Flotation grade prediction method based on dynamic hypergraph and Informer-GRU

Abstract

The invention discloses a flotation grade prediction method based on a dynamic hypergraph and Informer-GRU, belongs to the technical field of industrial process intelligent monitoring, and solves the technical problems that in the complex flotation process, multi-source heterogeneous variables cannot be integrated, and the high-order space-time dependency relationship cannot be effectively captured. According to the technical scheme, the method is characterized in that a DHIG unified framework is constructed, and dynamic hypergraph modeling and an Informer-GRU network are integrated; designing a self-adaptive hypergraph construction mechanism based on a sliding time window, and realizing effective representation of evolution correlation among multi-source heterogeneous variables; capturing a long-range dependency relationship by adopting a sparse self-attention mechanism of an Informer module, and processing short-term fluctuation caused by operation change in combination with a GRU component; and multi-scale feature fusion is realized through a cascade Informer-GRU structure. The method is mainly used for intelligent monitoring and grade prediction of the flotation process, the prediction precision and the calculation efficiency of the method are superior to those of an existing method in an industrial grade flotation data set test, and a technical support is provided for process optimization in a complex industrial environment.

Description

Technical Field

[0001] The present invention relates to grade prediction in a mine flotation process, and in particular to a flotation grade prediction method based on a dynamic hypergraph and an Informer-GRU architecture. Background Art

[0002] The continued decline in ore grade has become a global challenge, with China's ore reserves generally below the global average. To meet the growing demand for high-quality ore materials, China has stepped up its efforts in the research and development of flotation technology. However, the primary goal of flotation technology remains to improve ore recovery, especially in the separation of low-grade ores, which remains a challenge. During the flotation process, multiple factors such as ore grade, flotation reagent concentration, and equipment performance directly influence the separation effect. When processing low-grade ores, traditional flotation technology often struggles to maintain stable and efficient flotation efficiency due to the small differences in hydrophobicity between minerals.

[0003] Currently, flotation grade monitoring and prediction primarily rely on manual judgment of froth quality and traditional analytical instruments (such as X-ray fluorescence analyzers). These methods not only limit detection accuracy and consistency but are also susceptible to operator experience and subjective factors, resulting in poor flotation process stability and difficulty maintaining consistent product quality. Furthermore, the flotation process exhibits significant nonlinear and time-varying characteristics, with complex dynamic coupling relationships between process parameters such as feed grade, flotation reagent concentration, and bubble characteristics. These interplays complicate flotation grade prediction, further highlighting the need for intelligent and accurate flotation process prediction.

[0004] To improve the accuracy of flotation grade prediction, numerous studies have focused on multi-source data fusion and the application of machine learning models. While these approaches have achieved some progress in certain areas, they still face significant challenges in handling the dynamic nature of the flotation process and complex, multi-source, heterogeneous data. In particular, the prediction accuracy of existing models remains insufficient when dealing with long sequences of data and capturing temporal dependencies.

[0005] Therefore, there is an urgent need for a novel method that can better cope with the complex data in the flotation process, provide accurate grade prediction, and optimize the flotation process. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a flotation grade prediction method based on dynamic hypergraph and Informer-GRU. By introducing the dynamic hypergraph model and Informer-GRU architecture, this method overcomes the shortcomings of traditional methods in processing complex data in the flotation process, can effectively capture the dynamic characteristics and nonlinear dependencies in the flotation process, and improve the accuracy of grade prediction.

[0007] The technical solutions of the invention are as follows: A flotation grade prediction method based on dynamic hypergraph and Informer-GRU includes the following steps: Step 1: preprocessing multi-source flotation process data, including bubble feature data and chemical composition data extracted from flotation foam images, to obtain normalized multivariate time series data through outlier processing, missing value processing, and feature normalization; Step 2: A dynamic hypergraph is constructed using a sliding window mechanism. The correlation matrix is ​​constructed by calculating the Pearson correlation coefficient between features within the time window, and a hyperedge set is dynamically generated based on an adaptive threshold mechanism. Step 3: Design a dynamic weight calculation module to perform weighted processing on the hypergraph; Step 4: The dynamic hypergraph represents the hypergraph structure through the association matrix and performs hypergraph convolution operation to achieve feature aggregation; Step 5: Construct an Informer-GRU cascade time series prediction network. The Informer module uses a probabilistic sparse self-attention mechanism to extract global trend features, and the GRU module learns local time series patterns and outputs flotation grade prediction results.

