A two-way timing fusion beneficiation grade multi-step prediction method

Abstract

This invention relates to the field of intelligent mineral sorting technology and discloses a multi-step prediction method for mineral processing grade based on bidirectional temporal fusion using a binary splitting approach. The method includes: constructing a binary splitting interactive temporal fusion unit according to a binary splitting ratio, where each branch is constructed based on a feature extraction block. The feature extraction block is constructed in a pyramid network structure based on cross-attention, and the cross-attention uses one or more of long-term and short-term models and linear mappings; using the binary splitting interactive temporal fusion unit as an attention branch, and combining it with cross-attention to construct a pyramid-structured feature fusion module; constructing a prediction model based on the feature fusion module, combined with residual connections and a multilayer perceptron; and predicting mineral grade based on the prediction model using time-series data from a screen-high magnetic separation process. This method solves the problems commonly found in existing mineral processing grade prediction technologies, such as low preprocessing fidelity, insufficient multi-scale quality coupling modeling, poor feature interaction stability, and lack of training optimization mechanisms.

Description

Technical Field

[0001] This invention relates to the field of intelligent mineral sorting technology, and in particular to a multi-step method for predicting mineral grade through bidirectional time-series fusion of two-way flow and separation. Background Technology

[0002] In the field of mineral processing engineering, the screen-magnetic separation process is the core step in the purification and separation of metallic minerals. Concentrate grade and tailings metal recovery rate are core quality indicators for evaluating the quality of mineral processing products, production stability, and economic benefits. Due to the significant time lag in mineral processing, grade cannot be directly measured in real time by sensors, and offline testing suffers from severe delays. Therefore, high-precision, long-term, stable, and reliable grade prediction is a key technological prerequisite for achieving intelligent control of mineral processing, stable control of product quality, early warning of abnormal operating conditions, and reduction of metal loss.

[0003] Currently, the industry commonly employs traditional statistical methods and single-data-driven time-series prediction models for predicting mineral processing grade and quality. These models primarily include LSTM, GRU, time-series state-space models, ordinary attention fitting networks, and single-modal denoising prediction methods. These methods establish a mapping relationship between parameters and grade by collecting industrial data such as slurry concentration, feed particle size, feed flow rate, separation pressure, and equipment operating parameters, thereby achieving quality trend prediction.

[0004] However, under actual industrial variable operating conditions, the screen-magnetic separation process has inherent characteristics such as non-stationary fluctuations in the properties of raw ore, strong coupling of multiple process parameters, frequent equipment disturbances, and severe quality response lag. Existing mainstream time series prediction technologies have several unavoidable technical defects, resulting in low product quality prediction accuracy, poor stability, and weak generalization ability, which cannot meet the needs of refined industrial quality control. Existing mineral processing grade quality prediction technologies generally have inherent defects such as low preprocessing fidelity, large feature fusion loss, insufficient multi-scale quality coupling modeling, poor feature interaction stability, and lack of training optimization mechanisms. Summary of the Invention

[0005] This invention provides a multi-step prediction method for mineral processing grade based on bidirectional time-series fusion and splitting, in order to solve the problems that existing mineral processing grade quality prediction technologies generally suffer from, such as low preprocessing fidelity, large feature fusion loss, insufficient multi-scale quality coupling modeling, poor feature interaction stability, and lack of training optimization mechanisms.

[0006] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a multi-step method for predicting mineral processing grade using a two-way time-series fusion method, comprising the following steps: Step 1: Construct a bi-branch interactive temporal fusion unit with a multi-branch structure according to the bi-branch ratio. Each branch in the multi-branch structure is constructed based on feature extraction blocks combined with residual connections. The feature extraction blocks are constructed in the form of a pyramid network structure based on cross attention. The attention branches of the cross attention are selected from one or more of long-time series models, short-time series models, and linear mappings. Step 2: Using the binary split-flow interactive temporal fusion unit as the attention branch, a pyramid-structured feature fusion module is constructed by combining cross-attention. Based on the feature fusion module, a prediction model is constructed by combining residual connections and multilayer perceptron. Based on the prediction model, the mineral grade is predicted using the time series data of the screen-strong magnetic separation process.

