A method for detecting corrosion of boiler heating surface

Abstract

The present invention relates to a method for detecting corrosion of boiler heating surfaces. Step 100: Parallelly decode the original video using a GPU to obtain grayscale images for each frame. Step 200: Divide the i-th frame image into N*N image blocks; parallelly perform DCT transformation on each image block using a GPU. Step 300: Parallelly convert the DCT coefficients of each image block into column vectors, and input each column vector into a trained improved BP neural network in sequence to obtain the determination result of whether there is corrosion in each image block. If there is corrosion in a certain image block, jump to Step 400; if there is no corrosion in all image blocks of the i-th frame image, let i = i + 1 and return to Step 200. Step 400: Input the i-th frame image into a trained improved Yolov5 model for corrosion detection to determine the specific location of corrosion, let i = i + 1 and return to Step 200. The accuracy of detecting the corrosion condition of the boiler heating surface is improved while reducing the amount of computation.

Description

Technical Field

[0001] The present invention relates to the technical field of boilers, and particularly relates to a method for detecting corrosion of boiler heating surfaces. Background Art

[0002] The pipes of the heating surfaces of boilers in thermal power generating units (mainly referring to boiler water walls, superheaters, reheaters and economizers) operate under high temperature and high pressure for a long time. The flue gas corrosion on their outer walls will cause the wall thickness of the pipes to decrease. When the thickness of the corroded pipe reaches the limit thickness under the current pressure, pipe explosion will occur, resulting in the non-stop operation of the unit, which seriously affects the safe operation of the boiler.

[0003] In the prior art, it is often necessary to conduct visual inspections on each pipe of the heating surfaces such as boiler water walls, superheaters, reheaters and economizers, and use an ultrasonic thickness gauge to detect the degree of corrosion. The above-mentioned manual methods have disadvantages such as being prone to missing corrosion locations, low detection accuracy, and low efficiency. Therefore, how to detect the corrosion conditions and specific locations of the heating surfaces such as boiler water walls, superheaters, reheaters and economizers with high efficiency and high accuracy, and how to reduce the calculation amount and improve the calculation speed are technical problems that urgently need to be solved. Summary of the Invention

[0004] In order to overcome the problems existing in the above-mentioned prior art, the present invention provides a method for detecting corrosion of boiler heating surfaces. First, it uses an improved BP neural network with a smaller calculation amount to detect whether there is corrosion in an image block. When corrosion exists in a certain image block, then the entire image is input into a trained improved Yolov5 model for corrosion detection to obtain the specific locations where the boiler heating surfaces are corroded in the entire image. This method can minimize the calculation amount while improving the detection accuracy. At the same time, the present application utilizes the parallel computing of the GPU to improve the operation speed.

[0005] The first aspect of the embodiments of the present application proposes a method for detecting corrosion of boiler heating surfaces, including the following steps:

[0006] Step 100: Photograph the boiler heating surface, transmit the original video output by the high-definition camera into the video memory of the graphics card, and the GPU in the graphics card performs parallel decoding on the original video to obtain each frame of grayscale image;

[0007] Step 200: Divide the i-th frame of grayscale image into N*N image blocks to obtain sum identical-sized image blocks; perform DCT transformation on each image block in parallel using the GPU; and obtain the DCT transformation coefficient matrix of each image block;

[0008] Step 300: Use GPU parallelism to convert the DCT transform coefficient matrix of each image block into an (N*N)×1-dimensional column vector, and input each column vector into the trained improved BP neural network in sequence to obtain the determination result of whether there is corrosion in each image block. If the determination result of the jth column vector corresponding image block is that there is corrosion, where j∈[1,sum], then jump to Step 400; if the determination results of all image blocks of the ith grayscale image are that there is no corrosion, let i = i + 1, and return to Step 200;

[0009] Step 400: Input the ith grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location of corrosion on the boiler heating surface in the ith frame. Let i = i + 1, and return to Step 200.

