Process space structure-based ore dressing process key indicator prediction method and system, device and medium

Abstract

The application discloses a kind of based on process space-time structure's ore dressing process key index prediction method and system, method includes: obtaining the process variable data related to key index in ore dressing process, obtain original data set;Original data set is preprocessed;Process variable time lag matrix and key index vector are constructed, the time lag correlation coefficient of each process variable is calculated, and the time lag of each process variable is determined according to the coefficient and the process variable is filtered;Each sub-process of ore dressing process is decoupled, and the space-time structure model of ore dressing process is established, and embedded deep learning network structure, and the space-time deep LSTM ore dressing process key index prediction model is established;According to the process variable after screening and time lag, the data set after screening process variable is time sequence matched, training data set is constructed using time sliding window, and the optimal prediction model is obtained by training, for ore dressing process key index online prediction.The application can realize the real-time, accurate prediction of long process key index in ore dressing process.

Description

Technical Field

[0001] This invention belongs to the field of key indicator measurement technology in mineral processing, specifically relating to a method and system for predicting key indicators in mineral processing based on the spatiotemporal structure of the process. Background Technology

[0002] In mineral processing, timely monitoring and control of key indicators are crucial for achieving green production, energy conservation, and improved enterprise efficiency. Grinding overflow particle size and concentrate grade are important performance indicators in the mineral processing process. Timely and accurate monitoring and analysis of these key indicators are essential guarantees and prerequisites for maintaining stable operation, ensuring smooth process control, and producing high-quality products. However, online monitoring equipment for key indicators in mineral processing is expensive and difficult to maintain. Currently, most mineral processing plants in China still rely on manual offline testing for key indicator monitoring. Offline testing involves a series of operations such as product sampling, drying, and analysis, resulting in long testing cycles, significant delays, and low testing frequency, making it difficult to meet the needs of real-time production control. Since workers cannot obtain real-time monitoring values ​​for key indicators, the control of the mineral processing process is usually determined by experienced workers visually observing flotation froth and process variables to judge the production status and make corresponding adjustments. This method of operation is characterized by strong worker subjectivity; different workers have different experiences, and there is no unified standard for judging the working conditions. This leads to deviations in operation among different workers, making it difficult for the mineral processing process to achieve optimal operating conditions. In modern industrial processes, soft sensors estimate key quality variables by establishing mathematical models and using easily measurable auxiliary variables, thereby providing crucial real-time information for process monitoring, optimization, and control. Compared to traditional hard measurement methods, soft sensors offer numerous advantages such as rapid response, low cost, diverse methods, convenient maintenance, and safe operation, leading to their widespread and successful applications in industries such as chemical, biochemical, metallurgical, and pharmaceutical manufacturing.

[0003] Mineral processing is a long-term, continuous, and dynamic production process, and key performance indicators (KPIs) are the result of the combined effects of all sub-processes over a period of time. Considering the impact of each stage of mineral processing on the final KPIs, establishing an online predictive model for KPIs based on multi-source time-series data of the mineral processing process can provide real-time references for on-site workers, thereby optimizing process parameters in real time and ensuring optimal production. This has significant practical implications for improving the product quality and economic benefits of mineral processing plants. Summary of the Invention

[0004] To address the technical problems of high cost and difficult maintenance of online detection equipment for key indicators in mineral processing, and the significant lag in offline grade detection through manual testing, which cannot meet the needs of real-time production control and optimization, this invention provides a method and system for predicting key indicators in mineral processing based on the spatiotemporal structure of the process, with high detection accuracy.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] A method for predicting key indicators in a mineral processing step based on the spatiotemporal structure of the process includes:

[0007] S1, Obtain process variable data related to key indicators during mineral processing to obtain the raw dataset;

[0008] S2, preprocessing the original dataset;

[0009] S3, construct the process variable time delay matrix and key indicator vector, calculate the time delay correlation coefficient of the process variables, and screen process variables based on the coefficient;

[0010] S4 decouples each sub-process in the mineral processing, establishes a spatiotemporal structure model of the mineral processing process, and embeds the spatiotemporal structure model of the mineral processing process into a deep learning network structure to establish a spatiotemporal depth LSTM prediction model for key indicators of the mineral processing process.

