Graph neural network-based overflow ball mill granularity prediction method and related device
Through the method based on graph neural network, an undirected graph of the rotary overflow particle size and particle size prediction model is constructed, which solves the problem of low particle size detection efficiency of the ball mill, real-time particle size prediction is realized, cost is reduced, and resource utilization and production efficiency are improved.
Patent Information
- Application Number
- CN202510412695.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-03
- Publication Date
- 2025-08-08
AI Technical Summary
In the prior art, the particle size detection efficiency of ball mills is low, and relying on manual inspection and online particle size instruments has problems such as low detection frequency, complex maintenance and high cost, making it difficult to meet the real-time optimization control needs.
Using a graph neural network-based method, we collect multi-source data of overflow ball mills, perform normalization and Kalman filtering, construct a rotary overflow particle size undirected graph, train a particle size prediction model, and realize real-time particle size prediction.
It improves the efficiency of particle size detection, reduces production costs, enhances resource utilization, optimizes the production process, and improves the separation efficiency and resource recovery rate of subsequent processes.
Smart Images
Figure CN120448722A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of engineering equipment, and more specifically, relates to an overflow ball mill particle size prediction method based on graph neural network, an overflow ball mill particle size prediction device, and a computer-readable storage medium. Background Art
[0002] In the field of mineral resource processing, grinding and classification plays a key role in ore separation. The process involves the coordinated operation of multiple links, including ball mills and cyclones. The ball mill, in particular, is a core piece of equipment. Its discharge particle size directly affects the separation efficiency and resource recovery rate of subsequent ore sorting and other processes, making it a critical process node with a ripple effect.
[0003] Among related technologies, particle size detection relies primarily on two methods: the first is traditional manual inspection, where operators periodically take samples and make tactile judgments. This method has inherent drawbacks such as high subjectivity, low detection frequency (usually every 2-4 hours), and poor data continuity. The second is online particle size analyzers based on optical principles. Although they can achieve automatic detection, they face multiple limitations in practical applications: First, the instruments are sensitive to water turbidity and are prone to measurement failure due to freezing of the circulating water in cold winter regions. Second, the equipment requires manual calibration and maintenance at least once a week, which places strict demands on the professional quality of the maintenance team. Third, the data collection interval of existing online systems is still as long as 15-30 minutes, which is difficult to meet the needs of real-time optimization and control. According to statistics, due to these technical limitations, approximately 38% of mineral processing plants worldwide are forced to adopt a hybrid manual and instrument detection model, resulting in an increase of 12-15% in annual comprehensive production costs.
[0004] Therefore, how to improve the efficiency of particle size detection, reduce costs in the production line, and improve resource utilization are key issues that technical personnel in this field are concerned about. Summary of the Invention
[0005] The purpose of this application is to provide an overflow ball mill particle size prediction method, overflow ball mill particle size prediction device, and computer-readable storage medium based on graph neural network to improve the efficiency of particle size detection, reduce costs in the production line, and improve resource utilization.
[0006] In response to the above defects or improvement needs of the prior art, the present invention provides a method for predicting particle size of an overflow ball mill based on a graph neural network, comprising:
[0007] Collect mill production process data based on the grinding and grading process flow data of the overflow ball mill; wherein the grinding and grading process flow data includes: process design data, process flow chart data, operation manual data, expert experience data, and literature data; the mill production process data includes: ball mill current data, rotary overflow particle size data, rotary feed pressure data, feed capacity data, mill load data, particle size data, rotary feed concentration data, rotary feed flow data, supplementary water flow data, and ball addition data;
[0008] performing normalization preprocessing and Kalman filter smoothing processing on the mill production process data to obtain processed production process data;
[0009] Constructing an undirected graph of overflow particle size according to the process data of the grinding and classification process, and constructing a particle size prediction model based on a graph neural network based on the undirected graph of overflow particle size;
[0010] Performing parameter training on the particle size prediction model based on the processed production process data to obtain an overflow ball mill particle size prediction model based on a graph neural network;
[0011] Real-time production data is acquired, the real-time production data is input into the overflow ball mill particle size prediction model, and key indicator prediction results are output; wherein the real-time production data is time series data.
[0012] Optionally, the normalization process includes:
[0013] The mill production process data is normalized using the Min-Max normalization formula to obtain normalized process data; wherein the mapping range of the Min-Max normalization formula is [-1, 1], and the formula is:
[0014] X={x i |=1, 2, 3...I}, X'={x' i |=1, 2, 3...I};
[0015] Among them, I is the total number of selected production process parameters, X is the original data set, and X' is the normalized data set.
