Large model-based flotation working condition intelligent evaluation method and storage medium

Abstract

The application discloses a kind of based on big model's flotation condition intelligent evaluation method and storage medium, the steps of the method include: step S1: original data pre-processing and input definition;The collected process parameters are organized as multivariate time series matrix, and multivariate time series matrix is standardized;Step S2: trend feature extraction;The potential change trend of flotation condition is automatically captured using frozen weight large language model LLM;Step S3: potential self-attention feature modeling step;S4: cross fusion with trend feature and original feature, to enhance the feedback effect of trend feature on original convolution feature;Step S5: intelligent condition evaluation output;Fusion features are mapped to interpretable index space, to provide decision basis for flotation process optimization and closed-loop control.The storage medium is used to store the computer program for executing the above method.The application has the advantages of higher intelligent degree, better controllability, wider application range, etc.

Description

Technical Field

[0001] This invention mainly relates to the field of mineral processing automation technology, specifically a method and storage medium for intelligent evaluation of flotation conditions based on a large model. Background Technology

[0002] Flotation is one of the most widely used and critical separation technologies in mineral processing. Its operating status directly affects concentrate grade, recovery rate, and the stability and economic benefits of the entire beneficiation process. With the declining grade of ore and the increasing complexity of mineral composition, the dynamic behavior of the flotation process has become more nonlinear and uncertain, significantly increasing the difficulty of operating condition control and evaluation. How to achieve accurate, real-time, and intelligent evaluation of flotation conditions has become an important research direction in the fields of mineral processing automation and intelligent manufacturing.

[0003] Currently, the evaluation of flotation conditions mainly relies on the following methods:

[0004] 1. Judgment based on human experience;

[0005] Operators typically rely on their accumulated experience to judge the operating conditions by observing the foam layer morphology, reagent dosage, and some online monitoring indicators. While this method is intuitive, it is highly dependent on personal experience, subjective, and difficult to establish unified, quantifiable evaluation standards. Furthermore, it is prone to misjudgment under complex operating conditions.

[0006] 2. Statistical modeling and mechanistic modeling methods;

[0007] Some studies have attempted to model the flotation process using multiple regression, principal component analysis (PCA), cluster analysis, or flotation kinetic equations. These methods can reveal certain correlations between variables, but they typically assume linear or weakly nonlinear relationships between variables, making it difficult to address the complex nonlinear coupling characteristics among multiple variables in the flotation process, and limiting the model's generalization ability.

[0008] 3. Traditional machine learning methods;

[0009] With the development of industrial data acquisition and sensing technologies, methods such as Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) networks have been gradually applied to flotation condition identification and prediction. These methods have improved the nonlinear modeling capabilities of the models to some extent, but the following problems still exist: a) They rely heavily on feature engineering, and the model performance is limited by manually designed input features; b) They lack the ability to mine potential patterns in long-term time-series data, making it difficult to capture the long-term trends of the operating conditions; c) The utilization rate of multivariate interaction features is low, making it difficult to fully reflect the evolution mechanism of the operating conditions.

[0010] In summary, existing methods for evaluating flotation conditions have the following technical shortcomings:

[0011] 1. Insufficient long-term time-dependent capture capability: Flotation conditions have significant dynamic and temporal characteristics. Traditional methods mostly focus on short-term data analysis, making it difficult to model the potential evolution of conditions over long time scales, resulting in unstable prediction results and discontinuous trends.

[0012] 2. Insufficient utilization of multivariate interaction features: The flotation process involves multiple key parameters such as grade, particle size, reagent dosage, and pH. There are complex nonlinear coupling relationships between these variables. Existing methods have failed to fully explore their interaction information, which limits the expressive power of the model.

[0013] 3. High reliance on human experience and lack of interpretability: The evaluation process often relies on the operator's experience and judgment, resulting in highly subjective results. It lacks a unified, structured quantitative indicator system and an interpretable quantitative indicator system, leading to highly subjective and poor repeatability of evaluation results, which makes it difficult to meet the needs of intelligent control systems.

