CN115861235A - Flotation tailing ash content prediction method based on multi-feature data fusion - Google Patents
Flotation tailing ash content prediction method based on multi-feature data fusion Download PDFInfo
- Publication number
- CN115861235A CN115861235A CN202211546363.6A CN202211546363A CN115861235A CN 115861235 A CN115861235 A CN 115861235A CN 202211546363 A CN202211546363 A CN 202211546363A CN 115861235 A CN115861235 A CN 115861235A
- Authority
- CN
- China
- Prior art keywords
- image
- layer
- flotation
- prediction
- feature data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention belongs to the technical field of soft measurement and the field of prediction of key indexes in coal dressing industry, and discloses a method for predicting ash content of flotation tailings based on multi-feature data fusion. The method predicts the ash content of the flotation tailing through the ore pulp image in the flotation production process, predicts the deep mining image characteristic data, and realizes the real-time accurate prediction of the ash content of the flotation tailing by compensating and predicting the color characteristic data. The average absolute error of the ash content value of the flotation tail coal is only 0.56%, real-time detection can be realized, flotation production can be guided according to the prediction result, the medicament consumption is reduced, the labor cost is saved, and a solution is provided for intelligent closed-loop control of flotation.
Description
Technical Field
The invention belongs to the technical field of soft measurement and the field of prediction of key indexes in coal dressing industry, and particularly relates to a method for predicting ash content of flotation tailings based on multi-feature data fusion.
Background
The intelligent construction of coal mines relates to a plurality of systems, and a coal preparation plant is used as a terminal system, and the intelligent construction level of the coal preparation plant directly influences the overall height and level of the intelligent construction of the coal mines. Flotation is the most efficient and economic method for processing fine coal, and the intelligent flotation control has great significance on intelligent roads of coal. The flotation tail coal is used as one of two major products of coal slime flotation, the ash content of the flotation tail coal is an important production index of a flotation system, the current operation condition of the flotation system and the clean coal recovery rate can be reflected, and the flotation tail coal flotation method is also significant for flotation intelligent control.
However, the coal slime flotation process is long in time, the detection has strong hysteresis, and the floating coal slime quantity, the equipment running state, the production process variable and the product quality cannot be detected in real time. In addition, in the coal slime flotation production process, the detection and control effects are unsatisfactory due to the problems of complex and changeable separation process (solid-liquid-gas three-phase coexistence, physical reaction, chemical reaction, energy conversion and the like), large flotation treatment capacity, harsh production site environment and the like. Most of the ash content detection of the flotation tail coal of the existing coal preparation plant still adopts a high-temperature ashing method, manual sampling and chemical examination are needed to be completed, 1-2 hours are needed, the detection result is seriously lagged, the production cannot be regulated and controlled in time, and the intelligent requirement of the coal preparation plant is difficult to meet.
At present, the flotation tailing ash content is predicted on line based on ore pulp image characteristic data, but the methods still need to confirm the image characteristics manually, so that the influence of subjective factors cannot be avoided, the model prediction precision is insufficient, and the production cannot be effectively guided.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a flotation tailing ash prediction method based on multi-feature data fusion.
