Deep learning network-based belt type roasting machine pellet compressive strength distribution prediction model and system thereof

Abstract

The invention discloses a belt type roasting machine pellet compressive strength distribution prediction model based on a deep learning network and a system thereof. According to the model, firstly, process variables in the continuous production process of a roasting machine and corresponding pellet compressive strength data are collected, through feature selection, the process variables most related to pellet compressive strength distribution characteristics are obtained to form a feature matrix, and then the feature matrix serves as an input quantity; and constructing a pellet compressive strength distribution prediction model based on the deep learning network, and outputting prediction probability distribution of pellet compressive strength distribution. A prediction system constructed based on the model can achieve accurate modeling and prediction of pellet compressive strength distribution, the learning ability and prediction precision of overall strength distribution are remarkably improved, the model can achieve accurate prediction on the minute level, the technical problem that prediction results are lagged in the prior art is effectively solved, and it is tested that the prediction result is more accurate. The error of the prediction result of the system is only 0.053, and the industrial requirements can be completely met.

Description

Technical Field

[0001] The present invention relates to a prediction model for the compressive strength distribution of pellets, and in particular to a prediction model and system for the compressive strength distribution of pellets on a traveling grate roaster based on a deep learning network, belonging to the field of iron ore pellet production. Background Art

[0002] The iron and steel industry is a basic industry of the national economy and one of the important indicators to measure a country's national strength. China's iron and steel industry still adopts the "blast furnace-converter" long-process production technology. Iron ore pellets are a major raw material for steel production and are widely used in long-process blast furnace ironmaking and short-process direct reduction ironmaking, with strong vitality and broad development prospects. Among them, the large-scale traveling grate roasting process is the main development trend of the pellet production process in recent years. It has high heat utilization efficiency, large single-machine production capacity, and strong raw material adaptability, meeting the requirements of large-scale and intensive production in the iron and steel industry. Pellet roasting is a key process affecting the efficiency of blast furnace smelting and product quality. The quality of pellet ore is usually measured by indicators such as compressive strength. However, during the traveling grate roasting process, due to internal mass transfer and heat transfer and the mutual coupling of different process sections, the quality distribution of pellets is uneven. Moreover, the quality inspection of pellet ore mainly relies on off-line chemical analysis, with low sampling frequency and long cycle, unable to provide real-time feedback for the production process, resulting in lag in process control and easy fluctuation of product quality. During the pellet roasting process, due to the complex on-site production environment, there is a common problem of multi-sampling rates. Some key quality indicators can only be obtained through off-line chemical analysis, with low sampling frequency and long cycle, unable to achieve real-time monitoring. This makes it difficult to timely grasp the relationship between process parameters and the quality distribution of pellet ore, and thus unable to effectively predict the quality distribution of pellet ore. Summary of the Invention

[0003] Aiming at the problems existing in the prior art, the first object of the present invention is to provide a prediction model for the compressive strength distribution of pellets on a traveling grate roaster based on a deep learning network. The prediction model collects the process variables in the roasting production at the minute level and their corresponding pellet compressive strength distribution data. Through feature selection, multiple key process variables most relevant to the pellet compressive strength distribution are screened out from the process parameters, and then input into a prediction model constructed by including a convolutional neural network and a mixture density network for training, so as to effectively predict the pellet compressive strength distribution.

[0004] The second object of the present invention is to provide a prediction system for the compressive strength distribution of pellets on a traveling grate based on a deep learning network. Based on the excellent response rate and prediction accuracy of the above prediction model, the prediction system constructed by it can achieve accurate modeling and prediction of the compressive strength distribution of pellets on the traveling grate, significantly improving the learning ability and prediction accuracy of the overall strength distribution, providing scientific and reliable technical support for optimizing the pellet production process and improving product quality, and at the same time being able to achieve accurate prediction at the minute level, effectively solving the technical problem of lagging prediction results in the prior art

[0005] To achieve the above technical object, the present invention provides a prediction model for the compressive strength distribution of pellets on a traveling grate based on a deep learning network, including:

[0006] Step S1: Obtain the process variables and their corresponding pellet compressive strength data during the continuous production process of the traveling grate, and perform normalization processing on the obtained data;

[0007] Step S2: Use the Gaussian mixture model to calculate the compressive strength distribution, use the peaks method to classify the calculated distribution, use the Spearman correlation coefficient to evaluate the features of the normalized data and the classification results, and combine the traveling grate pelletizing process to obtain the process variables most relevant to the pellet compressive strength distribution characteristics to form a feature matrix;

[0008] Step S3: Use the feature matrix as the input quantity to construct a prediction model for the compressive strength distribution of pellets based on a deep learning network, and output the probability density function of the compressive strength distribution of pellets;

[0009] The deep learning network includes a convolutional neural network and a mixture density network.