[0007] Furthermore, the binary splitting and interactive temporal fusion unit performs linear processing on the input features, and then proportionally divides the linearly processed input features to obtain the first feature and the second feature; The first feature is input into the branch constructed based on the feature extraction block to obtain the enhanced feature, which is then concatenated with the unprocessed second feature and linearly mapped to obtain the third feature. The third feature is then added to the linearly processed input feature to obtain the output feature of the attention branch corresponding to the binary splitting interactive temporal fusion unit in the feature fusion module.

[0008] Furthermore, the feature extraction block uses a long-time series model, a short-time series model, and a linear mapping to process the first feature to obtain branch features. The branch features of the linear mapping are used as query feature inputs, and the branch features processed by the long-time series model and the short-time series model are used as key-value feature inputs. Then, cross-attention is combined to obtain the first attention feature and the second attention feature. The first attention feature and the second attention feature are concatenated to obtain the enhanced feature.

[0009] Furthermore, in the cross-attention of the feature extraction block, attention weights are generated based on branch features, and calibration weights are obtained by combining the attention weights with the value features and query feature inputs. The branch features output by the attention branch corresponding to the linear mapping are used as a benchmark, and the calibration weights are combined to calibrate the branch features output by the attention branches corresponding to the long-term model and the short-term model respectively.

[0010] Furthermore, the calibration weight obtained based on the attention weight combined with the value feature and the query feature input is expressed by the following formula: ; in, Indicates calibration weights; Indicates attention weight; Indicates value characteristics; This indicates the input of query features.

[0011] Furthermore, the prediction model preprocesses the time series data to obtain pure operating condition time series features, and then predicts the mineral grade based on the pure operating condition time series features.

[0012] Furthermore, the preprocessing sequentially includes outlier removal constraints and variational mode time-frequency decoupling; The variational mode time-frequency decoupling includes: performing time-frequency decoupling on the features to obtain the mid-frequency steady-state effective operating condition component and the high-frequency detail residual component; performing wavelet high-frequency noise purification on the high-frequency detail residual component and splicing it with the mid-frequency steady-state effective operating condition component to obtain the pure operating condition time series feature.

[0013] Through the above design, a multi-level collaborative noise reduction mechanism of outlier removal, time-frequency mode decoupling and layering, and wavelet residual fine purification is implemented. First, fault-tolerant correction of abnormal and abrupt data is performed, then noise components are stripped away layer by layer, and finally, pure time series features are reconstructed. This achieves the preprocessing of high-fidelity time series data under strong interference mineral processing conditions.

[0014] Furthermore, the prediction model uses a bisection-splitting interactive time-series fusion unit, linear mapping, and self-attention to process the pure working condition time-series features to obtain branch features. The branch features of the linear mapping are used as query feature inputs, and the branch features processed by the bisection-splitting interactive time-series fusion unit and self-attention are used as key-value feature inputs. Then, cross-attention is combined to obtain third and fourth attention features. The third and fourth attention features are then concatenated, linearly mapped, and residual connected in sequence, and finally, the mineral grade prediction result is obtained through a multilayer perceptron.

[0015] Furthermore, in the self-attention process, the query features are corrected based on the query features and key features to obtain accurate query features, and branch features corresponding to the self-attention are generated based on the accurate query features, key features, and value features.

[0016] Furthermore, obtaining precise query features includes the following steps: The query features and key features are sequentially expanded in dimension, concatenated, and mapped to obtain an adaptive matching matrix. The key features are sorted and filtered based on the adaptive matching matrix to obtain the first key feature and the second key feature. The query deviation is calculated based on the first key feature, the second key feature, and the query features using a deviation calculation function. The accurate query features are obtained based on the query deviation and the query features.

