Sintering condition prediction method based on Bayesian neural network

Abstract

The invention relates to the field of iron ore sintering production process prediction and control, in particular to a Bayesian neural network-based sintering condition prediction method, which comprises the following steps of: firstly, performing random forest extraction on importance characteristics by using currently obtained cleaned data; then, the sintering condition is divided according to the obtained position of the sintering end point, and the time series data are clustered through a fuzzy C-means clustering algorithm; and finally, training the data of different clusters by using a Bayesian neural network, and weighting the membership degree by using the obtained time sequence data of each air bellow so as to obtain a prediction result of the sintering working condition. According to the method, high-accuracy and high-robustness result prediction can be realized, so that managers and workers can understand the production condition more deeply, the stability of the sintering result is maintained, and the efficiency and income of enterprises are improved.

Description

Technical Field

[0001] The present invention relates to the field of prediction and control in the iron ore sintering production process, and particularly to a method for predicting sintering conditions based on a Bayesian neural network. Background Art

[0002] Sintering condition prediction is to predict the future sintering conditions of the machine during its operation by using the currently obtained data, so as to guide the staff to adjust the production of the machine and maintain the stability of the entire production process. The sintering conditions are mainly divided into three types: overburning, underburning, and normal burning. Using model algorithms to predict the sintering conditions in the production process enables managers and staff to better understand the production situation, maintain the stability of the sintering conditions, and improve the efficiency and benefits of the enterprise.

[0003] As a traditional industrial process, the sintering process has always attracted the attention of scholars. It is a time-consuming process with a complex flow. Due to numerous parameters, it is difficult for operators to make real-time adjustments to the operating state. If there are problems in the production process, at least a batch of products will be wasted, and at worst, the operation of the entire machine will be affected. This has hindered the improvement of the quality of sintered ore and the reduction of production costs. Therefore, developing a prediction of the results of the sintering process has important economic value.

[0004] When the entire production process shows a tendency to deviate from normal sintering due to reasons such as the environment and materials, how to use the currently known data to predict the future sintering conditions and provide the staff with information about the current production line is an urgent problem to be solved. Summary of the Invention

[0005] In order to solve the problems of low quality of sintered ore, high production cost, low accuracy of prediction results, poor robustness, and difficulty for operators to make real-time adjustments to the operating state in existing prediction methods, the present invention provides a method for predicting sintering conditions based on a Bayesian neural network, which mainly includes:

[0006] S1: Preprocess the time series data of the sintering end point and the time series data of the detection parameters;

[0007] S2: Use random forest to screen the importance of various time series data and construct a feature library;

[0008] S3: According to the constructed feature library, use fuzzy C-means clustering to cluster each time series data;

[0009] S4: Establish multiple Bayesian neural networks for the clustering results, and fuse the results output by the multiple Bayesian neural networks according to the membership degrees of the data to different clusters to achieve accurate prediction of the sintering conditions.

[0010] A computer device includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the above method.

[0011] A computer-readable storage medium stores a computer program, which when executed by a processor, implements the steps of the above method.

[0012] A computer program product includes a computer program or instruction, which when executed by a processor, implements the steps of the above method.

[0013] The beneficial effects brought by the technical solution provided by the present invention are as follows: The present invention uses the currently obtained data that has been cleaned to extract important features through random forest. Then, the fuzzy C-means clustering algorithm is used to cluster the time-series data. Finally, the Bayesian neural network is used to train the data of different clusters, and the results will be weighted by membership degrees to obtain the prediction result of the sintering working condition. By using the random forest algorithm to extract feature importance, the present invention excludes interfering features and makes the prediction result more accurate. The FCM is used to cluster the data to obtain membership degrees, and the method of weighted summation of multiple Bayesian neural networks is used to obtain the result, which has good robustness and stability. Using the Bayesian neural network to predict the sintering working condition has a high accuracy rate and robustness, enabling managers and staff to better understand the production status, maintain the stability of the sintering result, and improve the efficiency and benefits of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0015] Figure 1 is a flowchart of a method for predicting sintering working conditions based on a Bayesian neural network in an embodiment of the present invention;

[0016] Figure 2 is a schematic diagram of the feature importance ranking calculated by the random forest in an embodiment of the present invention;

[0017] Figure 3 is a schematic diagram of the confusion matrix of the Bayesian neural network for predicting the sintering working conditions in the next 1 minute in an embodiment of the present invention;

[0018] Figure 4 is a schematic diagram of the operation of the hardware device in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] In order to have a clearer understanding of the technical features, objectives, and effects of the present invention, the specific embodiments of the present invention will be described in detail below with reference to the drawings.

