Method for online prediction of sinter chemical quality
By acquiring sintering process parameters and tail section image data, extracting keyframe features using the Unet network, and combining them with a quality prediction model, the problem of insufficient accuracy in online prediction of sinter chemical quality was solved, achieving higher accuracy prediction and production stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2022-12-28
- Publication Date
- 2026-06-26
AI Technical Summary
In existing technologies, the accuracy of online prediction of sinter chemical quality is poor, resulting in frequent fluctuations in production indicators, large amounts of returned ore, and high energy consumption.
By determining multiple parameters related to chemical quality during the sintering process, acquiring time-series features and sintering machine tail image data, using the Unet segmentation network to extract the depth features of key frames, and combining them with a quality prediction model for prediction.
This greatly improves the accuracy of online prediction of sinter chemical quality, more accurately reflects the current chemical quality, and stabilizes production and optimizes control.
Smart Images
Figure CN115859829B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of quality prediction technology, and in particular relates to an online prediction method for the chemical quality of sintered ore. Background Technology
[0002] Sintering is a key process in steel production. Its main function is to produce sinter from various iron ore powders, fluxes, and fuels through processes such as mixing and batching, adding water for granulation, charging and ignition, sintering with ventilation, and screening and sizing. This process continuously provides high-quality iron-containing raw materials for the blast furnace. As the main raw material for the blast furnace, stabilizing the chemical quality of sinter plays a crucial role in improving the furnace's output and quality. However, the sintering process is a complex industrial process characterized by nonlinearity, strong coupling, and large time lag. Due to limitations in testing instruments, current analysis of the chemical composition of sinter is mainly conducted offline. This time lag leads to frequent fluctuations in various production indicators, large amounts of returned ore, and high energy consumption.
[0003] To address the above issues, scholars both domestically and internationally have conducted research on methods for predicting the quality of the sintering process. Among these, mechanistic models reflect the main laws of sintering production, but they contain numerous assumptions and simplifications, resulting in low model accuracy. Existing data-driven models are mainly based on sintering process production operation data, indirectly predicting sintering quality by mapping the relationship between operational variables and quality indicators. However, due to external interference and the influence of some unmeasurable variables, the model prediction accuracy is still insufficient.
[0004] With the development of machine vision technology, many steel companies have installed industrial cameras and other inspection equipment to collect images of the sintering machine tail end. Since the sintering machine tail image is a direct representation of the internal characteristics of the sintering bed at the end of the sintering process, it can more directly and realistically reflect the current sintering conditions compared to process data that indirectly reflects sintering quality. Therefore, some scholars have tried to use sintering machine tail images for sinter quality prediction. However, due to insufficient utilization of image feature information, the accuracy of the established sinter chemical quality prediction model still cannot meet the requirements of on-site production.
[0005] Therefore, how to combine the actual production situation of the sintering process, fully integrate the time sequence characteristics of the sintering process data and the image data of the sintering machine tail, and quickly and accurately predict the chemical quality of the sintering process online, so as to stabilize the production and optimize the control of the sintering process, improve the quality and yield of sintered ore and reduce energy consumption, is an urgent problem to be solved in the sintering field.
[0006] In practical applications, the aforementioned limitations result in poor prediction accuracy for online prediction of sinter chemical quality. Summary of the Invention
[0007] This application provides an online prediction method for the chemical quality of sintered ore, which aims to solve the problem of poor accuracy in online prediction of the chemical quality of sintered ore.
[0008] This application provides an online method for predicting the chemical quality of sintered ore, including:
[0009] Determine multiple sintering process parameters related to the chemical quality of sintered ore during sintering;
[0010] Obtain the temporal characteristics of the multiple sintering process parameters at the current prediction time;
[0011] The sintering machine tail image data from the previous prediction time to the current prediction time is acquired, and the key frame sequence of the sintering machine tail image data is determined by using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image.
[0012] The Unet segmentation network is used to extract the primary features of each key frame in the key frame sequence, and the secondary features of the sintering machine tail key frame at the current prediction time are obtained based on the primary features of each key frame.
[0013] The time-series features, the secondary features of the keyframe at the tail of the sintering machine, and the chemical quality value of the sinter at the previous prediction time are input into the quality prediction model to perform quality prediction, thereby obtaining the predicted value of the chemical quality of the sinter at the current prediction time.
[0014] Optionally, the period from the previous prediction time to the current prediction time includes multiple sintering cycles, and the sintering machine tail image data includes sintering machine tail image data corresponding to each sintering cycle.
[0015] The step of using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image to determine the keyframe sequence of the sintering machine tail image data includes:
[0016] For each sintering cycle, the sintering machine tail image data corresponding to the sintering cycle is input into the Unet segmentation network for processing to obtain the combustion zone shape corresponding to each sintering machine tail image in the sintering cycle, and the sintering machine tail image corresponding to the combustion zone shape with the largest area is taken as the key frame of the sintering cycle.