[0017] Through the above design, the method of manually constructing reference features such as mean pooling and fixed perceptron is abandoned. Instead, an adaptive matching matrix generated by query features and key features is used to adaptively select the globally optimal and individually suboptimal reference samples. This fully matches the inherent correlation features of the mineral processing time series samples, eliminates the feature offset problem caused by manual prior design, and improves the matching degree of working condition features.

[0018] Furthermore, the deviation calculation function is expressed by the following formula: ; in, Indicates the query deviation; Indicates query characteristics; Indicates the second bond feature; Indicates the first key feature; All of these are learnable constants constrained by the Sigmoid function.

[0019] Furthermore, the branch features generated based on precise query features, key features, and value features for self-attention include: The first attention weight is generated based on the accurate query features, key features, and feature dimensions. The adaptive matching matrix is ​​activated using the Sigmoid function, and then weighted and linearly processed with the first attention weight to generate the second attention weight. The value features are weighted based on the second attention weight to generate the branch features corresponding to the self-attention.

[0020] Through the above design, the adaptive matching matrix is ​​embedded into the complete attention computation link, which dynamically suppresses the weight ratio of invalid keys with low matching degree, allowing the model to accurately focus on the relevant features of effective working conditions in mineral processing, filter out redundant residual noise features, and greatly improve the quality of time series feature aggregation.

[0021] Furthermore, a multi-objective loss function is constructed, and the prediction model is optimized using the multi-objective loss function; The multi-objective loss function includes a prediction true value fitting error term, a temporal continuity smoothing constraint term, and an attention parameter steady-state regularization constraint term; The true value fitting error term is used to constrain the deviation between the multi-step grade prediction results and the actual process label. The temporal continuity smoothing constraint term is used to suppress non-process-related jumps and gradual accumulation of errors in long-term time-series predictions; The steady-state regularization constraint term of the attention parameter is used to constrain the parameter update magnitude in cross-attention.

[0022] Beneficial effects: This invention provides a bidirectional time-series fusion method for multi-step prediction of mineral processing grade, achieving high-fidelity noise reduction of industrial strong interference time series, fixed-ratio low-loss time series splitting, bidirectional benchmark calibration of long and short time series, double-layer residual anti-degradation fidelity, and multi-constraint process adaptation optimization. A dedicated time series prediction system adapted to the large lag, non-stationarity, multiple disturbances, and slow evolution characteristics of the screen-strong magnetic separation process is constructed, completely solving the technical bottlenecks of traditional models such as insufficient time series modeling, fragmentation of long and short features, severe detail loss, long-step error accumulation, and poor generalization ability under varying operating conditions. It achieves high-precision and highly stable long-time series prediction of mineral processing grade under all operating conditions, and sets up an adapted engineering solution for the specific scenario of the screen-strong magnetic separation process. By using a binary splitting mechanism, the computational optimization branch is strictly distinguished from the original detail retention branch which is frozen and does not perform any calculations. One branch completes multi-scale temporal iterative optimization, while the other branch retains the original working condition details for later compensation and repair, effectively solving the problem of information loss in temporal fusion from a structural perspective. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the network structure of the two-way splitting interactive timing fusion unit according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall network structure of the prediction model in an embodiment of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a," and similar terms, do not indicate a quantity limitation, but rather indicate the presence of at least one.

[0026] This application provides a multi-step method for predicting mineral processing grade using a two-way time-series fusion method, including the following steps: Step 1: Construct a multi-branch interactive temporal fusion unit with a bisection splitting structure according to the bisection ratio. Any branch in the multi-branch structure is constructed based on feature extraction blocks combined with residual connections. The feature extraction blocks are constructed in the form of a pyramid network structure based on cross attention. The attention branches of the cross attention are selected from one or more of long temporal models, short temporal models and linear mappings. Please see Figure 1 In the binary splitting and interactive temporal fusion unit, the input features are linearly processed, and then the linearly processed input features are proportionally split to obtain the first feature and the second feature. The first feature is input into the branch constructed based on the feature extraction block to obtain the enhanced feature and concatenate it with the unprocessed second feature. After linear mapping, the third feature is obtained. Then, the third feature is added to the linearly processed input feature to obtain the output feature of the attention branch corresponding to the binary splitting interactive temporal fusion unit in the feature fusion module. Specifically, the feature extraction block uses a long-term model, a short-term model, and a linear mapping to process the first feature to obtain branch features. The long-term model is exemplarily the SSM state-space model, while the short-term model is exemplarily the BiLSTM bidirectional long short-term memory network.