In order to achieve the purpose, the invention adopts the main technical scheme that:
a flotation tail coal ash content prediction method based on multi-feature data fusion comprises the following steps:
s1: collecting images of tailing pulp in a flotation process, preprocessing the collected images, testing actual ash values H (t) of real-time tailing pulp of the corresponding images, and establishing a data set according to the images and the corresponding actual ash values;
furthermore, the tailing coal pulp image is shot by a built image acquisition system, the position and parameters of a CMOS industrial camera are fixed, the intensity and angle of a light source are fixed, the rotating speed of a stirrer is fixed, the equipment is placed in an imaging black box, black paint is sprayed in the box to prevent reflection, and the image acquisition process is not interfered by the outside world;
further, the acquired image is subjected to preliminary preprocessing:
step 1: cutting the image, and fixing the cutting position and size;
step 2: screening the cut images, and removing the images containing abnormal bright spots;
and step 3: denoising the screened image to ensure the image quality;
s2: building a convolutional neural network prediction main model, taking the image preprocessed in the step S1 as model input and the actual grey value H (t) as output, training the convolutional neural network prediction main model, determining parameters of each node in the convolutional neural network, and calculating a preliminary predicted grey value H (t);
furthermore, the convolutional neural network mainly comprises an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer;
furthermore, the convolution layer is a core layer for constructing a convolution neural network, each layer consists of a plurality of convolution kernels, and point multiplication operation is carried out between the input items and the convolution kernels, so that the characteristics of the image are extracted; after the pooling layer is formed in a convolution layer, the feature dimensionality obtained after convolution is large generally, the pooling layer can reduce the dimensionality of data, reduce the number of parameters in a network, and also can effectively prevent overfitting, and average pooling and maximum pooling are common; the fully connected layer is usually before the output layer, and can synthesize all extracted features, and the CNN usually performs dimensionality reduction by arranging multiple fully connected layers at the end;
further, the constructed convolutional neural network comprises 16 convolutional layers, 5 pooling layers and 3 full-connection layers; the two layers of convolution layers of the first part adopt 64 convolution kernels with the size of 3 multiplied by 3, the step length is set to be 1, and then dimension reduction is carried out on the extracted image characteristics through an activation function and a pooling layer with the size of 2 multiplied by 2 and the step length of 2; the second partial two-layer convolutional layer continues to use 3 x 3 convolutional kernels, again with step size set to 1, but with the number of convolutional kernels increased to 128, and the pooling layer is the same as the first partial. The number of convolution kernels in the third part is increased to 256, the number of layers is four, and the pooling layer is unchanged; the number of convolution kernels of the fourth part and the fifth part is 512, and the number of layers is still four; the sixth part is three full-connection layers, parameters are 4096, 4096 and 1000 respectively, and a dropout layer is added to prevent the neural network from being over-fitted; meanwhile, the ReLU activation function is used for the convolution layer and the full connection layer to improve the network propagation speed and avoid gradient explosion;
further, the ore pulp image with the size of 224 × 224 after pretreatment is used as an input layer of the convolutional neural network, the model is trained by taking the actual ash value H (t) as an output label, parameters of each node in the neural network are determined, and a preliminary predicted ash value H (t) is calculated.
S3: extracting color characteristic data of the image preprocessed in the step S1, and normalizing the extracted characteristic data;
furthermore, the color of the ore pulp image has strong correlation with the ash content of the ore pulp image, and workers in a coal preparation plant can roughly estimate the ash content of the flotation tailing coal by observing the color of the ore pulp, because the main components in the ore pulp are gangue and coal slime which have different reflectivities, and the reflectivity of the gangue is far greater than that of the coal, so that the ore pulp with high ash content and the ore pulp with low ash content have difference, and the image color characteristic data can be used as a data set of a compensation model;
further, the extracted color feature data comprises 12 feature data, namely image gray feature data, RGB feature data and HIS feature data;
further, the image gray characteristic data comprises a gray mean, a variance, smoothness, skewness, energy and entropy; the RGB characteristic data comprises an R mean value, a G mean value and a B mean value; the HIS characteristic data comprises an H mean value, an S mean value and an I mean value;
further, the normalization processing adopts a maximum and minimum normalization method.