[0017] The set of keyframes from all sintering cycles is taken as the keyframe sequence of the sintering machine tail image data.
[0018] Optionally, the step of extracting the primary features of each keyframe in the keyframe sequence using the Unet segmentation network, and obtaining the secondary features of the sintering machine tail keyframe at the current prediction time based on the primary features of each keyframe, includes:
[0019] For each keyframe in the keyframe sequence, the keyframe is input into the encoder of the Unet segmentation network for processing to obtain the first-level features of the keyframe.
[0020] For each keyframe, the first-level features are input into the Flatten layer for processing, and the data output from the Flatten layer is input into the first Dense layer for processing to obtain the feature vector of the keyframe.
[0021] The average value of the feature vectors of all key frames is calculated and input into the second Dense layer for processing to obtain the temporal features of the sintering machine tail image frame at the current prediction time.
[0022] The temporal features of the sintering machine tail image frame at the current prediction time are expanded to obtain the secondary features of the key frame of the sintering machine tail at the current prediction time.
[0023] Optionally, the formula for calculating the average of all feature vectors is:
[0024]
[0025] in, f is the average of all eigenvectors k Let K be the feature vector of the key frame of the k-th sintering cycle, where k = 1, ..., K T K T The number of the plurality of sintering cycles.
[0026] Optionally, the step of inputting the time-series features, the secondary features of the sintering machine tail keyframe, and the chemical quality value of the sinter at the previous prediction time into the quality prediction model to perform quality prediction and obtain the predicted value of the chemical quality of the sinter at the current prediction time includes:
[0027] The time-series characteristics of the multiple sintering process parameters at the current prediction time, the secondary characteristics of the sintering machine tail keyframe at the current prediction time, and the chemical quality value of the sintered ore at the previous prediction time are input into the quality prediction model to perform quality prediction, and a two-dimensional matrix of the predicted values of the chemical quality of the sintered ore at the current prediction time is obtained.
[0028] Calculate the average value of the two-dimensional matrix and use the average value as the predicted value of the chemical quality of the sinter at the current prediction time.
[0029] Optionally, the two-dimensional matrix of the predicted values of the chemical quality of the sinter at the current prediction time. The expression is:
[0030]
[0031] The calculation of the average value of the two-dimensional matrix includes:
[0032] Through formula Calculate the average value of the two-dimensional matrix;
[0033] in, For the The value of Q in the Mth row and Mth column. T The average value of the two-dimensional matrix. For the The value in the i-th row and j-th column is N, where N is the number of data points contained in any sintering process parameter from the previous prediction time to the current prediction time.
[0034] Optionally, before determining the keyframe sequence of the sintering machine tail image data using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image, the prediction method further includes:
[0035] Eliminate noise in the sintering machine tail image data;
[0036] Color correction is performed on the sintering machine tail image data after noise elimination;
[0037] Dehazing was performed on the color-corrected image data of the sintering machine tail section.
[0038] The above-mentioned solution in this application has the following beneficial effects:
[0039] In the embodiments of this application, multiple sintering process parameters related to the chemical quality of sintered ore are determined during the sintering process. The temporal characteristics of these multiple sintering process parameters at the current prediction time are obtained. Sintering machine tail image data from the previous prediction time to the current prediction time is obtained. The Unet segmentation network used to extract the combustion zone morphology of the sintering machine tail image is used to determine the key frame sequence of the sintering machine tail image data. The Unet segmentation network is used to extract the first-level features of each key frame in the key frame sequence. The second-level features of the sintering machine tail key frame at the current prediction time are obtained based on the first-level features of each key frame. The temporal characteristics, the second-level features of the sintering machine tail key frame, and the chemical quality value of sintered ore at the previous prediction time are input into the quality prediction model for quality prediction to obtain the predicted value of the chemical quality of sintered ore at the current prediction time. This application utilizes the image information from the tail of the sintering machine, extracts the depth features of the key frames of the tail of the sintering machine based on the Unet neural network, and effectively integrates the temporal features of the sintering process parameters and the depth features of the key frames of the tail of the sintering machine, thereby more fully and realistically reflecting the chemical quality of the sinter at the current moment and greatly improving the prediction accuracy of online prediction of the chemical quality of sinter.
[0040] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 A flowchart illustrating an embodiment of the online prediction method for the chemical quality of sintered ore provided in this application;
[0043] Figure 2 A flowchart for obtaining the timing characteristics of multiple sintering process parameters at the current prediction time, provided as an embodiment of this application;
[0044] Figure 3 A flowchart for data preprocessing of sintering machine tail image data from the previous prediction time to the current prediction time, provided as an embodiment of this application;
[0045] Figure 4 A flowchart for determining the keyframe sequence of sintering machine tail image data provided in one embodiment of this application;
[0046] Figure 5 This is a schematic diagram of the structure for extracting secondary features of keyframes at the tail of a sintering machine according to an embodiment of this application;
[0047] Figure 6 This is a flowchart illustrating an embodiment of the present application for obtaining a predicted value of the chemical quality of sinter at the current prediction time. Detailed Implementation
[0048] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0049] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.