[0027] The branch features of the linear mapping are used as query feature inputs, and the branch features processed by the long-term and short-term models are used as key feature inputs respectively. In other words, the features of the steady-state baseline output by the linear mapping are used as query feature inputs, and the query features need to be obtained by performing another linear mapping on the baseline. Similarly, the long-period time-series features output by the long-term model are linearly mapped to obtain corresponding value and key features, and the short-period time-series features output by the short-term model are linearly mapped to obtain corresponding value and key features. Then, by combining cross-attention to obtain the first attention feature and the second attention feature, the first attention feature and the second attention feature are concatenated to obtain the enhanced feature; It should be noted that the above description uses a two-step operation of "first concatenating the first and second attention features to generate the enhanced feature, and then concatenating it with the unprocessed second feature," which is only for the purpose of illustrating the principle. In actual industrial deployments, a more efficient approach is to directly concatenate the first attention feature, the second attention feature, and the original second feature in parallel, and then obtain the third feature through linear mapping, thereby reducing intermediate steps and improving inference efficiency.

[0028] Specifically, in the cross-attention of the feature extraction block, attention weights are generated based on branch features. That is, attention weights are generated by combining query features, key features, and feature dimensions with the softmax function. Based on the attention weights, the value features and query features are combined to obtain calibration weights. The branch features output by the attention branch corresponding to the linear mapping are used as the benchmark, and the calibration weights are combined to calibrate the branch features output by the attention branches corresponding to the long-term model and the short-term model respectively to obtain the first attention feature and the second attention feature. The calibration weights are obtained using the following formula: ; in, Indicates calibration weights; Indicates attention weight; Indicates value characteristics; Indicates the query feature input; Step 2: Using the binary split-flow interactive temporal fusion unit as the attention branch, a pyramid-structured feature fusion module is constructed by combining cross-attention. Based on the feature fusion module, a prediction model is constructed by combining residual connections and multilayer perceptron. Based on the prediction model, the mineral grade is predicted using the time series data of the screen-strong magnetic separation process.

[0029] Please see Figure 2 In the prediction model, the prediction model preprocesses the time series data of the screen-strong magnetic separation process to obtain the time series characteristics of the pure working condition, and then predicts the mineral grade based on the time series characteristics of the pure working condition. The preprocessing includes outlier removal constraints and variational mode time-frequency decoupling. The outlier removal constraints are common to industrial scenarios. Anomaly detection constraints remove data points from time series data whose absolute deviation from the mean is greater than 3 times the standard deviation. Variational mode time-frequency decoupling involves decoupling the features to obtain the mid-frequency steady-state effective operating condition component and the high-frequency detail residual component. After wavelet high-frequency noise purification of the high-frequency detail residual component, it is spliced ​​with the mid-frequency steady-state effective operating condition component to obtain the pure operating condition time series feature. The preprocessing of the prediction model is achieved through a structured chain of anomaly-tolerant initial screening → time-frequency layered decoupling → residual fine noise reduction → effective component reconstruction. This process peels away industrial impulse noise and time-frequency aliasing interference layer by layer at the physical level, fully preserving the real working condition evolution characteristics of screen clogging, material adhesion hysteresis, and magnetic separation parameter fluctuations. This solves the shortcomings of traditional single noise reduction methods, such as incomplete noise removal and easy loss of effective features, and significantly improves the model's generalization ability under different working conditions.

