Prediction method of ash content of clean coal from dense medium separation based on multi-channel line graph of time series data
Through multi-channel linear graphing processing and improved Swin Transformer model, the problem that traditional prediction methods cannot adapt to dynamic sampling frequency is solved, and high-precision coal ash prediction is achieved, which improves production efficiency and stability.
Patent Information
- Application Number
- CN202411497947.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2025-02-28
- Estimated Expiration
- 2044-10-25
AI Technical Summary
The traditional coal ash prediction method relies on a constant sampling frequency and cannot adapt to the demand for coal preparation plants to dynamically adjust the sampling frequency during production, resulting in a reduction in prediction accuracy.
A multi-channel linear graph processing method based on time series data is adopted, combined with the improved Swin Transformer deep learning model, irregularly sampled time series data are processed, and a multi-channel composite data line chart is generated to achieve accurate prediction of coal ash.
It significantly improves the accuracy and reliability of coal ash prediction, adapts to the operating conditions of irregular sampling, shortens the prediction time, and improves production efficiency and stability of coal preparation process.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a method for predicting ash content of clean coal separated by heavy medium separation based on multi-channel line graphing of time series data. Background Art
[0002] In the production process of coal preparation plants, accurate prediction of clean coal ash content is crucial to optimizing the heavy medium separation process. However, the currently widely used rapid ash determination methods have obvious limitations, including long detection cycles, delayed results, and excessive reliance on the experience of on-site workers. These problems not only reduce production efficiency but also lead to waste of resources. Coal preparation plants have accumulated a wealth of historical data, in which clean coal ash content is closely related to key process parameters such as sorting density. Through advanced big data analysis technology, in-depth mining and comprehensive analysis of these historical data can effectively construct a mapping model between process parameters and clean coal ash content. This is not only feasible, but also provides coal preparation plants with an accurate basis for process optimization and quality control.
[0003] Traditional time series prediction methods usually assume that the data sampling frequency is fixed, that is, predictions are made based on regular sampling. However, in the actual production process, the coal preparation plant will dynamically adjust the sampling frequency of the raw coal ash content according to the sorting effect of the clean coal, resulting in an unstable sampling frequency, which affects the accuracy of the traditional prediction method. This irregular sampling pattern is contrary to the constant sampling frequency that traditional prediction methods rely on, which may reduce the prediction accuracy of these methods. To overcome this challenge, the present invention has developed a clean coal ash prediction method that adapts to irregular sampling, aiming to improve prediction accuracy and adapt to changes in actual working conditions.
[0004] In data preprocessing, line graphing refers to converting data into a line graph, which displays the changing trend of variables by connecting a series of data points. This visualization method makes the relationship between data more intuitive and easy to understand. In order to more comprehensively display the association between variables and improve the processing efficiency of the model, line graphing of irregular sampling data is used to generate multi-channel line graphs. Multi-channel line graphs can simultaneously display the changing trends of multiple variables and their mutual relationships by superimposing multiple related line graphs as different channels in the same graph. Combined with artificial intelligence technology, this method significantly improves the accuracy and reliability of clean coal ash content prediction, providing strong data support and decision-making basis for coal preparation plants. Summary of the invention
[0005] In view of the above-mentioned technical problems existing in the prior art, the present invention proposes a method for predicting the ash content of clean coal by heavy medium separation based on multi-channel line graphing of time series data, introduces an advanced time series multi-channel line graph representation technology, and can flexibly handle the discontinuity of data for the sampling problem of raw coal ash content and key process parameters at irregular time intervals, thereby improving the application capability of artificial intelligence in coal preparation technology.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] A method for predicting ash content of clean coal by heavy medium separation based on multi-channel line graphing of time series data is provided, which converts time series data into multi-channel line graphs to achieve accurate prediction of clean coal ash content, and comprises the following steps:
[0008] Step 1: Collect historical big data of the coal preparation plant and build a coal preparation plant database;
[0009] Step 2: Perform data preprocessing and divide the preprocessed data into training set and test set in proportion;
[0010] Step 3, generate a multi-channel composite data line chart;
[0011] Step 4: Design a deep learning model based on the improved Swin Transformer;
[0012] Step 5: Train the deep learning model based on the improved Swin Transformer;
[0013] Step 6: Model evaluation and application.
