Condenser condensing pipe plug identification and positioning method and system based on deep learning

Abstract

The application discloses a kind of condenser condensing pipe plug identification and positioning method, system based on deep learning, method includes: collecting condenser tube plate image containing plug, and the position and category of plug in image are marked;Image is carried out multi-stage morphological pretreatment and edge enhancement;Utilize the training data set of preprocessed image and its marked information construction;Plug target detection model is trained, model is improved YOLOv8 model, improvement is: using RFAConv replaces C2F module in main network;Add P2 detection branch in Neck structure, adopt BiFPN instead of PANet, and use lossless compression module SPDConv to replace original down sampling;Finally, the same pre-processing is carried out to the tube plate image to be identified, and input into the trained plug target detection model to carry out real-time plug identification and positioning.The application can realize the automatic identification of plug, and the identification efficiency is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image recognition, and in particular to a condenser condensing pipe plug identification and positioning method and system based on deep learning. BACKGROUND

[0002] The condenser is an important device in a large power station, which is composed of tens of thousands of small-diameter condensing pipes. It is mainly used to condense the low-pressure steam discharged by the steam turbine into water and maintain a certain vacuum degree at the steam exhaust port of the steam turbine. The main functions of the condenser are as follows: improving the thermal efficiency of the steam turbine cycle; reducing steam and water loss; and improving the safety and reliability of the entire power generation system.

[0003] To ensure the reliability of the condenser during operation, the defects of the condensing pipes found during inspection are all treated by plugging (divided into plastic plug plugging and welding plugging), and because the most surface condensing pipe bundle of the condenser is impacted by the exhaust steam or other equipment engineers consider that the pipe bundle has a higher risk of damage, the power plant adopts the measure of using plastic plugs for preventive plugging.

[0004] The plugs in the condenser tube plate image are arranged particularly densely, the number is extremely large, the structure is regular, the shape is repeated (such as Figure 1 indicated), the spacing is small, it is easy to be blocked, it is obviously affected by light, and the size of the tube plate is large, while the size of the plug is small, the size of the plug is at least 10 3 orders of magnitude smaller than the size of the tube plate; and the proportion of abnormal plugs on the tube plate is low, which is difficult to detect, and is often ignored due to slight shape differences or displacement.

[0005] The existing inspection method for old plugs completely relies on manual visual inspection, and each is checked one by one by referring to historical plugging drawings, which is extremely time-consuming, and long-time repetitive work is easy to cause personnel negligence and omission, resulting in human errors, and there is an urgent need for a method that can automatically identify and position the three types of plugs, plastic, welding, and no plug, according to the tube plate plugging image during the overhaul of the steam turbine, and the missing or incorrectly plugged plugs can be automatically identified to ensure the safe operation of the equipment. SUMMARY

[0006] In view of the low efficiency and poor accuracy of the existing manual detection of old plugs in the condenser tube plate, the present application proposes a condenser condensing pipe plug identification and positioning method and system based on deep learning, and the specific technical solutions are as follows:

[0007] A condenser condensing pipe plug identification and positioning method based on deep learning, comprising the following steps:

[0008] Step 1: Collect condenser tube plate images containing plugs, and label the positions and categories of the plugs in the condenser tube plate images to construct a data set;

[0009] Step two: multi-stage morphological preprocessing is performed on the images in the data set to solve the problems of reflection, occlusion and noise, and edge enhancement is performed, and all morphological operations are directly applied to the RGB three-channel image to retain the original color information; a training data set is constructed using the preprocessed image and its labeled information;

[0010] Step three: a plug target detection model is constructed, and the plug target detection model is trained using the training data set; the plug target detection model is an improved YOLOv8 model, and the specific improvements are:

[0011] For the backbone network in the YOLOv8 model, the receptive field attention convolution module RFAConv is used to replace the C2F module;

[0012] For the Neck structure in the YOLOv8 model, a P2 detection branch is added to improve the detection capability of dense small plugs; an improved bidirectional feature pyramid network BiFPN is used to replace the PANet, and a lossless compression module SPDConv module is used to replace the original downsampling operation;

[0013] The input of the plug target detection model is the preprocessed image, and the output is the plug class, plug position and confidence whether it contains the target;

[0014] Step four: after the condenser tube sheet image to be recognized is preprocessed in the same way as step two, the trained plug target detection model is input to perform real-time plug recognition and positioning.

