Methods, systems, and storage media for multivariate time-series feature extraction and grade prediction in flotation processes.

Abstract

This invention discloses a method, system, and storage medium for multivariate time-series feature extraction and grade prediction in flotation processes. The method includes the following steps: Step S1: Raw data input and encoding; encoding and structured input of time-series data composed of various process variables; Step S2: Extracting dynamic features; Step S3: Condition-guided encoding modeling; Step S4: Enhancing the saliency of key variables and important time segments; employing a multi-head attention mechanism to match key-value pairs generated from target-guided query vectors and multi-scale features; Step S5: Outputting a prediction module; performing feature fusion and nonlinear mapping on the attention mechanism output to output the predicted concentrate grade and recovery rate for future times. The system and storage medium are both based on the above method. This invention has advantages such as higher intelligence, better controllability, and improved prediction accuracy and model adaptability for key indicators in the flotation process.

Description

Technical Field

[0001] This invention mainly relates to the field of intelligent control technology for mineral processing, specifically a method, system, and storage medium for multivariate time-series feature extraction and grade prediction in flotation processes. Background Technology

[0002] With the continuous advancement of automation and intelligence in the mining industry, flotation, as a crucial fine separation process in mineral processing, has seen its process control level become a key bottleneck in improving ore processing efficiency and product quality.

[0003] Flotation systems are influenced by a combination of factors, including fluctuations in raw ore grade, operational disturbances, reagent ratios, and particle size distribution, resulting in a complex industrial process characterized by multivariate coupling, strong nonlinearity, and significant dynamics. This complexity makes accurately modeling the variable relationships in the flotation process and achieving precise prediction of key output indicators (such as concentrate grade and recovery rate) a critical technical challenge that urgently needs to be addressed in the field of intelligent control in mining.

[0004] In existing technologies, practitioners have proposed several typical methods for modeling the flotation process:

[0005] 1. Unidirectional Long Short-Term Memory Network (LSTM);

[0006] As a variant of traditional recurrent neural networks, LSTM effectively captures long-term dependencies in time series through gating mechanisms and is widely used in industrial process modeling. However, when faced with the complex and ever-changing variable interactions and perturbation patterns in flotation operations, LSTM is insufficient in feature extraction capabilities, responds slowly to local short-term fluctuations, and has low training efficiency, making it difficult to fully utilize the coupling characteristics between multiple variables, thus limiting the improvement of prediction performance.

[0007] 2. Transformer encoder structure;

[0008] Transformers utilize self-attention mechanisms to model global dependencies in time series, offering advantages in parallel computation and capturing long-range dependencies, and have achieved good results in some industrial forecasting tasks. However, they lack sensitivity to local time series structures and multi-level feature extraction capabilities. When faced with multi-source perturbations and noise during the flotation process, they are easily affected by redundant information, leading to a decrease in model stability and generalization ability.

[0009] 3. One-dimensional convolutional neural network (CNN-1D);

[0010] CNNs significantly improve the efficiency of feature extraction by extracting local temporal features through sliding convolutional kernels, making them a common method for industrial time series modeling. However, most existing CNNs use single-scale convolutions, lacking the ability to simultaneously perceive short-term perturbations and long-term trends, thus limiting the dimensionality of feature representation and making it difficult to cope with the complex dynamic characteristics of multi-timescale interleaved evolution in flotation systems.

[0011] Furthermore, the aforementioned methods generally lack a target-guided mechanism, failing to integrate key output indicators such as concentrate grade into the model construction process. Because the models lack target-driven conditional modeling, they struggle to flexibly adjust their focus on input features according to production needs, resulting in insufficient consistency between prediction accuracy and control strategies. This limits the models' applicability in actual production and their engineering promotion value.

[0012] In summary, existing traditional methods still have some technical shortcomings:

[0013] 1. Insufficient feature extraction capability for multivariate time series data;

[0014] Traditional models struggle to effectively capture the complex nonlinear time-series relationships among various key process variables during flotation, resulting in low prediction accuracy and an inability to accurately reflect the dynamic changes in the process.

[0015] 2. Lack of multi-scale feature fusion mechanism;

[0016] Most existing methods employ single-scale feature extraction, which cannot simultaneously take into account information on short-term fluctuations and long-term trends, thus limiting the model's ability to represent multi-level temporal characteristics of the flotation process.

[0017] 3. Failure to fully utilize target product indicators to guide modeling;

[0018] Most models do not incorporate target concentrate grade information as a guiding condition and lack a target-driven feature attention mechanism, making it difficult for the models to be flexibly adjusted and optimized for different process objectives.

[0019] 4. The prediction objective is singular, making joint optimization difficult;

[0020] Traditional forecasting methods often focus on a single indicator (such as concentrate grade), neglecting the need for synergistic optimization of recovery rate and grade, making it difficult to meet the comprehensive quality control requirements of the production process.

