A surface defect detection system for water control gates

Abstract

The application discloses a kind of water conservancy gate surface defect detection systems, belong to gate surface defect detection technical field, including following module: adaptive surface image acquisition module: can be attached gate board face, weld and curved surface are inspected;Defect image pre-processing enhancement module: to gate corrosion, crack, peeling, wear and tear and other defect images are denoised, anti-reflection, correction optimization;Lightweight multi-scale defect identification module: improved YOLO network is used;Defect quantification early warning module: accurately positioning and quantifying size parameters to detected defects;Data operation and maintenance interactive module: store image, defect data and historical detection ledger;The application identifies network integrated multi-scale feature pyramid and channel attention structure, can extract different size disease characteristics such as fine crack, large area rust simultaneously, autonomously filter out invalid background interference such as water stain, stain, through channel screening pruning and fixed-point quantization to network lightweight processing, eliminate redundant convolution parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of gate surface defect detection, and specifically, relates to a water conservancy gate surface defect detection system. Background Art

[0002] Steel gates supporting reservoirs, rivers and pumping stations are core components of water conservancy hubs for water storage, water retaining, scheduling and flood control; the gates are in a working environment full of water vapor, alternating dry and wet conditions, and continuous sediment scouring all year round, so the gate plate surfaces, splicing welds, and edge and corner areas are extremely prone to damages such as rust, coating peeling and falling off, fine cracks, and plate surface wear; if such surface damages are not detected and disposed of in time, they will continue to expand inward, continuously weaken the overall structural strength of the gate, and further cause leakage and plate surface deformation, and in severe cases directly result in component failure, threatening the operation safety of water conservancy projects; therefore, regularly carrying out defect inspection on the gate surface is an indispensable work link in the daily operation and maintenance of water conservancy.

[0003] Therefore, the present invention is made to improve the related problems, and a water conservancy gate surface defect detection system is proposed. Summary of the Invention

[0004] In view of the above problems existing in the prior art, the purpose of the present invention is to provide a water conservancy gate surface defect detection system.

[0005] To solve the above problems, the technical solution adopted by the present invention is as follows: A water conservancy gate surface defect detection system, comprising the following modules: Adaptive surface image acquisition module: capable of performing inspection by fitting to the gate plate surface, welds and curved surfaces, adaptively adjusting imaging parameters according to the on-site illumination and water mist working conditions, and collecting high-definition images of the gate surface in full coverage. Defect image preprocessing enhancement module: performing noise reduction, reflection removal, correction and optimization on defect images of gate rust, cracks, spalling, wear and the like to eliminate environmental interference; generating defect samples through an improved GAN network to supplement the data set, solving the problem of insufficient samples and improving the model detection adaptation capability. Lightweight multi-scale defect identification module: adopting an improved YOLO network, combining multi-scale feature fusion and attention mechanism to accurately extract various gate surface defect features; reducing the computing power consumption of equipment through model lightweight processing. Defect quantification and early warning module: accurately locating the detected defects and quantifying the size parameters, determining the risk level according to industry standards, performing acousto-optic early warning on defects exceeding standards, and automatically generating a detection report containing defect information and maintenance suggestions. Data operation and maintenance interaction module: storing images, defect data and historical detection accounts, supporting local and cloud data storage, capable of tracing and comparing the change trend of diseases, and providing data support for gate operation and maintenance and safety assessment.

[0006] Preferably, the adaptive surface image acquisition module includes a magnetic walking mechanism, an angle-adjustable high-definition camera component, a ring-shaped soft light supplementary lighting component, and an attitude sensing unit. The magnetic walking mechanism uses permanent magnet adsorption to adhere to the steel gate and walks without the need for a preset track. It can adapt to the inspection movement of the gate's flat surface, curved surface, and weld seam unevenness. Even if the gate surface has dust accumulation, moisture, or slight corrosion, it can still ensure firm adsorption, stable movement, and no slippage deviation. The camera component can flexibly adjust the angle according to the shooting position, and the ring-shaped soft light supplementary lighting improves the on-site lighting conditions.