[0010] As a preferred solution, the process variables include the speed of the drying exhaust fan, the speed of the main induced draft fan, the speed of the regenerative air fan, the temperature of each air box, the temperature at the inlet of the regenerative air fan, the temperature of the cold section hood, the pressure of each air box, the pressure of the drying section hood, the material thickness, and the machine speed.

[0011] As a preferred solution, the process of the normalization processing is:

[0012] Equation 1:

[0013] In Equation 1: x' represents the value of a single data, min is the minimum value of the data in the column, and max is the maximum value of the data in the column.

[0014] As a preferred solution, the process of obtaining the feature sequence in the feature matrix is:

[0015] Step S2-1: Perform Gaussian mixture model classification on the normalized compressive strength data to obtain unimodal distribution and multimodal distribution labels;

[0016] Step S2-2: Perform Spearman correlation coefficient analysis on the classification results obtained after step S2-1 and the normalized process variables to obtain a feature sequence arranged by correlation;

[0017] Step S2-3: Perform Spearman correlation coefficient analysis again on the characteristic sequence obtained after processing in step S2-2, combine the process characteristics of belt roasting pellets, remove redundant characteristic variables, and obtain the characteristic sequence most relevant to the compressive strength distribution characteristics.

[0018] As a preferred solution, the process of calculating the compressive strength distribution is: performing the expectation step and the maximization step on the normalized compressive strength data in sequence, and repeating the above steps until the parameter model converges, that is, the calculation process is:

[0019] Formula 2:

[0020] Formula 3: Update mean

[0021] Equation 4: Update variance

[0022] Formula 5: Update weight

[0023] Formula 6:

[0024] The process of using the peaks method to classify the calculated distribution is:

[0025] Formula 7:

[0026] Formula 8:

[0027] The processing process of the Spearman correlation coefficient is:

[0028] Formula 9: d i =R(X i )-R(Y i );

[0029] Formula 10:

[0030] Formula 11: F Spearman =F(x1,x2,...,x n ,Y)=[ρ1,ρ2,...,ρ g ];

[0031] In formulas 2 to 11: i is the weight of the i-th Gaussian component, μ i is the mean of the i-th component, σ i is the standard deviation of the i-th component, is the probability density function of the normal distribution, g j is the true probability density of the sample, Y is the peak classification of the pellet compressive strength corresponding to the process variable, γ ij is the sample y j The posterior probability of belonging to the i-th Gaussian component. d i is the rank difference of the i-th data point, n is the total number of data points, R(X i ) is X i The level of R(Y i ) is Y i The level, ρ g represents the correlation coefficient of the g-th process parameter, F Spearman is the set of correlation coefficients of all variable parameters.

[0032] As a preferred solution, the process of obtaining the characteristic matrix composed of the most relevant process variables is as follows:

[0033] Formula 12: W = F Spearman +F knowledge ;

[0034] Formula 13: x=[X1,X2,...,X N ];

[0035] In Equations 12 and 13: W is the characteristic matrix of the comprehensive selection of the Spearman correlation coefficient combined with the belt roasting pellet process, F knowledge It is the knowledge of belt roasting pellet process, X N is the Nth variable finally selected.

[0036] As a preferred solution, the convolutional neural network in the deep learning network is used to extract spatial distribution features from input features, and the mixture density network is used to establish Gaussian mixture model parameters for the distribution of pellet compressive strength.