[0021] 5. The model lacks generalization ability and stability;

[0022] Some deep learning models rely excessively on training data and lack effective regularization and multi-objective constraints, resulting in unstable performance in real-world production environments and limited application and promotion. Summary of the Invention

[0023] 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 method, system and storage medium for multivariate time series feature extraction and grade prediction in flotation process with higher intelligence, better controllability and improved prediction accuracy and model adaptability of key indicators in flotation process.

[0024] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0025] A method for multivariate time-series feature extraction and grade prediction in flotation processes, comprising the following steps:

[0026] Step S1: Raw data input and encoding; encoding and structured input of time series data composed of various process variables;

[0027] Step S2: Extract dynamic features; extract dynamic features at different time scales through a multi-layer one-dimensional convolutional network to capture the short-term fluctuations and long-term trends of process variables;

[0028] Step S3: Condition-guided coding modeling; embed the target concentrate grade information into the feature representation to form a guided query vector, guiding the subsequent modeling process to focus on target-related features;

[0029] Step S4: Enhance the significance of key variables and important time segments; employ a multi-head attention mechanism to match key-value pairs generated based on the target-guided query vector and multi-scale features;

[0030] Step S5: Output prediction module; perform feature fusion and nonlinear mapping on the output results of the attention mechanism to output the predicted results of concentrate grade and recovery rate at future time.

[0031] As a further improvement of the present invention: the process variables in step S1 include one or more of the following: raw ore grade, tailings grade, particle size, collector, foaming agent, modifier, and inhibitor.

[0032] As a further improvement of the present invention: the process in step S1 includes:

[0033] The original input is a tensor , used to represent multivariate time-series data in the flotation process;

[0034] in: Each sample sequence corresponds to an independent production data point within a batch; each time step contains Key process variables include raw ore grade, tailings grade, particle size, collector dosage, frother dosage, modifier dosage, and inhibitor dosage; Indicates the batch number; Indicates timing length; Indicates the number of process variables

[0035] Input target concentrate grade , serving as an auxiliary target variable to guide the model's focus direction; It is a scalar time-invariant objective used to provide guidance in subsequent feature modeling.

[0036] As a further improvement of the present invention: step S2 introduces a four-layer one-dimensional convolution operation to extract temporal features at different scales, the process of which includes:

[0037] Step S201: Definition of convolution operation:

[0038]

[0039]

[0040]

[0041]

[0042] in, This represents a one-dimensional convolution operation. This represents the step size parameter, which controls the sparsity of feature sampling. This indicates the size of the convolution kernel, which affects the receptive field of the model.

[0043] Step S202: Perform a linear mapping on the output of each convolutional layer:

[0044]

[0045] in, Indicates the first The feature representation of the output of a convolutional layer after linear transformation;

[0046] Step S203: The outputs of all layers are concatenated to form the final depth feature representation:

[0047]

[0048] Here, "splicing" refers to a splicing operation performed along the direction of the feature channel. Specifically, It consists of feature maps from different layers (or different convolutional modules, etc.), which are concatenated along the channel dimension (usually the last dimension) while maintaining a consistent spatial size, to obtain a new feature map. Its number of channels is the sum of the number of channels of all input feature maps.

[0049] Step S204: Extract the key vector and value vector from the deep feature:

[0050]

[0051] in: This is the key vector used for matching; The vector is used for weighted fusion; both are obtained from the same depth feature through linear transformation.

[0052] As a further improvement of the present invention: the process of step S3 includes:

[0053] Step S301: Adjust the target concentrate grade Embedded as a guiding vector through two layers of linear transformation:

[0054]

[0055] in, This is a characteristic that guides taste;

[0056] Step S302: Output from each of the four convolutional layers. Extract the query vector:

[0057]

[0058] Step S303: After concatenating all query vectors with the target grade embedding vector, the final query input is formed:

[0059]

[0060] Step S304: Obtain the final query vector through linear transformation:

[0061]

[0062] in: This represents the query vector used for attention calculation.

[0063] As a further improvement of the present invention: the process of step S4 includes:

[0064] Step S401: Define the multi-head attention structure:

[0065]

[0066] in: This indicates a bullish attention-based strategy; For query vector; The key vector; It is a value vector;

[0067] Step S402: Attention mechanism through matching and To calculate the weights, and use those weights to... The output is obtained by weighted aggregation;

[0068] Step S403: Perform the following operations within each attention head:

[0069]

[0070] in: The first The query, key, and value in the attention head; Let the key vector be the dimension of each attention head; Used to generate normalized attention weights;

[0071] Step S404: The outputs of all heads are concatenated and then linearly transformed. Finally, the normalized output is obtained through layer normalization.

