Water-cooled wall ash deposition detection and cleaning system based on computer vision

Abstract

The invention provides a garbage incinerator water-cooled wall ash deposition detecting and cleaning system based on computer vision. Computer vision is used for detecting whether ash deposits on the water-cooled wall or not, so that the mechanical arm is controlled to clean the ash deposits on the water-cooled wall, and automatic and intelligent ash cleaning is achieved. According to the scheme, an image acquisition module is used for acquiring a water-cooled wall image; the ash deposition recognition module is used for detecting the water-cooled wall image to obtain an ash deposition position; the dust deposition identification module comprises an image preprocessing network, a backbone network, a neck network and a head network; the image preprocessing network is used for performing image enhancement processing on the water-cooled wall image to obtain a processed image; the backbone network is used for extracting different levels of feature information of the processed image; the neck network is used for performing feature fusion on the different levels of feature information to obtain a fused image; and the head network is used for performing target prediction based on the fused image to obtain the dust deposition position.

Description

Technical Field

[0001] The invention relates to the technical field of water-cooling wall dust accumulation cleaning, and in particular to a water-cooling wall dust accumulation detection and cleaning system based on computer vision. Background Art

[0002] As one of the key evaluation indicators of waste incineration power plants, the continuous operation time of the waste incinerator directly determines its economic benefits. Since the glass, ash and salt components in my country's urban domestic waste are relatively high, the fly ash is easy to melt and sinter under high temperature conditions, which leads to the phenomenon of ash accumulation in the incinerator, which not only reduces the thermal efficiency of the boiler, but also easily causes high-temperature corrosion. In severe cases, it may even cause the flow area of ​​the incinerator throat to decrease or the superheater tube to burst, forcing the incinerator to be shut down for maintenance. Its operation data shows that the water-cooled walls around the second and third flues of the waste heat furnace and the evaporation heating surface at the tail are prone to serious ash accumulation.

[0003] Traditional water-cooled wall cleaning is usually done manually, which requires a lot of manpower and material resources to complete. At the same time, a large amount of dust will be generated when cleaning the dust on the water-cooled wall, which will have a serious impact on the health of the workers. Therefore, it is of great theoretical and practical significance to propose a safe, reliable and low-cost method to clean the dust on the surface of the water-cooled wall. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of low efficiency of manual cleaning of dust accumulation on the water-cooled wall and great harm to workers during cleaning. A system for detecting and cleaning dust accumulation on the water-cooled wall of a garbage incinerator based on computer vision is proposed. Computer vision is used to detect whether there is dust accumulation on the water-cooled wall, so as to control the robot arm to clean the dust accumulation on the water-cooled wall, and realize automatic intelligent dust cleaning. It effectively improves dust removal efficiency and saves company costs.

[0005] The technical solution of the present invention is:

[0006] A water-cooled wall dust accumulation detection system based on computer vision.

[0007] It includes an image acquisition module and a dust accumulation recognition module;

[0008] The image acquisition module is used to acquire the water-cooled wall image; the dust accumulation recognition module is used to detect the water-cooled wall image and obtain the dust accumulation position;

[0009] The dust accumulation recognition module includes: an image preprocessing network, a backbone network, a neck network and a head network; the image preprocessing network is used to perform image enhancement processing on the water-cooled wall image to obtain a processed image; the backbone network is used to extract different levels of feature information of the processed image; the neck network is used to perform feature fusion of the different levels of feature information to obtain a fused image; the head network is used to perform target prediction based on the fused image to obtain the dust accumulation position. The controller controls the spray gun to clean according to the dust accumulation position, re-predicts after cleaning, controls the cleaning process, and evaluates the cleaning effect.

[0010] Furthermore, the image preprocessing network is to perform adaptive histogram enhancement and gamma correction on the image, and then fuse the processed images into one image through the Laplacian pyramid. Its structure is as follows: Figure 6 shown.

[0011] The adaptive histogram enhancement is to adaptively equalize the R, G, and B channels in the image respectively, and then merge the enhanced R, G, and B channels together.

[0012] The gamma correction is to normalize the image and then perform nonlinear operations on the grayscale value of the image, which can make the grayscale value of the output image have an exponential relationship with the grayscale value of the input image. The formula is

[0013] f(x)=x γ

[0014] The Laplacian pyramid fusion: Figure 5 As shown in the figure, first, downsample the two images that need Laplacian fusion. Second, upsample the downsampled image. Third, calculate the difference between the processed image and the original image, and adjust the residual to be in the range of 0-255. Fourth, calculate the sum of the residuals of each layer of the pyramid. Finally, add the total residuals of each layer starting from the top of the pyramid. After each layer, it is necessary to make the value of the image pixel between 0-255.

[0015] Furthermore, the backbone network has a total of 13 layers of network structure:

[0016] The first layer is CBH, which contains a Conv2d layer, a Batchnorm2d and an activation function Hardwish, with 3 convolution kernels, 2 strides, 1 pooling layer, 3 input channels, and 8 output channels;

[0017] The second layer is the Mobilenetv3 layer, with a convolution kernel size of 3, a stride of 2, a pooling of 1, an input channel of 8, an output channel of 8, and a hidden layer hidden_dim of 16;

[0018] The third layer is the Mobilenetv3 layer, with a convolution kernel size of 3, a stride of 2, a pooling of 0, an input channel of 8, an output channel of 8, and a hidden layer hidden_dim of 72;

[0019] The fourth layer is the Mobilenetv3 layer, with a convolution kernel size of 3, a stride of 1, a pooling of 0, an input channel of 8, an output channel of 8, and a hidden layer hidden_dim of 88;

[0020] The fifth layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 2, a pooling of 1, an input channel of 8, an output channel of 16, and a hidden layer hidden_dim of 96;

[0021] The sixth layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 240;

[0022] The 7th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 240;

[0023] The 8th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 120;

[0024] The 9th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 144;

[0025] The 10th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 2, a pooling of 1, an input channel of 16, an output channel of 24, and a hidden layer hidden_dim of 288;

[0026] The 11th layer is a Mobilenetv3 layer with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 24, an output channel of 24, and a hidden layer hidden_dim of 576;

[0027] The 12th layer is a Mobilenetv3 layer with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 24, an output channel of 24, and a hidden layer hidden_dim of 576;

[0028] The 13th layer is the AWGAM attention mechanism layer, with 24 input channels, 7 convolution kernels, 3 pooling layers, and an output feature size of 20*20*576.

[0029] Furthermore, the neck network has a total of 12 layers of network structure:

[0030] The first layer is the upsample layer, which takes the 13th layer output of the backbone network as input, and the output feature size is 40*40*24;

[0031] The second layer is the concat layer, which takes the output of the first layer and the 9th layer of the backbone network as input, and the output feature size is 40*40*40;

[0032] The third layer is C2f, which takes the output of the second layer as input. The output feature size is 40*40*64, and Bottleneck has a shortcut.

