Fire pool defect detection method and system

Abstract

The invention provides a fire pool defect detection method and system. The method comprises the following steps: acquiring image data and angle data of a pool wall and pose data of a robot; performing image processing on the image data, and performing defect identification on an image processing result by using a convolutional neural network to obtain defect information; the defect information comprises the pixel offset of the defect position and the defect type; and performing coordinate calculation according to the pixel offset, the angle data and the pose data to obtain defect positioning. According to the invention, the robot is utilized to obtain the image data, angle data and pose data of the pool wall, and the convolutional neural network is combined to carry out defect identification and positioning, so that an automatic and high-precision detection process is realized. The problems of potential safety hazards, low efficiency, low accuracy and high operation cost caused by dependence on manual operation in a traditional fire pool defect inspection method are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fire protection engineering, and particularly to a method and system for detecting defects in fire pools. Background Art

[0002] In traditional methods for inspecting defects in fire pools, manual operation is mainly relied on. Inspectors need to wear diving equipment and enter the interior of the pool to check for defects in the coating on the inner wall of the pool through visual observation and manual operation. However, this method has many problems. From a safety perspective, manual inspection requires operation in water, posing a diving risk. Especially in the case of turbid water and low visibility, the operating environment is complex, increasing the work hazard and prone to causing personal injuries or accidents. Considering from the aspect of efficiency, manual operation is slow and time-consuming. When faced with large fire pools, it requires a large amount of manpower and time, with low inspection efficiency and difficult to meet the rapid detection requirements. From the aspect of accuracy, the results of manual inspection rely on the subjective judgment of inspectors, and may overlook minor defects due to human factors or misjudge due to individual experience differences, resulting in low accuracy of inspection results. Analyzing from the aspect of operation cost, manual operation requires professional diving equipment and systematic training and certification for inspectors to ensure their operation skills and safety awareness. Maintenance personnel also need to participate in training and certification regularly to maintain professional levels, all of which increase the operation cost. Summary of the Invention

[0003] The present invention provides a method and system for detecting defects in fire pools to solve the problem of misjudgment due to individual experience differences in traditional methods for inspecting defects in fire pools.

[0004] The technical solution adopted by the present invention to solve its technical problems is: to provide a method for detecting defects in a fire pool, including the following steps:

[0005] Step S1: Obtain image data, angle data of the pool wall, and pose data of the robot;

[0006] Step S2: Perform image processing on the image data, and use a convolutional neural network to identify defects on the result of the image processing to obtain defect information; the defect information includes the pixel offset of the defect position and the defect type;

[0007] Step S3: Perform coordinate calculation based on the pixel offset, the angle data, and the pose data to obtain defect positioning.

[0008] In one embodiment, the performing image processing on the image data includes:

[0009] Perform image processing on the image data based on a local contrast enhancement algorithm to obtain an enhanced image;

[0010] Performing image processing on the enhanced image based on a defogging algorithm to obtain the image processing result.

[0011] In one embodiment, performing image processing on the image based on a local contrast enhancement algorithm to obtain the enhanced image includes:

[0012] Performing image block division on the image to obtain a plurality of small regions;

[0013] Performing histogram equalization on the plurality of small regions and performing contrast limitation on the histogram of each small region to obtain a small region image processed by contrast limitation;

[0014] Performing interpolation processing on the small region image processed by contrast limitation to obtain the enhanced image.

[0015] In one embodiment, performing histogram equalization on the plurality of small regions includes:

[0016] Performing gray histogram statistics on each small region, calculating the distribution of its gray values to obtain a gray histogram;

[0017] Calculating the cumulative distribution function of each small region based on the gray histogram;

[0018] Using the cumulative distribution function to convert the original gray values of the pixels in each small region into new gray values to achieve histogram equalization.

[0019] In one embodiment, the defect information further includes a defect type;

[0020] Performing defect recognition on the image processing result using a convolutional neural network to obtain the defect information includes:

[0021] Inputting the result of the processed image into a pre-trained convolutional neural network model to obtain the bounding box coordinates of the defect, the defect type, and confidence data;

[0022] Adopting a non-maximum suppression algorithm to remove overlapping detection boxes based on the bounding box coordinates of the defect and the confidence data to obtain the final defect information.

