A high-temperature corrosion image detection method and system for a water-cooled wall of a coal-fired boiler

Abstract

The present application relates to a kind of high-temperature corrosion image detection method and system of coal-fired boiler water wall, belong to thermal power equipment maintenance technical field.The following steps are included:collecting water wall surface image, input preliminary result obtained by pre-training high-temperature corrosion image detection model, final result is obtained after processing;The model is based on the mask convolution neural network constructed by pyramid network and multi-scale anchor region generation network of embedding convolution block attention module.In training stage, the model is optimized by custom dataset, multi-scale feature fusion, attention reinforcement and dynamic training strategy.The present application effectively improves the recognition accuracy of micro, small pattern corrosion defects, suppresses the interference information on site, reduces the missed detection and misdiagnosis rate, has high stability and robustness of industrial scene deployment.

Description

Technical Field

[0001] This invention relates to a high-temperature corrosion image detection method and system for water-cooled walls of coal-fired boilers, belonging to the technical field of thermal power equipment maintenance. Background Technology

[0002] In thermal power generating units, the furnace water-cooled wall is a core heat exchange component. It is in a harsh environment with high temperature, high pressure and rich in corrosive gases for a long time, and is extremely prone to high temperature corrosion, which can lead to serious equipment failures such as pipe leakage and rupture. This not only affects power generation efficiency, but may also cause safety accidents.

[0003] Traditional high-temperature corrosion detection of water-cooled walls relies on manual visual inspection, which has significant drawbacks: on the one hand, the detection process is time-consuming and labor-intensive, resulting in low maintenance efficiency; on the other hand, the internal structure of the furnace is complex, making it difficult for manual inspection to cover the entire area, resulting in blind spots and making it impossible to comprehensively investigate potential corrosion risks. Existing technologies have the following shortcomings: One type of technology focuses on general corrosion detection (such as corrosion detection of chemical pipelines), using traditional image recognition models, which do not consider the special scenario of "multiple forms of corrosion defects such as pitting, erosion, and cracking coexisting and superimposed" in water-cooled walls of coal-fired boilers, resulting in poor model adaptability; Another type of technology focuses on the detection of surface defects in other equipment (such as wear detection of wind turbine blades), and its model structure and training strategy are designed for smooth blade surfaces and single wear patterns, which differs significantly from the scenario of "rough surfaces and multiple corrosion patterns superimposed" in water-cooled wall tubes, and cannot meet the detection requirements; Existing models generally have weak resistance to environmental interference and low accuracy in recognizing small targets, making it difficult to deal with false defects caused by attachments such as coking, scale, floating ash, and coal dust on the surface of water-cooled wall tubes in the furnace, as well as the problem of missing detection of tiny corrosion defects with a diameter of less than 2 mm.

[0004] Therefore, there is an urgent need for a high-temperature corrosion image detection method and system suitable for the accurate identification, location and classification of high-temperature corrosion defects on the surface of water-cooled wall tubes inside the furnace of thermal power generating units. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method and system for high-temperature corrosion image detection of water-cooled walls in coal-fired boilers.

[0006] The technical solution of the present invention is as follows: On the one hand, this invention proposes a method for high-temperature corrosion image detection of water-cooled walls in coal-fired boilers, comprising the following steps: On the one hand, this invention proposes a method for high-temperature corrosion image detection of water-cooled walls in coal-fired boilers, comprising the following steps: Acquire images of the water-cooled wall surface of a coal-fired boiler; The surface image of the water-cooled wall is input into a pre-trained high-temperature corrosion image detection model to obtain preliminary high-temperature corrosion image detection results of the water-cooled wall. The preliminary high-temperature corrosion image detection results are input into the target detection post-processing algorithm. After redundancy removal, the final high-temperature corrosion image detection results are obtained. The pre-trained high-temperature corrosion image detection model is constructed based on a mask region convolutional neural network with a region generation network configured with multi-scale anchor points and a pyramid network embedded with a convolutional attention module.