[0033] The fourth layer is the upsample layer, which takes the output of the third layer as input and the output feature size is 80*80*64;

[0034] The fifth layer is the concat layer, which takes the output of the fourth layer and the fourth layer of the backbone network as input, and the output feature size is 80*80*72;

[0035] The sixth layer is the C2f layer, which takes the output of the fifth layer as input. The output feature size is 80*80*32, and there is a shortcut for Bottleneck.

[0036] The 7th layer is the Conv layer, which takes the output of the 6th layer as input, has a convolution kernel of 3, a stride of 2, a pooling layer of 1, and an output feature size of 40*40*32;

[0037] The 8th layer is the concat layer, which takes the output of the 7th and 3rd layers as input, and the output feature size is 40*40*72;

[0038] The 9th layer is the C2f layer, which takes the output of the 8th layer as input. The output feature size is 40*40*64, and there is a shortcut for Bottleneck.

[0039] The 10th layer is the Conv layer, which takes the output of the 9th layer as input, has a convolution kernel of 3, a stride of 2, a pooling layer of 1, and an output feature size of 20*20*64;

[0040] The 11th layer is the concat layer, which takes the output of the 10th layer and the 13th layer of the backbone network as input, and the output feature size is 20*20*88;

[0041] The 12th layer is the C2f layer, which takes the output of the 11th layer as input. The output feature size is 20*20*128, and Bottleneck has a shortcut.

[0042] Furthermore, the neck network has a total of 12 layers of network structure:

[0043] The head network has a total of 3 layers of network structure:

[0044] The first layer is the Detect layer, which takes the output of the sixth layer of the neck network as input;

[0045] The second layer is the Detect layer, which takes the output of the 9th layer of the neck network as input;

[0046] The third layer is the Detect layer, which takes the output of the 12th layer of the neck network as input.

[0047] A water-cooled wall dust cleaning device comprises a dust detection system and a cleaning module for cleaning the dust.

[0048] Furthermore, the cleaning module includes a mechanical arm and a soot blower installed on the mechanical arm.

[0049] Furthermore, the image acquisition device includes a 360° control pan-tilt fixed on the robotic arm, a camera installed on the control pan-tilt and a pan-tilt controller. The pan-tilt controller is used to 360° control the pan-tilt to drive the high-definition visible light camera to rotate and acquire the image of the water-cooled wall in the front field of view.

[0050] First, make an initial prediction - then control the soot blower to clean, and after cleaning, make a second prediction. If the position has been cleared, clean the next area. If there is still an unclean area, control the spray gun to clean it again. After cleaning, re-judge. If a position has been cleaned many times but is still not clean, store the position and the picture, skip the area and clean the next area. It is convenient for targeted cleaning or manual cleaning later.

[0051] The beneficial effect of the present invention is that by taking real-time photos of the water-cooled wall and then inputting the captured image into the algorithm, it is possible to effectively identify whether there is dust accumulation on the water-cooled wall. Thanks to the Mobilenetv3 network structure in the model, the water-cooled wall dust accumulation detection algorithm of the present invention has a parameter amount of only 1.2 megabytes. It does not occupy too many computing resources and is very suitable for use in scenarios where computing resources are limited. It can be well applied to embedded devices. The water-cooled wall dust accumulation detection algorithm of the present invention can replace manual detection of dust accumulation on the water-cooled wall, which can improve the intelligence of the water-cooled wall cleaning robot arm, reduce the labor cost of the enterprise, improve the efficiency of cleaning dust accumulation on the water-cooled wall, and provide construction companies with sustainable development capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0053] Figure 1 This is the training result of AWGAM-YOLOv8n.

[0054] Figure 2 This is a comparison chart of images using different image enhancement algorithms;

[0055] Figure 3 This is the effect diagram of improving model detection;

[0056] Figure 4 These are some sample graphs of the dataset;

[0057] Figure 5 It is a Laplace pyramid fusion network structure diagram of the present invention;

[0058] Figure 6 It is a structural diagram of the image enhancement algorithm based on multi-scale fusion of the present invention;

[0059] Figure 7 It is a computer vision algorithm structure diagram of the present invention (network framework scheme of AWGAM-YOLOv8n model);

[0060] Figure 8 This is the core module diagram of Mobilenetv3;

[0061] Fig. 9 This is the structure diagram of the attention mechanism AWGAM of the present invention;

[0062] Fig.10 is a schematic diagram of the system structure of an embodiment of the present application;

[0063] Fig.11 2 is a structural diagram of the mechanical arm of this embodiment.

[0064] Fig.12 Shown is a picture of water-cooled wall pipes with dust accumulation.

[0065] Fig.13 Shown is a picture of the water-cooled wall tubes without dust accumulation.

[0066] The numbers in the attached drawings are: 1 is the air blowing port of the soot blowing system carried by the robotic arm, 2 is the camera and image recognition component carried by the robotic arm, and 3 is the air inlet of the soot blowing system carried by the robotic arm. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0068] Example 1

[0069] like Fig.10 , 11 As shown, the water-cooled wall dust cleaning robot arm based on computer vision includes a water-cooled wall dust accumulation recognition module and a dust cleaning system;

[0070] The water-cooled wall dust accumulation recognition module is used to identify the dust accumulation position based on the dust accumulation detection model. The water-cooled wall dust accumulation recognition module includes: an image acquisition module and a dust accumulation recognition module.

[0071] The image acquisition module is used to acquire the water-cooled wall image; the dust accumulation recognition module is used to detect the water-cooled wall image based on the dust accumulation detection model to obtain the dust accumulation position. In this embodiment, the image acquisition module includes a high-definition visible light camera, a 360° control pan-tilt and a pan-tilt controller. The pan-tilt controller is used to control the pan-tilt 360° to drive the high-definition visible light camera to rotate and acquire the water-cooled wall image in the front field of view; the dust accumulation recognition device includes an embedded microprocessor, which uses the dust accumulation detection model to identify the dust accumulation position on the water-cooled wall in the image and record the dust accumulation position information at the detection time.

[0072] The dust accumulation detection model includes: image preprocessing network, backbone network, neck network and head network; the image preprocessing network is used to enhance the water-cooled wall image to obtain the processed image; the backbone network is used to perform convolution pooling operations on the processed image to obtain feature images of different levels; the neck network is used to fuse the feature images of different levels to obtain the fused image; the head network is used to predict the target based on the fused image to obtain the dust accumulation position.