[0014] Furthermore, in step 1, the clean coal ash content, raw coal ash content and key process parameters in the heavy medium separation process are collected through the production monitoring system of the coal preparation plant and manual records; the key process parameters include but are not limited to sampling date, sampling time, magnetic material content, coal slime content, separation density, cyclone inlet pressure, and combined medium barrel liquid level; combined with big data analysis technology, the data is deeply mined and analyzed.
[0015] Furthermore, in step 2, data preprocessing includes correcting sensor errors, correcting recording errors, and formatting the data into a structure used by subsequent models.
[0016] Furthermore, in the step 3, the preprocessed data is used to generate multiple indicator line graphs based on time series. When generating, the key process parameters are set as the x-axis, the raw coal ash content is set as the y-axis, and the data points in the data set are represented as turning points of the line graph. The turning points are connected by straight lines to form a line graph showing the change of raw coal ash content over time; then, these line graphs are superimposed in the same graph in a multi-channel form according to the time series to construct a multi-channel composite data line graph.
[0017] Furthermore, in step 4, the deep learning model based on the improved Swin Transformer captures local and global information by constructing a hierarchical representation, which starts from small-sized blocks in early layers and gradually merges adjacent blocks in deeper layers; the input of the model is the multi-channel composite data line chart generated in step 3, and the output of the model is the ash content of clean coal; the deep learning model based on the improved Swin Transformer includes four stages, and stages one, two, and four all include a group of Swin Transformer blocks; stage three includes three groups of Swin Transformer blocks connected in series; a group of Swin Transformer blocks includes a window-based multi-head attention mechanism layer and an offset window-based multi-head attention mechanism layer connected in series in sequence; a residual-based multi-layer perceptron with a GELU activation function is connected between the window-based multi-head attention mechanism layer and the offset window-based multi-head attention mechanism layer.
[0018] Furthermore, the specific working process of the deep learning model based on the improved Swin Transformer is:
[0019] Step 4.1, input the multi-channel composite data line graph as an input image into the model, and divide the input image into blocks;
[0020] Step 4.2: Perform the first-stage linear embedding and Swin Transformer block processing to obtain the output features of the first stage;
[0021] Step 4.3, block merging and Swin Transformer block processing are performed on the output features of the first stage to obtain the output features of the second stage;
[0022] Step 4.4, continue to perform block merging and Swin Transformer block processing on the output features of the second stage to obtain the output features of the third stage;
[0023] Step 4.5: Continue to perform block merging and Swin Transformer block processing on the output features of the third stage to obtain the output features of the fourth stage, which is the final output of the model.
[0024] Furthermore, the calculation process of the Swin Transformer block is as follows:
[0025] (1);
[0026] (2);
[0027] (3);
[0028] (4);
[0029] in, For blocks The output features of the window-based multi-head attention mechanism layer; For blocks The output features of the offset window multi-head attention mechanism layer; , , Respectively for blocks ,piece ,piece Output features of residual-based multilayer perceptron; Represents a window-based multi-head attention mechanism layer; Represents a multi-head attention mechanism layer based on an offset window; represents a residual-based multilayer perceptron; Representation layer normalization;
[0030] In the Swin Transformer block, the input image is divided into several non-overlapping local windows based on the window multi-head attention mechanism layer, and then the local attention of each local window is calculated to obtain the local window self-attention matrix, which is specifically composed as follows:
[0031] (5);
[0032] in, represents the self-attention matrix; , , Represent the query matrix, key matrix, and value matrix respectively; Represents the Softmax function; represents the mapping matrix; is the bias matrix;
[0033] Then it is input into the offset window based multi-head attention mechanism layer through the residual based multi-layer perceptron.
[0034] Furthermore, in step 5, the deep learning model based on the improved Swin Transformer is trained based on the training set. The model uses the mean square error as the empirical loss to measure the difference between the predicted value and the actual value. The calculation formula is:
[0035] (6);
[0036] in, is the mean square error value; is the total number of samples; Representative The actual clean coal ash value of each sample; It is Predicted clean coal ash value for samples.