[0007] Preferably, the high-temperature corrosion image detection model is trained according to the following steps: Based on the characteristics of high-temperature corrosion on the surface of water-cooled walls of coal-fired boilers, multi-scale anchor points are set in the convolutional neural network of the mask region, and a convolutional attention module is embedded. Combined with the pyramid network, a high-temperature corrosion image detection model is constructed. Collect sample images of the water-cooled wall surface of a coal-fired boiler, use annotation tools to annotate the actual high-temperature corrosion defect boxes, and construct the original dataset; The original dataset is processed using data augmentation strategies to obtain an augmented dataset; The enhanced dataset is input into the high-temperature corrosion image detection model, and the high-temperature corrosion image detection model is trained using a preset training strategy to obtain a pre-trained high-temperature corrosion image detection model.

[0008] Preferably, the data augmentation strategy includes at least one or more combinations of the following: randomly rotating, horizontally flipping, vertically flipping, adjusting brightness, adjusting contrast, and adding Gaussian noise to the extracted sample images of the water-cooled wall surface of the coal-fired boiler.

[0009] Preferably, the step of inputting the enhanced dataset into the high-temperature corrosion image detection model and training the high-temperature corrosion image detection model using a preset training strategy specifically involves: The enhanced dataset is input into the high-temperature corrosion image detection model; The samples in the augmented dataset are passed through the backbone network of the mask region convolutional neural network to obtain original feature maps at different levels. The original feature maps at each level are fused using a pyramid network. A fused feature map is generated by upsampling the upper-layer feature map and superimposing it with the lower-layer feature map. The fused feature map is then obtained by convolution. The predicted feature map is then input into the convolutional block attention module to obtain the attention-enhanced predicted feature map. Based on the hierarchical scale differences of the attention-enhanced predicted feature maps and the common size range of high-temperature corrosion defects in water-cooled walls, multiple sets of anchor boxes with different pixel sizes and aspect ratios are pre-configured. They are assigned to the corresponding hierarchical feature maps according to the principle of matching small-sized anchors to low-level feature maps and large-sized anchors to high-level feature maps. Then, the cross-union algorithm is used to calculate the matching degree between the anchor boxes and the actual corrosion defect boxes. Multi-scale anchor boxes with cross-union ratios greater than a preset threshold and their corresponding annotation information are selected to obtain defect candidate boxes. Based on the defect candidate boxes and corresponding annotation information, the total training loss of the model is calculated; the total training loss of the model is backpropagated using the backpropagation algorithm through an optimizer with a preset weight decay coefficient, and the training parameters of the high-temperature corrosion image detection model are updated. After training according to the preset training strategy and iteration rounds, a pre-trained high-temperature corrosion image detection model is obtained.

[0010] Preferably, the predicted feature map is calculated according to the following steps: The samples in the augmented dataset are passed through the backbone network of the mask region convolutional neural network to obtain the first, second, third and fourth level original feature maps in sequence; Starting from the fourth-level original feature map, a 1×1 convolution operation is performed to adjust the number of channels to obtain the first-level intermediate feature map. The first-level intermediate feature map is upsampled by 2 times, and the third-level original feature map is then subjected to a 1×1 convolution to adjust the number of channels before being superimposed with the first-level intermediate feature map to obtain the second-level intermediate feature map. The second-level intermediate feature map is upsampled by 2 times, and the second-level original feature map is then subjected to a 1×1 convolution to adjust the number of channels before being superimposed with the second-level intermediate feature map to obtain the third-level intermediate feature map. The third-level intermediate feature map is upsampled by 2 times, and the first-level original feature map is then subjected to a 1×1 convolution to adjust the number of channels before being superimposed with the third-level intermediate feature map to obtain the fourth-level intermediate feature map. The intermediate feature maps of the four levels are processed by 3×3 convolution to obtain the predicted feature maps of the four levels.

