A method for generating spatiotemporal full-modal data of a mineral processing plant

Abstract

This invention discloses a method for generating spatiotemporal full-modal data for a mineral processing plant, relating to the field of digital twin technology. The method involves dividing the mineral processing process into crushing, grinding and classification, flotation, and thickening / dewatering sections, and configuring dedicated agents in each section. Each agent receives explicit monitoring data from its section and generates implicit data tensors. Adjacent sections construct a cross-section physical causal interlocking network through shared state tensors, jointly fitting upstream implicit data with local downstream explicit data. When numerical convergence fails, a physical causal inconsistency residual vector is extracted, triggering inverse correction by the upstream agent. For abnormal drift or fault monitoring data, a cross-modal prediction reconstruction operator is used to generate predicted true values ​​for replacement. Finally, through iterative game theory involving multiple agents throughout the entire process, full-modal data generation under a global steady-state equilibrium is achieved. This method improves the physical consistency, reliability, and robustness of the generated data for the entire mineral processing process.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method for generating spatiotemporal full-modal data of a mineral processing plant. Background Technology

[0002] With the continuous advancement of intelligent, digital, and unmanned mining construction, mineral processing plants are evolving from traditional manual experience-driven models to data-driven, model-driven, and autonomous collaborative control models. The mineral processing process typically includes multiple continuous stages such as crushing, grinding and classification, flotation, and thickening / dewatering. Significant material transfer relationships, physical causal relationships, and process coupling relationships exist between these stages. Due to frequent fluctuations in ore properties, complex equipment operating conditions, and diverse sources of operational disturbances, the entire mineral processing process exhibits strong nonlinearity, multivariate coupling, cross-stage correlation, and dynamic time-varying characteristics. Existing mineral processing data acquisition methods largely rely on distributed sensors, online instruments, and localized video monitoring. While these methods can acquire some explicit monitoring data, they still suffer from incomplete monitoring dimensions, difficulty in directly observing key implicit states, weak correlation between heterogeneous data, and isolated data across stages.

[0003] On the other hand, traditional data modeling methods typically focus on local analysis centered on a single work section, paying more attention to the mapping relationship between parameters and output indicators of that section. They lack constraints on the physical causal consistency between the input states of upstream sections and the actual observations of downstream sections, easily leading to a disconnect between model projections and real operating conditions. When monitoring data experiences drift, missing data, faults, or noise contamination, conventional methods often rely on posterior correction or simple interpolation compensation, making it difficult to maintain full-process physical consistency in complex process chains. Furthermore, existing multimodal fusion methods mostly focus on the surface fusion of image and sensor data, and have not yet formed a full-process spatiotemporal full-modal data generation mechanism that can combine mineral processing mechanism equations, cross-section state sharing, conflict detection, reverse correction, and abnormal data reconstruction. Summary of the Invention

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

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for generating spatiotemporal full-modal data of a mineral processing plant, which includes dividing the processing technology of the mineral processing plant into: crushing section, grinding and classification section, flotation section and thickening and dewatering section; In the crushing section, large pieces of raw ore are crushed to a specified particle size, and the feed particle size, discharge port width, and real-time power and image information of the crusher are collected simultaneously. In the grinding and classification section, the ore output from the crushing section is refined by the mill, and the balling rate, rotational speed, filling rate, hydrocyclone feed pressure and concentration are collected simultaneously. In the flotation section, the differences in the physicochemical properties of the mineral surface are used for separation, and the pH value of the slurry, aeration rate, slurry flow rate, reagent dosage and foam morphology characteristics are collected simultaneously. In the thickening and dewatering section, the concentrate from the flotation is subjected to solid-liquid separation, and the concentration of solid particles, mud interface height and underflow concentration are collected simultaneously. In each work section, a dedicated intelligent agent is configured to generate full-modal data for a single batch of materials sequentially; between intelligent agents in adjacent work sections, physical causal interlocking is achieved through sharing state tensors, so that the four intelligent agents in the whole process form a cross-work section physical causal interlocking network. During the continuous processing cycle of materials, adjacent intelligent agents execute a synchronization and co-decision protocol: The current work section's intelligent agent extracts the monitoring data it has collected, and uses the data shared by the intelligent agent of the previous work section as prior inference features to generate the current latent features; By using a deep interlocking equilibrium operator, the agents of the current work section and the previous work section are forced to synchronously search for the physical self-consistent interlocking point of the global state within the same calculation iteration cycle. Each dedicated intelligent agent extracts the equilibrium state parameters of the physically self-consistent interlocking point, combines them with the actual control data of this section, and drives the multimodal generation network within this section to achieve full-modal data generation.

[0006] As a preferred embodiment of the spatiotemporal full-modal data generation method for ore dressing plants described in this invention, the dedicated intelligent agent synchronously receives real-time sensor time-series monitoring data and spatial image monitoring data of this section. The temporal features of sensor data and the spatial features of image data are extracted through the feature embedding layer. A conditional attention mechanism is used to map all heterogeneous monitoring data features to a local feature space of a unified dimension, and then concatenates them to construct a global monitoring feature matrix, which serves as the current operating condition state vector. The current operating condition state vector is used as the dynamic condition evaluation benchmark to adaptively calculate the dynamic attention weight for each monitoring data stream. The global features, after being dynamically weighted and adaptively integrated, are input into a nonlinear mapping network specific to this work section, and the output is an implicit data tensor representing the physical process of the current work section.

[0007] As a preferred embodiment of the spatiotemporal full-modal data generation method for the mineral processing plant described in this invention, the step of achieving physical causal interlocking through shared state tensors includes: generating a globally unique discrete physical batch identifier for the currently processed single batch of materials; performing hard variable sharing in memory space, and directly defining the state tensor derived by the previous process agent for the discrete physical batch identifier as the prior deduction input data when the next process agent executes the synchronization and decision-making protocol in the physical memory address; The state tensor is a composite tensor obtained by orthogonally splicing an explicit data tensor that can be directly observed and a latent data tensor that is derived, after correlation filtering. It is used to represent the material state input to the next intelligent agent. Among them, correlation screening is based on the correlation between each data point and the next work section.

[0008] As a preferred embodiment of the spatiotemporal full-modal data generation method for ore dressing plants described in this invention, the deep interlocking equilibrium operator includes reading explicit monitoring data collected in this section as local observation constraints. The preset set of physical mechanism equations for mineral processing, including mass balance equations and energy conservation operators, is used to fit the mechanism to local observation constraints using the a priori deduction features. If the implicit data generated in the previous stage and the local explicit monitoring data cannot achieve numerical convergence in the physical mechanism equation set, then logical interference will occur on the feature inference path of the current agent; the energy conservation operator extracts the physical energy difference that cannot be closed in the inference process, and quantizes it into a high-dimensional physical causal violation residual vector through the variational mapping algorithm. Using the physical causal conflict residual vector as a trigger signal, the current agent's normal forward inference process is interrupted instantly, and the correction program of the previous agent is activated in reverse, entering the synchronous co-decision protocol.

