A coal quality component detection method and system based on deep learning
Through the coal quality component detection method based on deep learning, sub-infrared sensors and deep learning models are used to predict the components of coal in furnaces in real time, solving the problem of coal quality detection in coal-fired power generation, improving combustion efficiency and reducing carbon emissions.
Patent Information
- Application Number
- CN202410560968.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-08
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-05-08
AI Technical Summary
The existing artificial test methods cannot grasp the coal quality of the coal entering the furnace in real time, which makes it difficult to accurately and scientifically produce coal combustion in coal-fired power generation, resulting in waste of coal-fired power generation and unnecessary carbon dioxide emissions.
Using deep learning-based coal composition detection method, coal absorbance data is obtained through sub-infrared sensors, pre-processing and characteristic polynomial crossover, combining the fully connected network and crossover network, the six major components of the coal entering the furnace are predicted in real time, including heat Value, fixed moisture, total moisture, volatile, ash and sulfur.
Real-time prediction of the components of coal entering the furnace is achieved, helping scientifically mix coal, improve combustion efficiency, and reduce carbon emissions.
Smart Images

Figure CN118425087B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of coal quality component detection, and in particular to a coal quality component detection method and system based on deep learning. Background Art
[0002] At present, coal-fired power generation is the main source of electricity and has attracted much attention. Reducing coal consumption for power generation and reducing carbon dioxide emissions are important goals of coal-fired power plants. In the boiler combustion stage of coal-fired power generation, multiple types of coal are blended for combustion, and a scientific blending ratio can reduce carbon dioxide emissions. However, the quality of the coal entering the furnace varies complexly, and the existing methods that rely on manual testing cannot grasp the quality of the coal entering the furnace in real time, making it difficult to accurately and scientifically blend coal for combustion, resulting in waste of coal for power generation and unnecessary carbon dioxide emissions. Summary of the invention
[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a coal quality composition detection method and system based on deep learning, which solves the problem of difficulty in detecting coal quality composition in the process of coal-fired power generation.
[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:
[0005] A method for detecting coal quality components based on deep learning is provided, which comprises the following steps:
[0006] S1. Obtaining the absorbance data of coal through a sub-infrared sensor;
[0007] S2, preprocessing the absorbance data of coal to obtain original features;
[0008] S3, perform feature polynomial crossover on the original features to obtain second-order crossover features;
[0009] S4, using the second-order cross feature as the input of the fully connected network to obtain the first detection feature;
[0010] S5, using the original feature as the input of the cross network to obtain the second detection feature;
[0011] S6. Merge the first detection feature and the second detection feature, and process the merged result through a fully connected layer to obtain a coal quality component detection result; wherein the coal quality component detection result includes calorific value, fixed moisture, total moisture, volatile matter, ash content and sulfur content.
[0012] Furthermore, the specific method of step S1 is:
[0013] The absorbance data of coal is obtained by using a sub-infrared sensor with six different wavelengths of sub-infrared rays to collect the absorbance value of coal.
[0014] Furthermore, the specific method of step S2 includes the following sub-steps:
[0015] S2-1. Remove outliers from the absorbance data of coal by using the box plot method: take the difference between the upper quartile and the lower quartile of the absorbance data of coal as the interquartile range, and determine the data points that are 1.5 times the interquartile range outside the upper and lower bounds of the box as outliers and remove them;
[0016] S2-2, delete the row data with absorbance value of 0;
[0017] S2-3, delete the row data with missing values;
[0018] S2-4, standardize the remaining data to obtain the original features;
[0019] The absorbance data of coal are in six columns and M rows, and each row includes the absorbance values of coal collected by sub-infrared rays of six different wavelengths.
[0020] Furthermore, the specific method of step S3 is:
[0021] Multiplication is performed within the same row of the original feature, and each absorbance value in the original feature is squared to obtain 21 enhanced features, that is, second-order cross features.