[0011] S5: Based on the selected process variables and time delays, perform time-series matching on the dataset after selecting process variables, construct a training dataset using a time sliding window, and obtain the optimal prediction model through training for online prediction of key indicators in the mineral processing process.

[0012] Furthermore, the process variables related to the key indicators include, but are not limited to: mill feed rate, water feed rate, mill power, hydrocyclone feed flow rate, hydrocyclone feed concentration, hydrocyclone feed pressure, thickness of the froth layer in each rougher cell, dosage of reagents in each rougher cell, aeration rate in each rougher cell, thickness of the froth layer in each cleaner cell, aeration rate, and dosage of reagents.

[0013] Further preprocessing of the original dataset includes missing value imputation and standardization.

[0014] The missing value imputation process employs different methods depending on the characteristics of the data from the ore dressing plant. For missing process data caused by on-site server failure or shutdown, the nearest neighbor imputation method (either forward or backward) is used. For missing data collected by sensors, the K-nearest neighbor imputation method is used. This method searches the entire dataset for the K groups of samples closest to the missing value sample, assigns distance weights between the sample and other samples based on the Euclidean distance, and calculates the dot product of the vectors of the non-missing values ​​and the weight vectors of the K groups of samples to obtain the missing value sample value. Finally, the imputed dataset is obtained. , where p is the number of process variables in the imputed dataset, and T is the number of data points;

[0015] The standardization process uses the Z-Score method to evaluate the imputed dataset. Standardize to obtain the standardized dataset. .

[0016] Furthermore, the method for determining the time delay of each process variable is as follows:

[0017] 1) Let the data sampling interval of the process variable be... The average residence time of materials in the sub-processes corresponding to each process variable are as follows: Where p is the number of process variables in the imputed dataset; and the maximum time delay of all process variables. The following is determined, among which This represents the floor function;

[0018] (1)

[0019] 2) Based on the maximum time delay Constructing process variables time delay matrix and key indicator vectors as follows:

[0020] (2)

[0021] (3)

[0022] in, Representing process variables exist Data at any given time indicates the key metrics at [time]. Data at any given time; time delay matrix The first line It is related to the key indicator vector The sequence has no time delay, the second row has a time delay of 1, and so on until the last row has a time delay. A sequence of order;

[0023] 3) Calculate process variables time delay matrix With key indicator vector Time-delay correlation coefficient vector obtained based on Pearson correlation ;

[0024] (4)

[0025] (5)

[0026] in, , For vectors and key indicator vectors Pearson correlation coefficient, It is a vector The average value, Key Indicator Vector The average value;

[0027] 4) Select The time delay corresponding to the maximum value of the time delay correlation coefficient in As a process variable The time lag;

[0028] 5) Time delay for the i-th process variable Perform correction to obtain the corrected time delay. The correction formula is as follows;

[0029] (6)

[0030] in, For the first The average dwell time in the subprocess corresponding to each process variable. This represents the function for rounding up.

[0031] Furthermore, process variables are filtered based on their maximum correlation coefficients. Specifically, it is determined whether the maximum correlation coefficient of each process variable is greater than a preset threshold. If it is, the process variable is retained; otherwise, it is removed, resulting in the final dataset. Where n is the number of filtered process variables, and T is the number of data points. The preset threshold can be set to 0.1.

[0032] Furthermore, the specific steps of step S4 are as follows:

[0033] 1) Divide the mineral processing process into m sub-processes according to the actual production flow sequence. The input process variables and output variables of any sub-process k are respectively... and ,in The number of input process variables for subprocess k, and the output of subprocess m. These are the key indicator variables in the mineral processing process;

[0034] 2) Based on the coupling between the forward and reverse flows in the mineral processing flow, i.e., a portion of the output slurry from a later sub-process will flow back to an earlier sub-process, let the input of the slurry flowing from a later sub-process back to an earlier sub-process be k. ,in This represents the number of variables that are returned to subprocess k from subsequent subprocesses.

[0035] 3) Unify the time delay between all process variables and key indicator variables in subprocess k as follows: ;

[0036] 4) Different processes in the mineral processing influence each other, and data from any single sub-process is insufficient to fully characterize the dynamic changes of the entire production process. Spatially, the output of the preceding sub-process will progressively affect the output of the subsequent sub-process, ultimately impacting key indicators. Temporally, the time lag between sub-processes results in a strong correlation between the current key indicator value and the historical time series of process variables. Therefore, the raw ore changes progressively in both space and time throughout the entire production process, and the process variables of each sub-process act on the key indicator variables progressively in both space and time.

[0037] However, there is no coupling relationship between the sub-process backflow and the final key performance indicator (KPI) variable in the mineral processing process. For the KPI variable, the sub-process backflow can be regarded as an internal loop of the preceding process. By decoupling the forward flow and backflow of the sub-process and considering the time lag of the backflow variable of the subsequent sub-process, the backflow of the subsequent sub-process can be used as the input variable for the backflow to the preceding sub-process. Therefore, based on the fact that the output of each sub-process is the result of the combined effects of the input of that sub-process, the backflow input of that sub-process, the output of the preceding sub-process, and the operating conditions of that sub-process, a spatiotemporal structure model of the mineral processing process can be established as shown below:

[0038] (7)

[0039] (8)

[0040] (9)

[0041] (10)

[0042] in, It is the selection function of subprocess k, which maps the non-linear relationship between the input and output of the subprocess and is learned through a deep learning network; Represents the k-th subprocess in td k All input variables at any given time, including the input process variables of the current k-th subprocess itself. And the variables that flow back from the backward subprocess to the k-th subprocess. ;d represents the input time step; It is the output of the kth subprocess; This represents the key performance indicator output value at time t, which is also the output of the m-th subprocess at time t. ; Historical information for key indicator variables;

[0043] (11)

[0044] 5) Establish an m+1 layer spatiotemporal depth LSTM network, and embed the constructed spatiotemporal structure model of the mineral processing process into the m+1 layer spatiotemporal depth LSTM network to obtain a spatiotemporal depth LSTM prediction model for key indicators of the mineral processing process, specifically:

[0045] The spatiotemporal depth LSTM network vertically represents the spatial progression sequence of the mineral processing technology from sub-process 1 to sub-process m, with the last layer representing the historical key indicator sequence. Vertically, the input to each layer consists of features from each sub-process, with these features progressing over time. The external input to the k-th layer is the relevant process variable data from sub-process k, and its hidden layer result is input to the (k+1)-th layer. The external input to the (k+1)-th layer is the process variable data from sub-process k+1, progressing step by step to the (m+1)-th layer. At the last time step of the (m+1)-th layer, the hidden layer information of the LSTM network is output to the fully connected layer. After passing through the activation function, the final predicted value of the key indicator is output. Horizontally, the network represents the unfolding of the entire process time series with a time step size of d, meaning each layer consists of d LSTM cell units. The input to the d-th cell unit is all the input variables of the current sub-process k corresponding to the key indicator variable at time t. The input variables from the d-th cell unit to the first cell unit are gradually delayed by one time point in the time step;

[0046] Because there is a time lag between each subprocess and the key variables, the key indicator variable at time t corresponds to time td. k The input of the k-th layer network at time k This input is used to predict the key performance indicator value. The information dissemination process is as follows:

[0047] Input Gate: (12)

[0048] Forgotten Gate: (13)

[0049] Output gate: (14)

[0050] (15)

[0051] (16)

[0052] (17)

[0053] Final key indicator forecasts: (18)

[0054] in, , , The k-th layer network at td kThe input gate, forget gate, and output gate at each time step; It is the sigmoid activation function. These are the external input variables of the k-th layer network, including the original input variables and return input variables of the corresponding subprocess k. and td k The hidden layer output and td of the (k-1)th layer of the network at time k k The output of the hidden layer of the k-th layer network at time -1 , , These are the weights of the input gate, forget gate, and output gate of the k-th layer network, respectively. , , , For the corresponding network bias, For td k Input the door status information at all times. and For the corresponding weights and biases, For the k-th layer network Cellular state at any given moment These are the weights for the fully connected layer. It is the tanh activation function. This indicates that corresponding elements in the corresponding vectors are multiplied together. This is the hidden layer output of the last layer at time t-2. This represents the true value of the key indicator at time t-1.

