CN-121999901-A - Method and system for predicting FeO content of sinter based on multi-source data fusion

CN121999901ACN 121999901 ACN121999901 ACN 121999901ACN-121999901-A

Abstract

The invention discloses a method and a system for predicting FeO content of a sintering ore based on multi-source data fusion, which belong to the technical field of sintering process control and comprise the following steps of S1, multi-source data acquisition and feature extraction, S2, image feature prediction, S3, temperature and process data fusion prediction and S4, and self-adaptive weighted fusion. The invention realizes real-time and high-precision prediction of the FeO content of the sintering ore by processing the characteristics of the tail section, the temperature distribution of the tail section and the production process data in parallel, utilizing the strong time sequence characteristic extraction capability of the TCN-BiLSTM network and by self-adaptive weighted average fusion of the two-path prediction results.

Inventors

TANG YALING
Qiao long
ZHANG XUEFENG
WU HAOHAO
LI YONGLIN

Assignees

安徽工业大学

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (9)

1. The method for predicting the FeO content of the sinter based on the multi-source data fusion is characterized by comprising the following steps of: s1, multi-source data acquisition and feature extraction Three heterogeneous data of the sintering process are acquired in parallel, wherein the three heterogeneous data comprise a cross section image of the sintering machine tail is acquired through an industrial vision system, and pre-defined image characteristics are extracted from the cross section image; s2 image feature prediction In the first prediction path, the image features extracted in the step S1 are used as a time sequence, and are input into a first TCN_ BiLSTM hybrid neural network model to perform time sequence feature learning and regression prediction, so that a first FeO content prediction result is obtained; s3, fusion prediction of temperature and process data In a second prediction path, carrying out data preprocessing and feature selection on the temperature distribution data and the production process data obtained in the step S1, fusing the selected features, and inputting the fused feature sequence into a second TCN_ BiLSTM hybrid neural network model to obtain a second FeO content prediction result; S4, self-adaptive weighted fusion Based on the real-time prediction performance of the first prediction path and the second prediction path, the self-adaptive weights of the first prediction path and the second prediction path are dynamically calculated, the first FeO content prediction result and the second FeO content prediction result are weighted and averaged, and the final sinter FeO content prediction value is output.
2. The agglomerate FeO content prediction method based on multi-source data fusion according to claim 1 is characterized in that in the step S1, the extracted predefined image features comprise morphological features, statistical features and distribution features, the morphological features at least comprise porosity and red fire layer duty ratio obtained through image segmentation and morphological operation, the statistical features at least comprise gray level average brightness and average gradient of the whole section area, the distribution features at least comprise vertical gravity center representing the combustion center position and material layer transverse uniformity representing combustion uniformity, the calculation of the porosity is to binarize a section image through global threshold segmentation or adaptive threshold segmentation and calculate the proportion of the pixels of a hole area to the pixels of the total section area, and the red fire layer duty ratio is obtained through mask processing in a HSV space definition set color range, extracting red fire pixels and calculating the area duty ratio.
3. The method for predicting the FeO content of the agglomerate based on the multi-source data fusion according to claim 2, wherein in the step S1, the temperature distribution data is derived from a section temperature matrix acquired by a thermal infrared imager, and thermal state characteristics including a section temperature extremum, a mean value and a ratio of each temperature interval are extracted through statistical analysis.
4. A method for predicting the FeO content of a sintered ore based on multi-source data fusion according to claim 3, wherein in the step S1, the production process data includes a raw material proportioning parameter, a blending ingredient parameter, a process operation parameter and a state progress parameter.
5. The method for predicting the FeO content of the sintered ore based on the multi-source data fusion according to claim 1, wherein in the step S3, the data preprocessing includes missing value filling, outlier processing and data normalization.
6. The method for predicting the FeO content of the agglomerate based on the multi-source data fusion according to claim 4, wherein the first TCN_ BiLSTM mixed neural network model and the second TCN_ BiLSTM mixed neural network model are both obtained based on TCN_ BiLSTM network training, the TCN_ BiLSTM network is a series mixed model and comprises a time domain convolution network module and a bidirectional long-short-time memory network module, wherein the time domain convolution network module is used for extracting local features and multi-scale dependency relations of an input time sequence, and the bidirectional long-time memory network module is used for capturing bidirectional long-time sequence dependency relations in the feature sequence extracted by the time domain convolution network module and outputting a final prediction result.
7. The method for predicting the FeO content of the agglomerate based on multi-source data fusion is characterized in that the time domain convolution network module comprises a causal expansion convolution layer, a residual block and a weight normalization layer, the time domain convolution network module takes the residual block as a core, the causal expansion convolution layer is cascaded to greatly expand the receptive field by using an exponentially-increased expansion factor on the premise of guaranteeing the causality of time sequences, meanwhile, the weight normalization layer is integrated behind each causal expansion convolution layer to accelerate model convergence and improve training stability, finally, fusion of identity mapping and feature extraction results is achieved through jump connection inside the residual block, deep network gradient disappearance is effectively prevented, and accordingly refined multi-scale local features are transmitted to the bidirectional long-short-term memory network module to complete bidirectional long-term dependency modeling and prediction of the FeO content.
8. The method for predicting the FeO content of the sintered ore based on the multi-source data fusion according to claim 1 or 7, wherein in the step S4, the adaptive weight is calculated as follows: S41, respectively calculating the predicted mean square error of the first predicted path and the second predicted path in the last time window And ; S42, self-adaptive weight And Inversely proportional to the mean square error of the corresponding path, calculated according to the following formula: ; ; ; Wherein, the As a result of the first FeO content prediction, And the second FeO content prediction result.
9. A system for predicting the FeO content of a sintered ore based on multi-source data fusion, which is characterized by being applied to the method as set forth in any one of claims 1 to 8, comprising: The data acquisition and feature extraction module is used for acquiring three heterogeneous data of the sintering process in parallel, and comprises the steps of acquiring a cross section image of the sintering machine tail through an industrial vision system, extracting predefined image features from the cross section image, acquiring temperature distribution data of the cross section of the sintering machine tail through a thermal infrared imager, and acquiring real-time production process data through a factory information system; the first prediction module is used for taking the image features extracted in the step S1 as a time sequence in a first prediction path, inputting the time sequence features into a first TCN_ BiLSTM hybrid neural network model for time sequence feature learning and regression prediction, and obtaining a first FeO content prediction result; The second prediction module is used for carrying out data preprocessing and feature selection on the temperature distribution data and the production process data obtained in the step S1 in a second prediction path, fusing the selected features, and inputting the fused feature sequence into a second TCN_ BiLSTM mixed neural network model to obtain a second FeO content prediction result; The weighted fusion module is used for dynamically calculating the self-adaptive weights of the first prediction path and the second prediction path based on the real-time prediction performance of the first prediction path and the second prediction path, carrying out weighted average on the first FeO content prediction result and the second FeO content prediction result, and outputting the final sinter FeO content prediction value.