[0007] Preferably, the defect image preprocessing enhancement module integrates multi-level image optimization algorithms, including Gaussian filtering for noise reduction, adaptive reflection suppression, lens distortion correction, and local contrast enhancement. Based on multi-level digital image processing technology, it smooths image noise through spatial filtering, and combines color space transformation and contrast stretching algorithms to weaken environmental interference, enhance weak defect features such as cracks and corrosion, and improve the image signal-to-noise ratio. It also filters out image noise caused by water mist and dust using a Gaussian filtering formula. ; In the formula, Image coordinates Gaussian filter weights at the location, These are the relative coordinates of the image pixels. The standard deviation of the Gaussian kernel determines the range of filtering and smoothing. The larger the image, the stronger the noise reduction and smoothing effect. It can adapt to different levels of water mist, dust and noise interference on site, and optimize on-site images with different noise intensities.

[0008] Preferably, the improved YOLO network of the lightweight multi-scale defect recognition module is equipped with a multi-scale feature pyramid and channel attention mechanism. Based on the principle of multi-scale feature hierarchical fusion, it takes into account both shallow details and deep semantic features. Through the principle of adaptive channel weight allocation, it autonomously selects effective defect features, shields invalid background interference such as water stains and dirt, and adapts to the detection of defects in gates of various sizes and shapes. By fusing deep and shallow image features through the multi-scale feature pyramid, it adapts to the recognition of defects of different sizes, based on the weight formula: ; In the formula, This represents the final weight value for a single channel. This represents the total number of network feature channels. For the first Feature parameters of each feature channel The activation function normalizes the accumulated feature parameters to achieve adaptive allocation of defect feature channel weights, thereby adaptively strengthening defect features and shielding background interference.

[0009] Preferably, the defect quantification and early warning module presets a pixel-to-physical size conversion coefficient. Based on image edge contour extraction technology and the pixel-to-physical mapping conversion principle, it accurately locks the defect boundary, converts two-dimensional pixel coordinates into actual physical dimensions, and combines industry standards to achieve defect quantification and risk classification management; relying on the defect quantification formula: , ; In the formula, L is the actual maximum span length of the defect; k is the scale conversion factor. , The pixel coordinates of the defect contour This is a scale conversion factor. The area of ​​a single pixel region. The length formula is used to calculate the actual length of the maximum span of the defect outline, which represents the total number of pixels covered by the defect. The area formula is used to calculate the actual total area of ​​the defect by accumulating pixels one by one, which can accurately adapt to the size quantification calculation of irregular defects.

[0010] Preferably, the data operation and maintenance interaction module adopts a dual backup architecture of local solid-state storage combined with 4G / 5G cloud transmission; based on the dual backup storage principle of local caching and cloud synchronization, it avoids the risk of data loss, and combines the principle of time series statistical analysis to compare and fit multi-cycle detection data, explore the evolution pattern of diseases, realize the visualization of operation and maintenance data and the predictability of trends, and realize the permanent retention of detection data and the traceability and retrieval at any time by caching various detection data locally in real time and archiving ledger data in the cloud for a long time.

[0011] Preferably, the improved GAN network of the defect image preprocessing enhancement module adopts a texture and contour dual discriminator structure. Based on the principle of alternating balance training of dual discriminators, it constrains the microscopic texture features and macroscopic contour morphology features of the defect respectively, and corrects the generation image error in both directions, greatly improving the realism and fit of the simulated defect sample; using the dual-branch loss formula: + ; In the formula, This represents the total network loss value. This is the texture feature loss function, responsible for constraining the microscopic texture details of defects. The contour morphology loss function is responsible for constraining the overall geometry of the defect. The weight coefficient is used to balance the training proportion of the two losses, ensuring that the generated sample texture and contour are both accurate, and is trained iteratively through dual discriminators.