[0073] In this embodiment, we have improved a new dust accumulation detection model. We replaced the backbone feature extraction network of YOLOv8n algorithm structure with the backbone network designed by ourselves. Specifically, the backbone feature extraction network of YOLOv8n was replaced with Mobilenetv3 network, which greatly reduced the parameters in the model. At the same time, we have also improved a new attention mechanism AWGAM based on the GAM attention mechanism. AWGAM extracts feature information of images in channels and spaces respectively, and then weights and merges the extracted feature information, and finally outputs the image to the Neck network. The AWGAM attention mechanism enables our backbone network to better extract feature information from images, thereby improving the accuracy of model detection. Thanks to the Mobilenetv3 network structure in the model, the improved model parameter volume is only 1.2 megabytes. It does not occupy too much computing resources, and is very suitable for scenarios with limited computing resources, and can be well applied to embedded devices. The dust accumulation detection model of the present invention includes four parts: an image preprocessing network, an input network, a backbone network, a neck network and a head network. The main recognition process includes: the image preprocessing network enhances the collected water-cooled wall image to obtain a processed image; the processed image will be input into the backbone network through the input network for feature extraction, and a series of feature maps of different levels of different sizes will be gradually obtained after operations such as convolution layers and pooling layers. The low-level feature maps pay more attention to image texture and edge information, while the high-level feature maps pay more attention to the semantic information of the image; in order to obtain more comprehensive language information, the neck network will fuse the different levels of feature maps generated by the backbone network to obtain a fused image, so as to obtain richer and more diverse feature information; after feature fusion, the YOLOv8n in the head network regards each grid unit of the fused image as a candidate target frame, and performs target prediction on each target frame. According to the water-cooled wall pipe pictures with dust accumulation and the water-cooled wall pipe pictures without dust accumulation used when we train the model, as shown in FIG. Fig.12 , 13 The algorithm of the present invention can determine whether there is dust accumulation on the water-cooled wall pipe, and if there is dust accumulation, the position information of the dust accumulation will be output. Fig.12 , 13These two pictures are photos taken on site that we annotated manually using the segmentanything model in a pixel-level layer-by-pixel manner. The photos are annotated with positive and negative examples. The positive example is the dust-covered area, and the negative example is the normal area. They are saved as a dataset in CoCo format as the dataset required for training and testing. Then the robotic arm is controlled to move to the dust accumulation location, and the dust is cleaned using the soot blower on the robotic arm. A guide rail can be set on one side of the water-cooled wall, and the robotic arm is controlled to move on the guide rail, and the water-cooled wall is inspected and dust is cleaned through the guide rail.

[0074] The backbone network includes: a first convolution unit, a second convolution unit and a third attention mechanism unit; the first convolution unit includes Conv2d; the second convolution unit includes an 11-layer Mobilenetv3 network; the third attention mechanism unit includes an AWGAM attention mechanism.

[0075] In this embodiment, the first convolution unit includes a layer of CBH units, the convolution kernel is 3, the step size is 2, the pooling layer is 1, and the input 3*640*640 image is output as 8*320*320; the second convolution unit is configured with an 11-layer Mobilenetv3 network.The first layer is the Mobilenetv3 layer, the convolution kernel size is 3, the stride is 2, the pooling is 1, the hidden layer hidden_dim is 16, and the input 8*320*320 image is output as 8*160*160; the second layer is the Mobilenetv3 layer, the convolution kernel size is 3, the stride is 2, the pooling is 0, the hidden layer hidden_dim is 72, and the input 8*160*160 image is output as 8*80*80; the third layer is the Mobilenetv3 layer, the convolution kernel size is 3, the stride is 1, the pooling is 0, the hidden layer hidden_dim is 88, and the input 8*80*80 image is output as 8*80*80; The 4th layer is a Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 2, a pooling of 1, and a hidden layer hidden_dim of 96. The input image of 8*80*80 is output as 16*40*40; the 5th layer is a Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, and a hidden layer hidden_dim of 240. The input image of 16*40*40 is output as 16*40*40; the 6th layer is a Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, and a hidden layer hidden_dim of 240. The input image of 16*40*40 is output as 16*40*40; The 7th layer is a Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, and a hidden layer hidden_dim of 120. The input 16*40*40 image is output as 16*40*40; the 8th layer is a Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, and a hidden layer hidden_dim of 144. The input 16*40*40 image is output as 16*40*40; the 9th layer is a Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 2, a pooling of 1, and a hidden layer hidden_dim of 288. The input 16*40*40 image is output as 24*20*2 0; the 10th layer is a Mobilenetv3 layer, the convolution kernel size is 5, the stride is 1, the pooling is 1, the hidden layer hidden_dim is 576, and the input 24*20*20 image is output as 24*20*20; the 11th layer is a Mobilenetv3 layer, the convolution kernel size is 5, the stride is 1, the pooling is 1, the hidden layer hidden_dim is 576, and the input 24*20*20 image is output as 24*20*20; the third attention mechanism unit AWGAM has 24 input channels, 7 convolution kernels, and 3 pooling layers. The input 24*20*20 image is output as 24*20*20, and the output enters the Neck network.

[0076] The overall structure of the above-mentioned CBH unit consists of Conv2d, Batchnorm2d and Hardwish.

[0077] The network structure diagram of the above AWGAM attention mechanism is as follows Figure 5 As shown, the specific process is as follows:

[0078] The input image is processed by the channel attention mechanism and the spatial attention mechanism respectively;

[0079] Multiply the processed images by the corresponding scaling coefficients. The scaling coefficients are generated based on the amount of feature information extracted by the channel attention mechanism and the spatial attention mechanism. This can better improve the model's ability to extract features from images.

[0080] Stitch the final processed images together.

[0081] The Neck network includes: a first fusion network and a second fusion network. The first fusion network includes 2 groups of UCC networks, the first letter U is the Upsample network, the second letter C is the Concat network, and the third letter C is the C2f network. The second fusion network includes 2 groups of CCC networks, the first letter C is the CBS network, the second letter C is the Concat network, and the third letter C is the C2f network.

[0082] In this embodiment, the Upsample in the first group of UCC networks of the first fusion network outputs the input 24*20*20 image as 24*40*40, the Concat network combines the input 24*40*40 image and the 16*40*40 image output by the 8th layer of the second convolutional network of the backbone network to output 40*40*40, and the C2f network outputs the input 40*40*40 network as 64*40*40; the Upsampling in the second group of UCC of the first fusion network The ample outputs the input 64*40*40 image as 64*80*80. The Concat network combines the input 64*80*80 image and the 8*80*80 image output by the third layer of the second convolutional network of the backbone network to output 72*80*80. The C2f network outputs the input 72*80*80 network as 32*80*80, and inputs the output image of this layer into the first layer Detect detection head of the Head network. The first group of CCC networks of the second fusion network The CBS in the network converts the input 72*80*80 network output to 32*40*40. The Concat network combines the input 32*40*40 image and the 64*40*40 image output by C2f in the first group of UCC in the first fusion network in the Neck network to output 96*40*40. The C2f network converts the input 96*40*40 network output to 64*40*40, and inputs the image output by this layer into the second layer Detect detection head of the Head network. The CBS in the second group of CCC networks in the fusion network converts the input 64*40*40 network output to 64*20*20. The Concat network combines the input 64*20*20 image and the 24*20*20 image output by the third attention mechanism unit in the backbone network to output 88*20*20. The C2f network converts the input 88*20*20 network output to 128*20*20. The image output by this layer is input into the third layer Detect detection head of the Head network.