[0037] As a preferred solution, the feature extraction process of the convolutional neural network is:

[0038] Step S3-1, the convolutional neural network extracts spatial features through multi-layer convolution operations. The output of each layer of convolution operation is expressed by the following formula:

[0039] Formula 14: H (l) =ReLU(W (l) *H (l-1) +b (l) );

[0040] Step S3-2: After each convolution operation, pooling is performed, and its calculation formula is:

[0041] Equation 15:

[0042] Step S3-3: After multiple layers of convolution and pooling processing, the output features are sent to the fully connected layer for feature mapping, and its calculation formula is:

[0043] Equation 16: z = W (fc) H (L) + b (fc) ;

[0044] The process of the mixture density network for establishing the Gaussian mixture model parameters of the pellet compressive strength distribution is as follows:

[0045] Step S3-4: The goal of improving the mixture density network is to learn the parameters of the mixture Gaussian distribution, and the model is trained by maximizing the Bhattacharyya loss function. The process is as follows:

[0046] Equation 17:

[0047] Equation 18:

[0048] Equation 19:

[0049] In Equations 17-19: π i is the weight of the i-th Gaussian component predicted, μ i is the mean of the i-th component predicted, σ i is the standard deviation of the i-th component predicted, is the probability density function of the normal distribution, p j is the probability density of the sample prediction.

[0050] As a preferred solution, the prediction model also introduces the Area Difference to compare the difference between the predicted density and the true density. The process is as follows:

[0051] Equation 20: Area Difference = ∫|PDF pred (x) - PDF true (x)|dx;

[0052] In Equation 20: PDF pred (x) is the predicted probability density, PDF true (x) is the true probability density, x is the sample point on the support range.

[0053] The present invention also provides a prediction system for the compressive strength distribution of pellets on a traveling grate roaster based on a deep learning network, which consists of a data acquisition device, a memory containing the prediction model described in any one of the above, and a processor; the data acquisition device is used to collect sintering production process data; the memory is used to store sintering production process data and the prediction model; the processor is used to read the sintering process production data and execute the prediction model.

[0054] Compared with the prior art, the beneficial technical effects of the technical solution of the present invention are as follows:

[0055] 1) The prediction model provided by the present invention collects the process variables in the roasting production at the minute level and the corresponding pellet compressive strength distribution data at the hourly level. Through the feature selection method, multiple key process variables most relevant to the pellet compressive strength distribution are screened out from the process parameters, and then input into the prediction model constructed by including a convolutional neural network and a mixture density network for training, so as to realize the effective prediction of the pellet compressive strength distribution.

[0056] 2) In the technical solution provided by the present invention, the prediction system constructed based on this model can realize the accurate modeling and prediction of the pellet compressive strength distribution, significantly improve the learning ability and prediction accuracy of the overall strength distribution, provide scientific and reliable technical support for optimizing the pellet production process and improving product quality, and at the same time can realize accurate prediction at the minute level, effectively solving the technical problem of lagging prediction results in the prior art. After testing, the error of the prediction result of this system is only 0.053, which can fully meet the industrial requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the prediction model for the compressive strength distribution of pellets on a traveling grate roaster based on a deep learning network provided in Embodiment 1 of the present invention;

[0058] Figure 2 It is a network structure diagram of the prediction model for the compressive strength distribution of pellets on a traveling grate roaster based on a deep learning network provided in Embodiment 1 of the present invention;

[0059] Figure 3 It is a schematic diagram of the comparison of the prediction effects of the prediction model for the compressive strength distribution of pellets on a traveling grate roaster based on a deep learning network provided in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0061] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0062] Example 1

[0063] This example provides a prediction model for the compressive strength distribution of pellets on a traveling grate pelletizer based on a deep learning network, and its construction process is as follows:

[0064] 1. Obtain multiple groups of historical data of the normal operation of the production process of the pellet traveling grate pelletizer as the data set (X, Y), where X represents the process variables and Y represents the pellet compressive strength;

[0065] The process variables include the speed of the drying exhaust fan, the speed of the main induced draft fan, the speed of the regenerative air fan, the temperature of each air box, the temperature at the inlet of the regenerative air fan, the temperature of the cold section hood, the pressure of each air box, the pressure of the drying section hood, the material thickness, and the machine speed.

