A method and system for detecting leaks in heating pipelines

Abstract

This invention discloses a method and system for detecting leaks in heating pipelines, relating to the field of heating pipeline detection. The method includes: acquiring infrared images of the heating pipeline using a drone; identifying the heating pipeline region in the infrared images using a heating pipeline identification model; the heating pipeline identification model is an improved R-value based on a training set. 3 Obtained by training a Det network; improved R 3 In the Det network, the convolution operation in the input layer of the feature pyramid network adopts variable convolution operation, which improves R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the R-activation function is improved. 3 Candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm; based on the heating pipeline area in the infrared image, an adaptive threshold algorithm is used to determine the location of leak points in the heating pipeline area. This invention improves the accuracy of leak detection.

Description

Technical Field

[0001] This invention relates to the field of pipeline inspection technology, and in particular to a method and system for detecting leaks in heating pipelines. Background Technology

[0002] There are three main laying methods for heating pipelines: overhead laying, trench laying, and direct burial. This invention mainly focuses on overhead laying heating pipelines, which can be divided into low-support, medium-support, and high-support systems. Low-support systems are 0.3 meters to 1.0 meters high, along walls or in areas that do not obstruct traffic; medium-support systems are 2.0 meters to 4.5 meters high, used to cross sidewalks and areas with high traffic; high-support systems are above 4.5 meters high, used to cross highways and major traffic arteries. Long-distance heating pipelines are prone to damage and leakage, resulting in abnormally high temperature distribution at and around the leak location. Pipeline leaks not only cause huge economic losses to enterprises but may also pose a threat to people in the surrounding area. Currently, relevant departments still use manual methods to detect leaks in heating pipelines, which requires climbing pipelines, posing significant risks and being inefficient. Therefore, leak detection of heating pipelines is crucial for their safe operation.

[0003] Existing methods for detecting leaks in heating pipelines mainly include water imbalance detection, pressure anomaly detection, temperature anomaly detection, and sound anomaly detection. Heating pipelines are closed-loop systems under pressure. Under normal operation, the water volume and pressure within the pipeline should maintain a dynamic balance. If a leak occurs, it will manifest as a significant reduction in water volume and a rapid decrease in pressure. Therefore, leaks can be detected by monitoring the water and pressure balance over a section of the pipeline. However, for detecting leaks in long-distance heating pipelines, there is a problem with accurately locating the leak point; only a section of the pipeline can be identified as leaking. Using pipeline robots for leak detection requires placing the robot inside the pipeline, which necessitates damaging the pipe and is extremely inconvenient.

[0004] The temperature of heating pipes is manually measured using infrared thermometers. This method is labor-intensive, involves sampling, and cannot comprehensively cover all areas of the pipes, potentially leading to missed detections. Furthermore, measurements taken from the ground on elevated sections of the pipes can introduce errors. The latest method uses drones equipped with infrared thermal imaging cameras to photograph the pipes, employing deep learning algorithms for leak detection within the images. However, this method suffers from low accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting leaks in heating pipelines, thereby improving the accuracy of leak detection.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for detecting leaks in heating pipelines, comprising:

[0008] Infrared images of heating pipelines were collected using drones;

[0009] A heating pipe recognition model is used to identify the infrared image, thereby obtaining the heating pipe region in the infrared image; the heating pipe recognition model is an improved R model based on a training set. 3 The improved R is obtained by training the Det network. 3 The convolution operations in the input layer of the feature pyramid network of the Det network employ variable convolution operations, and the improved R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the improved R... 3 The candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm;

[0010] Based on the heating pipeline area in the infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipeline area.

[0011] Optionally, it also includes:

[0012] Based on the UAV positioning information carried in the infrared image, the positioning information of the heating pipeline in the infrared image is determined.

[0013] Optionally, based on the heating pipe area in the infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipe area, specifically including:

[0014] The actual ambient temperature is collected from the infrared image, and the temperature information in the infrared image is corrected using the actual ambient temperature.

[0015] Based on the heating pipeline area in the corrected infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipeline area.