Description

Method and system for predicting FeO content of sinter based on multi-source data fusion Technical Field The invention relates to the technical field of sintering process control, in particular to a method and a system for predicting FeO content of a sintered ore based on multi-source data fusion. Background FeO content in the sinter is one of key indexes for measuring the quality of the sinter, and has important influence on the stability and efficiency of the subsequent blast furnace smelting process. At present, the detection of FeO content mainly depends on manual sampling and chemical analysis, and has the problems of large hysteresis, incapability of realizing real-time control and the like. Traditional prediction methods based on physical models or statistics are difficult to accurately capture complex nonlinear dynamic characteristics in the sintering process. In recent years, deep learning has advanced in industrial prediction, but it is often difficult for a single model to fully utilize multi-source heterogeneous data (such as production process parameters, image features, temperature distribution, etc.) in a sintering process. Therefore, how to effectively integrate the multi-source data and realize high-precision real-time prediction of the FeO of the sinter by using a deep learning model with time sequence processing capability is a technical problem to be solved. Therefore, the invention provides a method and a system for predicting the FeO content of a sinter based on multi-source data fusion. Disclosure of Invention The technical problems to be solved by the invention are as follows: how to solve the problems of large hysteresis, insufficient utilization of multi-source heterogeneous data, difficult accurate capture of complex dynamic characteristics and the like in the existing sinter FeO content prediction method, the method provides a sinter FeO content prediction method based on multi-source data fusion. The invention solves the technical problems through the following technical proposal, and the invention comprises the following steps: s1, multi-source data acquisition and feature extraction Three heterogeneous data of the sintering process are acquired in parallel, wherein the three heterogeneous data comprise a cross section image of the sintering machine tail is acquired through an industrial vision system, and pre-defined image characteristics are extracted from the cross section image; s2 image feature prediction In the first prediction path, the image features extracted in the step S1 are used as a time sequence, and are input into a first TCN_ BiLSTM hybrid neural network model to perform time sequence feature learning and regression prediction, so that a first FeO content prediction result is obtained; s3, fusion prediction of temperature and process data In a second prediction path, carrying out data preprocessing and feature selection on the temperature distribution data and the production process data obtained in the step S1, fusing the selected features, and inputting the fused feature sequence into a second TCN_ BiLSTM hybrid neural network model to obtain a second FeO content prediction result; S4, self-adaptive weighted fusion Based on the real-time prediction performance of the first prediction path and the second prediction path, the self-adaptive weights of the first prediction path and the second prediction path are dynamically calculated, the first FeO content prediction result and the second FeO content prediction result are weighted and averaged, and the final sinter FeO content prediction value is output. Further, in the step S1, the extracted predefined image features comprise morphological features, statistical features and distribution features, wherein the morphological features at least comprise porosity and red fire layer duty ratio obtained through image segmentation and morphological operation, the statistical features at least comprise gray level average brightness and average gradient of the whole section area, the distribution features at least comprise vertical gravity centers representing the combustion center position and material layer transverse uniformity representing combustion uniformity, the calculation of the porosity is to binarize a section image through global threshold segmentation or adaptive threshold segmentation and calculate the proportion of the pixels of the hole area to the pixels of the section total area, and the red fire layer duty ratio is obtained through mask processing in HSV space definition setting color range, extracting red fire pixels and calculating the area duty ratio. Further, in the step S1, the temperature distribution data is derived from a section temperature matrix acquired by the thermal infrared imager, and thermal state characteristics including a section temperature extremum, a mean value and a ratio of each temperature interval are extracted through statistical analysis. Further, in the step S1, the produc