[0014] 4. Problems with the lack of smoothness and trend consistency in prediction results: Existing methods are prone to curve jumps, trend reversals or excessive fluctuations when predicting flotation indicators. They lack constraints on the continuity and smoothness of time series, which affects the engineering usability of the prediction results.

[0015] Therefore, existing flotation condition evaluation technologies are insufficient to meet the comprehensive requirements of modern mineral processing processes for intelligence, real-time performance, interpretability, and controllability. Summary of the Invention

[0016] The technical problem to be solved by this invention is: in view of the technical problems existing in the prior art, this invention provides a more intelligent, controllable, and widely applicable intelligent evaluation method and storage medium for flotation conditions based on a large model.

[0017] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0018] A method for intelligent evaluation of flotation conditions based on a large model, comprising the following steps:

[0019] Step S1: Raw data preprocessing and input definition; organize the collected process parameters into a multivariate time series matrix and standardize the multivariate time series matrix;

[0020] Step S2: Trend feature extraction; Automatically capture potential changing trends of flotation conditions using a frozen weighted large-scale language model (LLM);

[0021] Step S3: Modeling of latent self-attention features; Establishing a self-attention mechanism within the trend feature space to enhance the ability to model long-term dependencies between latent trends;

[0022] Step S4: Cross-fuse the trend features with the original features to enhance the feedback effect of the trend features on the original convolutional features;

[0023] Step S5: Intelligent operating condition evaluation output; the fused features are mapped to an interpretable index space to provide a decision-making basis for flotation process optimization and closed-loop control.

[0024] As a further improvement of the present invention: the multivariate time series matrix in step S1 is:

[0025]

[0026] Where R is the real number field, indicating that all elements in the matrix are real numbers; Time steps, i.e., the number of consecutive sampling time points; Process variables include grade, particle size, collector, foaming agent, modifier, inhibitor, and pH.

[0027] : The original multivariate time series data matrix, used to describe the state of the flotation operation at each time step.

[0028] As a further improvement of the present invention: the standardization processing of the multivariate time series matrix in step S1 includes normalization processing, specifically including:

[0029]

[0030] in, The mean of each feature dimension, used to shift the center of the data distribution; The standard deviation of each feature dimension is used to scale the range of the data distribution. The normalized input matrix ensures the numerical stability of convolution and attention operations.

[0031] As a further improvement of the present invention: the process of step S2 includes:

[0032] Step S201: Based on the normalized matrix Constructing flotation condition analysis prompts: Normalize the multivariate time series matrix. Input prompt word construction function Generate structured cue vectors ;

[0033]

[0034] in, The prompt word construction function maps multivariate temporal information into a vector representation that can be understood by a large language model; Used to input frozen weighted language models (LLMs) to carry historical operating conditions and trend characteristics;

[0035] Step S202: Extract LLM trend features; input the prompt words into the frozen weighted LLM and extract the last hidden vector:

[0036]

[0037] in, : A feature matrix that implies long-term operating condition changes and can capture potential patterns that change over time;

[0038] Step S203: Local temporal convolution downsampling; normalization matrix Perform one-dimensional convolution downsampling:

[0039]

[0040] in, : Kernel size controls the length of the local time window; : Stride, used for time downsampling; The convolution output matrix contains local temporal pattern information; It is a one-dimensional convolutional layer;

[0041] Step S204: Multi-head cross-attention fusion; using trend features as the query and convolutional features as the key and value:

[0042]

[0043]

[0044] in, , , A learnable linear mapping matrix can be used to adjust the contribution of each feature in attention calculation; The fused feature matrix combines trend and local time series information.

[0045] As a further improvement of the present invention: In step S3, potential self-attention is constructed.

[0046]

[0047]

[0048] in, Learnable mapping matrix; : Latent feature matrix, used to capture global long-term dependencies.

[0049] As a further improvement of the present invention: when performing cross-fusion of trend features and original features in step S4, the cross-attention operation is as follows:

[0050]

[0051]

[0052] in, Learnable linear mapping matrices; The output matrix is ​​fused, combining trends with the original features.