[0009] As a preferred embodiment of the spatiotemporal full-modal data generation method for the ore dressing plant described in this invention, the preceding process agent, after receiving the reverse-transmitted physical causal conflict residual vector, initiates a reverse anchoring adversarial algorithm: The nonlinear mapping network of the previous stage updates the implicit data tensor input to the current agent by fine-tuning the network parameters, thereby reducing the magnitude of the inconsistency residual vector. At the same time, the agent of the previous stage extracts its own explicit data tensor as the truth anchor point and forces a judgment on whether the corrected implicit data deviates from its own physical reality. By iteratively playing a game between the physical authenticity of the previous stage agent and the magnitude of the inconsistency residual vector of the current agent, until an equilibrium point that minimizes the sum of the residuals of both sides is found, the physical self-consistent interlocking point is taken as the physical self-consistent interlocking point. By utilizing the network parameters of the previous work segment's agent when the physical self-consistent interlocking point is reached, the previous work segment is induced to output the corrected implicit data. At the same time, after sensing that the interlocking point has been reached, the current work segment's agent uses the corrected shared data and combines it with the local explicit data tensor to generate the current work segment's implicit data tensor.

[0010] As a preferred embodiment of the spatiotemporal full-modal data generation method for the mineral processing plant described in this invention, wherein: when the intelligent agent of the previous section outputs the corrected implicit data, the intelligent agent of the previous section uses the corrected implicit data tensor to perform positive physical reconstruction verification on the local explicit data tensor. If the overall loss of the reconstruction verification converges, but the error gradient is extremely concentrated on some explicit data and exceeds the preset boundary, then it is determined that the corresponding explicit data itself has hardware-level distortion or drift. After tracing the source of the distorted or drifting data, the confidence level of the explicit data collected by the abnormal sensor node is downgraded and the anomaly is corrected. After the anomaly is corrected, a hard rollback protocol is triggered to restore the feature inference network parameters of the previous process agent to the initial network state, and to re-derive the implicit data tensor and update the adversarial relationship with the current agent. The anomaly correction includes extracting the remaining valid explicit data other than the fault dimension, as well as the updated implicit data tensor, inputting them into the built-in cross-modal prediction reconstruction operator, generating the current predicted true value for the current batch of materials, and replacing the original distorted data of the fault node.

[0011] As a preferred embodiment of the spatiotemporal full-modal data generation method for ore dressing plants described in this invention, the cross-modal prediction and reconstruction operator includes the construction of two branches in the horizontal and vertical directions using a conditional mapping operator; Lateral cross-modal spatial causal inference branch: Through the conditional mapping operator, extract the remaining effective explicit data tensors except for the fault dimension at the current time, as well as the implicit data tensors after verification and update, as multi-dimensional prior conditions, and derive the spatial prediction truth value that conforms to the current transient process mechanism on the spatial physical manifold. The temporal fluctuation evolution branch of longitudinal historical inertia: extract the continuous historical sequence of the fault dimension within a preset time window before the abnormal drift occurs; separate and extract the temporal fluctuation characteristics that characterize mechanical oscillation and the local trend characteristics that characterize the equipment operating benchmark in the historical sequence, and derive the true value of the temporal evolution prediction that conforms to the physical operating inertia of the equipment. The first derivative of the remaining valid explicit data tensor is calculated in real time to assess the physical volatility of the current operating condition; dynamic gating coefficients are generated based on the physical volatility, and the two-branch prediction results are weighted and jointly decided to output the current true prediction value.

[0012] As a preferred embodiment of the spatiotemporal full-modal data generation method for ore dressing plants described in this invention, the method involves: using multiple dedicated intelligent agents configured throughout the entire process to perform upward correction and updates, and downward sharing of updated data, thereby achieving global iteration across work sections. Calculate the differential difference of the hidden data tensor output by each agent in two consecutive iterations; When the change in the correction results output by all agents in the entire process during the adversarial update continues to decay and eventually reaches the convergence criterion, the game of the cross-section physical causal interlocking network is determined to end, and a global steady-state equilibrium state is reached; and the parameters that reach the global steady-state equilibrium state are used as the equilibrium state parameters.

[0013] The beneficial effects of this invention are as follows: By configuring dedicated intelligent agents in each stage of crushing, grinding and classification, flotation, and thickening and dewatering, and constructing a cross-stage physical causal interlocking network using shared state tensors, this invention enables each stage to no longer be analyzed in isolation, but rather to be collaboratively generated under unified physical constraints, significantly improving the consistency and reliability of the entire process data. By introducing mass balance equations, energy conservation operators, and deep interlocking equilibrium mechanisms, it can promptly identify discrepancies and conflicts between upstream implicit deductions and downstream explicit observations, and reversely correct upstream implicit data, preventing erroneous information from propagating along the process. Furthermore, this invention can also perform cross-modal prediction and reconstruction of abnormal drift or faulty sensor data, achieving abnormal data replacement and operating condition recovery, improving the system's robustness to complex industrial environments. Compared to existing technologies, this invention combines the advantages of cross-stage collaboration, implicit state deduction, anomaly correction, full-modal generation, and global steady-state convergence, providing a highly reliable data foundation for intelligent monitoring, digital twins, and optimized control in mineral processing plants. Attached Figure Description

[0014] 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.

[0015] Figure 1 This is a flowchart of a method for generating spatiotemporal full-modal data for a mineral processing plant. Detailed Implementation

[0016] 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.

[0017] 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.

[0018] 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.

[0019] Reference Figure 1 As one embodiment of the present invention, this embodiment provides a method for generating spatiotemporal full-modal data of a mineral processing plant, comprising the following steps: The processing technology of the mineral processing plant is divided into: crushing section, grinding and classification section, flotation section and thickening and dewatering section.

[0020] In the crushing section, the main focus is on initial particle size crushing of large ore blocks entering the beneficiation process to obtain feed conditions that meet the requirements of subsequent grinding. To effectively characterize the processing status of this section, feed particle size, discharge port width, crusher real-time power, and real-time image information can be collected simultaneously. Feed particle size can be acquired online using industrial cameras, laser scanning devices, or image recognition devices installed at the feed belt or crusher inlet to characterize the ore block size distribution. Discharge port width can be obtained through displacement sensors, opening detection elements, or feedback values ​​from control actuators to reflect the current target discharge control status of the crusher. Crusher real-time power can be collected in real-time using current sensors, power transmitters, or drive control units to reflect the actual energy consumption and equipment load level during ore crushing. Real-time image information can be continuously collected by industrial cameras deployed above the feed or discharge ends to identify ore block size, shape, and edge features in real time, enabling state analysis of the entire process and allowing the intelligent agent to make more accurate predictions about the material state.

[0021] In the grinding and classification section, the ore output from the crushing section is further refined and separated into different particle sizes using mills and classification equipment to achieve the particle size and concentration conditions required for flotation. To reflect the processing mechanism and operating status of this section, data such as ball loading rate, rotational speed, filling rate, hydrocyclone feed pressure, and concentration can be collected simultaneously. The ball loading rate can be calculated or estimated based on ball loading records, ball bin weighing information, and mill volume model, characterizing the configuration of the grinding media. The rotational speed can be obtained through encoders, speed sensors, or frequency converter feedback, reflecting the mill's operating speed. The filling rate can be comprehensively calculated based on power, current, feed rate, and empirical models, characterizing the degree of filling of the mill with ore and media. The hydrocyclone feed pressure can be obtained online through pressure sensors installed on the feed pipeline, reflecting the classification driving force status. The hydrocyclone feed concentration can be obtained through online concentration meters, density meters, or a combination of flow and density calculations, reflecting the solid-liquid ratio of the slurry before classification. The above data can comprehensively characterize the energy input, material filling, fluid pressure and slurry concentration characteristics of the grinding and classification section, and serve as an important basis for inferring implicit states such as grinding efficiency, particle size evolution and classification effect.