S4: taking the color characteristic data after normalization processing as the input of a compensation model BP neural network, taking the difference value c (t) between the actual gray value H (t) and the preliminary prediction gray value H (t) as an output label, training the compensation model, determining the parameters of each node of the compensation model, and calculating the compensation prediction value of the actual gray value H (t) and the preliminary prediction gray value H (t)
S5: combining the prediction main model trained in the step S2 with the compensation model trained in the step S4, carrying out the preprocessing process S1 on the tailing pulp image collected in real time, inputting the preprocessed image into the convolutional neural network main model trained in the step S2, outputting a preliminary prediction gray value h (t), extracting color characteristic data of the image, and inputting the color characteristic data after normalization processingS4, outputting a compensation predicted value by the trained compensation model
H (t) and->
Add up to get the final predicted gray value->
According to the preferred embodiment of the present invention, in step S1: the distance between the CMOS industrial camera and the liquid level of the ore pulp is 15cm, so that the image acquisition process is not influenced by the stirring of the ore pulp; the light source irradiates downwards at an angle of 45 degrees, so that the liquid level of the ore pulp can be uniformly illuminated; when the rotating speed of the stirring device is 500r/min, the solution is fully stirred, and the fluctuation of the liquid level is small.
According to the preferred embodiment of the present invention, in step S2: the convolutional neural network adopts a large number of small convolution kernels of 3 multiplied by 3, and compared with larger convolution kernels, the number of layers of the neural network is increased, and the parameter quantity is reduced.
According to the preferred embodiment of the present invention, in step S5: when the trained network is used for application prediction, the industrial camera can be connected with a computer, the acquired image is directly read and input into a preprocessing program, then the image and the characteristic data are respectively input into the main model and the compensation model, and the predicted ash content value is output
The whole process can be fast and accurate without manual participation.
The beneficial effects of the invention are:
the method disclosed by the invention has the advantages that the average absolute error of the prediction result of the ash content value of the flotation tail coal is only 0.56%, the real-time detection can be realized, the flotation production can be guided according to the prediction result, the medicament consumption is reduced, the labor cost is saved, and a solution is provided for the intelligent closed-loop control of the flotation.
Drawings
Fig. 1 an image acquisition apparatus.
FIG. 2 is an algorithm structure of the prediction method of the present invention.
FIG. 3 shows a flow chart of a flotation tailing ash prediction algorithm.
FIG. 4 shows a convolutional neural network structure constructed by the present invention.
FIG. 5 ash prediction values are compared to actual assay values.
Detailed Description
For a better understanding of the present invention, reference is made to the following detailed description of the invention, which is to be read in connection with the accompanying drawings and the specific embodiments. The method utilizes the convolution neural network prediction main model based on the image and the BP neural network prediction compensation model based on the characteristic data to realize real-time prediction of the ash content of the flotation tailing by collecting the flotation tailing pulp image, the prediction result can guide flotation production, the medicament consumption is reduced, the labor cost is saved, and a solution is provided for the flotation intelligent closed-loop control. 300 groups of flotation pulp images of a certain coal preparation plant are selected, 200 groups are used for training, 50 groups are used for verification, and 50 groups are used for testing.
As shown in fig. 2-3, the flotation tailing ash prediction algorithm of the invention includes four parts, namely data acquisition and transmission, image and characteristic data preprocessing, prediction main model based on convolutional neural network, and compensation model based on BP neural network is established by combining main model prediction value H (t), offline test value H (t) and preprocessed data. Using compensation model output values
Compensating the predicted value h (t) of the main model to obtain the final prediction result of flotation clean coal->
The following details are given for each part and their relationship, respectively:
the first step is as follows: data acquisition and transmission
1. Image acquisition and transmission
As shown in fig. 1, image data is collected by the built image collecting device, the collected tailing pulp is put into a container, the container is placed on a stirring device, light sources are positioned above the left side and the right side, an industrial camera is positioned right above the stirring device and is connected with a computer, and the collected images are transmitted and stored.
2. Assay data collection and transmission
The test data refers to the actual ash value obtained by performing a timing test on the flotation tail coal. The tailings products are first collected and then subjected to filtration, drying, sample preparation and ash burning. The filtered coal sample is dried in an oven for at least about half a day. Then, the sample preparation is carried out according to the national standard. Finally, the coal samples were weighed in porcelain boats and placed in muffle furnaces for combustion, which took about two and a half hours. The process of raising the temperature from 100 ℃ to 500 ℃ needs 30min, then the temperature is kept at 500 ℃ for 30min, finally the temperature is raised to (815 +/-10) ℃, and finally the accurate value of the ash content of the tail coal is obtained.
