YOLOv3-based gasification furnace ash identification method

Abstract

The invention belongs to the field of gasification furnace ash identification, and the method comprises the following steps: learning that an ash picture which is manually collected and marked is associated with a real carbon residue rate of a laboratory by using a YOLOv3 network structure to obtain a model capable of automatically identifying the ash carbon residue rate, and then loading the model to return an identification result and the carbon residue rate of the picture in real time. According to the method, the defects of experience conclusion obtained by manually identifying the ash and serious hysteresis of a laboratory in the prior art are overcome, the ash is rapidly and accurately identified, target detection is rapidly achieved, meanwhile, the accuracy rate superior to that of manual experience is obtained, the carbon residue rate can be rapidly identified and obtained through the model obtained through the method, and the method is suitable for large-scale popularization and application. And the device can be widely applied to related gasifiers.

Description

Technical Field

[0001] The invention belongs to the technical field of gasifier ash identification, and specifically relates to photographic identification of gasifier ash, and especially to a gasifier that requires manual inspection of the ash. Background Art

[0002] As the leading device of coal chemical enterprises, the safety, stability and economic operation of gasifiers are the key points for coal chemical enterprises to ensure safe production, reduce energy consumption and improve economic benefits. As an important indicator for evaluating the operation status of gasifiers, the carbon content of ash has a significant impact on production consumption and energy conservation and emission reduction. The carbon content of gasifier ash is high, which directly affects the high consumption of raw coal and high production costs of the system. At present, the identification of residual carbon in ash is divided into laboratory measurement and visual estimation. After taking the ash, it is ground, dried, sampled, and burned. The real residual carbon rate is obtained. However, the residual carbon rate test process is cumbersome and the measurement process takes a long time. It usually takes 6 hours to get the result. At present, it is estimated based on the laboratory. Experienced operators judge the residual carbon rate based on experience, which can provide guidance for faster improvement of gasifier production.

[0003] At present, in the operation of the gasifier, an important work content is that every 8 hours, the operator needs to open the inlet ash discharge hole on the vertical ash pipe, dump part of the ash on the ground, and then spread it flat, check the size, color and shape of the ash, and pour some water on the ash when necessary. Finally, the ash is collected and sent to the ash ditch. The main disadvantage of this is the time lag, and the parameters such as the steam-oxygen ratio cannot be adjusted in time according to the ash information. On the other hand, the labor intensity is high, the working conditions on site are relatively poor, and the identification of ash is very dependent on the experience and ability of the operator. Different operators come to different conclusions on the same batch of ash. With the help of modern computer technology, the ash is identified and standardized, the accuracy of ash diagnosis is improved, and more accurate guidance is provided for the operation of the gasifier. Summary of the invention

[0004] The purpose of this invention is to provide a gasifier ash recognition method based on YOLOv3.

[0005] The purpose of the present invention is achieved through the following technical solutions A gasifier ash recognition method based on YOLOv3, comprising the following steps: Step 1: Collect coal ash images and test the residual carbon rate: Use a camera to take pictures, and conduct a carbon residue laboratory analysis on each sample to obtain the actual residual carbon rate. Take a picture of each ash sample first, and then take another picture after spraying water. In particular, the picture taken after spraying water to remove the floating dust can clearly reflect the actual condition of the ash.

[0006] Step 2: Image annotation and data set division: Use the software labelimg to annotate the pictures, and get the ash and slag types for communication with the operator to select and mark them into 5 categories: a lot of fine ash, some fine ash, no fine ash and no obvious agglomeration, some slag, and severe slag. Get the image annotation information file, and then organize the data and the corresponding annotation files according to the format of the data set to annotate the position and category of the object boundary box in the photo to ensure the balance of image types. Divide the data set into training set: validation set: test set, with a ratio of 7:2:1.

[0007] Step 3: Image data enhancement: Preprocess the training set images, rotate, crop, and increase the number of training set images to improve the generalization ability of the model.

[0008] Step 4: Set the main parameters of model training: image width and height are set to 416×416, number of channels is set to 3, learning_rate learning efficiency is set to 0.001, momentum parameter momentum is set to 0.9, training epoch is set to 300, batch is set to 16, learning adjustment strategy policy is set to steps, and learning efficiency decay is set to 0.1. Train the processed data set, calculate the loss function in each iteration, and update the parameter value to minimize the value of the loss function until the model converges. At the same time, to prevent overfitting, the number of iterations does not exceed 300.

[0009] Step 5. After completing the model training, save the model weight parameters and set the format to .pt. Reload the saved model weight file and use this weight file to detect the images in the test set.

[0010] The working process and principle of the present invention are: The ash recognition method based on the YOLOv3 algorithm uses labelimg for annotation in the sample labeling stage; the YOLOv3 network structure is used to learn the ash recognition through the actual residual carbon rate and manual experience, and a model that can automatically identify ash and derive the residual carbon rate is obtained. Then the model is embedded in the intelligent production system to achieve real-time shooting and real-time return of recognition results and residual carbon rate. It solves the drawbacks of time-consuming and inefficient manual inspection and inconsistent manual experience, making the gasifier operation more efficient and safe. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Model training process

[0012] Figure 2 Original photo of ash

[0013] Figure 3 The test result diagram for the test set Implementation

[0014] 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.

[0015] A gasifier ash recognition method based on YOLOv3, comprising the following steps: Step 1: Collect coal ash images and test the residual carbon rate: Use a camera to take pictures and perform a carbon residual laboratory analysis on each sample to obtain the actual residual carbon rate.

[0016] Take a photo of each ash sample first, and then take another photo after spraying water. Especially the photo after spraying water to remove the floating dust can clearly reflect the actual condition of the ash.