[0053] As a further improvement of the present invention: in step S5, when outputting the intelligent working condition evaluation, the fused features are mapped to the interpretable index space:

[0054]

[0055] in, Output mapping matrix; Bias vector; Multidimensional working condition evaluation index matrix.

[0056] As a further improvement of the present invention, it also includes: constructing a comprehensive objective function for multi-objective collaborative optimization, the process of which includes:

[0057] Step S100: Use multidimensional index regression error term; that is, construct regression error term for core monitoring indicators to quantify the deviation between model prediction value and actual operating condition value;

[0058] Step S200: Employ time-series trend and smoothing constraints to suppress abrupt changes and drastic fluctuations in the prediction curve, thereby improving overall smoothness;

[0059] Step S300: Constructing the overall objective function; combining the two types of losses, the final objective function is defined as follows:

[0060]

[0061] in, , These are hyperparameters used to balance the importance of different loss terms; they are minimized. This achieves synergistic optimization of prediction accuracy and trend smoothness.

[0062] As a further improvement of the present invention: step S200 includes:

[0063] A first-order difference trend constraint is introduced to prevent "jumps" or "trend reversals" in the prediction results:

[0064]

[0065] in, For trend-constrained weights, This represents the first difference of the predicted value. The first difference representing the true value;

[0066] A second-order difference smoothing constraint is introduced to suppress drastic fluctuations in the prediction curve:

[0067]

[0068] in, To smooth the constraint weights, This represents the second difference of the predicted value.

[0069] The present invention further provides a storage medium that can be read by a computer or processor, wherein the storage medium stores a computer program for performing any of the above methods.

[0070] Compared with the prior art, the advantages of the present invention are as follows:

[0071] 1. The intelligent evaluation method and storage medium for flotation conditions based on a large model of the present invention have a higher degree of intelligence, better controllability, and wider applicability. The present invention can simultaneously capture the short-term local features and long-term potential trend features of flotation conditions, and can make full use of multivariate interaction information. While reducing dependence on human experience, it realizes an objective, accurate, interpretable, smooth and stable comprehensive intelligent evaluation method for flotation conditions, thereby providing reliable support for the optimization and closed-loop control of flotation process.

[0072] 2. The intelligent evaluation method and storage medium for flotation conditions based on a large model of the present invention automatically extracts the potential trend features of flotation conditions by introducing a large language model (LLM) with frozen weights, which can capture the changing patterns of indicators such as grade and recovery rate on medium- and long-term scales. Compared with traditional statistical modeling and shallow machine learning methods, the present invention has stronger robustness and predictive stability in long-term time-series dependency modeling.

[0073] 3. The intelligent evaluation method and storage medium for flotation conditions based on a large model of the present invention uses one-dimensional convolutional downsampling to extract short-term time-series features and combines them with long-term trend features extracted by LLM to achieve dual modeling of local fluctuations and global trends, effectively avoiding the problem of a single model being overly sensitive to short-term noise or ignoring long-term trends.

[0074] 4. The intelligent evaluation method and storage medium for flotation conditions based on a large model of the present invention achieves efficient interaction and fusion of original time-series features and trend features through a bidirectional cross-attention mechanism and a potential self-attention module. It can fully explore the nonlinear coupling relationship between multi-dimensional process parameters such as grade, particle size, reagent dosage, and pH, and improve the comprehensiveness and accuracy of condition evaluation.

[0075] 5. The intelligent evaluation method and storage medium for flotation conditions based on a large model of the present invention, by introducing trend constraints and smoothing constraints into the objective function, ensures the continuity and rationality of the prediction results in the time series, avoids the common phenomena of prediction curve jumps, trend reversals or excessive oscillations in traditional methods, and improves the engineering usability of the prediction results.

[0076] 6. The intelligent evaluation method and storage medium for flotation conditions based on a large model of the present invention maps deeply fused features to an interpretable index space such as grade trend, recovery rate prediction, stability index, and separation efficiency, providing intuitive and quantitative evaluation results for operators and process optimization systems, which facilitates closed-loop control and decision support.