[0055] Furthermore, when training the spatiotemporal depth LSTM prediction model for key indicators in the mineral processing process using the training dataset, the mean squared error (MSE) is selected as the loss function during training:

[0056] (19)

[0057] in The number of training samples, The true value of the key indicator at time t. Here are the predicted values ​​of key indicators at time t.

[0058] Furthermore, a time-based error backpropagation algorithm and the Adam optimizer are used to train a spatiotemporal depth LSTM prediction model for key indicators in the mineral processing process. Even better, an early stopping method is employed to avoid overfitting and improve the model's generalization ability, until the model converges or reaches the maximum number of iterations, resulting in a well-trained prediction model for key indicators in the mineral processing process.

[0059] Furthermore, for the grinding process in the mineral processing, the key indicator is the second-stage overflow particle size; and / or, for the flotation process in the mineral processing, the key indicator is the concentrate grade.

[0060] A key indicator prediction system for a mineral processing process based on a spatiotemporal structure includes a mineral processing site data acquisition server, a data storage server, a deep learning server, and multiple display terminals for displaying predicted values ​​of key indicators. The deep learning server includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor enables the processor to implement the key indicator prediction method for a mineral processing process based on a spatiotemporal structure as described above.

[0061] Beneficial effects

[0062] This invention can integrate the temporal and spatial correlation of materials between sub-processes in the mineral processing into the construction of a key indicator prediction model, thereby achieving real-time and accurate prediction of key indicators for long-process mineral processing. Attached Figure Description

[0063] Figure 1 This is a flowchart of the present invention;

[0064] Figure 2 This is a flow chart of the grinding process in Example 1;

[0065] Figure 3 The spatiotemporal depth LSTM network structure diagram is shown in the embodiment.

[0066] Figure 4 This is a schematic diagram of the time-series processing of mineral processing data;

[0067] Figure 5 This is a comparison chart of the predicted and actual values ​​of 1200 sets of grinding overflow particle sizes obtained in Example 1 of the present invention.

[0068] Figure 6 This is a flowchart of the flotation process in Example 2;

[0069] Figure 7 This is a comparison chart of the predicted and actual values ​​of 784 groups of concentrate grades obtained from two examples of this invention. Detailed Implementation

[0070] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0071] The embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and intended to illustrate the principles and methods of the present invention, and are not intended to limit the scope of the present invention.

[0072] Example 1

[0073] This embodiment 1 presents a method for predicting key indicators of a grinding process based on its spatiotemporal structure, comprising the following steps: First, collecting relevant process variable data during the grinding process to form an original dataset; preprocessing the original dataset, including imputation of missing values ​​and standardization; constructing a process variable time lag matrix and a key indicator vector, calculating the time lag correlation coefficient of the process variables, and filtering process variables based on this coefficient; performing mechanistic analysis on the grinding process, establishing a spatiotemporal structure model of the grinding process, and embedding a deep LSTM network to establish a spatiotemporal deep LSTM overflow particle size prediction model; based on the time lag, performing time-series matching on the filtered dataset, constructing a training dataset using a time sliding window, and obtaining the optimal prediction model through training to achieve online prediction of the two-stage overflow particle size of the grinding process.

[0074] S1: Figure 2 This is the grinding process flow diagram for Example 1. The original dataset consists of 15 process variables and one second-stage overflow particle size data related to the grinding process. The data sampling period is 1 hour.

[0075] S2: Preprocess the original dataset, including imputation of missing values ​​and standardization.

