinates corresponding to each process data The raw process data collected by various sensors and online monitoring instruments are time-corrected and offset-compensated according to the standard time coordinates of the corresponding processes. This eliminates data misalignment caused by material transport delays and achieves synchronous alignment of process data across the entire process on a unified time axis. For the first Standard timestamps after data alignment for each process step; This is the initial reference time when the raw ore enters the mineral processing system; For the first The processing time of each step is delayed.

[0021] Furthermore, in S3, the full-process state vector includes: ore property parameter sub-vector, equipment operating state parameter sub-vector, process control parameter sub-vector, product quality index sub-vector, and economic operation index sub-vector;

[0022] The specific sub-vectors of ore property parameters are: raw ore grade, raw ore particle size distribution, raw ore moisture content, raw ore mudification index, ore grindability index, and ore floatability index.

[0023] The specific sub-vectors of the equipment operating status parameters are: crusher main shaft current, crusher discharge port opening, mill cylinder vibration signal, mill ball loading, classifier overflow concentration, flotation machine impeller speed, flotation machine aeration volume, thickener rake frame torque, and filter vacuum degree.

[0024] The process control parameter sub-vectors are specifically: crusher feed rate setting value, mill water feed rate setting value, mill feed rate setting value, classifier makeup water flow rate setting value, flotation reagent addition amount setting value, flotation liquid level setting value, thickener underflow pump frequency setting value, and filter belt speed setting value.

[0025] The specific product quality index sub-vectors are: particle size of crushed products in each stage, fineness of grinding products, fineness of classification overflow, thickness of froth layer in each flotation cell, concentrate grade, metal recovery rate, tailings grade, concentrate moisture, and turbidity of return water.

[0026] The specific sub-vectors of the economic operation indicators are: unit raw ore power consumption, unit concentrate steel ball consumption, unit concentrate reagent consumption, unit concentrate production cost, equipment comprehensive utilization rate, and concentrate output rate.

[0027] Furthermore, in S4, the time-delay-aware fusion attention temporal model specifically includes:

[0028] Input preprocessing layer: used to perform outlier removal, normalization, and format conversion on the time-series state vector;

[0029] Material delay graph feature extraction layer: Based on the material flow delay graph, a process topology directed graph is constructed, and a time-series graph convolutional network is used to extract the process spatial coupling features and material transport delay features;

[0030] Multi-head temporal attention layer: Adaptively allocates temporal feature weights to highlight key operating conditions and abnormal features;

[0031] Spatiotemporal fusion coding layer: A bidirectional long short-term memory network is used to fuse spatiotemporal features to capture the long-term evolution law of the mineral processing process;

[0032] Multi-objective decision output layer: Combining multi-objective reinforcement learning strategies, it outputs the optimal control setpoints for each process.

[0033] Furthermore, in S4, the training of the time-delay perception fusion attention time-series model is based on historical production process data of the ore dressing plant. After dynamic alignment of the time axis and construction of the full-process state vector, a model training dataset is formed. The model training dataset takes maximizing the concentrate grade, maximizing the metal recovery rate, and minimizing the unit concentrate production cost as optimization objectives. A comprehensive fitness function is constructed as the basis for judging the performance of individual models. The specific formula is as follows:

[0034]

[0035] Where: F: Individual comprehensive fitness value, the larger the value, the better the control scheme; This represents the current concentrate grade; The maximum concentrate grade allowed by the process; This represents the current metal recovery rate; The maximum metal recovery rate allowed by the process; This represents the current unit cost of concentrate production. The lowest production cost allowed by the process; , , These are the weighting coefficients corresponding to concentrate grade, metal recovery rate, and production cost, respectively, satisfying... .

[0036] Furthermore, in S5, the simulation verification of the optimal control setpoints for each process specifically involves:

[0037] The optimal control setpoints of each process output by the time-delay perception fusion attention timing model are input into the mineral processing digital twin model. The operating conditions of the entire process over a future period are deduced with a preset simulation step size. The operating parameters of each process equipment, product quality indicators and economic operation indicators are obtained through synchronous simulation.

[0038] The simulation results are judged by combining the three evaluation indicators: concentrate grade, metal recovery rate, and unit concentrate production cost.