[0008] As a further solution of the present invention, the multi-source flotation process data in step 1 specifically includes image feature data, chemical composition data and prediction targets, wherein the image feature data includes bubble size mean and variance, contrast mean, correlation mean, energy mean, homogeneity mean, bubble velocity mean and variance, bubble rupture area and volume mean; the chemical composition data includes Fe content, Zn content and Pb content in the original ore; and the prediction target is the marked concentrate zinc grade content.

[0009] The specific operations of preprocessing flotation data in step 1 include: outlier processing, missing value processing, and feature normalization. The data sample is shown in Table 1, and the basic process is as follows: Table 1 Example of flotation grade data sample (after Z-value normalization) (1) (2) use In principle, if the actual value Data points that differ from the sample mean by more than two standard deviations are considered outliers and are removed; is the average value, is the number of values ​​contained in a channel; is the standard deviation; When a feature value in the data set is missing, it is filled using the interpolation method of the adjacent time steps; for isolated missing points, the average value of the adjacent valid data before and after is used to fill it: (3) In the formula is where the missing value is located, and It is the valid data before and after the missing value. For the case of continuous missing values, multiple interval differences are used to fill in the missing values.

[0010] The data set is normalized and the model is: (4) In the formula is the original data value, and are the minimum and maximum values ​​in the dataset, respectively. is the normalized data value.

[0011] As a further solution of the present invention, the node definition of the dynamic hypergraph in step 2 takes the time series of each feature as a node in the graph; the node feature is represented as: (5) Where, For nodes At time step The feature representation of is the sliding window size.

[0012] The Pearson correlation coefficient calculation described in step 2 uses the data segments in the sliding window to implement dynamic correlation analysis; the dynamic correlation coefficient calculation formula is: (6) The hyperedge generation described in step 2 adopts a correlation-driven rule; the hyperedge generation rule is: (7) Where, Represents the time step The hyperedge set of is a feature subset, is a node set, Features and The correlation coefficient between is the threshold parameter.

[0013] As a further solution of the present invention, the dynamic weight calculation described in step 3 is based on the intra-superedge feature correlation; the weight calculation formula is: (8) Where, For super edge The weight of For super edge The contained node collection.

[0014] As a further solution of the present invention, the dynamic hypergraph in step 4 represents the hypergraph structure through an association matrix; the association matrix is ​​defined as: (9) Where, is the time step The correlation matrix, represents the node index, Represents the hyperedge index.

[0015] As a further solution of the present invention, the hypergraph convolution operation described in step 4 realizes feature aggregation through the correlation matrix; the convolution operation formula is: (10) Where, is the activation function, and are the node degree matrix and the hyperedge degree matrix respectively, is the incidence matrix, is the weight matrix. Represents the total number of convolutional layers, and the first convolution input is the hypergraph node feature: , No. The input features of the layer are the output features of the previous layer: .

[0016] As a further solution of the present invention, the Informer module described in step 5 adopts an encoder-decoder architecture, and both the encoder and decoder use the ProbSparse self-attention mechanism; the input and output of the Informer module are defined as: (11) (12) (13) Where, represents the encoder input sequence length and the decoder history target sequence length, is the global feature of compression.

[0017] The ProbSparse self-attention mechanism described in step 5 reduces the computational complexity by probabilistic sparseness, The complexity is optimized to ,in For the sequence length, efficient modeling of long sequence temporal dependencies is achieved.

[0018] The cascade structure described in step 5 adopts the sequence arrangement of Informer in front and GRU in the back. The output of the Informer module It is directly used as the input of the GRU network. The GRU learns local time series patterns and outputs prediction results based on global trend features.

[0019] The GRU network described in step 5 implements hidden state updates through the reset gate, update gate, and candidate state; the calculation formula of the GRU network is: (14) (15) (16) (17) Where, for activation function, is the hyperbolic tangent activation function, Represents element-wise multiplication.

[0020] The dataset is divided into a training set and a test set. The training set is used to test the model performance, and the test set is used for verification. The root mean square error (RMSE) and mean absolute error (MAE) are used as indicators to evaluate the model performance.

[0021] The beneficial effects of the present invention are as follows: The present invention proposes a flotation grade prediction method based on DHIG (dynamic hypergraph and Informer-GRU). By combining the dynamic hypergraph architecture with the Informer-GRU model, the prediction accuracy of flotation grade is significantly improved. In particular, during the flotation process, the method can effectively model various spatiotemporal characteristics of the flotation process and capture the complex dynamic relationships between complex multi-source features. Compared with traditional models, the method has stronger adaptability and accuracy. At the same time, the method exhibits superior performance on different types of data sets, especially on highly nonlinear, highly periodic, and highly random data, showing strong robustness and stability. Finally, the method provides a high-precision, efficient, and robust solution for the flotation grade prediction task. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a basic diagram of the structure of the present invention; Figure 2 The present invention is based on the process of integrating multi-source data features of the dynamic hypergraph; Figure 3Flowchart of the DHIG algorithm constructed for the present invention; DETAILED DESCRIPTION The following will be combined Figure 1-Figure 3 The technical solution of the present invention is described in further detail.

[0023] Example 1 This invention addresses the strong coupling and nonlinear relationships between parameters in the flotation process. Traditional methods, when processing multi-source heterogeneous data, are prone to low prediction accuracy, the optimization process easily falling into local optimal solutions, and excessive training time. To address this, the present invention proposes an innovative dynamic hypergraph structure, combined with an Informer-GRU network, that efficiently captures spatiotemporal features and integrates the complex relationships between multi-source features, thereby achieving accurate flotation grade prediction.

[0024] A flotation grade prediction method based on dynamic hypergraph and Informer-GRU Figure 1 As shown in the figure, data preprocessing is performed based on multi-source data from the flotation process (bubble feature data and chemical composition data extracted from flotation foam images). This process includes outlier handling, missing value filling, and feature normalization to ensure data quality and standardize the input data. Next, a sliding window mechanism is used to segment the data. The Pearson correlation coefficient between features within each time window is calculated to construct a correlation matrix between features. A dynamic hyperedge set is generated based on an adaptive threshold mechanism, thereby constructing a dynamic hypergraph. A dynamic weight calculation module is then designed to assign weights to each hyperedge in the hypergraph based on the correlation between features. This weighted algorithm emphasizes the influence of important features. Next, a convolution operation is performed on the dynamic hypergraph to aggregate local information from each feature. Finally, an Informer-GRU cascade network is used to predict flotation grade. The Informer module uses a sparse self-attention mechanism to extract global trend features, while the GRU module learns local temporal patterns. The combination of the two effectively captures long-term trends and short-term fluctuations in the flotation process, outputting the final flotation grade prediction results. Error analysis and result optimization further improve the model's prediction accuracy and practical application performance.

[0025] The specific implementation steps of a flotation grade prediction method based on dynamic hypergraph and Informer-GRU are as follows: Step S1: The flotation process involves complex multi-source data, including bubble size mean and variance, contrast mean, correlation mean, energy mean, homogeneity mean, bubble velocity mean and variance, bubble rupture area and volume mean. Chemical composition data include Fe content, Zn content, and Pb content in the feed grade. The prediction target is the marked zinc grade content. In order to integrate the multi-source data and construct a dynamic hypergraph, this method will preprocess the data set. The specific operations for preprocessing flotation data include: outlier processing, missing value processing, and feature normalization. The data sample is shown in Table 1, and the basic process is as follows: Table 1 Example of flotation grade data sample (after Z-value normalization) (18) (19) use In principle, if the actual value Data points that differ from the sample mean by more than two standard deviations are considered outliers and are removed; is the average value, is the number of values ​​contained in a channel; is the standard deviation; When a feature value in the data set is missing, it is filled using the interpolation method of adjacent time steps. For isolated missing points, the average value of the adjacent valid data before and after the midpoint is used to fill it.

[0026] (20) in and Is a missing value For the cases of continuous missing data, multi-segment difference is used to fill in the missing data.

[0027] The data set is normalized and the model is: (twenty one) Among them is the original data value, and are the minimum and maximum values ​​in the dataset, respectively. is the normalized data value.

[0028] Each dataset is divided into training set, validation set and test set in the ratio of 12:3:5.

[0029] Step S2: Define the nodes of the dynamic hypergraph and take the time series of each feature as a node in the graph; the node feature is represented as: (twenty two) in, For nodes At time step The feature representation of is the sliding window size.

[0030] The Pearson correlation coefficient is calculated using the data segments within the sliding window to achieve dynamic correlation analysis; the dynamic correlation coefficient calculation formula is: (twenty three) By subsetting the features There are strong pairwise correlation features in the , to construct a hyperedge; the hyperedge generation rule is: (twenty four) in, Represents the time step The hyperedge set of is a feature subset, is a node set, Features and The correlation coefficient between is the threshold parameter.

[0031] Step S3: Dynamic weight calculation is based on the correlation of features within the hyperedge; the weight calculation formula is: (25) in, For super edge The weight of For super edge The contained node collection.

[0032] The dynamic hypergraph represents the hypergraph structure through an incidence matrix; the incidence matrix is ​​defined as: (26) in, is the time step The correlation matrix, represents the node index, Represents the hyperedge index.

[0033] Step S4, as Figure 2 As shown in Figure 1, hypergraph convolution operation is used to perform feature aggregation to obtain a comprehensive representation of multi-dimensional features. The convolution operation formula is: (27) in, is the activation function, and are the node degree matrix and the hyperedge degree matrix respectively, is the incidence matrix, is the weight matrix. Represents the total number of convolutional layers, and the first convolution input is the hypergraph node feature: , No. The input features of the layer are the output features of the previous layer: .

[0034] Step 5, such as Figure 3 As shown, an Informer-GRU cascade time series prediction network is constructed. The Informer module extracts global trend features through the probabilistic sparse self-attention mechanism, and the GRU module learns local time series patterns, and finally outputs the flotation grade prediction result. The specific steps of step 5 are as follows; The Informer module adopts an encoder-decoder architecture, and both the encoder and decoder use the ProbSparse self-attention mechanism; the input and output of the Informer module are defined as: (28) (29) (30) in, represents the encoder input sequence length and the decoder history target sequence length, is the global feature of compression.

[0035] The cascade structure adopts the sequence arrangement of Informer in front and GRU in the back. The output of the Informer module It is directly used as the input of the GRU network. The GRU learns local time series patterns and outputs prediction results based on global trend features.

[0036] The GRU network described in step 5 implements hidden state updates through the reset gate, update gate, and candidate state; the calculation formula of the GRU network is: (31) (32) (33) (34) in, for activation function, is the hyperbolic tangent activation function, Represents element-wise multiplication.

[0037] In order to verify the accuracy and effectiveness of the DHIG method proposed in this invention for predicting concentrate grade in flotation process, the root mean square error (RMSE) and mean absolute error (MAE) were used to evaluate the model performance.

[0038] The comparison results of the indicators obtained by using different algorithms to predict the zinc concentrate grade are shown in Table 1: Table 2 Comparison of different methods for zinc concentrate grade prediction of flotation data sets As can be seen from Table 2, on the flotation dataset, DHIG shows significant improvements: compared with CNN-GRU, MAE is reduced by (from 0.3933 to 0.1549), which is lower than that of Transformer. Compared with HGNN-GRU, it reduces Most notably, DHIG surpasses the state-of-the-art Informer model with a MAE reduction of , RMSE is reduced , demonstrating enhanced capabilities in capturing complex spatiotemporal dependencies. Stability analysis shows that the error range of DHIG on flotation data is ( ), compared to Informer's ( ), indicating that DHIG can provide more reliable forecasts in industrial process control. Experimental results show that DHIG provides a high-precision and robust solution for complex multivariate time series forecasting tasks.

[0039] By using the benchmark PM2.5 and temperature datasets, we verify the adaptability of the model in fields with different temporal dynamics. The comparison results of the obtained indicators are shown in Tables 3 and 4: Table 3 Comparison of indicators of PM2.5 content prediction of different methods for PM2.5 dataset.

[0040] Table 4 Comparison of temperature prediction indicators of different methods for Temperature dataset In terms of temperature data, DHIG improves The MAE of PM2.5 data decreased from 1.32 to 1.05. (From 12.32 to 10.65.) This consistent performance across multiple datasets with different characteristics—high nonlinearity for flotation data, strong periodicity for temperature data, and high randomness for PM2.5 data—demonstrates the adaptability and robustness of DHIG.

[0041] In summary, the present invention proposes a flotation grade prediction method based on dynamic hypergraph and Informer-GRU. By introducing a dynamic hypergraph construction mechanism algorithm, the complex nonlinear relationship of multi-source data in the flotation process is effectively captured, and the Informer-GRU network is used to improve the capture of long-term trends and short-term fluctuations in the flotation process. The model training process is further optimized, the accuracy of flotation grade prediction is enhanced, and an effective solution is provided for flotation process optimization in industrial applications.

Claims

1. A flotation grade prediction method based on dynamic hypergraph and Informer-GRU, characterized in that: The following steps are involved: Step 1: preprocessing multi-source flotation process data, including bubble characteristic data and chemical composition data extracted from flotation foam images, to obtain normalized multivariate time series data through data cleaning and standardization; Step 2: A dynamic hypergraph is constructed using a sliding window mechanism. The correlation matrix is ​​constructed by calculating the Pearson correlation coefficient between features within the time window, and a hyperedge set is dynamically generated based on an adaptive threshold mechanism. Step 3: Design a dynamic weight calculation module to perform weighted processing on the hypergraph; Step 4: Perform hypergraph convolution operation to achieve feature aggregation; Step 5: Construct an Informer-GRU cascade time series prediction network. The Informer module uses a probabilistic sparse self-attention mechanism to extract global trend features, and the GRU module learns local time series patterns and outputs flotation grade prediction results.

2. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1 is characterized in that: The specific operations of preprocessing the multi-source flotation process data in step 1 include outlier processing, missing value processing, and feature normalization, as follows: use The principle is to identify and eliminate abnormal data points. Suppose the data of a certain feature channel is , whose sample mean is , the standard deviation is : (1) If a data point satisfy: (2) It is considered as an outlier and removed; When a feature value in the data set is missing, it is filled using the interpolation method of adjacent time steps. For isolated missing points, the average value of the adjacent valid data before and after is used to fill it: (3) Where: and Is a missing value For the valid data before and after, if there are consecutive missing data, multi-segment difference is used to fill in the missing data; All feature data are normalized to eliminate the dimensional differences between different features. The normalization formula is: (4) In the formula is the original data value, and are the minimum and maximum values ​​in the data set, respectively. is the normalized value.

3. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1 is characterized in that: The hyperedge generation in step 2 adopts a correlation-driven rule, and the hyperedge set generation process is as follows: The time series of each feature is regarded as a node in the graph, and the node feature is represented as: (5) Where, For nodes At time step The feature representation of is the sliding window size; The Pearson correlation coefficient is calculated using the data segments within the sliding window to achieve dynamic correlation analysis; the Pearson correlation coefficient calculation formula is: (6) The hyperedge generation rule is: (7) Where, Represents the time step The hyperedge set of is a feature subset, is a node set, Characterized by and The correlation coefficient between is the threshold parameter.

4. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1 is characterized in that: described The dynamic weight calculation formula in step 3 is: (8) Where, For super edge The weight of For super edge The contained node collection.

5. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 4 is characterized in that: The hypergraph convolution operation in step 4 realizes feature aggregation through the correlation matrix; The incidence matrix is ​​defined as: (9) Where, is the time step The correlation matrix, represents the node index, represents the hyperedge index; The convolution operation formula is: (10) Where, is the activation function, and are the node degree matrix and the hyperedge degree matrix respectively, is the incidence matrix, is the weight matrix, Represents the total number of convolutional layers, and the first convolution input is the hypergraph node feature: , No. The input features of the layer are the output features of the previous layer: .

6. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1 is characterized in that: The Informer module in step 5 adopts an encoder-decoder architecture, and both the encoder and decoder use the ProbSparse self-attention mechanism; the input and output of the Informer module are defined as: (11) (12) (13) Where, represents the encoder input sequence length and the decoder history target sequence length, is the global feature of compression.

7. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 6 is characterized in that: The cascaded time series prediction network in step 5 adopts the sequence arrangement of Informer in front and GRU in the back. The output of the Informer module Directly as the input of the GRU network, the GRU performs local temporal pattern learning and prediction output based on the global trend features. The GRU network updates the hidden state through the reset gate, update gate and candidate state. The calculation formula of the GRU network is: (14) (15) (16) (17) Where, for activation function, is the hyperbolic tangent activation function, Represents element-wise multiplication.

8. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 2 is characterized in that: The data set is divided into a training set and a test set. The training set is used to test the model performance, and the test set is used for verification. The root mean square error and mean absolute error are used to evaluate the model performance.

Images