[0030] The prediction model uses a bisection-splitting interactive time series fusion unit, linear mapping, and self-attention to process the pure working condition time series features to obtain branch features. The branch features of linear mapping are used as query feature inputs. The branch features processed by the bisection-splitting interactive time series fusion unit and self-attention are used as key value feature inputs. Then, cross-attention is combined to obtain third and fourth attention features. The third and fourth attention features are then concatenated, linearly mapped, and residual connected in sequence. Finally, the mineral grade prediction results are obtained through a multilayer perceptron. To address the current self-attention mechanism's reliance on backpropagation for iterative updates of query and key features, which results in a limited optimization space and susceptibility to local optima, and the use of fixed inner products to calculate feature similarity fails to adaptively quantify the matching relationship between query and key features based on mineral processing time-series conditions. This leads to a large number of low-association, invalid key vectors continuously participating in weight allocation, introducing feature noise, masking effective processing-related information, and resulting in insufficient model feature extraction accuracy, poor multi-variable coupled feature representation capabilities under complex conditions, and severe accumulation of long-step prediction errors. This embodiment proposes an adaptive matching matrix based on query and key features to optimize these issues, as detailed below: In the self-attention branch, the query features and key features are sequentially expanded in dimension, concatenated, and mapped to obtain an adaptive matching matrix. Based on the adaptive matching matrix, the key features are sorted and filtered to obtain the first key feature and the second key feature. Based on the first key feature, the second key feature, and the query features, the query deviation is calculated using the deviation calculation function. Based on the query deviation and the query features, the accurate query features are obtained. The first attention weight is generated based on the precise query features, key features, and feature dimensions. The adaptive matching matrix is ​​activated using the Sigmoid function. The first attention weight is then weighted and linearly processed to generate the second attention weight. The value features are then weighted based on the second attention weight to generate the branch features corresponding to the self-attention. The deviation calculation function is expressed by the following formula: ; in, Indicates the query deviation; Indicates query characteristics; Indicates the second bond feature; Indicates the first key feature; All are learnable constants constrained by the Sigmoid function; By using a closed-loop constraint of adaptive matching matrix, deviation calculation constraint, and feature offset correction, combined with learnable constants that are continuously iterated and updated during model training, the feature update amplitude is dynamically constrained to avoid feature mutations caused by operating condition fluctuations, thus ensuring stable and controllable parameter updates.

[0031] The subsequent generation of the second attention weights, which introduced adaptive matching weights, achieved a hierarchical closed loop, completely solving the problems of rigid matching, limited parameter optimization, and interference from invalid features in traditional attention, and accurately adapting to the dynamic coupling of multiple variables and time-lag conditions in mineral processing.

[0032] For training the prediction model, a multi-objective loss function is constructed and used to optimize the parameters of the prediction model. The multi-objective loss function includes a prediction true value fitting error term, a temporal continuity smoothing constraint term, and an attention parameter steady-state regularization constraint term. The multi-objective loss function is expressed by the following formula: ; in, This represents the loss function for end-to-end training of the model; This represents the error term in the true value fitting. This represents the temporal continuity smoothing constraint term; This represents the steady-state regularization constraint term for the attention parameter; Represents the time-series smoothing weighting coefficients, preferably ; Represents the regularization weight coefficient of the parameter, preferred. .

[0033] The true value fitting error term is used to constrain the deviation between the multi-step grade prediction results and the actual process label, and is expressed by the following formula: ; in, This refers to the number of training batches. The total time series step size for multi-step prediction; This represents the predicted mineral processing grade output by the model. These are the actual grade label values ​​collected on-site. This ensures the model's basic ability to fit the evolution of mineral processing timelines and reduces static prediction bias. The temporal continuity smoothing constraint term is used to suppress non-process-related jumps and gradual error accumulation in long-term forecasts, and is expressed by the following formula: ; in, To predict the grade output value for the next time step; Predict the grade output value for the current moment. Force a smooth transition in the gradient of the continuous prediction results, which conforms to the slow and continuous changes in the mineral processing technology and effectively suppresses the oscillation distortion of long-step prediction; The steady-state regularization constraint term for the attention parameter is used to constrain the parameter update magnitude in cross-attention, and is expressed by the following formula: ; in, To preserve learnable coefficients of the original features of the query features, the preferred approach is to... ; For individual instantaneous operating condition deviation correction coefficients, the preferred ones are... ; The optimal global steady-state process deviation correction coefficient is... ; The overall correction step size control coefficient for query features is preferred. ; This is an L2 norm regularization operation. This structure restricts the iterative update range of the attention correction parameters, stabilizes the correction mechanism of dual-reference query features, and significantly improves the model's generalization ability and training convergence stability under varying mineral properties and equipment disturbance conditions.

[0034] To verify the proposed bidirectional time-series fusion method for multi-step prediction of mineral processing grade and the corresponding prediction model, multi-source time-series data from a real-world mineral processing plant screen-high-intensity magnetic separation scenario were used to perform predictions in four schemes, including this embodiment, as follows: Comparison Scheme 1: Conventional MLP fully connected prediction scheme, which adopts the traditional multilayer perceptron structure, relies only on single-layer global feature mapping to achieve quality fitting, without multi-scale temporal feature extraction, long and short temporal interaction calibration, and attention adaptive correction mechanism, and only completes temporal prediction through shallow data mapping.

[0035] Comparison Scheme 2: Single time series network prediction scheme, which adopts the traditional LSTM / SSM single-branch time series prediction structure, has basic long and short time series fitting capabilities, but does not build a bi-splitting low-loss fusion mechanism, and does not have a dual-reference attention bias correction structure. It cannot achieve bi-directional constraints on long and short time series, and has the defects of single time series characteristics, weak adaptability to operating conditions, and serious accumulation of long step errors.

[0036] Comparison Scheme 3: The hybrid time series prediction scheme without bidirectional calibration adopts a multi-branch time series fusion fitting method and has the ability to extract features at multiple scales. However, it does not set a fixed ratio bidirectional splitting fidelity structure, does not have a shallow benchmark bidirectional mutual calibration mechanism, and does not have a multi-constraint composite target optimization strategy. The feature fusion loss is large, the time series coupling modeling is insufficient, and the generalization ability is insufficient under complex fluctuation conditions.

[0037] Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination ( ) As core evaluation indicators, the mean absolute error (MAE) characterizes the average deviation between the model's predicted grade and the actual process grade, directly reflecting the prediction accuracy; the root mean square error (RMSE) focuses on characterizing the cumulative error level of long-term predictions, reflecting the model's stability; the coefficient of determination (CME) is used as the core evaluation indicator. The model is used to assess the fit of the multivariate temporal coupling and process lag patterns, reflecting the feature modeling capability under complex working conditions. The test results are shown in Table 1.

[0038] Table 1: Horizontal comparison of core evaluation indicators between the comparative scheme and the scheme in this embodiment

[0039] All three traditional comparative schemes exhibit varying degrees of deficiencies in time series modeling and limitations in adapting to different operating conditions. Specifically, the conventional MLP fully connected prediction scheme can only learn shallow static correlation features, failing to capture the temporal lag coupling patterns and dynamic fluctuation characteristics of the mineral processing process. Its ability to mine multi-scale time series information is insufficient, resulting in the highest MAE and RMSE error values. The single-branch time-series network prediction scheme exhibits the lowest goodness of fit and the worst overall prediction accuracy and adaptability to operating conditions. While possessing basic time-series memory capabilities and improving accuracy compared to static fitting models, it is limited by the single-branch feature extraction structure, failing to distinguish between long-cycle mineral evolution patterns and short-cycle equipment disturbance characteristics. Interference between long and short time-series features leads to significant accumulation of long-step prediction errors, limiting fitting accuracy under complex and fluctuating operating conditions. Although the hybrid time-series prediction scheme without bidirectional calibration introduces multi-branch feature fusion logic and achieves preliminary extraction of multi-scale time-series information, it lacks a low-loss, fidelity-preserving mechanism for bidirectional splitting and shallow benchmark mutual calibration constraints. This results in severe detail loss during time-series fusion, the inability of long and short time-series features to mutually correct each other, insufficient modeling of process lag coupling relationships, and a significant shortcoming in generalization ability under strong disturbance conditions.