[0011] Preferably, the attention-enhanced prediction feature map is obtained according to the following steps: For each level of predicted feature map, convolutional block attention modules are used for feature enhancement processing, sequentially through channel attention and spatial attention mechanisms: The channel attention mechanism performs global average pooling and global max pooling on the predicted feature map of each layer. The pooling results are input into a multilayer perceptron, transformed, summed, and then passed through an activation function to generate a channel attention weight map. This weight map is then multiplied by the predicted feature map of each layer to obtain the feature map after channel attention processing. The spatial attention mechanism performs channel-dimensional average pooling and max pooling on the feature maps after channel attention processing. The pooling results are concatenated and then generated by convolution and activation functions to produce a spatial attention weight map. This weight map is then multiplied with the feature maps after channel attention processing to obtain the predicted feature maps after attention enhancement at each level.

[0012] Preferably, the defect candidate box containing high-temperature corrosion is obtained according to the following steps: Configure multiple scale anchor points with different pixel sizes to adapt to corrosion defects of different scales; For each anchor point, preset position and size parameters are used. Combined with the pixel scale of the predicted feature map after attention enhancement at the anchor point level, a coordinate conversion algorithm is used to obtain the specific coverage area of ​​each anchor point on the predicted feature map after attention enhancement at the corresponding level. Each multi-scale anchor box is compared with the corresponding real corrosion defect box. By calculating the ratio of the intersection area to the union area of ​​the two, multi-scale anchor boxes that meet the preset matching conditions are selected and their positive and negative sample attributes are determined. The selected positive sample multi-scale anchor point boxes are subjected to bounding box regression processing to correct their position and size parameters, thereby obtaining defect candidate boxes containing high-temperature corrosion.

[0013] On the other hand, the present invention also proposes a high-temperature corrosion image detection system for water-cooled walls of coal-fired boilers, which includes the following modules: Image acquisition module: Acquires images of the water-cooled wall surface of a coal-fired boiler; Corrosion detection module: Input the surface image of the water-cooled wall into the pre-trained high-temperature corrosion image detection model to obtain the preliminary high-temperature corrosion image detection results of the water-cooled wall; Detection result denoising module: The preliminary high-temperature corrosion image detection result is input into the target detection post-processing algorithm. After redundancy removal, the final high-temperature corrosion image detection result is obtained.

[0014] In another aspect, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in any embodiment of the present invention.

[0015] In another aspect, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0016] The present invention has the following beneficial effects: (1) This invention proposes a method and system for detecting high-temperature corrosion images of water-cooled walls in coal-fired boilers. It uses a pyramid network to fuse multi-level original feature maps from top to bottom, combined with channel and spatial attention enhancement from a convolutional block attention module. This preserves the semantic information of high-level features while enhancing the detail expression of low-level features, effectively solving the problem of insufficient recognition capability of traditional models for corrosion defects at different scales. Simultaneously, the attention mechanism suppresses interference information such as coking, floating ash, and oxide scale on the water-cooled wall surface, significantly improving the targeting and accuracy of feature extraction, thereby greatly improving the detection accuracy and robustness of high-temperature corrosion defects.

[0017] (2) This invention proposes a high-temperature corrosion image detection method and system for water-cooled walls of coal-fired boilers. By configuring multi-scale anchor points based on the scale distribution characteristics of high-temperature corrosion defects, the anchor point coverage of small corrosion defects is strengthened. Combined with the intersection-union-ratio matching algorithm and bounding box regression processing, the position and size of the anchor point box can be accurately screened and corrected, effectively solving the problems of high false negative rate and insufficient positioning accuracy of traditional anchor point mechanism for small defects. This mechanism can significantly improve the generation quality of defect candidate boxes, providing a more accurate basis for subsequent model detection, thereby reducing the false negative rate and false positive rate of high-temperature corrosion defects.