[0020] Embodiment 1

[0021] Please refer to Figure 1 , Figure 1 which is a flowchart of a method for predicting sintering conditions based on a Bayesian neural network in an embodiment of the present invention. First, the present invention divides sintering conditions by obtaining the position of the sintering end point and classifies the data. Then, the obtained time series data of each air box is used to predict the sintering conditions. In the figure, BNN1 is a network obtained by training a Bayesian neural network using the first type of data set generated by clustering, BNN2 is a network obtained by training a Bayesian neural network using the second type of data set generated by clustering, and BNN3 is a network obtained by training a Bayesian neural network using the third type of data set generated by clustering.

[0022] The specific steps are as follows:

[0023] (1) Collect historical sintering production data to obtain original sample data

[0024] The historical sintering production data is stored in the local database of the industrial control computer in the operation room in the form of a daily report. According to the data in the daily report, collect historical data for one month such as the sintering end point and the exhaust gas temperature of the i-th air box (i = 1, 2, 3, 5, 7, 9, 11, 13, 15, 17, 18, 19, 20, 21, 22, 23, 24) to form the original sample data. There are a total of 18 detection parameters.

[0025] (2) Preprocess the time series data of the sintering end point and the detection parameters

[0026] Preprocess the collected original sample data. Specifically, eliminate the zero detection data caused by sensor failures and shutdown conditions; the sampling interval of the data is 60 seconds. For the time series data of the sintering end point and the detection parameters, the length of each time subsequence is 10 minutes, containing 10 data points, and establish a sample database with these sample data;

[0027] (3) Construct a feature library

[0028] Now there are a total of 17 air box temperatures, which are T1, T2, T3, T5, T7, T9, T 11 , T 13 , T 15 , T 17 , T 18 , T 19 , T 20 , T 21 , T 22 , T 23 , T 24 ;

[0029] Statistic represents the average change in node splitting impurity of the j-th variable in all trees of the random forest; when analyzing using models based on decision trees such as random forest, the bellows temperature data obtained in S2.1 is used as the feature variable of the sample; by calculating the partitioning situation of each node based on the Gini index, the importance of these temperature variables in the model is evaluated, and the Gini index GI m The calculation formula of

[0030]

[0031] K is the number of categories in the bootstrap sample set, and m is a specific partitioning node in the decision tree, corresponding to a subset of the samples, is the probability estimate value that the samples at node m belong to the k-th category.

[0032] When the sample is three-class data, the Gini index GI of node m m is

[0033]

[0034] where is the probability estimate value that the sample belongs to any category at node m; k represents the category; the sample refers to the data sampled from the time series data of the sintering end point and the time series data of the detection parameters.

[0035] The variable T j appears M times in the i-th tree, then the importance of the variable T j in the i-th tree is

[0036]

[0037] According to the importance on the tree, calculate the Gini importance of the variable T j in the random forest,

[0038]

[0039] That is, the variable T j The Gini importance in the entire random forest is defined as

[0040]

[0041] where n is the number of classification trees in the random forest; represents the average change in node splitting impurity of the j-th variable in the i-th tree of the random forest, represents the average change in node splitting impurity of the m-th variable in the j-th tree of the random forest.

[0042] The importance of various types of time series data calculated using the random forest model is asFigure 2 As shown, according to this importance, each characteristic variable is screened, and the time series data with the score of all important indicators greater than 2 is used as the feature library.

[0043] (4) Use fuzzy C-means clustering (FCM) to cluster each time series data

[0044] (4-1) Obtained from step (2), a time subsequence with a length of 10 minutes and a sampling interval of 1 minute. Denote a set of time series data as X = {x1, x2, x3, …, x 10}, where each x i ∈ R d represents a time series with a length of d (d = 1).