[0012] Preferably, the lightweight multi-scale defect recognition module achieves model lightweight optimization through structured pruning and fixed-point quantization. It eliminates invalid and redundant parameters based on the convolutional channel contribution screening and pruning principle, and reduces model storage and computational overhead by combining fixed-point quantization compression principles. This improves model inference speed while maintaining recognition accuracy, making it suitable for low-computing-power terminal devices. The channel scoring formula is as follows: ; In the formula, The feature response score for a single convolutional channel. , These represent the height and width of the feature map, respectively. , The feature value is the feature value of the corresponding pixel position in the feature map. The higher the score, the greater the feature contribution of the channel. Conversely, the lower the score, the more redundant channels can be removed. This provides an accurate basis for model pruning, eliminating redundant parameters and invalid network layers.

[0013] Preferably, the defect quantification and early warning module presets the coordinate parameters of each structural area of ​​the gate. Based on the preset structural coordinate calibration technology and the Euclidean distance matching positioning principle, it completes the accurate matching and classification of defect areas through pixel coordinate distance calculation, realizing refined point positioning of defects; through the coordinate matching formula: ; In the formula, The distance between the defect center and the reference points of each section of the gate is expressed in Euclidean form. The coordinates of the defect center are... Using the gate partition reference coordinates, the gate partition corresponding to the minimum distance is matched by calculating the straight-line distance between pixel coordinates to determine the structural area to which the defect belongs.

[0014] Preferably, the data operation and maintenance interaction module is equipped with dual-mode inspection control logic; it realizes unmanned routine inspection based on timed and sequential triggering logic, and realizes fixed-point precise re-inspection by combining manual command interaction control logic. The dual modes complement each other to adapt to various operation and maintenance conditions, improve the system's practicality and flexibility, and support switching between timed automatic inspection and manual inspection. The timed mode can autonomously complete the unmanned inspection process of full-domain collection, identification, and archiving, while the manual mode supports fixed-point re-inspection and area re-scanning.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This system recognizes a network that integrates multi-scale feature pyramids and channel attention structures. It can simultaneously extract features of different sizes of defects, such as fine cracks and large-area corrosion. It can autonomously filter out invalid background interference such as water stains and dirt. Through channel filtering pruning and fixed-point quantization, the network is lightweighted and redundant convolution parameters are eliminated. While ensuring recognition accuracy, the computational power requirement of the model is reduced. It can be equipped with an embedded terminal on site to achieve real-time detection, solving the problems of high missed detection and false detection rates, high hardware costs, and difficulty in real-time on-site recognition in traditional detection methods.

[0016] 2. The system establishes a pixel-to-actual physical size conversion model and automatically calculates the actual length and total area of ​​defects using a contour coordinate algorithm. It pre-stores the reference coordinates of each zone of the gate and accurately locates the plate surface, weld, and corner area to which the defect belongs through pixel Euclidean distance matching. Based on water conservancy industry standards, the system classifies the risk level of defects. Defects exceeding the standard trigger audible and visual warnings and automatically generate inspection reports that record the location, size, and maintenance suggestions of defects. This overcomes the shortcomings of manual visual inspection, which can only make qualitative judgments, and realizes the quantitative archiving of defect data, thereby improving the standardization of gate safety operation and maintenance. Attached Figure Description

