Intelligent sensing method and system for sintering mixture moisture and its change trend

Abstract

The application discloses an intelligent sensing method and system for sintering mixing water content and its change trend, obtains a sintering mixing image, extracts surface visual features, deep semantic features and strong correlation process variable features of the sintering mixing image, fuses the features, constructs a mixing water content detection model based on the fused features, simultaneously collects mixing water content time series data in continuous time, constructs a water content change trend prediction model based on the mixing water content time series data and a correlation process variable matrix, and predicts the sintering mixing water content change trend, so that the technical problem that the prior art cannot accurately detect and predict the sintering mixing water content is solved, the intelligibility of the mixing water content intelligent detection method is strengthened, the detection accuracy and reliability are improved, and timely and effective auxiliary information is provided for on-site water observation operators.

Description

Technical Field

[0001] This invention mainly relates to the field of blast furnace smelting technology, specifically to an intelligent sensing method and system for the moisture content of sintering mixtures and its changing trends. Background Technology

[0002] Sintering is a crucial process for transforming various powdered iron-containing raw materials into a dense mass to improve their metallurgical properties. The main components of the sintering mix are various iron ores, cold return ore, coke powder, quicklime, etc. The mixing and water addition process aims to obtain the necessary material temperature and a favorable sintering particle size distribution. Therefore, the moisture content of the mix is ​​one of the key parameters in sintering. Insufficient moisture leads to low granulation efficiency, dust generation on-site, resulting in metal loss and deteriorating working conditions. Conversely, excessive moisture causes a large amount of steam mixed with dust to escape, and reduces the permeability of the sintered ore, making ignition difficult and increasing sintering time and costs. Therefore, accurate and real-time detection of the moisture content of the mix is ​​of great significance for improving the production quality and economic efficiency of sintering and subsequent metallurgical processes.

[0003] Currently, the real-time performance and accuracy of methods for measuring moisture content in mixed materials urgently need improvement. Existing moisture measurement methods can be categorized into two types: direct measurement and indirect measurement. Direct measurement methods, such as thermal drying and rapid weight loss methods, offer high accuracy but suffer from significant time lag, and are only used as a means of calibrating and correcting moisture content in mixed materials. Indirect measurement methods, such as infrared spectroscopy and microwave scattering, are highly susceptible to the effects of on-site dust, water mist, changes in light, and frequent fluctuations in material flow, resulting in significant deviations between the measured results and laboratory test accuracy. This leads to serious inaccuracies in real-time moisture measurement of mixed materials. Furthermore, if moisture change trends can be predicted in advance during the mixing process, it will help operators adjust the amount of water added in a timely manner, thereby reducing production costs and ensuring sintering quality.

[0004] Therefore, this invention proposes an intelligent sensing method for the moisture content and its changing trend of sintering mixtures. On the one hand, based on the mixing process mechanism and expert experience in moisture assessment, a mixture moisture detection model jointly driven by empirical knowledge, process mechanism, and production data is constructed using a feature-level fusion strategy. On the other hand, this invention performs subsequence decomposition of the mixture moisture time-series data using variational mode decomposition, combines relevant process variables, and designs an interactive correlation gate gating module based on long short-term memory network units to construct a mixture moisture prediction model guided by trend consistency, thereby achieving prediction of the changing trend of mixture moisture content.

[0005] Patent application CN115222661A discloses an intelligent online detection method and system for moisture in sintered materials. By acquiring images of sintered materials, the method extracts the color features of the images in HSV space, uses a gray-level co-occurrence matrix to extract the texture features of the images, and extracts the clumping shadows of the images as morphological features. Based on these features, a moisture detection model for sintered materials is established, which improves the stability and accuracy of moisture detection in sintered materials.

[0006] This invention incorporates the experience of experts in mixing and adding water, but does not take into account the changes in sintering-related process variables, and extracts features only through mixing images, which introduces errors caused by the subjectivity of human experience.

[0007] The invention patent with publication number CN113447392A proposes a calibration and verification method for a sintering mixture moisture measuring device. The relative true value of the moisture content of the mixture is obtained by drying method, and the moisture percentage of the mixture is expressed by linguistic variables or fuzzy mathematics by worker / expert experience method to verify the consistency of the dynamic measurement of large / small or slightly large / slightly small values ​​of the device. Finally, based on the above, the continuous measurement value of the device is calibrated and verified by big data analysis relative verification method.

[0008] The invention forms a relatively complete measurement, calibration and verification scheme based on scientific computing and big data processing. However, the fuzzy mathematical representation in the scheme requires tactile means such as gripping and pressing the mixed materials to make judgments. Expert experience is relatively abstract and difficult to quantify scientifically. Summary of the Invention

[0009] The present invention provides an intelligent sensing method and system for the moisture content and its changing trend of sintering mixtures, which solves the technical problem that the existing technology cannot accurately detect and predict the moisture content of sintering mixtures.

[0010] To address the aforementioned technical problems, this invention proposes an intelligent sensing method for the moisture content and its changing trend in sintering mixes, comprising:

[0011] Acquire images of the sintering mixture and preprocess them.

[0012] Extract surface visual features and deep semantic features from sintering mixture images.

[0013] Obtain the relevant process variable matrix related to the sintering mixing process, and extract the features of strongly correlated process variables closely related to the moisture content of the mixture based on the relevant process variable matrix.

[0014] A combination of bilinear pooling and channel attention mechanisms is used to fuse surface visual features, deep semantic features, and strongly correlated process variable features to obtain fused features.

[0015] Based on the fusion features, a mixture moisture detection model is constructed to detect the moisture content of the mixture.

[0016] Collect continuous time-series data on the moisture content of the mixture, and construct a moisture change trend prediction model based on the mixture moisture time-series data and related process variable matrices to predict the moisture change trend of the sintering mixture.

[0017] Furthermore, the surface visual features extracted from the sintering mixture image include:

[0018] Extract the grayscale run-length matrix features from the mixed material image.

[0019] Extract LBP image entropy features from the mixture image.

[0020] Ridge features are extracted from the mixture image. These ridge features are used to characterize the degree of agglomeration and moisture content of the sintered mixture.

[0021] The quicklime white spot feature is extracted from the mixture image. The quicklime white spot feature is used to characterize the distribution of quicklime that undergoes a digestion reaction with water in the sintering mixture.

[0022] The surface visual feature vector is obtained based on the gray-scale run-length matrix features, LBP image information entropy features, ridge features, and quicklime white spot features.

[0023] Furthermore, extracting ridge features from the mixture image includes:

[0024] The grayscale image of the sintering mixture is binarized, and the projection vector is obtained by one-dimensional projection in the vertical direction.

[0025] The peak features of the histogram corresponding to the projection vector are obtained by the local maximum method. The peak features are used to characterize the ridge features, and the peak features include the number of peaks, the mean and maximum values ​​of peak heights, and the total peak area.

[0026] Furthermore, the features of white spots in the quicklime in the mixed image include:

[0027] Perform gamma transformation on the mixture image.

[0028] Perform a closing operation on the mixed image after gamma transformation.

[0029] The edges of the white patches are extracted using the Canny operator to obtain the edge extraction results.

[0030] Based on the edge extraction results, the candidate white spots are outlined and filled, and the pixel area is used as the filtering condition to obtain the distribution map of quicklime.

[0031] The number of white spots, the total area of ​​white spot pixels, and the average area in the mixed material image are used as characteristics of white spots in quicklime.

[0032] Furthermore, a matrix of relevant process variables related to the sintering mixing process is obtained, and based on this matrix, features of strongly correlated process variables closely related to the moisture content of the mixture are extracted, including:

[0033] Acquire process data related to the sintering and mixing process, and integrate them into a matrix of relevant process variables through splicing operations;

[0034] Time delay correction is performed on the relevant process variable matrix.

[0035] Based on the time-delay corrected matrix of relevant process variables, the strongly correlated process variables closely related to the sintering and mixing process are selected by using the maximum information coefficient.

