A flotation grade prediction method based on dynamic hypergraph and informer-gru

Abstract

The application discloses a flotation grade prediction method based on a dynamic hypergraph and an Informer-GRU, belongs to the technical field of intelligent monitoring of industrial processes, and solves the technical problems that integration of multi-source heterogeneous variables cannot be realized in a complex flotation process and high-order space-time dependent relationships cannot be effectively captured. Technical solution points are as follows: a DHIG unified framework is constructed, dynamic hypergraph modeling and an Informer-GRU network are integrated, an adaptive hypergraph construction mechanism based on a sliding time window is designed, effective representation of evolution correlation between multi-source heterogeneous variables is realized, a sparse self-attention mechanism of an Informer module is adopted to capture long-range dependent relationships, short-term fluctuations caused by operation changes are processed in combination with a GRU component, and multi-scale feature fusion is realized through cascading of the Informer-GRU structure. Main uses are intelligent monitoring and grade prediction of the flotation process, and in industrial-grade flotation dataset tests, the prediction accuracy and calculation efficiency are superior to those of existing methods, thereby providing technical support for process optimization in a complex industrial environment.

Description

TECHNICAL FIELD

[0001] The present application relates to grade prediction in the mine flotation process, in particular to a flotation grade prediction method based on dynamic hypergraph and Informer-GRU architecture. BACKGROUND

[0002] The continuous decline of ore resource grade has become a challenge faced by the world, and the ore resource reserves in China are generally lower than the world average. In order to cope with the increasing demand for high-quality ore materials, China has increased its efforts in the research and development of flotation technology. However, the main goal of flotation technology is still to improve the recovery rate of ore, especially in the separation of low-grade ore, which is still a difficulty. In the flotation process, ore grade, flotation reagent concentration and equipment performance and other factors directly affect the separation effect. When dealing with low-grade ore, traditional flotation technology often has difficulty in maintaining stable and efficient flotation efficiency due to the small difference in hydrophobicity between minerals.

[0003] Currently, the grade monitoring and prediction of the flotation process mainly relies on manual judgment of foam quality and traditional analytical instruments (such as X-ray fluorescence analyzer). These methods not only limit the accuracy and continuity of detection, but also are easily affected by the experience and subjective factors of operators, resulting in poor stability of the flotation process and difficulty in maintaining consistent product quality. At the same time, the flotation process has significant nonlinearity and time-varying characteristics, and there is a complex dynamic coupling relationship between process parameters, such as feed grade, flotation reagent concentration and bubble characteristics. These mutual influences make the prediction of flotation grade more difficult, further highlighting the necessity of intelligent and accurate prediction of the flotation process.

[0004] In order to improve the prediction accuracy of flotation grade, many studies have been devoted to the application of multi-source data fusion and machine learning models. Although these methods have made some progress in certain aspects, they still face great challenges in dealing with the dynamic changes and complex multi-source heterogeneous data in the flotation process. Especially in dealing with long sequence data and capturing time-dependent relationships, the prediction accuracy of existing models is still insufficient.

[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

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

[0007] The technical solutions of the application are as follows:

[0008] A flotation grade prediction method based on dynamic hypergraph and Informer-GRU, comprising the following steps:

[0009] Step 1, preprocessing multi-source flotation process data, the multi-source data including bubble feature data extracted from flotation froth images and chemical composition data, obtaining standardized multivariate time series data through outlier processing, missing value processing and feature normalization;

[0010] Step 2, adopting a sliding window mechanism to construct a dynamic hypergraph, constructing a correlation matrix by calculating the Pearson correlation coefficient between features within a time window, and dynamically generating a hyperedge set based on an adaptive threshold mechanism;

[0011] Step 3, designing a dynamic weight calculation module to perform weighted processing on the hypergraph;

[0012] Step 4, the dynamic hypergraph represents the hypergraph structure through a correlation matrix, and performs hypergraph convolution operation to realize feature aggregation;

[0013] Step 5, constructing an Informer-GRU cascaded time series prediction network, the Informer module adopts a probabilistic sparse self-attention mechanism to extract global trend features, and the GRU module learns local time series patterns to output flotation grade prediction results.