[0017] Figure 1 This application provides an architecture diagram of a surface defect detection system for hydraulic gates. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0019] A surface defect detection system for hydraulic gates includes the following modules; Adaptive surface image acquisition module: It can be fitted to the gate plate, weld and curved surface for inspection. It adaptively adjusts the imaging parameters according to the on-site lighting and water mist conditions, and acquires high-definition images of the gate surface with full coverage, eliminating detection blind spots and image interference problems. Defect image preprocessing enhancement module: Denoise reduction, de-glare removal, correction and optimization of defect images such as rust, cracks, peeling, and wear of gates, and removal of environmental interference; Defect samples are generated by improving the GAN network to supplement the dataset, solve the problem of insufficient samples, and improve the model's detection and adaptation capabilities; Lightweight multi-scale defect recognition module: It adopts an improved YOLO network, combined with multi-scale feature fusion and attention mechanism, to accurately extract the surface defect features of various gates; through lightweight model processing, it reduces the computing power consumption of the equipment, realizes real-time on-site detection, and reduces the phenomenon of missed and false defects. Defect Quantification and Early Warning Module: Accurately locates and quantifies the size parameters of detected defects, determines the risk level according to industry standards, provides audible and visual warnings for defects exceeding the standard, and automatically generates a test report containing defect information and maintenance suggestions; Data operation and maintenance interaction module: Stores images, defect data and historical inspection logs, supports local and cloud data storage, can trace and compare disease change trends, and provides data support for gate operation and maintenance and safety assessment.

[0020] Furthermore, the adaptive surface image acquisition module includes a magnetic walking mechanism, an angle-adjustable high-definition camera component, a ring-shaped soft light supplementary lighting component, and an attitude sensing unit. The magnetic walking mechanism uses permanent magnet adsorption to adhere to the steel gate and walks without the need for a pre-set track. It can adapt to the inspection movement of the gate's flat, curved, and welded uneven positions. Even if the gate surface has dust accumulation, moisture, or slight corrosion, it can still ensure firm adsorption, stable movement, and no slippage deviation. The camera component can flexibly adjust its angle according to the shooting position. With the ring-shaped soft light supplementary lighting, it can improve the on-site lighting conditions, eliminate strong light reflection, backlight, and shadow problems, and combine with the real-time attitude feedback from the attitude sensing unit to unify the imaging angle and brightness of each inspection position, realizing the acquisition of images of the entire gate area without blind spots.

[0021] Furthermore, the defect image preprocessing and enhancement module integrates multi-level image optimization algorithms, including Gaussian filtering for noise reduction, adaptive reflection suppression, lens distortion correction, and local contrast enhancement. Based on multi-level digital image processing technology, it smooths image noise through spatial filtering, and combines color space transformation and contrast stretching algorithms to weaken environmental interference, enhance weak defect features such as cracks and corrosion, and improve the image signal-to-noise ratio. It also filters out image noise caused by water mist and dust using a Gaussian filtering formula. ; In the formula, Image coordinates Gaussian filter weights at the location, These are the relative coordinates of the image pixels. The standard deviation of the Gaussian kernel determines the range of filtering and smoothing. The larger the image, the stronger the noise reduction and smoothing effect. It can adapt to different levels of water mist, dust and noise interference on site, and optimize on-site images with different noise intensities. It effectively enhances the features of subtle defects, removes environmental interference, and provides high-quality image input for defect recognition models.

[0022] Furthermore, the improved YOLO network of the lightweight multi-scale defect recognition module incorporates a multi-scale feature pyramid and channel attention mechanism. Based on the principle of multi-scale feature hierarchical fusion, it balances shallow details and deep semantic features. Through the principle of adaptive channel weight allocation, it autonomously selects effective defect features, shielding invalid background interference such as water stains and dirt, thus adapting to the detection of defects in gates of various sizes and shapes. By fusing deep and shallow image features through the multi-scale feature pyramid, it adapts to the recognition of defects of different sizes, based on the weight formula: ; In the formula, This represents the final weight value for a single channel. This represents the total number of network feature channels. For the first Feature parameters of each feature channel The activation function normalizes the accumulated feature parameters to achieve adaptive allocation of defect feature channel weights, adaptively enhances defect features, and shields background interference. Combined with lightweight structure optimization, it balances recognition accuracy and real-time detection performance, reducing the probability of missed detections and false detections.