[0072]

[0073] in, This indicates that each sample is normalized.

[0074] As a further improvement of the present invention: after the multi-head attention mechanism in step S5, a final representation integrating multi-scale temporal features and target information has been obtained, denoted as... This vector is used for the joint prediction of flotation product indicators;

[0075] Will Inputting data into a multilayer perceptron (MLP), a joint estimation of future concentrate grade and recovery rate is performed:

[0076]

[0077] in: This represents a feedforward neural network that includes several linear transformations and nonlinear activation functions.

[0078] The model prediction results include two target indicators: concentrate grade and recovery rate.

[0079] As a further improvement to the present invention, it also includes: constructing a joint loss function that comprehensively considers concentrate grade and recovery rate, wherein the joint loss function uses weighted mean square error (MSE) as the error metric, and combines regularization constraints on model parameters.

[0080]

[0081] in, The overall loss function reflects the model's predictive performance and complexity. Total number of training samples; The model predicts the first The grade of concentrate in each sample; : No. The actual concentrate grade of each sample; The model predicts the first Recovery rate per sample; : No. True recovery rate of each sample; The weight of the concentrate grade error adjusts the influence of this error term in the loss function; The weight of the recovery rate error adjusts the influence of this error term in the loss function; Regularization coefficient: controls the penalty applied to the complexity of model parameters; The set of all trainable parameters of the model; : Regularization term to prevent overfitting.

[0082] The present invention further provides a system for multivariate time-series feature extraction and grade prediction in flotation processes, comprising:

[0083] The first module is used for raw data input and encoding; it encodes and structures time-series data composed of various process variables.

[0084] The second module is used to extract dynamic features; it uses a multi-layer one-dimensional convolutional network to extract dynamic features at different time scales, capturing the short-term fluctuations and long-term trends of process variables.

[0085] The third module is used for condition-guided coding modeling; it embeds the target concentrate grade information into the feature representation to form a guided query vector, which guides the subsequent modeling process to focus on the target-related features;

[0086] The fourth module is used to enhance the significance of key variables and important time segments; it employs a multi-head attention mechanism to match key-value pairs generated based on target-guided query vectors and multi-scale features.

[0087] The fifth module is used to output the prediction module; it performs feature fusion and nonlinear mapping on the output results of the attention mechanism to output the predicted results of concentrate grade and recovery rate at future time points.

[0088] 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.

[0089] Compared with the prior art, the advantages of the present invention are as follows:

[0090] 1. The method, system, and storage medium for multivariate time-series feature extraction and grade prediction in the flotation process of the present invention have a higher degree of intelligence and better controllability. They can improve the prediction accuracy and model adaptability of key indicators in the flotation process. The method integrates multi-scale convolutional neural networks and target-guided attention mechanisms to achieve efficient modeling and joint prediction of key parameters of the flotation process. It belongs to the field of mine automation, process intelligence, and industrial big data analysis technology.

[0091] 2. The method, system, and storage medium for multivariate time-series feature extraction and grade prediction in the flotation process of the present invention can integrate advanced modeling methods of multi-scale time-series perception and target condition guidance. While taking into account local dynamic response, this method can enhance the ability to capture global time-series dependence and enhance the predictive targeting of the model through a target-driven mechanism, thereby achieving high-precision joint prediction of future concentrate grade and recovery rate, significantly improving the intelligent control level and process optimization effect of the flotation process, and helping the intelligent upgrading of the mining industry.

[0092] 3. The multivariate time-series feature extraction and grade prediction method, system, and storage medium of this invention for the flotation process employ a multi-layer one-dimensional convolutional structure, which can simultaneously capture short-term fluctuations and long-term trends in the flotation process, enhancing the model's ability to express complex time-series dynamics and improving the accuracy and robustness of predictions. This is the first innovative use in the industry of a four-layer progressive one-dimensional convolutional network, combined with different phase lengths and convolutional kernel sizes, which can simultaneously extract local short-term fluctuations and global long-term trend features of the flotation process at multiple scales, significantly improving the richness and accuracy of time-series feature representation.

[0093] 4. The multivariate time-series feature extraction and grade prediction method, system, and storage medium for flotation processes of the present invention, by introducing an embedding vector of the target concentrate grade to form a conditional query, effectively guides the model to focus on key features related to the target, thereby improving the relevance of features and the interpretability of the model. Specifically, the present invention innovatively introduces an embedding vector of the target concentrate grade as conditional information, concatenating it with multi-scale time-series features to construct a guiding query vector. This drives the subsequent attention mechanism to focus on feature regions highly correlated with the target, achieving conditional feature selection and dynamic adjustment, and improving the model's adaptability to target-driven process control.