[0050] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0051] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."
[0052] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0053] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.
[0054] To address the issue of insufficient accuracy in current sinter chemical quality prediction models, this application proposes a method that determines multiple sintering process parameters related to the chemical quality of sintering ore during the sintering process, obtains the temporal characteristics of these parameters at the current prediction time, acquires sintering mill tail image data from the previous prediction time to the current prediction time, and uses the Unet segmentation network for extracting the combustion zone morphology of the sintering mill tail image to determine the keyframe sequence of the sintering mill tail image data. The Unet segmentation network is then used to extract the primary features of each keyframe in the keyframe sequence, and the secondary features of the sintering mill tail keyframe at the current prediction time are obtained based on the primary features of each keyframe. The temporal characteristics, the secondary features of the sintering mill tail keyframe, and the sinter chemical quality value from the previous prediction time are input into the quality prediction model for quality prediction, resulting in a predicted value for the sinter chemical quality at the current prediction time. This application utilizes the image information from the tail of the sintering machine, extracts the depth features of the key frames of the tail of the sintering machine based on the Unet neural network, and effectively integrates the temporal features of the sintering process parameters and the depth features of the key frames of the tail of the sintering machine, thereby more fully and realistically reflecting the chemical quality of the sinter at the current moment and greatly improving the prediction accuracy of online prediction of the chemical quality of sinter.
[0055] The online prediction method for chemical quality of sinter provided in this application will be described below with reference to specific embodiments.
[0056] This application provides an online prediction method for the chemical quality of sintered ore. This method can be executed by a terminal device or by a device (such as a chip) applied within the terminal device. The following embodiments use the execution of this method by a terminal device as an example. As an example, the terminal device can be a tablet, server, or laptop, etc., and this application does not limit this to any particular type.
[0057] like Figure 1 As shown in the embodiments of this application, the online prediction method for the chemical quality of sintered ore includes the following steps:
[0058] Step 101: Determine multiple sintering process parameters related to the chemical quality of the sintered ore during the sintering process.
[0059] Specifically, the composition of sintering raw materials is collected as chemical quality data through the Data Collection System (DCS), and process data such as wind box negative pressure and exhaust gas temperature are used as sintering process parameters of sintered ore. Through sintering mechanism and correlation analysis, multiple sintering process parameters related to the quality of sintered ore are determined.
[0060] In some embodiments of this application, determining multiple sintering process parameters related to the quality of sintered ore during sintering is to facilitate the subsequent obtaining of the temporal characteristics of the sintering process parameters.
[0061] Step 102: Obtain the temporal characteristics of multiple sintering process parameters at the current prediction time.
[0062] Specifically, based on time registration and data cleaning methods, multiple sintering process parameters related to sinter quality can be preprocessed to obtain one-dimensional time series data of the sintering process parameters. Then, based on Gramian Angular Field (GAF), the one-dimensional time series data of the sintering process parameters at the current prediction time can be converted into a two-dimensional matrix to extract the time series features of the sintering process parameters.
[0063] In some embodiments of this application, the temporal characteristics of the sintering process parameters are obtained for use in the online prediction model of sinter chemical quality, thereby improving the prediction accuracy of the online prediction model of sinter chemical quality.
[0064] Step 103: Obtain the sintering machine tail image data from the previous prediction time to the current prediction time, and use the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image to determine the key frame sequence of the sintering machine tail image data.
[0065] Specifically, the sintering machine tail image data is acquired and preprocessed to obtain preprocessed sintering machine tail image data. Then, the sintering machine tail image data from the previous prediction time to the current prediction time is determined, and the key frame sequence of the sintering machine tail image data is extracted based on the Unet segmentation network (Unet segmentation network is a commonly used segmentation network) used to extract the combustion zone morphology of the sintering machine tail image.
[0066] In some embodiments of this application, obtaining sintering machine tail image data from the previous prediction time to the current prediction time, and determining the key frame sequence of the sintering machine tail image data based on the Unet segmentation network used to extract the combustion zone morphology of the sintering machine tail image, is to facilitate the subsequent acquisition of secondary features of the sintering machine tail key frames.
[0067] It should be noted that the interval between the previous prediction time and the current prediction time is the time between the previous prediction time and the current prediction time, and this interval can be one hour.
[0068] Step 104: Use the Unet segmentation network to extract the first-level features of each keyframe in the keyframe sequence, and obtain the second-level features of the sintering machine tail keyframe at the current prediction time based on the first-level features of each keyframe.