[0016] Optionally, based on the heating pipe area in the corrected infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipe area, specifically including:

[0017] According to the formula Determine the leakage temperature threshold;

[0018] Where V represents the leakage temperature threshold, N represents the number of pixels in the heating pipe area, T1 represents the temperature value of the first pixel, T2 represents the temperature value of the second pixel, and T... nThis represents the temperature value of the nth pixel, α is the number of the highest temperature values ​​in the heating pipe area, and K is the average temperature offset.

[0019] Pixels with temperatures exceeding the leakage temperature threshold are identified as leakage points.

[0020] Optionally, the variable convolution operation is represented as:

[0021]

[0022] Where, x s Let y(p0) represent the input feature map, y(p0) represent the convolution output at position p0, R represent the sampling region, and p0 represent the center point of the sampling region. n w(p) represents the nth point in the sampling region. n ) represents the convolution kernel weight coefficient corresponding to the nth point in the sampling region, Δp n Indicates the offset, Δm k This represents the weighting coefficient.

[0023] Optionally, the infrared image is an infrared image in R-JPEG format.

[0024] Optionally, the H-Swish activation function is expressed as:

[0025]

[0026] Where x represents the input to the H-Swish activation function, and ReLU6() represents the ReLU6 activation function.

[0027] This invention also discloses a heating pipeline leak detection system, comprising:

[0028] Infrared image acquisition module, used to acquire infrared images of heating pipelines via drone;

[0029] A heating pipe identification model is used to identify the heating pipe region in the infrared image; the heating pipe identification model is an improved R model based on a training set. 3 The improved R is obtained by training the Det network. 3 The convolution operations in the input layer of the feature pyramid network of the Det network employ variable convolution operations, and the improved R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the improved R... 3 The candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm;

[0030] The leak location determination module is used to determine the location of the leak point in the heating pipeline area based on the infrared image and using an adaptive threshold algorithm.

[0031] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0032] This invention employs a drone equipped with an infrared thermal imaging camera to capture infrared images at intervals along the heating pipeline, and uses the acquired infrared images to detect pipeline leak locations, thus improving detection efficiency; it also utilizes an improved R... 3 The Det rotating box target detection algorithm can accurately identify and select rotating boxes for heating pipelines with large aspect ratios, enabling more precise identification and selection of pipeline areas. By using an adaptive threshold method to detect high-temperature anomalies, accurate leak points can be obtained. The detection method provided by this invention can handle leak detection for most overhead heating pipelines and has good generalization ability. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This invention provides a schematic flowchart of a method for detecting leaks in heating pipelines.

[0035] Figure 2 R is provided for the embodiments of the present invention. 3 Schematic diagram of Det network structure;

[0036] Figure 3 R is provided for the embodiments of the present invention. 3 A schematic diagram of the feature optimization part in the Det network structure;

[0037] Figure 4 A schematic diagram of the feature refinement module structure is provided for embodiments of the present invention;

[0038] Figure 5 A schematic diagram of variable convolution operation is provided for embodiments of the present invention;

[0039] Figure 6 A flowchart illustrating the process of determining candidate anchor boxes using the K-means++ clustering algorithm is provided for embodiments of the present invention.

[0040] Figure 7 A schematic diagram of the camera field of view of an unmanned aerial vehicle (UAV) is provided for embodiments of the present invention;

[0041] Figure 8 An improved R is provided for the embodiments of the present invention. 3 A schematic diagram of the detection results from the Det network;

[0042] Figure 9 This invention provides a schematic diagram of a heating pipeline leak detection system. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] The purpose of this invention is to provide a method and system for detecting leaks in heating pipelines, thereby improving the accuracy of leak detection.

[0045] To make the above-mentioned objects, 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.

[0046] Example 1

[0047] like Figure 1 As shown in the figure, this embodiment provides a method for detecting leaks in heating pipelines, including the following steps.

[0048] Step 101: Collect infrared images of the heating pipeline using a drone.

[0049] Specifically, step 101 includes: the drone is equipped with an infrared thermal imaging camera, which collects infrared images of the heating pipes.

[0050] After flying to a safe altitude, the drone cruises along the overhead heating pipeline to capture infrared images of the pipeline in R-JPEG format, which contain temperature information.