[0036] The strongly correlated process variables are normalized, and a process feature extraction network based on a multilayer perceptron is constructed to obtain the feature vector of strongly correlated process variables closely related to the moisture content of the mixture.

[0037] Furthermore, a combination of bilinear pooling and channel attention mechanisms is used to fuse surface visual features, deep semantic features, and strongly correlated process variable features, resulting in fused features including:

[0038] The first feature is obtained by concatenating the surface visual feature vector with the feature vector of the strongly correlated process variable at the corresponding time, and the deep semantic feature is denoted as the second feature.

[0039] Perform a tensor product operation on the first and second features to obtain the tensor feature matrix.

[0040] Based on the tensor product feature matrix, a preliminary fusion matrix is ​​obtained through summation and pooling.

[0041] The initial fusion matrix is ​​expanded to obtain the fusion feature vector, which is then used as the input vector for the mixture moisture detection model.

[0042] Furthermore, time-series data on the moisture content of the mixture over a continuous period are collected. Based on this data and the relevant process variable matrix, a moisture change trend prediction model is constructed to predict the moisture change trend of the sintering mixture, including:

[0043] Collect time-series data of mixture moisture content over a continuous period of time, and use a mixture moisture detection model to fill in missing values ​​in the time-series data of mixture moisture content over a continuous period of time, thereby obtaining the mixture moisture sequence.

[0044] Obtain the sequence of relevant process variable matrices corresponding to the moisture content sequence of the mixture.

[0045] The variational mode decomposition method is used to decompose the mixture moisture sequence into subsequences to obtain the mixture water molecule sequence, which includes the overall trend component, local fluctuation component, and residual term of the mixture moisture.

[0046] The autocorrelation estimates of the overall trend component and fluctuation component of the mixture moisture content are calculated separately to obtain the first autocorrelation estimate and the second autocorrelation estimate.

[0047] The cross-correlation estimates of the overall trend component of the mixture moisture content and its corresponding related process variable characteristics are calculated separately, as well as the cross-correlation estimates of the local fluctuation component and its corresponding related process variable characteristics, to obtain the first cross-correlation estimate and the second cross-correlation estimate. The related process variable characteristics are obtained from the related process variable matrix.

[0048] Based on the first autocorrelation estimate, the second autocorrelation estimate, the first cross-correlation estimate, and the second cross-correlation estimate, the output gate of the long short-term memory network unit is calculated to obtain the interactively related gated long short-term memory network unit.

[0049] Based on the interactive association-gated long short-term memory network unit, a prediction sub-model corresponding to the water molecule sequence of the mixture is established.

[0050] Based on the forecast results of the forecast sub-model, a moisture change trend forecasting model is constructed to predict the moisture change trend of the sintering mixture.

[0051] Furthermore, based on the first autocorrelation estimate, the second autocorrelation estimate, the first cross-correlation estimate, and the second cross-correlation estimate, the formula for calculating the output gate of the Long Short-Term Memory (LSTM) network unit is as follows:

[0052]

[0053] Among them, o t Indicates the output gate, W o and b o These represent the weights and biases of the output gate, respectively. and Let represent the implicit states of the overall trend component, local fluctuation component, and residual term of the mixture moisture content at time t, respectively. and These represent the implicit states of the overall trend component, local fluctuation component, and residual term of the mixture moisture content at time t-1, respectively. and These represent the first autocorrelation estimate and the second autocorrelation estimate, respectively. and These represent the first cross-correlation estimate and the second cross-correlation estimate, respectively. and These represent the overall trend component, local fluctuation component, and residual term of the mixture moisture content, respectively. and These represent the autocorrelation weights of the overall trend component and the local fluctuation component of the mixture moisture content, respectively. and These represent the autocorrelation weights of the overall trend component and local fluctuation component of the mixture moisture content relative to their respective latent states at the previous time step. This represents the cross-correlation weight between the overall trend component of the mixture's moisture content and its corresponding relevant process variable characteristics. This represents the cross-correlation weight between the local fluctuation component of the mixture moisture content and its corresponding relevant process variable characteristics. and b represents the cross-correlation weights between the overall trend component and the local fluctuation component of the mixture moisture content and their respective latent states at the previous time step. Th1 and b Th2 Let b represent the autocorrelation bias and cross-correlation bias used for updating the hidden state of the overall trend component of the mixed moisture content, respectively. Fh1 and b Fh2 W represents the autocorrelation bias and cross-correlation bias used for updating the hidden state of the local fluctuation component of the mixture moisture content, respectively. Ry W represents the autocorrelation weight of the moisture residual term in the mixture. Rh b represents the correlation weight between the moisture residual term of the mixture and its previous hidden state. Rh ω represents the bias of the hidden state of the moisture residual term in the mixture, ω represents the weight hyperparameter for the autocorrelation and cross-correlation terms, tanh[·] is the hyperbolic tangent activation function, and ⊙ represents the Hadamard product operation.

[0054] Furthermore, after establishing the prediction sub-model corresponding to the water molecule sequence of the mixture, the following steps are also included:

[0055] A loss function is constructed to optimize the prediction sub-model, and the formula for calculating the loss function is as follows:

[0056]

[0057] in, y and y represent the predicted value output by the forecast sub-model and the corresponding actual moisture content of the mixture, respectively. Loss MSE and Loss trend Let represent the root mean square loss function and the trend loss function, respectively; ξ and ζ represent the weights of the root mean square loss function and the trend loss function, respectively; γ and η represent the first and second hyperparameters that determine the importance of trend consistency, respectively; n and m represent the input window length and output window length of the forecast sub-model, respectively; and (Δy) i and (Δ 2 y) j These represent the i-th first-order difference value and the j-th second-order difference value of the mixed water molecule sequence in the input window data, respectively. and These represent the i-th first-order difference value and the j-th second-order difference value of the output window data, respectively.

[0058] The present invention provides an intelligent sensing system for the moisture content and its changing trend of sintering mixtures, comprising:

[0059] The present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the intelligent sensing method for the moisture content and its changing trend of sintering mixtures provided by the present invention.

[0060] This invention focuses on the mixing process in a sintering plant and proposes an intelligent sensing method for the moisture content and trends of the mixed materials. Firstly, it quantifies surface information such as the morphology and texture of the mixed materials using the aforementioned method. Secondly, it identifies strongly correlated process variables based on the process mechanism and extracts relevant features using a multilayer perceptron. Then, it extracts deep semantic features from the mixed material images using a residual neural network, avoiding the subjectivity and randomness of human experience-based judgment. Finally, it constructs a mixed material moisture detection model by complementary fusion of the aforementioned multimodal features using a strategy combining bilinear pooling and channel attention mechanisms, achieving accurate detection of mixed material moisture. Secondly, it utilizes the mixed material moisture sequence and time-series data of related process variables, combining variational mode decomposition and an improved long short-term memory network to construct a moisture prediction model guided by interactive correlation information mining and trend consistency, achieving accurate characterization of the mixed material moisture change trend.

[0061] This invention is the first to construct a mixture moisture perception model driven by expert knowledge, process mechanism and production data, which solves the problem of difficulty in quantifying the visual features of mixtures, enhances the interpretability of intelligent mixture moisture detection methods, improves detection accuracy and reliability, and proposes a mixture moisture trend prediction method, which enhances the prediction model's ability to understand the correlation of different time series data, thereby providing timely and effective auxiliary information for on-site water monitoring operators.

[0062] The beneficial effects of this invention specifically include:

[0063] (1) To address the problem of inaccurate moisture measurement in mixed materials, a moisture detection model driven by a combination of empirical knowledge, process mechanism, and production data is proposed. First, expert experience is effectively introduced, and combined with the process mechanism of adding water to mixed materials, various surface visual features of mixed materials are scientifically quantified. Then, process variables strongly correlated with moisture in mixed materials are screened with sufficient interpretability, and a multilayer perceptron is constructed to extract the correlation between moisture in mixed materials and strongly correlated process variables. Second, deep semantic features of mixed material images are mined through residual neural networks to achieve effective complementarity with artificial features, thereby constructing an intelligent moisture sensing model for mixed materials.