[0022] In the flotation stage, the separation of valuable minerals from gangue minerals is mainly achieved by utilizing the differences in the physicochemical properties of mineral surfaces. To accurately characterize the reaction environment and separation state of this stage, the following parameters can be collected simultaneously: pulp pH, aeration rate, pulp flow rate, reagent dosage, and foam morphology. Specifically, pulp pH can be detected in real time using online pH electrodes and transmitters to characterize the chemical environment of the flotation system; aeration rate can be obtained through air flow meters, blower feedback, or gas mass flow rate detection devices to characterize bubble generation conditions; pulp flow rate can be measured online using electromagnetic or ultrasonic flow meters to reflect the pulp flow and residence state; reagent dosage can be obtained through metering pump feedback, valve opening, or dosing system control records to characterize the input levels of collectors, frothers, and modifiers; and foam morphology characteristics can be extracted by continuously acquiring images from an industrial camera positioned above the flotation cell, followed by image processing methods to extract parameters such as foam color, size, density, flow state, and breakage frequency. The above data can characterize the explicit operating features of the flotation section from multiple dimensions such as chemical conditions, fluid state, reagent action and visual appearance, and provide support for subsequent deduction of implicit data such as mineral floatability, liberation state and separation efficiency.

[0023] In the thickening and dewatering section, the concentrate obtained from flotation is mainly subjected to solid-liquid separation and slurry concentration to meet the requirements of subsequent filtration, transportation, or storage. To accurately reflect the settling and dewatering status of this section, solid particle concentration, mud interface height, and underflow concentration can be collected simultaneously. Solid particle concentration can be obtained through online concentration meters, densitometers, turbidity detectors, or automatic sampling and analysis methods to characterize the overall solid-liquid ratio of the slurry. Mud interface height can be measured online using ultrasonic mud level gauges, pressure interface detectors, or radar level devices to reflect the boundary between the settling layer and the clarifying layer within the thickener. Underflow concentration can be obtained through online concentration meters or density detectors installed on the underflow discharge pipeline to characterize the degree of concentration of the underflow discharge slurry. These data can be used to determine whether the thickener is operating stably, whether the settling process is normal, and whether the underflow meets the equipment requirements of the next section, and provide a basis for generating implicit data related to settling trends, dewatering efficiency, and equipment load.

[0024] S1: In each work section, a dedicated intelligent agent is configured to generate full-modal data for a single batch of materials sequentially.

[0025] Each dedicated intelligent agent is independently trained based on historical process data, equipment operation data, image monitoring data, and process tag data for its corresponding work section. Due to significant differences in the physical mechanisms, equipment types, monitoring parameters, and operating conditions of each work section, the training sets of each dedicated intelligent agent differ in data distribution, feature structure, and target tasks. Therefore, after training, each dedicated intelligent agent possesses the ability to identify and extrapolate the specific operating conditions of its work section.

[0026] In this embodiment, the data processed by the intelligent agents in each process section includes two categories: explicit data and implicit data. Explicit data refers to measurable data that can be directly acquired through sensors, online instruments, industrial cameras, control system feedback values, or manual detection devices; it represents direct observation results of equipment operation and material conditions. For example, in the crushing section, feed particle size, discharge port width, crusher real-time power, and real-time ore image information are explicit data; in the grinding and classification section, balling rate, rotational speed, filling rate, hydrocyclone feed pressure, and concentration are explicit data; in the flotation section, slurry pH, aeration rate, slurry flow rate, reagent dosage, and foam morphology characteristics are explicit data; and in the thickening and dewatering section, solid particle concentration, mud interface height, and underflow concentration are explicit data. The characteristic of this type of data is that it can be directly collected and recorded, and can intuitively reflect the surface conditions of the current process section. Because the explicit data types, implicit data labels, and process rules in the training sets of each work section are different, each dedicated intelligent agent can form a differentiated recognition capability for the material status and process characteristics of the corresponding work section.

[0027] Implicit data refers to data that is difficult to measure directly in real time using existing detection devices, but which objectively exists and has a significant impact on the processing status of the current stage or the performance of downstream stages. This type of data typically requires inference based on explicit data, process mechanisms, and causal relationships across stages. For example, the true hardness distribution of ore, the mass distribution of particles of different sizes, and the degree of particle fragmentation in the crushing stage can be considered implicit data; the grindability of ore, particle liberation degree, and critical particle size distribution in the grinding and classification stage can be considered implicit data; the surface floatability of minerals, bubble carrying stability, reagent effectiveness, and particle adhesion probability in the flotation stage can be considered implicit data; and particle settling trends, flocculation state, underflow compression degree, and solid-liquid separation efficiency in the thickening and dewatering stage can also be considered implicit data. Although this type of data cannot be directly observed in real time, it can reflect the internal physical properties of materials, process evolution trends, and causal relationships between multiple stages at a deeper level. Therefore, it is an important object for achieving full-modal data generation and cross-stage physical self-consistent inference.

[0028] The dedicated intelligent agent synchronously receives real-time sensor time-series monitoring data and spatial image monitoring data for this work section. The sensor time-series monitoring data reflects the continuous fluctuations of equipment operating parameters, process variables, and material states over time, characterizing the dynamic changes of the current work section. The spatial image monitoring data visually reflects spatial structural features such as material morphology, particle distribution, foam state, settling interfaces, or on-site accumulation conditions, characterizing the apparent morphological information of the current work section. By synchronously receiving these two types of data, the dedicated intelligent agent can simultaneously grasp both "numerical changes" and "spatial manifestations," thereby more completely identifying the actual working conditions of the work section, improving the accuracy of inferring latent states, and providing a more sufficient input basis for subsequent full-modal data generation.

[0029] The feature embedding layer extracts temporal features from sensor data and spatial features from image data. This layer, a feature encoding module positioned between the raw input data and the subsequent analysis network, maps different types of raw monitoring data into a unified high- or low-dimensional feature representation, facilitating subsequent feature fusion, state deduction, and data generation.

[0030] A conditional attention mechanism is used to map all heterogeneous monitoring data features to a local feature space of a unified dimension, and then concatenate them to construct a global monitoring feature matrix, which serves as the current operating condition state vector. The current operating condition state vector is used as the dynamic condition evaluation benchmark to adaptively calculate the dynamic attention weight for each monitoring data stream.