[0066] 2. Perform normalization processing on the data set. The specific process is as follows:

[0067]

[0068] Among them, x′ represents the value of a single data, min is the minimum value of the column where the data is located, and max is the maximum value of the column where the data is located;

[0069] 3. Use the Gaussian mixture model to calculate the compressive strength distribution, use the peaks method to classify the calculated distribution, use the Spearman correlation coefficient to evaluate the features of the data after normalization processing and the classification results, and combine the traveling grate pelletizing process to obtain the feature matrix composed of the process variables most relevant to the pellet compressive strength distribution characteristics:

[0070] 3-1. The process of obtaining the feature sequence in the feature matrix is as follows:

[0071]

[0072] Update the mean

[0073] Update the variance

[0074] Update the weight

[0075]

[0076] The process of using the peaks method to classify the calculated distribution is as follows:

[0077] Formula 7:

[0078] Formula 8:

[0079] The processing procedure of the Spearman correlation coefficient is as follows:

[0080] Formula 9: d i = R(X i ) - R(Y i );

[0081] Formula 10:

[0082] Formula 11: F Spearman = F(x1, x2,..., x n , Y) = [ρ1, ρ2,..., ρ g ;

[0083] In Formulas 2 to 11: π i is the weight of the i-th Gaussian component, μ i is the mean of the i-th component, σ i is the standard deviation of the i-th component, is the probability density function of the normal distribution, g j is the true probability density of the sample, Y is the classification of the peak pellet compressive strength corresponding to the process variable, γ ij is the posterior probability that the sample y j belongs to the i-th Gaussian component. d i is the rank difference of the i-th data point, n is the total number of data points, R(X i ) is the rank of X i , R(Y i ) is the rank of Y i , ρ g represents the correlation coefficient of the g-th process parameter, F Spearman is the set of correlation coefficients of all variable parameters.

[0084] 3-2. The process of obtaining the characteristic matrix composed of the most relevant process variables is as follows:

[0085] Formula 12: W = F Spearman + F knowledge ;

[0086] Formula 13: x = [X1, X2,..., X N ;

[0087] In Formulas 12 and 13: W is the characteristic matrix comprehensively selected by combining the Spearman correlation coefficient with the grate-kiln pelletizing process, F knowledge is the knowledge of the grate-kiln pelletizing process, XN is the Nth variable finally selected.

[0088] 4. Using the most relevant process variables to form a feature matrix as the input, a prediction model for the pellet compressive strength distribution based on a deep learning network is constructed. The specific process is as follows:

[0089] The prediction model includes a Convolutional Neural Network (CNN) and a Mixture Density Network (MDN). The CNN part extracts features through convolutional layers and pooling layers, and maps the extracted features to a fixed-length vector through a fully connected layer. These features are passed to the MDN part to fit the Gaussian mixture distribution parameters of the compressive strength distribution, including the mean, standard deviation, and mixture weights;

[0090] 4-1. CNN Feature Extraction

[0091] The input data is x = [X1, X2,..., X N , where N is the number of samples. The CNN model extracts spatial features through multiple convolutional operations. The output of each convolutional operation is represented by the following formula:

[0092] H (l) = ReLU(W (l) * H (l-1) + b (l) );

[0093] After each convolutional operation, pooling (max pooling or average pooling) is performed, and its calculation formula is:

[0094]

[0095] After multiple convolutional and pooling processes, the output features are fed into a fully connected layer for feature mapping and further processing:

[0096] z = W (fc) H (L) + b (fc) ;

[0097] where W (l) is the convolutional kernel of the l-th layer, * represents the convolutional operation, b (l) is the bias term, H (l-1) is the output of the (l-1)-th layer, ReLU is the activation function; H (L) is the final feature map after all convolutional and pooling layers, W (fc) is the weight matrix of the fully connected layer, and b (fc) is the bias.

[0098] 4-2. Improved MDN Prediction Network

[0099] The MDN receives the feature vectors extracted by the CNN in part and outputs the Gaussian mixture distribution parameters describing the compressive strength distribution, including the mean (μ), standard deviation (σ), and weight (π).