[0037] Furthermore, in step 6, after completing the model training, the model is evaluated using a test set; finally, the model with the best performance is applied to actual coal preparation production to achieve accurate prediction of clean coal ash content.
[0038] The beneficial technical effects brought about by the present invention are as follows.
[0039] 1. In the heavy medium coal preparation process, the traditional clean coal ash content prediction method mainly relies on the experience of on-site workers, and roughly estimates the clean coal ash content based on a series of indicators. This method has the problems of long detection cycle, delayed results, and over-reliance on workers' experience, resulting in low production efficiency and waste of resources. In order to solve these problems, the present invention introduces big data mining technology to conduct in-depth analysis of a large amount of historical data accumulated by the coal preparation plant, and combines it with a data-driven deep learning model to significantly improve the accuracy of clean coal ash content prediction, while improving production efficiency and reducing resource waste.
[0040] 2. Traditional time series prediction methods usually assume that the sampling time interval is fixed, that is, prediction is based on regular sampling. However, in actual production, the coal preparation plant will dynamically adjust the sampling frequency of the raw coal ash content according to the changes in the clean coal ash content, resulting in irregular sampling time intervals for the raw coal ash content, which poses a challenge to the accurate prediction of the clean coal ash content. To address this problem, the present invention adopts artificial intelligence technology, which can maintain a high prediction accuracy when processing irregularly sampled time series data. Compared with traditional manual prediction methods, the present invention not only greatly shortens the prediction time, but also significantly improves the prediction accuracy of the clean coal ash content, while effectively enhancing the stability of the coal preparation process. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a flow chart of the method for predicting ash content of clean coal separated by heavy medium separation based on multi-channel line diagram of time series data of the present invention.
[0042] Figure 2 It is a structural diagram of the deep learning model based on the improved Swin Transformer of the present invention.
[0043] Figure 3 It is a connection diagram between consecutive blocks in the Swin Transformer model of the present invention. DETAILED DESCRIPTION
[0044] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0045] A method for predicting ash content of clean coal separated by dense medium separation based on multi-channel line diagram of time series data, the process is as follows: Figure 1 As shown, the following steps are included:
[0046] Step 1, collect historical big data of the coal preparation plant and build a coal preparation plant database; the data in the embodiment of the present invention are collected from a coal preparation plant under Shandong Energy Zaozhuang Mining Group, specifically including the clean coal ash content, raw coal ash content and key process parameters in the heavy medium separation process, combined with big data analysis technology, these data are deeply mined and analyzed, and a solid foundation is laid for the subsequent clean coal ash prediction. Due to the huge amount of data, the present invention only shows part of the data. Table 1 is an example of regular sampling data of the data set, and Table 2 is an example of irregular sampling data of the data set. As shown in Tables 1 and 2, the key parameters of the heavy medium separation process cover multiple dimensions such as sampling date, sampling time, magnetic content, coal slime content, separation density, cyclone inlet pressure and combined medium barrel liquid level. These data are continuously sampled at intervals of 1 minute. In the daily operation of the coal preparation plant, production activities are divided into three 8-hour shifts. Each shift will conduct 3 to 5 raw coal floating and sinking experiments to measure the raw coal ash content according to actual production needs. Since the timing of the floating and sinking experiment has a certain uncertainty, the sampling time of the raw coal ash content is irregular, which in turn brings certain challenges to the prediction of the clean coal ash content. In view of the fact that historical data contains information that is useful for prediction, the present invention constructs samples in the following way: using the process parameters and raw coal ash data within 8 hours before the clean coal ash sampling time point to predict the clean coal ash content at that moment. Through this sample construction method, 18 months of heavy medium coal preparation process data were collected and sorted, and a data set containing 3290 samples was constructed.
[0047] Table 1 Data set rule sampling data example table
[0048] .
[0049] Table 2. Example table of irregular sampling data of the dataset
[0050] .
[0051] Step 2: preprocess the data and divide the preprocessed data into training set and test set in proportion; in order to ensure the quality and consistency of the data, the collected data is strictly preprocessed, including outlier and missing value preprocessing. In the present invention, it specifically includes steps such as correcting sensor errors and recording errors, and formatting the data into a structure suitable for subsequent model use to ensure high-quality data input. Data preprocessing can effectively prevent the reduction of clean coal ash content prediction accuracy due to partial data errors and improve prediction accuracy.