[0051] R-JPEG format infrared images contain actual temperature values, offering superior accuracy compared to the previous method of distinguishing temperature conditions by color in infrared images.

[0052] Step 102: Use a heating pipe recognition model to identify the infrared image and obtain the heating pipe area in the infrared image; the heating pipe recognition model is an improved R model based on the training set. 3 The improved R is obtained by training the Det network. 3The convolution operations in the input layer of the feature pyramid network of the Det network employ variable convolution operations, and the improved R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the improved R... 3 The candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm.

[0053] The improved R of this invention 3 The Det network incorporates variable convolution operations, H-Swish activation function, and K-Means++ target box clustering algorithm. The improved R... 3 The Det rotating target detection algorithm can obtain more accurate and efficient rotating target selection.

[0054] R 3 The Det network is an improvement on RetinaNet. The overall RetinaNet network structure consists of a backbone network (ResNet), a neck network (Feature Pyramid Network, FPN), and a head network using a fully convolutional network (FCN). The backbone network (ResNet) is responsible for computing convolutional feature maps across the entire input image and is an independent convolutional network. The neck network connects the spine and head, and it improves or reconfigures the original feature maps generated by the backbone network. The FPN enhances and utilizes the multi-scale features generated by ResNet, resulting in feature maps that are more expressive and contain multi-scale target region information.

[0055] R 3 The Det network can rotate the bounding box on top of the horizontal bounding box, enabling target detection with rotated bounding boxes. (Original R...) 3 Det network structure as follows Figure 2 As shown, Figure 2 The structure of the feature optimization part within the dashed box is as follows: Figure 3 As shown. A single-stage rotation detector based on the RetinaNet network, R 3 The Det network is primarily designed for refined rotating target detection of objects with large aspect ratios and arbitrary rotation. It uses a high-precision, arbitrarily rotating single-stage detector to achieve high-accuracy detection. The overall network structure is as follows: First, horizontal anchor points are used to detect the preliminary positions of all heating pipes in the image, ensuring high recall. Then, in the refinement stage, feature optimization is performed, and the refined rotating anchors are used to achieve precise rotating bounding boxes for the target object.

[0056] To address the feature misalignment issue arising from refined single-stage detectors, R3 The Det network employs a Feature Refinement Module (FRM), which uses feature interpolation to obtain the positional information corresponding to the refined anchor points, while simultaneously reconstructing the entire feature map pixel by pixel. The specific process of the Feature Refinement Module is as follows: Figure 4 As shown, Figure 4 for Figure 3 The specific structure diagram shows that by adding feature maps through bidirectional convolution, a new large convolutional kernel can be obtained, consisting of a 5×1 and a 1×5 dual convolutional kernel, and a 1×1 single convolutional kernel. Addressing the feature mismatch issue in current refined single-level detectors, a feature refinement module is employed. This module improves detection precision by acquiring more accurate features. The refinement stage retains only the bounding box with the highest score for each feature point, and each feature point corresponds to only one refined bounding box. For each feature point in the feature map, a corresponding feature vector is obtained on the feature map based on the five coordinates of the refined center point and the four vertices of the bounding box. A more accurate feature vector is then obtained through bilinear interpolation. The five feature vectors are added together and used to replace the current feature map. This process is repeated for all feature points, and the entire feature map is reconstructed. Finally, the reconstructed feature map is added to the original feature map and output. The feature refinement module significantly reduces the number of refined bounding boxes after horizontal anchor point localization, thereby accelerating the model's computation speed.

[0057] The improved R of this invention 3 The Det network introduces variable convolution operations, such as Figure 5 As shown, the original convolution kernels are arranged in a regular pattern. By adding an offset to each convolution kernel, they are offset and distributed according to the shape of the target being detected, and finally a new feature map is output. The standard two-dimensional convolution operation is to slide sampling on a regular feature map through a regular grid R, and the sampled values ​​are summed according to the weights. The grid R defines the size of the sampling region and the weight w. The output y(p0) for each position p0 in the input feature map x is given by formula (1); while in deformable convolution, an offset is added to each sampling point p, and the calculation process is given by formula (2); however, after adding the offset in deformable convolution, the receptive field range is larger than the target range, so a weight coefficient is added to each sampling point to ensure effective feature extraction, as given by formula (3).