[0023] Furthermore, the defect quantification and early warning module presets pixel-to-physical size conversion coefficients. Based on image edge contour extraction technology and pixel-to-physical mapping conversion principles, it accurately locks the defect boundary, converts two-dimensional pixel coordinates into actual physical dimensions, and combines industry standards to achieve defect quantification and risk classification management; relying on the defect quantification formula: , ; In the formula, L is the actual maximum span length of the defect; k is the scale conversion factor. , The pixel coordinates of the defect contour This is a scale conversion factor. The area of ​​a single pixel region. The total number of pixels covered by the defect is used to calculate the actual length of the maximum span of the defect outline using the length formula. The area formula calculates the actual total area of ​​the defect by accumulating pixels one by one. It can accurately adapt to the size quantification calculation of irregular defects. Based on industry standards, the quantified defects are classified into risk levels, and audible and visual early warning operations are performed according to the level. A complete inspection report with maintenance suggestions is automatically generated.

[0024] Furthermore, the data operation and maintenance interaction module adopts a dual backup architecture combining local solid-state storage with 4G / 5G cloud transmission. Based on the dual backup storage principle of local caching and cloud synchronization, it avoids the risk of data loss. Combining the principle of time-series statistical analysis, it compares and fits multi-period detection data to explore the evolution pattern of defects, realizes the visualization of operation and maintenance data and the predictability of trends. By caching various detection data locally in real time and archiving ledger data in the cloud for a long time, it realizes the permanent retention of detection data and the ability to trace and retrieve it at any time. At the same time, by comparing multi-period detection data, it analyzes the evolution pattern of gate defects and generates trend statistical reports, providing complete data support for the routine operation and maintenance of gates, defect rectification and safety assessment.

[0025] Furthermore, the improved GAN network in the defect image preprocessing enhancement module adopts a texture and contour dual discriminator structure. Based on the principle of alternating balance training of the dual discriminators, it constrains the microscopic texture features and macroscopic contour morphology features of the defects respectively, and corrects the generation image error in both directions, greatly improving the realism and fit of the simulated defect samples; using the dual-branch loss formula: + ; In the formula, This represents the total network loss value. This is the texture feature loss function, responsible for constraining the microscopic texture details of defects. The contour morphology loss function is responsible for constraining the overall geometry of the defect. The weight coefficient is used to balance the training proportion of the two losses, ensuring that the generated sample texture and contour are both accurate. Through alternating iterative training with dual discriminators, defect samples that fit the real working conditions are generated, effectively supplementing the dataset, solving the problem of scarce on-site defect samples, and improving the model's generalization ability.

[0026] Furthermore, the lightweight multi-scale defect recognition module achieves model lightweight optimization through structured pruning and fixed-point quantization. It eliminates invalid and redundant parameters based on the convolutional channel contribution screening and pruning principle, and reduces model storage and computational overhead by combining fixed-point quantization compression principles. This improves model inference speed while maintaining recognition accuracy, making it suitable for low-computing-power terminal devices. The channel scoring formula is as follows: ; In the formula, The feature response score for a single convolutional channel. , These represent the height and width of the feature map, respectively. , The feature value is the feature value of the corresponding pixel position in the feature map. The higher the score, the greater the feature contribution of the channel. Conversely, the channel can be removed as a redundant channel. This provides an accurate basis for model pruning, removes redundant parameters and invalid network layers, and compresses the model size with 8-bit fixed-point quantization. This improves inference speed without reducing recognition accuracy and is suitable for real-time detection scenarios with low computing power on site.