[0015] Further, when training the plug target detection model, the loss function is the weighted sum of three types of losses of boundary box regression loss L box , classification loss L cls and target loss L obj .

[0016] The boundary box regression loss L box is the loss L Shacpe-IoU that focuses on the matching degree of the predicted boundary box and the real boundary box contour, and the weighted sum of the original loss function L CIoU .

[0017] The classification loss L cls uses the Focal Loss loss function.

[0018] Further, when training the plug target detection model, the weight coefficient of the CIoU loss function is higher than that of the Shape-IoU loss function in the early training stage to ensure that the model converges quickly and completes preliminary positioning; the weight coefficient of the Shape-IoU loss function is gradually increased in the middle and later stages of training to improve the boundary fitting accuracy of dense small plugs.

[0019] Further, the step two comprises the following sub-steps:

[0020] S2.1: Median filtering is performed on the original color image to suppress high-frequency noise and retain edge details;

[0021] S2.2: Top-hat transformation is independently performed on each channel of the color image using an elliptical structure element to extract and suppress the highlight area, and then an enhanced image with suppressed highlights is generated by channel weighting fusion, and the overall contrast is further improved by brightness weighting superposition;

[0022] S2.3: The image obtained in S2.2 is converted into a grayscale image, and the dark spot features of the black hat transformation are extracted to generate a mask of the occluded area, and the mask is optimized by threshold segmentation and closing operation to identify the area in the tube plate image that is occluded by soil, and the occluded area is expanded by dilation operation, and finally the soil occluded area in the color image is repaired;

[0023] S2.4: Morphological opening operation is performed on the repaired color image obtained in S2.3 using an elliptical structure element to eliminate isolated small-scale noise points while maintaining the edge and structural integrity of the target area;

[0024] S2.5: The image obtained in S2.4 is converted into a grayscale image, and the gradient amplitude is extracted by a Sobel operator to generate an edge-enhanced binary feature map; the edge-enhanced binary feature map is superimposed on the original image by weight to form an enhanced image.

[0025] Further, the type of the plug includes plastic plug, welded plug and no plug.

[0026] Further, the condenser tube plate image containing the plug collected in the step one includes design image data of the condenser tube plate and spliced condenser tube plate real image.

[0027] Further, it further comprises step five:

[0028] The confidence of whether the target is contained in the output of the trained plug target detection model is compared with the confidence threshold, and when the confidence output by the model is less than the confidence threshold, the current recognition result is marked as "suspicious target" for subsequent manual review.

[0029] A condenser condensing pipe plug recognition and positioning system based on deep learning, the system comprises:

[0030] A data preprocessing module is configured to perform multi-stage morphological preprocessing on the collected condenser tube sheet images, solve the problems of reflection, obstruction and noise, and perform edge enhancement, and all morphological operations are directly performed on RGB three-channel images to retain original color information.

[0031] A plug target detection module is configured to identify plugs in the condenser tube sheet images by using a trained plug target detection model.

[0032] For the backbone network in the YOLOv8 model, a receptive field attention convolution module RFAConv is used to replace the C2F module; for the Neck structure in the YOLOv8 model, a P2 detection branch is added to improve the detection capability of dense small plugs; an improved bidirectional feature pyramid network BiFPN is used to replace the PANet, and a lossless compression module SPDConv is used to replace the original downsampling operation.

[0033] The input of the plug target detection model is the preprocessed image, and the output is the plug category, plug position and confidence of whether the target is contained.

[0034] Further, a post-processing module is further included to compare the confidence of whether the target is contained output by the trained plug target detection model with a confidence threshold, and when the confidence output by the model is less than the confidence threshold, the current recognition result is marked as a "suspicious target" for subsequent manual review.

[0035] A computer-readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement a condenser condensing pipe plug recognition and positioning method based on deep learning.

[0036] The beneficial effects of the present application are as follows:

[0037] The present application provides a condenser condensing pipe plug recognition and positioning method and system, which can effectively extract the position and category of the plug by performing targeted preprocessing on the condenser tube sheet image and constructing a plug recognition model. Compared with the manual detection method, the present application can improve the detection efficiency while ensuring the recognition accuracy, and reduce the labor cost and the risk of missed detection. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The figure is a schematic diagram of the condenser tube sheet image in the background art.

[0039] Figure 2 The figure is a flowchart of the condenser condensing pipe plug recognition and positioning method based on deep learning according to an embodiment of the present application.

[0040] Figure 3 The figure shows the results of each stage of the image preprocessing of the condenser pipe plug in the embodiment of the present application, wherein (a) is the original color image, (b) is the median filter denoising image, (c) is the multi-scale top-hat reflection suppression image, (d) is the black hat guided repair occlusion image, (e) is the open operation cleaning image, and (f) is the final edge enhancement result image.

[0041] Figure 4 The network architecture diagram of the improved YOLOv8 model in the embodiment of the present application.

[0042] Figure 5 The condenser pipe plug recognition result diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0043] The purposes and effects of the present application will become more apparent from the following detailed description of the preferred embodiments of the present application with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0044] Explanation of technical terms:

[0045] Sobel: Sobel Operator, Sobel operator, is an edge detection operator;

[0046] YOLOv8: You Only Look Once version 8, single-stage target detection algorithm of the eighth version of YOLO series;

[0047] RFAConv: Receptive Field Attention Convolution, receptive field attention convolution;

[0048] C2F: Cross Stage Partial with 2 Feature Fusion, cross-stage partial connection double-branch feature fusion structure;

[0049] BiFPN: Bidirectional Feature Pyramid Network, bidirectional feature pyramid network;

[0050] PANet: Path Aggregation Network, path aggregation network;

[0051] SPDConv: Spatial Pyramid Dilated Convolution, spatial pyramid dilated convolution;

[0052] Shape-IoU: Shape Intersection over Union, shape intersection over union

[0053] CIoU: Complete Intersection over Union, complete intersection over union

[0054] As Figure 2 shown, in one aspect, an embodiment of the present application provides a condenser pipe plug identification and positioning method based on deep learning, comprising the following steps:

[0055] Step 1: Collect condenser tube plate images containing plugs and label the plugs in the condenser tube plate images to construct a data set; the labeling content includes the position and category of the plug, wherein the category of the plug includes plastic plug, welded plug and no plug.

[0056] The condenser tube plate image of the plug includes the design image data of the condenser tube plate, the spliced real image, etc.

[0057] Step 2: Perform multi-stage morphological preprocessing on the images in the data set to solve the problems of reflection, occlusion and noise, and perform edge enhancement, and all morphological operations are directly applied to the RGB three-channel image to retain the original color information; use the preprocessed image and its labeled information to construct a training data set;

[0058] Step 2 is implemented through the following sub-steps:

[0059] (2.1) Median filter denoising: median filter is performed on the original color image (RGB three channels) to suppress high-frequency noise and retain edge details;

[0060] (2.2) Multi-scale top-hat transformation reflection suppression: an elliptical structure element is used to independently perform top-hat transformation on each channel of the color image to extract and suppress the reflection area, and then a reflection-suppressed enhanced image is generated through channel weighting fusion, and the overall contrast is further improved through brightness weighting superposition;

[0061] (2.3) Black hat guided repair of occluded areas: convert the image to a grayscale image, extract the dark spot features of the black hat transformation to generate a mask of the occluded area, combine threshold segmentation and closing operation to optimize the mask, identify the area in the tube plate image that is occluded by soil, and expand the occluded area through dilation operation, and finally repair the soil occluded area in the color image;

[0062] (2.4) Open operation to clean up residual noise: for the repaired color image, an elliptical structure element is used to perform morphological open operation to eliminate isolated small-scale noise points while maintaining the edge and structural integrity of the target area;

[0063] (2.5) Sobel Edge Enhancement: The processed image is converted to a grayscale image, and the gradient magnitude is extracted using the Sobel operator to generate a binarized feature map for edge enhancement. The binarized feature map for edge enhancement is then superimposed onto the original image using weights to form an enhanced image. During subsequent model training, this enhanced image and the plug category in the image are used as the input to the model.

[0064] like Figure 3 As shown, the original image suffers from uneven lighting, strong reflections, and blurred edges. Some endcaps have unclear boundaries due to reflections or occlusions, such as... Figure 3 In (a), median filtering reduces high-frequency noise and makes the image smoother while retaining edge information, slightly improving the overall image quality, especially making the pipe hole area more uniform. Figure 3 (b) In this context, after multi-scale top-cap reflection suppression, the reflection effect in bright areas of the image is suppressed, the white reflection of the aperture is significantly reduced, and the structure is more uniform and prominent, as shown in (b). Figure 3 (c) Then, through black hat guided repair of occlusion, the occluded and dark areas were repaired and compensated, making the edge contour more complete and better restoring the blockage area that was obscured by light or dirt, such as Figure 3 (d); then, an opening operation is performed to clean up the image, removing small noise, burrs, and isolated bright spots. This makes the extraction of structured elements clearer, which is beneficial for subsequent edge detection, such as... Figure 3 (e) Finally, edge reinforcement is performed, resulting in a significant enhancement of the plug edge and a more prominent circular structure, such as... Figure 3 (f) in the equation provides a good foundation for subsequent detection or segmentation algorithms.

[0065] pass Figure 3 The difference between the original image (a) and the preprocessed image (f) is obvious. In the preprocessed image, the edges of each plug are clear, the shape is uniform, the background noise interference is small, and the image structure is prominent. The improvement in image quality helps to make the plug features more identifiable.

[0066] Step 3: Construct a blockage target detection model and train it using the training dataset. The blockage target detection model is an improved YOLOv8 model, replacing the C2F structure in the backbone of the original YOLOv8 model to enhance shape awareness. It introduces a P2 detection branch and a BiFPN-SPDConv fusion module to improve the detection capability of small targets. A new loss function is constructed by combining Shape-IoU and CIoU to adapt to the bounding box regression problem of dense blockages.

[0067] like Figure 4 As shown, the specific improvements to the original YOLOv8 model are as follows:

[0068] For the backbone network Backbone in the YOLOv8 model, the receptive field attention convolution module RFAConv is used to replace the C2F module, the extraction ability of local geometry is enhanced, the sensitivity of the round plug is improved, and in the actual scene of high density and slight deformation, the overall detection accuracy is improved.

[0069] According to the characteristics that the plugs in the condenser tube plate image are arranged densely and the spacing is extremely small, the detection branch starting from the P3 layer in the traditional structure is easy to cause detail loss, and the identification accuracy of small size targets is affected, therefore, for the Neck structure in the YOLOv8 model, a new P2 detection branch is added, more high-resolution features are retained for detection, and the detection ability of dense small plugs is improved.

[0070] In addition, the improved bidirectional feature pyramid network BiFPN is used to replace the PANet, and the lossless compression module SPDConv module is used to replace the original down-sampling operation; by using the weighted fusion mechanism of BiFPN and the SPDConv module to optimize the down-sampling, the detection performance of the model on the dense small targets such as the condenser plug is improved on the premise of ensuring the detection efficiency.

[0071] Finally, when training the improved YOLOv8 model, the loss function is designed as follows:

[0072] The original loss function of the YOLOv8 model is:

[0073]

[0074] Among them, L box represents the boundary box regression loss, which is used to measure the position difference between the predicted frame and the real frame, the original method uses CIoU loss function, the present application modifies it, designs a new regression loss function with adjustable weight, introduces Shape-IoU loss function and original CIoU loss function for combination, forms a composite loss function for dense and repeated small plugs, and pays more attention to the shape and scale of the boundary box, and is more sensitive to the specific shape, so as to help optimize the regression effect of the boundary box, therefore:

[0075]

[0076] Among them, L Shacpe-IoU is an improved IoU that pays attention to the shape contour matching degree of the predicted boundary box and the real boundary box, compared with CIoU, Shape-IoU introduces the shape information of the target, which makes up for the deficiency that CIoU does not fully consider the specific shape contour of the target, so as to adapt to the task of regular graphics and small targets.

[0077] Among them, L Shacpe-IoU is the specific calculation formula as follows:

[0078]

[0079] wherein, w p , h p represent the width and height of the prediction box, w p , h p represent the width and height of the GT (ground truth) box, and β x , β y represent the weighting coefficients related to the shape of the GT box, reflecting the sensitivity of the width and height to the error.

[0080] L cls is the classification loss, which measures the difference between the predicted class and the true class, and is improved using the Focal Loss for the classification branch to alleviate the imbalance problem between the classes and improve the detection sensitivity of the minority abnormal plugs.

[0081]

[0082] L obj represents the target loss, which determines whether there is an object in the anchor box.

[0083] Therefore, the loss function of the improved YOLOv8 model in the present application is:

[0084]

[0085] Step four: after the same pre-processing as step two is performed on the condenser tube plate image to be identified, the trained plug target detection model is input to perform real-time plug identification and positioning.

[0086] In order to further improve the identification accuracy and reliability of the trained plug target detection model, step five can also be added:

[0087] Step five: compare the confidence whether the target is contained output by the trained plug target detection model with the confidence threshold, and when the confidence output by the model is less than the confidence threshold, mark the identification result this time as "suspicious target" for subsequent manual review. The historical data is used to assist in judgment in the manual review stage, and the manual review result is automatically returned to the training database to trigger the incremental learning to update the model.

[0088] On the other hand, another embodiment of the present application provides a condenser condensing pipe plug identification and positioning system based on deep learning, which is used to realize the condenser condensing pipe plug identification and positioning method based on deep learning. The system comprises a data preprocessing module, a plug target detection module and a post-processing module.

[0089] The data preprocessing module is used for multi-stage morphological preprocessing of the collected condenser tube plate image, solves the problems of reflection, shielding and noise, and performs edge enhancement, and all morphological operations are directly performed on the RGB three-channel image to retain the original color information.

[0090] The plug target detection module is built-in with a trained plug target detection model, which is used to identify the plug in the condenser tube plate image.

[0091] For the backbone network in the YOLOv8 model, the receptive field attention convolution module RFAConv is used to replace the C2F module; for the Neck structure in the YOLOv8 model, the P2 detection branch is added to improve the detection capability of dense small plugs; the improved bidirectional feature pyramid network BiFPN is used to replace the PANet, and the lossless compression module SPDConv is used to replace the original downsampling operation; the input of the plug target detection model is the preprocessed image, and the output is the plug category, plug position and confidence of whether the target is contained.

[0092] The post-processing module is used for comparing the confidence of whether the target is contained output by the trained plug target detection model with a confidence threshold, and when the confidence output by the model is less than the confidence threshold, the current recognition result is marked as a “suspicious target” for subsequent manual review.

[0093] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts are described in the method embodiment. The system embodiment described above is only illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. According to actual needs, some or all of the modules can be selected to achieve the purpose of the present application. Those skilled in the art can understand and implement it without creative labor.

[0094] The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the condenser condensing pipe plug identification and positioning method based on deep learning in the above embodiment.

[0095] The computer readable storage medium can be an internal storage unit of any of the aforementioned data processing capable devices, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device, such as a plug-in hard disk, a SmartMedia card (SMC), an SD card, a Flash card, etc. Further, the computer readable storage medium can include both an internal storage unit of any of the aforementioned data processing capable devices and an external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the aforementioned data processing capable devices, and can also be used to temporarily store data that has been output or will be output.

[0096] Figure 5 A result diagram of the plug identification using the method and system of the embodiment, Figure 5 The left diagram includes the whole condenser tube plate area, including the center structure and multiple plugs densely distributed, showing the complete detection target area and the actual application scenario. Figure 5 The right diagram includes a local enlarged view of the red frame area, showing the actual appearance of multiple plugs and the visualization effect after the detection algorithm identification, showing multiple detection boxes (red circles / boxes) and the occlusion area. Figure 5 It can be seen that the method of the embodiment has good visualization effect for plug identification, and can accurately and clearly identify the plugs.

[0097] Those skilled in the art can understand that the above description is only preferred examples of the application and is not used to limit the application, although the application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or make equivalent replacements for part of the technical features. Any modification, equivalent replacement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A deep learning-based condenser condensing pipe plug identification and positioning method, characterized in that, The method comprises the following steps: Step 1: Collecting the image of the condenser tube plate containing the plug, and labeling the position and category of the plug in the image of the condenser tube plate to construct a data set; Step 2: The image in the data set is subjected to multi-stage morphological preprocessing to solve the problems of reflection, occlusion and noise, and edge enhancement is performed, and all morphological operations are directly applied to the RGB three-channel image to retain the original color information; A training data set is constructed by using the preprocessed image and its labeled information; Step 3: A plug target detection model is constructed, and the plug target detection model is trained by using the training data set; the plug target detection model is an improved YOLOv8 model, and the specific improvement is as follows: For the backbone network in the YOLOv8 model, the receptive field attention convolution module RFAConv is used to replace the C2F module; For the Neck structure in the YOLOv8 model, a P2 detection branch is added to improve the detection capability of dense small plugs; an improved bidirectional feature pyramid network BiFPN is used to replace the PANet, and a lossless compression module SPDConv is used to replace the original downsampling operation; The input of the plug target detection model is the preprocessed image, and the output is the plug category, plug position and confidence whether the target is contained; Step 4: After the condenser tube plate image to be recognized is preprocessed in the same way as step 2, the trained plug target detection model is inputted to perform real-time plug recognition and positioning.