[0045] (4-2) For each iteration t = 1, 2, …, T, the cluster center point matrix V (t-1) can be calculated according to the membership matrix U (t) at the previous moment, where the cluster center point is calculated according to the following formula,

[0046]

[0047] is the membership corresponding to this data point at the previous moment, x i is the time series data at this moment, and m is the fuzzy coefficient. C is the number of clusters and can be evaluated by the fuzzy classification coefficient.

[0048] (4-3) Since the sampling intervals of the time series data are equal, the Euclidean distance is selected to calculate the distance matrix D i from the data point x to the center (t) , where the Euclidean distance is calculated according to the following formula,

[0049]

[0050] where is the cluster center point corresponding to the data at this moment, x i is the time series data at this moment, and m is the fuzzy coefficient.

[0051] (4-4) Update the membership matrix U (t) at this moment according to the distance matrix D (t) calculated in step (4-3), where the membership is calculated according to the following formula,

[0052]

[0053] where, is the Euclidean distance of the data point at this moment, and m is the fuzzy coefficient.

[0054] (4 - 5) The membership degree matrix U at this moment calculated according to step (4 - 4) (t) and the membership degree matrix U at the previous moment (t-1) are compared. If the maximum iteration number T or the minimum error is reached, stop; otherwise, repeat (4 - 2), (4 - 3), and (4 - 4) for iteration. Output the clustering center V and the membership degree matrix U.

[0055] (5) Sintering condition prediction experiment

[0056] (5 - 1) Design the structure of the Bayesian neural network, which consists of two hidden layers, an input layer, and an output layer. Set the parameters of each layer to 128 and 64 neurons respectively.

[0057] (5 - 2) Build the overall model architecture. Divide the results obtained from different clusterings into a training set and a test set. Input the training set into the Bayesian neural network and update the weights of the Bayesian neural network using the gradient descent method until the set number of iteration rounds is reached to obtain a trained Bayesian neural network.

[0058] (5 - 3) Train multiple Bayesian neural networks respectively to obtain different weight parameters. Let the current data be X, and the membership degrees of this data for different clusterings be u1, u2, …, u n , place this data in different Bayesian neural networks for calculation, and obtain different results as c1, c2, …, c n , perform weighted calculation according to the obtained membership degrees, and the obtained prediction result is

[0059]

[0060] Use the method of weighted summation of multiple Bayesian neural networks to obtain the result, so as to predict the sintering condition. The prediction result of the Bayesian neural network for the sintering condition in the next 1 minute is as Figure 3 shown. The green part in the figure represents the number of correctly predicted samples, such as 190, 193, 196, and other colors represent the number of incorrectly predicted samples, such as 4, 7, 10. Through calculation, the overall correct rate of the prediction result can reach 97.2%. It can be seen that using this method can well predict the sintering condition, which has important economic value and application value.

[0061] Example 2

[0062] A computer device includes a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the above method.

[0063] Example 3

[0064] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above method.

[0065] Embodiment 4

[0066] A computer program product includes a computer program or instructions, which, when executed by a processor, implement the steps of the above method.

[0067] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting sintering conditions based on a Bayesian neural network, characterized in that, include: S1: Preprocessing the time series data of sintering endpoint and detection parameter; S2: Use random forest to screen the importance of various time series data and build a feature library; S3: Based on the constructed feature library, the time series data are clustered using fuzzy C-means clustering; S4: Multiple Bayesian neural networks are established based on the clustering results. The output results of multiple Bayesian neural networks are fused according to the membership of different clusters according to the data to achieve accurate prediction of sintering conditions.

2. The method for predicting sintering conditions based on a Bayesian neural network according to claim 1, wherein The preprocessing described in S1 is to eliminate zero detection data caused by sensor failure and shutdown, set the sampling interval of the data, and select the length of each time subsequence and the number of data points.