[0010] The process of inputting each column vector into the trained improved BP neural network in sequence to obtain the determination result of whether there is corrosion in each image block includes:

[0011] Step 310: Use training sample set 1 to train the improved BP neural network to obtain the trained improved BP neural network. The specific process is as follows:

[0012] Set the training termination criteria: one is to reach the specified maximum number of iterations T, the second is that the error function value is less than the specified standard ε, and the third is that the weight change amount is less than ε;

[0013] Set the training pause criteria: when the training termination criteria are not met;

[0014] Training process:

[0015] Step1: Initialize the network structure, termination and pause criteria, and other network parameters;

[0016] Step2: Start training. If the termination criteria are met, the training process ends. If the training pause criteria are met, go to Step3 for pruning;

[0017] Step3: If the pause criteria are met, then start pruning, calculate the minimum longitudinal grey relational degree, use it to determine the pruning connection. If the minimum longitudinal grey relational degree < ε, then take its corresponding connection as the pruning connection of the output neuron, and go to Step4. If they are all not less than the defined value, then go to Step2 and continue training;

[0018] Step4: Select the corresponding merging connection for each pruning connection. Using the network horizontal grey relational analysis method, calculate the maximum network horizontal grey relational degree of the input value sequence, and the merging connection is its corresponding connection;

[0019] Step 5: Pruning connection deletion, modifying the weights of the merged connections, and adjusting the weights using weighted averaging. Adjust the rated correlation degree ε = ε + k, where k is a constant selected according to the specific problem.

[0020] Step 6: Pause the pruning process, go to Step 2, and continue training.

[0021] Step 7: Determine the network structure and end the training process.

[0022] Step 320: Use the trained improved BP neural network to determine whether to send corrosion for each column vector.

[0023] Input the i-th frame grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location where the boiler heating surface is corroded in the i-th frame. The specific steps are as follows:

[0024] Step 410: Improve the Yolov5 model: In the Yolov5 model, retain the ordinary convolutions in the backbone feature extraction network, and replace the ordinary feature layers in the neck network with FSConv convolutions. The specific implementation method of the FSConv convolution is as follows: First, generate high-dimensional features through ordinary convolutions, then use depthwise separable convolutions to transform the high-dimensional features, then splice the two features obtained in the above two steps, and through a shuffle operation, shuffle the feature maps of different groups together to enhance the information interaction between different groups; Improve the CA attention structure used in Yolov5 to obtain the improved CA attention structure ICA. The specific implementation method is: Add a branch to the CA attention, use max pooling to extract the key features of this part, and merge with the average pooling branch at the end. The two branches jointly act on the output feature map to strengthen the final classification and localization effects.

[0025] Step 420: Use the training sample set 2 to train the improved Yolov5 model to obtain the trained improved Yolov5 model.

[0026] Step 430: Input the i-th frame grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location where the boiler heating surface is corroded in the i-th frame.

[0027] According to an embodiment of the present application, after obtaining the specific location where the boiler heating surface is corroded, it further includes: obtaining the actual location where the boiler heating surface is corroded according to the specific location where the boiler heating surface is corroded in the image, and sending a reminder to measure the pipeline at the actual location where the boiler heating surface is corroded.

[0028] According to an embodiment of the present application, it further includes: detecting the actual thickness of the pipeline at the actual position where the boiler heating surface is corroded by an ultrasonic thickness gauge. If the actual thickness is greater than or equal to the first preset thickness, the corrosion level of the pipeline at the actual position where the boiler heating surface is corroded belongs to mild corrosion and does not require treatment, and a prompt of "mild corrosion, continuous attention" is output; if the actual thickness is less than the first preset thickness and greater than or equal to the second preset thickness, the corrosion level of the pipeline at the actual position where the boiler heating surface is corroded belongs to moderate corrosion, and the operation and maintenance personnel are reminded to optimize the operation control; if the actual thickness is less than the second preset thickness, the corrosion level of the pipeline at the actual position where the boiler heating surface is corroded belongs to severe corrosion, and the operation and maintenance personnel are reminded to conduct a metallographic structure test, replace the pipeline at the corroded position of the boiler heating surface, and optimize the operation control; wherein, the first preset thickness is greater than the second preset thickness.

[0029] By adopting the above technical solution, the corrosion detection of the boiler heating surface is realized.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] First, the image is divided into blocks and the DCT transform coefficients of each image block are used as the input of the improved BP neural network, so as to better mine the frequency domain information of the image; second, the improved BP neural network with less computational complexity is used to detect whether there is corrosion in the image block, and when corrosion exists in a certain image block, the judgment on whether the subsequent image blocks are corroded is terminated in time, and then the whole image is input into the trained improved Yolov5 model for corrosion detection to obtain the specific position where the boiler heating surface is corroded in the whole image, so as to minimize the computational complexity while improving the detection accuracy. It should be noted that although the Yolov5 model will also judge whether there is corrosion, in this application, the BP neural network with less computational complexity is first used to preliminarily judge whether the boiler heating surface is corroded, and when there is no corrosion, the Yolov5 model with large computational complexity is avoided for improvement, thus saving computational complexity, and only when there is corrosion in the preliminary judgment result, the improved Yolov5 model is used for corrosion detection to obtain the specific position where the boiler heating surface is corroded; finally, the parallel computing of the GPU is utilized to improve the operation speed. Description of the Drawings

[0032] Figure 1 It is a flow chart of a method for detecting corrosion of a boiler heating surface according to the present invention.

[0033] Figure 2 It is a CA attention structure diagram.

[0034] Figure 3 It is an ICA attention structure diagram. Detailed Embodiments

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0036] Figure 1 The schematic flowchart of a method for detecting corrosion of a boiler heating surface according to an embodiment of the present invention is shown.

[0037] As Figure 1 described, a method for detecting corrosion of a boiler heating surface according to an embodiment of the present invention includes:

[0038] Step 100, photograph the boiler heating surface, transmit the original video output by the high-definition camera into the video memory of the graphics card, and the GPU in the graphics card performs parallel decoding on the original video to obtain each frame of grayscale image;

[0039] Step 200, divide the i-th frame of grayscale image into N*N image blocks to obtain sum image blocks of the same size; perform DCT transformation on each image block in parallel using the GPU; obtain the DCT transformation coefficient matrix of each image block;

[0040] It should be noted that when the grayscale cannot be divided into an integer number of N*N-sized image blocks, a supplementary operation can be performed on the grayscale image. The most common supplementary operation is zero-padding or padding according to the pixel values at the image edges.

[0041] Step 300, use the GPU to parallelly convert the DCT transformation coefficient matrix of each image block into a (N*N)×1-dimensional column vector, input each column vector into the trained improved BP neural network in sequence, and obtain the determination result of whether there is corrosion in each image block. If the determination result of the j-th column vector corresponding image block is that there is corrosion, j∈[1,sum], then jump to Step 400; if the determination results of all image blocks of the i-th frame of grayscale image are that there is no corrosion, let i = i + 1, and return to Step 200;

[0042] Step 400, input the i-th frame of grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific position where the boiler heating surface is corroded in the i-th frame, let i = i + 1, and return to Step 200.

[0043] According to an embodiment of the present application, the DCT transform is performed on each image block in parallel using a GPU; the DCT transform coefficient matrix of each image block is obtained, including the GPU allocating a thread block to each image block and allocating 128 or 256 threads within each thread block to perform the DCT transform on the image block in parallel.

[0044] According to an embodiment of the present application, the step of inputting each column vector into the trained improved BP neural network in sequence to obtain the determination result of whether there is corrosion in each image block includes:

[0045] Step 310, training the improved BP neural network using training sample set 1 to obtain the trained improved BP neural network. The specific process is as follows:

[0046] Set the training termination criteria: one is to reach the specified maximum number of iterations T, the second is that the error function value is less than the specified standard ε, and the third is that the weight change amount is less than ε;

[0047] Set the training pause criteria: when the training termination criteria are not met;

[0048] Training process:

[0049] Step 1: Initialize the network structure, termination and pause criteria, and other network parameters;

[0050] Step 2: Start training. If the termination criteria are met, the training process ends. If the training pause criteria are met, go to Step 3 for pruning;

[0051] Step 3: If the pause criteria are met, then start pruning. Calculate the minimum vertical grey relational degree and use it to determine the pruning connection. If the minimum vertical grey relational degree < ε, then take its corresponding connection as the pruning connection of the output neuron, go to Step 4. If all are not less than the defined value, then go to Step 2 and continue training;

[0052] Step 4: Select the corresponding merging connection for each pruning connection. Using the network horizontal grey relational analysis method, calculate the maximum network horizontal grey relational degree of the input value sequence, and the merging connection is its corresponding connection;

[0053] Step 5: Delete the pruning connection, modify the weights of the merging connection, and perform weight adjustment using weighted average. Adjust the rated relational degree ε = ε + k, where k is a constant selected according to the specific problem;

[0054] Step 6: Pause the pruning process, go to Step 2, and continue training;

[0055] Step 7: Determine the network structure and end the training process.

[0056] Step 320: Use the trained improved BP neural network to determine whether to send erosion determination for each column vector respectively.

[0057] It should be noted that each sample in the training sample set 1 consists of a column vector composed of DCT transform coefficients of an image block and a label. The label includes two situations: the existence of erosion and the non - existence of erosion. The improved BP neural network algorithm used in this application is not limited to the improved BP neural network algorithm proposed above, and can also be other existing improved BP neural network algorithms.

[0058] According to an embodiment of the present application, inputting the i - th frame grayscale image into the trained improved Yolov5 model for erosion detection, so as to obtain the specific position where the boiler heating surface is corroded in the i - th frame, specifically includes the following steps:

[0059] Step 410: Improve the Yolov5 model. In the Yolov5 model, retain the ordinary convolution in the backbone feature extraction network, and replace the ordinary feature layers in the neck network with FSConv convolution. The specific implementation method of FSConv convolution is as follows: First, generate high - dimensional features through ordinary convolution, then use depth - separable convolution to transform the high - dimensional features, then splice the two features obtained in the above two steps, and through a shuffle operation, shuffle the feature maps of different groups together to enhance the information interaction between different groups; Improve the CA attention structure used in Yolov5 to obtain the improved CA attention structure ICA. The specific implementation method is: Add a branch to the CA attention, use max - pooling to extract this part of the key features, and finally merge with the average - pooling branch. The two branches act together on the output feature map to strengthen the final classification and localization effect.

[0060] It should be noted that CA attention is a lightweight attention structure in the prior art, and its structure diagram is as Figure 2 shown. CA attention uses average - pooling for one - dimensional feature encoding. The sampling method of average - pooling can better contain the background information of the object, which is helpful for the classification of the model, but will ignore some details. While the sampling method of max - pooling can focus on details such as the texture of the target. Therefore, in the present invention, a branch is added to the CA attention, max - pooling is used to extract this part of the key features, and finally merged with the average - pooling branch. The two branches act together on the output feature map to strengthen the final classification and localization effect. It is named the ICA attention module, and the ICA attention structure is as Figure 3As shown in the figure. For the input feature map, first use two one-dimensional average pooling kernels of (1, W) and (H, 1) and two max pooling kernels of the same size to encode the channel information respectively. After this step, four groups of feature tensors will be obtained. Then, according to the pooling method, the four groups of features are concatenated into two groups to fuse the information from the horizontal and vertical directions, and a group of 1×1 convolutions are used to reduce the number of channels and the number of parameters. To reduce the number of parameters, the number of channels is reduced in this layer, and then the input is passed through a non-linear activation function. After that, the two groups of obtained features are sliced according to the width and height dimensions to obtain 4 groups of features: fh1, fh2, fw1, fw2. The sum operation is used to merge the feature information from the width and height corresponding to the two branches, and a convolution is used to restore the same number of channels as the feature map. The process is shown in formulas (1) and (2):

[0061] g h = σ[F h (fh1 + fh2)] (1)

[0062] g w = σ[F w (fw1 + fw2)] (2)

[0063] where σ is the sigmoid activation function. F h and F w are two different convolutional layers, which respectively receive the feature information from the width and height and restore the number of channels. The two groups of obtained feature tensors are used as the learned weights to act on the feature map to complete the enhanced representation of the features.

[0064] When the input feature map is x(i, j), the output through the entire ICA attention structure can be expressed as formula (3):

[0065] y(i,j) = x(i,j) × g h (i) × g w (j) (3)

[0066] Thus, the process of using the attention module to weight the features can be realized, enabling the model to pay more attention to important features while suppressing irrelevant information such as the background.

[0067] Step 420: Use the training sample set 2 to train the improved Yolov5 model to obtain the trained improved Yolov5 model;

[0068] Step 430: Input the i-th grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location of the boiler heating surface corrosion in the i-th frame.

[0069] It should be noted that the samples in the training sample set 2 are boiler heating surface images in various situations, and the labels are the classification of whether there is corrosion and the coordinates of the corrosion position.

[0070] According to an embodiment of the present application, after obtaining the specific position where the boiler heating surface is corroded, it further includes: obtaining the actual position where the boiler heating surface is corroded based on the specific position where the boiler heating surface is corroded in the image, and sending a reminder to measure the thickness of the pipeline at the actual position where the boiler heating surface is corroded.

[0071] Specifically, after obtaining the specific position where the boiler heating surface is corroded through the improved Yolov5 model, based on the one-to-one correspondence between each point in the image and the actual position coordinates of the boiler heating surface, the actual position where the boiler heating surface is corroded can be obtained according to the specific position where the boiler heating surface is corroded in the image. When the actual position where the boiler heating surface is corroded is determined, it is also necessary to further remind the operation and maintenance personnel to measure the thickness of the actual position where the boiler heating surface is corroded, so as to further determine the severity of the corrosion.

[0072] According to an embodiment of the present application, it further includes: detecting the actual thickness of the pipeline at the actual position where the boiler heating surface is corroded by an ultrasonic thickness gauge. If the actual thickness is greater than or equal to the first preset thickness, the corrosion level of the pipeline at the actual position where the boiler heating surface is corroded belongs to mild corrosion and does not need to be treated, and a prompt of "mild corrosion, continuous attention" is output; if the actual thickness is less than the first preset thickness and greater than or equal to the second preset thickness, the corrosion level of the pipeline at the actual position where the boiler heating surface is corroded belongs to moderate corrosion, and the operation and maintenance personnel are reminded to optimize the operation control; if the actual thickness is less than the second preset thickness, the corrosion level of the pipeline at the actual position where the boiler heating surface is corroded belongs to severe corrosion, and the operation and maintenance personnel are reminded to conduct a metallographic structure test, replace the pipeline at the corrosion position of the boiler heating surface, and optimize the operation control; wherein, the first preset thickness is greater than the second preset thickness.

[0073] Specifically, after the boiler heating surface is corroded, after cleaning the corroded part, measure the thickness of the pipeline at the actual location where the boiler heating surface is corroded. Different degrees of corrosion correspond to different thicknesses of the metal pipelines of the boiler heating surface, and different treatment measures should be adopted. If the actual thickness is greater than or equal to the first preset thickness, the corrosion level of the pipeline at the actual location where the boiler heating surface is corroded belongs to mild corrosion and does not require treatment, and a prompt of "mild corrosion, continuous attention" is output. If the actual thickness is less than the first preset thickness and greater than or equal to the second preset thickness, the corrosion level of the pipeline at the actual location where the boiler heating surface is corroded belongs to moderate corrosion, and the operation and maintenance personnel are reminded to optimize the operation control. The optimization of the operation control specifically includes, according to the boiler startup curve, strictly controlling the quantity and ratio of air and coal, maintaining a stable combustion rate, reasonably using desuperheated water, and avoiding local temperature mutations of the metal pipe materials; during the normal operation stage of the unit, strengthening operation monitoring, controlling the flue gas temperature at the furnace outlet, maintaining an appropriate amount of desuperheated water, and ensuring that the metal wall temperature of the boiler heating surface does not exceed the limit, etc. If the actual thickness is less than the second preset thickness, the corrosion level of the pipeline at the actual location where the boiler heating surface is corroded belongs to severe corrosion, and the operation and maintenance personnel are reminded to conduct a metallographic structure test, replace the pipeline part at the actual location where the boiler heating surface is corroded, and optimize the operation control. Specifically, determine whether the material of the heating surface meets the design requirements through the metallographic structure test, and analyze the specific factors causing severe corrosion. When the actual thickness is less than the second preset thickness, it indicates that the thickness of the corroded pipeline exceeds the limit thickness for safe operation and there is a risk of pipe burst, and it is necessary to directly replace this pipe section to improve the safety of the heating surface. And, when the actual thickness is less than the second preset thickness, it indicates that there are serious deficiencies in the operation control of the boiler and it is necessary to optimize the operation control.

[0074] It should be understood that the various forms of the processes shown above can be used, steps can be reordered, added, or deleted. For example, the steps described in the present invention can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved, and this is not limited herein.

[0075] The above specific implementation manners do not constitute a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub - combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting corrosion of boiler heating surfaces, characterized in that, Including: Step 100: Photograph the boiler heating surface, transmit the original video output by the high-definition camera into the video card memory, and the GPU in the video card performs parallel decoding on the original video to obtain each frame of grayscale image. Step 200: Divide the i-th frame of grayscale image into N*N image blocks to obtain sum image blocks of the same size; perform DCT transformation on each image block in parallel using the GPU; obtain the DCT transformation coefficient matrix of each image block. Step 300: Use the GPU to parallelly convert the DCT transformation coefficient matrix of each image block into a (N*N)×1-dimensional column vector, input each column vector into the trained improved BP neural network in sequence, and obtain the determination result of whether there is corrosion in each image block. If the determination result of the j-th column vector corresponding image block is that there is corrosion, j ∈ [1, sum], then jump to Step 400. If the determination results of all image blocks of the i-th frame of grayscale image are that there is no corrosion, let i = i + 1, and return to Step 200. Step 400: Input the i-th frame of grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location where the boiler heating surface is corroded in the i-th frame, let i = i + 1, and return to Step 200.