[0052] For example, the values with sorting density lower than 1000 are deleted. Since the irregular sampling time series prediction method is adopted, there is no need to interpolate the deleted data set.
[0053] Step 3, generate a multi-channel composite data line graph; after the data preprocessing in step 2 is completed, generate multiple time series-based indicator line graphs (such as a line graph of raw coal ash content changing over time). Subsequently, these line graphs are superimposed in the same graph in a multi-channel form according to the time series to construct a comprehensive multi-channel composite data line graph to provide input data for model training and testing. The present invention draws a two-dimensional line graph for the raw coal ash and each key process parameter to intuitively display their changing trends over time. When generating, the key process parameter is set to the x-axis, the raw coal ash is set to the y-axis, and the data points in the data set are represented as turning points of the line. The turning points are connected by straight lines to form a line graph showing the change of raw coal ash content over time. Taking the raw coal ash-time line graph as an example, the time is set to the x-axis, the raw coal ash is set to the y-axis, the data points in the data set are represented as turning points of the line, and the turning points are connected by straight lines to form a line graph showing the change of raw coal ash content over time. Drawing a line graph can accurately reflect the characteristics of the time series. For another example, in the sorting density-time line graph, time is the x-axis and the sorting density is the y-axis; in the clean coal ash-time line graph, time is the x-axis and the clean coal ash is the y-axis. In order to enable the model to better distinguish turning points from interpolated data points, some necessary marks are made for the turning points when drawing the line graph, such as the size and format of the points, and the thickness of the line can also help the model to achieve better performance. The same method is also used to draw line graphs of other process parameters changing over time. In this way, the present invention obtains multiple line graphs based on time series. Next, these line graphs are superimposed in the same graph in the form of multiple channels to construct a comprehensive multi-channel composite data line graph. This composite data line graph will serve as the final input of the model to visualize the irregular sampling time series, helping the model to better understand and predict the changing trends of process parameters and raw coal ash over time.
[0054] Step 4: Design a deep learning model based on the improved Swin Transformer; Design a deep learning model based on the improved Swin Transformer for the characteristics of multi-channel composite data line graphs. The model captures local and global information by building a hierarchical representation that starts with small-sized blocks in the early layers and gradually merges adjacent blocks in deeper layers. The structure of the model is shown in the figure below. Figure 2 As shown in the figure, there are four stages in total. Stage 1, stage 2, and stage 4 each contain a set of Swin Transformer blocks; stage 3 includes three sets of Swin Transformer blocks connected in series. A set of Swin Transformer blocks contains a Window-based Multi-head Self-attention (W-MSA) layer and a Shifted Window-based Multi-head Self-attention (SW-MSA) layer connected in series. A residual-based multi-layer perceptron ResMLP with a GELU activation function is connected between the W-MSA layer and the SW-MSA layer.
[0055] The specific working process of the deep learning model based on the improved Swin Transformer is:
[0056] Step 4.1, input the image into the model (i.e., multi-channel composite data line graph, also called multi-channel feature graph), and perform patch division (i.e., block division) on the image;
[0057] Step 4.2: Perform the first-stage linear embedding and Swin Transformer block processing to obtain the output features of the first stage;
[0058] Step 4.3, perform Patch merging (i.e., block merging) and Swin Transformer block processing on the output features of the first stage to obtain the output features of the second stage;
[0059] Step 4.4, continue to perform Patch merging (i.e., block merging) and SwinTransformer block processing on the output features of the second stage to obtain the output features of the third stage;
[0060] Step 4.5: Continue to perform Patch merging (i.e. block merging) and SwinTransformer block processing on the output features of the third stage to obtain the output features of the fourth stage, which is the final output of the model.