[0040] Compared to the three existing technical solutions, the bidirectional interactive temporal fusion modeling method proposed in this invention has significant technical advantages and performance improvements. First, this invention uses a fixed-ratio bidirectional isolation computational structure to distinguish between optimized computational branches and original detail retention branches, maximizing the preservation of subtle, slowly changing details in mineral processing conditions and effectively reducing information loss during the temporal fusion process, thus structurally improving the completeness of feature representation. Second, through a three-scale parallel temporal extraction mechanism, it simultaneously covers three core temporal features: shallow steady-state benchmarks, long-cycle mineralization trends, and short-cycle equipment fluctuations, comprehensively adapting to the multi-dimensional dynamic characteristics of mineral processing conditions. Simultaneously, relying on a self-developed lightweight residual cross-attention mechanism to construct a bidirectional temporal calibration mechanism with shallow benchmark constraints, it achieves bidirectional correction and complementary constraints of long and short temporal features, completely solving the defects of traditional models such as temporal fragmentation and insufficient lag modeling, and accurately fitting the lag coupling correlation law of the screen-magnetic separation process. Finally, through the collaborative optimization of inner and outer double-layer residual fault-tolerant structures and multi-constraint composite objective functions, it effectively suppresses the accumulation of long-step prediction errors, improving model training stability and generalization ability under varying conditions.

[0041] Experimental data show that the MAE and RMSE error in the proposed method are significantly lower than those of existing comparative methods, and the prediction bias is greatly reduced. The fitting coefficient reached 0.967, which significantly improved the time series fitting accuracy and working condition adaptability compared with the traditional time series model. It can maintain high accuracy and high stability in grade prediction under stable working conditions, process fluctuation conditions, and strong equipment disturbance conditions. This fully verifies the effectiveness and engineering superiority of the multi-scale time series modeling, bidirectional feature calibration and multi-constraint optimization mechanism of the present invention, and fully meets the production needs of high accuracy and long-term stable prediction under all working conditions in industrial mineral processing.

[0042] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A beneficiation grade multi-step prediction method of two-way time fusion of two-way time series, characterized by, Includes the following steps: Step 1: Construct a bi-branch interactive temporal fusion unit with a multi-branch structure according to the bi-branch ratio. Each branch in the multi-branch structure is constructed based on feature extraction blocks combined with residual connections. The feature extraction blocks are constructed in the form of a pyramid network structure based on cross attention. The attention branches of the cross attention are selected from one or more of long-time series models, short-time series models, and linear mappings. Step 2: Using the binary split-flow interactive temporal fusion unit as the attention branch, a pyramid-structured feature fusion module is constructed by combining cross-attention. Based on the feature fusion module, a prediction model is constructed by combining residual connections and multilayer perceptron. Based on the prediction model, the mineral grade is predicted using the time series data of the screen-strong magnetic separation process.

2. The method according to claim 1, wherein, The binary splitting and interactive temporal fusion unit performs linear processing on the input features, and then proportionally divides the linearly processed input features to obtain the first feature and the second feature. The first feature is input into the branch constructed based on the feature extraction block to obtain the enhanced feature, which is then concatenated with the unprocessed second feature and linearly mapped to obtain the third feature. The third feature is then added to the linearly processed input feature to obtain the output feature of the attention branch corresponding to the binary splitting interactive temporal fusion unit in the feature fusion module.

3. The method according to claim 2, wherein, The feature extraction block uses a long-time series model, a short-time series model, and a linear mapping to process the first feature to obtain branch features. The branch features of the linear mapping are used as query feature inputs. The branch features processed by the long-time series model and the short-time series model are used as key-value feature inputs. Then, cross-attention is combined to obtain the first attention feature and the second attention feature. The first attention feature and the second attention feature are concatenated to obtain the enhanced feature.