[0076] S3: Based on the grinding process mechanism and the sampling interval of process variable data, the maximum time lag of the process variable is determined to be 2. The time lag matrix of the process variable and the time lag vector of the key indicator are established. The maximum time lag correlation coefficient of each process variable and the key indicator and its corresponding time lag are calculated, as shown in Table 1.

[0077] Table 1. Time lags and correlation coefficients of various process variables in the grinding process relative to the overflow particle size.

[0078]

[0079] Table 1 shows that the correlation between the feed flow rate of hydrocyclone #2, the overflow flow rates of hydrocyclones #1 and #2, and the overflow particle size of the second grinding stage is less than 0.1. Therefore, these three process variables are removed, and the remaining 12 process variables are selected as auxiliary process variables. Meanwhile, the time lag between the feed concentration, feed flow rate, hydrocyclone pressure, and overflow concentration of hydrocyclone #1 and the overflow particle size of the second grinding stage is 1. According to the mineral processing mechanism, the average residence time of the semi-autogenous mill feed slurry to the hydrocyclone overflow is greater than 30 minutes, while the data sampling interval for this study is 1 hour. Therefore, the time lag of the above four process variables is corrected to 0.

[0080] S4: Mechanism analysis of the grinding process. The ore undergoes three stages from feed slurry to overflow slurry: semi-autogenous mill, pump pool, and hydrocyclone. For the overflow slurry at any given moment, the hydrocyclone underflow and underflow makeup water need to pass through the ball mill-pump pool-hydrocyclone again to become qualified overflow slurry. Since the average residence time of hydrocyclone underflow in the ball mill is 10 minutes, much shorter than the data sampling interval, the grinding process is simplified, and the hydrocyclone underflow makeup water is considered as the input to the pump pool. Furthermore, the overflow flow rate and overflow concentration of the hydrocyclone are directly related to the overflow particle size. Therefore, this embodiment divides the grinding process into four sub-processes: semi-autogenous mill, pump pool, hydrocyclone feed, and hydrocyclone discharge, establishing a spatiotemporal structure model of the grinding process. Corresponding to each sub-process of the grinding process, a 5-layer spatiotemporal depth LSTM overflow particle size prediction model is established, with the model structure as follows: Figure 3 As shown in Table 2.

[0081] (1)

[0082] Table 2. Structure of the spatiotemporal depth LSTM grinding overflow particle size prediction model with domain embedding

[0083]

[0084] S5: Perform time-series matching on the filtered dataset based on time delay, and construct the training dataset using a time sliding window, such as... Figure 4 As shown, the mean squared error (MSE) was chosen as the loss function during training, with a training learning rate of 0.0001 and a maximum number of iterations of 200. A time-based backpropagation algorithm and the Adam optimizer were used to train the spatiotemporal deep LSTM model, minimizing the loss function. Early stopping was employed to avoid overfitting and improve the model's generalization ability until convergence or the maximum number of iterations was reached, resulting in a well-trained prediction model for key indicators in the mineral processing process.

[0085] In this embodiment, continuous sampling was conducted at the industrial production site, with sampling performed once every hour, resulting in a total of 6048 samples. The ratio of training set, validation set, and test set data was 6:2:2. Therefore, 5048 sets of data were used for model training, and 1200 sets of samples were used for model testing. Figure 5 The graph shown is a comparison of predicted grinding overflow particle size. The accuracy of predicted grinding overflow particle size with a relative error of 5% reaches 98.5%.

[0086] Example 2

[0087] This embodiment 2 presents a method for predicting key indicators in the flotation process using a spatiotemporal deep LSTM network based on domain knowledge embedding. The method includes the following steps: First, collecting relevant process variable data during the flotation process to form an original dataset; preprocessing the original dataset, including imputation of missing values ​​and standardization; constructing a time-delay matrix for process variables and a vector for key indicators, calculating the time-delay correlation coefficient of the process variables, and filtering process variables based on this coefficient; performing a mechanism analysis of the flotation process, establishing a spatiotemporal structure model of the flotation process, and embedding it into a deep LSTM network to establish a spatiotemporal deep LSTM concentrate grade prediction model; based on the time delay, performing time-series matching on the filtered dataset, constructing a training dataset using a time sliding window, and obtaining the optimal prediction model through training to achieve online prediction of concentrate grade in the flotation process.