[0083] The overall structure of the above CBS unit consists of Conv2d, Batchnorm2d and SiLU.

[0084] The head network includes: 1 layer of Detect detection head. Detect is divided into two parts. The first part is to calibrate the target bounding box, which is divided into 4 layers. The first layer is CBS, the second layer is CBS, the third layer is Conv2d convolution layer, and the fourth layer is Bbox_loss bounding box loss function, which outputs the detection box of the target; the second part is to determine the target category, which is divided into 4 layers. The first layer is CBS, the second layer is CBS, the third layer is Conv2d convolution layer, and the fourth layer is Cls_loss classification loss function, which outputs the classification category of the target.

[0085] In this example, the 32*80*80 image output by C2f in the second group UCC network of the first fusion network in the Neck network is input into the first layer Detect detection head; the 64*40*40 image output by C2f in the first group CCC network of the second fusion network in the Neck network is input into the second layer Detect detection head; the 128*20*20 image output by C2f in the second group CCC network of the second fusion network in the Neck network is input into the third layer Detect detection head; then the target frames and target categories output by the three layers are merged, the target detection frame and target classification category with the maximum probability are selected, and finally the detected image is output.

[0086] The cleaning system is used to clean dust based on the dust accumulation position. The cleaning system includes: a robotic arm and a soot blower.

[0087] The mechanical arm is used to control the cleaning system to move to the dust accumulation position. The mechanical arm includes: a rotating motor. The working process of the mechanical arm includes: the position information is improved by the image recognition module, and then the rotating motor controls the mechanical arm to rotate so that the air blowing port of the soot blower is aimed at the dust accumulation on the water-cooled wall.

[0088] The soot blower uses high-pressure gas to clean the soot accumulated on the water-cooled wall.

[0089] In summary, in this example, we replaced the backbone feature extraction network of the YOLOv8n algorithm with the backbone network designed by ourselves based on its structure. At the same time, we also improved a new attention mechanism AWGAM based on the GAM attention mechanism, so that our backbone network can better extract feature information in the image. At the same time, we refer to the structural improvement design of the existing robotic arm, and equip the robotic arm with a soot blower, identify and locate the dust accumulation on the water-cooled wall through the dust accumulation detection model, and then control the rotation of the robotic arm according to the obtained identification distance result, so that the blowing port of the soot blower is aimed at the dust accumulation on the water-cooled wall, and the dust accumulation on the water-cooled wall is cleaned by the high-pressure gas of the soot blower. The present invention can control the operation of the robotic arm by detecting whether there is dust accumulation on the water-cooled wall, realize the intelligent cleaning of the water-cooled wall, and has higher cleaning efficiency and lower cleaning cost than the traditional water-cooled wall cleaning.

[0090] The embodiments described above are only descriptions of the preferred methods of the present application, and are not intended to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements made to the technical solutions of the present application by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present application.

[0091] 1. Introduction of the present invention:

[0092] One of the key evaluation indicators of a waste incinerator is the continuous operation time, which directly determines the economic benefits of the incinerator. my country's urban domestic waste contains a large amount of glass, ash, salt and other components, and the fly ash produced by its incineration is easy to burn and sinter under high temperature conditions. This forms a large amount of ash accumulation in the incinerator, which will bring a series of unfavorable factors to the incinerator. First, the ash will form an insulating layer on the furnace wall and the heat transfer surface, thereby hindering the transfer of heat from the high temperature area to the low temperature area and significantly reducing the heat transfer efficiency. Secondly, it reduces the effective volume of the furnace, reduces the processing capacity of the incinerator, and slows down the waste treatment speed. In addition, it will disrupt the airflow distribution, lead to abnormal combustion, increase the residence time of flue gas, and have an adverse effect on combustion efficiency and emission quality. In severe cases, the superheater pipe burst accident will force the incinerator to shut down for maintenance.

[0093] Traditional manual cleaning of water-cooled walls requires a lot of manpower and material resources. In addition, a large amount of dust will be generated during the cleaning process, which seriously affects the health of the workers. Therefore, it is very necessary to propose a safe, reliable and low-cost method for cleaning the ash on the surface of the water-cooled wall, which has important practical significance. In order to detect the ash accumulation on the water-cooled wall of the waste incinerator, we plan to use a high-temperature resistant robotic arm equipped with an imaging system. The target detection algorithm will be used to identify and locate the ash on the water-cooled wall. When a large area of ​​ash is detected, the robotic arm will be directed to clean it. Therefore, the identification and positioning of ash on the water-cooled wall is the key point of this study. The deep learning network can effectively extract the feature information on the image, and then complete the detection and classification of the target based on the extracted feature information. At present, the target detection algorithms based on deep learning are mainly divided into two categories. One is the single-stage target detection algorithm based on regression represented by YOLO 1 (You Only Look Once), SSD 2 There are two types of target detection algorithms: regression-based single-stage target detection algorithms represented by SingleShot MultiBox Detector and others. Another type is the two-stage target detection algorithm based on candidate regions represented by R-CNN. 3 (Regions with Convolutional NeuralNetwork features)、SPP-Net 4 (Spatial Pyramid Pooling Network) etc.

[0094] The YOLO series of algorithms is a single-stage target detection algorithm proposed by Redmon et al. in 2016. The initial version was YOLOv1, which has a very fast detection speed. As soon as it appeared, it solved the problem of slow detection speed of two-stage detection networks such as R-CNN, and also met the requirements of real-time detection. But its disadvantage is that it has poor detection ability for small targets. Subsequently, a series of iterative versions of YOLO continued to appear, and the target detection accuracy and speed of the model have been improved. Among them, YOLOv8n is a very classic single-stage target detection algorithm with fewer parameters and shorter inference time. It is particularly suitable for deployment on resource-constrained devices, such as robotic arms 5 .

[0095] The YOLOv8n algorithm is a good algorithm framework, not a specific algorithm, which is very helpful for us to improve the algorithm for different detection targets. Zhang 6 et al. improved the YOLOv8n algorithm for automatic detection and segmentation of tunnel defects and objects. 7 Wang et al. used the improved YOLOv8n algorithm to detect underground dangerous areas. 8 et al. used the improved YOLOv8n algorithm to detect road surface defects so as to repair damaged roads in time and ensure road safety. 9 et al. improved the YOLOv8n algorithm to achieve real-time monitoring of urban fires. This speeds up response and reduces the risk of property and life loss.