[0018] (3) This invention proposes a method and system for detecting high-temperature corrosion images of water-cooled walls of coal-fired boilers. By constructing a customized dataset that fits the scene of water-cooled walls of coal-fired boilers, and combining targeted data augmentation strategies such as random rotation and brightness adjustment, the problem of insufficient labeled samples in industrial scenarios is effectively alleviated. At the same time, the training strategy of pre-trained weight initialization, dynamic learning rate and early stopping mechanism is adopted, which not only improves the convergence speed and generalization ability of the model, but also avoids the risk of overfitting. This whole training scheme enables the model to better adapt to the complex high-temperature corrosion scene in industrial sites and has stronger stability and reliability in actual deployment. Attached Figure Description

[0019] Figure 1 This is a flowchart of a high-temperature corrosion image detection method for water-cooled walls of a coal-fired boiler, as proposed in Embodiment 1 of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0022] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0024] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0025] Example 1: See Figure 1 This embodiment proposes a method for high-temperature corrosion image detection of water-cooled walls in coal-fired boilers, including the following steps: S100. During unit shutdown and maintenance periods, images of the water-cooled wall surface are acquired using two acquisition methods: First, professional testing personnel were arranged to collect images on-site using high-definition image acquisition equipment; Secondly, data was collected using a modified DJI Mavic 3 Enterprise drone platform. Specifically: The drone's flight altitude is controlled within the range of 0.5-1.5 meters from the water-cooled wall tube, and the acquisition angle fully covers 0°, 30°, and 45° to ensure that corrosion characteristics from different perspectives can be captured. The acquired images need to cover samples of water-cooled walls in normal condition as well as samples of various degrees of high-temperature corrosion, including: Slightly corroded samples correspond to pits with a diameter <2mm and a depth <0.5mm; The corresponding pits for moderately corroded samples are: 2mm ≤ diameter ≤ 5mm, 0.5mm ≤ depth ≤ 1mm; Severely corroded samples have pits with a diameter >5mm and a depth >1mm. No fewer than 5,000 images were collected for each state, ultimately constructing a diverse original dataset.

[0026] Furthermore, the original dataset is labeled. The labeling process uses the professional labeling software Labelme to accurately label the eroded regions in each image. The labeling content includes the outline bounding box of the eroded region and the corresponding erosion degree annotation. The erosion degree annotation follows the above-mentioned mild, moderate and severe classification rules to provide a reference for model training.

[0027] S200. Data augmentation processing employs online augmentation strategies on the original training set to generate augmented datasets. Specifically: Random rotation: Rotation angle range -15°~15°, automatically fills background pixels after rotation; Horizontal and vertical flipping: each with a 50% execution probability, maintaining the semantics of corrosion defects unchanged; Brightness adjustment: Randomly adjusts the original brightness from 0.8 to 1.2 times. Contrast adjustment: Randomly adjust the original contrast ratio from 0.8 to 1.2 times. Add Gaussian noise: noise standard deviation range 0.01~0.03, to avoid masking corrosion features. Meanwhile, the augmented dataset was divided into training, validation, and test sets in a 7:2:1 ratio, resulting in a dataset that is three times the size of the original dataset.