[0088] S1: Figure 6 This is a flow chart of the flotation process in Example 2. The original dataset consists of 24 process variables and 1 concentrate grade data related to the flotation process. The data sampling period is 1 hour.

[0089] S2: Preprocess the original dataset, including imputation of missing values ​​and standardization.

[0090] S3: Based on the flotation process mechanism and the sampling interval of process variable data, the maximum time delay of the process variable is determined to be 2. The process variable time delay matrix and the key indicator time delay vector are established. The maximum time delay correlation coefficient of each process variable and key indicator and its corresponding time delay are calculated, as shown in Table 3.

[0091] S4: Within the cleaning line, the underflow from Cleaner I to Cleaner III is used as backflow input to the previous process. There is some coupling between the reverse backflow and the forward flotation flow. However, from the roughing to the end of the cleaning line, there is only one output: concentrate. These backflows are not coupled with the final concentrate. For the final concentrate grade, the backflow can be regarded as an internal circulation of the preceding process. Therefore, the function of multi-stage backflow is essentially to circulate and improve the grade of the slurry, and to discharge some of the lower-grade slurry. Since the low-grade slurry in the cleaning line only has one outlet at the roughing stage, the low-grade slurry generated after the backflow of Cleaner II and Cleaner III can be regarded as being discharged from the roughing cell through a series of countercurrents. Decoupling the slurry after backflow and dividing it into two parts, the high-grade overflow slurry is incorporated into the forward slurry flow, and the low-grade underflow is discharged with the roughing outlet. This is equivalent to the backflow only containing low-grade tailings. A time step of 1 is selected to establish a spatiotemporal structure model of the flotation process. A six-layer spatiotemporal depth LSTM concentrate grade prediction model is established for each sub-process of the flotation process. The model structure is shown in Table 4.

[0092]

[0093] Table 3. Time lag and correlation coefficient of each process variable in the flotation process relative to concentrate grade

[0094]

[0095] S5: Time-series matching is performed on the filtered dataset based on time delays. A time sliding window is used to construct the training dataset. The mean squared error (MSE) is selected as the loss function during training, with a training learning rate of 0.0001 and a maximum number of iterations of 200. A time-based error backpropagation algorithm and the Adam optimizer are used to train a spatiotemporal deep LSTM model, minimizing the loss function. Early stopping is employed to avoid overfitting and improve the model's generalization ability until the model converges or reaches the maximum number of iterations, resulting in a well-trained prediction model for key indicators in the mineral processing process.

[0096] Table 4. Structure of the Spatiotemporal Depth LSTM Flotation Concentrate Grade Prediction Model with Domain Embedding

[0097]

[0098] In this embodiment, continuous sampling was conducted at the industrial production site, with sampling performed once every hour, resulting in a total of 3922 samples. The ratio of training set, validation set, and test set data was 6:2:2. Therefore, 3138 sets of data were used for model training, and 784 samples were used for model testing. Figure 7 The chart shown is a comparison of flotation concentrate grade predictions. The accuracy rate for predicting flotation concentrate grade with a relative error within 10% reaches 87.24%, and the accuracy rate for predicting with a relative error within 15% reaches 94.01%.

[0099] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various modifications or improvements based on these embodiments. Without departing from the overall concept of this application, such modifications or improvements should fall within the scope of protection claimed in this application.