[0058]

[0059]

[0060]

[0061] Where, x s Let y(p0) represent the input feature map, y(p0) represent the convolution output at position p0, R represent the sampling region, and p0 represent the center point of the sampling region. nw(p) represents the nth point in the sampling region. n ) represents the convolution kernel weight coefficient corresponding to the nth point in the sampling region, Δp n Indicates the offset, Δm k This represents the weighting coefficient.

[0062] The variable convolution operation of this invention adopts formula (3).

[0063] The original ResNet (ResNet101) network uses the ReLU activation function, as shown in formula (4). The improved version uses the H-swish activation function, as shown in formula (5). It is a variant of the Swish activation function, replacing the Sigmoid function with the less computationally expensive ReLU and selecting the H-swish function as the main activation function. This approach increases model accuracy while reducing the number of times the network accesses memory, thus improving network training speed.

[0064] f(x) = x·sigmoid(βx) (4);

[0065] Where x represents the input to the ReLU activation function, β is the parameter, and sigmoid() represents the Sigmoid function.

[0066] The H-Swish activation function is expressed as follows:

[0067]

[0068] Where x represents the input to the H-Swish activation function, and ReLU6() represents the ReLU6 activation function.

[0069] To more efficiently find anchor frames that match the pipeline target, this invention employs the K-Means++ algorithm for clustering candidate anchor frames. Since each anchor frame has vectors in both x and y directions, the K-Means++ algorithm is used to calculate the magnitudes of these two vectors, R. 3 The Det network is optimized using pre-generated anchor boxes, as follows: Figure 6 As shown, the anchor boxes saved to the configuration file are used as improved R. 3 Candidate anchor boxes from the Det network. The K-Means++ target box clustering algorithm is used primarily to select ideal anchor box sizes through clustering, improving recognition accuracy and accelerating training. This module uses distance as a metric for similarity between data objects; smaller distances indicate they are likely to belong to the same cluster. The center of each cluster is calculated using the mean of the values ​​contained within the cluster, and the method for selecting the center points is optimized.

[0070] The specific process is as follows: First, a sample is randomly selected as the initial cluster center O1. Then, the shortest distance from each sample to the current cluster center is calculated. The probability of each sample becoming a cluster center is calculated sequentially, and cluster centers are selected according to these probabilities. This process is repeated to calculate the probability of the remaining samples becoming the next cluster center, until all cluster centers are selected. Next, samples outside the cluster centers are assigned to the cluster corresponding to the shortest-distance cluster center. Finally, for each cluster, the coordinates of its cluster center are repeatedly calculated until the cluster center position no longer changes, ultimately selecting K initial cluster centers. The probability calculation formula is as follows:

[0071]

[0072] Where p is the probability that the i-th sample becomes the cluster center, D(x) i ) represents the shortest distance from the i-th sample to the current cluster center.

[0073] The K-means++ clustering algorithm can discover significant patterns and clusters in samples, grouping samples with similar or identical features into the same category based on similarity. This method greatly reduces the impact of the K value on the clustering effect, effectively solving the problems caused by the shortcomings of the K-means clustering algorithm.

[0074] The heating pipe recognition model of this invention achieves the following effect in recognizing heating pipes from infrared images: Figure 8 As shown, Figure 8 Images (a), (b), (c), and (d) are renderings of the identified heating pipeline. The identification results include the identified heating pipeline outline and the confidence level.

[0075] Before step 102, the infrared image is preprocessed, specifically to enhance the contrast of the infrared image.

[0076] Improved R based on training set 3 Training the Det network specifically includes:

[0077] Infrared images containing heating pipes were acquired, and the heating pipes in the infrared images were rotated and labeled to form a training set.

[0078] Improved R based on training set 3 A Det network is used to obtain a heating pipe recognition model. This model is used for accurate rotational recognition and bounding selection of heating pipe areas.

[0079] Step 103: Based on the heating pipeline area in the infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipeline area.