[0031] It's worth noting that in the field of multimodal data fusion, attention mechanisms (such as standard self-attention) typically rely on dot product calculations within each input feature to obtain correlations. This approach mainly focuses on local similarity or static mapping relationships between feature streams, lacking a global "condition awareness" perspective. In the mineral processing process, the operating state of the equipment is highly dynamic and nonlinear. If a traditional attention mechanism is used, the network can only perform mechanical feature fusion and cannot perceive "what state the factory is in at this moment," thus failing to dynamically adjust the focus of attention on different sensor or image features according to drastic changes in operating conditions. In this embodiment, the traditional mechanism is upgraded, with the core being the introduction of a global perspective of "conditional state." Instead of directly allowing various heterogeneous data (sensor time series and image spatial features) to blindly interact internally, it first uses feature embedding and local mapping to stitch all the data together to construct a complete "global monitoring feature matrix." This matrix is ​​creatively defined as the "current operating condition state vector." It is not merely a collection of data, but represents the macroscopic physical state benchmark (i.e., the "current macro environment") of the entire section's equipment operation within this small time slice. After establishing the "current working condition state vector," the algorithm uses it as the benchmark for dynamic condition evaluation (equivalent to a dynamic global query) to re-examine each independent data stream. Since any fluctuation in the mineral processing conditions will cause a shift in the distribution of monitoring data in the multi-dimensional feature space, this mechanism makes the attention operator no longer static, but capable of sensitively following changes in the distribution of data in different dimensions. For example, when a sudden change in the working condition causes a drastic change in the distribution of foam morphology in the image modality, while the data distribution in other dimensions remains stable, the system, using the condition state vector as a reference, can instantly detect the abnormal distribution in the image dimension and adaptively allocate attention weights towards that dimension. Through this improved architecture of "global macro-state guiding local micro-feature weighting," the system achieves global features after dynamic weight integration, which can most realistically and sensitively reflect the physical evolution of the current section, laying a solid algorithmic foundation for deriving high-confidence implicit data tensors in nonlinear mapping networks.

[0032] The global features, after being dynamically weighted and adaptively integrated, are input into a nonlinear mapping network specific to this work section, and the output is an implicit data tensor representing the physical process of the current work section.

[0033] In this embodiment, the nonlinear mapping network is preferably a multilayer feedforward neural network. This multilayer feedforward neural network includes an input layer, several hidden layers, and an output layer. The input layer receives the global features after dynamic weight adaptive integration. The hidden layers transform the global features layer by layer using a fully connected weight matrix and a nonlinear activation function to extract higher-order coupling relationships from the explicit monitoring data. The output layer outputs a latent data tensor representing the physical process of the current work section. In this embodiment, the nonlinear activation function can be any one or a combination of ReLU, Sigmoid, or Tanh functions. The advantages of using a multilayer feedforward neural network are its clear structure, simple implementation, and ease of direct interface with the global features output by the aforementioned conditional attention mechanism. It is suitable for uniformly mapping the overall state after the fusion of multiple heterogeneous monitoring data within the current work section.

[0034] S2: Physical causal interlocking is achieved between intelligent agents in adjacent work sections by sharing state tensors, so that the four intelligent agents in the whole process form a cross-work section physical causal interlocking network.

[0035] Furthermore, the physical causal interlocking achieved by sharing the state tensor includes: generating a globally unique discrete physical batch identifier for the single batch of materials currently being processed; performing hard variable sharing in the memory space, and directly defining the state tensor derived by the previous process agent for the discrete physical batch identifier as the prior deduction input data when the next process agent executes the synchronization and decision-making protocol in the physical memory address; ; The identifier representing the physical memory storage address of the variable; This indicates that the addresses are equal; This represents the input for the priori deduction of the next work section; This represents the state tensor (a composite tensor containing explicit and implicit data) generated and shared by the previous work section. This represents the zero-time-difference spatial feature projection operator matrix.

[0036] The state tensor is a composite tensor obtained by orthogonally concatenating directly observable explicit data tensors and derived implicit data tensors after correlation filtering. It is used to represent the material state input to the next agent. The correlation filtering is based on the correlation between each data type and the next processing stage.

[0037] In continuous heavy industrial production lines, upstream data often has extremely high dimensionality, but only a small fraction of variables truly provide physical causal guidance for downstream processes. Transmitting all explicit and implicit data indiscriminately downstream would not only lead to the curse of dimensionality but also introduce significant computational noise. Therefore, the "relevance filtering" in this embodiment achieves "refinement and elimination" by constructing a dynamic routing gateway driven by both mutual information entropy and dynamic gradient sensitivity. This mechanism first performs a static benchmark evaluation based on physical history priors. The system extracts massive amounts of historical batch data and calculates the mutual information entropy between each feature dimension in the current process's output tensor and key process indicators (such as recovery rate and concentrate grade) of the next process. Mutual information can accurately capture highly nonlinear physical correlations, and the system generates a static physical correlation benchmark vector accordingly, directly eliminating marginal hardware noise data lacking causal correlation from the underlying physical logic.

[0038] After establishing a static baseline, a real-time dynamic gradient sensitivity detection mechanism based on the current operating conditions is introduced. Since even highly correlated physical quantities can dynamically shift in importance under different transient operating conditions, the downstream AI sends a detection gradient to the upstream system before each iteration. The system calculates in real-time the absolute values ​​of the partial derivatives of the downstream nonlinear mapping network's output with respect to each input feature dimension of the current operating condition. This derivative value characterizes the feature's "sensitivity," i.e., the degree of influence of a small feature perturbation on the downstream simulation results, thereby accurately assessing the real-time importance of the feature under the current specific operating condition and generating a dynamic sensitivity vector accordingly.

[0039] Finally, the system mathematically fuses and orthogonally filters the aforementioned dual evaluation mechanisms. The algorithm weights and fuses static mutual information entropy and dynamic gradient sensitivity to calculate a comprehensive relevance score for each feature dimension. This score is then input into activation functions such as Sigmoid to generate an adaptive feature mask. This mask is then used to perform a Hadamard product (element-wise multiplication) with the original massive composite state tensor, directly setting dimensions with relevance scores below the blocking threshold to zero. Through this filtering process, the system successfully intercepts invalid mechanistic noise, ensuring that only core features with extremely high information density and strong causal correlation are retained for orthogonal concatenation and passed as an extremely pure prior input tensor to the dedicated agent in the next stage.

[0040] Comprehensive relevance score calculation (causal value assessment): ; Adaptive feature mask generation (non-linear gated filtering): ; Feature orthogonal filtering and tensor simplification (Hadamard product execution): ; Indicates the first In the current processing batch, the tensor output by the current section is the first... A comprehensive relevance score based on each feature dimension. This represents the original feature dimension to be evaluated (covering explicit physical quantities or derived implicit states). This represents the set of key process indicators for the next stage (such as target vectors for recovery rate, concentrate grade, etc.). Representation of features With downstream targets The mutual information entropy between them is used to measure the strength of the statistical physical association between them (static benchmark). This indicates that the next work segment's intelligent agent is in the previous iteration cycle. The mapping function (neural network model). These represent preset weighting coefficients, used to balance the proportions of static physical mechanism correlation and dynamic operating condition sensitivity.

[0041] This indicates that the downstream agent has a positive view of the features. The real-time partial derivative gradient represents the sensitivity of this feature to downstream inference results (dynamic benchmark).