[0039] Furthermore, the specific formula for the mineral processing digital twin model is as follows:

[0040]

[0041] in: The simulation output vector of the digital twin model contains simulation results such as concentrate grade, metal recovery rate, unit concentrate production cost, equipment operating parameters, and process status within the next T time period, corresponding to product quality, economic, and equipment indicators; This is a simulation operator that integrates the entire mineral processing process, representing the serial coupling relationship between the sub-models of crushing, grinding, classification, flotation, concentration, and filtration. The input vector at time t represents the optimal control setpoints for each process output by the collaborative control model. The state vector of the entire process at time t includes the sub-vectors of the ore property parameters, the sub-vectors of the equipment operating status parameters, the sub-vectors of the process control parameters, the sub-vectors of the product quality indicators, and the sub-vectors of the economic operation indicators. The material transfer time delay vector is taken from the material flow time delay map and represents the material transfer delay constraint between each process. A fixed parameter set for the model, including static parameters such as equipment rated parameters, process safety thresholds, and model training weights; The preset simulation duration.

[0042] Furthermore, in S7, the closed-loop feedback control specifically involves: sending the optimal control setpoint to the programmable logic controller (PLC) of each process through an industrial communication network; the PLC comparing the optimal control setpoint with the real-time detection values ​​of each process; calculating using a PID control algorithm; and outputting corresponding control commands to the frequency converter, electric actuator, pneumatic regulating valve, or reagent dosing pump to achieve precise adjustment of the process variables of each process.

[0043] Beneficial effects:

[0044] This invention establishes a time-delay graph of the entire material flow in a mineral processing plant, dynamically aligning the sampling data of each process according to the actual arrival time of the material. This effectively eliminates the problem of data spatiotemporal misalignment caused by material transmission delays, enabling the constructed full-process state vector to truly reflect the actual state of the same batch of material in each process, providing an accurate state perception basis for subsequent control decisions. Furthermore, by constructing the full-process state vector and employing a time-delay perception fusion attention timing model to globally optimize and match the control parameters of each process, this invention overcomes the problem of relatively independent control systems for each process and the lack of a global collaborative optimization mechanism in traditional control methods, achieving global optimization of the overall operating performance of the mineral processing plant. Simulation verification of the optimal control setpoints for each process effectively avoids production fluctuations caused by unreasonable control setpoints output by the time-delay perception fusion attention timing model, reducing the application risk of intelligent control systems. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention; Detailed Implementation

[0046] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0047] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0048] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0049] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0050] Combination Figure 1 As shown, this invention provides an artificial intelligence-driven fully automated collaborative control method for a mineral processing plant, comprising the following steps:

[0051] S1: Construct a material flow time delay map of the entire process of the mineral processing plant to quantify the material transfer delay time between each process from raw ore input, crushing, grinding, classification, flotation, concentration, filtration, tailings discharge to concentrate output;

[0052] The material flow time delay diagram includes: material transfer path for each process, process node number, single process time delay duration, inter-process coupling time delay, and total process cumulative time delay. The specific formula is as follows:

[0053]

[0054] in: The total material transport time delay in the entire mineral processing process is expressed in time units (s / min / h). It represents the total time it takes for a single batch of raw ore to go from entering the system to forming concentrate and tailings. It is the core parameter characterizing the time delay characteristics of the entire process. This represents the total number of mineral processing steps, including all processes such as crushing, grinding, classification, flotation, concentration, filtration, tailings discharge, and concentrate output. It is a fixed statistical value. This is the process sequence number, with a value range of 1 ≤ ≤ This is used to iterate through each production process. For the first Material handling lag in a process refers to the inherent time required for materials to remain in the equipment of that process to complete processes such as processing, sorting, and concentration. It is determined by the equipment specifications, process system, and throughput. This refers to the number of material transfer nodes between adjacent processes, including all material transfer and conveying nodes such as belt conveyors, silos, chutes, and pumps. For the first The material transfer interval delay at each transfer node refers to the time delay caused by the transfer, conveying, and temporary storage of materials between two adjacent processes. It is the sum of the processing delays of all processes, reflecting the total time that materials stay in various processing equipment; It is the sum of the transfer delays between all process nodes, reflecting the total time for materials to move and be temporarily stored between equipment.

[0055] The material flow time delay map uses the process topology map as a carrier to visualize and structure all the time delay parameters, process numbers, and transmission paths mentioned above, serving as the basic database for full-process time-series data processing.

[0056] S2: Through online detection instruments and sensors distributed in each process, process data of each process is collected in real time. Based on the material flow time delay map, the process data of each process is dynamically aligned on the time axis to obtain the standard timestamp after the data of each process is aligned.