[0069] Specifically, the key frame sequence of the sintering machine tail image data can be processed based on the Unet segmentation network used to extract the combustion zone morphology of the sintering machine tail image to obtain the first-level features of the sintering machine tail key frame sequence. Then, the second-level features of the sintering machine tail key frame at the current moment can be extracted through methods such as averaging and dimensional transformation.
[0070] In some embodiments of this application, the secondary features of the keyframe at the tail of the sintering machine are obtained for use in the online prediction model of the chemical quality of sintered ore, thereby improving the prediction accuracy of the online prediction model of the chemical quality of sintered ore in this application.
[0071] Step 105: Input the time series features, the secondary features of the key frame at the tail of the sintering machine, and the chemical quality value of the sinter at the previous prediction time into the quality prediction model to make quality prediction and obtain the predicted value of the chemical quality of the sinter at the current prediction time.
[0072] Specifically, the system integrates the temporal characteristics of the sintering process parameters at the current prediction time, the secondary characteristics of the key frame of the sintering machine tail at the current prediction time, and the chemical quality value of the sintered ore at the previous time. Based on the online prediction model of the chemical quality of the sintered ore, it outputs the chemical quality value of the sintered ore at the current prediction time.
[0073] It is worth mentioning that, by utilizing the image information of the sintering machine tail and extracting the depth features of the key frames of the sintering machine tail based on the Unet neural network, this application effectively integrates the temporal features of the sintering process parameters and the depth features of the key frames of the sintering machine tail, thereby more fully and realistically reflecting the chemical quality of the sinter at the current moment and greatly improving the prediction accuracy of the online prediction of the chemical quality of the sinter.
[0074] The following describes step 102 by way of example with reference to specific embodiments.
[0075] In some embodiments of this application, such as Figure 2 As shown, the specific implementation method for obtaining the temporal characteristics of multiple sintering process parameters at the current prediction time includes the following steps:
[0076] Step 201: Based on the different spatial distributions of multiple sintering process parameters, construct a time-series registration model, and use the time-series registration model to perform time-series registration on multiple sintering process parameters.
[0077] Specifically, the temporal registration model constructed above is as follows:
[0078]
[0079] Among them, Y T =TFe,FeO,R] T T represents the current predicted time of the sinter, TFe represents the total iron content in the sinter at the current predicted time, FeO represents the ferrous iron content in the sinter at the current predicted time, R represents the basicity of the sinter, A1 represents the exhaust gas temperature of channel A1, A1 represents the wind box negative pressure of channel A1, A represents the total pipe temperature of channel A, U1,…,U i ,…,U MM represents the time lag between the sintered ore product and the sintering process parameters, M represents the number of sintering process operation parameters, and N represents the number of data points for a certain sintering process parameter within a certain time period.
[0080] Step 202: Perform data cleaning on the multiple sintering process parameters after time registration to obtain the time series data of the multiple sintering process parameters.
[0081] Specifically, based on the box plot, data cleaning can be performed on multiple sintering process parameters after time registration to obtain time-series data of multiple sintering process parameters.
[0082] In some embodiments of this application, data cleaning is performed on multiple sintering process parameters after time registration to obtain time-series data of multiple sintering process parameters in order to remove outliers in the multiple sintering process parameters after time registration.
[0083] Step 203: Determine the time series data of multiple sintering process parameters at the current prediction time, and obtain the time series characteristics of multiple sintering process parameters at the current prediction time based on the time series data of multiple sintering process parameters at the current prediction time.
[0084] Specifically, let X be the denoted X. T This represents a time series of data for a sintering process parameter x that matches the chemical quality value of the sinter at time T.
[0085] X T ={x z z = TU Z ,…,TU z +N-1}
[0086] Among them, X T Let x be a time series of data that matches the chemical quality value of the sintered ore at time T, representing a certain sintering process parameter x. z For X T The z-th data point, where T is the current prediction time for the sinter, and U... z Let N be the time lag between the sintered ore product and a certain sintering process parameter, and let N be the number of data points contained in any of the above sintering process parameters from the previous prediction time to the current prediction time. It should be noted that the number of data points contained in each of the above sintering process parameters is the same from the previous prediction time to the current prediction time.
[0087] For a time series of data for the sintering process parameter x mentioned above, the following formula is used for standardization:
[0088]
[0089] in, For X T The z-th data after standardization, ave(X)T ) represents X T The mean, std(X) T ) represents X T The variance.
[0090] The standardized sintering process parameters described above are normalized using the following formula:
[0091]
[0092] in, for The z-th data after normalization, This is the standardized timing data for this type of sintering process. for The minimum value, for The maximum value.
[0093] The above The following formula can be used to convert it to polar coordinates:
[0094]
[0095] in, and r z They are respectively The corresponding angular and radial coordinates in polar coordinates.
[0096] The sum of the angular coordinates of two data points is used to represent the temporal relationship between the two data points, i.e., the GASF matrix, which is expressed as follows:
[0097]
[0098] Wherein, GASF is the time-series feature matrix of a certain sintering process parameter at the current prediction time. for The sum of the angular coordinates of the polar coordinates corresponding to the i-th and j-th data points. For the The corresponding cosine values, j∈[1,N], i∈[1,N], This is the normalized timing data for this type of sintering process.