[0077] 7. The intelligent evaluation method and storage medium for flotation conditions based on large models of the present invention have strong versatility and scalability. It is not only applicable to the evaluation of flotation conditions, but can also be extended to other industrial multivariate time series data scenarios, such as pulp concentration prediction, chemical process monitoring, and metallurgical process optimization. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the present invention in a specific embodiment.

[0079] Figure 2 This is a schematic diagram of the network structure in a specific embodiment of the present invention.

[0080] Figure 3 This is a schematic diagram of the structured intelligent analysis prompts for flotation conditions in a specific embodiment of the present invention. Detailed Implementation

[0081] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] like Figure 1 and Figure 2 As shown, this invention discloses an intelligent evaluation method for flotation conditions based on a large model, the steps of which include:

[0083] Step S1: Raw data preprocessing and input definition;

[0084] The collected process parameters are organized into a multivariate time series matrix, and the multivariate time series matrix is ​​standardized to eliminate dimensional differences and enhance the stability of model training.

[0085] Step S2: Trend feature extraction;

[0086] The frozen weight large language model LLM is used to automatically capture the potential changing trends of flotation conditions, so as to improve the model's ability to perceive the implicit operating conditions.

[0087] Step S3: Modeling latent self-attention features;

[0088] A self-attention mechanism is established within the trend feature space to enhance the ability to model long-term dependencies between potential trends.

[0089] This invention enables the model to establish global dependencies in the trend feature space through a self-attention mechanism, thereby enhancing its ability to capture long-term operating condition changes and suppressing the interference of short-term fluctuations on prediction.

[0090] Step S4: Cross-merge the trend features with the original features;

[0091] It is used to enhance the feedback effect of trend features on the original convolutional features, thereby improving the model's sensitivity to real-time operating condition fluctuations and prediction accuracy.

[0092] Step S5: Intelligent operating condition evaluation output;

[0093] The fused features are mapped to an interpretable index space to provide a basis for decision-making in flotation process optimization and closed-loop control. That is, a linear mapping is used to transform deep features into interpretable indices, which facilitates visualization analysis, closed-loop control, and operating condition optimization.

[0094] In a specific application example, the multivariate time series matrix in step S1 is:

[0095]

[0096] Where R is the real number field, indicating that all elements in the matrix are real numbers; Time steps, i.e., the number of consecutive sampling time points;

[0097] Process variables include grade, particle size, collectors, foaming agents, modifiers, inhibitors, and pH.

[0098] : The original multivariate time series data matrix, used to describe the state of the flotation operation at each time step.

[0099] In a specific application example, the standardization process performed on the multivariate time series matrix in step S1 includes normalization, specifically including:

[0100]

[0101] in, The mean of each feature dimension, used to shift the center of the data distribution;

[0102] The standard deviation of each feature dimension is used to scale the range of the data distribution.

[0103] The normalized input matrix ensures the numerical stability of convolution and attention operations.

[0104] The standardization process described above not only balances the contribution of each feature but also prevents gradient vanishing or exploding, thereby improving the accuracy and convergence speed of subsequent deep learning feature extraction. See also... Figure 3 This is a schematic diagram of the structured intelligent analysis prompts for flotation conditions in a specific embodiment of the present invention.

[0105] In a specific application example, the detailed process of step S2 may include:

[0106] Step S201: Based on the normalized matrix Constructing flotation condition analysis prompts: Normalize the multivariate time series matrix. Input prompt word construction function Generate structured cue vectors ;

[0107]

[0108] in, The prompt word construction function maps multivariate temporal information into a vector representation that can be understood by a large language model. Used to input frozen weighted language models (LLMs) to carry historical operating conditions and trend characteristics.

[0109] Step S202: Extract LLM trend features;

[0110] Input the cue words into the LLM with frozen weights and extract the last hidden vector:

[0111]

[0112] in, : A feature matrix that implies long-term operating condition changes can capture the potential patterns of changes in grade, recovery rate, etc. over time.

[0113] Step S203: Local temporal convolution downsampling;

[0114] For normalized matrix Perform one-dimensional convolution downsampling:

[0115]

[0116] in, : Kernel size, controls the length of the local time window. : Stride, used for time downsampling. The output matrix of the convolution contains local temporal pattern information. It is a one-dimensional convolutional layer. : Number of convolutional channels, each channel corresponds to a local pattern feature.

[0117] Step S204: Multi-head cross-attention fusion;

[0118] Use trend features as the query and convolutional features as the key and value:

[0119]

[0120]

[0121] in, , A learnable linear mapping matrix can be used to adjust the contribution of each feature in attention calculation; Attention model dimension, defining the size of the feature space processed by each attention head; The fused feature matrix combines trend and local time series information.

[0122] In a specific application example, step S3 involves constructing potential self-attention:

[0123]

[0124]

[0125] in, Learnable mapping matrix; : Latent feature matrix, used to capture global long-term dependencies.

[0126] In a specific application example, when performing cross-fusion of trend features and original features in step S4, the cross-attention operation is as follows:

[0127]

[0128]

[0129] in, Learnable linear mapping matrices; The output matrix is ​​fused, combining trends with the original features.

[0130] In a specific application example, in step S5, when outputting the intelligent working condition evaluation, the fused features are mapped to the interpretable indicator space:

[0131]

[0132] in, Output mapping matrix; Bias vector; Multidimensional working condition evaluation index matrix.

[0133] In a preferred embodiment, the present invention further includes: constructing a comprehensive objective function for multi-objective collaborative optimization. This is to achieve collaborative prediction of multi-dimensional flotation operating condition indicators and ensure the accuracy and stability of the prediction results in medium-term (typically 12–48 hours) prediction tasks. This objective function ensures the reliability and engineering applicability of the model prediction results through joint regulation of the regression accuracy and temporal trend consistency of multiple indicators.

[0134] Specifically, the process of constructing the comprehensive objective function for multi-objective collaborative optimization can include:

[0135] Step S100: Use multidimensional index regression error term; that is, construct regression error term for core monitoring indicators to quantify the deviation between model prediction value and actual operating condition value;

[0136] For example, for the core monitoring indicators of the flotation process (grade, recovery rate, stability index, separation efficiency), a regression error term is constructed to quantify the deviation between the model's predicted values ​​and the actual operating conditions.

[0137]

[0138] in, To predict the number of time steps, For time indexing, Indexed by indicator dimensions (1–4, corresponding to grade, recovery rate, stability index, and sorting efficiency). and These are the actual value and the predicted value, respectively. The weights of each indicator satisfy the following conditions: This item is the standard weighted mean square error, used to ensure prediction accuracy.

[0139] Step S200: Use time-series trend and smoothing constraints to suppress abrupt changes and drastic fluctuations in the prediction curve and improve overall smoothness.

[0140] For example, flotation conditions have continuous dynamic characteristics. To avoid "jumps" or "trend reversals" in the prediction results, a first-order difference trend constraint is introduced:

[0141]

[0142] in, For trend-constrained weights, This represents the first difference of the predicted value. This represents the first difference of the true value. This term ensures the continuity and reasonableness of the forecast curve by constraining the consistency between the predicted trend and the actual trend.

[0143] Furthermore, to suppress drastic fluctuations in the prediction curve, a second-order difference smoothing constraint is introduced:

[0144]

[0145] in, To smooth the constraint weights, This represents the second difference of the predicted value. This term is used to reduce oscillations in the prediction curve and improve overall smoothness.

[0146] Step S300: Construction of the overall objective function;

[0147] Combining the two types of losses mentioned above, the final objective function is defined as follows:

[0148]

[0149] in, , These are hyperparameters used to balance the importance of different loss terms. This is achieved by minimizing... This achieves synergistic optimization of prediction accuracy and trend smoothness.

[0150] The method described above in this invention is scalable and universal, meaning it can be applied to the intelligent analysis and index evaluation of other industrial multivariate time series data. Specifically, the method is suitable for the intelligent evaluation of multivariate industrial time series data, not limited to flotation conditions, and can be extended to industrial scenarios such as slurry concentration prediction and chemical process monitoring. In practical applications, LLM can be fine-tuned according to different processes or frozen weights can be used directly. The kernel size, stride, and number of attention heads are adjustable to adapt to different sampling frequencies and feature dimensions.

[0151] The present invention further provides a storage medium that can be read by a computer or processor, wherein the storage medium stores a computer program for performing the above-described method.

[0152] Those skilled in the art will understand that the above embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0153] The detailed process of this method will be described below with reference to its implementation in a specific application.

[0154] Data preparation;

[0155] Data source and collection: In the flotation workshop of a lead-zinc beneficiation plant, multiple online sensors and automatic sampling devices were deployed. The collected process parameters include:

[0156] Lead concentrate grade, zinc concentrate grade; pulp particle size distribution; collector dosage; frother dosage; modifier dosage; inhibitor dosage; pH value;

[0157] The sampling frequency was 1 minute / time, and the data was collected continuously for 20 days, resulting in approximately 29,000 multivariate time series data points.

[0158] The table below shows an example of multivariate time-series sample data collected under flotation conditions at a lead-zinc beneficiation plant. Each data point represents a time step (e.g., sampling once per minute) and includes 8 key process parameters:

[0159] Table 1. Multivariate time-series sample data collected under flotation conditions at a lead-zinc beneficiation plant.

[0160]

[0161] Parameter description:

[0162] Lead concentrate grade (%): The percentage of lead by mass in the concentrate after flotation.

[0163] Zinc concentrate grade (%): The percentage of zinc by mass in the concentrate after flotation.

[0164] Particle size (μm): The average particle size of particles in the slurry.

[0165] Collector dosage (g / t): The amount of collector added per unit of ore.

[0166] Foaming agent dosage (g / t): The amount of foaming agent added per unit of ore.

[0167] Adjuster dosage (g / t): The dosage of reagent used to adjust the properties of slurry.

[0168] Inhibitor dosage (g / t): The amount of reagent used to inhibit impurity minerals.

[0169] pH value: The acidity or alkalinity of the pulp, which affects the selectivity of flotation.

[0170] Data preprocessing;

[0171] Organize the original data into a multivariate time series matrix ,in , .

[0172] Standardize the parameters of each dimension:

[0173]

[0174] in, and These are the mean and standard deviation of each feature dimension, respectively.

[0175] The processed data is used as input to the model to ensure that different parameters contribute in a balanced manner on the same numerical scale.

[0176] Key parameters for model training

[0177] In the implementation of this invention, to ensure the convergence, stability, and prediction accuracy of the intelligent evaluation model for flotation conditions, it is necessary to reasonably set the key parameters for model training. Specifically:

[0178] (1) Data partitioning and batch processing;

[0179] Dataset partitioning: The collected lead-zinc flotation data were divided into training, validation, and test sets in a ratio of 7:2:1.

[0180] Batch size: Set to 64 to improve computational efficiency while ensuring training stability.

[0181] (2) Optimizer and learning rate strategy;

[0182] Optimizer: The AdamW optimizer is used, which can suppress overfitting while maintaining weight decay.

[0183] Initial learning rate: set to 1×10 -4 .

[0184] Learning rate scheduling: A cosine annealing strategy is adopted to gradually reduce the learning rate during training, avoiding oscillations and accelerating convergence.

[0185] (3) Training rounds and early stop mechanism;

[0186] Maximum training epochs: set to 100.

[0187] Early Stopping Strategy: If the validation set loss does not decrease within 10 consecutive rounds, training is terminated early to prevent overfitting.

[0188] (4) Regularization and normalization;

[0189] Dropout ratio: Introduce 0.2–0.3 Dropout in fully connected layers and attention layers to enhance the model's generalization ability.

[0190] Weight normalization: Layer normalization is used in convolutional and attention layers to ensure numerical stability.

[0191] (5) Loss function weight settings;

[0192] Comprehensive loss function:

[0193]

[0194] Weight parameters: , This was determined through cross-validation.

[0195] Technical significance: To ensure that the forecast results achieve a balance between accuracy, trend consistency and smoothness.

[0196] (6) Hardware and operating environment;

[0197] Hardware platform: NVIDIA 4090 GPU, 24GB video memory.

[0198] Software environment: Python 3.12, PyTorch 2.2 deep learning framework.

[0199] Training time: Under the above configuration, a single complete training session takes approximately 6 hours to converge.

[0200] Testing and evaluation;

[0201] After completing the model training, the intelligent evaluation model for flotation conditions proposed in this invention needs to be verified and compared with existing methods to prove its effectiveness and superiority.

[0202] (1) Validate the dataset;

[0203] A test set was selected from 20% of the non-training operating data continuously collected from a lead-zinc beneficiation plant, including multivariate time-series data such as lead and zinc grades, recovery rates, reagent dosage, and pH.

[0204] The test set covers three typical operating conditions: normal operating conditions, fluctuating operating conditions, and abnormal operating conditions, ensuring that the verification results are representative.

[0205] (2) Evaluation indicators;

[0206] Mean Squared Error (MSE): Measures the deviation between the predicted value and the actual value.

[0207] Trend Consistency Coefficient (TCC): Measures the consistency between the predicted curve and the actual curve in the trend direction.

[0208] Smoothness Index (SI): The smoothness of the prediction curve is calculated by second-order difference. The smaller the value, the more stable the prediction result.

[0209] Accuracy of comprehensive working condition evaluation (Acc): The overall accuracy of judgment based on four output indicators (grade trend, recovery rate, stability index, and sorting efficiency).

[0210] (3) Comparison method;

[0211] Random Forest (RF): Represents a traditional machine learning method.

[0212] Long Short-Term Memory (LSTM) network: represents a deep learning temporal modeling method.

[0213] The method of this invention is based on large model trend feature extraction + convolutional local feature modeling + attention fusion.

[0214] (4) The experimental results are shown in the table below:

[0215] Table 2 Experimental Results

[0216]

[0217] (5) Results analysis;

[0218] The method of this invention significantly outperforms RF and LSTM in prediction accuracy (MSE), reducing the error by approximately 35%–60%. In terms of trend consistency (TCC), this invention achieves 0.88, a 15% improvement over LSTM, indicating that the predicted curve better reflects the actual operating conditions.

[0219] The present invention achieves the best results in terms of smoothness index (SI), with small fluctuations in the prediction curve, making it more suitable for engineering applications.

[0220] In terms of overall accuracy (Acc), this invention achieves 91.2%, which is nearly 10 percentage points higher than LSTM and nearly 20 percentage points higher than RF.

[0221] (6) Engineering application effect

[0222] In actual production, the method of this invention can predict potential fluctuations in lead-zinc flotation conditions 1.5 hours in advance, assisting operators in adjusting the dosage of reagents in a timely manner.

[0223] By integrating with a closed-loop control system, this method helped the concentrator achieve the following during the trial operation phase:

[0224] Lead concentrate grade increased by 1.0%;

[0225] The zinc concentrate grade increased by 0.8%;

[0226] The overall recovery rate increased by 2.3%;

[0227] Drug consumption decreased by 5%.

[0228] In summary, the practical applications described above verify the effectiveness of the method of this invention in actual lead-zinc flotation conditions. Compared with traditional methods, this invention not only significantly improves prediction accuracy and trend consistency, but also performs better in terms of smoothness and engineering usability, providing reliable support for the intelligent, stable, and energy-efficient operation of the flotation process.