[0023] In one embodiment, the defect information includes a defect result and confidence data; adopting a non-maximum suppression algorithm to remove overlapping detection boxes based on the bounding box coordinates of the defect and the confidence data to obtain the final defect information includes:

[0024] The first step: sorting the bounding box coordinates of the defect according to the confidence data;

[0025] Step 2: Select the bounding box coordinates with the highest confidence data as the reference box and add them to the final defect information;

[0026] Step 3: Calculate the intersection over union (IoU) of the remaining bounding box coordinates and the reference box;

[0027] Step 4: If the calculated IoU is greater than the preset threshold, remove the corresponding bounding box coordinates from the final defect information;

[0028] Repeat the above Step 2 to Step 4 until all bounding box coordinates are processed to obtain the final defect information.

[0029] In one embodiment, the method further includes:

[0030] Save the defect type, the defect location, the confidence data, and the image data in a preset database.

[0031] In one embodiment, the pose data includes the yaw angle and the three-dimensional coordinates relative to the reference point; the angle data includes the pitch angle;

[0032] The step S3 includes:

[0033] Obtain the defect localization according to the pixel offset, the yaw angle, the three-dimensional coordinates, and the pitch angle in combination with a preset boundary function.

[0034] In one embodiment, the method further includes:

[0035] Select a corresponding processing method according to the defect type and the defect location;

[0036] Control the robot to move to the defect location and perform the corresponding processing operation.

[0037] This application also provides a fire pool defect detection system, including a processor and a memory storing a computer program, and the processor implements the steps of the fire pool defect detection method described in any one of the above when executing the computer program.

[0038] Implementing the present invention has the following beneficial effects: The present application provides a method and system for detecting defects in a fire pool. The steps of the method include obtaining image data, angle data, and pose data of the pool wall; performing image processing on the image data, and using a convolutional neural network to identify defects in the image processing result to obtain defect information; the defect information includes the pixel offset of the defect position and the defect type; coordinate calculation is performed based on the pixel offset, angle data, and pose data to obtain defect positioning. By using a robot to obtain image data, angle data, and pose data of the pool wall, and combining a convolutional neural network for defect identification and positioning, the present invention realizes an automated and highly accurate detection process. It effectively solves the problems of potential safety hazards, low efficiency, low accuracy, and high operating costs caused by manual operation in the traditional method for inspecting defects in a fire pool. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for the description of the present invention will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings:

[0040] Figure 1 is a flowchart of a method for detecting defects in a fire pool according to the present application;

[0041] Figure 2 is a schematic diagram of the movement structure of the robot according to the present application;

[0042] Figure 3 is a schematic diagram of the defect identification process according to the present application;

[0043] Figure 4 is a schematic diagram of automatic defect marking according to the present application;

[0044] Figure 5 is a top view of obtaining the three-dimensional coordinates of the robot according to the present application;

[0045] Figure 6 is a front view of obtaining the three-dimensional coordinates of the robot according to the present application;

[0046] Figure 7 is a top view of automatic defect marking of the defect in the field of view of the robot according to the present application;

[0047] Figure 8 is a schematic diagram of the robot system structure according to the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] To make the objectives, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0049] As Figure 1 shown, Figure 1 This is a schematic flow chart of a method for detecting defects in a fire pool of this application.

[0050] This application provides a method for detecting defects in a fire pool, including the following steps:

[0051] Step S1: Obtain image data, angle data, and pose data of the pool wall;

[0052] In this step, it should be noted that by using sensors and cameras installed on the robot, image data, angle data, and pose data of the pool wall are obtained. The image data is used to reflect the surface condition of the pool wall, and the angle data and pose data provide a reference basis for subsequent defect location. Among them, the pose data includes one or more of position data and attitude data.

[0053] Step S2: Perform image processing on the image data, and use a convolutional neural network to identify defects on the image processing result to obtain defect information; the defect information includes the pixel offset of the defect position and the defect type;

[0054] In this step, it should be noted that first, preprocessing is performed on the obtained image data, including operations such as denoising and enhancing contrast to improve the image quality. Then, the processed image is input into a pre-trained convolutional neural network model, which can automatically learn the features in the image and perform defect identification. Through the analysis of the convolutional neural network, defect information is obtained, including the pixel offset of the defect position and the defect type. Among them, the pixel offset is used to determine the specific position of the defect in the image, and the defect type helps to evaluate the severity of the defect.

[0055] Step S3: Perform coordinate calculation based on the pixel offset, angle data, and pose data to obtain defect location.

[0056] In this step, it should be noted that according to the pixel offset obtained in step S2, combined with the angle data and pose data obtained in step S1, through coordinate transformation and calculation, the position of the defect in the image is converted into actual physical coordinates, thereby realizing precise positioning of the defect.