[0016] Optionally, the Kalman filter smoothing process includes:
[0017] The grinding and classification process data is smoothed by Kalman filtering to obtain process data from which noise has been filtered.
[0018] Optionally, parameter training is performed on the particle size prediction model based on the processed production process data to obtain an overflow ball mill particle size prediction model based on a graph neural network, including:
[0019] The processed production process data is used as a training sample; wherein the label of each training sample is the fault condition data corresponding to the system when the last set of data in the multi-source time series historical data is obtained;
[0020] Parameter training is performed on the particle size prediction model based on the training samples to obtain an overflow ball mill particle size prediction model based on a graph neural network.
[0021] Optionally, the process of acquiring the mill load data includes:
[0022] The mill load data is obtained by calculating the grinding sound collected by the intelligent grinding sound measuring instrument, the mill current, and the amplitude obtained by the vibration sensor; wherein the mill load data includes perfect load, good load, average load, poor load, and poor load.
[0023] Optionally, the structure of the granularity prediction model includes:
[0024] The input layer is the training data set after input processing;
[0025] The GCN layer is a graph neural network layer that calculates and outputs results based on the shape of the graph and the input;
[0026] The activation function is the Rule function;
[0027] The output layer outputs the prediction result, which is the predicted value.
[0028] Optionally, the pooling polymerization process of the overflow ball mill particle size prediction model includes:
[0029] Aggregate related nodes in the graph, merge the values of related devices, and use the average pooling method for pooling;
[0030] After the values of each device are aggregated, the aggregated graph is aggregated again according to the grinding process, and the key indicator prediction results are output.
[0031] Optionally, the training process of the granularity prediction model uses a cross entropy loss function to optimize neural network parameters.
[0032] The present application also provides an overflow ball mill particle size prediction device, comprising:
[0033] memory for storing computer programs;
[0034] A processor is used to implement the steps of the overflow ball mill particle size prediction method as described above when executing the computer program.
[0035] The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the overflow ball mill particle size prediction method as described above are implemented.
[0036] The present application provides a method for predicting the particle size of an overflow ball mill based on a graph neural network, comprising: collecting mill production process data according to the grinding and grading process flow data of the overflow ball mill; wherein the grinding and grading process flow data include: process design book data, process flow chart data, operation manual data, expert experience data, and literature data; the mill production process data include: ball mill current data, rotary overflow particle size data, rotary feed pressure data, feed processing capacity data, mill load data, block size data, rotary feed concentration data, rotary feed flow data, supplementary water flow data, and ball addition amount data; for the The mill production process data is normalized preprocessed and Kalman filter smoothed to obtain processed production process data; an undirected graph of overflow particle size is constructed according to the grinding and grading process flow data, and a particle size prediction model based on a graph neural network is constructed based on the overflow particle size undirected graph; parameter training is performed on the particle size prediction model based on the processed production process data to obtain an overflow ball mill particle size prediction model based on a graph neural network; real-time production data is acquired, the real-time production data is input into the overflow ball mill particle size prediction model, and key indicator prediction results are output; wherein, the real-time production data is time series data.
[0037] It has the following beneficial effects:
[0038] By comprehensively collecting multi-source data such as process design documents, operating manuals, and expert experience, rich information is provided for prediction. After normalization and Kalman filtering, data quality is improved, making model training more stable and efficient. The construction of an undirected graph and graph neural network model for vortex overflow particle size can deeply explore the complex potential relationships between data and accurately grasp the factors affecting particle size. The prediction of real-time production data allows operators to obtain the prediction results of key particle size indicators in a timely manner, adjust production parameters accordingly, optimize the production process, improve the separation efficiency and resource recovery rate of subsequent processes, reduce costs, enhance the company's market competitiveness, effectively solve the shortcomings of traditional particle size detection methods, and bring significant benefits to mining production. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.
[0040] Figure 1A flowchart of a method for predicting particle size of an overflow ball mill based on a graph neural network provided in an embodiment of the present application;
[0041] Figure 2 A schematic diagram of the structure of an overflow ball mill particle size prediction device based on a graph neural network provided in an embodiment of the present application;
[0042] Figure 3 This is a schematic diagram of the structure of the overflow ball mill particle size prediction device provided in the embodiment of the present application. DETAILED DESCRIPTION
[0043] The core of this application is to provide an overflow ball mill particle size prediction method, overflow ball mill particle size prediction device, and computer-readable storage medium based on graph neural network to improve the efficiency of particle size detection, reduce costs in the production line, and improve resource utilization.