[0022] Further, the deep learning model includes a fully connected network and a cross network, and a merge output module for merging the fully connected network output and the cross network output;
[0023] The fully connected network includes n basic modules connected in sequence, and a first fully connected layer connected to the last basic module, and the output end of the first fully connected layer is the output end of the fully connected network;
[0024] The cross network includes n+1 basic modules, the first n basic modules are connected in sequence, and the output end of the nth basic module is connected to the second fully connected layer; after the output end of the fully connected layer in the first basic module in the cross network is calculated with the output of the second fully connected layer, vector addition is performed with the output of the nth basic module, and the vector addition result is the input of the n+1th basic module; the output of the n+1th basic module is the output of the cross network;
[0025] Each basic module consists of a fully connected layer, a batch normalization layer, and a LeakyReLU activation function that are connected sequentially.
[0026] Furthermore, the first detection feature and the second detection feature are combined by splicing.
[0027] A coal quality component detection system based on deep learning is provided, which includes a sub-infrared sensor, a data preprocessing module, a feature expansion module, a component detection module and a merging output module;
[0028] A sub-infrared sensor is installed above the coal conveyor belt for collecting the absorbance data of the coal entering the furnace through sub-infrared rays;
[0029] The data preprocessing module is used to preprocess the absorbance data of coal to obtain the original features;
[0030] The feature expansion module is used to perform feature polynomial crossover on the original features to obtain second-order crossover features;
[0031] A component detection module, used to use the second-order cross feature as an input of a fully connected network to obtain a first detection feature; and use the original feature as an input of the cross network to obtain a second detection feature;
[0032] The merging output module is used to merge the first detection feature and the second detection feature, and process the merging result through a fully connected layer to obtain a coal quality component detection result.
[0033] Furthermore, the component detection module and the merge output module constitute a deep learning model; wherein the fully connected network includes a first fully connected layer connected to n basic modules in sequence and connected to the last basic module, and the output end of the first fully connected layer is the output end of the fully connected network;
[0034] The cross network includes n+1 basic modules, the first n basic modules are connected in sequence, and the output end of the nth basic module is connected to the second fully connected layer; after the output end of the fully connected layer in the first basic module in the cross network is calculated with the output of the second fully connected layer, vector addition is performed with the output of the nth basic module, and the vector addition result is the input of the n+1th basic module; the output of the n+1th basic module is the output of the cross network;
[0035] Each basic module consists of a fully connected layer, a batch normalization layer, and a LeakyReLU activation function that are connected sequentially.
[0036] Furthermore, the deep learning model is constructed and trained by Keras. The trained deep learning model is saved in .h5 format and deployed on the industrial computer. The industrial computer uses the socket module in python3 for UDP communication to obtain the data to be tested and feedback the test results. During the training process, the loss function of the deep learning model adopts L1 loss, which is expressed as: ,in is the loss value, represents the true label of the i-th sample, represents the detection label of the deep learning model for the i-th sample, N is the total number of training samples; the optimizer in the training process is the Adam optimizer, the learning rate is 0.01, and the batch size is 256.
[0037] The beneficial effects of the present invention are as follows: the present invention can predict the six major component values of the coal entering the furnace in real time, help to scientifically and reasonably blend the coal for combustion, improve combustion efficiency, and reduce carbon emissions. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 is a schematic diagram of the process of this method;
[0039] Figure 2 Schematic diagram of the deep learning model. DETAILED DESCRIPTION
[0040] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0041] like Figure 1 As shown, the coal quality component detection method based on deep learning includes the following steps:
[0042] S1. Obtaining the absorbance data of coal through a sub-infrared sensor;
[0043] S2, preprocessing the absorbance data of coal to obtain original features;
[0044] S3, perform feature polynomial crossover on the original features to obtain second-order crossover features;
[0045] S4, using the second-order cross feature as the input of the fully connected network to obtain the first detection feature;
[0046] S5, using the original feature as the input of the cross network to obtain the second detection feature;
[0047] S6. Merge the first detection feature and the second detection feature, and process the merged result through a fully connected layer to obtain a coal quality component detection result; wherein the coal quality component detection result includes calorific value, fixed moisture, total moisture, volatile matter, ash content and sulfur content.