[0042] This represents the generated adaptive feature mask vector, whose elements take values ​​in the range of... between; This represents the sigmoid activation function, used to map linear scores to probabilistic gating weights. The threshold for blocking correlation is the physical red line that the system uses to determine whether a feature is retained. This represents the gain operator (temperature parameter) used to adjust the "hardness" of the filter. The larger the value, the closer the selection is to a "0-1" hard truncation; The smaller the value, the smoother the filtering process. This indicates that the score is based on all dimensions. The resulting comprehensive relevance score vector. It represents the massive composite state tensor of the original input, containing the initial dimensions of all acquisitions and derivations. This represents the Hadamard Product, which is the element-wise multiplication operation between corresponding tensors. This represents the final simplified state tensor. The noise dimension, which is irrelevant to the downstream, has been set to zero, retaining only the highly correlated physical causal features.

[0043] Furthermore, the cross-segment physical causal interlocking network includes: a zero-time-difference spatial feature projection operator established between dedicated agents in adjacent segments; after the feature tensor undergoes dimensional transformation through the zero-time-difference spatial feature projection operator, its mapping result and the prior inference input data of the next segment agent point to the same physical memory address, thereby using the uniqueness of the memory address to force causal interlocking of cross-segment states.

[0044] It should be noted that the zero-time-difference spatial feature projection operator matrix employs a generalized linear minimum mean square error spatial projection algorithm based on process mechanism topological constraints. The specific model construction and calculation methods are divided into the following four core stages: The first stage involves constructing the augmented state-space model. The algorithm uses absolute physical time slices as a benchmark, aligning and orthogonally concatenating the manifolds of all explicit sensor monitoring data in the same transient state (all in a healthy state) with the implicit physical states (such as ore liberation degree) obtained through cross-segment agent game convergence. This constructs an augmented feature vector representing the current transient global macroscopic working condition. This model unifies the previously scattered surface data and deep mechanism data onto the same high-dimensional feature space.

[0045] The second stage involves calculating the expected value of the underlying statistical covariance. Based on a massive batch dataset of historical steady-state operation, the system uses mathematical expectation calculation to measure the autocovariance matrix of the augmented eigenvectors and the cross-covariance vector between the augmented eigenvectors and the actual observations of the target fault sensor. This calculation method aims to extract data-driven linkage patterns and supporting weights between variables in high-dimensional space using big data.

[0046] The third stage involves introducing a physical prior regularization penalty algorithm. To overcome the "physical spurious correlation" defect easily caused by pure statistical fitting, the algorithm incorporates a structured "physical mechanism penalty bias matrix" into the traditional covariance inversion process (similar to Tikhonov regularization or ridge regression mechanisms). This penalty matrix is ​​obtained using an adaptive instantiation mechanism of "static prior skeleton superimposed with dynamic operating condition mask": First, the system deeply analyzes the piping and instrumentation diagram (P&ID) and mass / energy conservation partial differential equations of the concentrator to construct a static topological baseline matrix, assigning basic penalty resistance values ​​to variable pairs with absolutely no physical exchange; then, at time t during the system's real-time closed-loop simulation, discrete switching signals such as equipment start-up / shutdown and valve opening are synchronously collected to generate a dynamic operating condition mask matrix that accurately reflects transient physical connectivity. The system combines the two using the Hadamard product operator to adaptively instantiate the physical mechanism penalty bias matrix specific to time t within milliseconds. The algorithm superimposes the bias matrix as an algebraic resistance onto specific diagonal or off-diagonal elements of the autocovariance matrix. This not only transiently increases the penalty weight of the corresponding variable pair to block its mathematical correlation when the physical path (such as a bypass valve) is forcibly cut off, but also forcibly changes the eigenvalue distribution of the original covariance matrix, forcibly distorting the pure mathematical projection space and confining it to the manifold hyperplane allowed by physical laws.

[0047] The fourth stage involves the analytical solution of the operator matrix and zero-time-difference algebraic operations. The algorithm calculates the generalized inverse of the autocovariance matrix superimposed with physical penalty terms and multiplies it with the cross-covariance vector to obtain a spatial projection operator matrix with an absolutely closed-form solution. In practical closed-loop control, once a sensor transient jam or drift occurs, the system's real-time calculation method degenerates into a simplified vector-matrix dot product operation: the system instantly extracts the remaining effective augmented eigenvectors at the moment of the fault and directly multiplies them by the aforementioned fixed operator matrix. This calculation process does not involve time-step recursive solutions, eliminates the time-delay inertia of historical data flow, and achieves absolutely zero-delay transient reconstruction output of the failure dimension data at the pure algebraic operation level.

[0048] S3: During the continuous processing cycle of materials, adjacent intelligent agents execute a synchronization and co-decision protocol: The current work segment's intelligent agent extracts the monitoring data it collects, and uses the data shared by the previous work segment's intelligent agent as prior inference features to generate the current latent features.

[0049] Upon receiving locally collected explicit objective monitoring data in real time, as well as shared prior data from the previous work section (including objective phenomena and implicit derivations from the previous work section), the current work section's agent does not immediately output the final result. Instead, the system internally constructs an extremely rigorous virtual simulation sandbox. The agent simultaneously injects the two data streams into the work section's specific "physical mechanism equation set" (covering the work section's unique mass balance, energy conservation, and fluid dynamics constraints), initiating a round of tentative simulation generation and derivation. During the simulation generation and derivation process, the system uses the physical mechanism equation set as an "absolute judge" to strongly verify the causal logic of the data. At this time, the derivation process often encounters great "physical resistance." This resistance is mathematically manifested as "numerical non-convergence." Essentially, it lies in the absolute conflict between the implicit derivations claimed by the previous work section (such as an extremely ideal degree of ore liberation) and the actual explicit physical reactions of the current work section (such as abnormally high instantaneous mill energy consumption or mismatched slurry concentration) in terms of energy or mass. This generates a "physical mechanism inconsistency residual" that characterizes causal conflicts. If this residual exceeds a preset tolerance threshold, the system determines that the current simulation has encountered resistance, refuses to generate final data, and uses this residual to trigger reverse game theory and parameter correction between upstream and downstream agents. The system continuously eliminates conflict resistance during the simulation process through multiple rounds of local parameter fine-tuning and cross-segment reverse correction. Only when the implicit prior data corrected in the previous segment and the local explicit monitoring data achieve absolute numerical convergence in the current physical mechanism equations (i.e., the physical causal inconsistency residual decays to near zero) does the system determine that the virtual simulation has entered a "global physical steady state." At this point, the system removes the simulation sandbox's interception mechanism and performs final dimensionality reduction decoding on the converged composite data stream through a nonlinear mapping network. Only on this steady-state benchmark, which fully conforms to objective physical laws, does the system formally "finalize" and generate the true implicit data characteristics of the current segment.

[0050] By using a deep interlocking balance operator, the agents of the current work section and the previous work section are forced to synchronously search for the physical self-consistent interlocking point of the global state within the same computational iteration cycle.

[0051] Furthermore, the deep interlocking equilibrium operator includes reading explicit monitoring data collected in this section as local observation constraints. A preset set of mineral processing physical mechanism equations, including mass balance equations and energy conservation operators, is used to fit the local observation constraints using the prior deduction features.

[0052] The mass balance equation characterizes a conservation law model where the input, output, and accumulated amounts of material in a continuous processing flow must be absolutely equal. Between upstream and downstream stages in mineral processing (such as grinding and flotation), this equation defines the macroscopic flow logic of materials. As the physical basis for energy conservation operator calculations, the mass balance equation is considered a prerequisite for analysis.

[0053] The energy conservation operator is a computational execution unit used to verify the energy conversion equivalence between external work done on a equipment system and internal changes in the physical / chemical state of materials. During ore crushing or grinding, the reduction in ore particle size (generating new surface area) must consume corresponding crushing work (e.g., an energy model based on the Bond work exponent). This operator is responsible for extracting implicit derivation features shared by the previous stage (e.g., claimed ore liberation degree / particle size distribution) and performing energy equivalence conversion calculations with explicit observation constraints actually collected in this stage (e.g., the actual instantaneous power consumption of the mill motor) to verify whether the mechanical principles between the data match.

[0054] Mechanism fitting refers to substituting the prior deduced characteristics from the previous process stage with the explicit monitoring data collected in the current process stage into a pre-set set of physical mechanism equations for mineral processing, and solving whether it can form a physical explanation process consistent with the actual working conditions of the current process stage. In practice, it involves analyzing whether the physical mechanism equations converge, applying the "upstream states of the material's condition" and the "downstream's actual current state" to the physical formulas to see if the intermediate processing stage is physically reasonable.

[0055] If the implicit data generated in the previous stage and the local explicit monitoring data cannot achieve numerical convergence in the physical mechanism equations, logical interference occurs on the feature deduction path of the current agent. The energy conservation operator extracts the physical energy difference that fails to close during the deduction process and quantizes it into a high-dimensional physical causal conflict residual vector using a variational mapping algorithm. This physical causal conflict residual vector is used as a trigger signal to instantly interrupt the current agent's normal forward deduction process and reversibly activate the correction program of the previous stage agent, entering the synchronous co-decision protocol.

[0056] The physical energy difference refers to the energy non-equilibrium remainder term formed during the fitting of the mechanistic equations to the current work section. Essentially, it reflects whether there is a physical logic gap between the "material state deduced from the implicit data of the previous work section" and the "actual energy consumption of the equipment in the current work section." Specifically, the energy requirement of the current batch of materials in this work section is theoretically obtained using the energy conservation operator, serving as the theoretical energy value. Then, the actual power, current, or inverter feedback value during equipment operation is extracted from the explicit monitoring data of this work section to obtain the actual observed energy value. Subsequently, the absolute deviation between the two is calculated using a differential comparison method to obtain the physical energy difference. This difference is essentially a scalar used to quantify the degree of discrepancy between the current deduced path and the actual working condition. When this difference exceeds a preset threshold, it indicates that the implicit state generated by the prior deduced features cannot be supported by the actual energy consumption of the current work section, thus demonstrating that the current implicit features do not conform to physical logic. The physical causal discrepancy residual vector is a high-dimensional control signal further formed based on the aforementioned physical energy difference. Its function is to transform the scalar discrepancy information, which originally only represented the "magnitude of the deviation," into a multi-dimensional residual expression that simultaneously characterizes the "direction of the deviation" and the "correction position." In specific implementation, the physical energy difference can be used as the loss target. Through variational mapping algorithms or gradient-based chain derivative methods, the partial derivatives of this energy difference with respect to each dimension of the latent feature vector of the previous process are calculated, thereby generating the corresponding physical causal discrepancy residual vector. Each component in this residual vector corresponds to the correction direction and correction intensity in a certain latent feature dimension. Therefore, it not only indicates the existence of a physical discrepancy in the system but also further indicates which type of latent state deviation the discrepancy mainly originates from. After the vector is generated, it is fed back to the agent of the previous process as a direct basis for subsequent parameter correction and state correction, enabling the previous process to update its model in the next calculation cycle along the direction of reducing the physical energy difference until global physical self-consistency is restored between the upstream and downstream processes.

[0057] After receiving the reverse-transmitted physical causal conflict residual vector, the intelligent agent of the previous section initiates the reverse anchoring adversarial algorithm: The nonlinear mapping network of the previous stage updates the implicit data tensor input to the current agent by fine-tuning the network parameters, thereby reducing the magnitude of the inconsistency residual vector. At the same time, the agent of the previous stage extracts its own explicit data tensor as the truth anchor point and forces a judgment on whether the corrected implicit data deviates from its own physical reality.

[0058] Furthermore, the calculation of physical authenticity can be described as follows: the agent from the previous work section reconstructs and verifies the consistency of the corrected implicit data tensor based on its own explicit data tensor. Specifically, after parameter fine-tuning, the nonlinear mapping network from the previous work section regenerates or reverse-engineers the corresponding implicit state representation using the explicit data of this work section, and compares the consistency of the re-analyzed implicit result with the currently corrected implicit data to calculate the authenticity deviation between the two. The smaller the deviation, the closer the corrected implicit data is to the actual physical state of the previous work section; the larger the deviation, the more the correction has deviated from the constraints of the actual explicit observations of the previous work section.

[0059] By iteratively playing a game between the physical authenticity of the agent in the previous work segment and the magnitude of the discrepancy residual vector of the current agent, until an equilibrium point that minimizes the sum of the residuals of both sides is found, this equilibrium point is designated as the physically self-consistent interlocking point. In this embodiment, the calculation process of the equilibrium point includes: taking the physical authenticity deviation of the previous work segment as a first cost term and the magnitude of the discrepancy residual of the current work segment as a second cost term, assigning corresponding weights to each, and then performing a weighted summation to obtain the total objective value; in each iteration, updating the parameters of the nonlinear mapping network of the previous work segment according to the direction of change of the total objective value, until the total objective value converges to a minimum. The corresponding implicit data correction result at this time is the physically self-consistent interlocking state corresponding to the equilibrium point.

[0060] By utilizing the network parameters of the previous work segment's agent when the physical self-consistent interlocking point is reached, the previous work segment is induced to output the corrected implicit data. At the same time, after sensing that the interlocking point has been reached, the current work segment's agent uses the corrected shared data and combines it with the local explicit data tensor to generate the current work segment's implicit data tensor.

[0061] Furthermore, when the previous stage agent outputs corrected implicit data, it uses the corrected implicit data tensor to perform a forward physical reconstruction check on the local explicit data tensor. If the overall loss of the reconstruction check converges, but the error gradient is extremely concentrated on a portion of the explicit data and exceeds a preset boundary, then the corresponding explicit data is determined to have hardware-level distortion or drift. After tracing the source of the distorted or drifted data, the confidence level of the explicit data collected by the abnormal sensor node is downgraded and anomaly corrected. After the anomaly correction, a hard rollback protocol is triggered to restore the feature inference network parameters of the previous stage agent to the initial network state, and the implicit data tensor is re-derived and the adversarial update with the current agent is performed.

[0062] It should be noted that real-time reconstruction verification is achieved by constructing an inverse reconstruction network that is a mirror image of the forward mapping network, combined with a physical consistency discriminator. Taking the generated implicit data tensor as input, it is projected back into the explicit physical space through transposed convolution or fully connected mapping layers. Then, by minimizing the reconstruction error loss function (which uses the explicit monitoring data collected in real-time in this section as the truth anchor), the authenticity and self-consistency of the implicit features in the physical generation logic are forcibly verified.

[0063] The anomaly correction includes extracting the remaining valid explicit data other than the fault dimension, as well as the updated implicit data tensor, inputting them into the built-in cross-modal prediction reconstruction operator, generating the current predicted true value for the current batch of materials, and replacing the original distorted data of the fault node.

[0064] The cross-modal prediction and reconstruction operator includes two branches constructed horizontally and vertically using a conditional mapping operator. The horizontal cross-modal spatial causal inference branch: using the conditional mapping operator, it extracts the remaining valid explicit data tensors (excluding the fault dimension) at the current moment, along with the implicit data tensors after verification and updates, as multi-dimensional prior conditions. It then derives the spatial prediction truth value conforming to the current transient process mechanism on the spatial physical manifold. The vertical historical inertia temporal fluctuation evolution branch: it extracts the continuous historical sequence of the fault dimension within a preset time window before the abnormal drift occurs; it separates and extracts the temporal fluctuation characteristics representing mechanical oscillations and the local trend characteristics representing the equipment operating baseline from the historical sequence, deriving the temporal evolution prediction truth value conforming to the physical operating inertia of the equipment.

[0065] In this embodiment, the conditional mapping operator is preferably a conditional coding neural network. This conditional coding neural network receives multi-source prior conditions and completes the mapping and reconstruction of target features under conditional constraints. Specifically, in the lateral cross-modal spatial causal inference branch, the conditional mapping operator can employ a conditional mapping network combining a multi-layer feedforward neural network and an attention mechanism. Its input includes the remaining valid explicit data tensors (excluding the fault dimension) at the current time, and the implicit data tensors after verification and updates. By weighted encoding and nonlinear mapping of conditional features from different sources, the spatial coupling relationship between each explicit variable and the implicit state is extracted, thereby outputting the spatial prediction truth value that conforms to the current transient process mechanism on the spatial physical manifold. The design purpose of this structure is to enable the network to learn the cross-modal spatial causal relationship between the fault dimension and other variables under the joint constraints of "remaining valid observation data" and "implicit state priors".

[0066] In the longitudinal historical inertia temporal fluctuation evolution branch, the conditional mapping operator is preferably any one of a recurrent neural network, a long short-term memory neural network, or a gated recurrent unit neural network, used to process the continuous historical sequence of the fault dimension before the occurrence of the anomaly. This network can extract short-term fluctuation features and local trend features from the historical sequence under conditional constraints, and further derive the true value of the temporal evolution prediction that conforms to the physical operating inertia of the equipment. That is, in this embodiment, the conditional mapping operator is not a single fixed structure, but rather adopts a neural network structure more suitable for the data characteristics of the two branches, horizontal and vertical: the horizontal branch emphasizes multimodal conditional fusion and spatial correlation modeling, while the vertical branch emphasizes time dependence and inertial evolution law modeling.

[0067] The first derivative of the remaining valid explicit data tensor is calculated in real time to assess the physical volatility of the current mineral processing condition. Based on the physical volatility, dynamic gating coefficients are generated: when the condition is stable, the time-series prediction truth value is given a higher fusion weight to maintain the smooth inertia of the control command; when the condition changes abruptly, the spatial prediction truth value is given a higher fusion weight to keep up with the physical causal changes; the two-branch prediction results are weighted and jointly decided by the dynamic gating coefficients to output the smooth and absolutely physically consistent current prediction truth value.

[0068] S4: Each dedicated intelligent agent extracts the equilibrium state parameters of the physically self-consistent interlocking point, combines them with the actual control data of this section, and drives the multimodal generation network in this section to achieve full-modal data generation.

[0069] Specifically, multiple dedicated intelligent agents configured throughout the entire process perform upward correction and updates, and downward sharing of updated data, achieving global iteration across work sections. The differential difference value of the implicit data tensor output by each intelligent agent during two consecutive iterations is calculated.

[0070] When the change in the correction results output by all agents in the entire process during the adversarial update continues to decay and eventually reaches the convergence criterion, the game of the cross-section physical causal interlocking network is determined to end, and a global steady-state equilibrium state is reached; and the parameters that reach the global steady-state equilibrium state are used as the equilibrium state parameters.

[0071] It's important to note that due to the significant cross-segment transitivity and causal coupling in the mineral processing process, the state of the previous segment directly affects the performance of the next segment, while the actual observations of the next segment can reveal whether there are any deviations in the projections of the previous segment. Therefore, relying solely on the output of a single segment can easily lead to hidden state distortion, broken causal chains, and inconsistencies between generated data and actual operating conditions. By setting up multiple dedicated intelligent agents to correct and update upwards and to share updated data downwards, a continuous iterative collaborative correction mechanism can be formed throughout the entire process, thereby gradually eliminating physical causal conflicts between different segments.

[0072] When the changes in the correction results output by each agent continuously decay and eventually reach the convergence criterion, it indicates that the causal conflicts between upstream and downstream processes have been fully coordinated, and the implicit states and explicit observations of each process have achieved stable consistency across the entire process. At this point, the game between the cross-process physical causal interlocking network can be considered over, and the system can be considered to have entered a global steady-state equilibrium state. The parameters in this state are used as equilibrium state parameters to ensure that the data basis for subsequently driving the multimodal generation network is not a local temporary correction result, but a stable parameter that simultaneously satisfies the authenticity of the current process, cross-process consistency, and overall physical self-consistency, thereby improving the reliability, continuity, and engineering usability of the final full-modal data generation results.

[0073] In summary, this invention divides the entire mineral processing plant into crushing, grinding and classification, flotation, and thickening / dewatering sections, and configures dedicated intelligent agents in each section. These agents then collaboratively extrapolate based on explicit monitoring data from their respective sections and shared state tensors from the previous section. Furthermore, it achieves joint modeling of explicit and implicit data through conditional attention mechanisms and nonlinear mapping networks. A cross-section physical causal interlocking network is constructed using shared state tensors. Finally, a deep interlocking equilibrium operator synchronously verifies the causal consistency between upstream and downstream sections under the joint constraints of the mass balance equation and energy conservation operator. When a physical discrepancy is found between the implicit inference results of the previous process and the explicit observations of the current process, the physical causal discrepancy residual vector is extracted and reverse correction is triggered, so that the previous process corrects the implicit data while maintaining its own physical authenticity. At the same time, when drift or distortion is detected in the explicit monitoring data, the cross-modal prediction and reconstruction operator is further used to replace and correct the abnormal data. Finally, through continuous iterative game between the multi-process agents, the entire process reaches a global steady-state equilibrium state, and the parameters of this equilibrium state drive the multi-modal generation network of each process to achieve full-modal data generation consistent with the actual process.