[0027] Furthermore, the defect quantification and early warning module presets the coordinate parameters of each structural area of ​​the gate. Based on the preset structural coordinate calibration technology and the Euclidean distance matching positioning principle, it completes the accurate matching and classification of defect areas through pixel coordinate distance calculation, realizing refined point positioning of defects; through the coordinate matching formula: ; In the formula, The distance between the defect center and the reference points of each section of the gate is expressed in Euclidean form. The coordinates of the defect center are... Using the gate zone reference coordinates, the gate zone corresponding to the minimum distance is matched by calculating the straight-line distance between pixel coordinates to determine the structural area to which the defect belongs, thus achieving refined positioning and classification of defects. By matching Euclidean distance, the specific location of the defect, such as the plate surface, weld, or corner, is determined, providing accurate location basis for maintenance and rectification.

[0028] Furthermore, the data operation and maintenance interaction module is equipped with dual-mode inspection control logic; based on timed and sequential triggering logic, it realizes unmanned routine inspection, and combined with manual command interaction control logic, it realizes precise re-inspection at fixed points. The dual modes complement each other and adapt to various operation and maintenance conditions, improving the system's practicality and flexibility. It supports switching between timed automatic inspection and manual inspection. The timed mode can autonomously complete the unmanned inspection process of full-domain collection, identification, and archiving, while the manual mode supports fixed-point re-inspection and regional supplementary scanning, adapting to various conditions such as daily operation and maintenance, flood season spot checks, and defect verification, ensuring the flexibility and practicality of the system's operation and maintenance detection.

[0029] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0030] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A surface defect detection system for hydraulic gates, characterized in that, Includes the following modules: Adaptive surface image acquisition module: It can be fitted to the gate plate, weld and curved surface for inspection, and adaptively adjust the imaging parameters according to the on-site lighting and water mist conditions to acquire high-definition images of the gate surface with full coverage; Defect image preprocessing and enhancement module: Denoise reduction, anti-glare, correction and optimization are performed on defect images such as rust, cracks, peeling, and wear of gates to remove environmental interference; defect samples are generated by improving the GAN network to supplement the dataset; Lightweight multi-scale defect recognition module: It adopts an improved YOLO network, combined with multi-scale feature fusion and attention mechanism, to accurately extract the surface defect features of various gates; and reduces the computing power consumption of the equipment through model lightweighting. Defect Quantification and Early Warning Module: Accurately locates and quantifies the size parameters of detected defects, determines the risk level according to industry standards, provides audible and visual warnings for defects exceeding the standard, and automatically generates a test report containing defect information and maintenance suggestions; Data operation and maintenance interaction module: Stores images, defect data and historical inspection records, supports local and cloud data storage, and can trace and compare disease change trends.

2. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The adaptive surface image acquisition module includes a magnetic walking mechanism, an angle-adjustable high-definition camera component, a ring soft light supplementary lighting component, and an attitude sensing unit. The magnetic walking mechanism uses permanent magnet adsorption to adhere to the steel gate and walk without the need for a pre-set track. The camera component can flexibly adjust its angle according to the shooting position, and the ring soft light supplementary lighting improves the on-site lighting conditions.

3. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The defect image preprocessing and enhancement module integrates multi-level image optimization algorithms, including Gaussian filtering for noise reduction, adaptive reflection suppression, lens distortion correction, and local contrast enhancement. Based on multi-level digital image processing technology, it smooths image noise through spatial filtering, and combines color space transformation and contrast stretching algorithms to weaken environmental interference, enhance weak defect features such as cracks and corrosion, and improve the image signal-to-noise ratio. It also filters out image noise caused by water mist and dust using a Gaussian filtering formula. ； In the formula, Image coordinates Gaussian filter weights at the location, These are the relative coordinates of the image pixels. The standard deviation of the Gaussian kernel determines the range of filtering and smoothing. The larger the image, the stronger the noise reduction and smoothing effect.

4. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The improved YOLO network in the lightweight multi-scale defect recognition module incorporates a multi-scale feature pyramid and channel attention mechanism. Based on the principle of multi-scale feature hierarchical fusion, it balances shallow details and deep semantic features. Through adaptive channel weight allocation, it autonomously selects effective defect features, filtering out invalid background interference such as water stains and dirt, thus adapting to the detection of defects in gates of various sizes and shapes. By fusing deep and shallow image features through a multi-scale feature pyramid, it adapts to the recognition of defects of different sizes, based on the weight formula: ； In the formula, This represents the final weight value for a single channel. This represents the total number of network feature channels. For the first Feature parameters of each feature channel This is the activation function, which normalizes the accumulated feature parameters.

5. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The defect quantification and early warning module presets pixel-to-physical size conversion coefficients. Based on image edge contour extraction technology and pixel-to-physical mapping conversion principles, it accurately locks defect boundaries, converts two-dimensional pixel coordinates into actual physical dimensions, and combines industry standards to achieve defect quantification and risk classification management; relying on the defect quantification formula: 、 ； In the formula, L is the actual maximum span length of the defect; k is the scale conversion factor. , The pixel coordinates of the defect contour This is a scale conversion factor. The area of ​​a single pixel region. The total number of pixels covered by the defect is represented by the length formula, which is used to calculate the actual length of the maximum span of the defect outline. The area formula calculates the actual total area of ​​the defect by accumulating pixels one by one.

6. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The data operation and maintenance interaction module adopts a dual backup architecture of local solid-state storage combined with 4G / 5G cloud transmission; based on the dual backup storage principle of local caching and cloud synchronization, it avoids the risk of data loss, and combines the principle of time series statistical analysis to compare and fit multi-cycle detection data, explore the evolution pattern of diseases, and realize the visualization of operation and maintenance data and the predictability of trends.

7. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The improved GAN network of the defect image preprocessing enhancement module adopts a texture and contour dual discriminator structure. Based on the principle of alternating balance training of dual discriminators, it constrains the micro-texture features and macro-contour morphology features of defects respectively, and corrects the generation image error in both directions, greatly improving the realism and fit of the simulated defect samples; using the dual-branch loss formula: + ； In the formula, This represents the total network loss value. This is the texture feature loss function, responsible for constraining the microscopic texture details of defects. The contour morphology loss function is responsible for constraining the overall geometry of the defect. These are the weighting coefficients.

8. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The lightweight multi-scale defect recognition module achieves model lightweight optimization through structured pruning and fixed-point quantization. It eliminates invalid and redundant parameters based on the convolutional channel contribution screening and pruning principle, and reduces model storage and computational overhead by combining fixed-point quantization compression principles. This improves model inference speed while maintaining recognition accuracy, making it suitable for low-computing-power terminal devices. The module uses a channel scoring formula: ； In the formula, The feature response score for a single convolutional channel. , These represent the height and width of the feature map, respectively. , The feature value is the feature value of the corresponding pixel position in the feature map. The higher the score, the greater the feature contribution of that channel. Conversely, the lower the score, the more redundant channels can be removed, providing an accurate basis for model pruning.

9. The surface defect detection system for hydraulic gates according to claim 1, characterized in that, The defect quantification and early warning module presets the coordinate parameters of each structural area of ​​the gate. Based on the preset structural coordinate calibration technology and the Euclidean distance matching positioning principle, it completes the accurate matching and classification of defect areas through pixel coordinate distance calculation, realizing the fine-grained point positioning of defects; through the coordinate matching formula: ； In the formula, The distance between the defect center and the reference points of each section of the gate is expressed in Euclidean form. The coordinates of the defect center are... Using the gate partition reference coordinates, the gate partition corresponding to the minimum distance is matched by calculating the straight-line distance between pixel coordinates to determine the structural area to which the defect belongs.

10. A surface defect detection system for hydraulic gates according to claim 1, characterized in that, The data operation and maintenance interaction module is equipped with dual-mode inspection control logic; it realizes unmanned routine inspection based on timed and sequential triggering logic, and realizes precise re-inspection at fixed points by combining manual command interaction control logic. The dual modes complement each other to adapt to various operation and maintenance conditions, improving the system's practicality and flexibility.

Images