[0048] The coal quality component detection system based on deep learning includes a sub-infrared sensor, a data preprocessing module, a feature expansion module, a component detection module and a merging output module;
[0049] A sub-infrared sensor is installed above the coal conveyor belt for collecting the absorbance data of the coal entering the furnace through sub-infrared rays;
[0050] The data preprocessing module is used to preprocess the absorbance data of coal to obtain the original features;
[0051] The feature expansion module is used to perform feature polynomial crossover on the original features to obtain second-order crossover features;
[0052] A component detection module, used to use the second-order cross feature as an input of a fully connected network to obtain a first detection feature; and use the original feature as an input of the cross network to obtain a second detection feature;
[0053] The merging output module is used to merge the first detection feature and the second detection feature, and process the merging result through a fully connected layer to obtain a coal quality component detection result.
[0054] In this embodiment, if Figure 2 As shown, the component detection module and the merge output module constitute a deep learning model; wherein the fully connected network includes n basic modules connected in sequence, and a first fully connected layer connected to the last basic module, and the output end of the first fully connected layer is the output end of the fully connected network;
[0055] The cross network includes n+1 basic modules, the first n basic modules are connected in sequence, and the output end of the nth basic module is connected to the second fully connected layer; after the output end of the fully connected layer in the first basic module in the cross network is calculated with the output of the second fully connected layer, vector addition is performed with the output of the nth basic module, and the vector addition result is the input of the n+1th basic module; the output of the n+1th basic module is the output of the cross network;
[0056] Each basic module consists of a fully connected layer, a batch normalization layer, and a LeakyReLU activation function that are connected sequentially.
[0057] It should be noted that the number of basic modules in the fully connected network and the cross network can be the same or different, and the "n" here only refers to the number in general, and does not mean that the number of basic modules in the cross network is always 1 more than the number of basic modules in the fully connected network.
[0058] In the specific implementation process, the specific method for obtaining the absorbance data of coal is: using six different wavelengths of sub-infrared rays through a sub-infrared sensor to collect the absorbance value of the coal to obtain the absorbance data of the coal.
[0059] The specific method for preprocessing the absorbance data of coal includes the following sub-steps:
[0060] S2-1. Remove outliers from the absorbance data of coal by using the box plot method: take the difference between the upper quartile and the lower quartile of the absorbance data of coal as the interquartile range, and determine the data points that are 1.5 times the interquartile range outside the upper and lower bounds of the box as outliers and remove them;
[0061] S2-2, delete the row data with absorbance value of 0;
[0062] S2-3, delete the row data with missing values;
[0063] S2-4, standardize the remaining data to obtain the original features;
[0064] The absorbance data of coal are in six columns and M rows, and each row includes the absorbance values of coal collected by sub-infrared rays of six different wavelengths.
[0065] Multiplication is performed within the same row of the original feature, and each absorbance value in the original feature is squared to obtain 21 enhanced features, that is, second-order cross features.
[0066] The deep learning model is constructed and trained by Keras. During the training process, the loss function of the deep learning model adopts L1 loss, which is expressed as: ,in is the loss value, represents the true label of the i-th sample, represents the detection label of the deep learning model for the i-th sample, and N is the total number of training samples.
[0067] The optimizer in the training process is the Adam optimizer, which includes a momentum term (exponential moving average based on the gradient) and an adaptive learning rate term. The learning rate of each parameter is adjusted in an adaptive manner to make it perform better in terms of stability and convergence speed. The learning rate is set to 0.01. The batch size is set to 256. The overall convergence speed of the deep learning model is fast. After dividing the data set, it can be judged that the deep learning model is in a state of overfitting or underfitting. If the deep learning model has a strong detection ability on the training set, but a weak detection ability on the test set, and there is a large gap between the two, then this is the overfitting of the model. You can reduce the model size, add BN layers, stop training early, etc. to alleviate overfitting; if the deep learning model has a weak detection ability on the training set and a weak detection ability on the test set, then this is the underfitting of the model. You can expand the model size, increase the number of training rounds, etc. to alleviate underfitting. Finally, after trying, the model can reach a convergence state when the epoch is set to 30.
[0068] In the training process of the deep learning model in this embodiment, in order to achieve the best generalization ability of the model and make the model fit the new data better, the data set to be tested with known labels is folded and divided into training set and test set alternately through K-Fold, and the entire data set is divided into K parts, of which K-1 parts are training sets and the remaining 1 part is used as the test set. This process is repeated K times. This process uses the KFold class in the sklearn library to implement cross-validation, and K is 5, that is, the data set is folded and divided into training set and test set alternately using the five-fold cross-validation K-Fold, and the data volume ratio is 4:1.
[0069] In the specific implementation process, this method / system only needs 8ms to complete data processing and detection. Combined with the speed of the coal conveyor belt entering the furnace, the absorbance data of the coal can be sent every 100ms to ensure the goal of real-time detection. After obtaining the corresponding component detection results of different coals, the amount of coal entering the furnace can be controlled by the industrial computer according to the set component ratio, thereby achieving scientific coal blending and combustion.
[0070] In one embodiment of the present invention, the method / system is compared with coal quality detection based on linear regression, coal quality detection based on random forest and coal quality detection based on fully connected network on the same data set, and the comparison results are shown in Table 1.
[0071] Table 1: Comparison results
[0072]
[0073] As can be seen from Table 1, the present invention is superior to the above three schemes in the detection of the six components. The present invention can not only learn the low-order cross features, but also explicitly calculate the high-order cross features, and the detection effect is more accurate.
[0074] In summary, the present invention can predict the six major component values of coal entering the furnace in real time, help to scientifically and reasonably blend coal for combustion, improve combustion efficiency, and reduce carbon emissions.
Claims
1. A coal quality component detection method based on deep learning, characterized in that: The following steps are involved: S1. Obtaining the absorbance data of coal through a sub-infrared sensor; S2, preprocessing the absorbance data of coal to obtain original features; S3, perform feature polynomial crossover on the original features to obtain second-order crossover features; S4, using the second-order cross feature as the input of the fully connected network to obtain the first detection feature; S5, using the original feature as the input of the cross network to obtain the second detection feature; S6. Merge the first detection feature and the second detection feature, and process the merged result through a fully connected layer to obtain a coal quality component detection result; wherein the coal quality component detection result includes calorific value, fixed moisture, total moisture, volatile matter, ash and sulfur; The specific method of step S1 is: The absorbance data of coal is obtained by using a sub-infrared sensor to collect the absorbance value of coal using six different wavelengths of sub-infrared rays; The specific method of step S2 includes the following sub-steps: S2-1. Remove outliers from the absorbance data of coal by using the box plot method: take the difference between the upper quartile and the lower quartile of the absorbance data of coal as the interquartile range, and determine the data points that are 1.5 times the interquartile range outside the upper and lower bounds of the box as outliers and remove them; S2-2, delete the row data with absorbance value of 0; S2-3, delete the row data with missing values; S2-4, standardize the remaining data to obtain the original features; The absorbance data of coal is in six columns and M rows, and each row includes the absorbance values of coal collected by sub-infrared rays of six different wavelengths; The specific method of step S3 is: Multiplication is performed in the same row of the original feature, and each absorbance value in the original feature is squared to obtain 21 enhanced features, that is, the second-order cross features; The fully connected network includes n basic modules connected in sequence, and a first fully connected layer connected to the last basic module, and the output end of the first fully connected layer is the output end of the fully connected network; The cross network includes n+1 basic modules, the first n basic modules are connected in sequence, and the output end of the nth basic module is connected to the second fully connected layer; after the output end of the fully connected layer in the first basic module in the cross network is calculated with the output of the second fully connected layer, vector addition is performed with the output of the nth basic module, and the vector addition result is the input of the n+1th basic module; the output of the n+1th basic module is the output of the cross network; Each basic module consists of a fully connected layer, a batch normalization layer, and a LeakyReLU activation function that are connected sequentially.