[0057] The specific steps for dynamically aligning the process data of each process on the time axis are as follows: relying on the inherent time delay parameters of each process in the material flow time delay diagram, and using the quantified raw ore input time as the global time reference. Combining the formula:

[0058]

[0059] Calculation yields the first Standard time coordinates corresponding to each process data The raw process data collected by various sensors and online monitoring instruments are time-corrected and offset-compensated according to the standard time coordinates of the corresponding processes. This eliminates data misalignment caused by material transport delays and achieves synchronous alignment of process data across the entire process on a unified time axis. For the first The standard timestamp after data alignment for each process represents the theoretical standard collection time of the corresponding batch of materials for that process, and is the core identifier of the unified timeline of data for the entire process. The initial reference time for the raw ore entering the beneficiation system is the initial moment when the raw ore officially enters the beneficiation system and the first process begins feeding. It is the zero point of the entire process. For the first The processing time of each step is delayed.

[0060] Based on the time of raw ore feeding Starting from the first step, the processing delays of all preceding processes and the inter-process transfer delays are added sequentially to calculate the standard arrival time of the material in the current process. ;According to the raw sensor data Offset compensation and timing correction are performed to ultimately achieve data synchronization and alignment of the same batch of materials in all processes.

[0061] S3: Construct a full-process status vector based on the standard timestamps after data alignment for each process.

[0062] The overall process state vector includes: ore property parameter subvector, equipment operating status parameter subvector, process control parameter subvector, product quality index subvector, and economic operation index subvector;

[0063] The specific sub-vectors of ore property parameters are: Ore grade, derived from online X-ray fluorescence analysis or laboratory testing, representing the mass percentage of useful metal elements in the ore. For example, copper ore grade is typically between 0.3% and 2.0%, while lead-zinc ore grade (Pb+Zn) is typically between 2% and 10%. This parameter is the most important feed quality indicator, directly determining the control strategies for subsequent processes and the expected concentrate output. Ore particle size distribution, derived from an online particle size analyzer, describes the proportional distribution of particles of different sizes in the ore. The particle size corresponding to a cumulative yield of 80% is commonly used as a representative indicator, or it can be fully characterized by the two characteristic parameters of the Rosin-Rammler distribution function. Ore moisture content, derived from a microwave moisture analyzer, represents the mass percentage of water contained in the ore. Excessive moisture content affects crushing efficiency and screening effect, while insufficient moisture content increases dust pollution. Generally, the moisture content of raw ore mined in open-pit mines is 3%–8%, while that in underground mines is 5%–12%. The ore slime index, derived from settling tests, is a comprehensive indicator characterizing the content of easily slime-forming minerals in the raw ore. Higher slime levels cause more severe interference with the flotation process, manifesting as slime covering the surface of valuable minerals, hindering reagent adsorption, increasing pulp viscosity, and reducing separation efficiency. The ore grindability index, derived from the standard Bond grindability test, quantitatively characterizes the ore's resistance to grinding. The Bond work index is an internationally recognized measure of ore grindability; a higher value indicates that the ore is more difficult to grind. The ore floatability index, derived from micro-flotation tests, comprehensively characterizes the natural floatability of valuable minerals in the ore. This index is influenced by various factors such as the type of valuable mineral, its particle size, oxidation level, and gangue mineral type.

[0064] The specific sub-vectors of equipment operating status parameters are as follows: Crusher main shaft current, sourced from a current transformer, is a core electrical parameter reflecting the actual load of the crusher. Excessive current indicates the presence of uncrushable foreign objects or excessive iron in the crushing chamber, or overload operation; insufficient current indicates insufficient material in the crushing chamber, leading to idling and wasted energy. Crusher discharge port opening, sourced from a discharge port displacement sensor, is the width of the closed-side discharge port of a cone crusher or jaw crusher. This parameter directly determines the maximum particle size of the crushed product and is one of the most important operating parameters in the crushing section. Grinding mill cylinder vibration signal, sourced from an acceleration sensor, is the amplitude of vibration acceleration on the surface of the ball mill / semi-autogenous mill cylinder. The trend of the vibration signal can indirectly reflect the impact state between the grinding media and the liner, and the mill load level. Grinding mill ball load, sourced from a ball load meter, is the percentage of the total filling volume of the grinding media to the effective volume of the mill. The ball charge directly affects grinding efficiency and product particle size distribution; the optimal ball charge for ball mills is typically 30%–45%. The overflow concentration of the classifier, derived from a nuclear concentration meter, indicates the solid mass concentration in the overflow product of the classifying hydrocyclone or spiral classifier. Excessive overflow concentration means decreased classification efficiency, while insufficient concentration may lead to inadequate subsequent flotation pulp concentration, affecting flotation efficiency. The impeller speed of the flotation machine, derived from a speed sensor, measures the rotational speed of the impeller in a mechanically agitated flotation machine. Speed ​​affects the agitation intensity and air dispersion capacity of the pulp. Excessively high speed, while enhancing pulp suspension, also increases energy consumption and impeller wear; excessively low speed may result in the loss of coarse mineral particles. The aeration rate of the flotation machine, derived from a flow meter, measures the volumetric flow rate of air drawn into the flotation machine per unit time. The aeration rate directly affects the bubble generation rate and the quality of mineralized foam formation, and is one of the key operating variables controlling the flotation rate constant. The rake torque of the thickener, derived from a torque sensor, measures the resistance torque borne by the central rotating rake of the thickener. Increased torque typically indicates higher concentration and thicker deposits in the underflow compression zone, serving as an important reference signal for adjusting the underflow pump discharge rate. The filter vacuum level originates from the vacuum pressure transmitter and the vacuum level inside the sealing casing of a disc vacuum filter or ceramic filter. Vacuum level directly affects the moisture removal efficiency and filter cake thickness. Excessive vacuum may damage the integrity of the filter cake, while insufficient vacuum will result in excessive moisture content.