[0229] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent evaluation of flotation conditions based on a large model, characterized by the following steps: include: Step S1: Raw data preprocessing and input definition; The collected process parameters are organized into a multivariate time series matrix, and the multivariate time series matrix is ​​standardized. Standardization of multivariate time series matrices includes normalization, specifically: in, The mean of each feature dimension, used to shift the center of the data distribution; The standard deviation of each feature dimension is used to scale the range of the data distribution. Normalized matrices can guarantee the numerical stability of convolution and attention operations; Step S2: Trend feature extraction; Automatically capture potential changing trends of flotation conditions using a frozen weighted large-scale language model (LLM); The process of step S2 includes: Step S201: Based on the normalized matrix Constructing flotation condition analysis prompts: Normalize the multivariate time series matrix. Input prompt word construction function Generate structured cue vectors ; in, The prompt word construction function maps multivariate temporal information into a vector representation that can be understood by a large language model; Used to input frozen weighted language models (LLMs) to carry historical operating conditions and trend features; Step S202: Extract LLM trend features; input the prompt words into the frozen weighted LLM and extract the last hidden vector: in, : A feature matrix that implies long-term operating condition changes and can capture potential patterns that change over time; Step S203: Local temporal convolution downsampling; normalization matrix Perform one-dimensional convolution downsampling: in, : Kernel size controls the length of the local time window; : Step size, used for time downsampling; The convolution output matrix contains local temporal pattern information; It is a one-dimensional convolutional layer; Step S204: Multi-head cross-attention fusion; using trend features as the query and convolutional features as the key and value: in, , A learnable linear mapping matrix can be used to adjust the contribution of each feature in attention calculation; The fused feature matrix combines trend and local time-series information. Step S3: Modeling of latent self-attention features; Establishing a self-attention mechanism within the trend feature space to enhance the ability to model long-term dependencies between latent trends; Step S4: Cross-fuse the trend features with the original features to enhance the feedback effect of the trend features on the original convolutional features; Step S5: Intelligent operating condition evaluation output; the fused features are mapped to an interpretable index space to provide a decision-making basis for flotation process optimization and closed-loop control.

2. The intelligent evaluation method for flotation conditions based on a large model according to claim 1, characterized in that, The multivariate time series matrix in step S1 is: Where R is the real number field, indicating that all elements in the matrix are real numbers; Time steps, i.e., the number of consecutive sampling time points; Process variables include grade, particle size, collector, foaming agent, modifier, inhibitor, and pH. : The original multivariate time series data matrix, used to describe the state of the flotation operation at each time step.

3. The intelligent evaluation method for flotation conditions based on a large model according to claim 1 or 2, characterized in that, In step S3, potential self-attention is constructed: in, Learnable mapping matrix; : Latent feature matrix, used to capture global long-term dependencies.

4. The intelligent evaluation method for flotation conditions based on a large model according to claim 1 or 2, characterized in that, When performing cross-fusion of trend features and original features in step S4, the cross-attention operation is as follows: in, Learnable linear mapping matrices; The output matrix is ​​fused, combining trends with the original features.

5. The intelligent evaluation method for flotation conditions based on a large model according to claim 1 or 2, characterized in that, In step S5, when outputting the intelligent working condition evaluation, the fused features are mapped to the interpretable index space: in, Output mapping matrix; Bias vector; Multidimensional working condition evaluation index matrix.

6. The intelligent evaluation method for flotation conditions based on a large model according to claim 1 or 2, characterized in that, Also includes: The process of constructing a comprehensive objective function for multi-objective collaborative optimization includes: Step S100: Use multidimensional index regression error term; that is, construct regression error term for core monitoring indicators to quantify the deviation between model prediction value and actual operating condition value; Step S200: Employ time-series trend and smoothing constraints to suppress abrupt changes and drastic fluctuations in the prediction curve, thereby improving overall smoothness; Step S300: Constructing the overall objective function; combining the two types of losses, the final objective function is defined as follows: in, , These are hyperparameters used to balance the importance of different loss terms; they are minimized. This achieves synergistic optimization of prediction accuracy and trend smoothness.

7. The intelligent evaluation method for flotation conditions based on a large model according to claim 6, characterized in that, Step S200 includes: A first-order difference trend constraint is introduced to prevent "jumps" or "trend reversals" in the prediction results: in, For trend-constrained weights, This represents the first difference of the predicted value. The first difference representing the true value; A second-order difference smoothing constraint is introduced to suppress drastic fluctuations in the prediction curve: in, To smooth the constraint weights, This represents the second difference of the predicted value.

8. A storage medium capable of being read by a computer or processor, characterized in that, The storage medium stores a computer program for performing the method as described in any one of claims 1-7.

Images