Claims

1. A method for predicting key indicators in a mineral processing step based on the spatiotemporal structure of the process, characterized in that, include: S1, Obtain process variable data related to key indicators during mineral processing to obtain the raw dataset; S2, preprocessing the original dataset; S3, construct the process variable time delay matrix and key indicator vector, calculate the time delay correlation coefficient of each process variable, determine the time delay of each process variable based on the coefficient, and screen the process variables; S4 decouples each sub-process in the mineral processing, establishes a spatiotemporal structure model of the mineral processing process, and embeds the spatiotemporal structure model of the mineral processing process into a deep learning network structure to establish a spatiotemporal depth LSTM prediction model for key indicators of the mineral processing process. S5: Based on the selected process variables and time delays, perform time-series matching on the dataset after selecting process variables, construct a training dataset using a time sliding window, and obtain the optimal prediction model through training for online prediction of key indicators in the mineral processing process. For the grinding process in mineral processing, the key indicator is the second-stage overflow particle size; and / or, for the flotation process in mineral processing, the key indicator is the concentrate grade. The method for determining the time delay of each process variable is as follows: 1) Let the data sampling interval of the process variable be... The average residence time of materials in the sub-processes corresponding to each process variable are as follows: Where p is the number of process variables in the imputed dataset; and the maximum time delay of all process variables. The following is determined, among which This represents the floor function; (1) 2) Based on the maximum time delay Constructing process variables time delay matrix and key indicator vectors as follows: (2) (3) in, Representing process variables exist Data at any given time Indicates key indicators at Data at any given time; time delay matrix The first line It is related to the key indicator vector The sequence has no time delay, the second row has a time delay of 1, and so on until the last row has a time delay. A sequence of order; 3) Calculate process variables time delay matrix With key indicator vector Time-delay correlation coefficient vector obtained based on Pearson correlation ; 4) Select The time delay corresponding to the maximum value of the time delay correlation coefficient in As a process variable The time lag; 5) Time delay for the i-th process variable Perform correction to obtain the corrected time delay. The correction formula is as follows; (6) in, For the first The average dwell time in the subprocess corresponding to each process variable. This represents the floor function; The specific steps of step S4 are as follows: 1) Based on the actual production process sequence, the mineral processing process is divided into m sub-processes. The input process variables and output variables of sub-process k are respectively... and ,in The number of input process variables for subprocess k, and the output of subprocess m. These are the key indicator variables in the mineral processing process, where T is the number of data points. 2) Based on the coupling between the forward and reverse flows in the mineral processing flow, i.e., a portion of the output slurry from a later sub-process will flow back to an earlier sub-process, let the input of the slurry flowing from a later sub-process back to an earlier sub-process be k. ,in This represents the number of variables that are returned to subprocess k from subsequent subprocesses. 3) Unify the time delay between all process variables and key indicator variables in subprocess k as follows: ; 4) Based on the premise that the output of each sub-process is the result of the combined effects of its input, the return input, the output of the previous sub-process, and the operating conditions of that sub-process, a spatiotemporal structure model of the mineral processing process is established: 5) Establish an m+1 layer spatiotemporal depth LSTM network, and embed the constructed spatiotemporal structure model of the mineral processing process into the m+1 layer spatiotemporal depth LSTM network to obtain a spatiotemporal depth LSTM prediction model for key indicators of the mineral processing process, specifically: The spatiotemporal depth LSTM network represents the spatial progression of the mineral processing technology from sub-process 1 to sub-process m vertically, with the last layer representing the historical key indicator sequence. The external input of layer k is the relevant process variable data from sub-process k, and its hidden layer results are input to layer k+1. The external input of layer k+1 is the process variable data from sub-process k+1, progressing step by step to layer m+1. At the last time step of layer m+1, the hidden layer information of the LSTM network is output to the fully connected layer, and after passing through the activation function, the final predicted value of the key indicator is output. Horizontally, the network represents the unfolding of the entire process time series with a time step size of d, meaning each layer consists of d LSTM cell units. The input of the d-th cell unit is all the input variables of the current sub-process k corresponding to the key indicator variable at time t. The input variables from the d-th cell unit to the first cell unit are gradually delayed by one time point in the time step.

2. The method for predicting key indicators in the mineral processing process according to claim 1, characterized in that, Process variables related to key indicators include: mill feed rate, water feed rate, mill power, hydrocyclone feed flow rate, hydrocyclone feed concentration, hydrocyclone feed pressure, froth layer thickness in each rougher cell, reagent dosage in each rougher cell, aeration rate in each rougher cell, froth layer thickness in each cleaner cell, aeration rate, and reagent dosage.