[0065] The process control parameter sub-vectors are as follows: The crusher feed rate setpoint, derived from the variable frequency speed control feed belt, represents the design target feed rate of the crusher feed belt. This value needs to be set in coordination with factors such as the crusher's rated power, ore hardness, and discharge port opening. The mill water feed rate setpoint, derived from the electric regulating valve, represents the target flow rate of new water added to the mill. The water feed rate directly affects the grinding concentration, thus affecting grinding efficiency and product particle size. The mill feed rate setpoint, derived from the variable frequency speed control feed belt, represents the target new feed rate to the mill. The classifier makeup water flow rate setpoint, derived from the electric regulating valve, represents the makeup water flow rate at the inlet of the classifier hydrocyclone or spiral classifier. Adjusting the makeup water flow rate can change the classification particle size and classification efficiency. The flotation reagent addition rate setpoint, derived from the metering pump, represents the target addition rate of flotation reagents, including the addition amounts of various reagents such as collectors, frothers, and modifiers. Each reagent may have an independent set value; here, it refers to the total reagent equivalent. The flotation liquid level set value is derived from the electric liquid level control valve, representing the target value for the slurry level height in the flotation cell. A stable liquid level is fundamental to ensuring a constant flotation froth layer thickness and stable concentrate grade. The thickener underflow pump frequency set value is derived from the frequency converter, specifically the frequency setting value of the thickener underflow discharge pump. The frequency directly determines the underflow discharge rate and underflow concentration. The filter belt speed set value is derived from the variable frequency speed control motor, representing the operating speed of the vacuum filter cake discharge belt. Belt speed affects filter cake thickness and moisture content—a faster speed results in a thinner filter cake with lower moisture content, while a slower speed results in a thicker filter cake with higher moisture content.

[0066] The specific product quality indicators are as follows: Particle size of each crushing stage, derived from an online particle size analyzer, representing the characteristic particle size of the coarse, medium, and fine crushing stages; Fineness of the grinding product, derived from an online particle size analyzer, representing the percentage of the grinding mill discharge product passing through a 200-mesh (75μm) or 325-mesh (45μm) sieve. Fineness is a key indicator for measuring the quality of grinding products; Fineness of the classifying overflow, derived from an online particle size analyzer, representing the fineness of the overflow product from the classifying equipment; Thickness of the foam layer in each flotation cell, derived from visual analysis of foam images, representing the vertical thickness of the foam layer in the k-th flotation cell. Foam layer thickness is an important reference for operators in the flotation section to visually judge the flotation status; Concentrate grade, derived from an online X-ray fluorescence analyzer, representing the percentage content of useful elements in the final concentrate product. This is one of the core product quality assessment indicators for a mineral processing plant; metal recovery rate, derived from laboratory results, is the percentage of useful metals recovered from the raw ore into the concentrate; tailings grade, derived from an online X-ray fluorescence analyzer, is the content of residual useful metals in the tailings waste product. Excessively high tailings grade indicates significant metal loss and should be reduced; concentrate moisture, derived from an infrared moisture analyzer, is the residual moisture content in the concentrate filter cake after filtration. Excessive moisture increases transportation costs and energy consumption in subsequent smelting processes; return water turbidity, derived from an online turbidity meter, is the degree of turbidity of the return water in the mineral processing plant, expressed in NTU (scattered turbidity unit) or mg / L (suspended solids concentration).