[0044] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0045] The following is an example to illustrate the overflow ball mill particle size prediction method based on graph neural network provided by this application.
[0046] Please refer to Figure 1 , Figure 1 This is a flowchart of a method for predicting particle size of an overflow ball mill based on a graph neural network provided in an embodiment of the present application.
[0047] In this embodiment, the method may include:
[0048] S101, collecting mill production process data based on the grinding and classification process data of the overflow ball mill; wherein the grinding and classification process data includes: process design data, process flow chart data, operation manual data, expert experience data, and literature data; the mill production process data includes: ball mill current data, rotary overflow particle size data, rotary feed pressure data, feed capacity data, mill load data, particle size data, rotary feed concentration data, rotary feed flow data, supplementary water flow data, and ball addition data;
[0049] This step collects data related to overflow ball mill production from various sources, comprehensively capturing information on various factors influencing ball mill discharge particle size, providing a data foundation for subsequent analysis and prediction. Grinding and grading process data encompasses multiple aspects of design, operation, and experience, while mill production process data represents key parameters reflecting the equipment's real-time operating status.
[0050] This step obtains relevant data by reading the process design document, analyzing the process flow chart, and consulting the operation manual; using sensors and instruments to collect real-time production data such as ball mill current and vortex overflow particle size; at the same time, integrating expert experience and data from relevant literature.
[0051] It can be seen that this step ensures the integrity and diversity of the data, making subsequent analysis and prediction more comprehensive and accurate, and providing rich data support for establishing a reliable particle size prediction model.
[0052] S102, performing normalization preprocessing and Kalman filter smoothing processing on the mill production process data to obtain processed production process data;
[0053] Based on S101, the normalization preprocessing in this step maps the mill production process data of different magnitudes and distribution ranges to a unified interval, eliminating the impact of data dimension differences on model training; the Kalman filter smoothing processing utilizes the characteristics of the Kalman filter algorithm to extract real signals from noisy data and improve data quality.
[0054] Among them, the normalization processing uses the Min-Max normalization formula to transform the original data set X according to the formula X′=λmax(X)-min(X)X-min(X) (the mapping range is [-1,1]) to obtain the normalized process data; the Kalman filter smoothing processing is based on the Kalman filter algorithm to iteratively calculate the grinding and grading process flow data to obtain the process data with filtered noise.
[0055] Among them, normalization processing makes model training more stable and converges faster, thereby improving model training efficiency; Kalman filtering smoothing processing removes noise interference, improves data accuracy, and helps to improve the accuracy of the granularity prediction model.
[0056] S103, constructing an undirected graph of overflow particle size based on the grinding and classification process data, and constructing a particle size prediction model based on a graph neural network based on the undirected graph of overflow particle size;
[0057] Based on S102, this step aims to construct an undirected graph of overflow particle size based on the grinding and grading process data, and represent the various parameters and their relationships in the ball mill production process in the form of a graph. The nodes in the graph represent different parameters, and the edges represent the associations between the parameters. Based on the undirected graph, a particle size prediction model based on a graph neural network is constructed. The graph neural network's ability to process graph-structured data is used to mine the complex hidden relationships between the data, thereby realizing the prediction of particle size.
[0058] Among them, the nodes and edges of the graph are determined according to the physical relationship and logical connection between the data, and an undirected graph with spiral overflow granularity is constructed; the model structure includes the input layer (the training data set after input processing), the GCN layer (the graph neural network layer, which calculates the output results based on the graph structure and input data), the activation function (the ReLU function is selected to enhance the nonlinear expression ability of the model) and the output layer (output predicted values).
[0059] It can be seen that this step can effectively process data with complex correlations and fully explore the potential connections between production process parameters. Compared with traditional models, it can more accurately capture the factors affecting particle size and improve the accuracy of particle size prediction.
[0060] S104, performing parameter training on a particle size prediction model based on the processed production process data to obtain a particle size prediction model for an overflow ball mill based on a graph neural network;
[0061] Building on S103, this step aims to use processed production process data as training samples and adjust model parameters to ensure that the model output is as close to the actual granularity data as possible. The fault condition data corresponding to the last set of data acquired from the multi-source time series historical data is used as a label, allowing the model to learn the relationship between fault conditions and production process data, thereby optimizing the model's ability to predict granularity.