[0094] 5. The present invention relates to a method, system, and storage medium for multivariate time-series feature extraction and grade prediction in the flotation process. The model design enables joint prediction of concentrate grade and recovery rate, meeting the comprehensive needs of quality control in the flotation process and promoting the comprehensiveness and scientific nature of process optimization. This invention combines a multi-head self-attention mechanism to dynamically weight and fuse features extracted by multi-scale convolution, enhancing the model's ability to select key variables and time periods, and improving the semantic richness and prediction accuracy of feature representation. This method differs from traditional single-head or unconditional attention mechanisms and has stronger modeling capabilities.

[0095] 6. The multivariate time-series feature extraction and grade prediction method, system, and storage medium of the present invention for the flotation process effectively control model complexity and enhance the model's generalization ability under different operating conditions and batch data by using a multi-objective weighted loss function and parameter regularization, thus ensuring the stability of the prediction results. Specifically, the weighted mean square error objective function designed in this invention, which includes both concentrate grade and recovery rate, and the introduction of parameter regularization constraints, achieves multi-objective collaborative training and improved generalization ability of the model, meeting the practical needs of comprehensive quality control in the flotation process.

[0096] 7. The multivariate time-series feature extraction and grade prediction method, system, and storage medium of this invention for the flotation process feature a modular model structure. The input variables cover a variety of key process parameters, exhibiting strong adaptability and allowing for flexible adjustment and expansion according to actual production needs, thus possessing significant engineering application value. The overall architecture of this invention is clearly modular; the input, feature extraction, condition-guided, attention fusion, and output prediction modules are coupled yet remain independent, facilitating adjustments to relevant parameters and module configurations for different ore types or process conditions, and demonstrating good potential for engineering promotion. Attached Figure Description

[0097] Figure 1 This is a flowchart illustrating the present invention in a specific embodiment.

[0098] Figure 2 This is a schematic diagram of the network structure in a specific embodiment of the present invention. Detailed Implementation

[0099] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0100] This invention discloses a method for multivariate time-series feature extraction and grade prediction in flotation processes. It utilizes deep learning technology to model key variables in complex flotation processes, achieving high-precision joint prediction of future concentrate grade and recovery. The innovation of this invention lies in its comprehensive consideration of the variable coupling, temporal dynamics, and target-guided characteristics present in the flotation process, constructing an end-to-end prediction model.

[0101] like Figure 1 and Figure 2 As shown, the method for multivariate time-series feature extraction and grade prediction in the flotation process according to the present invention includes the following steps:

[0102] Step S1: Input and encoding of raw data;

[0103] Encode and structure the time-series data consisting of various process variables such as raw ore grade, tailings grade, particle size, collector, frother, modifier, and inhibitor;

[0104] Step S2: Extract dynamic features;

[0105] By extracting dynamic features at different time scales through multi-layer one-dimensional convolutional networks, the short-term fluctuations and long-term trends of process variables can be effectively captured.

[0106] Step S3: Condition-guided coding modeling;

[0107] The target concentrate grade information is embedded into the feature representation to form a guided query vector, which guides the subsequent modeling process to focus more on the target-related features;

[0108] Step S4: Enhance the significance of key variables and important time segments;

[0109] A multi-head attention mechanism is adopted to match key-value pairs generated by multi-scale features based on the target-guided query vector, thereby enhancing the significance of key variables and important time segments;

[0110] Step S5: Output the prediction module;

[0111] The output of the attention mechanism is subjected to feature fusion and nonlinear mapping to produce predictions of concentrate grade and recovery rate at future time points.

[0112] In a specific application example, the process in step S1 includes:

[0113] The original input is a tensor , is used to represent multivariate time-series data in the flotation process.

[0114] Specifically: Each sample sequence corresponds to an independent production data point within a batch; each time step contains Key process variables include raw ore grade, tailings grade, particle size, collector dosage, frother dosage, modifier dosage, and inhibitor dosage; Indicates the batch number; Indicates timing length; This indicates the number of process variables, for example, 7 in this case.

[0115] Enter the target concentrate grade as well. , which serves as an auxiliary target variable to guide the model's focus. It is a scalar time-invariant objective used to provide guidance in subsequent feature modeling.

[0116] In a specific application example, to enhance the model's ability to model the dynamic changes in the flotation process, step S2 introduces a four-layer one-dimensional convolution operation to extract temporal features at different scales. The process includes:

[0117] Step S201: Definition of convolution operation:

[0118]

[0119]

[0120]

[0121]

[0122] in, This represents a one-dimensional convolution operation. This represents the step size parameter, which controls the sparsity of feature sampling. This indicates the size of the convolution kernel, which affects the receptive field of the model.

[0123] Step S202: To unify the dimensions of subsequent processing, perform a linear mapping on the output of each convolutional layer:

[0124]

[0125] in, Indicates the first The feature representation of the output of the convolutional layer after linear transformation.