[0067] The specific sub-vectors of the economic operation indicators are as follows: Electricity consumption per unit of raw ore: the total electrical energy consumed to process one ton of raw ore (kWh / t), covering all electrical equipment such as crushing, grinding, flotation, and conveying; Steel ball consumption per unit of concentrate: the mass of steel balls consumed to produce one ton of concentrate (kg / t concentrate). Steel balls are one of the main consumables in a concentrator, and their consumption directly affects operating costs; Reagent consumption per unit of concentrate: the total cost of flotation reagents consumed to produce one ton of concentrate; Production cost per unit of concentrate: the total cost to produce one ton of concentrate (yuan / t concentrate) after comprehensively considering factors such as electricity, steel balls, reagents, labor, depreciation, and maintenance. This is the most core economic evaluation indicator; Equipment utilization rate: the proportion of the average operating time of key equipment in the entire plant to the total calendar time (%). This reflects the reliability level and maintenance management level of the equipment; Concentrate yield: the percentage of the quality of the concentrate product obtained after beneficiation of the raw ore to the total quality of the raw ore (%). Also known as "concentrate productivity".

[0068] S4: Construct a time-delay perception fusion attention time-series model, train the time-delay perception fusion attention time-series model, input the full-process state vector into the trained time-delay perception fusion attention time-series model for inference calculation, and output the optimal control setpoint for each process.

[0069] The time-delay-aware fusion attention time-series model specifically includes:

[0070] Input preprocessing layer: used to perform outlier removal, normalization, and format conversion on the temporal state vector. Outlier removal: uses the 3σ criterion or IQR method to identify and replace outlier data that exceeds the reasonable range. Normalization: maps heterogeneous data of each dimension to a unified numerical range of [0,1] or [-1,1]. Commonly used normalization methods include Min-Max normalization and Z-Score normalization. Format conversion: reshapes the preprocessed vector into the tensor format required for model input, usually a three-dimensional tensor shape of [batch_size, sequence_length, feature_dim].

[0071] Material delay graph feature extraction layer: Based on the material flow delay graph, a process topology directed graph is constructed. A time-series graph convolutional network is used to extract process spatial coupling features and material transport delay features. Process spatial coupling features: Capture the implicit dependencies between adjacent and non-adjacent processes caused by material associations. For example, how changes in crushed product particle size ultimately affect the fluctuation of flotation recovery rate through a series of intermediate steps. Material transport delay features: Encode the edge weights in the delay graph as edge attributes of the graph, enabling graph convolution operations to perceive the time decay effect during information transmission.

[0072] Multi-head temporal attention layer: Adaptively allocates temporal feature weights to highlight key operating conditions and abnormal features;

[0073] The working mechanism of the multi-head temporal attention layer is as follows: It linearly maps the input temporal feature sequence into three tensors: query (Q), key (K), and value (V); and calculates the attention weight matrix. Multiple parallel and independent "heads" are used to focus on different temporal patterns; the outputs of the multiple heads are concatenated and then linearly projected to obtain the final temporal weighted features;

[0074] The multi-head time-series attention layer can automatically highlight time steps that contain key operating condition turning points or abnormal signs (such as sudden drops in quality, equipment alarm precursors, etc.), suppress information redundancy during normal stable periods, thereby improving the model's sensitivity to important events.

[0075] Spatiotemporal fusion coding layer: A bidirectional long short-term memory network is used to fuse spatiotemporal features to capture the long-term evolution law of the mineral processing process;

[0076] The characteristic of bidirectional long short-term memory networks is that they maintain a forward LSTM and a backward LSTM at the same time: the forward LSTM reads the sequence in forward chronological order to capture the causal dependency relationship of "past → present"; the backward LSTM reads the sequence in reverse chronological order to capture the contextual relationship relationship of "future → present"; the hidden states of the two are spliced ​​together at each time step to form a complete encoding of the bidirectional information before and after that moment.

[0077] Multi-objective decision output layer: Combining multi-objective reinforcement learning strategies, it outputs the optimal control setpoints for each process.

[0078] The output dimension of this layer is consistent with the dimension of the process control parameter sub-vector, with each dimension corresponding to the optimal setpoint of a controllable variable. The range of output values ​​is limited by the physical constraints of each actuator, which are implemented through the output activation function and post-processing pruning.