[0064] (2) To address the difficulty in representing the morphology and texture of mixed materials, surface visual features such as ridge line features, quicklime white spot features, grayscale run length matrix features, and texture information entropy were designed. Ridge line related statistical features can effectively describe the degree of agglomeration of mixed materials; quicklime white spot features indirectly reflect its key influence on the moisture content of mixed materials by detecting the amount of quicklime remaining in the mixed material image; while grayscale run length matrix statistical features and texture information entropy can analyze the grayscale texture information and complexity of mixed materials, thereby reflecting the agglomeration effect and moisture content of mixed materials.

[0065] (3) To fully utilize the multimodal data related to the mixture, a strategy combining bilinear pooling and channel attention mechanism is used to complement and fuse them. Specifically, the bilinear pooling method is used to capture the interaction relationship between the pairs of features by utilizing the second-order statistical information of shallow artificial features and deep semantic features. The channel attention mechanism is introduced to learn the importance of the fused features and adaptively assign different weights, thereby achieving accurate detection of moisture in the mixture.

[0066] (4) To achieve timely and accurate forecasting of mixed material moisture content, an interactive correlation gated module was designed, and a moisture forecasting model guided by trend consistency was constructed. Specifically, after performing variational mode subsequence decomposition on the mixed material moisture content, a forecasting sub-model was established for each mixed material water molecule sequence and related process variable data. An interactive correlation calculation mechanism was designed, and the interaction relationship between variables and variables was explicitly mined based on the LSTM forecasting sub-model to provide global state correlation information for the model. Furthermore, a trend loss function was proposed, and a mixed material moisture forecasting model guided by trend consistency was constructed, thereby realizing the forecasting of seven trends in mixed material moisture content and providing effective auxiliary information for decision-making by on-site water monitoring personnel. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the implementation steps of the mixture moisture detection and prediction method in Embodiment 2 of the present invention.

[0068] Figure 2 The results of the moisture detection experiment in Embodiment 2 of the present invention;

[0069] Figure 3 This is a comparison of the fusion feature effects in Embodiment 2 of the present invention;

[0070] Figure 4 The results of the moisture forecasting experiment in Example 2 of this invention;

[0071] Figure 5 This is a structural block diagram of an intelligent sensing system for the moisture content and its changing trend in sintered mixtures, according to an embodiment of the present invention.

[0072] Figure label:

[0073] 10. Memory; 20. Processor. Detailed Implementation

[0074] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to the accompanying drawings and preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0075] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0076] Example 1

[0077] The present invention provides an intelligent sensing method for the moisture content and its changing trend in sintering mixtures, comprising:

[0078] Step S101: Obtain the sintering mixture image and preprocess the sintering mixture image.

[0079] Step S102: Extract the surface visual features and deep semantic features of the sintering mixture image.

[0080] Step S103: Obtain the relevant process variable matrix related to the sintering mixing process, and extract the strongly correlated process variable features closely related to the moisture content of the mixture based on the relevant process variable matrix.

[0081] Step S104: Use a combination of bilinear pooling and channel attention mechanisms to fuse surface visual features, deep semantic features, and strongly correlated process variable features to obtain fused features.

[0082] Step S105: Based on the fusion features, construct a mixture moisture detection model to detect the moisture content of the mixture.

[0083] Step S106: Collect time-series data of moisture content in the mixture over a continuous period of time, and construct a moisture change trend prediction model based on the time-series data of moisture content in the mixture and the relevant process variable matrix to predict the moisture change trend of the sintering mixture.

[0084] The intelligent sensing method for moisture content and its changing trend in sintering mixtures provided in this invention acquires images of the sintering mixture, extracts surface visual features, deep semantic features, and strongly correlated process variable features from these images, and fuses these features using a combination of bilinear pooling and channel attention mechanisms to obtain fused features. Based on these fused features, a moisture content detection model is constructed to detect the moisture content of the mixture and collect time-series data of the moisture content over a continuous period. Furthermore, based on the time-series data of the moisture content and the relevant process variable matrix, a moisture change trend prediction model is constructed to predict the moisture change trend of the sintering mixture. This method solves the technical problem of existing technologies being unable to accurately detect and predict the moisture content of sintering mixtures, enhances the interpretability of intelligent moisture content detection methods, improves detection accuracy and reliability, and enhances the prediction model's understanding of the correlation between different time-series data, thereby providing timely and effective auxiliary information for on-site moisture monitoring personnel.

[0085] Example 2

[0086] like Figure 1 As shown in the figure, the specific implementation steps of the intelligent sensing method for the moisture content and its changing trend of sintering mixture proposed in this embodiment are as follows.

[0087] (1) Use the visible light camera equipment on site to obtain visible light images of the sintering mixture and preprocess the image data.

[0088] (2) Based on the experience and knowledge of on-site water inspection experts, the surface visual features such as the morphology and degree of clumping of the mixture are quantified. At the same time, the deep semantic features of the image are extracted using a residual neural network as the input vector of the mixture moisture detection model.

[0089] (3) Based on the process mechanism of mixing and adding water, process variables that are strongly correlated with the moisture content of the mixture are selected, and the correlation features between the two are learned through a multilayer perceptron and used as the input vector of the mixture moisture detection model.

[0090] (4) By combining bilinear pooling with channel attention mechanism, the above multimodal features are fused to establish a mixed material moisture detection model.

[0091] (5) The time series data of moisture in the mixture is decomposed into subsequences by variational mode decomposition method. The water molecule sequence and related process variables are used for modeling. An interactive correlation gate gating mechanism is designed based on LSTM unit to construct a trend prediction model of moisture in the mixture and describe the changing trend.

[0092] The specific implementation plan is as follows.

[0093] (1) Image dataset construction and preprocessing:

[0094] A visible light camera and related equipment are installed on the cantilever beam above the conveyor belt after the first mixing and water addition process to acquire images of the mixture. The region of interest is obtained through angle correction and region cropping, followed by noise reduction processing. An adaptive histogram equalization method is then used to highlight the morphological details of the mixture. The steps are as follows:

[0095] Step 1: Since the feeding belt is fixed and the position of the object is known, a rectangle is selected to frame the main body of the mixture in the image; at the same time, in order to reduce the complexity of the model parameters, the image f(x,y) is obtained by downsampling and then grayscale operation is performed to obtain g(x,y).

[0096] Step 2: The mixed material image suffers from blurred details and low contrast due to the influence of water mist and dust on site. Therefore, an adaptive histogram equalization method with contrast suppression is used to enhance the image and improve the local contrast of the mixed material image. On the other hand, median filtering is used to suppress image noise while preserving its edge characteristics. The filtered image is denoted as h(x,y).

[0097] (2) Extracting shallow visual features of the mixture based on expert experience:

[0098] Step 1: Extract the grayscale run-length matrix features of the mixed image.

[0099] This invention reflects the distribution of material texture by constructing a grayscale run-length matrix and extracting relevant statistical features. The texture of the mixture refers to the visual feature formed by the grayscale difference between the shadows created by the agglomeration of the mixture and the surrounding flat surface. When the moisture content of the mixture is high, resulting in more agglomerates, the shadow areas of the surface image are complex and the texture is rough; conversely, the texture is smoother and more uniform.

[0100] The grayscale run-length matrix reflects the texture distribution by describing the continuous length of grayscale values. Coarse textures have a larger number of pixels with the same grayscale value around them, resulting in more frequent long grayscale runs; conversely, fine textures have fewer pixels with the same grayscale value around them, resulting in more frequent short grayscale runs. Therefore, this method can analyze the local texture information of the mixed grayscale image, thereby reflecting the degree of agglomeration and moisture content of the mixture. The specific process is as follows.

[0101] First, pixel normalization is performed on the image h(x,y) using the following formula, where X is the gray value of the image pixel.

[0102]

[0103] Then, a gray-level run-length matrix is ​​constructed for the mixed image. The matrix element p(i,j|θ) represents the statistical measure of a pixel with gray value i appearing j times consecutively in the x-direction, where θ takes values ​​of 0°, 45°, 90°, and 135°. Relevant statistical features are calculated using the following formula:

[0104] 1) Short run advantage: describes the frequency of a certain gray value appearing for a short continuous length, that is, it measures the small-scale, delicate mixed texture in the image.

[0105]

[0106] 2) Long run advantage: It reflects the frequency of a certain gray value appearing for a relatively long continuous length, that is, it measures a larger range of coarser mixed texture.

[0107]

[0108] 3) Gray-scale non-uniformity: This parameter describes the uniformity of gray-scale values, that is, it reflects the uniformity of the size of clumps in the gray-scale image of the mixture.

[0109]

[0110] 4) Run unevenness: The larger this parameter is, the higher the run unevenness of each gray value in the mixed image, and vice versa.

[0111]

[0112] Finally, the statistical features of the above gray-level run-length matrix are concatenated, as shown in the following formula:

[0113] f GLCM =concat[SRE,LRE,GLN,RLN] (6)

[0114] Step 2: Extract the LBP image entropy features from the mixture image.

[0115] To measure the complexity of the texture in a mixed image, the mixed image is first converted into a local binary mode that highlights the texture representation, and then its information entropy is calculated. The details are as follows.

[0116] The basic idea of ​​LBP (Local Binary Patterns) is to use the gray value of the center pixel in a neighborhood as a reference, compare it with the gray values ​​of its surrounding pixels, and represent the local texture using binary encoding, as shown in the following formula:

[0117]

[0118]

[0119] Where I(c) is the gray value of the center pixel, and I(p) is the gray value of the surrounding pixels in the neighborhood.

[0120] The advantage of LBP lies in its good expression and differentiation of different mixed texture features. It also has grayscale invariance for uneven lighting and dust and fog interference in local areas. That is, when the grayscale value of the local image increases or decreases at the same time due to lighting or water fog, its encoded value remains almost stable.

[0121] Therefore, the mixed image is converted into a local binary mode to enhance its texture representation. Then, a one-dimensional information entropy is calculated on the texture image, and the entropy value is used to measure the complexity of the mixed texture. The formula is as follows:

[0122]

[0123] Where, p i The proportion of the LBP mixture image occupied by pixels with grayscale value i.

[0124] Step 3: Extract the ridge features from the mixture image.

[0125] This invention defines for the first time a ridge feature to describe the morphology of a mixture and characterize its clumping degree and moisture content. The principle is as follows: According to expert experience, in the morphology of a mixture with high moisture content, the clumping part can be observed to be raised, and there is a clear boundary line between the bright part and the shadow part, resembling a mountain ridge. Moreover, there is a significant difference in the gray value of the pixels on both sides of the ridge line. Therefore, its one-dimensional projection in the vertical direction can effectively reflect the existence of the ridge line, thereby characterizing the clumping condition and moisture content of the mixture.

[0126] First, the grayscale image of the mixture is binarized, and a projection vector v is obtained through one-dimensional projection in the vertical direction. The i-th element corresponds to the cumulative amount of black pixels in the i-th column of the mixture image. Then, the peaks of the histogram of the projection vector are obtained through the local maximum method to characterize the ridge features, including the number of peaks, the mean and maximum values ​​of peak heights, and the sum of peak areas. The process is as follows.

[0127] 1) Smooth and denoise the projection vector v. Use Savitzky-Golay smoothing filter to fit the data within the sliding window with a k-th order polynomial to avoid noise points being misidentified as peaks.

[0128] 2) Candidate local peaks were detected using the adjacent comparison method, and the following screening criteria were applied:

[0129] height: Crest height threshold, meaning data points below a specified crest height are not considered. The threshold is set to [value missing]. Where N v Let v be the length of vector v.

[0130] distance: Set the minimum horizontal distance between adjacent peaks to an empirical value of 20, thereby removing smaller peaks and avoiding noise interference.

[0131] width: The width of the peak. Peaks that are too wide and exceed the range will not be considered.

[0132] 3) After obtaining the peaks that meet the conditions through the above steps, calculate the ridge features of the mixture image. The area of ​​each peak is approximated by the product of its height and half-width, defined using the minimum peak height. The final ridge features include the number of peaks P. num Mean peak height P height Mean peak width P width The mean area of ​​the wave crest, A mean and maximum value A max The ridge feature vector is obtained by concatenation, as shown in the following formula:

[0133] f ridge =concat[P num ,P height ,P width A mean A max (10)

[0134] Step 4: Extract the white spot features of quicklime from the mixed material image.

[0135] According to the process mechanism of adding water to the mixture, quicklime in the mixture undergoes a digestion reaction with water to generate Ca(OH)2 colloidal particles, which enhance the binding properties and granulation effect of the mixture. However, quicklime breaks down and disperses upon contact with water, leaving behind white particles that are the quicklime that has not yet reacted with water. Excessive residual quicklime indicates insufficient moisture in the mixture, while insufficient residual quicklime indicates excessive moisture. Therefore, the distribution of quicklime in the image is of significant reference value for analyzing the moisture content of the mixture.

[0136] Since the mixture is usually dark brown, there is a significant gray-scale gradient between the quicklime and the surrounding area. Therefore, this invention proposes a quicklime feature extraction method based on the Canny edge operator, which indirectly reflects the moisture content of the mixture by quantifying the remaining quicklime in the mixture image.

[0137] First, grayscale image enhancement is performed using gamma transformation, where parameter γ = 1.5.

[0138] g′(x,y)=g(x,y) γ (11)

[0139] Secondly, morphological closing operations are used to improve the integrity and continuity of the white patches' edges. Specifically, dilation operations are used to fill in smaller edge discontinuities, and erosion operations are used to reduce the expansion of the white patch's connected domain boundaries and the erroneous increase in area caused by dilation operations.

[0140] Then, the edges of the white spots are extracted using the Canny operator. The separation of the quicklime white spots from the mixed image mainly involves pixel gradient calculation, non-maximum suppression, and double threshold detection. The main formulas are as follows, where G is the pixel gradient intensity matrix and Θ is the pixel gradient direction matrix:

[0141]

[0142] Θ(i,j)=arctan(G y (i,j) / G x (i,j)) (13)

[0143] Finally, based on the edge extraction results described above, the candidate white spots are outlined and filled, with pixel area used as the filtering criterion to avoid over-detection or under-detection, thereby obtaining the distribution map of quicklime. Finally, the number N white spots in the mixed image is counted. CaO Total pixel area P CaO and average area Avg CaO The formula for identifying white spots in quicklime is as follows:

[0144] f spot =concat[N CaO ,P CaO Avg CaO (14)

[0145] The surface visual features of the mixed image are summarized by stitching together the visual representations obtained in Steps 1-4 above, as shown in the following formula:

[0146] f vis =concat[f GLCM ,f LBPEnt ,f ridge ,f spot (15)

[0147] (3) Obtain deep semantic features of the mixture based on deep learning methods

[0148] Since the aforementioned surface visual features of the mixed materials are extracted based on expert experience, they are often limited by the subjectivity of human experience, making it difficult to uncover deep correlations within the data. This results in low prediction accuracy and limited generalization performance of the model. In contrast, the high-level semantic features learned by deep neural networks can effectively supplement shallow feature information.

[0149] Therefore, a ResNet50 model is constructed as a deep semantic feature extractor for mixed material images. The enhanced image h(x,y) is used as input, fully utilizing the autonomous learning advantage of deep convolutional neural networks to obtain the deep semantic feature vector of the mixed material image. Furthermore, this method adaptively adjusts the original network structure, reducing the convolutional kernel size and adjusting the number of channels in the convolutional layers to make it more suitable for mixed material moisture detection tasks. The model uses the Adam optimizer with mean squared error as the loss function, as shown in the following formula.

[0150]

[0151] Where y is the actual moisture content of the mixture. is the model detection value, and N is the number of samples.

[0152] (4) Extracting features of strongly correlated process variables based on process mechanisms:

[0153] Iron ore powder and related ingredients are the main raw materials in the mixing process, and related process variables have a significant impact on the moisture content of the mixture. On the one hand, the moisture content of the mixture is closely related to the characteristics of the iron ore, specifically including its surface morphology, roughness, and internal pore structure. For example, limonite has a rough, porous surface and good hydrophilicity; as its proportion increases, the moisture content of the mixture also increases. Magnetite, on the other hand, is relatively smooth and has a regular shape, thus having poor hydrophilicity. On the other hand, the slaking reaction of quicklime consumes a significant amount of moisture.

[0154] Therefore, based on the above-mentioned process mechanism, a multilayer perceptron was constructed to extract the data dependency relationship between process variables and the moisture content of the mixture. First, multiple data points, such as the flow rates of iron ore powder silos, cold return ore silos, and quicklime silos, were selected and integrated into a matrix through a splicing operation, which was denoted as the relevant process variable matrix u. Second, considering the time difference between the data recording time of the upstream sintering process and the moisture detection time at the outlet of the mixing drum in actual production, and the delay in data transmission, this time lag was corrected. Finally, strongly correlated process variables were selected based on the maximum information coefficient (MIC). Variables with a MIC ≥ 0.25 were included in the selection.

[0155] Then, the selected strongly correlated process variables are normalized, and a feature extraction network for strongly correlated process variables based on a multilayer perceptron is constructed. Here, a multilayer perceptron with three hidden layers is defined, and the ReLU activation function is used. Finally, the feature vector f of the strongly correlated process variables of the mixture moisture is obtained. proc .

[0156] (5) Construct a mixture moisture detection model based on multimodal feature fusion:

[0157] After obtaining the surface visual features, deep semantic features, and strongly correlated process variable features of the image based on the above steps, feature fusion is performed by combining bilinear pooling and channel attention mechanisms.

[0158] Specifically, for a mixed image I at a certain moment, the surface visual feature vector f is first... vis The eigenvector f of the process variable strongly correlated with the corresponding time proc splicing to obtain f A The deep semantic features learned by ResNet50 are denoted as f. B Then, the tensor product of the two is taken to obtain the tensor characteristic matrix b, as shown in the following formula:

[0159] b(I,f A ,f B )=f A (I) T f B (I) (17)

[0160] Then, the initial fusion matrix ξ(I) is obtained through summation and pooling. This matrix is ​​expanded to obtain the intermediate vector x. After processing, the final fusion feature vector r is obtained, as shown in the following formula:

[0161]

[0162] x=vec(ξ(I)) (19)

[0163]

[0164] r = s / ||s||2 (21)

[0165] Where I is the mixing image at a certain moment, vec(·) represents the vector expansion operation, and sign(·) is the sign function.

[0166] This method captures the pairwise interactions between features by calculating the second-order statistical information of both surface features and deep semantic features. Next, the importance of the fused features needs to be measured; therefore, the aforementioned fused feature vector *r* is used as the subsequent input vector, and a channel attention mechanism is introduced to achieve adaptive weighted fusion of feature channels. First, feature compression is performed using channel descriptors, as shown in the following formula:

[0167]

[0168] Then, feature activation is performed. By learning the parameter W, the importance of each channel of the fused feature vector r is calculated and weights s are generated. These weights s are then multiplied with the original feature vector to obtain the weighted feature vector. The formula is as follows:

[0169] s = Fex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (23)

[0170]

[0171] in, Let σ and δ be the weights to be learned, and let σ and δ be the sigmoid and ReLU activation functions, respectively.

[0172] After mapping the above vectors through three fully connected layers, the final result of the mixture moisture detection is obtained.

[0173] (6) A water change trend forecasting model based on interactive correlation information mining and trend consistency guidance:

[0174] First, variational mode decomposition (VM) is used to decompose the time-series moisture data of the mixture into subsequences. Then, corresponding relevant process variable sequences are introduced. Data modeling is performed using the mixture water molecule sequence and the relevant process variable sequences. Interactive correlation gates are designed to improve the long short-term memory network units, thereby establishing a mixture moisture forecasting model guided by trend consistency. Finally, the moisture forecasting results are described in terms of trend. Specifically:

[0175] Step 1: Process time-series data using a sliding window operation. This involves constructing a continuous-time sequence of mixture moisture content and related process variables at one-minute sampling intervals, and imputing missing values ​​based on a mixture moisture detection model. The data is divided into n+m sliding windows, with the historical n-step mixture moisture content and related process variable matrix u as input, and the moisture values ​​for the next m steps as output. That is, the input window length is n, and the output window length is m.

[0176] Step 2: Subsequence decomposition of the mixture moisture sequence is performed using variational mode decomposition. Variational mode decomposition can adaptively determine the number of modes for a given sequence. Its constrained variational model is shown in the following formula, which is to find the mode function y that minimizes the sum of the bandwidths of the center frequencies of each mode component. (k) and center frequency ω (k) .

[0177]

[0178] Among them, y (k) Let ω be the eigenmode function of the mixture moisture sequence y. (k) For the corresponding center frequency, and T is the sequence length.

[0179] Setting the modal decomposition number to an empirical value of K = 3, the optimal solution to the constrained variational problem is iteratively searched, separating the eigenmode functions, i.e., three sequences of water molecules in the mixture. The low-frequency components correspond to the overall trend of the mixture's moisture content, the next lowest frequency components correspond to local fluctuations in moisture content, and the high-frequency components are the residual terms. Therefore, for the moisture content y of the mixture at a certain time t... t It can be expressed by the following formula.

[0180]

[0181] in, This indicates the overall trend of moisture content in the mixture. This indicates the local fluctuation in moisture content of the mixture. This represents the residual moisture content of the mixture.

[0182] Step 3: Design interactive association gates to improve the gating mechanism of long short-term memory network units, analyze the temporal dependencies of data, and build a trend consistency-guided mixed moisture trend forecasting model.

[0183] This embodiment improves the gating mechanism of the traditional LSTM cell internal structure by proposing an interactive correlation gate, which can more effectively mine the historical correlation of the mixture moisture itself and the interrelationship between the mixture moisture and process variables, thereby providing global state correlation information for the model. The autocorrelation and cross-correlation are transmitted and memorized through the hidden layer state, thereby improving the accuracy of mixture moisture prediction.

[0184] First, to reflect the causal interaction between the overall trend component and local fluctuation component of the mixture moisture series and its own historical state, the autocorrelation estimates of the overall trend component and local fluctuation component of the mixture moisture series are calculated respectively. and The formula is shown below.

[0185]

[0186]

[0187] in, b Tp These represent the autocorrelation weights and biases of the overall trend component of the mixture moisture content, respectively. b Fp These represent the autocorrelation weights and biases of the local fluctuation components of the mixture moisture content, respectively. These are the overall trend component of the mixed moisture content, the local fluctuation component, and the autocorrelation weights of their respective previous hidden states.

[0188] Then, the relevant process variable matrix u is normalized and convolutionally processed to learn the weights and obtain the dimensionality-reduced relevant process variable features u′, as shown in the formula below. Where Conv is the convolution operation, σ is the Sigmoid activation function, and T... u T represents the number of categories of the relevant process variables, and T represents the sequence length.

[0189]

[0190] Furthermore, in order to explore the interaction and influence between the overall trend component and local fluctuation component of the mixture moisture content and the characteristics of relevant process variables, and to achieve more efficient joint modeling among variables, the cross-correlation between the overall trend component of the mixture moisture content and the characteristics of relevant process variables is calculated. And the cross-correlation between the local fluctuation component of the mixture moisture content and the characteristics of relevant process variables. The formula is as follows:

[0191]

[0192]

[0193] in, b Tq These represent the cross-correlation weights and biases between the overall trend component of the mixture moisture content and the characteristics of relevant process variables. b Fq These represent the cross-correlation weights and biases between the local fluctuation component of the mixture moisture content and the characteristics of relevant process variables. These are the cross-correlation weights of the overall trend component, the local fluctuation component, and their respective latent states at the previous time step.

[0194] Then, the calculation results of the interaction correlation gate are used to update the model's hidden state, thereby learning and passing on the aforementioned autocorrelation and cross-correlation through the hidden state weights, enabling the model to memorize global state correlation information, as shown in the following formula:

[0195]

[0196]

[0197]

[0198] Among them, b Th1 b Th2 These are the autocorrelation bias and cross-correlation bias used for updating the hidden state of the overall trend component of the mixed moisture content, respectively. Fh1 b Fh2 These are the autocorrelation bias and cross-correlation bias used for updating the latent state of local fluctuations in the moisture content of the mixture, respectively. Ry W represents the autocorrelation weight of the moisture residual term.Rh b represents the association weight between the moisture residual term and its previous hidden state. Rh ω represents the bias of the hidden state of the moisture residual term, ω represents the weight hyperparameter for the autocorrelation and cross-correlation terms, tanh[·] is the hyperbolic tangent activation function, and ⊙ is the Hadamard product operation.

[0199] Finally, the improved output gate o t The calculation formula is as follows:

[0200]

[0201] In the above formulas (27)-(35), u t ′ represents the characteristics of the relevant process variables at time t. Let W represent the hidden states of the overall trend component of the mixture moisture content at time t, the hidden states of the local fluctuation component of the mixture moisture content, and the hidden state of the residual term. o b o These are the weights and biases of the output gate, respectively.

[0202] For the improved Long Short-Term Memory network units, a prediction sub-model LSTM is constructed respectively. k (k=1,2,3), Prediction Sub-model LSTM k The sequence of the k-th mixed water molecule containing n historical steps and the related process variable matrix. The input window data is used as the prediction value for the next m steps of the k-th mixed water molecule sequence. This is for the output window data. Considering the rationality of the forecast period, this method sets the input window length n=5 and the output window length m=3. Therefore, the k-th forecast sub-model LSTM... k The input-output mathematical expression is as follows:

[0203]

[0204] Furthermore, traditional forecasting methods typically only predict specific numerical values ​​while ignoring the trend changes in time-series data. This can easily lead to situations where the predicted numerical error is small, but the trend does not match the actual trend, which can mislead on-site decisions regarding water addition adjustments. Therefore, this invention proposes for the first time a trend consistency-guided loss function, as shown in the following formula:

[0205]

[0206]

[0207] in, y and y represent the predicted value output by the forecast sub-model and the corresponding actual moisture content of the mixture, respectively. Loss MSE and Loss trendLet represent the root mean square loss function and the trend loss function, respectively; ξ and ζ represent the weights of the root mean square loss function and the trend loss function, respectively; γ and η represent the first and second hyperparameters that determine the importance of trend consistency, respectively; n and m represent the input window length and output window length of the forecast sub-model, respectively; and (Δy) i and (Δ 2 y) j These represent the i-th first-order difference value and the j-th second-order difference value of the mixed water molecule sequence in the input window data, respectively. and These represent the i-th first-order difference value and the j-th second-order difference value of the output window data, respectively.

[0208] The improved loss function not only focuses on constraining numerical errors during training, but also guides the direction and magnitude of model training changes by constraining the difference between the actual moisture sequence and the predicted sequence. This improves the shape similarity between the predicted sequence and the actual moisture sequence, thus avoiding misreporting an upward trend as a downward trend or other similar situations. This makes it easier for operators at the mixing site to make the correct decision to increase or decrease the amount of water added based on the trend prediction results.

[0209] Step 4: Calculate the forecast results of each sub-model and describe the trend changes. Predicted moisture content of the mixture at time t. The sum of the outputs of the forecast sub-models in the previous step is shown in the following formula.

[0210]

[0211] in, Let LSTM represent the predicted values ​​of the overall trend component, local fluctuation component, and residual term of the mixture moisture content at time t. k This is the k-th forecast sub-model.

[0212] The trend description of the moisture forecast results for the mixture is based on the results, and the specific process is as follows.

[0213] First, a trend forecast window is set. This involves using historical two-minute data as the observation baseline and combining it with the predicted moisture content of the mixture for the next three minutes to determine the trend characterization window.

[0214] Then, the least squares method is used to fit the window data. Zero-order, first-order, and second-order polynomials are used sequentially to fit the data, and the fitting result with the smallest root mean square error is selected as the trend curve.

[0215] Finally, the trend is described according to the following rules:

[0216] a) When the fitting result is a constant c, the moisture change trend is range-stable;

[0217] b) When the fitting result is a linear equation, calculate its first derivative f′ and determine its sign. If f′>0, the trend is uniformly increasing; if f′<0, the trend is uniformly decreasing.

[0218] c) When the fitting result is a quadratic equation, calculate the first derivative f′ and the second derivative f″ at the midpoint. If f′>0 and f″>0, the moisture trend is accelerating upward. If f′>0 and f″<0, the moisture trend is decelerating upward. If f′<0 and f″>0, the moisture trend is decelerating downward. If f′<0 and f″<0, the moisture trend is accelerating downward.

[0219] This invention focuses on the mixing process in a sintering plant and proposes an intelligent sensing method for the moisture content and trends of the mixed materials. Firstly, it quantifies surface information such as the morphology and texture of the mixed materials using the aforementioned method. Secondly, it identifies strongly correlated process variables based on the process mechanism and extracts relevant features using a multilayer perceptron. Then, it extracts deep semantic features from the mixed material images using a residual neural network, avoiding the subjectivity and randomness of human experience-based judgment. Finally, it constructs a mixed material moisture detection model by complementary fusion of the aforementioned multimodal features using a strategy combining bilinear pooling and channel attention mechanisms, achieving accurate detection of mixed material moisture. Secondly, it utilizes the mixed material moisture sequence and time-series data of related process variables, combining variational mode decomposition and an improved long short-term memory network to construct a moisture prediction model guided by interactive correlation information mining and trend consistency, achieving accurate characterization of the mixed material moisture change trend.

[0220] This invention is the first to construct a mixture moisture perception model driven by expert knowledge, process mechanism and production data, which solves the problem of difficulty in quantifying the visual features of mixtures, enhances the interpretability of intelligent mixture moisture detection methods, improves detection accuracy and reliability, and proposes a mixture moisture trend prediction method, which enhances the prediction model's ability to understand the correlation of different time series data, thereby providing timely and effective auxiliary information for on-site water monitoring operators.

[0221] This invention proposes an intelligent sensing method for the moisture content and its changing trends in sintering mixtures. This method overcomes the problems of low accuracy and large lag in on-site moisture measurement methods and provides moisture trend information for on-site moisture monitoring operators. On one hand, it extracts surface visual features from mixture images based on expert experience and obtains deep semantic features using deep learning. Then, it analyzes the process mechanism to filter process variable information and constructs a mixture moisture detection model driven by experience knowledge, process mechanism, and production data using a feature-level fusion strategy, thereby improving the accuracy and reliability of intelligent mixture moisture detection methods. On the other hand, it fully utilizes historical moisture information and the temporal characteristics of related process variables, using variational mode decomposition to obtain subsequences of mixture moisture. It also designs an interactive correlation gate control mechanism to mine global state correlation information, thereby establishing a moisture forecasting model guided by trend consistency. A trend forecasting window is constructed using historical and forecasted moisture values, and the results of multinomial fitting are used to characterize seven changing trends of mixture moisture, thus providing timely and effective auxiliary information for water monitoring operators to adjust the amount of water added.

[0222] Specifically, in response to the current problems of low accuracy, large lag, poor interpretability, and inability to predict moisture content in sintering mixes, this invention proposes an intelligent sensing method for moisture content and its changing trends in sintering mixes.

[0223] On the one hand, in the task of detecting moisture in sintering mixtures, traditional methods rely on infrared moisture meters, microwave moisture meters, etc., which have low accuracy and large time lag. On-site, detection, calibration, and correction are usually carried out based on expert judgment and thermal drying methods. Newer methods mainly involve direct data modeling based on relevant process variables, or extracting mixture features and modeling them using existing image texture feature calculation methods. These methods rarely incorporate the process mechanism of water mixing and expert experience, and do not consider the interaction relationships between multi-source heterogeneous features. Therefore, existing methods have poor interpretability, are susceptible to errors caused by the subjectivity of human experience, have weak characterization capabilities for mixture image features, and struggle to fully utilize mixture image data, process variable data, and the correlation information between them.

[0224] Therefore, to fully utilize the multimodal data related to the mixing process, this invention first addresses the difficulty in representing the morphology and texture of the mixture. Based on expert experience and the reaction mechanism of water addition during mixing, it designs surface visual features such as ridge features, quicklime white spot features, grayscale run-length matrix features, and texture information entropy. Specifically, ridge features can effectively describe the degree of clumping and bulging in the mixture; quicklime white spot features indirectly reflect the amount of water consumed in the digestion reaction by detecting the amount of quicklime remaining in the mixture image; while grayscale run-length matrix statistical features and texture information entropy can analyze the grayscale texture information and complexity of the mixture image, thereby reflecting the degree of clumping and moisture content. Furthermore, to reduce the random errors caused by the subjectivity of feature extraction based on human experience, this invention utilizes deep learning methods to obtain deep semantic features of the mixture image, enhancing the generalization performance of the mixture moisture detection model. In addition, by analyzing the physicochemical properties of iron ore, such as surface roughness, pore structure, and hydrophilicity, as well as the process mechanism of water addition during mixing, this invention selects relevant process variables and extracts corresponding features with strong interpretability. Finally, targeting the aforementioned surface visual features, deep semantic features, and strongly correlated process variable features of the mixed material image, this invention constructs a mixed material moisture detection model driven by experience knowledge, process mechanism, and production data using a feature-level fusion strategy. Specifically, it employs a strategy combining bilinear pooling and channel attention mechanisms to complementarily fuse features, capturing the interaction relationships between pairs of feature elements, and introduces a channel attention mechanism to learn the importance of fused features, thereby improving the accuracy and reliability of the intelligent mixed material moisture detection method.

[0225] On the other hand, addressing the technological gap in predicting the moisture trend of sintering mixes, this invention designs a prediction model based on interactive correlation information mining and trend consistency guidance. Because this process involves complex steps such as batching, primary mixing with water, secondary mixing with water, and sintering, it is characterized by multiple variables, nonlinearity, and significant time delays. Therefore, there are varying degrees of time delays and interactions between the process variables and the target variable, moisture. For example, different iron ore grades during mineral processing, uneven mixing, and fluctuations in related process variables can all cause time-delayed fluctuations in the mix moisture at different scales. Thus, the time-series data of related process variables and the time-series data of mix moisture have different degrees of correlation. However, existing methods for modeling based on process data typically only select process variables with high correlation coefficients for direct modeling, ignoring the global state correlation information of the production process, thus having certain limitations. In addition, according to the mixing process mechanism and actual production requirements, the moisture content of the mixture will not change drastically, but should show a relatively smooth transition. Therefore, the moisture content of the mixture has autocorrelation within a certain time interval, that is, the moisture content of the mixture at future moments changes based on the current and historical moisture content. However, the current method does not utilize the autocorrelation of the time series data of the moisture content of the mixture, so it is difficult to reasonably and effectively explore the causal relationship of the trend change of the moisture time series data.

[0226] Therefore, considering the aforementioned industrial and technological background, this patent decomposes the time series data of mixture moisture into an overall trend component, a local fluctuation component, and a residual term using a variational mode decomposition method. This explicitly separates the trend and periodicity of the mixture moisture time series, reducing the complexity of data modeling to some extent. Furthermore, different forecast sub-models are constructed to predict the overall trend, local fluctuation, and residual of mixture moisture based on the characteristics and patterns of different sub-sequences, improving prediction accuracy. Simultaneously, this patent proposes an interactive correlation gate to improve the gating mechanism of the LSTM unit. The main idea of ​​the interactive correlation gate is to not only calculate the autocorrelation estimates of the overall trend component and local fluctuation component of mixture moisture, but also to calculate the cross-correlation estimates between the overall trend component and local fluctuation component of mixture moisture and the characteristics of relevant process variables. This allows the model to adaptively learn which process variables are more correlated with moisture trends and which are more correlated with moisture fluctuations. The aforementioned autocorrelation and cross-correlation are learned and passed through hidden state weights, enabling the model to memorize global state correlation information, which is beneficial for the model to better explore the mixture moisture itself and its interaction with various variables. Compared to traditional time series neural networks, this method overcomes their inability to efficiently acquire the correlation between target and state variables, thus learning the different contributions of various process variables to the trend and fluctuation of moisture content in the mixture, while also offering better interpretability. Furthermore, this invention proposes a trend consistency-guided loss function. Compared to traditional numerical error loss functions, the improved trend loss function guides the model's convergence direction, preventing situations where the model's predicted numerical error is small but the predicted moisture change trend does not match the actual trend. This improves the accuracy of the mixture moisture trend forecast, avoids erroneous adjustments to the mixing water dosage, and ultimately benefits on-site operators in making decisions to increase or decrease the water dosage based on the moisture trend forecast results.

[0227] This embodiment uses the first mixing process in the mixing area of ​​a sintering plant as the experimental object. A camera is installed on a suspended beam above the conveyor belt behind a mixing drum machine to collect real-time visible light images of the mixing process. Simultaneously, real process variables related to the moisture content of the mixing material are acquired from the plant's sintering database. Moisture content experts judge the mixing material's moisture content based on experience and periodically calibrate the values ​​using a drying method. The above real data and corresponding characteristics are applied to the model proposed in this invention to obtain the mixing material moisture detection results and trend prediction results, and to calculate evaluation indicators. The specific experimental results are as follows: Figure 2 The moisture detection results shown are as follows: Figure 3 The comparison of fusion feature effects shown and Figure 4 The experimental results for moisture forecasting are shown. The results demonstrate that the method described in this embodiment can effectively improve the accuracy and reliability of moisture detection in mixed materials and provide effective trend forecasting results.

[0228] Reference Figure 5 The intelligent sensing system for the moisture content and its changing trend of sintering mixtures proposed in this embodiment of the invention includes a memory 10, a processor 20, and a computer program stored in the memory 10 and executable on the processor 20. When the processor 20 executes the computer program, it implements the steps of the intelligent sensing method for the moisture content and its changing trend of sintering mixtures proposed in this embodiment.

[0229] The specific working process and working principle of the intelligent sensing system for the moisture content and its changing trend of sintered mixtures in this embodiment can be referred to the working process and working principle of the intelligent sensing method for the moisture content and its changing trend of sintered mixtures in this embodiment.

[0230] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent sensing of moisture content and its trend in sinter mix, characterized in that, The method includes: Acquire sintering mixture images and preprocess them; Extract surface visual features and deep semantic features from sintering mixture images; Obtain the relevant process variable matrix related to the sintering mixing process, and extract the features of strongly correlated process variables closely related to the moisture content of the mixture based on the relevant process variable matrix; A combination of bilinear pooling and channel attention mechanisms is used to fuse surface visual features, deep semantic features, and strongly correlated process variable features to obtain fused features. Based on the fusion features, a mixture moisture detection model is constructed to detect the moisture content of the mixture. Collecting continuous time-series data on the moisture content of the sintering mixture, and constructing a moisture change trend prediction model based on the time-series data and relevant process variable matrices to predict the moisture change trend of the sintering mixture, includes: Collect time-series data of mixture moisture content over a continuous period of time, and use a mixture moisture detection model to fill in missing values ​​in the time-series data of mixture moisture content over a continuous period of time, thereby obtaining the mixture moisture sequence. Obtain the sequence of process variable matrices corresponding to the moisture content sequence of the mixture; The variational mode decomposition method is used to decompose the mixture moisture sequence into subsequences to obtain the mixture water molecule sequence, which includes the overall trend component, local fluctuation component and residual term of the mixture moisture; The autocorrelation estimates of the overall trend component and fluctuation component of the mixture moisture content are calculated separately to obtain the first autocorrelation estimate and the second autocorrelation estimate. The cross-correlation estimates of the overall trend component of the mixture moisture content and its corresponding related process variable characteristics are calculated separately, as well as the cross-correlation estimates of the local fluctuation component and its corresponding related process variable characteristics, to obtain the first cross-correlation estimate and the second cross-correlation estimate. The related process variable characteristics are obtained from the related process variable matrix. Based on the first autocorrelation estimate, the second autocorrelation estimate, the first cross-correlation estimate, and the second cross-correlation estimate, the output gate of the long short-term memory network unit is calculated to obtain the interactively related gated long short-term memory network unit. Based on the interactive relational gating long short-term memory network unit, a prediction sub-model corresponding to the water molecule sequence of the mixture is established; Based on the forecast results of the forecast sub-model, a moisture change trend forecasting model is constructed to predict the moisture change trend of the sintering mixture.

2. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 1, characterized in that, The surface visual features extracted from the sintering mixture image include: Extract the grayscale run-length matrix features from the mixed image; Extract the LBP image entropy features from the mixture image; The ridge features of the mixture image are extracted, and the ridge features are used to characterize the degree of agglomeration and moisture content of the sintered mixture; Extract the quicklime white spot feature from the mixed image. The quicklime white spot feature is used to characterize the distribution of quicklime that undergoes a digestion reaction with water in the sintering mixture. Based on the grayscale run-length matrix features, LBP image information entropy features, ridge features, and quicklime white spot features, the surface visual feature vector is obtained.

3. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 2, characterized in that, Extracting ridge features from the mixture image includes: The grayscale image of the sintering mixture is binarized, and the projection vector is obtained by one-dimensional projection in the vertical direction. The peak features of the histogram corresponding to the projection vector are obtained by the local maximum method. The peak features are used to characterize the ridge features, and the peak features include the number of peaks, the mean and maximum values ​​of peak heights, and the total peak area.

4. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 3, characterized in that, The features of white spots in quicklime extracted from the mixed image include: Perform gamma transformation on the mixture image; Perform a closing operation on the gamma-transformed mixture image; The edges of the white patches are extracted using the Canny operator to obtain the edge extraction results; Based on the edge extraction results, the candidate white spots are outlined and filled, and the pixel area is used as the filtering condition to obtain the distribution map of quicklime. The number of white spots, the total area of ​​white spot pixels, and the average area in the mixed material image are used as characteristics of white spots in quicklime.

5. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 4, characterized in that, Obtain the matrix of relevant process variables related to the sintering mixing process, and extract the features of strongly correlated process variables closely related to the moisture content of the mixture based on the matrix: Acquire process data related to the sintering and mixing process, and integrate them into a matrix of relevant process variables through splicing operations; Time delay correction is performed on the relevant process variable matrix; Based on the relevant process variable matrix after time delay correction, the strongly correlated process variables closely related to the sintering and mixing process are screened out by the maximum information coefficient; The strongly correlated process variables are normalized, and a process feature extraction network based on a multilayer perceptron is constructed to obtain the feature vector of strongly correlated process variables closely related to the moisture content of the mixture.

6. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 5, characterized in that, A combination of bilinear pooling and channel attention mechanisms is used to fuse surface visual features, deep semantic features, and strongly correlated process variable features, resulting in fused features including: The first feature is obtained by concatenating the surface visual feature vector with the feature vector of the strongly correlated process variable at the corresponding time, and the deep semantic feature is recorded as the second feature. Perform a tensor product operation on the first feature and the second feature to obtain the tensor feature matrix; Based on the tensor product feature matrix, a preliminary fusion matrix is ​​obtained through summation and pooling. The initial fusion matrix is ​​expanded to obtain the fusion feature vector, and the fusion feature vector is used as the input vector of the mixture moisture detection model.

7. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 6, characterized in that, Based on the first autocorrelation estimate, the second autocorrelation estimate, the first cross-correlation estimate, and the second cross-correlation estimate, the formula for calculating the output gate of a Long Short-Term Memory (LSTM) network unit is as follows: ， in, Indicates the output gate. and These represent the weights and biases of the output gate, respectively. , and They represent The implicit state of the overall trend component, local fluctuation component, and residual term of the mixed material moisture content at all times. , and They represent The implicit state of the overall trend component, local fluctuation component, and residual term of the mixed material moisture content at all times. and These represent the first autocorrelation estimate and the second autocorrelation estimate, respectively. and These represent the first cross-correlation estimate and the second cross-correlation estimate, respectively. , and These represent the overall trend component, local fluctuation component, and residual term of the mixture moisture content, respectively. and These represent the autocorrelation weights of the overall trend component and the local fluctuation component of the mixture moisture content, respectively. and These represent the autocorrelation weights of the overall trend component and local fluctuation component of the mixture moisture content relative to their respective latent states at the previous time step. This represents the cross-correlation weight between the overall trend component of the mixture's moisture content and its corresponding relevant process variable characteristics. This represents the cross-correlation weight between the local fluctuation component of the mixture moisture content and its corresponding relevant process variable characteristics. and These represent the cross-correlation weights between the overall trend component and the local fluctuation component of the mixture moisture content and their respective latent states at the previous time step. and These represent the autocorrelation bias and cross-correlation bias used for updating the hidden state of the overall trend component of the mixed moisture content, respectively. and These represent the autocorrelation bias and cross-correlation bias used for updating the hidden state of the local fluctuation component of the mixed moisture content, respectively. This indicates the autocorrelation weight of the moisture residual term in the mixture. The weighted correlation between the moisture residual term of the mixture and its hidden state at the previous time step is given. This indicates the bias in the implicit state of the residual moisture term in the mixture. This represents the weight hyperparameters for the autocorrelation and cross-correlation terms. The hyperbolic tangent activation function is used. This represents the Hadamard product operation.

8. The intelligent sensing method for moisture content and its changing trend in sintered mixtures according to claim 7, characterized in that, After establishing the prediction sub-model corresponding to the water molecule sequence of the mixture, the following is also included: A loss function is constructed to optimize the prediction sub-model, and the formula for calculating the loss function is as follows: ， in, and These represent the predicted value output by the forecast sub-model and the corresponding actual moisture content of the mixture, respectively. and Let these represent the root mean square loss function and the trend loss function, respectively. and These represent the weights of the root mean square loss function and the trend loss function, respectively. and These represent the first and second hyperparameters, respectively, which determine the importance of trend consistency. and These represent the input window length and output window length of the forecast sub-model, respectively. and These represent the sequence numbers of the mixed water molecules in the input window data. The first-order difference and the... Each second-order difference value and These represent the first and second parts of the output window data, respectively. The first-order difference and the... Two second-order difference values.

9. An intelligent sensing system for the moisture content and its changing trend of sintering mixtures, the system comprising: The memory (10), the processor (20), and the computer program stored in the memory (10) and executable on the processor (20) are characterized in that the processor (20) implements the steps of the method according to any one of claims 1 to 8 when executing the computer program.

Images