[0017] Step 2: Image annotation and data set division: Use the software labelimg to annotate the pictures, and get the ash and slag types for communication with the operator to select and mark them into 5 categories: a lot of fine ash, some fine ash, no fine ash and no obvious agglomeration, some slag, and severe slag. Get the image annotation information file, and then organize the data and the corresponding annotation files according to the format of the data set to annotate the position and category of the object boundary box in the photo to ensure the balance of image types. Divide the data set into training set: validation set: test set, with a ratio of 7:2:1.

[0018] Step 3: Image data enhancement: Preprocess the training set images, rotate, crop, and increase the number of training set images to improve the generalization ability of the model.

[0019] Step 4: Set the main parameters of model training: image width and height are set to 416×416, number of channels is set to 3, learning_rate learning efficiency is set to 0.001, momentum parameter momentum is set to 0.9, training epoch is set to 300, batch is set to 16, learning adjustment strategy (policy) is set to steps, and learning efficiency decay is set to 0.1. Train the processed data set; calculate the loss function in each iteration and update the parameter value to minimize the value of the loss function until the model converges. At the same time, to prevent overfitting, the number of iterations does not exceed 300.

[0020] Step 5. After completing the model training, save the model weight parameters and set the format to .pt. Reload the saved model weight file and use this weight file to detect the images in the test set.

[0021] In the field of deep learning object detection, common object detection methods can be divided into two models: two-stage and one-stage. The traditional two-stage model needs to generate a large number of candidate boxes during the object detection process, and then use convolutional neural networks to classify these candidate boxes. However, this method has the problems of high computational cost and low efficiency. To overcome these problems, YOLOv3 introduced a one-stage model, which converts the positioning problem of the target bounding box into a regression problem, avoiding the candidate box generation process, thereby greatly improving the speed of the algorithm.

[0022] The working process and principle of the present invention are: The ash recognition method based on the YOLOv3 algorithm uses labelimg for annotation in the sample labeling stage; the YOLOv3 network structure is used to learn the ash recognition through the actual residual carbon rate and manual experience, and a model that can automatically identify ash and derive the residual carbon rate is obtained. Then the model is embedded in the intelligent production system to achieve real-time shooting and real-time return of recognition results and residual carbon rate. It solves the drawbacks of time-consuming and inefficient manual inspection and inconsistent manual experience, making the gasifier operation more efficient and safe.

[0023] The gasification device includes but is not limited to Lurgi furnace, Texaco water-coal slurry gasification furnace, and aerospace furnace. The main parameter adjustment of step 4 includes learning rate, learning rate decay, momentum parameter, batch size, batch size division and threshold, etc., with the purpose of the fastest and best accuracy.

[0024] The beneficial effects of the present invention are: There are many advantages to automating the identification of gasifier ash. Here are some of the main ones: Improved efficiency: Ash identification can achieve fast and accurate identification and classification to obtain residual carbon rate, which can greatly improve production efficiency. Compared with laboratory processing, the time to obtain residual carbon rate is greatly shortened. Accuracy and consistency: Accurately detect and identify the composition and characteristics of ash. This eliminates the errors that may be introduced by different manual experience and ensures consistent results. Quality control: Through ash identification, better quality control can be achieved. The system can monitor and record the composition, quality and other relevant parameters of ash, thereby helping to implement timely adjustments and improvements in the production process. Data analysis and decision support: Collect and generate statistical information and trend analysis about ash. These data can help companies make more informed decisions, optimize production processes, and implement more effective resource management strategies.

[0025] Overall, ash identification can increase production efficiency, provide precise and consistent results, improve quality control, and provide data analysis support. These advantages can help companies reduce costs, improve product quality, and gain a competitive advantage in a highly competitive market.

Claims

1. A gasifier ash recognition method based on YOLOv3, comprising the following steps: Step 1: Collect coal ash images and test the residual carbon rate: Use a camera to take pictures, and conduct a carbon residual laboratory analysis on each sample to obtain the actual residual carbon rate. Take a picture of each ash sample first, and then take another picture after spraying water. In particular, the picture after spraying water to remove the floating ash can clearly reflect the actual condition of the ash; Step 2: Image annotation and data set division: Use the software labelImg to annotate the pictures, and get the ash types communicated with the operator for frame selection and labeling classification, which are divided into 5 categories: a lot of fine ash, some fine ash, no fine ash and no obvious agglomeration, some slag, and severe slag. Obtain the image annotation information file, and then organize the data and the corresponding annotation files according to the format of the data set. Annotate the position and category of the object boundary box in the photo to ensure the balance of image types, and divide the data set into training set: validation set: test set, with a ratio of 7:2:1; Step 3: Image data enhancement: The training set images Preprocess, rotate, crop, and increase the number of training set images to improve the generalization ability of the model; Step 4: Set the main parameters of model training: set the image width and height to 416×416, the number of channels to 3, the learning_rate learning efficiency to 0.001, the momentum parameter momentum to 0.9, the training epoch to 300, the batch to 16, the learning adjustment strategy policy to steps, the learning efficiency decay to 0.1, and train the processed data set; calculate the loss function in each iteration and update the parameter value to minimize the value of the loss function until the model converges. At the same time, to prevent overfitting, the number of iterations does not exceed 300; Step 5. After completing the model training, save the model weight parameters, set the format to .pt format, reload the model weight file saved, and use this weight file to detect the pictures of the test set.

2. The method according to claim 1, wherein the gasification device includes but is not limited to a Lurgi furnace, a Texaco water-coal slurry gasification furnace, and an aerospace furnace.

3. The method according to claim 1, wherein the main parameter adjustment in step 4 includes learning rate, learning rate decay, momentum parameter, batch size, batch size division and threshold, etc., with the purpose of achieving the fastest and best accuracy.

Images