[0057] As Figure 2As shown in the figure, implementing this application can improve operation safety. The robot replaces personnel to enter the water tank, reducing dangerous operations such as diving; the inspection efficiency is greatly improved. The autonomous inspection robot can continuously operate, eliminating the waiting time for manual alternating operations; the accuracy is also improved. Introducing artificial intelligence eliminates the subjectivity of subjective judgment and ensures the accuracy of defect location; at the same time, the labor cost is reduced. The robot replaces most of the manual operations, reducing the labor cost and training time; the operation is simplified. The operator only needs to control the robot and does not need to directly enter the water tank. In summary, this application makes full use of intelligent technology and autonomous inspection robots. As shown in Figure 2 the figure, the schematic diagram of the robot structure, combined with image processing and artificial intelligence recognition, realizes the advantages of high efficiency, accuracy, and safety in the defect inspection of the fire water tank, providing a beneficial solution for the upgrade of traditional methods.

[0058] As Figure 3 shown in the figure, further, the image processing of the image data includes:

[0059] Performing image processing on the image data based on the local contrast enhancement algorithm to obtain an enhanced image;

[0060] Performing image processing on the enhanced image based on the dehazing algorithm to obtain the image processing result.

[0061] It should be noted that when performing defect detection on the fire water tank, first, image acquisition is carried out. A high-definition variable magnification pan-tilt camera is used to photograph the inside of the fire water tank. The camera has high resolution and variable magnification functions, which can adapt to the shooting requirements at different distances and angles to ensure that the collected images are clear and cover comprehensively; then, image preprocessing work is carried out. First, local contrast enhancement processing is performed on the collected image data. The local contrast enhancement algorithm is used to highlight the details and edge information in the image, improve the clarity and contrast of the image, and then dehazing processing is performed on the enhanced image. The dehazing algorithm is used to remove the fog and blur effects in the image to further improve the clarity and quality of the image; finally, defect recognition and positioning are carried out. The preprocessed image is subjected to artificial intelligence recognition using a trained convolutional neural network model. The model can automatically identify the tiny defects on the coating of the water tank wall by learning a large amount of image data, and then the precise location of the identified defects is determined through a deep learning algorithm to determine the specific location and scope of the defects.

[0062] Further, performing image processing on the image based on the local contrast enhancement algorithm to obtain an enhanced image includes:

[0063] Dividing the image into image blocks to obtain multiple small regions;

[0064] Histogram equalization is performed on multiple small regions, and the histogram of each small region is contrast-limited to obtain a small region image after contrast-limiting processing;

[0065] Interpolation processing is performed on the small region image after contrast-limiting processing to obtain an enhanced image.

[0066] Specifically, the input image is divided into multiple 16x16 pixel small regions (tiles) for image block division, and each small region is independently processed to adapt to the illumination and contrast changes in different regions; then, histogram equalization is performed on each small region. By statistically analyzing the gray histogram and calculating the cumulative distribution function (CDF), the original gray values are converted into new gray values to make the gray distribution more uniform; then, contrast limitation is carried out. The excessively high pixel values in the histogram are cropped and the cropped part is evenly distributed throughout the histogram to avoid over-enhancement in local regions and maintain the overall contrast balance of the image; finally, bilinear interpolation is used to smooth the pixels between small regions to eliminate the discontinuity at the region boundaries and ensure the overall consistency of the image.

[0067] Furthermore, performing histogram equalization on multiple small regions includes:

[0068] Performing gray histogram statistics on each small region, calculating the distribution of its gray values to obtain a gray histogram;

[0069] Based on the gray histogram, calculating the cumulative distribution function of each small region;

[0070] Using the cumulative distribution function to convert the original gray values of the pixels in each small region into new gray values to achieve histogram equalization.