[0062] In this step, the processed production process data is divided into training samples, each of which contains production process parameters for multiple time steps. The cross-entropy loss function is used to calculate the difference between the model prediction results and the labels. The weight parameters of the model are continuously adjusted through the back-propagation algorithm until the model converges, resulting in an overflow ball mill particle size prediction model based on a graph neural network.
[0063] It can be seen that this step enables the model to adapt to the data characteristics and laws in actual production, improves the accuracy and reliability of the model in predicting the particle size of overflow ball mills, and provides strong support for particle size control in actual production.
[0064] S105, acquiring real-time production data, inputting the real-time production data into the overflow ball mill particle size prediction model, and outputting key indicator prediction results; wherein the real-time production data is time series data.
[0065] Based on S104, this step aims to obtain real-time production data (time series data) and input it into the trained overflow ball mill particle size prediction model. The model processes and analyzes the input data according to the learned rules, outputs the key indicator prediction results, and realizes the real-time prediction of the ball mill discharge particle size.
[0066] In this step, real-time production data of the ball mill, such as ball mill current and rotary pressure, are collected through real-time monitoring equipment. After preprocessing these data according to the model input requirements, they are input into the trained model. After calculation, the model outputs the predicted particle size and other key indicator results in the output layer.
[0067] It can be seen that this step can grasp the particle size of the ball mill discharge in real time, help operators adjust production parameters in time, optimize the production process, improve the separation efficiency and resource recovery rate of subsequent magnetic separation, flotation and other processes, and reduce production costs.
[0068] In summary, this embodiment provides rich information for prediction by comprehensively collecting multi-source data such as process design documents, operation manuals, and expert experience. After normalization and Kalman filtering, the data quality is improved, making the model training more stable and efficient. The undirected graph and graph neural network model of vortex overflow particle size are constructed to deeply explore the complex potential relationships between data and accurately grasp the factors affecting particle size. The real-time production data is predicted, allowing operators to obtain the prediction results of key particle size indicators in a timely manner, and adjust the production parameters accordingly, optimize the production process, improve the separation efficiency and resource recovery rate of subsequent processes, reduce costs, enhance the market competitiveness of enterprises, effectively solve the shortcomings of traditional particle size detection methods, and bring significant benefits to mining production.
[0069] Optional normalization process, including:
[0070] The Min-Max normalization formula is used to normalize the mill production process data to obtain normalized process data. The mapping range of the Min-Max normalization formula is [-1, 1], and the formula is:
[0071] X={x i |i=1, 2, 3...I}, X'={x' i |i=1, 2, 3...I};
[0072] Among them, I is the total number of selected production process parameters, X is the original data set, and X' is the normalized data set.
[0073] In this alternative, different production process data have different dimensions and ranges, which can affect the effectiveness and efficiency of model training. The Min-Max normalization formula, based on the principle of linear transformation, maps the original data to the interval [-1, 1] to eliminate dimensional differences in the data, allowing different data to be compared and processed on the same scale. This helps the model better learn the relationships between data features.
[0074] This option uses a formula to calculate mill process data. First, the maximum and minimum values for each process parameter are determined. Then, each raw data point is substituted into the formula to calculate the normalized data. This normalizes the mill process data and produces normalized process data.
[0075] It can be seen that after normalization in this step, the data distribution is more even, which avoids the dominant role of certain features in model training due to their large values, improves the stability and convergence speed of model training, and thus improves the prediction accuracy and efficiency of the granular prediction model based on graph neural network.
[0076] Optional Kalman filter smoothing process, including:
[0077] The grinding and classification process data is smoothed by Kalman filtering to obtain process data with filtered noise.
[0078] In this alternative, the data from the grinding and grading process will inevitably be mixed with noise during collection and transmission. This noise will interfere with the data's true characteristics and hinder the model's learning of data patterns. The Kalman filter is an algorithm that uses the linear system state equation and observed input and output data to optimally estimate the system state. It iteratively updates the state estimate based on the previous state estimate and the current observed value, effectively filtering out noise and preserving the data's true trends.
[0079] This alternative takes grinding and grading process data as input and processes it according to the Kalman filter algorithm. First, the system is initialized to determine the initial state estimate and covariance matrix. Then, at each time step, the Kalman gain is calculated based on the current observation data and the state estimate from the previous time step. The state estimate is then updated using the Kalman gain to obtain the optimal state estimate for the current time step. Finally, the covariance matrix is updated to prepare for calculations in the next time step. This iterative process results in noise-filtered process data.