2. The boiler heating surface corrosion detection method according to claim 1, characterized in that: The process of inputting each column vector into the trained improved BP neural network in sequence to obtain the determination result of whether there is corrosion in each image block includes: Step 310: Use training sample set 1 to train the improved BP neural network to obtain the trained improved BP neural network. The specific process is as follows: Set the training termination criteria: one is to reach the specified maximum number of iterations T, the second is that the error function value is less than the specified standard ε, and the third is that the weight change amount is less than ε. Set the training pause criteria: when the training termination criteria are not met. Training process: Step 1: Initialize the network structure, termination and pause criteria, and other network parameters. Step 2: Start training. If the termination criteria are met, the training process ends. If the training pause criteria are met, go to Step 3 for pruning. Step 3: If the pause criteria are met, then start pruning, calculate the minimum longitudinal grey correlation degree, use it to determine the pruning connection. If the minimum longitudinal grey correlation degree < ε, then take its corresponding connection as the pruning connection of the output neuron, go to Step 4. If they are all not less than the defined value, then go to Step 2 and continue training. Step 4: Select the corresponding merging connection for each pruning connection. Using the network horizontal grey correlation analysis method, calculate the maximum network horizontal grey correlation degree of the input value sequence, and the merging connection is its corresponding connection. Step 5: Delete the pruning connection, modify the weight of the merging connection, and use weighted average for weight adjustment. Adjust the rated correlation degree ε = ε + k, where k is a constant selected according to the specific problem. Step 6: Pause the pruning process, go to Step 2, and continue training. Step 7: Determine the network structure and end the training process. Step 320: Use the trained improved BP neural network to determine whether corrosion occurs for each column vector respectively.

3. The boiler heating surface corrosion detection method according to claim 2, characterized in that: Input the i-th frame grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location of corrosion on the boiler heating surface in the i-th frame. The specific steps are as follows: Step 410: Improve the Yolov5 model. In the Yolov5 model, retain the ordinary convolutions in the backbone feature extraction network, and replace the ordinary feature layers in the neck network with FSConv convolutions. The specific implementation method of FSConv convolution is as follows: First, generate high-dimensional features through ordinary convolutions, then use depthwise separable convolutions to transform the high-dimensional features, then splice the two features obtained in the above two steps, and through a shuffle operation, shuffle the feature maps of different groups together to enhance the information interaction between different groups; Improve the CA attention structure used in Yolov5 to obtain the improved CA attention structure ICA. The specific implementation method is: Add a branch to the CA attention, use max pooling to extract the key features of this part, and merge with the average pooling branch at the end. The two branches jointly act on the output feature map to strengthen the final classification and localization effects; Step 420: Use the training sample set 2 to train the improved Yolov5 model to obtain the trained improved Yolov5 model; Step 430: Input the i-th frame grayscale image into the trained improved Yolov5 model for corrosion detection, so as to obtain the specific location of corrosion on the boiler heating surface in the i-th frame.

4. The boiler heating surface corrosion detection method according to claim 3, wherein: After obtaining the specific location of corrosion on the boiler heating surface, it further includes: obtaining the actual location of corrosion on the boiler heating surface based on the specific location of corrosion on the boiler heating surface in the image, and sending a reminder to measure the pipeline at the actual location of corrosion on the boiler heating surface.

5. The boiler heating surface corrosion detection method according to claim 4, characterized in that: It further includes: Detect the actual thickness of the pipeline at the actual location of corrosion on the boiler heating surface by an ultrasonic thickness gauge. If the actual thickness is greater than or equal to the first preset thickness, the corrosion level of the pipeline at the actual location of corrosion on this boiler heating surface belongs to mild corrosion and does not need to be processed, and output a prompt of "Mild corrosion, continuous attention"; If the actual thickness is less than the first preset thickness and greater than or equal to the second preset thickness, the corrosion level of the pipeline at the actual location of corrosion on this boiler heating surface belongs to moderate corrosion, and remind the operation and maintenance personnel to optimize the operation control; If the actual thickness is less than the second preset thickness, the corrosion level of the pipeline at the actual location of corrosion on this boiler heating surface belongs to severe corrosion, and remind the operation and maintenance personnel to conduct a metallographic structure test, replace the pipeline at the corrosion location of this boiler heating surface, and optimize the operation control; where the first preset thickness is greater than the second preset thickness.

Images