3. The method for predicting key indicators in the mineral processing process according to claim 1, characterized in that, Preprocessing of the original dataset includes imputation of missing values ​​and standardization. The missing value imputation process employs different methods depending on the characteristics of the data from the ore dressing plant. For missing process data caused by on-site server failure or shutdown, forward and backward imputation methods are used. For missing data collected by sensors, the K-nearest neighbor imputation method is used. This method searches for the K nearest neighbors to the missing value sample in the entire dataset, assigns distance weights between the sample and other samples based on the Euclidean distance, and calculates the dot product of the non-missing value vector and the weight vector of the K samples to obtain the missing value sample value. Finally, the imputed dataset is obtained. , where p is the number of process variables in the imputed dataset, and T is the number of data points; For the imputed dataset p process variables in the data; For the imputed dataset T data samples; The standardization process uses the Z-Score method to evaluate the imputed dataset. Standardize to obtain the standardized dataset. .

4. The method for predicting key indicators in the mineral processing process according to claim 1, characterized in that, The process variables are filtered based on their maximum correlation coefficient. Specifically, it is determined whether the maximum correlation coefficient of each process variable is greater than a preset threshold. If it is, the process variable is retained; otherwise, it is removed, resulting in the final dataset. , where n is the number of process variables after filtering, and T is the number of data.

5. The method for predicting key indicators in the mineral processing process according to claim 1, characterized in that, The spatiotemporal structure model of the mineral processing process is established as shown below: (7) (8) (9) (10) in, It is the selection function of subprocess k, which maps the non-linear relationship between the input and output of the subprocess and is learned through a deep learning network; Represents the k-th subprocess in td k All input variables at any given time, including the input process variables of the current k-th subprocess itself. And the variables that flow back from the backward subprocess to the k-th subprocess. ; d represents the input time step; It is the output of the kth subprocess; This represents the key performance indicator output value at time t, which is also the output of the m-th subprocess at time t. ; Historical information for key indicator variables; (11) Because there is a time lag between each subprocess and the key variables, the key indicator variable at time t corresponds to time td. k The k-th layer network input at time k This input is used to predict the key performance indicator value. The information dissemination process is as follows: Input Gate: (12) Forgotten Gate: (13) Output gate: (14) (15) (16) (17) Final key indicator forecasts: (18) in, , , The k-th layer network at td k The input gate, forget gate, and output gate at each time step; It is the sigmoid activation function. These are the external input variables of the k-th layer network, including the original input variables and return input variables of the corresponding subprocess k. and td k The hidden layer output and td of the (k-1)th layer of the network at time k k The output of the hidden layer of the k-th layer network at time -1 , , These are the weights of the input gate, forget gate, and output gate of the k-th layer network, respectively. , , , For the corresponding network bias, For td k Input the door status information at all times. and For the corresponding weights and biases, For the k-th layer network td k Cell state at time -1 These are the weights for the fully connected layer. It is the tanh activation function. This indicates that corresponding elements in the corresponding vectors are multiplied together. This is the hidden layer output of the last layer at time t-2. This represents the true value of the key indicator at time t-1.

6. The method for predicting key indicators in the mineral processing process according to claim 1, characterized in that, When training a spatiotemporal depth LSTM model for predicting key indicators in mineral processing using a training dataset, the mean squared error (MSE) is chosen as the loss function during training. (19) in The number of training samples, The true value of the key indicator at time t. Here are the predicted values ​​of key indicators at time t.

7. The method for predicting key indicators in the mineral processing process according to claim 1, characterized in that, A spatiotemporal depth LSTM prediction model for key indicators in mineral processing was trained using a time-based error backpropagation algorithm and the Adam optimizer.

8. A key indicator prediction system for mineral processing based on the spatiotemporal structure of the process, characterized in that, The system includes a mineral processing site data acquisition server, a data storage server, a deep learning server, and multiple display terminals for displaying predicted values ​​of key indicators. The deep learning server includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor implements the method as described in any one of claims 1 to 7.

Images