[0074] 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 method for generating spatiotemporal full-modal data of a mineral processing plant, characterized in that: This includes dividing the processing technology of the ore dressing plant into: crushing section, grinding and classification section, flotation section and thickening and dewatering section; In the crushing section, large pieces of raw ore are crushed to a specified particle size, and the feed particle size, discharge port width, and real-time power and image information of the crusher are collected simultaneously. In the grinding and classification section, the ore output from the crushing section is refined by the mill, and the balling rate, rotational speed, filling rate, hydrocyclone feed pressure and concentration are collected simultaneously. In the flotation section, the differences in the physicochemical properties of the mineral surface are used for separation, and the pH value of the slurry, aeration rate, slurry flow rate, reagent dosage and foam morphology characteristics are collected simultaneously. In the thickening and dewatering section, the concentrate from the flotation is subjected to solid-liquid separation, and the concentration of solid particles, mud interface height and underflow concentration are collected simultaneously. In each work section, a dedicated intelligent agent is configured to generate full-modal data for a single batch of materials sequentially; between intelligent agents in adjacent work sections, physical causal interlocking is achieved through sharing state tensors, so that the four intelligent agents in the whole process form a cross-work section physical causal interlocking network. During the continuous processing cycle of materials, adjacent intelligent agents execute a synchronization and co-decision protocol: The current work section's intelligent agent extracts the monitoring data it has collected, and uses the data shared by the intelligent agent of the previous work section as prior inference features to generate the current latent features; By using a deep interlocking equilibrium operator, the agents of the current work section and the previous work section are forced to synchronously search for the physical self-consistent interlocking point of the global state within the same calculation iteration cycle. Each dedicated intelligent agent extracts the equilibrium state parameters of the physically self-consistent interlocking point, combines them with the actual control data of this section, and drives the multimodal generation network within this section to achieve full-modal data generation.

2. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 1, characterized in that: The dedicated intelligent agent synchronously receives real-time sensor time-series monitoring data and spatial image monitoring data of this section; The temporal features of sensor data and the spatial features of image data are extracted through the feature embedding layer. A conditional attention mechanism is used to map all heterogeneous monitoring data features to a local feature space of a unified dimension, and then concatenates them to construct a global monitoring feature matrix, which serves as the current operating condition state vector. The current operating condition state vector is used as the dynamic condition evaluation benchmark to adaptively calculate the dynamic attention weight for each monitoring data stream. The global features, after being dynamically weighted and adaptively integrated, are input into a nonlinear mapping network specific to this work section, and the output is an implicit data tensor representing the physical process of the current work section.

3. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 2, characterized in that: The method of achieving physical causal interlocking through shared state tensors includes: generating a globally unique discrete physical batch identifier for the currently processed single batch of materials; performing hard variable sharing in memory space, and directly defining the state tensor derived by the previous process agent for the discrete physical batch identifier as the prior deduction input data when the next process agent executes the synchronization and decision-making protocol in physical memory address; The state tensor is a composite tensor obtained by orthogonally splicing an explicit data tensor that can be directly observed and a latent data tensor that is derived, after correlation filtering. It is used to represent the material state input to the next intelligent agent. Among them, correlation screening is based on the correlation between each data point and the next work section.

4. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 3, characterized in that: The deep interlocking equilibrium operator includes reading explicit monitoring data collected in this section as local observation constraints; The preset set of physical mechanism equations for mineral processing, including mass balance equations and energy conservation operators, is used to fit the mechanism to local observation constraints using the a priori deduction features. If the implicit data generated in the previous stage and the local explicit monitoring data cannot achieve numerical convergence in the physical mechanism equation set, then logical interference will occur on the feature inference path of the current agent; the energy conservation operator extracts the physical energy difference that cannot be closed in the inference process, and quantizes it into a high-dimensional physical causal violation residual vector through the variational mapping algorithm. Using the physical causal conflict residual vector as a trigger signal, the current agent's normal forward inference process is interrupted instantly, and the correction program of the previous agent is activated in reverse, entering the synchronous co-decision protocol.

5. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 4, characterized in that: After receiving the reverse-transmitted physical causal conflict residual vector, the intelligent agent of the previous section initiates the reverse anchoring adversarial algorithm: The nonlinear mapping network of the previous stage updates the implicit data tensor input to the current agent by fine-tuning the network parameters, thereby reducing the magnitude of the inconsistency residual vector. At the same time, the agent of the previous stage extracts its own explicit data tensor as the truth anchor point and forces a judgment on whether the corrected implicit data deviates from its own physical reality. By iteratively playing a game between the physical authenticity of the previous stage agent and the magnitude of the inconsistency residual vector of the current agent, until an equilibrium point that minimizes the sum of the residuals of both sides is found, the physical self-consistent interlocking point is taken as the physical self-consistent interlocking point. By utilizing the network parameters of the previous work segment's agent when the physical self-consistent interlocking point is reached, the previous work segment is induced to output the corrected implicit data. At the same time, after sensing that the interlocking point has been reached, the current work segment's agent uses the corrected shared data and combines it with the local explicit data tensor to generate the current work segment's implicit data tensor.

6. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 5, characterized in that: When the intelligent agent of the previous section outputs the corrected implicit data, the intelligent agent of the previous section uses the corrected implicit data tensor to perform positive physical reconstruction verification on the local explicit data tensor. If the overall loss of the reconstruction verification converges, but the error gradient is extremely concentrated on some explicit data and exceeds the preset boundary, then it is determined that the corresponding explicit data itself has hardware-level distortion or drift. After tracing the source of the distorted or drifting data, the confidence level of the explicit data collected by the abnormal sensor node is downgraded and the anomaly is corrected. After the anomaly is corrected, a hard rollback protocol is triggered to restore the feature inference network parameters of the previous process agent to the initial network state, and to re-derive the implicit data tensor and update the adversarial relationship with the current agent. The anomaly correction includes extracting the remaining valid explicit data other than the fault dimension, as well as the updated implicit data tensor, inputting them into the built-in cross-modal prediction reconstruction operator, generating the current predicted true value for the current batch of materials, and replacing the original distorted data of the fault node.

7. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 6, characterized in that: The cross-modal prediction reconstruction operator includes two branches constructed horizontally and vertically using a conditional mapping operator; Lateral cross-modal spatial causal inference branch: Through the conditional mapping operator, extract the remaining effective explicit data tensors except for the fault dimension at the current time, as well as the implicit data tensors after verification and update, as multi-dimensional prior conditions, and derive the spatial prediction truth value that conforms to the current transient process mechanism on the spatial physical manifold. The temporal fluctuation evolution branch of longitudinal historical inertia: extract the continuous historical sequence of the fault dimension within a preset time window before the abnormal drift occurs; separate and extract the temporal fluctuation characteristics that characterize mechanical oscillation and the local trend characteristics that characterize the equipment operating benchmark in the historical sequence, and derive the true value of the temporal evolution prediction that conforms to the physical operating inertia of the equipment. The first derivative of the remaining valid explicit data tensor is calculated in real time to assess the physical volatility of the current operating condition; dynamic gating coefficients are generated based on the physical volatility, and the two-branch prediction results are weighted and jointly decided to output the current true prediction value.

8. The method for generating spatiotemporal full-modal data of a mineral processing plant as described in claim 7, characterized in that: By configuring multiple dedicated intelligent agents throughout the entire process, corrective actions and updates are performed upwards, and updated data is shared downwards, enabling global iteration across work sections. Calculate the differential difference of the implicit data tensor output by each agent in two consecutive iterations; When the change in the correction results output by all agents in the entire process during the adversarial update continues to decay and eventually reaches the convergence criterion, the game of the cross-section physical causal interlocking network is determined to end, and a global steady-state equilibrium state is reached; and the parameters that reach the global steady-state equilibrium state are used as the equilibrium state parameters.

Images