[0061] Figure 2The Swin Transformer block in stage one is marked with ×2, indicating that the block has two parts, namely the W-MSA layer and the SW-MSA layer. The W-MSA layer and the SW-MSA layer use the transfer multi-head self-attention window and the multi-head self-attention window respectively. When using the W-MSA layer, the input image is first divided into windows of size M×M (M represents the length or width of the window), and then self-attention is performed on each window separately to learn the information inside a single window, while the SW-MSA layer can learn the information between multiple W-MSA layers through offset. The SW-MSA layer is followed by the W-MSA layer, which is a series relationship. Similarly, the ×6 in stage three means that there are 3 groups of W-MSA layers and SW-MSA layers connected in series. The structure of the Swin Transformer block is as follows. Figure 3 As shown, the input is a multi-channel feature map and the output is a feature map. The specific calculation process is as follows:
[0062] (1);
[0063] (2);
[0064] (3);
[0065] (4);
[0066] in, For blocks The output features of the window-based multi-head attention mechanism layer; For blocks Output features of the SW-MSA layer; , , Respectively for blocks ,piece ,piece Output features of residual-based multilayer perceptron; Represents a window-based multi-head attention mechanism layer; Represents a multi-head attention mechanism layer based on an offset window; represents a residual-based multilayer perceptron; Representation layer normalization;
[0067] In the Swin Transformer block, the W-MSA layer divides the input image into several non-overlapping local windows, and then calculates the local attention of each local window to obtain the local window self-attention matrix, which is specifically composed as follows:
[0068] (5);
[0069] in, represents the self-attention matrix; , , Represent the query matrix, key matrix, and value matrix respectively; Represents the Softmax function; Represents the mapping matrix, that is and The vector dimension of is the bias matrix.
[0070] Then it is input to the SW-MSA layer through a residual-based multilayer perceptron ResMLP with a GELU activation function. The function of the SW-MSA layer is to introduce cross-window connections through sliding windows, so that the information between different local windows can interact and improve the ability to extract global features. The residual-based multilayer perceptron ResMLP of the present invention replaces the MLP of the traditional SwinTransformer, improves model accuracy, reduces memory usage, and simplifies training complexity. Step 5, train the deep learning model based on the improved Swin Transformer; use the multi-channel composite data line chart generated in step 3 as the model input, and the clean coal ash content as the output, and train the deep learning model based on the improved Swin Transformer based on the training set to improve its prediction ability for clean coal ash content. The model uses mean square error (MSE) as the empirical loss to measure the difference between the predicted value and the actual value. The calculation formula is:
[0071] (6);
[0072] in, is the mean square error value; is the total number of samples; Representative The actual clean coal ash value of each sample; It is Predicted clean coal ash value for samples.
[0073] In addition, the model's hyperparameters were carefully adjusted to further optimize performance: the learning rate was 0.001; the optimizer was the Adam algorithm; the batch size was 32; and the number of training rounds was 50. At the same time, an early stopping mechanism was introduced to monitor the training process. Once the model performance no longer improved, the training was stopped to prevent the model from overfitting.
[0074] Step 6: Model evaluation and application: After completing the model training, the model is evaluated using the test set. By comparing the predicted results with the actual clean coal ash content, key indicators such as mean square error (MSE), relative absolute error (RSE) and relative absolute error (RAE) are calculated to measure the performance of the model. Finally, the model with the best training effect is applied to the actual coal preparation production to achieve accurate prediction of clean coal ash content and further improve the efficiency and stability of the coal preparation process.