2. The deep learning-based condenser condensate tube block identification and positioning method according to claim 1, characterized in that, When training the plug target detection model, the loss function is a boundary box regression loss L box , a classification loss L cls , and a target loss L obj a weighted sum of the three losses. The bounding box regression loss L box Loss for attention prediction bounding box and real bounding box contour matching degree L Shacpe-IoU Weighted sum of the original loss function L CIoU ​ The classification loss L cls The Focal Loss loss function is used.

3. The deep learning-based condenser condensate tube block identification and localization method of claim 2, wherein, When training the plug target detection model, the weight coefficient of the CIoU loss function is higher than that of the Shape-IoU loss function in the early training stage to ensure that the model converges quickly and completes the preliminary positioning; the weight coefficient of the Shape-IoU loss function is gradually increased in the middle and later training stages to improve the boundary fitting accuracy of dense small plugs.

4. The deep learning-based condenser condensate tube block identification and localization method of claim 1, wherein, The step 2 comprises the following substeps: S2.1: The original color image is subjected to median filtering to suppress high-frequency noise and retain edge details; S2.2: The top-hat transformation is independently performed on each channel of the color image by using an elliptical structure element to extract and suppress the reflection area, and then an enhanced image with suppressed reflection is generated by channel weighting fusion, and the overall contrast is further improved by brightness weighting superposition; S2.3: The image obtained in S2.2 is converted into a gray image, the dark spot features of the bottom-hat transformation are extracted to generate a mask of the occlusion area, the mask is optimized by combining threshold segmentation and closing operation, the area of the tube plate image with soil occlusion is identified, and the soil occlusion area in the color image is repaired by expansion operation; S2.4: The repaired color image obtained in S2.3 is subjected to morphological opening operation by using an elliptical structure element to eliminate isolated small-scale noise points while maintaining the edge and structural integrity of the target area; S2.5: The image obtained in S2.4 is converted into a gray image, the gradient amplitude is extracted by a Sobel operator to generate an edge-enhanced binary feature map; and the edge-enhanced binary feature map is superimposed on the original image by weighting to form an enhanced image.

5. The deep learning-based condenser condensate tube block identification and localization method of claim 1, wherein, The plug types include plastic plugs, welded plugs and no-plug.

6. The deep learning-based condenser condensate tube block identification and localization method of claim 1, wherein, The image of the condenser tube plate containing the plug collected in step one includes design image data of the condenser tube plate and a spliced physical image of the condenser tube plate.

7. The deep learning-based condenser condensate tube block identification and localization method of claim 1, wherein, Step five is further included: The confidence whether the target is contained output by the trained plug target detection model is compared with a confidence threshold, and when the confidence output by the model is less than the confidence threshold, the current recognition result is marked as a "suspicious target" for subsequent manual review.

8. A deep learning based condenser condensing pipe plug identification and positioning system, characterized in that, The system comprises: A data preprocessing module is configured to perform multi-stage morphological preprocessing on the collected condenser tube plate image, solve the problems of reflection, occlusion and noise, and perform edge enhancement, and all morphological operations are directly performed on RGB three-channel images to retain original color information; A plug target detection module is configured to have a trained plug target detection model built-in and recognize plugs in the condenser tube plate image; the plug target detection model is an improved YOLOv8 model, and the specific improvements are as follows: For the main network in the YOLOv8 model, a receptive field attention convolution module RFAConv is used to replace the C2F module; for the Neck structure in the YOLOv8 model, a P2 detection branch is added to improve the detection capability for dense small plugs; an improved bidirectional feature pyramid network BiFPN is used to replace the PANet, and a lossless compression module SPDConv is used to replace the original downsampling operation; The input of the plug target detection model is the preprocessed image, and the output is the plug type, the plug position and the confidence whether the target is contained.

9. The deep learning-based condenser condensate tube block identification and localization system of claim 8, wherein, A post-processing module is further included to compare the confidence whether the target is contained output by the trained plug target detection model with a confidence threshold, and when the confidence output by the model is less than the confidence threshold, the current recognition result is marked as a "suspicious target" for subsequent manual review.

10. A computer-readable storage medium, characterized in that, A program is stored thereon, and the program is executed by the processor to implement the deep learning-based condenser condensing tube plug recognition and positioning method in any one of claims 1-7.

Images