2. The coal quality component detection method based on deep learning according to claim 1 is characterized in that: The first detection feature and the second detection feature are combined by splicing.
3. A coal quality component detection system based on deep learning, characterized in that: It includes a sub-infrared sensor, a data preprocessing module, a feature expansion module, a component detection module and a merging output module; A sub-infrared sensor is installed above the coal conveyor belt for collecting the absorbance data of the coal entering the furnace through sub-infrared rays; The data preprocessing module is used to preprocess the absorbance data of coal to obtain the original features; The feature expansion module is used to perform feature polynomial crossover on the original features to obtain second-order crossover features; A component detection module, used to use the second-order cross feature as an input of a fully connected network to obtain a first detection feature; and use the original feature as an input of the cross network to obtain a second detection feature; A merging output module is used to merge the first detection feature and the second detection feature, and process the merging result through a fully connected layer to obtain a coal quality component detection result; The specific method for collecting the absorbance data of the coal entering the furnace is: The absorbance data of coal is obtained by using a sub-infrared sensor to collect the absorbance value of coal using six different wavelengths of sub-infrared rays; The specific method of preprocessing the absorbance data of coal to obtain the original features includes the following sub-steps: S2-1. Remove outliers from the absorbance data of coal by using the box plot method: take the difference between the upper quartile and the lower quartile of the absorbance data of coal as the interquartile range, and determine the data points that are 1.5 times the interquartile range outside the upper and lower bounds of the box as outliers and remove them; S2-2, delete the row data with absorbance value of 0; S2-3, delete the row data with missing values; S2-4, standardize the remaining data to obtain the original features; The absorbance data of coal is in six columns and M rows, and each row includes the absorbance values of coal collected by sub-infrared rays of six different wavelengths; The specific method of performing feature polynomial crossover on the original features to obtain the second-order crossover features is: Multiplication is performed in the same row of the original feature, and each absorbance value in the original feature is squared to obtain 21 enhanced features, that is, the second-order cross features; The fully connected network includes n basic modules connected in sequence, and a first fully connected layer connected to the last basic module, and the output end of the first fully connected layer is the output end of the fully connected network; The cross network includes n+1 basic modules, the first n basic modules are connected in sequence, and the output end of the nth basic module is connected to the second fully connected layer; after the output end of the fully connected layer in the first basic module in the cross network is calculated with the output of the second fully connected layer, vector addition is performed with the output of the nth basic module, and the vector addition result is the input of the n+1th basic module; the output of the n+1th basic module is the output of the cross network; Each basic module consists of a fully connected layer, a batch normalization layer, and a LeakyReLU activation function that are connected sequentially.
4. The coal quality component detection system based on deep learning according to claim 3 is characterized in that: The component detection module and the merge output module constitute the deep learning model.
5. The coal quality component detection system based on deep learning according to claim 4 is characterized in that: The deep learning model is built and trained by Keras. The trained deep learning model is saved in .h5 format and deployed on the industrial computer. The industrial computer uses the socket module in python3 for UDP communication to obtain the data to be tested and feedback the test results. During the training process, the loss function of the deep learning model adopts L1 loss, which is expressed as: ,in is the loss value, represents the true label of the i-th sample, represents the detection label of the deep learning model for the i-th sample, N is the total number of training samples; the optimizer in the training process is the Adam optimizer, the learning rate is 0.01, and the batch size is 256.
Citation Information
Patent Citations
Power budget evaluation method based on gate structure deep cross network
CN113222722A
KR20220033673A