[0080] Since the pipeline area is relatively large, the leak location usually only occupies a small portion of it. Therefore, the highest temperature value in the pipeline area selected by the identification box is removed, the average temperature of the remaining pipeline area is calculated, and then an offset is added based on the temperature of the leak location. This offset is used as the threshold for determining whether a leak has occurred.

[0081] Step 103 specifically includes:

[0082] The actual ambient temperature is used to acquire infrared images, and the temperature information in the infrared images is corrected using this actual ambient temperature. This ensures that the temperature information in the acquired infrared images matches the actual temperature.

[0083] Based on the heating pipe area in the corrected infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipe area. This step specifically includes:

[0084] According to the formula Determine the leakage temperature threshold.

[0085] Where V represents the leakage temperature threshold, N represents the number of pixels in the heating pipe area, T1 represents the temperature value of the first pixel, T2 represents the temperature value of the second pixel, and T... n This represents the temperature value of the nth pixel, α is the number of the highest temperature values ​​in the heating pipe area, and K is the average temperature offset.

[0086] Pixels with temperatures exceeding the leakage temperature threshold are identified as leakage points.

[0087] This embodiment of a method for detecting leaks in heating pipelines further includes: determining the location information of the heating pipeline in the infrared image based on the drone location information carried by the infrared image.

[0088] This embodiment uses the drone as the center (based on known GPS positioning information) and calculates the specific location of the leak point based on the linear ratio between the camera's field of view, the infrared image, and the actual captured image field distance. Figure 7 As shown, Figure 7 In the image, (a) represents the camera's field of view (FOV), and (b) represents the horizontal field of view (FOV) of the camera's field of view. X (c) represents the vertical field of view (FOV) of the camera. Y ).

[0089] The distances in the x and y directions are estimated using equation (6) (where the altitude h needs to be measured by the UAV's distance sensor):

[0090] Horizontal distance vertical distance

[0091] The relationship between distance and the ratio of pixels in an infrared image is linear, which can be expressed as:

[0092]

[0093] P x This indicates the number of pixels in the x-direction of the infrared image.

[0094]

[0095] P y The scale represents the number of pixels in the y-direction of an infrared image. x The scale represents the ratio between the distance in the x-direction and the number of pixels in the infrared image. y This represents the ratio between the distance in the y-direction and the pixels in the infrared image.

[0096] Assuming the leak is located at pixel (x, y) in the image, the offset of the leak location from the image center is:

[0097]

[0098] offset target Offset coordinates representing the actual distance;

[0099] For the transformation from world coordinates to a camera frame with an angle ψ, the rotation matrix is ​​defined as follows:

[0100]

[0101] In the formula, ψ is the yaw angle. Therefore, the position offset in the world frame can be expressed as:

[0102]

[0103] Therefore, the target's GPS coordinates can be used as follows:

[0104]

[0105] Among them, f x and f y P represents the distance expressed in longitude and latitude, respectively. E P represents the positional offset between the camera frame and the world frame along the east longitude direction. N Indicates the positional offset between the camera frame and the world frame along the north latitude direction, GPS cam This refers to the location information stored in the infrared images captured by the drone, GPS. tar get This indicates the location information of the leaked target.

[0106] This invention flies the drone to a safe altitude and flies in a straight line along the direction of the pipeline, using an improved R-type... 3The Det rotating box target detection algorithm first accurately selects the area of ​​the heating pipeline, and then detects high-temperature leak points within that area, greatly improving the accuracy of the detection results.

[0107] This invention controls a drone to fly along a heating pipeline, with the drone equipped with an infrared thermal imaging camera to acquire R-JPEG format infrared images of the pipeline. The acquired infrared images are then processed and analyzed using an improved R-JPEG-based method. 3 The Det rotating target detection algorithm trains and identifies bounding boxes in the heating pipeline area, then detects high-temperature anomalies within the heating pipeline area in infrared images and calculates and marks the leakage range. It then correlates GPS location information from infrared images showing leakage locations, finally outputting an image of the leak-marked heating pipeline along with its location information. This invention is non-contact, and inspectors only need to control a drone to fly along the pipeline, eliminating the need to damage the heating network pipeline structure or manually climb the pipeline for inspection. This significantly improves the efficiency and accuracy of inspecting heating pipeline leaks.