[0079] Training the time-delay-aware fusion attention time-series model is based on historical production process data from the ore dressing plant. After dynamic alignment of the time axis and construction of the full-process state vector, a model training dataset is formed. The model training dataset aims to maximize concentrate grade, maximize metal recovery rate, and minimize unit concentrate production cost. A comprehensive fitness function is constructed as the criterion for evaluating the performance of individual models. The specific formula is as follows:

[0080]

[0081] Where: F: Individual overall fitness value, a scalar score used to evaluate the overall quality of a set of control settings. The larger the F value, the better the corresponding control scheme. In policy gradient optimization in reinforcement learning or selection operations in genetic algorithms, the F value directly determines the survival probability or selection probability of the individual; This is the predicted value after running the system with the current concentrate grade and using the set control values. This is the maximum concentrate grade allowed by the process, based on the upper limit of concentrate grade requirements stipulated in the smelter's procurement contract or industry standards. This value is used as the normalization denominator to make... It becomes a dimensionless ratio between 0 and 1; Given the current metal recovery rate, this is the predicted useful metal recovery rate after operating with this set of control setpoints. A higher recovery rate means higher resource utilization and less metal loss. This represents the maximum metal recovery rate allowed by the process. This limit is constrained by inherent factors such as ore selectivity, embedding characteristics, and monomer liberation degree, and cannot be achieved by any arbitrary control strategy. This value is also used as the normalization denominator. This represents the current unit concentrate production cost, and the predicted comprehensive production cost per ton of concentrate after operating with this set of control setpoints. Lower costs are more advantageous. Note that this term is in the denominator, therefore, lower costs mean higher returns. A larger ratio leads to a larger F-value, which aligns with the optimization principle of "the lower the cost, the better." This represents the lowest theoretically achievable production cost allowed by the process, while ensuring product quality and recyclability. This value, as the numerator, makes... It becomes a dimensionless ratio between 0 and 1; The weighting coefficient corresponding to the concentrate grade determines the relative importance of the concentrate grade target in the comprehensive evaluation; The weighting coefficients corresponding to the metal recovery rate determine the relative importance of the recovery rate target; The weighting coefficients corresponding to production costs determine the relative importance of cost-saving targets, satisfying... .

[0082] S5: Perform simulation verification of the optimal control setpoints for each process and obtain the simulation verification results;

[0083] The simulation verification of the optimal control setpoints for each process is as follows:

[0084] The optimal control setpoints of each process output by the time delay perception fusion attention time sequence model are input into the mineral processing digital twin model. The operating conditions of the entire process in the future are deduced by the preset simulation step size, and the equipment operating parameters, product quality indicators and economic operation indicators of each process are obtained by synchronous simulation.

[0085] The simulation results are evaluated by combining three evaluation indicators: concentrate grade, metal recovery rate, and unit concentrate production cost.

[0086] In an embodiment of the present invention, the specific formula for the mineral processing digital twin model is as follows:

[0087]

[0088] in: The simulation output vector of the digital twin model contains simulation results such as concentrate grade, metal recovery rate, unit concentrate production cost, equipment operating parameters, and process status within the next T time period, corresponding to product quality, economic, and equipment indicators; This is a simulation operator that integrates the entire mineral processing process, representing the serial coupling relationship between the sub-models of crushing, grinding, classification, flotation, concentration, and filtration. The input vector at time t represents the optimal control setpoints for each process output by the collaborative control model. Let t be the state vector of the entire process at time t, which includes subvectors of ore property parameters, equipment operating status parameters, process control parameters, product quality indicators, and economic operating indicators. The material transfer time delay vector is taken from the material flow time delay map and represents the material transfer delay constraint between each process. A fixed parameter set for the model, including static parameters such as equipment rated parameters, process safety thresholds, and model training weights; The preset simulation duration.

[0089] S6: When the simulation verification results show an anomaly, block the issuance of the optimal control setpoint for each process to prevent the abnormal setpoint from reaching the field PLC and actuator; and regenerate a new optimized control setpoint by time delay perception fusion attention timing model.

[0090] S7: When the simulation verification results show that everything is normal, closed-loop feedback control is completed based on the optimal control setpoints of each process to achieve coordinated and optimized operation of the entire mineral processing plant.

[0091] Closed-loop feedback control specifically involves sending the optimal control setpoint to the programmable logic controller (PLC) of each process via an industrial communication network. The PLC compares the optimal control setpoint with the real-time detection values ​​of each process, calculates the values ​​using a PID control algorithm, and outputs corresponding control commands to the frequency converter, electric actuator, pneumatic regulating valve, or reagent dosing pump to achieve precise adjustment of the process variables of each process.

[0092] Workflow:

[0093] The optimal control setpoints are sent to the programmable logic controllers (PLCs) of each process through the industrial communication network. Each process is equipped with one or more independent PLCs, which are responsible for local logic control and data acquisition within the scope of that process.

[0094] After receiving the optimal control setpoint from the host computer, the programmable logic controller (PLC) of each process compares it with the real-time detected value of the corresponding execution variable for that process and calculates the control deviation. ,in The control deviation at the current moment, The optimal control setpoint (given by the TLAFA™ model) This represents the current value of the process variable as measured by the field sensors.