The second step is that: image and feature data preprocessing
1. Image data pre-processing
According to the method, the tailing pulp image is acquired through the CMOS industrial camera, the position is determined firstly, the image is cut, the cut image is screened, abnormal samples are removed, then the image is subjected to denoising processing, and the quality of a sample set is guaranteed.
2. Image feature data preprocessing
The method extracts the gray characteristic data, RGB characteristic data and HIS characteristic data of the preprocessed image, performs normalization preprocessing on the extracted data, and adopts a maximum and minimum normalization method.
The third step: constructing prediction main model based on convolutional neural network
As shown in fig. 4, the convolutional neural network constructed by the present invention comprises 16 convolutional layers, 5 pooling layers and 3 fully-connected layers; the two layers of convolution layers of the first part adopt 64 convolution kernels with the size of 3 multiplied by 3, the step length is set to be 1, and then dimension reduction is carried out on the extracted image characteristics through an activation function and a pooling layer with the size of 2 multiplied by 2 and the step length of 2; the second two-layer convolutional layer continues to use 3 x 3 convolutional kernels, again with step size set to 1, but the number of convolutional kernels increases to 128, and the pooling layer is the same as the first. The number of convolution kernels in the third part is increased to 256, the number of layers is four, and the pooling layer is unchanged; the number of convolution kernels of the fourth part and the fifth part is 512, and the number of layers is still four; the sixth part is three full-connection layers, parameters are 4096, 4096 and 1000 respectively, and a dropout layer is added to prevent overfitting of the neural network; meanwhile, the ReLU activation function is used for the convolution layer and the full-link layer to improve the network propagation speed and avoid gradient explosion.
Taking the pretreated ore pulp image with the size of 224 multiplied by 224 as the input of a convolutional neural network, taking the actual gray value H (t) as an output label, taking 200 images for training, taking 50 images for verification, carrying out back propagation on the error obtained by gradient descent by the convolutional neural network, updating the parameters of each layer of the convolutional neural network layer by layer, and finally determining the parameters and the structure of a prediction main model through multiple rounds of iterative training.
The fourth step: construction of prediction compensation model based on BP neural network
In order to improve the prediction precision, the invention constructs a prediction compensation model based on a BP neural network, takes the preprocessed image gray characteristic data, RGB characteristic data and HIS characteristic data as input, takes the difference value c (t) of the actual gray value H (t) and the preliminary prediction gray value H (t) as an output label, trains the compensation model, and determines the parameters of each node of the compensation model.
The fifth step: prediction using trained models
Inputting 50 groups of untrained samples into a constructed model, inputting an image into a main model based on a convolutional neural network, outputting an ash content prediction initial value h (t), normalizing extracted image characteristic data, inputting a compensation model after processing, and outputting an ash content prediction compensation value
Finally obtaining the real-time ash content predicted value of the tail coal>
As shown in a of fig. 5, a comparison graph of the flotation tail coal ash content predicted by the prediction model method of the present invention and the actual test value shows that the predicted tail coal ash content value is very consistent with the actual value, and b of fig. 5 shows that the average absolute error is only 0.56%, which indicates that the ash content detection method of the present invention has high accuracy and can be used for on-site pulp ash content detection.