[0080] It can be seen that this optional solution removes the noise interference in the data, making the data more realistically reflect the production process conditions of the ball mill, providing a more reliable data basis for the subsequent particle size prediction model, helping to improve the accuracy of model prediction and reduce prediction errors.
[0081] Optionally, parameter training of the particle size prediction model is performed based on the processed production process data to obtain an overflow ball mill particle size prediction model based on a graph neural network, including:
[0082] Step 1: Use the processed production process data as training samples; the label of each training sample is the fault condition data corresponding to the system when the last set of data in the multi-source time series historical data is obtained;
[0083] Step 2: Parameter training of the particle size prediction model is performed based on the training samples to obtain an overflow ball mill particle size prediction model based on graph neural network.
[0084] This option uses processed production process data as training samples because, after preprocessing, this data better reflects the true patterns of the production process. The system fault condition data corresponding to the last set of data acquired from the multi-source time series historical data is used as labels. This allows the model to learn the association between production process data and fault conditions, thereby establishing a relationship between production process data and particle size prediction results. During training, model parameters are continuously adjusted to ensure that the model's predictions are as close to the labels as possible, achieving model optimization.
[0085] This optional solution first organizes the processed production process data according to a certain time series and format, and divides it into multiple training samples. Each training sample contains production process data for multiple time steps, such as ball mill current data, rotary pressure data, etc. Then, the label corresponding to each training sample is determined, that is, the fault condition data corresponding to the system when the last set of data in the multi-source time series historical data is obtained. Next, the training sample is input into the particle size prediction model, and the cross-entropy loss function is used to calculate the difference between the model prediction result and the label. Through the backpropagation algorithm, the weight parameters of the model are adjusted according to the loss value, and the training is iterated continuously until the model converges, resulting in an overflow ball mill particle size prediction model based on a graph neural network.
[0086] This training method enables the model to fully learn the complex relationship between production process data, fault conditions, and particle size, improving its generalization and prediction accuracy. By integrating fault condition data into training, the model can better account for various potential factors that may affect particle size when predicting particle size, providing a more reliable basis for particle size prediction in actual production.
[0087] Optionally, the process of obtaining mill load data includes:
[0088] The mill load data is obtained by calculating the grinding sound collected by the intelligent grinding sound measuring instrument, the mill current, and the amplitude obtained by the vibration sensor; among them, the mill load data is perfect load, good load, average load, poor load, and poor load.
[0089] In this option, mill load is a key factor affecting ball mill particle size. Mill load data is calculated using a comprehensive set of information, including grinding noise, mill current, and vibration amplitude captured by a vibration sensor. This is because these parameters reflect the mill's internal operating conditions from different perspectives. Changes in grinding noise reflect the grinding process, mill current is related to mill load, and vibration amplitude captured by the vibration sensor reflects mill operating stability. This combined information allows for a more accurate assessment of mill load.
[0090] This optional intelligent grinding noise meter collects grinding noise signals in real time, converts them into electrical signals for analysis and processing, collects mill current via a current sensor, and collects mill amplitude data via a vibration sensor. The collected grinding noise, mill current, and amplitude data are fed into a pre-defined calculation model, which performs a comprehensive calculation based on a specific algorithm (such as weighted average or other more complex algorithms). The results are then classified into different levels (perfect load, good load, average load, poor load, and poor load) according to pre-defined standards, generating mill load data.
[0091] According to the relationship between mill current and grinding noise, when the current is maximum, mill efficiency is highest, theoretically resulting in the best particle size and a perfect mill load. As grinding noise increases, the current decreases, the ball mill begins to idle, and the particle size decreases from coarse to fine. Similarly, as grinding noise decreases, the current decreases, the ball mill begins to grind fully, and the particle size decreases from coarse to fine. Based on the particle size and process requirements, four values are assigned: good, fair, poor, and poor. The vibration method and soft sensing method also use this method to label mill load as perfect, good, fair, poor, and poor.
[0092] This optional solution accurately captures mill load data, providing a more comprehensive understanding of the ball mill's operating status and more accurate parameter information for particle size prediction. By monitoring mill load, operators can promptly identify abnormalities in mill operation and adjust production parameters appropriately to ensure stable ball mill operation, thereby improving the accuracy of particle size prediction and the stability of the production process.
[0093] Optionally, the structure of the granularity prediction model includes:
[0094] The input layer is the training data set after input processing;
[0095] The GCN layer is a graph neural network layer that calculates and outputs results based on the shape of the graph and the input;
[0096] Activation function selects Rule function:
[0097] The output layer finally outputs the result, which is the predicted value.
[0098] The input layer of the granularity prediction model in this optional solution is responsible for receiving the processed training dataset, providing a data foundation for subsequent model calculations. Based on the graph structure and input data, the GCN layer (graph neural network layer) aggregates and transforms node features through graph convolution operations, mining the relationships between nodes in the graph and learning the intrinsic feature representation of the data. The ReLU function, as an activation function, can introduce nonlinear factors, breaking the limitations of linear models, enhancing the model's expressive power, and enabling the model to learn more complex functional relationships. The output layer outputs the final predicted value based on the calculation results of the previous layers, achieving granularity prediction.
[0099] At the input layer, this optional solution organizes and inputs the production process data, which has undergone preprocessing such as normalization and filtering, according to the model's input format requirements. The GCN layer, based on the structure of the undirected graph of vortex granularity and the input data, defines a graph convolution kernel and a weight matrix to perform convolution operations on node features and continuously update the node's feature representation. During the calculation process, the ReLU function is used to activate the output of the GCN layer, that is, the output of each neuron takes the maximum value (0, x). Finally, the output layer performs a linear transformation (such as weighted summation) on the activated result to obtain the predicted granularity value.
[0100] This alternative model structure is well-designed, leveraging the advantages of graph neural networks in processing graph-structured data to effectively learn the complex relationships between production process parameters. The introduction of the ReLU function enhances the model's nonlinear expression capabilities, enabling it to more accurately fit the particle size variations observed in actual production, improving the accuracy and reliability of particle size predictions.
[0101] Optional, pooled aggregation process for overflow ball mill particle size prediction model, including:
[0102] Step 1: Aggregate the relevant nodes in the graph, merge the values of related devices, and use the average pooling method for pooling;
[0103] Step 2: After the values of each device are aggregated, the aggregated graph is aggregated again according to the grinding process, and the key indicator prediction results are output.
[0104] This optional solution aggregates graph-related nodes and combines the values of related devices in the particle size prediction model to reduce data dimensionality, extract key information, and preserve the relationships between the data. The average pooling method, based on the principle of averaging, processes the features of aggregated nodes, enabling the model to reduce its computational complexity while retaining key information. After the values of each device are aggregated, they are aggregated again based on the grinding process. This allows for a comprehensive consideration of the impact of different devices on particle size from the perspective of the overall production process, resulting in key indicator prediction results that are more consistent with actual production conditions.
[0105] First, based on the graph structure and the relationships between nodes, the relevant nodes and devices to be aggregated are identified. For each set of nodes to be aggregated, average pooling is used to calculate the average of the node features in the set to obtain the aggregated node features. The aggregated results for each device are then grouped according to the grinding process. Average pooling or other appropriate aggregation operations are again performed on each data set to obtain the predicted key indicators and output them.
[0106] This optional approach reduces data redundancy through aggregation, lowering the model's computational workload and improving its efficiency. Furthermore, by further aggregating the data from the perspective of the grinding process as a whole, it better reflects the combined impact of each step in the production process on particle size, making the predictions more aligned with actual production conditions and providing a more valuable reference for production decision-making.
[0107] Optionally, during the training process of the granularity prediction model, a cross entropy loss function is used to optimize the neural network parameters.
[0108] In this alternative, the cross-entropy loss function measures the difference between the model's predictions and the true labels. During granular prediction model training, the goal is to make the predictions as close as possible to the actual granular data. The cross-entropy loss function is based on information theory: the closer the predictions are to the true labels, the smaller the loss; conversely, the smaller the loss is. By minimizing the cross-entropy loss function, the model can continuously adjust its parameters to optimize its predictions.
[0109] During model training, the model's predicted granularity and the actual granularity label are fed into a cross-entropy loss function for calculation. Assuming the predicted granularity is p and the actual label is q, the cross-entropy loss function is calculated as \(H(p,q) = -\sum_{i}q_{i}\log(p_{i})\). During backpropagation, the model's weight parameters are adjusted using an optimization algorithm (such as stochastic gradient descent) based on the gradient of the cross-entropy loss function. Training is then iterated continuously to gradually reduce the cross-entropy loss value until the model converges.
[0110] The cross-entropy loss function effectively measures the difference between model predictions and actual results, providing a clear optimization direction for model training. By using the cross-entropy loss function to optimize neural network parameters, the model's prediction accuracy can be improved, enabling it to better fit actual production particle size data, thereby enhancing the performance of the graph neural network-based overflow ball mill particle size prediction model.
[0111] The following is an introduction to an overflow ball mill particle size prediction device based on a graph neural network provided in an embodiment of the present application. The overflow ball mill particle size prediction device based on a graph neural network and the overflow ball mill particle size prediction method based on a graph neural network described below can be referenced to each other.
[0112] Please refer to Figure 2 , Figure 2 A schematic structural diagram of an overflow ball mill particle size prediction device based on a graph neural network provided in an embodiment of the present application.
[0113] In this embodiment, the device may include:
[0114] The data acquisition module 100 is used to collect mill production process data based on the grinding and classification process data of the overflow ball mill; wherein the grinding and classification process data includes: process design data, process flow chart data, operation manual data, expert experience data, and literature data; the mill production process data includes: ball mill current data, rotary overflow particle size data, rotary feed pressure data, feed processing capacity data, mill load data, particle size data, rotary feed concentration data, rotary feed flow data, supplementary water flow data, and ball addition data;
[0115] The data preprocessing module 200 is used to perform normalization preprocessing and Kalman filter smoothing processing on the mill production process data to obtain processed production process data;
[0116] A model building module 300 is used to construct an undirected graph of overflow particle size based on the grinding and classification process data, and to construct a particle size prediction model based on a graph neural network based on the undirected graph of overflow particle size;
[0117] A model training module 400 is used to perform parameter training on a particle size prediction model based on processed production process data to obtain a particle size prediction model for an overflow ball mill based on a graph neural network;
[0118] The real-time prediction module 500 is used to obtain real-time production data, input the real-time production data into the overflow ball mill particle size prediction model, and output key indicator prediction results; wherein the real-time production data is time series data.
[0119] This application also provides overflow ball mill particle size prediction equipment, please refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of an overflow ball mill particle size prediction device provided in an embodiment of the present application. The overflow ball mill particle size prediction device may include:
[0120] memory for storing computer programs;
[0121] The processor can implement the steps of any of the above-mentioned overflow ball mill particle size prediction methods based on graph neural network when executing the computer program.
[0122] like Figure 3 FIG. 1 is a schematic diagram of the structure of an overflow ball mill particle size prediction device, which may include a processor 10, a memory 11, a communication interface 12, and a communication bus 13. The processor 10, the memory 11, and the communication interface 12 communicate with each other via the communication bus 13.
[0123] In the embodiment of the present application, the processor 10 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field programmable gate array, or other programmable logic devices.
[0124] The processor 10 may call a program stored in the memory 11 . Specifically, the processor 10 may execute the operations in the embodiment of the abnormal IP identification method.
[0125] The memory 11 is used to store one or more programs. The program may include program code, and the program code includes computer operating instructions. In the embodiment of the present application, the memory 11 stores at least a program for implementing the following functions:
[0126] Grinding mill production process data is collected based on the grinding and grading process data of the overflow ball mill; the grinding and grading process data includes: process design data, process flow chart data, operation manual data, expert experience data, and literature data; the grinding mill production process data includes: ball mill current data, rotary overflow particle size data, rotary feed pressure data, feed capacity data, mill load data, particle size data, rotary feed concentration data, rotary feed flow data, supplementary water flow data, and ball addition data;
[0127] Perform normalization preprocessing and Kalman filter smoothing on the mill production process data to obtain processed production process data;
[0128] An undirected graph of overflow particle size is constructed based on the grinding and classification process data, and a particle size prediction model based on a graph neural network is constructed based on the undirected graph of overflow particle size.
[0129] The particle size prediction model was trained based on the processed production process data to obtain a particle size prediction model for overflow ball mills based on graph neural networks.
[0130] Real-time production data is acquired, input into the overflow ball mill particle size prediction model, and key indicator prediction results are output; wherein the real-time production data is time series data.
[0131] In one possible implementation, the memory 11 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function, etc.; the data storage area may store data created during use.
[0132] In addition, the memory 11 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device or other volatile solid-state storage device.
[0133] The communication interface 12 may be an interface of a communication module, and is used to connect to other devices or systems.