[0014] As a further scheme of the application, the multi-source flotation process data in step 1 specifically includes image feature data, chemical composition data and a prediction target, wherein the image feature data includes bubble size mean and variance, contrast mean, correlation mean, energy mean, homogeneity mean, bubble speed mean and variance, bubble breakage area and volume mean, the chemical composition data includes Fe content, Zn content and Pb content in raw ore, and the prediction target is labeled concentrate zinc grade content.

[0015] The specific operations of preprocessing the 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:

[0016] Table 1: Example of flotation grade data sample (processed by Z value standardization)

[0017]

[0018] (1)

[0019] (2)

[0020] Adopting Principle, if the actual value Data points that differ from the sample mean by more than two standard deviations are identified as outliers and removed; is the average value, is the number of values contained in a certain channel; is the standard deviation;

[0021] When a certain feature value in the data set is missing, interpolation is used to fill in the adjacent time step; for isolated missing points, the average of the adjacent valid data before and after is used to fill in:

[0022] (3)

[0023] In the formula, is the position of the missing value, and is the valid data before and after the missing value. For continuous missing cases, multiple piecewise differences are used to fill in.

[0024] The data set is normalized, and the model is:

[0025] (4)

[0026] In the formula, is the original data value, and are the minimum and maximum values in the data set, respectively. is the normalized data value.

[0027] As a further scheme of the present application, step 2 defines the node of the dynamic hypergraph as a time series of each feature in the graph; the node feature is represented as:

[0028] (5)

[0029] In the formula, is the node characteristic representation at time step , and is the size of the sliding window.

[0030] Step 2 uses the data segment in the sliding window to calculate the Pearson correlation coefficient, and realizes dynamic correlation analysis; the dynamic correlation coefficient calculation formula is:

[0031] (6)

[0032] The hyperedge generation in step 2 uses a correlation-driven rule; the hyperedge generation rule is:

[0033] (7)

[0034] wherein, denotes the time step , the set of hyper-edges, is the feature subset, is the node set, is the feature and the correlation coefficient between is the threshold parameter.

[0035] As a further scheme of the present application, the dynamic weight calculation in step 3 is based on the correlation of features within the hyper-edge; the weight calculation formula is:

[0036] (8)

[0037] wherein, is the weight of the hyper-edge , and is the node set contained in the hyper-edge .

[0038] As a further scheme of the present application, the dynamic hypergraph in step 4 represents the hypergraph structure through an association matrix; the association matrix is defined as:

[0039] (9)

[0040] wherein, is the association matrix at time step , denotes the node index, denotes the hyper-edge index. As a further scheme of the present application, the hypergraph convolution operation in step 4 realizes feature aggregation through an association matrix; the convolution operation formula is:

[0041]

[0042] (10) wherein,

[0043] is the activation function, and are the node degree matrix and the hyper-edge degree matrix respectively, is the association matrix, is the weight matrix. denotes the total number of convolution layers, the first-layer convolution input is the hypergraph node feature: , the input feature of the th layer is the output feature of the previous layer: .

[0044] ​As a further aspect of the present invention, the Informer module described in step 5 adopts an encoder-decoder architecture, where both the encoder and decoder use the ProbSparse self-attention mechanism; the input and output of the Informer module are defined as follows:

[0045] (11)

[0046] (12)

[0047] (13)

[0048] In the formula, This represents the length of the encoder input sequence and the length of the decoder's historical target sequence. This is a global feature for compression.

[0049] The ProbSparse self-attention mechanism described in step 5 reduces computational complexity through probabilistic sparsity, thus simplifying the traditional self-attention mechanism. The complexity is optimized to ,in Given the sequence length, this enables efficient modeling of time dependencies in long sequences.

[0050] The cascaded structure described in step 5 uses an Informer-first, GRU-last sequence arrangement, with the Informer module's output... The data is directly used as input to the GRU network, which learns and predicts local time-series patterns based on global trend features.

[0051] The GRU network described in step 5 updates the hidden state through a reset gate, an update gate, and candidate states; the calculation formula for the GRU network is:

[0052] (14)

[0053] (15)

[0054] (16)

[0055] (17)

[0056] In the formula, for Activation function The hyperbolic tangent activation function is used. This indicates element-wise multiplication.

[0057] The data set is divided into a training set and a test set, the model performance is tested by using the training set, and the test set is verified, and the root mean square error (RMSE) and the mean absolute error (MAE) are used as the indicators for evaluating the model performance.