4. The method according to claim 3, wherein, In the cross-attention of the feature extraction block, attention weights are generated based on branch features. Calibration weights are obtained by combining the attention weights with the value features and the query feature input. The branch features output by the attention branch corresponding to the linear mapping are used as a benchmark and the calibration weights are combined to calibrate the branch features output by the attention branch corresponding to the long-term model and the short-term model, respectively.

5. The method according to claim 4, wherein, The calibration weight, obtained by combining attention weights with value features and query features, is expressed by the following formula: ； wherein, denotes a calibration weight; denotes an attention weight; denotes a value feature; denotes a query feature input.

6. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 1, characterized in that, The prediction model preprocesses the time series data to obtain pure operating condition time series features, and then predicts the mineral grade based on the pure operating condition time series features.

7. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 6, characterized in that, The preprocessing includes outlier removal constraints and variational mode time-frequency decoupling in sequence; The variational mode time-frequency decoupling includes: performing time-frequency decoupling on the features to obtain the mid-frequency steady-state effective operating condition component and the high-frequency detail residual component; performing wavelet high-frequency noise purification on the high-frequency detail residual component and splicing it with the mid-frequency steady-state effective operating condition component to obtain the pure operating condition time series feature.

8. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 6, characterized in that, The prediction model uses a binary split-flow interactive time series fusion unit, linear mapping, and self-attention to process the pure working condition time series features to obtain branch features. The branch features of the linear mapping are used as query feature inputs, and the branch features processed by the binary split-flow interactive time series fusion unit and self-attention are used as key value feature inputs. Then, cross-attention is combined to obtain third and fourth attention features. The third and fourth attention features are then concatenated, linearly mapped, and residual connected in sequence. Finally, the mineral grade prediction result is obtained through a multilayer perceptron.

9. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 8, characterized in that, The self-attention process corrects the query features based on query features and key features to obtain accurate query features, and generates branch features corresponding to the self-attention based on the accurate query features, key features, and value features.

10. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 9, characterized in that, The acquisition of precise query features includes the following steps: The query features and key features are sequentially expanded in dimension, concatenated, and mapped to obtain an adaptive matching matrix. The key features are sorted and filtered based on the adaptive matching matrix to obtain the first key feature and the second key feature. The query deviation is calculated based on the first key feature, the second key feature, and the query features using a deviation calculation function. The accurate query features are obtained based on the query deviation and the query features.

11. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 10, characterized in that, The deviation calculation function is expressed by the following formula: ； in, Indicates the query deviation; Indicates query characteristics; Indicates the second bond feature; Indicates the first key feature; All of these are learnable constants constrained by the Sigmoid function.

12. The multi-step prediction method for mineral processing grade based on bidirectional time-series fusion of two-way flow separation according to claim 10, characterized in that, The branch features generated based on precise query features, key features, and value features include: The first attention weight is generated based on the accurate query features, key features, and feature dimensions. The adaptive matching matrix is ​​activated using the Sigmoid function, and then weighted and linearly processed with the first attention weight to generate the second attention weight. The value features are weighted based on the second attention weight to generate the branch features corresponding to the self-attention.

13. The method for multi-step prediction of mineral processing grade by bidirectional time-series fusion of two-way flow separation according to claim 1, characterized in that, Construct a multi-objective loss function and use it to optimize the parameters of the prediction model; The multi-objective loss function includes a prediction true value fitting error term, a temporal continuity smoothing constraint term, and an attention parameter steady-state regularization constraint term; The true value fitting error term is used to constrain the deviation between the multi-step grade prediction results and the actual process label. The temporal continuity smoothing constraint term is used to suppress non-process-related jumps and gradual accumulation of errors in long-term time-series predictions; The steady-state regularization constraint term of the attention parameter is used to constrain the parameter update magnitude in cross-attention.

Images