[0096] Using data augmentation algorithms to enhance detailed information of images is an effective way to improve the accuracy of model training. 10 et al. designed an adaptive color correction method for enhancing underwater images. It improves the clarity of the image and further improves the accuracy of the underwater object detection model. 11 et al. used an image enhancement technique to improve the clarity and contrast of the image, thereby improving the detection accuracy of the model. 12 et al. used the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm to process images, which effectively improved the clarity and contrast of the images and enriched the image details. 13 et al. used a deep Laplacian pyramid network to achieve super-resolution reconstruction of images and improve image quality.

[0097] This paper proposes an improved YOLOv8n algorithm for detecting the location of water-cooled wall ash. The main work and contributions are as follows.

[0098] A multi-scale image fusion algorithm is proposed to improve the clarity and contrast of the image and reduce the impact of environmental factors such as lighting on the image.

[0099] Mobilenetv3 replaces the backbone feature extraction network of YOLOv8n, reducing the number of parameters and model size while ensuring model accuracy.

[0100] Based on the GAM attention mechanism, a new attention mechanism AWGAM is proposed. Then, AWGAM is added to the backbone network of YOLOv8n, which effectively improves the generalization ability and accuracy of the model.

[0101] Experiments show that our algorithm outperforms other algorithms in terms of parameter quantity, precision, recall, mAP50, and mAP50-95.

[0102] 2. Experimental results and analysis of the present invention

[0103] 2.1 Model Training

[0104] When training the water-cooled wall dust detection network model, the size of the input image is uniformly adjusted to 640×640×3. The AdamW (Adam with Weight Decay Fix optimizer) is used and trained 600 times in total. The batch size is 16 and the initial learning rate is 0.01. In addition, in order to improve the accuracy of the model, Mosaic is turned off in the model. 14 Data enhancement technology. The training results are as follows Figure 1 As shown in the figure. "box_loss" represents the difference between the predicted bounding box generated by the model and the actual bounding box. "cls_loss" (classification loss) is used to measure the performance of the model in classification tasks and is optimized during training. "dfl_loss" (distribution focus loss) is a loss function used to deal with class imbalance problems and can better handle class imbalance situations.

[0105] 2.2 Image enhancement algorithm based on multi-scale fusion

[0106] In order to prove the effectiveness of the image enhancement algorithm proposed in this paper, we compared the original image with the image enhanced by different image enhancement algorithms. The specific experimental comparison results are shown in Table 1. It can be clearly seen from Table 1 that our algorithm performs much better than other algorithms in terms of mAP50, mAP50-95 and accuracy. This fully proves the effectiveness of the image enhancement algorithm proposed in this paper. Image comparison of different image enhancement algorithms Figure 2 shown.

[0107] Table 1 Performance comparison of different image enhancement algorithms

[0108]

[0109] 2.3AWGAM Attention Mechanism

[0110] In order to verify the effectiveness of the AWGAM attention mechanism, this paper adds AWGAM and GAM to the model network for comparative experiments. The experimental results are shown in Table 2, where Params is the algorithm parameter in M ​​megabytes).

[0111] Table 2 Comparison of experimental results

[0112]

[0113] Experimental results show that the improved AWGAM-Mobilenetv3 network has higher detection accuracy. Compared with GAM-Mobilenetv3, the accuracy is improved by 7.6%, the recall rate is improved by 2.0%, the mAP50 is improved by 2.3%, and the mAP50-95 is improved by 3.5%. This fully proves the effectiveness of AWGAM.

[0114] 2.4 Ablation experiments on each part

[0115] The main improvements of this paper to the YOLOv8n model include replacing the backbone feature extraction network of YOLOv8n with Mobilenetv3 and integrating the AWGAM attention mechanism into the backbone network of YOLOv8n. In order to verify the impact of each component, this paper conducted an ablation experiment.

[0116] As can be seen from Table 2, the original YOLOv8n model has a precision of 94.1%, a recall of 93.5%, and an average precision mAP50 of 98.9% on the enhanced dataset.

[0117] After replacing the backbone feature extraction network with Mobilenetv3, the number of model parameters was reduced by 60.5%, the precision was reduced by 5.8%, the recall was increased by 3.1%, and the mAP50 was reduced by 2.2%.

[0118] After adding the AWGAM attention mechanism to the YOLOv8n backbone network, the precision is improved by 3.5%, the recall rate is improved by 3.2%, and the mAP50 is reduced by 0.1%.

[0119] Finally, by applying the above two models, the number of parameters is reduced by 59.9%, the precision is improved by 3.9%, the recall is improved by 3.2%, and the mAP50 is improved by 0.1%.

[0120] In conclusion, the improved model not only reduces the number of parameters but also improves the accuracy of the model. The results show that the performance of the model is better.

[0121] 2.5 Comparative experiments of various models

[0122] In order to verify the effectiveness of the improved model, this paper compares it with YOLOv5n, YOLOv6n, YOLOv7, YOLOv9-t and YOLOv10n algorithms. The experimental results are shown in Table 3. It can be clearly seen that the proposed algorithm is more complete and superior to other algorithms in all aspects. It fully proves the effectiveness of the improved algorithm.

[0123] Table 3 Performance comparison of different algorithms

[0124] Accuracy (%) Recall rate (%) mAP50(%) mAP50-95(%) Parameters / M YOLOv5n 96.9 89.0 94.7 79.7 1.8 YOLOv6n 94.8 93.5 96.7 90.2 4.7 YOLOv7 76.1 86.1 90.0 74.6 37.2 YOLOv8n 93.8 84.2 96.0 83.7 3.0 YOLOv9-t 91.8 96.7 96.6 91.0 26.2 YOLOv10n 94.8 89.8 98.0 87.4 2.7 This algorithm 98.0 96.7 99.0 95.4 1.2

[0125] 2.6 Analysis of the detection algorithm results of the present invention

[0126] In order to verify the detection effect of the model, we randomly selected some images from the test set and the network to test the accuracy of the model. The test results are as follows: Figure 3 As shown. The above picture shows three pictures of other water-cooled walls we found on the Internet, and the following picture shows three pictures in our test set. In the picture, "guanandao_dusk" means there is dust on the water-cooled wall, and "guanandao_nodusk" means there is no dust on the water-cooled wall. Figure 3 The value shown in represents the confidence of the detected target. When the confidence is close to 1, the model will become more and more confident that there is a target object in the box. If the confidence is close to 0, the model will believe that there is no recognizable target within the range of the box. Figure 3 As shown in the figure, the confidence level of the detection results is significantly higher, indicating that the model is able to skillfully identify dust deposits on the water-cooled wall. This finding highlights the significance of this study.