[0100] The goal of the MDN is to learn the parameters of the mixture Gaussian distribution and train the model by maximizing the Bhattacharyya loss function. Assuming we have M Gaussian components, the formula output by the MDN is as follows:

[0101]

[0102] where, π i is the weight of the i-th Gaussian component, μ i is the mean of the i-th component, σ i is the standard deviation of the i-th component, is the probability density function of the normal distribution, defined as:

[0103]

[0104] The output of the model is μ = [μ1, μ2,..., μ M , σ = [σ1, σ2,..., σ M , π = [π1, π2,..., π M ;

[0105] During the training process, the model parameters are optimized through the Bhattacharyya loss function, and the formula of the loss function is:

[0106]

[0107] 4-3. Design of Evaluation Index for Strength Distribution Prediction

[0108] As a preferred solution, the prediction model also designs an evaluation index for the prediction of the pellet compressive strength distribution: Area Difference to compare the difference between the predicted density and the true density:

[0109] Area Difference = ∫|PDF pred (x) - PDF true (x)|dx;

[0110] Based on the above prediction model, this embodiment also constructs a prediction system for the pellet compressive strength distribution of the traveling grate pelletizer based on a deep learning network, which consists of a data acquisition device, a memory containing the above prediction model, and a processor; the data acquisition device is used to collect the sintering production process data; the memory is used to store the sintering production process data and the prediction model; the processor is used to read the sintering process production data and execute the prediction model.

[0111] In order to further verify the response rate and accuracy of the system, the present invention also used the above system to conduct the following tests:

[0112] In this embodiment, a pellet grate-kiln of a certain iron and steel plant was used as the research object for verification. A feature selection model, a Gaussian mixture model, and a neural network model constructed by CNN and MDN were established using on-site data to predict the probability density of the pellet compressive strength distribution. Figure 3 It is a schematic diagram comparing the prediction effects of Model 1 adopted in the present invention with Model 2 and Model 3 adopted in the prior art. From Figure 3 it can be seen that the technical solution provided by the present invention can effectively predict the pellet compressive strength distribution online with high precision, and its error is only 0.053. However, affected by the loss function, the traditional maximum likelihood estimation loss and Wasserstein loss cannot capture the complete true distribution information, making it difficult to achieve online precise prediction of the pellet compressive strength distribution.

Claims

1. A prediction model for the compressive strength distribution of pellets in a grate-kiln based on a deep learning network, characterized in that, Including: Step S1: Obtain the process variables and their corresponding pellet compressive strength data during the continuous production process of the traveling grate pelletizer, and normalize the obtained data. Step S2: Use the Gaussian mixture model to calculate the compressive strength distribution, use the peaks method to classify the calculated distribution, use the Spearman correlation coefficient to evaluate the characteristics of the normalized data and the classification results, and combine the traveling grate pelletizing process to obtain the characteristic matrix composed of process variables most relevant to the pellet compressive strength distribution characteristics. Step S3: Use the characteristic matrix as the input quantity to construct a prediction model of the pellet compressive strength distribution based on a deep learning network, and output the probability density function of the pellet compressive strength distribution. The deep learning network includes a convolutional neural network and an improved mixture density network.

2. The prediction model for the compressive strength distribution of pellets in a grate-kiln based on a deep learning network according to claim 1, characterized in that: The process variables include the speed of the drying exhaust fan, the speed of the main induced draft fan, the speed of the regenerative air fan, the temperature of each air box, the temperature at the inlet of the regenerative air fan, the temperature of the cold section hood, the pressure of each air box, the pressure of the drying section hood, the material thickness, and the machine speed.

3. The prediction model for the pellet compressive strength distribution of a grate-kiln based on a deep learning network according to claim 1, wherein: The process of the normalization process is as follows: Formula 1: In Equation 1: x' represents the value of a single data, min is the minimum value of the data in the column, and max is the maximum value of the data in the column.

4. The prediction model for the pellet compressive strength distribution of a grate-kiln based on a deep learning network according to claim 1, wherein: The process of obtaining the feature sequence in the characteristic matrix is as follows: Step S2-1: Classify the normalized compressive strength data using the Gaussian mixture model to obtain the labels of the unimodal distribution and the multimodal distribution. Step S2-2: Perform Spearman correlation coefficient analysis on the classification results obtained after the processing in Step S2-1 and the normalized process variables to obtain the feature sequence arranged according to the correlation size. Step S2-3: Perform Spearman correlation coefficient analysis on the feature sequence obtained after the processing in Step S2-2 again, and combine the characteristics of the traveling grate pelletizing process to remove redundant feature variables and obtain the feature sequence most relevant to the compressive strength distribution characteristics.

5. A prediction model for the compressive strength distribution of pellets in a grate-kiln based on a deep learning network according to claim 1 or 4, characterized in that: The process of calculating the compressive strength distribution is as follows: Perform the expectation step and the maximization step on the normalized compressive strength data in sequence, and repeat the above steps until the parameter model converges, that is, obtained. The calculation process is as follows: Equation 2: Equation 3: Updated mean Equation 4: Updated variance Equation 5: Updated weight Equation 6: The process of classifying the calculated distribution using the peaks method is as follows: Formula 7: Formula 8: The processing procedure of the Spearman correlation coefficient is as follows: Formula 9: d i = R(X i ) - R(Y i ); Formula 10: Equation 11: F Spearman = F(x1, x2,..., x n , Y) = [ρ1, ρ2,..., ρ g ; In Formulas 2 to 11: π i is the weight of the i-th Gaussian component, μ i is the mean of the i-th component, σ i is the standard deviation of the i-th component, is the probability density function of the normal distribution, g j is the true probability density of the sample, Y is the classification of the peak pellet compressive strength corresponding to the process variable, γ ij is the posterior probability that the sample y j belongs to the i-th Gaussian component. d i is the rank difference of the i-th data point, n is the total number of data points, R(X i ) is the rank of X i , R(Y i ) is the rank of Y i , ρ g represents the correlation coefficient of the g-th process parameter, F Spearman is the set of correlation coefficients of all variable parameters.

6. The prediction model for the pellet compressive strength distribution of a traveling grate pelletizer based on a deep learning network according to claim 5, characterized in that: The process of obtaining the most relevant process variable composition feature matrix is as follows: Equation 12: W = F Spearman + F knowledge ; Equation 13: x = [X1, X2,..., X N ; In Equation 12 and Equation 13: W is the feature matrix comprehensively selected by combining the Spearman correlation coefficient with the grate-kiln pelletizing process, F knowledge is the knowledge of the grate-kiln pelletizing process, and X N is the Nth variable finally selected.

7. The prediction model for the pellet compressive strength distribution of a traveling grate pelletizer based on a deep learning network according to claim 1, characterized in that: The convolutional neural network in the deep learning network is used to extract the spatial distribution characteristics from the input features, and the mixture density network is used to establish the Gaussian mixture model parameters of the pellet compressive strength distribution.

8. A prediction model for the compressive strength distribution of pellets in a grate-kiln based on a deep learning network according to claim 7, characterized in that: The feature extraction process of the convolutional neural network is as follows: Step S3-1: The convolutional neural network extracts spatial features through multiple convolutional operations. The output of each convolutional operation is represented by the following formula: Equation 14: H (l) = ReLU(W (l) * H (l-1) + b (l) ); Step S3-2: Pooling is performed after each convolutional operation, and its calculation formula is: Formula 15: Step S3-3: After multi-layer convolution and pooling processing, the output features are sent to the fully connected layer for feature mapping, and its calculation formula is: Equation 16: z = W (fc) H (L) + b (fc) ; In formulas 14 to 16: W (l) is the l-th layer convolutional kernel, * represents the convolution operation, b (l) is the bias term, H (l-1) is the output of the (l - 1)-th layer, and ReLU is the activation function; H (L) is the final feature map after passing through all convolutional layers and pooling layers, W (fc) is the weight matrix of the fully connected layer, b (fc) is the bias.

9. The prediction model for the pellet compressive strength distribution of a grate-kiln based on a deep learning network according to claim 1, characterized in that: During the training process of the prediction model, an improved mixture density network is also introduced to establish the Gaussian mixture model parameters of the pellet compressive strength distribution. The process is as follows: Step S3-4: The goal of the improved mixture density network is to learn the parameters of the mixture Gaussian distribution, and the model is trained by maximizing the Bhattacharyya loss function. The process is as follows: Formula 17: Formula 18: Formula 19: In Formulas 17 to 19: π i is the weight of the i-th predicted Gaussian component, μ i is the mean of the i-th predicted component, σ i is the standard deviation of the i-th predicted component, is the probability density function of the normal distribution, p j is the probability density of the sample prediction.

10. A prediction system for the compressive strength distribution of pellets in a traveling grate pelletizer based on a deep learning network, characterized in that: It consists of a data acquisition device, a memory containing the prediction model according to any one of claims 1 to 8, and a processor; the data acquisition device is used to collect sintering production process data; the memory is used to store sintering production process data and the prediction model; the processor is used to read the sintering process production data and execute the prediction model.

Images