[0126] Step S203: The outputs of all layers are concatenated to form the final depth feature representation:

[0127]

[0128] Here, "splicing" refers to a splicing operation performed along the direction of the feature channel. Specifically, It consists of feature maps from different layers (or different convolutional modules, etc.), which are concatenated along the channel dimension (usually the last dimension) while maintaining a consistent spatial size, to obtain a new feature map. Its number of channels is the sum of the number of channels of all input feature maps.

[0129] Step S204: Extract the key vector and value vector from the deep feature:

[0130]

[0131] in: This is the key vector used for matching; The vector is used for weighted fusion; both are obtained from the same depth feature through linear transformation.

[0132] Step S3 aims to improve the model's ability to identify target grade-related information by introducing a conditional guidance mechanism. This mechanism enables the model to integrate semantic information of the predicted target when constructing query features. In a specific application example, the process in step S3 includes:

[0133] Step S301: Adjust the target concentrate grade Embedded as a guiding vector through two layers of linear transformation:

[0134]

[0135] in, This is a characteristic that guides taste.

[0136] Step S302: Output from each of the four convolutional layers. Extract the query vector:

[0137]

[0138] Step S303: After concatenating all query vectors with the target grade embedding vector, the final query input is formed:

[0139]

[0140] Step S304: Obtain the final query vector through linear transformation:

[0141]

[0142] in: This represents the query vector used for attention computation; this structure integrates target information with multi-scale features, improving the responsiveness of the attention mechanism to key variables.

[0143] In step S4, a multi-head self-attention mechanism is used to process the query vector. key-value pairs Information matching is performed to highlight the impact of key time segments and process variables on the target grade. In a specific application example, the process in step S4 includes:

[0144] Step S401: Define the multi-head attention structure:

[0145]

[0146] in: This indicates a bullish attention-based strategy; For query vector; The key vector; It is a value vector;

[0147] Step S402: Attention mechanism through matching and To calculate the weights, and use those weights to... The output is obtained by weighted aggregation;

[0148] Step S403: Perform the following operations within each attention head:

[0149]

[0150] in: The first The query, key, and value in the attention head; Let the key vector be the dimension of each attention head; Used to generate normalized attention weights.

[0151] Step S404: The outputs of all heads are concatenated and then linearly transformed. Finally, the normalized output is obtained through layer normalization.

[0152]

[0153] in, This indicates that normalization is performed on each sample, which helps improve training stability and enhance the model's generalization ability.

[0154] In step S5, after the multi-head attention mechanism, a final representation integrating multi-scale temporal features and target information is obtained, denoted as... This vector is used for joint prediction of flotation product indicators.

[0155] Will Inputting into a multi-layer perceptron (MLP), a joint estimation of future concentrate grade and recovery rate is completed:

[0156]

[0157] in: This represents a feedforward neural network that includes several linear transformations and nonlinear activation functions.

[0158] The model prediction results include two target indicators: predicted concentrate grade and predicted recovery rate.

[0159] Step S5 compresses the high-dimensional features resulting from attention fusion into a low-dimensional output corresponding to the business objective, ensuring the model has end-to-end predictive capabilities. Through joint modeling, the inherent coupling relationship between concentrate grade and recovery rate can be fully captured, improving prediction accuracy and stability.

[0160] As a preferred embodiment, for the multi-objective prediction problem in the flotation process, the present invention further includes: constructing a joint loss function that comprehensively considers concentrate grade and recovery rate. The joint loss function uses weighted mean squared error (Weighted MSE) as the error metric and combines it with the regularization constraints of model parameters to achieve a high-precision, stable and generalizable prediction effect.

[0161]

[0162] in, The overall loss function reflects the model's predictive performance and complexity. Total number of training samples; The model predicts the first The grade of concentrate in each sample; : No. The actual concentrate grade of each sample; The model predicts the first Recovery rate per sample; : No. True recovery rate of each sample; The weight of the concentrate grade error adjusts the influence of this error term in the loss function; The weight of the recovery rate error adjusts the influence of this error term in the loss function; Regularization coefficient: controls the penalty applied to the complexity of model parameters; The set of all trainable parameters of the model; : Regularization term to prevent overfitting.

[0163] The present invention further provides a system for multivariate time-series feature extraction and grade prediction in flotation processes, comprising:

[0164] The first module is used for raw data input and encoding; it encodes and structures time-series data composed of various process variables.

[0165] The second module is used to extract dynamic features; it uses a multi-layer one-dimensional convolutional network to extract dynamic features at different time scales, capturing the short-term fluctuations and long-term trends of process variables.

[0166] The third module is used for condition-guided coding modeling; it embeds the target concentrate grade information into the feature representation to form a guided query vector, which guides the subsequent modeling process to focus on the target-related features;

[0167] The fourth module is used to enhance the significance of key variables and important time segments; it employs a multi-head attention mechanism to match key-value pairs generated based on target-guided query vectors and multi-scale features.

[0168] The fifth module is used to output the prediction module; it performs feature fusion and nonlinear mapping on the output results of the attention mechanism to output the predicted results of concentrate grade and recovery rate at future time points.

[0169] 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.

[0170] 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 1The 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.

[0171] The detailed process of this method will be described below with reference to its implementation in a specific application.

[0172] Step S1000: Data Preparation

[0173] Data Acquisition: Real-time acquisition of multivariate time-series data from the flotation production line, including raw ore grade, tailings grade, particle size, collector dosage, frother dosage, modifier dosage, and inhibitor dosage.

[0174] Target data acquisition: Collect concentrate grade and recovery rate within the corresponding time period as the supervision target of the model.

[0175] Data cleaning: Remove outliers and missing data to ensure the integrity and validity of the training data.

[0176] Normalization: Normalize or standardize all input and target variables to eliminate differences in units and improve the stability and convergence speed of model training.

[0177] Data partitioning: Divide the data into training, validation and test sets to ensure the scientific nature of model training and evaluation.

[0178] Step S2000: Key parameters for model training

[0179] In this embodiment, to ensure the convergence speed, stability, and generalization ability of the model during the modeling process of floating multivariate time series data, the following key hyperparameters are set for the training phase:

[0180] First, the learning rate is set to 0.001, using an initial fixed value combined with a later dynamic adjustment strategy (such as learning rate decay or an adaptive scheduler) to adapt to the parameter convergence requirements at different training stages. This setting helps to quickly approach the optimal region in the early stages of training and maintain stable updates in the later stages, avoiding oscillations.

[0181] Secondly, the batch size was set to 64 to balance training efficiency and the model's generalization ability. This value demonstrates good stability and computational resource utilization in medium-scale industrial flotation data and is suitable for most practical production scenarios.

[0182] For the number of training epochs, the maximum number of training epochs is set to 100. The model performs a parameter update evaluation after each epoch that fully traverses the training set. An early stopping mechanism is also introduced. If the validation set loss does not decrease significantly for several consecutive epochs, the training process is terminated early to prevent overfitting.

[0183] In terms of model regularization, L2 regularization (weight decay) is adopted with a coefficient set to 1e-4, which effectively suppresses model overfitting and improves the model's generalization ability on unseen data.

[0184] To achieve sufficient modeling of key temporal features, a multi-head attention structure is used in the attention mechanism, with the number of attention heads set to 4. This setting allows the model to focus on information from different feature subspaces in parallel, enhancing its ability to identify complex temporal features.

[0185] In addition, the number of hidden dimensions (feature dimension) of each linear layer and attention layer in the model is uniformly set to 128 to maintain the consistency of data dimensions between modules and to balance representation capability and computational burden.

[0186] Regarding the optimizer, the Adam optimizer was selected for parameter updates, with its momentum factor parameters set to β1=0.9 and β2=0.999. It has adaptive gradient adjustment capability and can achieve faster convergence on non-stationary objective functions.

[0187] To prevent gradient explosion during training, a gradient clipping mechanism is introduced, with a clipping threshold of 5.0 to ensure the numerical stability of gradients during backpropagation.

[0188] The above hyperparameter configurations, based on experimental verification, can achieve stable and effective model performance in most typical flotation process scenarios. They can also be flexibly adjusted in practical applications according to the specific characteristics of the mineral, the data scale, or the deployment resources.

[0189] In this embodiment, the data used for model training comes from a typical flotation production line, covering multiple shifts and operating conditions. The total training data consists of approximately 45,000 records, each containing sampling points over 60 consecutive time steps. At each time point, seven process variables are recorded (raw ore grade, tailings grade, particle size, collector, frother, modifier, and depressant). Furthermore, each record corresponds to a monitoring label for concentrate grade and recovery rate.

[0190] In terms of data partitioning, the training set accounts for 70%, the validation set for 15%, and the test set for 15%, ensuring the temporal continuity of the data and preventing information leakage. All feature inputs are normalized using Z-score standardization before being fed into the model to ensure numerical balance of the input dimensions and improve model convergence efficiency.

[0191] Model training was performed on a workstation equipped with an NVIDIA RTX 4090 GPU (24GB VRAM). The training framework was implemented using PyTorch 2.1, and the Python version was 3.10. Under the above hardware environment, each training epoch took approximately 15-20 seconds, and the total training time for 100 epochs was approximately 30 minutes. The model achieved convergence on the validation set, and the training process was generally stable.

[0192] The model's performance on the test set shows that the mean square error (MSE) for predicting concentrate grade remains at a low level (e.g., below 0.004), and the prediction error for recovery rate is also controlled within a reasonable range, meeting the dual requirements of accuracy and response time in industrial scenarios.

[0193] The above training process and hardware configuration provide a feasibility guarantee for model deployment, while also having good transferability. Data can be retrained and transferred to learn according to different mining areas and different flotation conditions, and it has strong engineering practicality and scalability.

[0194] Based on the above examples, the present invention is tested and evaluated:

[0195] The following evaluation metrics were mainly used in this experiment to quantitatively assess the model performance:

[0196] Mean Squared Error (MSE)

[0197] Mean Absolute Error (MAE)

[0198] Coefficient of determination (R² Score)

[0199] The three indicators comprehensively reflect the model's ability to predict grade and recovery rate in the flotation process from the perspectives of error magnitude, deviation degree, and goodness of fit.

[0200] Comparative experimental setup

[0201] To verify the effectiveness of the method of the present invention, a comparison group was designed as shown in Table 1:

[0202] Table 1 Comparative Experiment Settings

[0203]

[0204] All methods use the same data partitioning, hyperparameter tuning, and training process to ensure fairness in the comparison.

[0205] Experimental Results and Analysis

[0206] Table 2. Prediction results of the four methods on the test set.

[0207]

[0208] Experimental results demonstrate that the multivariate temporal feature modeling method proposed in this invention exhibits significant advantages in flotation process prediction tasks. Compared with traditional temporal modeling methods (such as LSTM, Transformer encoders, and single-scale convolutional networks), this method shows a significant improvement in accuracy for both concentrate grade and recovery prediction tasks. Specifically, the model achieves the lowest mean squared error (MSE) and mean absolute error (MAE) on the test set, as well as the highest coefficient of determination (R²), fully demonstrating that the proposed structure has stronger predictive power and better generalization performance.

[0209] Firstly, in terms of feature extraction, this method introduces a multi-scale one-dimensional convolutional structure. Through a hierarchical design of the receptive field of the convolutional kernels, it effectively captures the dynamic changes in the process where short-term perturbations and long-term trends coexist. Experimental results show that this structure is particularly effective in processing complex flotation data containing multiple physical processes and perturbation frequencies, significantly outperforming traditional convolutional networks that only use a single scale.

[0210] Secondly, this invention incorporates a target-guided mechanism in its structural design. By embedding the target concentrate grade into a model and constructing an attention query vector together with the output of multiple convolutional layers, the model is guided to focus more on time-series segments and variable channels highly correlated with the target output when calculating attention weights. This design not only improves the interpretability of the model but also enhances its sensitivity to key variables (such as particle size, tailings grade, and collector dosage), thereby improving overall prediction performance.

[0211] Furthermore, in terms of inference efficiency, the average time for inferring a single sample is only 4.7 milliseconds, which meets the requirements of industrial control systems for real-time performance and high-frequency computation, demonstrating promising prospects for engineering deployment. Simultaneously, this method exhibits stable performance under different operating conditions and time periods, demonstrating strong robustness and cross-scenario transferability.

[0212] In summary, this invention establishes a structurally innovative, predictive, and fast-responding flotation process modeling method by integrating multi-scale convolutional feature extraction with a target-guided attention mechanism. This method not only theoretically achieves deep modeling of complex variable relationships in the flotation process but also demonstrates broad application potential and technological advantages in practical data testing.

[0213] 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 multivariate time-series feature extraction and grade prediction in flotation processes, characterized in that the steps include: include: Step S1: Input and encoding of raw data; Encode and structure the time series data composed of multiple process variables; Step S2: Extract dynamic features; By extracting dynamic features at different time scales through multi-layer one-dimensional convolutional networks, we can capture the short-term fluctuations and long-term trends of process variables. Step S3: Condition-guided coding modeling; The target concentrate grade information is embedded into the feature representation to form a guided query vector, which guides the subsequent modeling process to focus on the target-related features; The process of step S3 includes: Step S301: Adjust the target concentrate grade Embedded as a guiding vector through two layers of linear transformation: in, This is a characteristic that guides taste; Step S302: Output from each of the four convolutional layers. Extract the query vector: Step S303: After concatenating all query vectors with the target grade embedding vector, the final query input is formed: Step S304: Obtain the final query vector through linear transformation: in: This represents the query vector used for attention calculation; Step S4: Enhance the significance of key variables and important time segments; employ a multi-head attention mechanism to match key-value pairs generated based on the target-guided query vector and multi-scale features; The process of step S4 includes: Step S401: Define the multi-head attention structure: in: This indicates a bullish attention-based strategy; For query vector; The key vector; It is a value vector; Step S402: Attention mechanism through matching and To calculate the weights, and use those weights to... The output is obtained by weighted aggregation; Step S403: Perform the following operations within each attention head: in: The first The query, key, and value in the attention header; Let the key vector be the dimension of each attention head; Used to generate normalized attention weights; Step S404: The outputs of all heads are concatenated and then linearly transformed. Finally, the normalized output is obtained through layer normalization. in, This indicates that each sample is normalized. Step S5: Output prediction module; perform feature fusion and nonlinear mapping on the output results of the attention mechanism to output the predicted results of concentrate grade and recovery rate at future time.