[0127] 3. Discussion on the technical solution of the present invention

[0128] Our research introduced an algorithm called AWGAM-YOLOv8n, which is specifically used to detect dust accumulation on water-cooled walls, and pioneered the application of target detection technology for this purpose. Compared with the traditional method of manually cleaning water-cooled walls, our method is smarter and more convenient. If this technology is applied to robotic arms, a lot of manpower costs can be saved. At the same time, it also plays a vital role in protecting the health of employees. The backbone network of YOLOv8n is replaced by Mobilenetv3, and a new attention mechanism-AWGAM is developed based on the GAM attention mechanism, which is incorporated into the backbone network of YOLOv8n. Compared with the application of ash recognition technology on solar panels, our model parameters are smaller 15, which is only 1.2 megabytes. Our algorithm also performs well in other aspects. The proposed algorithm achieves a precision of 98.0%, a recall of 96.7%, and a mAP50 of 99.0%. It is proven that the proposed algorithm has excellent performance in detecting dust accumulation on water cooling walls.

[0129] In future work, we will try to improve the model and increase its reasoning speed, such as using knowledge distillation to compress the model, and try to embed the algorithm into the robotic arm control system to achieve full intelligence and automation of water-cooled wall cleaning.

[0130] 4. About the detection method of the present invention

[0131] 4.1 Dataset Creation

[0132] The dataset of this article was created by us. We used a camera to collect ash samples from the water-cooled wall, and then took corresponding sample photos based on the recorded video, totaling 300 photos. Due to the simple characteristics of the water-cooled wall ash, the number of effective samples obtained is small. In order to further expand the dataset, first, we randomly select pictures from the original dataset and rotate them 45 degrees. Secondly, crop the images in the original dataset to 640×640 size. Third, randomly grayscale the images in the dataset. Finally, we expand the dataset to 700 pictures. Then use Labelme 16 The software is used for annotation, and the annotated images are divided into training set, validation set and test set in the ratio of 8:1:1. Figure 4 A sample of some of the images collected are shown.

[0133] 4.2 Evaluation Metrics

[0134] This experiment uses the Pytorch-GPU 2.0.0 deep learning framework, the processor is: 13th Gen Intel(R) Core(TM) i5-13600KF 3.50GHz, the graphics card is NVIDIA GeForce RTX 4060, the operating system is Windows 11 Professional Edition, the NVIDIA driver version is 31.0.15.4633, the CUDA version is 11.8, and the cuDNN version is 8.7.0.

[0135] This paper uses precision, recall, average precision (mAP) and parameters (Params) as evaluation criteria for model performance.

[0136] Precision is mainly used to measure the percentage of correctly predicted targets; high precision can reduce the probability of false detection. The calculation formula is shown in formula (1).

[0137]

[0138] TP (True Positive) represents the true target, that is, the number of positive samples correctly predicted as positive samples by the model. FP (False Positive) represents the false target, that is, the number of negative samples incorrectly predicted as positive by the model.

[0139] Recall rate refers to the ratio of the number of samples that can be correctly predicted as positive to the number of actual positive samples. The calculation formula is shown in formula (2).

[0140]

[0141] FN (false negative) indicates the number of negative samples that are mistakenly identified as positive samples.

[0142] As shown in formula (3), the average precision of multiple categories is calculated.

[0143]

[0144] mAP50 represents the average precision of the model when the IoU critical value is 50.0%. mAP50-95 represents the average precision of the IoU critical value in units of 5.0%, from 50.0% to 95.0%.

[0145] Params are the number of adjustable parameters that need to be trained in the model. In the field of deep learning, the number of parameters is usually one of the indicators to measure the size and complexity of the model. The more parameters there are, the more computing resources are required to train and execute the model.

[0146] 4.3 Image enhancement algorithm based on multi-scale fusion in the present invention

[0147] 4.3.1CLAHE Algorithm

[0148] When collecting images on a water-cooled wall, the local brightness of the collected image will be very high due to the need for lighting. CLAHE can effectively alleviate the problem of high local brightness, improve image clarity, and enrich image detail information.

[0149] The principle of CLAHE is to set a threshold for the histogram. This threshold is called the extreme contrast value, and the values ​​above this threshold will be clipped. The clipped parts are then evenly distributed to other values ​​to reconstruct the histogram. 17 In this paper, the CLAHE method is used for image enhancement.

[0150] 4.3.2 Gamma Correction

[0151] In order to further improve the overall brightness of the image, this paper uses the gamma correction algorithm to process the image. Gamma correction is an image enhancement method that performs nonlinear operations on the grayscale values ​​of the input image. It makes the grayscale value of the output image exponentially related to the grayscale value of the input image. The expression of the gamma correction principle is shown in formula (4).

[0152] f(x)=x γ (4)

[0153] f(x) is the output image; x is the normalized image; γ is the image enhancement parameter.

[0154] 4.3.3 Laplace Pyramid Hybrid Method

[0155] In order to better improve the clarity and contrast of the image and ensure the integrity of the image detail information, this paper uses the Laplacian pyramid mixing method to mix the images. This method can effectively fuse two different images and retain the detail information between different images. 18 The principle is as follows: First, downsample the two images that need to be Laplace fused. Second, upsample the downsampled image. Third, calculate the difference between the processed image and the original image, and adjust the residual to between 0 and 255. Fourth, calculate the sum of the residuals of each layer of the pyramid. Finally, starting from the top of the pyramid, add up the total residuals of each layer, and after each layer, the image pixels need to be between 0 and 255. The principle of Laplace pyramid fusion is as follows Figure 5 shown.

[0156] 4.3.4 Algorithm Flow

[0157] In summary, the image enhancement algorithm in this paper is Figure 6 As shown in the figure. First, gamma correction is performed on the image to enhance the brightness of the image. Then the image is processed by adaptive histogram equalization to enhance the local detail information of the image and enhance the local contrast of the image. Then the two processed images are fused using the Laplacian pyramid fusion method to obtain the final enhanced image.

[0158] 4.4AWGAM-YOLOv8n Algorithm

[0159] In order to enhance the detection capability of ash in the water-cooled wall, we developed a new algorithm AWGAM-YOLOv8n based on the YOLOv8n framework. We replaced the backbone feature extraction network of YOLOv8n with Mobilenetv3, which significantly reduced the model size. In addition, in order to improve the feature extraction capability of the model, we integrated AWGAM into the YOLOv8n algorithm. Experimental results show that this method can quickly and accurately identify whether there is ash accumulation in the water-cooled wall. The structure of the AWGAM-YOLOv8n algorithm is shown in the figure. Figure 7 shown.

[0160] 4.5 Mobilenetv3 Network

[0161] Compared with heavyweight networks, lightweight networks have the advantages of fewer parameters, smaller models, shorter inference time, and faster running speed. They are suitable for scenarios with limited storage space and power consumption, and are more suitable for the working scenarios of the robotic arm in this article. Commonly used lightweight networks include Mobilenetv3 19 、Shufflenet 20 、Ghostnet 21 In order to select a lightweight network suitable for the detection target of this paper, this paper adds these three lightweight networks to the YOLOv8n model for comparative testing. The results are shown in Table 4, where FLOPs (floating point operations) is used to measure the complexity of the algorithm, in G gigabytes). The core modules of Mobilenetv3 are as follows Figure 8 shown.