[0099] The following describes step 103 by way of specific embodiments.
[0100] In some embodiments of this application, such as Figure 3 As shown, the specific implementation method before determining the keyframe sequence of the sintering machine tail image data using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image includes the following steps:
[0101] Step 301: Eliminate noise in the sintering machine tail image data.
[0102] Specifically, all sintering machine tail images are acquired using industrial cameras, and a nonlocal mean algorithm is introduced to eliminate noise in all sintering machine tail images.
[0103] The above-mentioned nonlocal mean algorithm for eliminating noise in the sintering machine tail image includes the following steps:
[0104] Step 1, let the input image for the nonlocal means algorithm be:
[0105] I c (x,y)
[0106] Among them, I c (x,y) is the input image of the nonlocal means algorithm, I c (x,y)∈[0,255], c∈{r,g,b}, where {r,g,b} are the R, G, and B color channels of the input image, and (x,y) are the coordinates of the pixels in the input image.
[0107] Step two, within the input image from step one, select the current pixel I. c The neighborhood centered at (u,v) and of size (2r+1)×(2r+1) is denoted as...
[0108] Step 3, in the neighborhood of step 2 In the middle, select pixel I c (m,n)∈ And with this pixel as the center, construct a neighborhood of size (2r+1)×(2r+1), denoted as
[0109] Step 4: Calculate the neighborhood and Similarity between them:
[0110]
[0111] in, As a whole, it represents two neighboring regions. and Similarity between them For two neighboring regions and The distance between them σ is the noise standard deviation, and h is the filtering parameter.
[0112] Step 5, using the neighborhood All pixels within and the current pixel The similarity is used as the weight to calculate the neighborhood. The estimated value of the current pixel can be obtained by taking the weighted average of all pixel values within the range. The mathematical expression is:
[0113]
[0114] in, For the current pixel I c The estimated value of (u,v).
[0115]
[0116] Step Six, for I c For each pixel in (x, y), repeat steps two through five to obtain the output image:
[0117] O c (x,y)
[0118] Among them, O c (x,y) is the output image of the nonlocal mean algorithm.
[0119] Step 302: Perform color correction on the sintering machine tail image data after noise elimination.
[0120] Specifically, adaptive color correction is performed on all sintering machine tail images obtained in step 301 based on a local color correction algorithm.
[0121] The above-mentioned adaptive color correction of the sintering machine tail image obtained in step 301 based on the local color correction algorithm includes the following steps:
[0122] Step 1, let the input image for the local color correction algorithm be:
[0123] O c (x,y)
[0124] Among them, O c (x,y) is the output image of the above nonlocal mean algorithm, O c (x,y)∈[0,255], c∈{r,g,b}, where {r,g,b} are the R, G, and B color channels of the input image, and (x,y) is O(0,255). c (x,y) are the coordinates of the pixels.
[0125] Step two, the mask image of the input image in step one can be represented as:
[0126]
[0127] Where M(x,y) is O c Mask image of (x,y), For a size of s g ×s gA Gaussian kernel with standard deviation σ.
[0128] Step 3: Process the mask image obtained in Step 2 using the following formula to obtain the output image:
[0129]
[0130] Among them, H c (x,y) is the output image of the local color correction algorithm.
[0131] Step 303: Dehaze the color-corrected sintering machine tail image data.
[0132] Specifically, based on the fast dehazing algorithm, all sintering machine tail images obtained in step 302 are dehazed to obtain all dehazed sintering machine tail images, and all the above dehazed sintering machine tail images are used as sintering machine tail image data.
[0133] The above-mentioned dehazing of the sintering machine tail image obtained in step 302 based on the fast dehazing algorithm includes the following steps:
[0134] Step 1: Calculate the global atmospheric light using the following formula:
[0135]
[0136] Where A represents global atmospheric light, and H represents... c (x,y) is the output image of the local color correction algorithm described above, which is represented as a foggy image, and ε is usually taken as 0.5.
[0137]
[0138]
[0139] For a size of s a ×s a The mean-filtered convolution kernel.
[0140] Step 2, the expression for ambient light is A(1-t(x,y)).
[0141] Where t(x,y) is the cutoff transmittance, calculated using the following formula:
[0142]
[0143] ρ is an adjustable parameter, h ave =Ave(H(x,y)).
[0144] Step 3: Obtain the dehazed image using a fast dehazing algorithm. The mathematical expression for the fast dehazing algorithm is:
[0145]
[0146] Among them, F c (x,y are the dehazed images, c∈{r,g,b}, {r,g,b} are the R, G, and B color channels of the dehazed image.)