Claims (7)
1. A flotation tail coal ash content prediction method based on multi-feature data fusion is characterized by comprising the following steps:
s1: collecting tailing slurry images in the flotation process, preprocessing the collected images, testing actual ash values H (t) of real-time tailing slurry corresponding to the images, and establishing a data set according to the images and the corresponding actual ash values;
s2: building a convolutional neural network prediction main model, taking the image preprocessed in the step S1 as model input and the actual grey value H (t) as output, training the convolutional neural network prediction main model, determining parameters of each node in the convolutional neural network, and calculating a preliminary predicted grey value H (t);
s3: extracting color characteristic data of the image preprocessed in the step S1, and normalizing the extracted characteristic data;
s4: taking the color characteristic data subjected to normalization processing in the step S3 as the input of a compensation model BP neural network, taking the difference c (t) between the actual gray value H (t) and the preliminary prediction gray value H (t) as an output label, training the compensation model, determining the parameters of each node of the compensation model, and calculating the compensation prediction value of the actual gray value H (t) and the preliminary prediction gray value H (t)
S5: combining the prediction main model trained in the step S2 with the compensation model trained in the step S4, enabling the tailing coal pulp image collected in real time to be subjected to the preprocessing process S1, inputting the preprocessed image into the convolutional neural network main model trained in the step S2, outputting a preliminary prediction gray value h (t), extracting color characteristic data of the image, inputting the normalized image into the compensation model trained in the step S4, and outputting a compensation prediction value
H (t) and->
Add up to get the final predicted gray value->
2. The method for predicting ash content of flotation tailings based on multi-feature data fusion according to claim 1, wherein in the step S1, the step of preprocessing the collected image comprises the following steps:
step 1: cutting the image, and fixing the cutting position and size;
and 2, step: screening the cut images, and removing the images containing abnormal bright spots;
and step 3: and denoising the screened image to ensure the image quality.
3. The method for predicting ash content of flotation tailings based on multi-feature data fusion of claim 1, wherein in the step S2, the convolutional neural network mainly comprises an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer;
the preprocessed ore pulp image with the size of 224 multiplied by 224 is used as an input layer of the convolutional neural network;
16 convolution layers, 5 pooling layers and 3 full-connection layers;
the preliminary predicted gray value h (t) is the output layer.
4. The method for predicting ash content of flotation tailings based on multi-feature data fusion according to claim 3, wherein specific parameters of the convolutional layer, the pooling layer and the fully-connected layer are as follows:
the two layers of convolution layers of the first part adopt 64 convolution kernels with the size of 3 multiplied by 3, the step length is set to be 1, and then dimension reduction is carried out on the extracted image characteristics through an activation function and a pooling layer with the size of 2 multiplied by 2 and the step length of 2;
the second part of two-layer convolutional layers continues to adopt a convolution kernel of 3 multiplied by 3, the step length is also set to be 1, but the number of the convolution kernels is increased to 128, and the pooling layers are the same as the first part;
the number of convolution kernels in the third part is increased to 256, the number of layers is four, and the pooling layer is unchanged;
the number of convolution kernels of the fourth part and the fifth part is 512, and the number of layers is still four;
the sixth part is three full-connection layers, parameters are 4096, 4096 and 1000 respectively, and a dropout layer is added to prevent overfitting of the neural network; meanwhile, the ReLU activation function is used for the convolution layer and the full connection layer to improve the network propagation speed and avoid gradient explosion.
5. The method for predicting ash content in flotation tailings based on multi-feature data fusion as claimed in claim 1, wherein in step S3, the extracted color feature data comprises 12 feature data, namely image gray feature data, RGB feature data and HIS feature data.
6. The method of claim 5, wherein the flotation tailings ash prediction method based on multi-feature data fusion,
the image gray characteristic data comprises a gray mean value, a variance, smoothness, skewness, energy and entropy;
the RGB characteristic data comprises an R mean value, a G mean value and a B mean value;
the HIS characteristic data comprises an H mean value, an S mean value and an I mean value.
