Online prediction method for ignition loss rate of light-burned magnesia powder in flash light-burning process of magnesite

Abstract

The invention discloses an online prediction method for the ignition loss of light-burned magnesia powder in the flash light-burning process of magnesite, which comprises the following steps: analyzing the flash light-burning process of magnesite, measuring the ignition loss of the light-burned magnesia powder as a dominant variable, and establishing a training sample set by taking important parameters influencing the dominant variable as auxiliary variables; carrying out standardization processing on training sample set data, carrying out dimension reduction processing on input data by utilizing a PCA method, and eliminating a redundant relationship among auxiliary variables; establishing a prediction model based on a WOA-XGBoost method, and performing iterative training in combination with training sample set data to obtain an ideal model for predicting the ignition loss rate of the light-burned magnesia powder; and correcting the ideal model of the light-burned magnesia powder ignition loss by adopting a method of combining deviation compensation correction and instant learning. The prediction model provided by the invention can predict the ignition loss of the light-burned magnesia powder on line as a result, is used for predicting the product quality in the flash light-burning process of the magnesite, and is used for feeding back and adjusting a production process link.

Description

Technical Field

[0001] The invention relates to the field of advanced control technology for a flash calcining process of magnesite, in particular to an online prediction method for the ignition loss rate of calcined magnesium powder in a flash calcining process of magnesite. Background Art

[0002] The flash calcining process of magnesite uses magnesite with a particle size of 100-200μm as raw material. After being dried and preheated, the magnesite is sent into the core equipment flash calcining furnace by a blower. The combustion air generated by the blower and the finished product blower is sent into the flash calcining furnace. At the same time, natural gas is sent into the flash calcining furnace at different heights for multi-stage combustion. Magnesite decomposition reaction and magnesium oxide sintering reaction occur in the high-temperature gas in the flash calcining furnace, and then it is collected by a cyclone separator and cooled to finally form the product light-burned magnesium powder.

[0003] The quality indicator of light-burned magnesium powder is the ignition loss rate. At present, the method used by manufacturers using the flash calcination process of magnesite to test the ignition loss rate of light-burned magnesium powder is: the test personnel manually sample the light-burned magnesium powder once every two hours, and then conduct manual testing to obtain the ignition loss rate. This testing method has the problems of low sampling frequency and long lag time, which makes it difficult to guide production in a timely manner, and seriously hinders the stability of product quality.

[0004] With the development of science and technology, process modeling and artificial intelligence technology provide a feasible solution for the online measurement of the ignition loss rate of light-burned magnesium powder. If an online prediction model of the ignition loss rate of light-burned magnesium powder with strong generalization ability can be established, and the ignition loss rate of light-burned magnesium powder can be measured online, it will be of great significance for stabilizing the production of the flash ignition process of magnesite and implementing intelligent optimization control of the flash ignition process of magnesite to improve the quality of light-burned magnesium powder. Summary of the invention

[0005] In view of the problem that the ignition loss rate of light-burned magnesium powder in the flash calcining process of magnesite cannot be detected online, the present invention provides an online prediction method for the ignition loss rate of light-burned magnesium powder in the flash calcining process of magnesite, which can provide online and continuous measurement values ​​of the ignition loss rate of light-burned magnesium powder, has certain prediction accuracy, generalization ability and adaptability, and reduces the risk of overfitting.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0007] An online prediction method for the ignition loss rate of light-burned magnesium powder in a flash calcining process of magnesite is provided. The following steps are performed to obtain a prediction model and perform online correction, which is used to predict the quality of unknown magnesite flash calcining process products, and the ignition loss rate of light-burned magnesium powder is output as a result for evaluating the product quality of the flash calcining process of magnesite. The method comprises the following steps:

[0008] The flash calcining process of magnesite was analyzed, the ignition loss rate of calcined magnesium powder was measured as the dominant variable, and the important parameters affecting the dominant variable were used as auxiliary variables to establish a training sample set;

[0009] The training sample set data is standardized, and the input data is reduced in dimension using the PCA method to eliminate the redundant relationships between the auxiliary variables;

[0010] A prediction model was established based on the WOA-XGBoost method, and iterative training was performed with the training sample set data to obtain an ideal model for predicting the ignition loss rate of light-burned magnesium powder.

[0011] The ideal model of ignition loss rate of light-burned magnesium powder was corrected by combining deviation compensation correction with real-time learning.

[0012] In the training sample set, {auxiliary variables are input and dominant variables are output}, and the auxiliary variables are represented by u 1 ′~u 11 ′ indicates that the raw ore powder flow rate, main combustion natural gas flow rate, first-layer compensation natural gas flow rate, second-layer compensation natural gas flow rate, air supply volume, raw powder air supply volume, finished product conveying air volume, combustion chamber outlet temperature, first-layer supplementary combustion outlet temperature, second-layer supplementary combustion outlet temperature, and flash light burning furnace outlet temperature are included as auxiliary variables.

[0013] The regularization method is used to standardize the training samples.

[0014] The prediction model established based on the WOA-XGBoost method includes:

[0015] Set the model parameters of the prediction model, including the learning rate b 1 , the minimum threshold b for the loss caused by node splitting 2 , minimum node weight threshold b 3 , L1 regularization coefficient b 4 , L2 regularization coefficient b 5 , and the threshold range of each parameter;

[0016] b 1 ~b 5 The vector composed of 1 ,b 2 ,...,b 5 ] as the position of the individual whale, randomly generate the positions of several individual whales within the range from the minimum to the maximum value of each parameter;

[0017] For each whale's individual position, parameter b 1 ~b 5 , using sample data and XGBoost method to build a model, and using WOA to adjust the parameter b in the model1 ~b 5 Perform iterative optimization; after S iterations, the parameters obtained by optimization are Substitute it into the XGBoost model to get the final prediction model.

[0018] In the sth (s=1,2,...,S)th iteration of the WOA optimization process, the position of the kth (k=1,2,...,K)th whale individual is used, and the XGBoost modeling method is adopted to establish an initialization objective function that considers the loss function and model complexity for the output, and the objective function is transformed using the second-order Taylor expansion; a decision tree is defined, and the tree structure is introduced into the objective function, and the optimal tree is constructed using the greedy algorithm. The above process is repeated t times to finally generate t decision trees; the outputs of these decision trees are added together and multiplied by the learning rate as the predicted output of the ignition loss rate prediction result of the dolomite flash light burning process obtained by the kth whale individual in the sth iteration.

[0019] The final prediction model is:

[0020]

[0021] Among them, u I Represents the newly arrived input data u I The results after standardization are Indicates the kth iteration of S * The output of the a-th decision tree of individuals, It represents the prediction result of the ignition loss rate of light-burned magnesium powder in the flash light-burning process of magnesite obtained based on the input data.

[0022] The method combining deviation compensation correction with real-time learning includes: calculating the root mean square error RMSE of the manual detection value and the predicted value of the ignition loss rate of light-burned magnesium powder within the latest continuous sampling time;

[0023] If RMSE < threshold 1, the original prediction model remains unchanged;

[0024] If threshold 1≤RMSE≤threshold 2, the deviation compensation method is used for correction, and the actual error is added to the predicted result as the final prediction result;

[0025] If RMSE>threshold 2, the just-in-time learning method is used for correction. The F sample data with the highest similarity to the newly arrived input data are selected for preprocessing, added to the training sample set, and the WOA-XGBoost method is continued to be used to rebuild the prediction model.

[0026] The present invention has the following advantages and beneficial effects:

[0027] 1. It can measure online the ignition loss rate of light-burned magnesium oxide powder in the flash-burning process of magnesite, which is very important and difficult to detect online by traditional methods, and provides the prerequisite for the implementation of intelligent optimization control of the flash-burning process of magnesite and the stability of product quality;

[0028] 2. The WOA-XGBoost method is used to establish a prediction model, which enhances the generalization ability of the model, overcomes the risk of overfitting, and improves the prediction accuracy of the model;

[0029] 3. Prediction model correction can perform online real-time correction on the established prediction model, and has the ability to adapt to changes in working conditions, greatly reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the schematic diagram of the flash burning process of magnesite;

[0031] Figure 2 This is a flow chart of the process for predicting the ignition loss rate of light-burned magnesium oxide powder in the flash calcining process of magnesite. DETAILED DESCRIPTION

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The present invention is further described in detail below in conjunction with the accompanying drawings and examples.

[0034] The method of the present invention provides an online prediction method for the ignition loss rate of light-burned magnesium powder in the flash ignition process of magnesite, which can continuously give the measured value of the ignition loss rate of light-burned magnesium powder online, has certain prediction accuracy, generalization ability and adaptability, and reduces the risk of overfitting. The specific process is as follows Figure 2 As shown, the following steps are included.

[0035] 1. Selection of auxiliary variables:

[0036] By analyzing the flash calcination process of magnesite, it can be determined that there are 11 important parameters that affect the ignition loss rate y of light-burned magnesium oxide powder: raw ore powder flow, main combustion natural gas flow, first-layer compensation natural gas flow, second-layer compensation natural gas flow, air supply, raw powder air supply, finished product conveying air, combustion chamber outlet temperature, first-layer supplementary combustion outlet temperature, second-layer supplementary combustion outlet temperature, flash calcination furnace outlet temperature. Therefore, these 11 variables are used as auxiliary variables of the prediction model, and u 1 ′~u 11 'express.

[0037] 2. Input data preprocessing:

[0038] The input data of the training samples are standardized and PCA dimension reduction is performed.

[0039] (1) Standardization:

[0040] The regularization method is used to standardize the training samples.

[0041]

[0042] In the formula, u ij represents the i-th input variable u i The jth sample value of The i-th input variable

[0043] Quantity u i The sample mean of i is the i-th input variable u i The sample standard deviation, u ij for u ij The standardized values ​​are: i=1,2,…,11; j=1,2,…,N; N is the number of samples.

[0044] (2) PCA dimensionality reduction of input variables:

[0045] Assume input data u ij The matrix U' is composed of the data u after standardization ij The matrix formed is U∈R N ×11 . Follow these steps:

[0046] Step 1: Calculate the covariance matrix

[0047] Step 2: Find the eigenvalues ​​and corresponding eigenvectors of the covariance matrix C, and arrange the eigenvalues ​​from large to small. Let the arranged eigenvalues ​​and their corresponding eigenvectors be θ i and p i , i=1,2,…,11.

[0048] Step 3: Calculation v is the number of new input variables obtained after the input variables are reduced in dimension. The specific value needs to be solved based on the inequality.

[0049] Step 4: Arrange the eigenvectors in columns from left to right according to the corresponding eigenvalues ​​to form a feature matrix P 0 , take P 0 The first v columns of form the matrix P.

[0050] Step 5: Calculate X = UP. X is the new input variable matrix after data preprocessing.

[0051] 3. Establishment of prediction model based on WOA-XGBoost method:

[0052] Follow these steps:

[0053] Step 1: Set the number of iterations s = 1, set the maximum number of iterations S = 30, and go to Step 2;

[0054] Step 2: Let k = 1, and the number of whales be K = 25. Let b i =b imin +rand ik (b imax -b imin ), where i = 1, 2, ..., 5, rand ik (i=1,2,...,5,k=1,2,...,K) is a random number uniformly distributed in the range [0,1]. The minimum and maximum values ​​of each parameter are: b 1min =10 -6 , b 2min =0.01, b 3min =0.01, b 4min =0.001, b 5min =0.001, b 1max =10 5 , b 2max =10, b 3max =20, b 4max =1000, b 5max = 1000. Randomly generate the positions of K whale individuals, and the position of each whale individual is represented by L s (k) = [b 1 ,b 2 ,...,b 5 ](k=1,2,...,K), go to Step 3;

[0055] Step 3: Let t = 1, set the maximum number of trees to T = 15, and go to Step 4:

[0056] Step 4: Initialize and construct the objective function.

[0057] Assume that the initialization objective function formula of the t-th decision tree of the k-th whale individual in the s-th iteration of the output is as follows:

[0058]

[0059] Among them, the loss function of the i-th (i=1,2,...,N) sample is y i Represents the i-th sample value of the output; represents the prediction value of the first t decision trees of the k-th whale individual for the i-th sample in the s-th iteration; represents the model complexity of the ath tree of the kth whale individual in the sth iteration. Go to Step 5;

[0060] Step 5: Objective function transformation.

[0061] Substituting formula (5) into formula (4), we can get

[0062]

[0063] Among them, x i represents the i-th row of X, And is a constant. Taylor expansion of equation (6) yields

[0064]

[0065] in They are all constants. After simplification and removing the constant term, the objective function is equation (8). Go to Step 6.

[0066]

[0067] Step 6: Define a decision tree and introduce the tree into the objective function.

[0068] The expression of a decision tree is defined as follows:

[0069] Bundle Defined as in Represents the result of a tree, which is used to convert the input x i ∈R s Mapped to a leaf node (assuming the tree has leaf nodes), Is the length A one-dimensional vector representing the weight of the leaf node (i.e., the predicted value of the decision tree).

[0070] The following formula is used to measure the complexity of a tree:

[0071]

[0072] in, for No. Therefore, the objective function is transformed into:

[0073]

[0074] All samples x belonging to the jth leaf node i , is classified into a leaf node sample set, mathematically expressed as: Calculate the loss of all leaf node samples as make

[0075] Then the objective function becomes To minimize the objective function, let right The partial derivative of is zero, so Determine whether it is satisfied or If satisfied, keep unchanged; if not satisfied, or Go to Step 7;

[0076] Step 7: Build the optimal tree.

[0077] Determination of candidate split points: For each new input variable, select quartiles based on the sample size, and each quartile is a candidate split point.

[0078] Use the greedy algorithm to split the tree nodes:

[0079] Starting from the depth of the tree being 0, a round of splitting is attempted for each leaf node in the tree, that is, considering all new input variables, splitting is performed at each candidate splitting point of each new input variable; after each split, the original leaf node continues to split into two left and right child leaf nodes, and the sample set in the original leaf node will be dispersed to the left and right leaf nodes according to the judgment rule of the node (that is, compared with the value at the candidate splitting point); after a new node is split, it is detected whether this split will bring gain to the loss function. The gain is defined as follows:

[0080]

[0081] The L subscript refers to the optimal value of the objective function when partitioning to the left subtree, and the R subscript refers to the optimal value of the objective function when partitioning to the right subtree. 2 , stop the split; if the number of samples of any leaf node is lower than the threshold Nu min , stopping this division, where Nu min A positive integer in the range [1, 5]; if the depth of the tree reaches the maximum value dep max , stop this split, where dep max The value range is a positive integer in the interval [2, 7]; otherwise, after all candidate split points are split, select Gain ≥ b 2 The largest candidate split point is taken as the best split point for this round of splitting. Then, at the next leaf node, all candidate split points of all new input variables are considered, and the above process is repeated to perform the next round of splitting on the leaf node until the conditions for stopping splitting are met. The final number of leaf nodes is Go to Step 8;

[0082] Step 8: Determine whether t reaches T. If so, go to Step 9. Otherwise, set t = t + 1 and go to Step 4.

[0083] Step 9: Get the kth model based on the position parameters of the kth whale individual Among them, u I Represents the newly arrived input data u I The results after standardization are represents the output of the a-th decision tree of the k-th individual in the s-th iteration, Represents the output of the kth model. The fitness function is used Calculate the fitness of the model where x i Indicates the i-th row of X. Determine whether k reaches K. If so, go to Step 10. Otherwise, set k = k + 1 and go to Step 3.

[0084] Step 10: Select The maximum value among the two is taken as the individual whale position corresponding to this value as the optimal individual whale position of the sth iteration and is recorded as L * (s) and The serial number of the whale individual is recorded as k * , determine whether s=S is satisfied, and go to Step 11;

[0085] Step 11: If satisfied, set the optimal whale individual position L *(s) is output as the final result and substituted into the XGBoost model. The obtained model is the prediction model of the ignition loss rate of light-burned magnesium powder in the flash light-burning process of magnesite. The model is established. I 'When it arrives, This is the final prediction model output, where u I Represents the newly arrived input data u I The results after standardization are Indicates the kth iteration of S * The output of the ath decision tree of individuals (i.e. the best individuals), represents the predicted result of the ignition loss rate of light-burned magnesium powder in the flash-burning process of magnesite obtained according to the input data; if it is not satisfied, let D k (s) = C k (s)L * (s)-L k (s), represents the position vector between the kth whale individual and the random whale individual; represents the distance vector between the kth whale individual and the optimal whale individual (i.e., the optimal solution); let A k (s) = 2a(s)r 1k (s)-a(s), a(s) is a parameter that decreases linearly from 2 to 0, and the expression is r 1k (s) is a random number in the range [0,1], C k (s) = 2r 2k (s), r 2k (s) is a random number in the range [0,1]. Generate a random number β(s) in the range [0,1]. Check whether β(s)≤0.5 and go to Step 12.

[0086] Step 12: If β(s)≤0.5, determine whether |A k (s)|<1, if satisfied, then use L k (s+1)=L * (s)-A k (s)D k (s) Update the position of the kth whale; if |A k (s)|≥1, then use L k (s+1)=L rand (s)-A k (s)D k (s) Update the position of the kth whale, L rand (s) represents the position of a randomly selected individual whale, go to Step 13;

[0087] Step 13: If β(s)>0.5, use Update the position of the kth whale individual; where b is the spiral shape parameter, and its value range is [0.2, 0.5]; l represents a random number uniformly distributed in the interval [-1, 1], and go to Step 14;

[0088] Step 14: Let s=s+1 and go to Step 2.

[0089] 4. Calibration of prediction model:

[0090] The prediction model of the ignition loss rate of light-burned magnesium powder is corrected by combining the method of bias compensation correction with real-time learning. The manual detection value of the ignition loss rate of light-burned magnesium powder is compared with the predicted value result. If the RMSE (root mean square error) of the latest 4 consecutive hours is less than the pre-defined threshold 1, the original prediction model is kept unchanged; if the RMSE is greater than the threshold 1 and less than the threshold 2, the deviation compensation method is used for correction, that is, the actual error is added to the predicted result as the final prediction result; if the RMSE is greater than the threshold 2, the F sample data with the highest similarity to the newly arrived input data are selected, the newly arrived input data is preprocessed, the prediction model is re-established using the WOA-XGBoost method, and the newly arrived input and output data are added to the training samples.

[0091] Among them, the threshold 1 range of the ignition loss rate of light-burned magnesium powder is between [0.1% and 0.2%], and the threshold 2 range is between (0.2% and 0.3%]. The similarity is calculated using formula (12). The value range of F is between 300 and 500.

[0092]

[0093] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An online prediction method for the ignition loss rate of light-burned magnesium powder in the flash light-burning process of magnesite, characterized in that: The following steps are performed to obtain a prediction model and perform online correction, which is used to predict the quality of unknown magnesite flash calcining process products, and output the ignition loss rate of calcined magnesium powder as a result to evaluate the product quality of magnesite flash calcining process products. The method comprises the following steps: The flash calcining process of magnesite was analyzed, the ignition loss rate of calcined magnesium powder was measured as the dominant variable, and the important parameters affecting the dominant variable were used as auxiliary variables to establish a training sample set; The training sample set data is standardized, and the input data is reduced in dimension using the PCA method to eliminate the redundant relationships between the auxiliary variables; A prediction model was established based on the WOA-XGBoost method, and iterative training was performed with the training sample set data to obtain an ideal model for predicting the ignition loss rate of light-burned magnesium powder. The ideal model of ignition loss rate of light-burned magnesium powder was corrected by combining deviation compensation correction with real-time learning.

2. The online prediction method for ignition loss rate of light-burned magnesium powder in the flash light-burning process of magnesite according to claim 1 is characterized in that: In the training sample set, {auxiliary variables are input and dominant variables are output}, and auxiliary variables are represented by u1′~u 11 ′ indicates that the raw ore powder flow rate, main combustion natural gas flow rate, first-layer compensation natural gas flow rate, second-layer compensation natural gas flow rate, air supply volume, raw powder air supply volume, finished product conveying air volume, combustion chamber outlet temperature, first-layer supplementary combustion outlet temperature, second-layer supplementary combustion outlet temperature, and flash light burning furnace outlet temperature are included as auxiliary variables.

3. The online prediction method for ignition loss rate of light-burned magnesium powder in a flash-burning process of magnesite according to claim 1 is characterized in that: The regularization method is used to standardize the training samples.

4. The online prediction method for ignition loss rate of light-burned magnesium powder in a flash-burning process of magnesite according to claim 1 is characterized in that: The prediction model established based on the WOA-XGBoost method includes: Set the model parameters of the prediction model, including the learning rate b1, the minimum threshold value of the loss caused by node splitting b2, the minimum node weight threshold b3, the L1 regularization coefficient b4, the L2 regularization coefficient b5, and the threshold range of each parameter; The vector [b1,b2,...,b5] consisting of b1 to b5 is used as the position of the individual whale, and the positions of several individual whales are randomly generated within the range of the minimum to maximum value of each parameter; For the parameters b1~b5 in each whale individual position, the model is established using sample data and the XGBoost method, and WOA is used to iteratively optimize the parameters b1~b5 in the model; after S iterations, the optimized parameters are Substitute it into the XGBoost model to get the final prediction model.

5. The method for online prediction of ignition loss rate of light-burned magnesium powder in a flash-burning process of magnesite according to claim 4, characterized in that: In the sth (s=1,2,...,S)th iteration of the WOA optimization process, the position of the kth (k=1,2,...,K)th whale individual is used, and the XGBoost modeling method is adopted to establish an initialization objective function that considers the loss function and model complexity for the output, and the objective function is transformed using the second-order Taylor expansion; a decision tree is defined, and the tree structure is introduced into the objective function, and the optimal tree is constructed using the greedy algorithm. The above process is repeated t times to finally generate t decision trees; the outputs of these decision trees are added together and multiplied by the learning rate as the predicted output of the ignition loss rate prediction result of the dolomite flash light burning process obtained by the kth whale individual in the sth iteration.

6. The method for online prediction of ignition loss rate of light-burned magnesium powder in the flash calcining process of magnesite according to any one of claims 1 to 5, characterized in that: The final prediction model is: Among them, u I Represents the newly arrived input data u I The results after standardization are Indicates the kth iteration of S * The output of the a-th decision tree of individuals, It represents the prediction result of the ignition loss rate of light-burned magnesium powder in the flash light-burning process of magnesite obtained based on the input data.

7. The method for online prediction of ignition loss rate of light-burned magnesium powder in a flash-burning process of magnesite according to claim 1, characterized in that: The method combining deviation compensation correction with real-time learning includes: calculating the root mean square error RMSE of the manual detection value and the predicted value of the ignition loss rate of light-burned magnesium powder within the latest continuous sampling time; If RMSE < threshold 1, the original prediction model remains unchanged; If threshold 1≤RMSE≤threshold 2, the deviation compensation method is used for correction, and the actual error is added to the predicted result as the final prediction result; If RMSE>threshold 2, the just-in-time learning method is used for correction. The F sample data with the highest similarity to the newly arrived input data are selected for preprocessing, added to the training sample set, and the WOA-XGBoost method is continued to be used to rebuild the prediction model.

Images