[0147] In some embodiments of this application, the specific implementation prior to utilizing the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image further includes the following steps:
[0148] Step 1: Based on the sintering machine tail image data, select a portion of the sintering machine tail image data for processing to obtain a dataset of sintering machine tail combustion zone morphology.
[0149] Specifically, a sintering machine tail image with a size of 1024×128×3 was selected and preprocessed to create a dataset of sintering machine tail combustion zone morphology.
[0150] Step 2: Perform data augmentation on the combustion zone morphology dataset at the tail of the sintering machine, and then divide the augmented dataset into a training set and a test set.
[0151] Specifically, the dataset was augmented by translation, rotation, and brightness changes, and then divided into training and testing sets.
[0152] Step 3: Train the Unet neural network using the training set to obtain the Unet segmentation network used to extract the combustion zone morphology of the sintering machine tail image.
[0153] Specifically, the training set from step two is input into the Unet neural network for training. During training, the cross-entropy function is used as the loss function, the Adaptive Moment Estimation (Adam) algorithm is used as the optimizer, and the learning rate is set to 3×10. -4 Once the Unet network is trained, a Unet segmentation network is obtained for extracting the combustion zone morphology from the sintering machine tail image.
[0154] For example, the sintered cross-sectional dimensions of the model input are 1024×128×3, and the overall structure includes four encoder layers and four corresponding decoder layers.
[0155] Each encoder layer contains two 3×3 convolutional layers (Conv), one batch normalization layer, and one max pooling layer. Each decoder layer contains one upsampling layer, one 3×3 convolutional layer, one concatenate layer, two 3×3 convolutional layers, and one batch normalization layer. Finally, after passing through two 3×3 convolutional layers and being activated by the sigmoid activation function, the output is a burning band shape with a size of 1024×128×1.
[0156] In some embodiments of this application, the period from the previous prediction time to the current prediction time includes multiple sintering cycles, and the sintering machine tail image data includes the sintering machine tail image data corresponding to each sintering cycle.
[0157] Specifically, because the sintering machine tail image data from the previous prediction time to the current prediction time includes K T If there are multiple sintering cycles and the video within each sintering cycle exhibits periodic changes, then the sintering machine tail image data from the previous prediction time to the current prediction time will be used according to K. T Each sintering cycle is divided into K T Image data of the tail end of the sintering machine during each sintering cycle.
[0158] like Figure 4 As shown, the specific implementation method for determining the keyframe sequence of sintering machine tail image data using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image includes the following steps:
[0159] Step 401: For each sintering cycle, the sintering machine tail image data corresponding to the sintering cycle is input into the Unet segmentation network for processing to obtain the combustion zone shape corresponding to each sintering machine tail image in the sintering cycle, and the sintering machine tail image corresponding to the combustion zone shape with the largest area is taken as the key frame of the sintering cycle.
[0160] Specifically, for each sintering cycle, all sintering tail images within that cycle are input into the Unet segmentation network used to extract the combustion zone morphology of the sintering tail images. This yields the combustion zone morphology corresponding to each tail image within that cycle, and the tail image corresponding to the maximum area of the combustion zone morphology is selected as the keyframe for that cycle.
[0161] Step 402: The set of keyframes from all sintering cycles is used as the keyframe sequence of the sintering machine tail image data.
[0162] Specifically, following the method above, a total of K is obtained. T Zhang keyframes, and K T Zhang keyframes are the keyframe sequences corresponding to the chemical quality values of sinter at the current prediction time.
[0163] The following describes step 104 by way of example with reference to specific embodiments.
[0164] In some embodiments of this application, the specific implementation of extracting the first-level features of each keyframe in the keyframe sequence using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image, and obtaining the second-level features of the sintering machine tail keyframe at the current prediction time based on the first-level features of each keyframe includes the following steps:
[0165] Step 1: For each keyframe in the keyframe sequence, input the keyframe into the encoder of the Unet segmentation network for processing to obtain the first-level features of the keyframe.
[0166] Specifically, for each keyframe in the keyframe sequence of the sintering machine tail image data, its feature vector f, after being encoded four times by the Unet network, is... g As a first-level feature of the keyframe at the tail of the sintering machine.
[0167] Step 2: For each keyframe, the first-level features are input into the Flatten layer for processing, and the output data of the Flatten layer is input into the first Dense layer for processing to obtain the feature vector of the keyframe.
[0168] Specifically, for each keyframe first-level feature of the sintering machine tail, the feature vector f is obtained after passing through a flatten layer and a first dense layer. k .
[0169] Step 3: Calculate the average value of the feature vectors of all keyframes and input the average value into the second Dense layer for processing to obtain the temporal features of the sintering machine tail image frame at the current prediction time.
[0170] In some embodiments of this application, the formula for calculating the average value of the feature vectors of all keyframes is as follows:
[0171]
[0172] in, f is the average of all eigenvectors k Let K be the feature vector of the key frame of the k-th sintering cycle, where k = 1, ..., K T K T This refers to the number of sintering cycles mentioned earlier, that is, the number of sintering cycles from the previous prediction time to the current prediction time.