Claims
1. A method for predicting ash content of clean coal by heavy medium separation based on multi-channel line graphing of time series data, characterized in that: The time series data is converted into a multi-channel line graph to achieve accurate prediction of clean coal ash content, including the following steps: Step 1: Collect historical big data of the coal preparation plant and build a coal preparation plant database; Step 2: Perform data preprocessing and divide the preprocessed data into training set and test set in proportion; Step 3, generate a multi-channel composite data line chart; In step 3, multiple time series-based indicator line graphs are generated using the preprocessed data. When generating, the key process parameters are set as the x-axis, the raw coal ash content is set as the y-axis, and the data points in the data set are represented as turning points of the line graph. The turning points are connected by straight lines to form a line graph showing the change of raw coal ash content over time; Subsequently, these line graphs are superimposed in the same graph in a multi-channel form according to the time series to construct a multi-channel composite data line graph; Step 4: Design a deep learning model based on the improved Swin Transformer; In the step 4, the deep learning model based on the improved Swin Transformer captures local and global information by constructing a hierarchical representation, which starts from small-sized blocks in the early layers and gradually merges adjacent blocks in deeper layers; the input of the model is the multi-channel composite data line chart generated in step 3, and the output of the model is the ash content of clean coal; the deep learning model based on the improved Swin Transformer includes four stages, and the first stage, the second stage, and the fourth stage all include a group of Swin Transformer blocks; the third stage includes three groups of Swin Transformer blocks connected in series; a group of Swin Transformer blocks includes a window-based multi-head attention mechanism layer and an offset window-based multi-head attention mechanism layer connected in series in sequence; A residual-based multi-layer perceptron with a GELU activation function is connected between the window-based multi-head attention mechanism layer and the offset window-based multi-head attention mechanism layer; The specific working process of the deep learning model based on the improved Swin Transformer is: Step 4.1, input the multi-channel composite data line graph as an input image into the model, and divide the input image into blocks; Step 4.2: Perform the first-stage linear embedding and Swin Transformer block processing to obtain the output features of the first stage; Step 4.3, block merging and Swin Transformer block processing are performed on the output features of the first stage to obtain the output features of the second stage; Step 4.4, continue to perform block merging and Swin Transformer block processing on the output features of the second stage to obtain the output features of the third stage; Step 4.5, continue to perform block merging and Swin Transformer block processing on the output features of the third stage to obtain the output features of the fourth stage, which is the final output of the model; The calculation process of the Swin Transformer block is as follows: in, Represents the output features of the window-based multi-head attention mechanism layer for block l; represents the output features of the offset window multi-head attention mechanism layer for block l+1; Z l 、z l-1 、z l+1 Represent the output features of residual multilayer perceptron for block l, block l-1, and block l+1 respectively; W-MSA(·) represents the window-based multi-head attention mechanism layer; SW-MSA(·) represents the offset window-based multi-head attention mechanism layer; ResMLP(·) represents the residual-based multilayer perceptron; LN(·) represents layer normalization; In the Swin Transformer block, the input image is divided into several non-overlapping local windows based on the window multi-head attention mechanism layer, and then the local attention of each local window is calculated to obtain the local window self-attention matrix, which is specifically composed as follows: Where Attention represents the self-attention matrix; Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; SoftMax(·) represents the Softmax function; d represents the mapping matrix; and B represents the bias matrix. Then it is input into the offset window based multi-head attention mechanism layer through the residual based multi-layer perceptron; Step 5: Train the deep learning model based on the improved Swin Transformer; Step 6: Model evaluation and application.
2. The method for predicting ash content of clean coal by dense medium separation based on multi-channel line graphing of time series data according to claim 1 is characterized in that: In the step 1, the clean coal ash content, raw coal ash content and key process parameters in the heavy medium separation process are collected through the production monitoring system and manual records of the coal preparation plant; the key process parameters include sampling date, sampling time, magnetic content, coal slime content, separation density, cyclone inlet pressure, and combined medium barrel liquid level; combined with big data analysis technology, the data is deeply mined and analyzed.
3. The method for predicting ash content of clean coal by dense medium separation based on multi-channel line graphing of time series data according to claim 1 is characterized in that: In step 2, data preprocessing includes correcting sensor errors, correcting recording errors, and formatting the data into a structure used by subsequent models.
4. The method for predicting ash content of clean coal by dense medium separation based on multi-channel line graphing of time series data according to claim 1 is characterized in that: In step 5, the deep learning model based on the improved Swin Transformer is trained based on the training set. The model uses the mean square error as the empirical loss to measure the difference between the predicted value and the actual value. The calculation formula is: Among them, MSE is the mean square error; n is the total number of samples; y i represents the true clean coal ash value of the i-th sample; is the predicted clean coal ash value of the i-th sample.
5. The method for predicting ash content of clean coal by dense medium separation based on multi-channel line graphing of time series data according to claim 1, characterized in that: In step 6, after completing the model training, the model is evaluated using a test set; finally, the model with the best performance is applied to actual coal preparation production to achieve accurate prediction of clean coal ash content.