[0108] Example 2

[0109] like Figure 9 As shown in the figure, this embodiment provides a heating pipeline leak detection system, which includes:

[0110] Infrared image acquisition module 201 is used to acquire infrared images of heating pipelines via drone.

[0111] The infrared image is an infrared image in R-JPEG format.

[0112] Heating pipe identification model 202 is used to identify the heating pipe region in the infrared image; the heating pipe identification model is an improved R model based on a training set. 3 The improved R is obtained by training the Det network. 3 The convolution operations in the input layer of the feature pyramid network of the Det network employ variable convolution operations, and the improved R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the improved R... 3 The candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm.

[0113] The leak location determination module 203 is used to determine the location of the leak point in the heating pipeline area based on the heating pipeline area in the infrared image and using an adaptive threshold algorithm.

[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0115] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting leaks in heating pipelines, characterized in that, include: Infrared images of heating pipelines were collected using drones; A heating pipe recognition model is used to identify the infrared image, thereby obtaining the heating pipe region in the infrared image; the heating pipe recognition model is an improved R model based on a training set. 3 The improved R is obtained by training the Det network. 3 The convolution operations in the input layer of the feature pyramid network of the Det network employ variable convolution operations, and the improved R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the improved R... 3 The candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm; Based on the heating pipe area in the infrared image, an adaptive threshold algorithm is used to determine the location of the leak point in the heating pipe area. Specifically, this includes: acquiring the actual ambient temperature of the infrared image, using the actual ambient temperature to correct the temperature information in the infrared image; and using the adaptive threshold algorithm to determine the location of the leak point in the heating pipe area based on the corrected infrared image. Based on the heating pipe area in the corrected infrared image, an adaptive threshold algorithm is used to determine the location of leak points in the heating pipe area, specifically including: according to the formula Determine the leakage temperature threshold; identify pixels with temperatures exceeding the leakage temperature threshold as leakage points. Where V represents the leakage temperature threshold, and N represents the number of pixels in the heating pipe area. This represents the temperature value of the first pixel. This represents the temperature value of the second pixel. This represents the temperature value of the nth pixel. K represents the number of the highest temperature values ​​in the heating pipeline area, and K is the average temperature offset. The variable convolution operation is represented as follows: ； in, Represents the input feature map, Indicates position The convolution output, R Indicates the sampling area. Indicates the center point of the sampling area. This represents the nth point in the sampling region. This represents the convolution kernel weight coefficient corresponding to the nth point in the sampling region. Indicates the offset. Indicates the weighting coefficient; The H-Swish activation function is expressed as follows: ； Where x represents the input to the H-Swish activation function, and ReLU6() represents the ReLU6 activation function.

2. The method for detecting leaks in heating pipelines according to claim 1, characterized in that, Also includes: Based on the UAV positioning information carried in the infrared image, the positioning information of the heating pipeline in the infrared image is determined.

3. The method for detecting leaks in heating pipelines according to claim 1, characterized in that, The infrared image is an infrared image in R-JPEG format.

4. A leak detection system for heating pipelines, characterized in that, The heating pipeline leak detection system uses the heating pipeline leak detection method of claim 1, and the heating pipeline leak detection system includes: Infrared image acquisition module, used to acquire infrared images of heating pipelines via drone; A heating pipe identification model is used to identify the heating pipe region in the infrared image; the heating pipe identification model is an improved R model based on a training set. 3 The improved R is obtained by training the Det network. 3 The convolution operations in the input layer of the feature pyramid network of the Det network employ variable convolution operations, and the improved R... 3 In the Det network, the ReLU activation function is replaced with the H-Swish activation function, and the improved R... 3 The candidate anchor boxes in the Det network are determined by clustering the labeled boxes in the training set using the K-means++ clustering algorithm; The leak location determination module is used to determine the location of the leak point in the heating pipeline area based on the infrared image and using an adaptive threshold algorithm.

Images