2. The method for multivariate time-series feature extraction and grade prediction in flotation process according to claim 1, characterized in that, The process variables in step S1 include one or more of the following: raw ore grade, tailings grade, particle size, collector, frother, modifier, and inhibitor.

3. The method for multivariate time-series feature extraction and grade prediction in flotation process according to claim 1, characterized in that, The process in step S1 includes: The original input is a tensor , used to represent multivariate time-series data in the flotation process; in: Each sample sequence corresponds to an independent production data point within a batch; each time step contains Key process variables include raw ore grade, tailings grade, particle size, collector dosage, frother dosage, modifier dosage, and inhibitor dosage; Enter the target concentrate grade , serving as an auxiliary target variable to guide the model's focus direction; It is a scalar time-invariant objective used to provide guidance in subsequent feature modeling.

4. The method for multivariate time-series feature extraction and grade prediction in flotation process according to any one of claims 1-3, characterized in that, Step S2 introduces four layers of one-dimensional convolution operations to extract temporal features at different scales. The process includes: Step S201: Definition of convolution operation: in, This represents a one-dimensional convolution operation. This represents the step size parameter, which controls the sparsity of feature sampling. This indicates the size of the convolution kernel, which affects the receptive field of the model. Each sample sequence corresponds to an independent production data point within a batch; Step S202: Perform a linear mapping on the output of each convolutional layer: in, Indicates the first The feature representation of the output of a convolutional layer after linear transformation; Step S203: The outputs of all layers are concatenated to form the final depth feature representation: The output splicing refers to a splicing operation performed along the feature channel direction; specifically... These are feature maps from different layers, which are connected along the channel dimension while maintaining consistent spatial dimensions to obtain a new feature map. Its number of channels is the sum of the number of channels of all input feature maps; Step S204: Extract the key vector and value vector from the deep feature: in: This is the key vector used for matching; This is the value vector used for weighted fusion; both are obtained from the same deep feature through a linear transformation.

5. The method for multivariate time-series feature extraction and grade prediction in flotation process according to any one of claims 1-3, characterized in that, In step S5, after the multi-head attention mechanism, a final representation integrating multi-scale temporal features and target information is obtained, denoted as... This vector is used for joint prediction of flotation product indicators; Will Inputting data into a multilayer perceptron (MLP) to perform a joint estimation of future concentrate grade and recovery rate: in: This represents a feedforward neural network that includes several linear transformations and nonlinear activation functions. The model prediction results include two target indicators: concentrate grade and recovery rate.

6. The method for multivariate time-series feature extraction and grade prediction in flotation process according to any one of claims 1-3, characterized in that, Also includes: A joint loss function is constructed that comprehensively considers concentrate grade and recovery rate. This joint loss function uses the weighted mean square error (MSE) as the error metric, combined with regularization constraints on the model parameters. in, The overall loss function reflects the model's predictive performance and complexity. Total number of training samples; The model predicts the first The grade of concentrate in each sample; : No. The actual concentrate grade of each sample; The model predicts the first Recovery rate per sample; : No. True recovery rate of each sample; The weight of the concentrate grade error adjusts the influence of this error term in the loss function; The weight of the recovery rate error adjusts the influence of this error term in the loss function; Regularization coefficient: controls the penalty applied to the complexity of model parameters; The set of all trainable parameters of the model; : Regularization term to prevent overfitting.

7. A system for multivariate time-series feature extraction and grade prediction in a flotation process, used to perform the method of any one of claims 1-6, characterized in that, include: The first module is used for raw data input and encoding; Encode and structure the time series data composed of multiple process variables; The second module is used to extract dynamic features; By extracting dynamic features at different time scales through multi-layer one-dimensional convolutional networks, we can capture the short-term fluctuations and long-term trends of process variables. The third module is used for condition-guided coding modeling. The target concentrate grade information is embedded into the feature representation to form a guided query vector, which guides the subsequent modeling process to focus on the target-related features; The fourth module is used to enhance the significance of key variables and important time segments; it employs a multi-head attention mechanism to match key-value pairs generated based on target-guided query vectors and multi-scale features. The fifth module is used to output the prediction module; The output of the attention mechanism is subjected to feature fusion and nonlinear mapping to produce predictions of concentrate grade and recovery rate at future time points.

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 any one of the methods of claims 1-6.

Images