[0058] The beneficial effects of the present application are as follows:

[0059] The present application provides a DHIG (based on dynamic hypergraph and Informer-GRU) flotation grade prediction method, which significantly improves the prediction accuracy of flotation grade by combining dynamic hypergraph architecture and Informer-GRU model, especially in the flotation process, can effectively model the various spatio-temporal features in the flotation process, capture the complex dynamic relationship between complex multi-source features, and has stronger adaptability and accuracy compared with traditional models; at the same time, the present application shows superior performance on different types of data sets, especially on high nonlinearity, strong periodicity and high randomness data, and shows strong robustness and stability; finally, the present application provides a high-precision, efficient and robust solution for the flotation grade prediction task. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The basic structure of the present application is shown in the figure;

[0061] Figure 2 The process of integrating multi-source data features based on dynamic hypergraph of the present application is shown in the figure;

[0062] Figure 3 The DHIG algorithm flowchart constructed by the present application is shown in the figure; DETAILED DESCRIPTION

[0063] The following will be combined Figures 1-3 The technical solutions of the present application will be further described in detail.

[0064] Example 1

[0065] The present application provides a kind of to the problem that parameter between flotation process has strong coupling and non-linear relationship, traditional method is prone to low prediction accuracy, optimization process is easy to fall into local optimal solution and training time is too long etc. when processing multi-source heterogeneous data.Therefore, the present application proposes an innovative dynamic hypergraph structure, and combines Informer-GRU network, can efficiently capture spatio-temporal features and integrate the complex relationship between multi-source features, so as to realize accurate flotation grade prediction.

[0066] A flotation grade prediction method based on dynamic hypergraph and Informer-GRU is as follows Figure 1The shown. First, according to the multi-source data in the flotation process (bubble feature data and chemical composition data extracted from the flotation froth image), data preprocessing is performed. This process includes outlier processing, missing value filling and feature normalization, ensuring data quality and standardizing input data. Next, a sliding window mechanism is used to split the data, calculate the Pearson correlation coefficient between features in each time window, build a correlation matrix between features, and generate a dynamic hyperedge set based on an adaptive threshold mechanism, thereby constructing a dynamic hypergraph. Then, a dynamic weight calculation module is designed to assign weights to each hyperedge in the hypergraph based on the correlation between features, and the influence of important features is strengthened through a weighted algorithm. Next, a convolution operation is performed on the dynamic hypergraph to aggregate local information of each feature. Finally, the Informer-GRU cascade network is used for flotation grade prediction, the Informer module uses a sparse self-attention mechanism to extract global trend features, while the GRU module learns local time series patterns, which can effectively capture long-term trends and short-term fluctuations in the flotation process, and output the final flotation grade prediction results. Through error analysis and result optimization, the prediction accuracy and practical application effect of the model are further improved.

[0067] The specific implementation steps of a flotation grade prediction method based on dynamic hypergraph and Informer-GRU are as follows:

[0068] 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 speed mean and variance, bubble breakage area and volume mean, chemical composition data including Fe content, Zn content, Pb content in feed grade, and the prediction target is the labeled zinc grade content. In order to integrate the multi-source data and construct a dynamic hypergraph, the method will preprocess the data set;

[0069] The specific operations for preprocessing the flotation data include: outlier processing, missing value processing, and feature normalization. The data samples are shown in Table 1, and the basic process is as follows:

[0070] Table 1: Example of flotation grade data samples (Z-value standardized)

[0071]

[0072] (18)

[0073] (19)

[0074] Using the principle, if the actual value differs from the sample mean by more than twice the standard deviation, the data point is determined to be an outlier and is removed; is the average value, is the number of values contained in a certain channel; is the standard deviation;

[0075] When a certain feature value in the data set is missing, it is filled in by interpolation of adjacent time steps. For isolated missing points, the average of the adjacent valid data before and after the midpoint is used to fill in.

[0076] (20)

[0077] wherein and are the valid data before and after the missing value. For continuous missing cases, multiple piecewise differences are used to fill in. The data set is normalized, and the model is:

[0078]

[0079] (21)

[0080] wherein is the original data value, and are the minimum and maximum values in the data set, respectively. is the normalized data value.

[0081] Each data set is divided into training set, validation set and test set according to the ratio of 12:3:5.

[0082] Step S2, the node definition of the dynamic hypergraph takes each feature time series as a node in the graph; the node feature is represented as:

[0083] (22)

[0084] wherein, is the feature representation of node at time step , and is the size of the sliding window.

[0085] The data segment in the sliding window is calculated by using the Pearson correlation coefficient to realize dynamic correlation analysis; the dynamic correlation coefficient calculation formula is:

[0086] (23)

[0087] By constructing a hyperedge with strong correlation in the feature subset , and the correlation degree exceeds the threshold ; the hyperedge generation rule is: ​

[0088] (24)

[0089] wherein, denotes the set of hyper-edges at time step , is the subset of features, is the set of nodes, is the feature and is the correlation coefficient between is the threshold parameter.

[0090] Step S3, the dynamic weight calculation is based on the correlation of features in the hyper-edge; the weight calculation formula is:

[0091] (25)

[0092] wherein, is the weight of the hyper-edge , is the set of nodes contained in the hyper-edge .

[0093] The dynamic hypergraph represents the hypergraph structure through the association matrix; the association matrix is defined as:

[0094] (26)

[0095] wherein, is the association matrix at time step , denotes the node index, denotes the hyper-edge index.

[0096] Step S4, as shown in Figure 2 , the feature aggregation is carried out by using the hypergraph convolution operation to obtain the comprehensive representation of the multi-dimensional feature; the convolution operation formula is:

[0097] (27)

[0098] wherein, is the activation function, and are the node degree matrix and the hyper-edge degree matrix respectively, is the association matrix, is the weight matrix. denotes the total number of convolution layers, the first layer convolution input is the hypergraph node feature: , the input feature of the layer is the output feature of the previous layer: .

[0099] Step 5, as shown in Figure 3As shown, the Informer-GRU cascaded time series prediction network is constructed, the Informer module extracts global trend features through the probabilistic sparse self-attention mechanism, the GRU module learns local time series patterns, and finally outputs the flotation grade prediction result, and the step 5 is specifically as follows.

[0100] The Informer module adopts an encoder-decoder architecture, and the encoder and the decoder both use ProbSparse self-attention mechanism; the input and output of the Informer module are defined as:

[0101] (28)

[0102] (29)

[0103] (30)

[0104] Wherein, represents the length of the encoder input sequence and the length of the decoder historical target sequence, is the compressed global feature.

[0105] The cascaded structure adopts the sequence arrangement of Informer in front and GRU behind, and the output of the Informer module is directly used as the input of the GRU network, and the GRU learns the local time series pattern based on the global trend feature and outputs the prediction result.

[0106] The GRU network in step 5 realizes the hidden state update through the reset gate, the update gate and the candidate state; the calculation formula of the GRU network is:

[0107] (31)

[0108] (32)

[0109] (33)

[0110] (34)

[0111] Wherein, is an activation function, is a hyperbolic tangent activation function, represents element-wise multiplication.

[0112] In order to verify the accuracy and effectiveness of the method DHIG proposed in the present application for concentrate grade prediction in the flotation process, the root mean square error (RMSE) and the mean absolute error (MAE) are used as the evaluation model performance.

[0113] The index comparison results obtained by using different algorithms to predict the grade of zinc concentrate are shown in Table 1:

[0114] Table 2 Index comparison of different methods for predicting the grade of zinc concentrate on the flotation data set

[0115]

[0116] As can be seen from Table 2, on the flotation data set, DHIG shows significant improvement: compared with CNN-GRU, MAE is reduced by (from 0.3933 to 0.1549), compared with Transformer, by . Compared with HGNN-GRU, by . Most significantly, DHIG outperforms the state-of-the-art Informer model, with a reduction in MAE of , and a reduction in RMSE of , demonstrating enhanced ability in capturing complex spatio-temporal dependencies. Stability analysis shows that the error range of DHIG on the flotation data is , compared with of Informer, indicating that DHIG can provide more reliable predictions in industrial process control. The experimental results show that DHIG provides a high-precision, robust solution for complex multivariate time series prediction tasks.

[0117] By using the benchmark data set PM2.5 and the temperature data set, the adaptability of the model in fields with different time dynamics is verified, and the index comparison results obtained are shown in Table 3 and Table 4:

[0118] Table 3 Index comparison of different methods for predicting PM2.5 content on the PM2.5 data set.