3. The method for predicting sintering conditions based on a Bayesian neural network according to claim 1, characterized in that, S2 includes the following processes: S2.1: Obtain the temperatures of several bellows: T1, T2, T3, T5, T7, T9, T 11 、T 13 、T 15 、T 17 、T 18 、T 19 、T 20 、T 21 、T 22 、T 23 、T 24 ; S2.2: When performing analysis using the random forest, use the bellows temperature data obtained in S2.1 as the characteristic variables of the samples; evaluate the importance of these temperature variables in the random forest by calculating the partitioning of each node based on the Gini index, where the Gini index GI m is calculated as where K is the number of classes of the self-help sample set, m is a specific division node in the decision tree, corresponding to a subset of the samples, is the probability estimation value that the sample of node m belongs to the k-th class; When the sample is three-class data, the Gini index GI of node m m is where k represents the category, is the probability estimate that the sample belongs to any category at node m; Variable T j The Gini importance in the random forest is defined as Among them, represents the average change in node splitting impurity of the j-th variable in the i-th tree of the random forest, represents the average change in node splitting impurity of the m-th variable in the j-th tree of the random forest, where n is the number of classification trees in the random forest; M is the number of times the variable T j appears in the i-th tree; S2.3: Screen each feature variable based on the calculated importance, and use the time series data with all important indicator scores greater than the preset threshold as the feature library.

4. The method for predicting sintering conditions based on a Bayesian neural network according to claim 1, wherein S3 includes the following processes: S3.1: Denote a set of time series data as X = {x1, x2, x3,..., x 10}, where each x i ∈ R d is represented as a time series of length d; S3.2: For each iteration \(e = 1, 2,\cdots, E\), calculate the cluster center matrix \(V\) of the current iteration according to the membership degree matrix \(U\) of the previous iteration (e-1) where the cluster centers are calculated according to the following formula (e) ​ Among them, is the membership degree corresponding to the data points of the previous iteration, x i is the time series data of the current iteration, m is the fuzzy coefficient, C is the number of clusters, and n is the number of samples; S3.3: Calculate the distance matrix D of the data point x i to the center using the Euclidean distance, where the Euclidean distance is calculated as follows: (e) ​ Among them, is the clustering center corresponding to the data at the current moment, is the data point x i to the center distance; S3.4: According to the distance matrix D calculated in step S3.3 (e) Update the membership matrix U at the current moment (e) , where the membership degree is calculated according to the following formula Among them, is the Euclidean distance of the data point at the current moment, is the membership degree of the data point to the j-th cluster center; S3.5: Compare the membership matrix U at the current moment (e) with the membership matrix U at the previous moment (e-1) If the maximum iteration number E or the minimum preset error is reached, stop the iteration; otherwise, repeat steps S3.2, S3.3, and S3.4 for iteration until the iteration stop condition is satisfied, and output the clustering center V and the membership matrix U.

5. The method for predicting sintering conditions based on a Bayesian neural network according to claim 1, wherein S4 includes the following processes: S4.1: Design the structure of a Bayesian neural network The Bayesian neural network includes two hidden layers, an input layer and an output layer; neurons on each layer are set; S4.2: Build the overall model architecture The results obtained from different clustering are divided into training sets and test sets, the training set is input into the Bayesian neural network, and the weights of the Bayesian neural network are updated using the gradient descent method until the set iteration round is reached to obtain a trained Bayesian neural network; S4.3: Train multiple Bayesian neural networks respectively to obtain different weight parameters; assume the current data is X, and the membership degrees of this data for different clusters are u1, u2, …, u n , place this data in different Bayesian neural networks for calculation, and obtain different results c1, c2, …, c n , perform weighted calculation according to the obtained membership degrees, and the obtained prediction result is Among them, C is the predicted sintering condition result, u i is the membership degree, i = 1, 2,..., n, where n is the number of Bayesian neural networks set, and c i is the prediction result of the i-th Bayesian neural network.

6. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that The processor executes the computer program to implement the steps of the method for predicting sintering conditions according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, A computer program is stored, and when the program is executed by a processor, the steps of the method for predicting sintering conditions according to any one of claims 1 to 5 are implemented.

8. A computer program product, characterized in that, The method comprises a computer program or an instruction, which, when executed by a processor, implements the steps of the method for predicting sintering conditions according to any one of claims 1 to 5.

Images