[0162] According to the data in Table 4, the Mobilenetv3 network is the smallest among the three network models. In this study, Mobilenetv3 is used as the backbone network for extracting features. 22 、Mobilenetv2 23 It is optimized and improved on the basis of Mobilenetv1. It inherits the depthwise separable convolution of Mobilenetv1 and the inverted residual and linear bottleneck structure of Mobilenetv2. The depthwise separable convolution can reduce the parameters of the model and reduce the complexity of the model. The inverted residual and linear bottleneck structure can improve the performance of the model. In addition, Mobilenetv3 also uses NAS (neural structure search) to obtain the optimal parameters, thereby improving the overall efficiency. In order to further enhance the feature representation capability, the SE channel attention module is also integrated into the model.

[0163] Table 4 Comparison of different lightweight networks

[0164] network Parameters / M FLOPs / G Mobilenetv3 1.2 2.8 Shufflenetv2 1.7 5.0 Ghostnet 2.3 6.5

[0165] 4.6 AWGAM Attention Mechanism

[0166] In order to further improve the algorithm's ability to extract image features, integrating the attention mechanism into the backbone network has proven to be a very effective strategy.

[0167] GAM 24The channel attention mechanism is effectively combined with the spatial attention mechanism. It enables the channel and space to influence each other, amplifies the global interaction ability, and thus improves the performance of deep neural networks. This method has the advantages of high efficiency, strong robustness, and wide applicability, but GAM ignores the inconsistency of feature extraction in different dimensions. Only combining channel with spatial attention may lead to serious loss of feature information.

[0168] Recognizing the inconsistency of image feature dimensional representation, we propose AWGAM, a new enhancement method based on GAM, which can adjust the feature extraction ratio of different dimensions more effectively than before by introducing weight parameters. Therefore, AWGAM improves the efficiency of image feature extraction, thereby improving the overall performance of the model. Fig. 9 shown.

[0169] 5 Reference Documents

[0170] [1]. Redmon, J., Divvala, S., Girshick, R. & Farhadi, A. You only look once: Unified, real-time object detection. in Proceedings of the IEEE conference on computer vision and pattern recognition. 779-788 (2016).

[0171] [2].Liu,W.et al.Ssd:Single shot multibox detector.in Computer Vision–ECCV 2016:14th European Conference,Amsterdam,The Netherlands,October 11–14,2016,Proceedings,Part I 14.21-37(2016).

[0172] [3].Cai, Z. & Vasconcelos, N. Cascade r-cnn: Delving into high quality object detection. in Proceedings of the IEEE conference on computer vision and pattern recognition. 6154-6162 (2018).

[0173] [4].He,K.,Zhang,X.,Ren,S.&Sun,J.Spatial pyramid pooling in deepconvolutional networks for visual recognition.IEEE transactions on patternanalysis and machine intelligence 37,1904-1916(2015).

[0174] [5].Zeng,J.&Zhong,H.YOLOv8-PD:an improved road damage detectionalgorithm based on YOLOv8n model.Scientific reports 14,12052(2024).

[0175] [6].Zhang,C.et al.Automated detection and segmentation of tunneldefects and objects using YOLOv8-CM.Tunnelling and Underground SpaceTechnology 150,105857(2024).

[0176] [7].Ni,Y.,Huo,J.,Hou,Y.,Wang,J.&Guo,P.Detection ofUndergroundDangerous Area Based on Improving YOLOV8.Electronics 13,623(2024).

[0177] [8].Wang,X.,Gao,H.,Jia,Z.&Li,Z.BL-YOLOv8:An improved road defectdetection model based on YOLOv8.Sensors 23,8361(2023).

[0178] [9].Talaat,F.M.&ZainEldin,H.An improved fire detection approach basedon YOLO-v8 for smart cities.Neural Computing and Applications 35,20939-20954(2023).

[0179]

[10] .Ke,K.et al.Underwater image enhancement via color correction andmulti-feature image fusion.Measurement Science and Technology 35(2024).

[0180]

[11] .Gupta,C.,Gill,N.S.,Gulia,P.,Yadav,S.&Chatterjee,J.M.A novelfinetuned YOLOv8 model for real-time underwater trash detection.Journal ofReal-Time Image Processing 21,48(2024).

[0181]

[12] .Ulutas,G.&Ustubioglu,B.Underwater image enhancement usingcontrast limited adaptive histogram equalization and layered differencerepresentation.Multimedia Tools and Applications 80,15067-15091(2021).

[0182]

[13] .Lai,W.-S.,Huang,J.-B.,Ahuja,N.&Yang,M.-H.Fast and accurate imagesuper-resolution with deep laplacian pyramid networks.IEEE transactions onpattern analysis and machine intelligence 41,2599-2613(2018).

[0183]

[14] .TAnM,P.&Efficientdet,L.Scalableandefficientobjectdetection.IEEE / CVFConferenceonComputerVi sionandPatternRecognition.Piscataway,USA:IEEE(2020).

[0184]

[15] .Su,R.,Wang,L.,Wang,Q.,Wang,F.&Zhang,Z.Research on theidentification method of photovoltaic panel ash accumulation state based onMobileViT model.Mechanical and Electrical Information 21-25,doi:10.19514 / j.cnki.cn32-1628 / tm.2023.14.005(2023).

[0185]

[16] .Torralba,A.,Russell,B.C.&Yuen,J.Labelme:Online image annotationand applications.Proceedings of the IEEE 98,1467-1484(2010).

[0186]

[17] .Chang,Y.,Jung,C.,Ke,P.,Song,H.&Hwang,J.Automatic contrast-limited adaptive histogram equalization with dual gamma correction.IeeeAccess 6,11782-11792(2018).

[0187]

[18] .Zeng,Q.et al.Detection of Coal and Gangue Based on ImprovedYOLOv8.Sensors 24,1246(2024).

[0188]

[19] .Howard,A.et al.Searching for mobilenetv3.in Proceedings of theIEEE / CVF international conference on computer vision.1314-1324(2019).

[0189]

[20] .Zhang,X.,Zhou,X.,Lin,M.&Sun,J.Shufflenet:An extremely efficientconvolutional neural network for mobile devices.in Proceedings of the IEEEconference on computer vision and pattern recognition.6848-6856(2018).

[0190]

[21] .Han,K.et al.Ghostnet:More features from cheap operations.inProceedings of the IEEE / CVF conference on computer vision and patternrecognition.1580-1589(2020).

[0191]

[22] .Andrew,G.&Menglong,Z.Efficient convolutional neural networks formobile vision applications,mobilenets.arXiv preprint arXiv:1704.04861(2017).

[0192]

[23] .Howard,A.,Zhmoginov,A.,Chen,L.-C.,Sandler,M.&Zhu,M.Invertedresiduals and linear bottlenecks:Mobile networks for classification,detectionand segmentation.in Proc.CVPR.4510-4520(2018).