[0028] S201. The detection model is built upon a Mask R-CNN (Mask-based Region Convolutional Neural Network) architecture, using the Detectron2 open-source framework. ResNet-50 is configured as the backbone network for extracting multi-scale raw features. The neck network employs a pyramid network FPN with embedded convolutional block attention modules (CBAM) for feature fusion and enhancement. The region generation network (RPN) is configured with multi-scale anchors to adapt to corrosion defects of different sizes. Overall, this constitutes a high-temperature corrosion image detection model. Specifically: First, multi-scale raw feature extraction is performed. The images from the augmented dataset are input into the model, and after convolution and pooling operations of the ResNet-50 backbone network, the raw feature maps of four levels are output, defined as follows: ; in, This represents the original feature map of the first level. This represents the original feature map of the second level. This represents the original feature map of the third level. This represents the original feature map of the fourth level; Subsequently, the pyramid network FPN multi-scale feature fusion is performed to generate predicted feature maps using a combination of top-down and lateral connections. First, a 1×1 convolution is performed on the original feature map of the fourth layer, adjusting the number of channels to 256, to obtain the intermediate feature map of the first layer, expressed by the formula: ; In the formula, This represents the intermediate feature map of the first level. This represents a 1×1 convolution operation; Next, perform a 2x bilinear interpolation upsampling on the first-level intermediate feature map, and simultaneously perform a 1×1 convolution on the original third-level feature map, adjusting the number of channels to 256; the two are then superimposed to obtain the second-level intermediate feature map, expressed by the formula: ; In the formula, This represents the intermediate feature map of the second level. This indicates a 2x bilinear interpolation upsampling operation; Following the same logic, the intermediate feature maps of the third and fourth layers are successively fused to obtain them, which can be expressed by the following formula: ; ; In the formula, This represents the intermediate feature map of the third level. This represents the intermediate feature map of the fourth level.

[0029] Furthermore, 3×3 convolutions are performed on the intermediate feature maps of the four levels respectively to eliminate the upsampling aliasing effect, resulting in predicted feature maps of the four levels, with dimensions corresponding to the intermediate feature maps. Figure 1 The number of channels is 256, expressed by the formula: ; In the formula, Indicates a hierarchical index. Indicates the first Hierarchical predicted feature maps Indicates the first Intermediate feature maps of the hierarchy, This represents a 3×3 convolution operation.

[0030] S202. Embed convolutional block attention modules in the predicted feature maps of each level, sequentially using channel attention and spatial attention mechanisms: The channel attention mechanism performs global average pooling and global max pooling on the predicted feature maps of each layer. The pooling results are then input into a multilayer perceptron for transformation, summed, and passed through an activation function to generate a channel attention weight map. This weight map is then multiplied by the predicted feature maps of each layer to obtain the channel attention-processed feature map, expressed by the formula: ; ; In the formula, This represents the channel attention weight map. Indicates global average pooling. Indicates global max pooling. This represents a multilayer perceptron. This represents the sigmoid activation function. Indicates the first Feature maps after channel attention processing at different levels; The spatial attention mechanism performs channel-dimensional average pooling and max pooling on the feature maps after channel attention processing. The pooling results are concatenated and then convolved with an activation function to generate a spatial attention weight map. This weight map is then multiplied by the channel attention-processed feature maps to obtain the final predicted feature maps after attention enhancement at each level, expressed by the formula: ; ; In the formula, Represents the spatial attention weight map. This represents a 7×7 convolution operation. Indicates the first Enhanced predictive feature maps at different levels.

[0031] S203. Based on the enhanced predicted feature map, generate multi-scale anchor points and filter candidate boxes; first set three anchor point sizes: 16×16, 32×32, and 64×64 pixels to adapt to corrosion defects of different scales, with a focus on enhancing anchor point coverage for small defects with a size of <5mm. Each size corresponds to three aspect ratios: 1:1, 1:2, and 2:1. Each feature map pixel generates nine anchor points. The parameters of the multi-size anchor point bounding box are represented as follows: ; in, This indicates the x-coordinate of the center point of the multi-size anchor point frame. This indicates the ordinate of the center point of the multi-size anchor point frame. Indicates the width of the anchor point frame for multiple sizes. Indicates the height of the anchor point frame with multiple dimensions; The intersection-union ratio (IUU) of the multi-scale anchor point frame and the actual corrosion defect frame is calculated using the following formula: ; In the formula, Represents a multi-scale anchor point frame. Represents the actual corrosion defect frame. This represents the intersection-union ratio (IoU) of the multi-scale anchor point bounding box and the actual corrosion defect bounding box. This represents the intersection area of ​​the multi-scale anchor point frame and the actual corrosion defect frame. This represents the area of ​​the union of the multi-scale anchor point frame and the actual corrosion defect frame; Set a matching threshold, for example when As a positive sample, The positive samples are negative samples, and the rest are ignored samples; regression correction is performed on the multi-scale anchor boxes of the positive samples to predict the offset, expressed by the formula: ; ; ; ; In the formula, This represents the correction value of the x-coordinate of the center point of the multi-scale anchor point frame. This represents the correction value of the ordinate of the center point of the multi-scale anchor point frame. This indicates the width correction value for the multi-scale anchor point frame. This indicates the height correction value for the multi-scale anchor point frame. This represents the x-coordinate of the center point of the actual corrosion defect frame. This represents the ordinate of the center point of the actual corrosion defect frame. This indicates the actual width of the corrosion defect frame. Indicates the actual height of the corrosion defect frame; By optimizing the offset using a regression model and correcting its position and size parameters, candidate boxes for defects including high-temperature corrosion are obtained.

[0032] S204. Loss Calculation and Training Parameter Update: The total model loss consists of classification loss, regression loss, and mask loss, expressed by the formula: ; In the formula, This represents the total loss of the model. Represents classification loss. Indicates regression loss, Indicates mask loss; The training parameters were configured using a transfer learning strategy. Weights pre-trained on the COCO dataset were used to initialize the model, followed by fine-tuning on a custom augmentation dataset. The Adam optimizer (weight decay coefficient 0.0001) was selected, with an initial learning rate of 0.001, decaying to 0.1 times the current value every 5 epochs, and an iteration cycle of 30 epochs. The batch size was 8. Early stopping was also implemented, monitoring the validation set loss value. Training was terminated when the loss did not decrease for 3 consecutive epochs to avoid overfitting and ensure training efficiency and model performance.

[0033] During the pre-trained model saving stage, after training is completed, the model weight file with the highest detection accuracy on the validation set is saved to obtain the pre-trained high-temperature corrosion image detection model.

[0034] S300: Directly call the pre-trained high-temperature corrosion image detection model and input the real-time collected water-cooled wall surface image into the model; the model automatically performs the feature extraction → FPN multi-scale feature fusion → CBAM attention enhancement → RPN candidate box generation → mask prediction inference process, and outputs preliminary detection results, including: bounding box coordinates, confidence, corrosion degree classification results of each high-temperature corrosion defect, and pixel-level mask of the corresponding corrosion area.

[0035] S301. Perform a post-processing algorithm on the preliminary detection results output by the model to eliminate redundancy and false detections and improve the accuracy of the results: First, candidate boxes are selected based on a confidence threshold of ≥0.5 to eliminate false positives with low confidence. Then, the non-maximum suppression (NMS) algorithm is used with an IoU threshold of 0.3 to calculate the intersection-union ratio among the remaining candidate boxes, suppressing highly overlapping redundant boxes and retaining the optimal bounding boxes. After post-processing optimization, the defect bounding box localization error is controlled within 1 pixel, resulting in the final high-temperature corrosion image detection result. The final high-temperature corrosion image detection results are transmitted to the power plant's distributed control system (DCS) in real time in a standardized data format, and the stability and real-time performance of data transmission are ensured through Docker containerization deployment and Kubernetes management. After receiving the data, the DCS system displays the corrosion distribution heat map, defect quantity statistics, and defect location coordinates of each area of ​​the water-cooled wall in real time on the monitoring terminal. Technicians can intuitively grasp the overall corrosion status of the water-cooled wall through the DCS system. When a severe corrosion defect is detected, the system will automatically trigger an early warning to prevent the equipment failure from escalating. Meanwhile, the final test results are simultaneously output to the customized report generation system, which automatically integrates defect information to generate a multi-dimensional test report. Technicians can quickly develop differentiated maintenance plans based on the report, achieving seamless integration of test results and equipment operation and maintenance.

[0036] Example 2: This embodiment proposes a high-temperature corrosion image detection system for the water-cooled wall of a coal-fired boiler, comprising the following modules: Image acquisition module: Acquires images of the water-cooled wall surface of a coal-fired boiler; Corrosion detection module: Input the surface image of the water-cooled wall into the pre-trained high-temperature corrosion image detection model to obtain the preliminary high-temperature corrosion image detection results of the water-cooled wall; Detection result denoising module: The preliminary high-temperature corrosion image detection result is input into the target detection post-processing algorithm. After redundancy removal, the final high-temperature corrosion image detection result is obtained.

[0037] Example 3: This embodiment proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in any embodiment of the present invention.

[0038] Example 4: This embodiment proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0039] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0040] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0041] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0042] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0043] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for high-temperature corrosion image detection of water-cooled walls in coal-fired boilers, wherein, Includes the following steps: Acquire images of the water-cooled wall surface of a coal-fired boiler; The surface image of the water-cooled wall is input into a pre-trained high-temperature corrosion image detection model to obtain preliminary high-temperature corrosion image detection results of the water-cooled wall. The preliminary high-temperature corrosion image detection results are input into the target detection post-processing algorithm. After redundancy removal, the final high-temperature corrosion image detection results are obtained. The pre-trained high-temperature corrosion image detection model is constructed based on a mask region convolutional neural network with a region generation network configured with multi-scale anchor points and a pyramid network embedded with a convolutional attention module.

2. The high-temperature corrosion image detection method for water-cooled walls of coal-fired boilers according to claim 1, wherein, The high-temperature corrosion image detection model is trained according to the following steps: Based on the characteristics of high-temperature corrosion on the surface of water-cooled walls of coal-fired boilers, multi-scale anchor points are set in the convolutional neural network of the mask region, and a convolutional attention module is embedded. Combined with the pyramid network, a high-temperature corrosion image detection model is constructed. Collect sample images of the water-cooled wall surface of a coal-fired boiler, use annotation tools to annotate the actual high-temperature corrosion defect boxes, and construct the original dataset; The original dataset is processed using data augmentation strategies to obtain an augmented dataset; The enhanced dataset is input into the high-temperature corrosion image detection model, and the high-temperature corrosion image detection model is trained using a preset training strategy to obtain a pre-trained high-temperature corrosion image detection model.

3. The high-temperature corrosion image detection method for water-cooled walls of coal-fired boilers according to claim 2, wherein, The data augmentation strategy includes at least one or more combinations of the following: randomly rotating, horizontally flipping, vertically flipping, adjusting brightness, adjusting contrast, and adding Gaussian noise to the extracted sample images of the water-cooled wall surface of a coal-fired boiler.

4. The high-temperature corrosion image detection method for water-cooled walls of coal-fired boilers according to claim 2, wherein, The process of inputting the enhanced dataset into the high-temperature corrosion image detection model and training the model using a preset training strategy specifically involves: The enhanced dataset is input into the high-temperature corrosion image detection model; The samples in the augmented dataset are passed through the backbone network of the mask region convolutional neural network to obtain original feature maps at different levels. The original feature maps at each level are fused using a pyramid network. A fused feature map is generated by upsampling the upper-layer feature map and superimposing it with the lower-layer feature map. The fused feature map is then obtained by convolution. The predicted feature map is then input into the convolutional block attention module to obtain the attention-enhanced predicted feature map. Based on the hierarchical scale differences of the attention-enhanced predicted feature maps and the common size range of high-temperature corrosion defects in water-cooled walls, multiple sets of anchor boxes with different pixel sizes and aspect ratios are pre-configured. They are assigned to the corresponding hierarchical feature maps according to the principle of matching small-sized anchors to low-level feature maps and large-sized anchors to high-level feature maps. Then, the cross-union algorithm is used to calculate the matching degree between the anchor boxes and the actual corrosion defect boxes. Multi-scale anchor boxes with cross-union ratios greater than a preset threshold and their corresponding annotation information are selected to obtain defect candidate boxes. Based on the defect candidate boxes and corresponding annotation information, the total training loss of the model is calculated; the total training loss of the model is backpropagated using the backpropagation algorithm through an optimizer with a preset weight decay coefficient, and the training parameters of the high-temperature corrosion image detection model are updated. After training according to the preset training strategy and iteration rounds, a pre-trained high-temperature corrosion image detection model is obtained.

5. The high-temperature corrosion image detection method for water-cooled walls of coal-fired boilers according to claim 4, wherein, The predicted feature map is calculated according to the following steps: The samples in the augmented dataset are passed through the backbone network of the mask region convolutional neural network to obtain the first, second, third and fourth level original feature maps in sequence; Starting from the fourth-level original feature map, a 1×1 convolution operation is performed to adjust the number of channels to obtain the first-level intermediate feature map. The first-level intermediate feature map is upsampled by 2 times, and the third-level original feature map is then subjected to a 1×1 convolution to adjust the number of channels before being superimposed with the first-level intermediate feature map to obtain the second-level intermediate feature map. The second-level intermediate feature map is upsampled by 2 times, and the second-level original feature map is then subjected to a 1×1 convolution to adjust the number of channels before being superimposed with the second-level intermediate feature map to obtain the third-level intermediate feature map. The third-level intermediate feature map is upsampled by 2 times, and the first-level original feature map is then subjected to a 1×1 convolution to adjust the number of channels before being superimposed with the third-level intermediate feature map to obtain the fourth-level intermediate feature map. The intermediate feature maps of the four levels are processed by 3×3 convolution to obtain the predicted feature maps of the four levels.

6. The high-temperature corrosion image detection method for water-cooled walls of coal-fired boilers according to claim 5, wherein, The attention-enhanced prediction feature map is obtained according to the following steps: For each level of predicted feature map, convolutional block attention modules are used for feature enhancement processing, sequentially through channel attention and spatial attention mechanisms: The channel attention mechanism performs global average pooling and global max pooling on the predicted feature map of each layer. The pooling results are input into a multilayer perceptron, transformed, summed, and then passed through an activation function to generate a channel attention weight map. This weight map is then multiplied by the predicted feature map of each layer to obtain the feature map after channel attention processing. The spatial attention mechanism performs channel-dimensional average pooling and max pooling on the feature maps after channel attention processing. The pooling results are concatenated and then generated by convolution and activation functions to produce a spatial attention weight map. This weight map is then multiplied with the feature maps after channel attention processing to obtain the predicted feature maps after attention enhancement at each level.

7. The high-temperature corrosion image detection method for water-cooled walls of coal-fired boilers according to claim 4, wherein, The defect candidate box containing high-temperature corrosion is obtained according to the following steps: Configure multiple scale anchor points with different pixel sizes to adapt to corrosion defects of different scales; For each anchor point, preset position and size parameters are used. Combined with the pixel scale of the predicted feature map after attention enhancement at the anchor point level, a coordinate conversion algorithm is used to obtain the specific coverage area of ​​each anchor point on the predicted feature map after attention enhancement at the corresponding level. Each multi-scale anchor box is compared with the corresponding real corrosion defect box. By calculating the ratio of the intersection area to the union area of ​​the two, multi-scale anchor boxes that meet the preset matching conditions are selected and their positive and negative sample attributes are determined. The selected positive sample multi-scale anchor point boxes are subjected to bounding box regression processing to correct their position and size parameters, thereby obtaining defect candidate boxes containing high-temperature corrosion.

8. A high-temperature corrosion image detection system for water-cooled walls of coal-fired boilers, wherein, Includes the following modules: Image acquisition module: Acquires images of the water-cooled wall surface of a coal-fired boiler; Corrosion detection module: Input the surface image of the water-cooled wall into the pre-trained high-temperature corrosion image detection model to obtain the preliminary high-temperature corrosion image detection results of the water-cooled wall; Detection result denoising module: The preliminary high-temperature corrosion image detection result is input into the target detection post-processing algorithm. After redundancy removal, the final high-temperature corrosion image detection result is obtained.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images