[0134] Of course, it needs to be explained that Figure 3 The structure shown does not constitute a limitation on the overflow ball mill particle size prediction device in the embodiment of the present application. In actual application, the overflow ball mill particle size prediction device may include Figure 3 More or fewer components than shown, or combinations of certain components.
[0135] The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of any of the above-mentioned overflow ball mill particle size prediction methods based on graph neural networks can be implemented.
[0136] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc., which can store program codes.
[0137] For an introduction to the computer-readable storage medium provided in this application, please refer to the above method embodiment, and this application will not go into details here.
[0138] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.
[0139] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0140] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
[0141] The above is a detailed introduction to the overflow ball mill particle size prediction method based on graph neural network, overflow ball mill particle size prediction device, and computer-readable storage medium provided by this application. This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method of this application and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to this application, and these improvements and modifications also fall within the scope of protection of the claims of this application.
Claims
1. A method for predicting particle size of overflow ball mill based on graph neural network, characterized in that: include: Collect mill production process data based on the grinding and grading process flow data of the overflow ball mill; wherein the grinding and grading process flow data includes: process design data, process flow chart data, operation manual data, expert experience data, and literature data; the mill production process data includes: ball mill current data, rotary overflow particle size data, rotary feed pressure data, feed capacity data, mill load data, particle size data, rotary feed concentration data, rotary feed flow data, supplementary water flow data, and ball addition data; performing normalization preprocessing and Kalman filter smoothing processing on the mill production process data to obtain processed production process data; Constructing an undirected graph of overflow particle size according to the grinding and classification process data, and constructing a particle size prediction model based on a graph neural network based on the undirected graph of overflow particle size; Performing parameter training on the particle size prediction model based on the processed production process data to obtain an overflow ball mill particle size prediction model based on a graph neural network; Real-time production data is acquired, the real-time production data is input into the overflow ball mill particle size prediction model, and key indicator prediction results are output; wherein the real-time production data is time series data.
2. The overflow ball mill particle size prediction method according to claim 1, characterized in that: The normalization process includes: The mill production process data is normalized using the Min-Max normalization formula to obtain normalized process data; wherein the mapping range of the Min-Max normalization formula is [-1, 1], and the formula is: Among them, I is the total number of selected production process parameters, X is the original data set, and X' is the normalized data set.
3. The overflow ball mill particle size prediction method according to claim 2, characterized in that: The Kalman filter smoothing process includes: The grinding and classification process data is smoothed by Kalman filtering to obtain process data from which noise has been filtered.
4. The overflow ball mill particle size prediction method according to claim 3, characterized in that: Parameter training is performed on the particle size prediction model based on the processed production process data to obtain an overflow ball mill particle size prediction model based on a graph neural network, including: The processed production process data is used as a training sample; wherein the label of each training sample is the fault condition data corresponding to the system when the last set of data in the multi-source time series historical data is obtained; Parameter training is performed on the particle size prediction model based on the training samples to obtain an overflow ball mill particle size prediction model based on a graph neural network.
5. The overflow ball mill particle size prediction method according to claim 4, characterized in that: The process of acquiring the mill load data includes: The mill load data is obtained by calculating the grinding sound collected by the intelligent grinding sound measuring instrument, the mill current, and the amplitude obtained by the vibration sensor; wherein the mill load data is perfect load, good load, average load, poor load, and poor load.
6. The overflow ball mill particle size prediction method according to claim 5, characterized in that: The structure of the particle size prediction model includes: The input layer is the training data set after input processing; The GCN layer is a graph neural network layer that calculates and outputs results based on the shape of the graph and the input; The activation function is the Rule function; The output layer outputs the prediction result, which is the predicted value.
7. The overflow ball mill particle size prediction method according to claim 6, characterized in that: The pooling polymerization process of the overflow ball mill particle size prediction model includes: Aggregate related nodes in the graph, merge the values of related devices, and use the average pooling method for pooling; After the values of each device are aggregated, the aggregated graph is aggregated again according to the grinding process, and the key indicator prediction results are output.
8. The overflow ball mill particle size prediction method according to claim 7, characterized in that: The training process of the particle size prediction model uses a cross entropy loss function to optimize the neural network parameters.
9. An overflow ball mill particle size prediction device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the overflow ball mill particle size prediction method according to any one of claims 1 to 8 when executing the computer program.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the overflow ball mill particle size prediction method according to any one of claims 1 to 8 are implemented.