[0193]

[24] .Liu,Y.,Shao,Z.&Hoffmann,N.Global attention mechanism:Retaininformation to enhance channel-spatial interactions.arXiv preprint arXiv:2112.05561(2021)。

Claims

1. A water-cooled wall dust accumulation detection system based on computer vision, characterized in that: It includes an image acquisition module and a dust accumulation recognition module; The image acquisition module is used to acquire water-cooled wall images; The dust accumulation recognition module is used to detect the water-cooled wall image and obtain the dust accumulation position; The dust accumulation recognition module includes: an image preprocessing network, a backbone network, a neck network and a head network; The image preprocessing network is used to perform image enhancement processing on the water-cooled wall image to obtain a processed image; the backbone network is used to extract different levels of feature information of the processed image; the neck network is used to perform feature fusion on the different levels of feature information to obtain a fused image; the head network is used to perform target prediction based on the fused image to obtain the dust accumulation position.

2. The water-cooled wall dust accumulation detection system based on computer vision according to claim 1 is characterized in that: The image preprocessing network performs adaptive histogram enhancement and gamma correction on the image, and then fuses the processed images into one image through a Laplacian pyramid.

3. The water-cooled wall dust accumulation detection system based on computer vision according to claim 1, characterized in that: The backbone network has a total of 13 layers of network structure: The first layer is CBH, which contains a Conv2d layer, a Batchnorm2d and an activation function Hardwish, with 3 convolution kernels, 2 strides, 1 pooling layer, 3 input channels, and 8 output channels; The second layer is the Mobilenetv3 layer, with a convolution kernel size of 3, a stride of 2, a pooling of 1, an input channel of 8, an output channel of 8, and a hidden layer hidden_dim of 16; The third layer is the Mobilenetv3 layer, with a convolution kernel size of 3, a stride of 2, a pooling of 0, an input channel of 8, an output channel of 8, and a hidden layer hidden_dim of 72; The fourth layer is the Mobilenetv3 layer, with a convolution kernel size of 3, a stride of 1, a pooling of 0, an input channel of 8, an output channel of 8, and a hidden layer hidden_dim of 88; The fifth layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 2, a pooling of 1, an input channel of 8, an output channel of 16, and a hidden layer hidden_dim of 96; The sixth layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 240; The 7th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 240; The 8th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 120; The 9th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 16, an output channel of 16, and a hidden layer hidden_dim of 144; The 10th layer is the Mobilenetv3 layer, with a convolution kernel size of 5, a stride of 2, a pooling of 1, an input channel of 16, an output channel of 24, and a hidden layer hidden_dim of 288; The 11th layer is a Mobilenetv3 layer with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 24, an output channel of 24, and a hidden layer hidden_dim of 576; The 12th layer is a Mobilenetv3 layer with a convolution kernel size of 5, a stride of 1, a pooling of 1, an input channel of 24, an output channel of 24, and a hidden layer hidden_dim of 576; The 13th layer is the AWGAM attention mechanism layer, with 24 input channels, 7 convolution kernels, 3 pooling layers, and an output feature size of 20*20*576.

4. The water-cooled wall dust accumulation detection system based on computer vision according to claim 1, characterized in that: The neck network has a total of 12 layers of network structure: The first layer is the upsample layer, which takes the 13th layer output of the backbone network as input, and the output feature size is 40*40*24; The second layer is the concat layer, which takes the output of the first layer and the 9th layer of the backbone network as input, and the output feature size is 40*40*40; The third layer is C2f, which takes the output of the second layer as input, and the output feature size is 40*40*64. Bottleneck has shortcut; The fourth layer is the upsample layer, which takes the output of the third layer as input and the output feature size is 80*80*64; The fifth layer is the concat layer, which takes the output of the fourth layer and the fourth layer of the backbone network as input, and the output feature size is 80*80*72; The sixth layer is the C2f layer, which takes the output of the fifth layer as input and outputs a feature size of 80*80*32. Bottleneck has shortcut; The 7th layer is the Conv layer, which takes the output of the 6th layer as input, has a convolution kernel of 3, a stride of 2, a pooling layer of 1, and an output feature size of 40*40*32; The 8th layer is the concat layer, which takes the output of the 7th and 3rd layers as input, and the output feature size is 40*40*72; The 9th layer is the C2f layer, which takes the output of the 8th layer as input and the output feature size is 40*40*64. Bottleneck has shortcut; The 10th layer is the Conv layer, which takes the output of the 9th layer as input, has a convolution kernel of 3, a stride of 2, a pooling layer of 1, and an output feature size of 20*20*64; The 11th layer is the concat layer, which takes the output of the 10th layer and the 13th layer of the backbone network as input, and the output feature size is 20*20*88; The 12th layer is the C2f layer, which takes the output of the 11th layer as input. The output feature size is 20*20*128, and Bottleneck has a shortcut.

5. The water-cooled wall dust accumulation detection system based on computer vision according to claim 4 is characterized in that: The neck network has a total of 12 layers of network structure: The head network has a total of 3 layers of network structure: The first layer is the Detect layer, which takes the output of the sixth layer of the neck network as input; The second layer is the Detect layer, which takes the output of the 9th layer of the neck network as input; The third layer is the Detect layer, which takes the output of the 12th layer of the neck network as input.

6. According to the water-cooled wall dust accumulation detection system based on computer vision in claim 2, the adaptive histogram enhancement is to adaptively equalize the R, G, and B channels in the image respectively, and then merge the enhanced R, G, and B channels together; The gamma correction is to normalize the image and then perform nonlinear operation on the grayscale value of the image. It can make the grayscale value of the output image have an exponential relationship with the grayscale value of the input image. The formula is f(x)=x γ 。 7. According to the water-cooled wall dust accumulation detection system based on computer vision in claim 2, the Laplacian pyramid fusion: first, down-sampling the two images that need Laplacian fusion; Secondly, the downsampled image is upsampled; Third, calculate the difference between the processed image and the original image, and adjust the residual to be in the range of 0-255; Fourth, calculate the sum of the residuals of each layer of the pyramid; Finally, the total residual of each layer is added starting from the top of the pyramid. After each layer, it is necessary to make the value of the image pixels between 0-255.

8. A water-cooled wall dust cleaning device, characterized in that: It comprises the dust accumulation detection system as described in any one of claims 1-7, and also comprises a cleaning module for cleaning the dust accumulation.

9. The water-cooled wall dust cleaning device according to claim 8, characterized in that: The cleaning module comprises a mechanical arm and a soot blower installed on the mechanical arm.

10. The water-cooled wall dust cleaning device according to claim 8, characterized in that: The image acquisition device includes a 360° control pan-tilt fixed on a mechanical arm, a camera installed on the control pan-tilt and a pan-tilt controller. The pan-tilt controller is used for the 360° control pan-tilt to drive the high-definition visible light camera to rotate and acquire the image of the water-cooled wall in the front field of view.

Images