[0119]

[0120] Table 4 Index comparison of different methods for predicting temperature on the Temperature data set

[0121]

[0122] On the temperature data, DHIG improves the MAE of (from 1.32 to 1.05), while on the PM2.5 data, the MAE is reduced by This consistent performance on multiple datasets with different characteristics - high nonlinearity of flotation data, strong periodicity of temperature data, and high randomness of PM2.5 data - demonstrates the adaptability and robustness of DHIG.

[0123] In summary, the present application 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. Through the Informer-GRU network, the long-term trend and short-term fluctuations in the flotation process are improved, further optimizing the model training process and enhancing the accuracy of flotation grade prediction, providing an effective solution for flotation process optimization in industrial applications.

Claims

1. A flotation grade prediction method based on dynamic hypergraph and Informer-GRU, characterized in that, Includes the following steps: Step 1: Preprocess the multi-source flotation process data, which includes bubble feature data and chemical composition data extracted from flotation foam images. Normalized multivariate time series data are obtained through data cleaning and standardization. Step 2: A dynamic hypergraph is constructed using a sliding window mechanism. A correlation matrix is ​​constructed by calculating the Pearson correlation coefficient between features within the time window, and a set of hyperedges 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 cascaded time-series prediction network. The Informer module uses a probabilistic sparse self-attention mechanism to extract global trend features, while the GRU module learns local time-series patterns and outputs the flotation grade prediction results. The hyperedge generation in step 2 adopts a correlation-driven rule, and the hyperedge set generation process is as follows: Each feature's time series is treated as a node in the graph, and the node feature is represented as follows: （5） In the formula, For nodes At time step Feature representation, To adjust the sliding window size; The Pearson correlation coefficient is calculated using data segments within a sliding window to achieve dynamic correlation analysis; the formula for calculating the Pearson correlation coefficient is as follows: （6） The rule for generating the hyperedge is as follows: （7） In the formula, Indicates time step The set of superedges For feature subset, For a set of nodes, Features and The correlation coefficient between them For threshold parameters; The Informer module in step 5 adopts an encoder-decoder architecture, with both the encoder and decoder using the ProbSparse self-attention mechanism; the input and output of the Informer module are defined as follows: （11） （12） （13） In the formula, This represents the length of the encoder input sequence and the length of the decoder's historical target sequence. For compression, global features; The cascaded time series prediction network in step 5 adopts an Informer-first, GRU-last sequence arrangement, with the Informer module's output... The input is directly used as the input to the GRU network. The GRU performs local temporal pattern learning and prediction based on global trend features. The GRU network updates the hidden state through a reset gate, an update gate, and candidate states. The calculation formula for the GRU network is: （14） （15） （16） （17） In the formula, for Activation function The hyperbolic tangent activation function is used. This indicates element-wise multiplication.

2. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1, characterized in that, The specific operations for preprocessing the multi-source flotation process data in step 1 include outlier handling, missing value handling, and feature normalization, as detailed below: use The principle is to identify and remove outlier data points. Let the data of a certain feature channel be... Its sample mean is The standard deviation is : （1） If a certain data point satisfy: （2） Then it is considered an outlier and removed; When a feature value in the dataset is missing, it is imputed using interpolation between adjacent time steps. For isolated missing points, the average of the preceding and following valid data is used for imputation. （3） In the formula: and Missing values For consecutive missing data, multiple interpolation methods are used to fill in the gaps. Min-max normalization is performed on all feature data to eliminate differences in dimensions between different features. The normalization formula is: （4） In the formula These are the original data values. and These are the minimum and maximum values ​​in the dataset, respectively. It is the normalized value.

3. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1, characterized in that, The dynamic weight calculation formula in step 3 is as follows: （8） In the formula, For super-edge The weight, For super-edge The set of nodes contained therein.

4. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 1, characterized in that, The hypergraph convolution operation in step 4 achieves feature aggregation through the correlation matrix; The correlation matrix is ​​defined as follows: （9） In the formula, For time steps The correlation matrix, Indicates the node index. Indicates a superedge index; The convolution operation formula is as follows: （10） In the formula, For activation function, and These are the node degree matrix and the hyperedge degree matrix, respectively. It is an incidence matrix. This is the weight matrix. This represents the total number of convolutional layers. The input to the first convolutional layer is the hypergraph node features. , No. The input features of the previous layer are the output features of the previous layer: .

5. The flotation grade prediction method based on dynamic hypergraph and Informer-GRU according to claim 2, characterized in that, 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 validation. The root mean square error and the mean absolute error are used to evaluate the model performance.

Images