[0071] It should be noted that in the image preprocessing stage of fire pool defect detection, the image is processed based on the local contrast enhancement algorithm to obtain the enhanced image. The specific process is as follows: First, the image is divided into blocks, and the input image is segmented into multiple tiles (small regions) of the same size. The size of each region is 16x16 pixels, and histogram equalization is performed independently for each region. Then, the gray histogram and CDF (cumulative distribution function) are calculated for each region, and the original gray values are mapped to new gray values according to the CDF to make the gray distribution of each region more uniform. Then, contrast limitation is performed. CLAHE, that is, contrast-limited adaptive histogram equalization, enhances the image contrast by locally applying histogram equalization in the image. In this process, a contrast limit (ClipLimit) is introduced to avoid over-enhancement in local regions. This limit clips the overly high pixel values in the histogram and evenly distributes the clipped part to the entire histogram to prevent excessive contrast in some regions. After that, interpolation processing is performed on the small-region images that have undergone contrast limitation processing. CLAHE uses bilinear interpolation to smooth the pixels between regions to avoid discontinuities at the boundaries between regions and ensure the overall consistency of the image, thereby obtaining the enhanced image. In terms of implementation steps, first, preprocess the input image, convert the input image to a grayscale image (when it is a color image), and denoise the image (such as Gaussian filtering or median filtering) to reduce the impact of noise on contrast enhancement. Then, set the CLAHE parameters, including the region Size (set to 16x16 pixels) and ClipLimit (controlling the degree of histogram clipping). Then, apply the CLAHE algorithm, perform histogram equalization on each region, clip and redistribute the histogram according to ClipLimit, and then use bilinear interpolation to smooth the pixels between regions to generate the final contrast-enhanced image. Finally, output the enhanced image for use by the subsequent defect detection module. In the fire pool defect detection scenario, due to problems such as uneven illumination and large local contrast differences on the inner wall of the pool, traditional global histogram equalization is likely to cause over-enhancement in some regions and insufficient enhancement in other regions. Therefore, CLAHE (adaptive histogram equalization) is more suitable. It divides the image into multiple small regions, performs independent histogram equalization on each region, enhances the contrast while avoiding over-enhancement.

[0072] Furthermore, the defect information also includes the defect type;

[0073] The defect recognition is performed on the image processing result using a convolutional neural network, and the obtained defect information includes:

[0074] The result of the image processing is input into a pre-trained convolutional neural network model to obtain the bounding box coordinates, defect type, and confidence data of the defect;

[0075] The non-maximum suppression algorithm is used to remove overlapping detection boxes based on the bounding box coordinates and confidence data of defects, and the final defect information is obtained.

[0076] As Figure 4 shown, specifically, in the process of detecting the inner wall defects of the fire pool, first, the high-definition variable magnification pan-tilt camera collects the video stream of the pool wall in real time, and transmits the video stream to the central processor through a high-speed data transmission line. Subsequently, the collected video stream is imported into the image enhancement algorithm for dehazing and enhancing the image contrast to improve the image quality for subsequent recognition. Then, in the robot, the artificial intelligence recognition algorithm is used. First, a calibrated historical defect data set is required. This data set can be collected on-site. The model is trained through a deep learning algorithm to obtain an initial deep learning model. Then, the preprocessed video stream is imported into the deep learning model, and the defect area and its credibility are obtained through calculation. As the number of recognized defects increases, the automatic recognition rate of defects will gradually increase. At the same time, in the underwater environment, the background environment is single and the interference is less. The defect image characteristics of the pool wall include obvious cracks, patches, and bulges. The preprocessed video stream is imported into the image recognition algorithm, and the image recognition algorithm detects the video stream in real time to obtain the area and location of the defects, as well as the credibility of the defects. The preprocessed image is subjected to defect detection. This module performs defect recognition based on CNN (Convolutional Neural Network). The CNN model can effectively identify defects such as cracks, patches, and bulges in the image through structures such as convolutional layers, pooling layers, and fully connected layers, and outputs the bounding box coordinates and types of the defects. To ensure the accuracy of the detection results, the system uses the non-maximum suppression (NMS) algorithm to remove overlapping detection boxes and avoid the same defect being detected multiple times.

[0077] In another embodiment, the deep learning algorithm is: a deep learning model based on the attention mechanism and Focal Loss.

[0078] Specific implementation steps:

[0079] Network structure:

[0080] Backbone network: Improved ResNet-50, embedded with a channel attention module (SE Block):

[0081] SE(x) = x·σ(W2·ReLU(W1·GAP(x)))

[0082] Among them, SE represents the output of the channel attention module (SEBlock), GAP is global average pooling, W1 and W2 are the weights of the fully connected layers, x represents the input feature map, σ: represents the activation function, ReLU: represents the activation function, and GAP(x): represents the global average pooling operation.

[0083] Multi-scale feature fusion: Aggregate features at different levels through FPN (Feature Pyramid Network).

[0084] Loss function: Use Focal Loss to address class imbalance: FL(p t ) = -α(1 - p t ) γ log(p t )

[0085] where p t is the probability predicted by the model, α is the weight for balancing positive and negative samples, and γ is the parameter for adjusting the weights of easy and hard samples.

[0086] Using the above deep learning model based on the attention mechanism and Focal Loss, through the multi-scale attention network (MSANet), combined with Focal Loss for optimized training, it addresses the insufficient feature extraction ability of traditional CNN in turbid water bodies, where small defects are easily submerged.

[0087] On the turbid water body test set, the average precision is increased to 92.3%, which is 18% higher than the baseline model.

[0088] Furthermore, the defect information includes defect results and confidence data; the non-maximum suppression algorithm is used to remove overlapping detection boxes based on the bounding box coordinates and confidence data of the defects, and the final defect information obtained includes:

[0089] Step 1: Sort the bounding box coordinates of the defects according to the confidence data;

[0090] Step 2: Select the bounding box coordinate with the highest confidence data as the reference box and add it to the final defect information;

[0091] Step 3: Calculate the intersection over union (IoU) of the remaining bounding box coordinates and the reference box;

[0092] Step 4: If the calculated IoU is greater than the preset threshold, remove the corresponding bounding box coordinate from the final defect information;

[0093] Repeat the above Step 2 to Step 4 until all bounding box coordinates are processed to obtain the final defect information.

[0094] In one embodiment, during the process of defect detection using the non-maximum suppression (NMS) algorithm, the specific steps are as follows: First, sort the bounding box coordinates of the defects according to the confidence data, which reflects the confidence level of the model in each defect detection result and usually ranges from 0 to 1. The higher the value, the greater the confidence of the model in the detection result. Then, select the bounding box coordinate with the highest confidence data as the reference box and add it to the final defect information. Next, calculate the intersection over union (IoU) between the remaining bounding box coordinates and the reference box, which is used to measure the overlap degree of two bounding boxes. If the calculated IoU is greater than the preset threshold, it indicates that the corresponding bounding box has a large degree of overlap with the reference box. To avoid detecting the same defect multiple times, remove the corresponding bounding box coordinate from the final defect information. Repeat the above steps until all bounding box coordinates are processed, and finally obtain accurate defect information. Specifically, during the defect detection process, the CNN model will output a confidence score for each detected defect. To further improve the reliability of the detection, the system will verify the same defect in consecutive multiple frames of images. If the same position is detected in multiple frames of images and the confidence scores are all higher than the preset threshold (such as 0.7), then the credibility of the defect is considered to be high.

[0095] In one embodiment, an improved non-maximum suppression (NMS) algorithm is adopted. By introducing soft non-maximum suppression (Soft-NMS), potential defect boxes are retained by dynamically adjusting the confidence, which solves the problem that traditional NMS directly removes overlapping detection boxes and easily causes small defects to be misdeleted.

[0096] Specific implementation steps:

[0097] Input parameters: The set of detection boxes B = {b1, b2,..., bn}, and each box contains the boundary coordinates (x min , y min , x max , y max ) and the confidence s i . The overlap threshold θ = 0.5, and the decay factor σ = 0.1.

[0098] Algorithm process:

[0099] 1. Sort the detection boxes in descending order of confidence;

[0100] 2. Select the box b max with the highest score and retain it in the result set;

[0101] 3. Calculate the intersection over union (IoU) between the remaining boxes and b max ;

[0102]

[0103] Among them, IoU represents the intersection over union, and Area represents the area, which is used to calculate the areas of the intersection and union of two bounding boxes.

[0104] ∩ represents the intersection, that is, the overlapping area of two bounding boxes.

[0105] ∪ represents the union, that is, the total area after merging two bounding boxes.

[0106] 4. For the bounding boxes with IoU > θ, decay their confidence scores according to the Gaussian function;

[0107]

[0108] Among them, si is the confidence score; exp represents the exponential function, and σ represents the variance or the speed of controlling the decay, which is used to adjust the influence degree of IoU on the weight.

[0109] 5. Repeat steps 2 - 4 until all bounding boxes are processed.

[0110] By decaying the confidence scores instead of directly deleting, small - sized defective bounding boxes are retained, and the recall rate is increased by more than 20%.

[0111] Use a confidence - calculation method that fuses multiple factors: Combine the bounding - box regression error and the multi - frame verification mechanism to construct a composite confidence score. Solve the problem that the existing confidence only depends on the class probability and ignores the spatial localization error.

[0112] Specific implementation steps: Single - frame confidence calculation: Define the bounding - box regression error e:

[0113] e = 1 - IoU(b pred , b gt )

[0114] Among them, b pred is the predicted bounding box, b gt is the predicted bounding box, and IoU is the intersection over union.

[0115] Composite confidence formula: s = p c ·exp(-λe)(λ = 0.5). Among them, p is the class probability, and λ controls the error weight.

[0116] Multi - frame verification mechanism: For defects at the same position in N = 5 consecutive frames, calculate the average confidence score: Among them, s final is the final sum or result.

[0117] Among them, only when s final ≥τs final ≥τ(τ = 0.7), it is determined as a valid defect. The false - detection rate is reduced by 35%, and the stability is significantly improved especially in low - contrast scenarios.

[0118] Further, the method further includes:

[0119] Storing the defect type, defect location, confidence data, and image data in a preset database.

[0120] Specifically, in the data storage and report generation link of the fire pool defect detection system, data storage is first carried out. The detected defect information, including key data such as the location, type, and confidence score of the defect, is stored in a preset database, which facilitates subsequent query and analysis work. Then, the report generation work is carried out. According to the defect information stored in the database, a detailed defect report is automatically generated. The report content includes defect type statistics, defect location distribution map, defect credibility score, and corresponding treatment suggestions. Through this report, users can comprehensively understand the defect situation of the inner wall of the fire pool and provide strong data support for subsequent maintenance and repair work.

[0121] Such as Figure 5 、 Figure 6 and Figure 7 As shown, further, the pose data includes the yaw angle and the three-dimensional coordinates relative to the reference point; the angle data includes the pitch angle;

[0122] Step S3 includes:

[0123] Obtaining the defect location according to the pixel offset, yaw angle, three-dimensional coordinates, and pitch angle in combination with a preset boundary function.

[0124] Specifically, the current yaw angle (Yaw) of the robot is read out through the inertial navigation module. The current pitch angle (θ) of the high-definition camera is read out through the pan-tilt head. Through the current image frame of the high-definition camera, the pixel offset (Δx, Δy) between the defect center and the camera view center is calculated. Defect location: A pixel coordinate system of the current image is formed with the camera field center as the coordinate origin. Using a convolutional neural network (CNN) as the deep learning model, features are learned through layers such as convolution and pooling to achieve automatic recognition of defects such as cracks, patches, and bulges. After obtaining the defect contour, the pixel coordinates (Δx, Δy) of the defect center are calculated, that is, the pixel offset of the defect center relative to the camera view center.

[0125] Combining the three-dimensional coordinates (x, y, z) of the robot relative to the reference point, the yaw angle (yaw) of the robot, the pitch angle (θ) of the camera, the pixel offset (Δx, Δy) of the defect, and the boundary function of the pool wall, the location where the defect is located is quickly located through the algorithm.

[0126] Specifically, the three-dimensional coordinates (x, y, z) of the robot relative to the reference point, the yaw angle of the robot, and the pitch angle (θ) of the high-definition camera. Combining these data with the boundary function of the pool wall enables precise positioning of the defect location. Assume the pool is a cuboid with length L, width D, and depth H. A three-dimensional coordinate system is established with its center as the coordinate origin, the x-axis along the length direction of the pool, the y-axis along the width direction, and the z-axis along the depth direction. To verify whether the defect location is within the pool range, the boundary functions of the five surfaces of the pool except the top surface are as follows:

[0127] Bottom surface: z = -H, -L / 2 ≤ x ≤ L / 2, -D / 2 ≤ y ≤ D / 2;

[0128] Side surface 1: x = L / 2, -D / 2 ≤ y ≤ D / 2, -H ≤ z ≤ 0;

[0129] Side surface 2: x = -L / 2, -D / 2 ≤ y ≤ D / 2, -H ≤ z ≤ 0;

[0130] Side surface 3: y = D / 2, -L / 2 ≤ x ≤ L / 2, -H ≤ z ≤ 0;

[0131] Side surface 4: y = -D / 2, -L / 2 ≤ x ≤ L / 2, -H ≤ z ≤ 0;

[0132] Through algorithm analysis, the pixel offset of the defect can be converted into the actual three-dimensional position, ensuring the accuracy of the defect detection on the inner wall of the fire pool.

[0133] Furthermore, the method also includes:

[0134] Selecting the corresponding treatment method according to the defect type and defect location;

[0135] Controlling the robot to move to the defect location and performing the corresponding treatment operation.

[0136] The defect treatment is divided into two methods: mechanical hand clamping and cutting and sucking. For the defect of large-area peeling, it can be clamped and torn off by the mechanical hand. For the defect of small-area non-peeling, it can be removed by the cutting and sucking device.

[0137] 1) Cutting and suction pump system: The system consists of an outer cylinder, a cutting blade, a motor, a filter screen, and a suction pump. By controlling the rotation of the motor, the cutting blade is driven to rotate, and the protruding coating is cut off by the rotating blade. The magnetic flakes are sucked through the suction pump. When foreign objects are blocked by the filter screen and gather at the bottom end cover, the water flows through the filter and is discharged through the suction pump into the pool, thus playing the role of collecting foreign objects. The robot can also dive to the bottom of the pool to suck foreign objects at the bottom of the pool. All the sucked foreign objects are collected inside the outer cylinder, and the internal foreign objects can be removed by opening the end cover after the water comes out.

[0138] 2) Gripping manipulator: The manipulator is a 2-DOF manipulator, mainly composed of pitch rotation, wrist rotation, and gripping fingers. The manipulator can change its posture with pitch and rotation movements to grip foreign objects in various postures. It can reliably grip the raised coating, rotate and pull it, and then twist it off the wall. At the same time, the manipulator can grip foreign objects that fall to the bottom of the pool.

[0139] Gripping robot: When a large peeling defect is detected, the gripping robot adjusts its posture according to the defect location data, grips and pulls the coating to separate it from the pool wall.

[0140] Cutting suction device: For defects that have not fallen off on the small surface, the cutting suction device is activated. The motor drives the cutting blade to rotate, cuts the coating and uses a suction pump to absorb the coating particles.

[0141] In a specific embodiment, the robot starts and dives into a pool, and accurately measures its own depth and relative position through a sonar sensor.

[0142] The inertial navigation module records the robot's yaw angle to ensure that the robot is facing the correct direction.

[0143] The high-definition camera begins scanning around the pool wall layer by layer, capturing image data at each layer.

[0144] The image processing module improves image quality based on image enhancement algorithms to provide more accurate input for deep learning models.

[0145] The deep learning model automatically identifies defect type and location using pre-training data.

[0146] The gripping robot performs gripping or processing operations based on the defect location information to achieve defect repair.

[0147] Through these embodiments, the present invention achieves optimization of multiple aspects of the intelligent pool defect inspection method, including defect identification, positioning, report generation and defect processing, thereby improving operational efficiency, accuracy and safety.

[0148] When implementing the fire pool defect detection method of this application, in terms of the defect recognition model, in addition to using convolutional neural networks, other deep learning architectures such as recurrent neural networks (RNNs) or Transformers can also be considered to adapt to different types of image data. In the defect localization process, more sensor data such as infrared sensors or laser sensors can be combined to improve the accuracy and stability of defect localization. For the gripper manipulator, in addition to pitch rotation and wrist rotation, more degrees of freedom can be added to achieve more flexible gripping and handling actions. In the autonomous inspection process, machine vision technologies such as object tracking and 3D reconstruction can be combined to further optimize the image processing and defect recognition processes. In the report generation scheme, in addition to statistical analysis, machine learning algorithms can also be introduced to automatically optimize the report content and format. In the defect handling scheme, other types of tools such as laser cutting or spraying devices can be considered to meet different types of defect handling requirements.

[0149] As Figure 8 shown, this application also provides a fire pool defect detection system, including a processor and a memory storing a computer program. When the processor executes the computer program, the steps of any of the above fire pool defect detection methods are implemented.

[0150] Mechanical structure and components: The robot main body shell is made of high-strength waterproof material, with a built-in high-definition variable magnification pan-tilt camera, surround sonar sensors, and an inertial navigation module. The gripper manipulator consists of pitch rotation and wrist rotation, and the end of the manipulator is equipped with gripper fingers.

[0151] Circuit connection and processing: The internal circuit of the robot includes an image processing module, a deep learning model operation unit, and a sonar signal processing unit. The high-definition camera and sonar sensors are connected to the central processor through high-speed data transmission lines. The image processing module uses a high-performance image processor for image enhancement and dehazing processing.

[0152] Workflow:

[0153] The robot enters the pool and obtains its relative position and water depth information through sonar sensors.

[0154] The inertial navigation module records the current course deviation angle to ensure that the robot is facing the correct direction.

[0155] The high-definition camera captures images of the pool wall, and image data is captured for each layer. It is transmitted to the image processing module in real time.

[0156] The image processing module uses local transmission and dehazing algorithms to improve the image quality. The image processing module improves the image quality according to the image enhancement algorithm, providing more accurate input for the deep learning model.

[0157] The deep learning model uses pre-trained data to identify the defect type and outputs the coordinates of the defect location.

[0158] The clamping manipulator performs clamping or processing operations according to the defect location information to achieve defect repair. Through these embodiments, the present invention realizes the optimization of multiple aspects of the intelligent pool defect inspection method, including defect identification, positioning, report generation, and defect processing, etc., improving the operation efficiency, accuracy, and safety.

[0159] It can be understood that the above embodiments only express the preferred implementation modes of the present invention, and their descriptions are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent of the present invention; it should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, the above technical features can be freely combined, and several deformations and improvements can also be made, and these all belong to the protection scope of the present invention; therefore, all equivalent transformations and modifications made to the scope of the claims of the present invention shall fall within the scope covered by the claims of the present invention.

Claims

1. A fire pool defect detection method, characterized in that: The following steps are involved: Step S1: Acquire image data, angle data of the pool wall and position data of the robot; Step S2: performing image processing on the image data, and using a convolutional neural network to perform defect recognition on the image processing result to obtain defect information; The defect information includes a pixel offset of a defect position and a defect type; Step S3: performing coordinate calculation according to the pixel offset, the angle data and the posture data to obtain the defect location.

2. The fire water pool defect detection method according to claim 1, characterized in that: The performing image processing on the image data comprises: Performing image processing on the image data based on a local contrast enhancement algorithm to obtain an enhanced image; The enhanced image is processed based on a defogging algorithm to obtain the image processing result.

3. The fire pool defect detection method according to claim 2, characterized in that: The image is processed based on the local contrast enhancement algorithm to obtain an enhanced image, which includes: Dividing the image into blocks to obtain a plurality of small areas; Performing histogram equalization on the multiple small areas, and performing contrast limitation on the histogram of each small area, to obtain a small area image subjected to contrast limitation processing; The small area image that has been subjected to contrast limitation processing is interpolated to obtain the enhanced image.

4. The fire water pool defect detection method according to claim 3, characterized in that: The performing histogram equalization on the multiple small areas comprises: Perform grayscale histogram statistics on each small area, calculate the distribution of its grayscale value, and obtain the grayscale histogram; Based on the grayscale histogram, calculating the cumulative distribution function of each small area; The original grayscale values ​​of the pixels in each small area are converted into new grayscale values ​​by using the cumulative distribution function to achieve histogram equalization.

5. The fire water pool defect detection method according to claim 4, characterized in that: The defect information also includes defect type; The image processing result is processed using a convolutional neural network to perform defect recognition, and the defect information obtained includes: Inputting the image processed result into a pre-trained convolutional neural network model to obtain the bounding box coordinates of the defect, the defect type and confidence data; A non-maximum suppression algorithm is used to remove overlapping detection frames based on the boundary box coordinates of the defect and the confidence data to obtain final defect information.

6. The fire water pool defect detection method according to claim 5, characterized in that: The defect information includes defect results and confidence data; The non-maximum suppression algorithm is used to remove overlapping detection frames based on the boundary box coordinates of the defect and the confidence data to obtain the final defect information, including: Step 1: sorting the bounding box coordinates of the defects according to the confidence data; Step 2: Select the bounding box coordinates with the highest confidence data as the reference box and add it to the final defect information; Step 3: Calculate the intersection ratio of the remaining bounding box coordinates and the reference box; Step 4: If the calculated intersection-over-union ratio is greater than the preset threshold, the corresponding bounding box coordinates are removed from the final defect information; Repeat the above steps 2 to 4 until all bounding box coordinates are processed to obtain the final defect information.

7. The fire water pool defect detection method according to claim 6, characterized in that: The method further comprises: The defect type, the defect location, the confidence data, and the image data are stored in a preset database.

8. The fire water pool defect detection method according to claim 7, characterized in that: The posture data includes the yaw angle and the three-dimensional coordinates relative to the reference point; the angle data includes the pitch angle; The step S3 comprises: The defect location is obtained according to the pixel offset, the yaw angle, the three-dimensional coordinates and the pitch angle in combination with a preset boundary function.

9. The fire water pool defect detection method according to claim 8, characterized in that: The method further comprises: Select a corresponding processing method according to the defect type and the defect location; The robot is controlled to move to the defect position and perform corresponding processing operations.

10. A fire pool defect detection system, comprising a processor and a memory storing a computer program, characterized in that: The processor implements the steps of the fire water pool defect detection method according to any one of claims 1 to 9 when executing the computer program.

Images