[0095] PID control algorithm based on deviation Calculate the control output:

[0096]

[0097] in: Output for PID control algorithm (control signal, such as 420mA current signal or 010V voltage signal); The proportional gain determines the strength of the response to the current deviation. This is the integral gain, used to eliminate static errors; The differential gain is used to suppress overshoot and predict the trend of change in prediction bias.

[0098] Output the corresponding control commands to the frequency converter, electric actuator, pneumatic regulating valve or reagent dosing pump to complete the precise adjustment of process variables in each process.

[0099] This invention provides an AI-driven, fully automated collaborative control method for a mineral processing plant. By establishing a time-delay graph of the material flow throughout the entire process, the sampling data of each process is dynamically aligned according to the actual arrival time of the material. This effectively eliminates the problem of data misalignment caused by material transmission delays, ensuring that the constructed full-process state vector truly reflects the actual state of the same batch of material in each process. This provides an accurate state perception basis for subsequent control decisions. Furthermore, by constructing the full-process state vector and employing a time-delay perception fusion attention timing model to globally optimize and match the control parameters of each process, this invention overcomes the problem of relatively independent control systems for each process and the lack of a global collaborative optimization mechanism in traditional control methods. This achieves global optimization of the overall operating performance of the mineral processing plant. Simulation verification of the optimal control setpoints for each process effectively avoids production fluctuations caused by unreasonable control setpoints output by the time-delay perception fusion attention timing model, reducing the application risk of intelligent control systems.

[0100] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An artificial intelligence-driven method for fully automated collaborative control of a mineral processing plant, characterized in that, Includes the following steps: S1: Construct a material flow time delay map of the entire process of the mineral processing plant to quantify the material transfer delay time between each process from raw ore input, crushing, grinding, classification, flotation, concentration, filtration, tailings discharge to concentrate output; S2: By using online detection instruments and sensors distributed in each process, process data of each process is collected in real time. Based on the material flow time delay map, the process data of each process is dynamically aligned on the time axis to obtain the standard timestamp after the data of each process is aligned. S3: Construct a full-process state vector based on the standard timestamps after aligning the data of each process. S4: Construct a time-delay perception fusion attention time-series model, train the time-delay perception fusion attention time-series model, input the full-process state vector into the trained time-delay perception fusion attention time-series model for inference calculation, and output the optimal control setpoint for each process. S5: Perform simulation verification on the optimal control setpoints for each process to obtain the simulation verification results; S6: When the simulation verification result shows an anomaly, block the issuance of the optimal control setting value for each process, and regenerate a new optimized control setting value from the time delay perception fusion attention timing model. S7: When the simulation verification results show that everything is normal, closed-loop feedback control is completed based on the optimal control setpoints of each process to achieve coordinated and optimized operation of the entire mineral processing plant.

2. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 1, characterized in that, In S1, the material flow time delay map includes: material transfer path for each process, process node number, single process time delay duration, inter-process coupling time delay, and total process cumulative time delay, with the specific formula as follows: in: The total material transport time delay for the entire mineral processing process; This represents the total number of mineral processing steps. This is the process sequence number, with a value range of 1 ≤ ≤ ; For the first Material handling delays in each process step; This refers to the number of material transfer nodes between adjacent processes. For the first Material transfer interval time delay at each transfer node.

3. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 2, characterized in that, In S2, the dynamic time axis alignment of the process data for each process step specifically involves: relying on the inherent time delay parameters of each process step in the material flow time delay diagram, and using the quantified raw ore input time as the global time reference. Combined with the formula: Calculate the first Standard time coordinates corresponding to each process data The raw process data collected by various sensors and online monitoring instruments are time-corrected and offset-compensated according to the standard time coordinates of the corresponding processes. This eliminates data misalignment caused by material transport delays and achieves synchronous alignment of process data across the entire process on a unified time axis. For the first Standard timestamps after data alignment for each process step; This is the initial reference time when the raw ore enters the mineral processing system; For the first The processing time of each step is delayed.

4. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 1, characterized in that, In S3, the full-process state vector includes: ore property parameter sub-vector, equipment operating state parameter sub-vector, process control parameter sub-vector, product quality index sub-vector, and economic operation index sub-vector; The specific sub-vectors of ore property parameters are: raw ore grade, raw ore particle size distribution, raw ore moisture content, raw ore mudification index, ore grindability index, and ore floatability index. The specific sub-vectors of the equipment operating status parameters are: crusher main shaft current, crusher discharge port opening, mill cylinder vibration signal, mill ball loading, classifier overflow concentration, flotation machine impeller speed, flotation machine aeration volume, thickener rake frame torque, and filter vacuum degree. The process control parameter sub-vectors are specifically: crusher feed rate setting value, mill water feed rate setting value, mill feed rate setting value, classifier makeup water flow rate setting value, flotation reagent addition amount setting value, flotation liquid level setting value, thickener underflow pump frequency setting value, and filter belt speed setting value. The specific product quality index sub-vectors are: particle size of crushed products in each stage, fineness of grinding products, fineness of classification overflow, thickness of froth layer in each flotation cell, concentrate grade, metal recovery rate, tailings grade, concentrate moisture, and turbidity of return water. The specific sub-vectors of the economic operation indicators are: unit raw ore power consumption, unit concentrate steel ball consumption, unit concentrate reagent consumption, unit concentrate production cost, equipment comprehensive utilization rate, and concentrate output rate.

5. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 1, characterized in that, In S4, the time-delay-aware fusion attention temporal model specifically includes: Input preprocessing layer: used to perform outlier removal, normalization, and format conversion on the time-series state vector; Material delay graph feature extraction layer: Based on the material flow delay graph, a process topology directed graph is constructed, and a time-series graph convolutional network is used to extract the process spatial coupling features and material transport delay features; Multi-head temporal attention layer: Adaptively allocates temporal feature weights to highlight key operating conditions and abnormal features; Spatiotemporal fusion coding layer: A bidirectional long short-term memory network is used to fuse spatiotemporal features and capture the long-term evolution law of the mineral processing process; Multi-objective decision output layer: Combining multi-objective reinforcement learning strategies, it outputs the optimal control setpoints for each process.

6. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 5, characterized in that, In S4, the training of the time-delay perception fusion attention time-series model is based on historical production process data of the ore dressing plant. After dynamic alignment of the time axis and construction of the full-process state vector, a model training dataset is formed. The model training dataset takes maximizing the concentrate grade, maximizing the metal recovery rate, and minimizing the unit concentrate production cost as optimization objectives. A comprehensive fitness function is constructed as the basis for judging the performance of individual models. The specific formula is as follows: Where: F: Individual comprehensive fitness value, the larger the value, the better the control scheme; This represents the current concentrate grade. The maximum concentrate grade allowed by the process; This represents the current metal recovery rate; The maximum metal recovery rate allowed by the process; This represents the current unit cost of concentrate production. The lowest production cost allowed by the process; , , These are the weighting coefficients corresponding to concentrate grade, metal recovery rate, and production cost, respectively, satisfying... .

7. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 6, characterized in that, In S5, the simulation verification of the optimal control setpoints for each process is specifically performed as follows: The optimal control setpoints of each process output by the time-delay perception fusion attention timing model are input into the mineral processing digital twin model. The operating conditions of the entire process over a future period are deduced with a preset simulation step size. The operating parameters of each process equipment, product quality indicators and economic operation indicators are obtained through synchronous simulation. The simulation results are judged by combining the three evaluation indicators: concentrate grade, metal recovery rate, and unit concentrate production cost.

8. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 7, characterized in that, The specific formula for the mineral processing digital twin model is as follows: in: The simulation output vector of the digital twin model contains simulation results such as concentrate grade, metal recovery rate, unit concentrate production cost, equipment operating parameters, and process status within the next T time period, corresponding to product quality, economic, and equipment indicators; This is a simulation operator that integrates the entire mineral processing process, representing the serial coupling relationship between the sub-models of crushing, grinding, classification, flotation, concentration, and filtration. The input vector at time t represents the optimal control setpoints for each process output by the collaborative control model. The state vector of the entire process at time t includes the sub-vectors of the ore property parameters, the sub-vectors of the equipment operating status parameters, the sub-vectors of the process control parameters, the sub-vectors of the product quality indicators, and the sub-vectors of the economic operation indicators. The material transfer time delay vector is taken from the material flow time delay map and represents the material transfer delay constraint between each process. A fixed parameter set for the model, including static parameters such as equipment rated parameters, process safety thresholds, and model training weights; The preset simulation duration.

9. The artificial intelligence-driven fully automated collaborative control method for a mineral processing plant according to claim 1, characterized in that, In S7, the closed-loop feedback control specifically involves: sending the optimal control setpoint to the programmable logic controller (PLC) of each process through an industrial communication network; the PLC comparing the optimal control setpoint with the real-time detection values ​​of each process; calculating using a PID control algorithm; and outputting corresponding control commands to the frequency converter, electric actuator, pneumatic regulating valve, or reagent dosing pump to achieve precise adjustment of the process variables of each process.

Images