[0173] Step 4: Expand the temporal features of the sintering machine tail image frame at the current prediction time to obtain the secondary features of the sintering machine tail key frame at the current prediction time.
[0174] Specifically, the temporal characteristics of the sintering machine tail image at the current prediction time will be used. Expanding, we obtain the secondary features of the sintering machine tail keyframe at the current prediction time, denoted as the N×N feature matrix F. T .
[0175] An exemplary structural diagram for extracting secondary features from the keyframe of the sintering machine tail is shown below. Figure 5 As shown, the first-level features of the sintering machine tail keyframe at the current prediction time are processed through the Flatten layer and the first Dense layer to obtain K. T eigenvectors f k Calculate K T eigenvectors f k average and average After passing through a second Dense layer with an input dimension of N×N, it is expanded (from... Figure 5 After processing by the Reshape function in the middle, the N×N secondary feature matrix F of the sintering machine tail keyframe at the current prediction time is obtained. T .
[0176] The following describes step 105 by way of example with reference to specific embodiments.
[0177] In some embodiments of this application, such as Figure 6 As shown, the specific implementation method of inputting the time series features, the secondary features of the keyframe of the sintering machine tail, and the chemical quality value of the sinter at the previous prediction time into the quality prediction model to obtain the predicted value of the chemical quality of the sinter at the current prediction time includes the following steps:
[0178] Step 1: Input the temporal characteristics of multiple sintering process parameters at the current prediction time, the secondary characteristics of the sintering machine tail keyframe at the current prediction time, and the chemical quality value of sintered ore at the previous prediction time into the quality prediction model to perform quality prediction, and obtain a two-dimensional matrix of the predicted values of the chemical quality of sintered ore at the current prediction time.
[0179] Specifically, the secondary features F of the sintering machine tail keyframe at the current prediction time, with dimensions N×N×1, are... T The timing feature matrix S of sintering process parameters at the current prediction time, with dimensions N×N×M. T The two-dimensional chemical quality matrix Q of the sinter at the previous prediction time is of dimension N×N×1. T-1 , as input to the quality prediction model.
[0180] Among them, the sintering process parameter timing feature matrix S with dimensions of N×N×M at the current prediction time is... TIt is represented as the temporal characteristics of M sintering process parameters (i.e., the multiple sintering process parameters mentioned in step one above) at the current prediction time.
[0181] In some embodiments of this application, a two-dimensional matrix of predicted values for the chemical quality of sinter at the current prediction time is provided. The expression is:
[0182]
[0183] in, This is a two-dimensional matrix representing the predicted chemical quality values of sinter at the current prediction time. For the The value in the Nth row and Nth column.
[0184] In some embodiments of this application, the quality prediction model can be a commonly used quality prediction model structure, which may specifically include: a first convolutional layer (Conv), a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, and an eighth convolutional layer connected in sequence.
[0185] In practical applications, to improve prediction accuracy, the parameters of the convolutional layers can be set as follows: the first, second, third, fifth, sixth, and seventh convolutional layers all use the tanh activation function and a 2×2×32 convolutional kernel, the fourth convolutional layer uses the tanh activation function and a 2×2×1 convolutional kernel, and the eighth convolutional layer uses only a 2×2×1 convolutional kernel.
[0186] It should be noted that when training the quality prediction model, the root mean square error can be used as the loss function, the learning rate can be dynamically adjusted using the exponential decay method, and the root mean square propagation gradient descent method (RMSprop) can be used as the optimizer for training the quality prediction model.
[0187] Step 2: Calculate the average value of the two-dimensional matrix and use the average value as the predicted value of the chemical quality of the sinter at the current prediction time.
[0188] In some embodiments of this application, the formula for calculating the average value of a two-dimensional matrix is as follows:
[0189]
[0190] Among them, Q T This represents the average value of a two-dimensional matrix, i.e., the predicted chemical quality of the sinter at the current prediction time. for The value in the Nth row and Nth column, for The value in the i-th row and j-th column is N, where N is the number of data points contained in any sintering process parameter from the previous prediction time to the current prediction time.
[0191] The following example illustrates the prediction accuracy of the online prediction method for the chemical quality of sintered ore in this application, using specific experimental data.
[0192] In this experiment, the input parameters of the model in this application are compared with those of different models, as shown in Table 1. The prediction results of each model for the total iron (TFe) content are compared in Table 2. The prediction results of each model for the ferrous iron (FeO) content are compared in Table 3. The prediction results of each model for the alkalinity (R) are compared in Table 4. Wherein, RMSE represents the root mean square error, and R... 2 This represents the correlation coefficient.
[0193] Table 1
[0194]
[0195] Table 2
[0196]
[0197] Table 3
[0198]
[0199]
[0200] Table 4
[0201]
[0202] Comparing the proposed method with the Convolutional Neural Networks (CNN) + temporal feature model, it can be seen that the proposed method, after incorporating the sintering mill tail features, significantly improves the prediction accuracy and hit rate of ferrous FeO content in sinter. Simultaneously, it can improve the prediction accuracy of total iron (TFe) content and basicity while maintaining the hit rate. Comparing the proposed method with the CNN + mean model, it can be seen that considering its temporal features can significantly improve the prediction accuracy and hit rate of sinter chemical quality compared to simply introducing the mean of sintering state data within the corresponding time period. Comparing the proposed method with the Deep Belief Networks (DBN) model, it can be seen that the proposed model can more accurately achieve online prediction of sinter chemical quality, providing more positive guidance for sintering operation optimization.
[0203] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for online prediction of the chemical quality of sintered ore, characterized in that, include: Determine multiple sintering process parameters related to the chemical quality of sintered ore during sintering; Obtain the temporal characteristics of the multiple sintering process parameters at the current prediction time; The sintering machine tail image data from the previous prediction time to the current prediction time is acquired, and the key frame sequence of the sintering machine tail image data is determined by using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image. The Unet segmentation network is used to extract the primary features of each key frame in the key frame sequence, and the secondary features of the sintering machine tail key frame at the current prediction time are obtained based on the primary features of each key frame. The time series features, the secondary features of the key frame of the sintering machine tail, and the chemical quality value of the sinter at the previous prediction time are input into the quality prediction model to perform quality prediction, and the predicted value of the chemical quality of the sinter at the current prediction time is obtained. The step of extracting the primary features of each keyframe in the keyframe sequence using the Unet segmentation network, and obtaining the secondary features of the sintering machine tail keyframe at the current prediction time based on the primary features of each keyframe, includes: For each keyframe in the keyframe sequence, the keyframe is input into the encoder of the Unet segmentation network for processing to obtain the first-level features of the keyframe. For each keyframe, the first-level features are input into the Flatten layer for processing, and the data output from the Flatten layer is input into the first Dense layer for processing to obtain the feature vector of the keyframe. The average value of the feature vectors of all key frames is calculated and input into the second Dense layer for processing to obtain the temporal features of the sintering machine tail image frame at the current prediction time. The temporal features of the sintering machine tail image frame at the current prediction time are expanded to obtain the secondary features of the key frame of the sintering machine tail at the current prediction time. The period from the previous prediction time to the current prediction time includes multiple sintering cycles, and the sintering machine tail image data includes the sintering machine tail image data corresponding to each sintering cycle. The step of using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image to determine the keyframe sequence of the sintering machine tail image data includes: For each sintering cycle, the sintering machine tail image data corresponding to the sintering cycle is input into the Unet segmentation network for processing to obtain the combustion zone shape corresponding to each sintering machine tail image in the sintering cycle, and the sintering machine tail image corresponding to the combustion zone shape with the largest area is taken as the key frame of the sintering cycle. The set of key frames from all sintering cycles is taken as the key frame sequence of the sintering machine tail image data; The step of inputting the time-series features, the secondary features of the sintering machine tail keyframe, and the chemical quality value of the sinter at the previous prediction time into the quality prediction model to perform quality prediction and obtain the predicted value of the chemical quality of the sinter at the current prediction time includes: The time-series characteristics of the multiple sintering process parameters at the current prediction time, the secondary characteristics of the sintering machine tail keyframe at the current prediction time, and the chemical quality value of the sintered ore at the previous prediction time are input into the quality prediction model to perform quality prediction, and a two-dimensional matrix of the predicted values of the chemical quality of the sintered ore at the current prediction time is obtained. Calculate the average value of the two-dimensional matrix and use the average value as the predicted value of the chemical quality of the sinter at the current prediction time.
2. The prediction method according to claim 1, characterized in that, The formula for calculating the average of all eigenvectors is: in, The average of all eigenvectors. For the first Feature vectors of key frames in each sintering cycle , The number of the plurality of sintering cycles.
3. The prediction method according to claim 1, characterized in that, The two-dimensional matrix of the predicted values of the chemical quality of the sinter at the current prediction time. The expression is: The calculation of the average value of the two-dimensional matrix includes: Through formula Calculate the average value of the two-dimensional matrix; in, For the The Middle Line number The value of the column, The average value of the two-dimensional matrix. For the The Middle Line number The value of the column, The number of data points contained in any one of the plurality of sintering process parameters from the previous prediction time to the current prediction time.
4. The prediction method according to claim 1, characterized in that, Before determining the keyframe sequence of the sintering machine tail image data using the Unet segmentation network for extracting the combustion zone morphology of the sintering machine tail image, the prediction method further includes: Eliminate noise in the sintering machine tail image data; Color correction is performed on the sintering machine tail image data after noise elimination; The image data of the sintering machine tail after color correction is dehazed.