7. The method for predicting ash content of flotation tailings based on multi-feature data fusion as claimed in claim 1, wherein in the step S3, the normalization process adopts a maximum-minimum normalization method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211546363.6A CN115861235A (en) | 2022-12-05 | 2022-12-05 | Flotation tailing ash content prediction method based on multi-feature data fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211546363.6A CN115861235A (en) | 2022-12-05 | 2022-12-05 | Flotation tailing ash content prediction method based on multi-feature data fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115861235A true CN115861235A (en) | 2023-03-28 |
Family
ID=85669782
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211546363.6A Pending CN115861235A (en) | 2022-12-05 | 2022-12-05 | Flotation tailing ash content prediction method based on multi-feature data fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115861235A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117092050A (en) * | 2023-09-13 | 2023-11-21 | 佛山科学技术学院 | Coal slime flotation ash content detection method and system based on spectrum multi-mode time sequence learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109269951A (en) * | 2018-09-06 | 2019-01-25 | 山西智卓电气有限公司 | Floating tail-coal ash content, concentration, coarse granule detection method of content based on image |
| CN114841453A (en) * | 2022-05-18 | 2022-08-02 | 中国矿业大学 | Clean coal ash content prediction method in flotation process |
-
2022
- 2022-12-05 CN CN202211546363.6A patent/CN115861235A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109269951A (en) * | 2018-09-06 | 2019-01-25 | 山西智卓电气有限公司 | Floating tail-coal ash content, concentration, coarse granule detection method of content based on image |
| CN114841453A (en) * | 2022-05-18 | 2022-08-02 | 中国矿业大学 | Clean coal ash content prediction method in flotation process |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117092050A (en) * | 2023-09-13 | 2023-11-21 | 佛山科学技术学院 | Coal slime flotation ash content detection method and system based on spectrum multi-mode time sequence learning |
| CN117092050B (en) * | 2023-09-13 | 2024-02-27 | 佛山科学技术学院 | Coal slime flotation ash content detection method and system based on spectrum multi-mode time sequence learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598736B (en) | Power equipment infrared image fault positioning, identifying and predicting method | |
| CN105678332B (en) | Converter steelmaking end point judgment method and system based on flame image CNN recognition modeling | |
| Xue et al. | A fast and easy method for predicting agricultural waste compost maturity by image-based deep learning | |
| CN110807760B (en) | Tobacco leaf grading method and system | |
| CN112288021B (en) | Medical wastewater monitoring data quality control method, device and system | |
| CN111638249B (en) | Water content measuring method based on deep learning and application of water content measuring method in oil well exploitation | |
| CN115861235A (en) | Flotation tailing ash content prediction method based on multi-feature data fusion | |
| CN112488082A (en) | Coal gangue intelligent sorting system based on deep learning | |
| CN115812634B (en) | Method and device for nondestructively sorting pre-hatching egg fertilization information | |
| CN109977747A (en) | A kind of coal rock detection system of comprehensive space-time characteristic | |
| CN112884379A (en) | Method for predicting coal mine underground human factor safety risk | |
| CN111649779B (en) | Oil well oil content and total flow rate measuring method based on dense neural network and application | |
| CN108931621B (en) | Zinc ore grade soft measurement method based on process texture characteristics | |
| CN115329853B (en) | Equipment parameter prediction and knowledge transfer method based on multi-source domain migration | |
| CN112749675A (en) | Potato disease identification method based on convolutional neural network | |
| CN116451848A (en) | Satellite telemetry data prediction method and device based on space-time attention mechanism | |
| CN113567159A (en) | Scraper conveyor state monitoring and fault diagnosis method based on edge cloud cooperation | |
| CN112947080B (en) | Scene parameter transformation-based intelligent decision model performance evaluation system | |
| CN110598326A (en) | Well testing interpretation method based on artificial intelligence | |
| CN112215410B (en) | Power load prediction method based on improved deep learning | |
| CN116312860B (en) | Agricultural product soluble solid matter prediction method based on supervised transfer learning | |
| CN116975645A (en) | Industrial process soft measurement modeling method based on VAE-MRCNN | |
| CN113295625B (en) | Machine vision dye concentration spectrum detection method based on extreme gradient promotion | |
| CN113468823B (en) | Optical module damage detection method and system based on machine learning | |
| CN113835964A (en) | Cloud data center server energy consumption prediction method based on small sample learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |