Full-automatic detection method for internal diseases and defects of pipe gallery

Abstract

The invention discloses a full-automatic detection method for internal disease defects of a pipe gallery, and the method comprises the steps: carrying out the manual flight of an unmanned aerial vehicle, autonomously modeling a pipe gallery model, and carrying out the region division of the pipe gallery according to the model; building a local area network in the divided area sections; a bionic robot is integrated; the bionic robot is put into the pipe gallery for autonomous inspection; image data, poisonous gas, temperature and humidity and mechanical knocking data in the pipe gallery are collected; the side end carries out rapid positioning and alarming on the pipe gallery disease defect and carries out data returning; the cloud carries out refined quantitative analysis on the returned disease data; and automatically generating a detailed detection report based on an analysis result. The pipe gallery internal disease defect detection method is more convenient, faster and lower in cost, the automation degree and the detection efficiency of pipe gallery disease detection are improved, the adaptability and the obstacle crossing ability of a detection system to the complex environment of the pipe gallery are enhanced, the detection precision and the recognition ability of the disease defects are improved, and the detection accuracy and the recognition ability of the disease defects are improved. And an intelligent early warning and report generation mechanism is established.

Description

Technical Field

[0001] This invention relates to the field of pipe gallery defect detection technology, and in particular to a fully automated method for detecting defects inside pipe galleries. Background Technology

[0002] According to the current national standard "Technical Specification for Urban Integrated Pipe Gallery Engineering" (GB 50838-2015), underground pipe galleries refer to structures built underground in cities to accommodate two or more types of urban engineering pipelines and their ancillary facilities, including municipal pipelines, communication pipelines, power pipelines and water supply and drainage pipelines and their ancillary facilities.

[0003] With the rapid development of my country's economy, the scale of urban pipeline construction is also increasing. Underground integrated utility tunnels have been included in key livelihood projects, aiming to solve problems such as "road zipper" and "overhead spider web" of overhead wires, and enhance urban resilience. More and more cities are choosing to build underground integrated utility tunnels to achieve unified management, inspection, and maintenance of various pipelines. At the same time, the Ministry of Housing and Urban-Rural Development, together with multiple departments, has issued technical guidelines and operation and maintenance standards to promote standardized construction and intelligent management.

[0004] However, the rapid development of underground space construction in China has brought enormous pressure and challenges to the safety of underground utility tunnels. Due to the complex and variable underground environment, structural defects such as cracking, water leakage, and large deformations caused by soil and rock loads are common during the operation and maintenance of utility tunnels, leading to various underground space accidents. Any type of defect can have a significant impact on the operation of integrated utility tunnels. Therefore, how to quickly and accurately detect defects in integrated utility tunnels and prevent unknown operational disasters in advance to ensure the safety of people's lives and property and the smooth operation of cities has become a key and hot research topic.

[0005] Currently, the methods for inspecting utility tunnels across the country include: manual inspection, real-time monitoring with surveillance equipment, and timely warnings with sensors. The most mainstream method is manual inspection, where the utility tunnel operation department sends professional personnel to conduct full-process inspections of the tunnel at specific intervals. Although manual inspection is more comprehensive, it has problems such as long inspection time, high consumption of manpower, material resources and financial resources, and inability to detect defects in real time. While the deployment of monitoring facilities and sensors can provide good real-time detection, it has problems such as limited detection range and high detection costs.

[0006] Therefore, in order to solve the problems of long inspection time, high consumption of manpower, material resources and financial resources, and inability to detect defects in real time in the existing pipe gallery defect detection methods, it is an urgent technical problem to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a fully automated method for detecting defects inside pipe corridors.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A fully automated method for detecting defects inside utility tunnels includes the following steps:

[0010] S0. Use a drone to manually fly and autonomously model the utility tunnel, and divide the utility tunnel area according to the model;

[0011] S1. Build a local area network in the designated area using drones, robot dogs, or manual deployment methods;

[0012] S2, integrates a bionic robot, enabling it to have path planning and target detection functions;

[0013] S3. Deploy bionic robots inside the utility tunnel for autonomous inspection;

[0014] S4. During the inspection process, the bionic robot collects image data, toxic gas, temperature and humidity, and mechanical impact data of the inside of the pipe gallery through various devices.

[0015] S5. The edge end can quickly locate and alarm for defects in the pipe gallery and transmit data back.

[0016] S6. The cloud performs refined quantitative analysis on the returned disease data;

[0017] S7. Automatically generate detailed test reports based on analysis results.

[0018] Preferably, in step S0, the modeling method is as follows: the drone is manually operated to fly inside the utility tunnel, and the entire utility tunnel is modeled by laser point cloud scanning. The tunnel is divided into three sections as a region, and road spikes are set as positioning points in each section.

[0019] Preferably, in step S1, the method for building a local area network is as follows: a drone or robot dog equipped with a local area communication network device is operated to serve as a local area network base station at the positioning point to provide network services for the autonomous inspection bionic robot. If the environment of the utility tunnel is not suitable for the passage of the drone or robot dog, the local area network device is installed at the positioning point by manual deployment.

[0020] Preferably, step S2 specifically includes the following steps:

[0021] S21. Design a biomimetic crawling robot. Each mechanical leg consists of an upper section and a lower section. A steering motor, a second servo joint, and a cushioning airbag are installed at the connection between the two sections. The bottom of the lower mechanical leg is equipped with an anti-slip pad, which enables it to walk stably on uneven ground and has the ability to cross obstacles.

[0022] S22, the bionic robot is equipped with a path planning system, which plans the path according to the unmanned aerial vehicle (UAV) pipe gallery modeling environment to optimize the route during the inspection process;

[0023] S23. The bionic robot is equipped with a side-mounted computer as the control host, which is internally deployed with a target detection and recognition model to perform real-time detection of the inspection process.

[0024] Preferably, in step S3, the autonomous inspection method is as follows: the bionic robot automatically inspects the pipe gallery according to the planned path. When the environment inside the pipe gallery changes, the height inside the pipe gallery also changes, and the robot automatically adjusts the lifting range of each foot to overcome obstacles.

[0025] Preferably, step S4 specifically includes the following steps:

[0026] S41. The bionic robot is equipped with high-power lighting equipment and binocular high-definition distortion-free cameras to collect image data inside the pipe gallery.

[0027] S42. The bionic robot is equipped with a toxic gas detection sensor. When toxic gas is present in the surrounding environment, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform via the local area network.

[0028] S43. The bionic robot is equipped with temperature and humidity sensors to record the temperature and humidity inside the pipe gallery area in real time. Once the temperature or humidity is too high, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform through the local area network.

[0029] S44. The bionic robot is equipped with a tapping detection device on its feet. As the robot moves, it continuously taps the concrete inside the pipe gallery to determine whether there are voids or cracks. Once a problem is detected, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform via the local area network.

[0030] Preferably, step S5 specifically includes the following steps:

[0031] S51. SimAM assesses the importance of a neuron by defining an energy function to measure the difference between it and its surrounding neurons. The energy function for a neuron is defined according to the following formula:

[0032]

[0033] in, The activation value of the target neuron. This represents the activation value of other neurons in the same channel. , Yes and linear transformation, , For the weights and biases of the linear transformation, Assign labels to the target neurons. The total number of neurons in each channel, and Labels for the target neuron and other neurons;

[0034] This energy function aims to measure the linear separability of the target neuron from other neurons. To prevent overfitting and simplify calculations, a regularization term is introduced. The energy function is modified to satisfy the calculation formula:

[0035]

[0036] Specifically, SimAM derives an analytical solution for the energy function that satisfies the calculation formula:

[0037]

[0038]

[0039] in, This represents the mean value excluding the target neuron. The variance excluding the target neuron is calculated using the following formula:

[0040]

[0041]

[0042] Specifically, the SimAM attention mechanism assumes that all neurons in each channel follow the same distribution during computation, and uses the mean and variance of the entire channel for estimation, satisfying the calculation formula:

[0043]

[0044] in, For minimum energy, and These are the mean and variance of the channel, respectively;

[0045] The lower the energy of the target neuron, the greater its difference from other neurons, and therefore its higher importance. Thus, specific applications of SimAM... As attention weights, SimAM calculates the attention-weighted feature map according to the following formula:

[0046]

[0047] in, The energy value for each neuron, Representing element-wise multiplication, the Sigmoid function is used to map attention weights to the (0,1) interval to prevent the influence of excessively large values;

[0048] S52, Variable Convolutional Network DCNv4, introduces learnable offsets into the convolutional network, introducing a learnable offset amount at each sampling point. At the same time, a modulation mechanism is introduced, and modulation coefficients are added. To control the influence of each sampling point, the importance of different sampling points is weighted, improving the model's ability to suppress redundant regions and further enhancing the model's performance in complex scenes. Lightweight group convolution and multi-scale fusion modules are used to improve efficiency. The generation of offsets no longer depends on a single convolutional layer, and the definition of deformable convolution satisfies the calculation formula:

[0049]

[0050] in, Learned by an additional convolutional layer network, The sampling points are typically floating-point numbers, and bilinear interpolation is used to sample the input feature map. , For a predefined, fixed sampling offset.

[0051] The introduction of S53, SimAM and DCNv4 has formed a powerful synergistic effect. DCNv4 is responsible for capturing the disease features with geometric invariance, while SimAM is responsible for screening and amplifying the most discriminative parts of these features. Together, they solve the problems of deformation adaptation and background interference in the detection of pipe gallery diseases.

[0052] S54. Once the bionic robot identifies defects in the utility tunnel, it will quickly locate and alarm at the edge. It will alarm the three types of defects identified through three different voices and record the location of the detected defects in the utility tunnel section.

[0053] S55. The raw data, detection results and environmental parameters collected by the robot are transmitted back to the cloud processing platform through the established local area network communication module of the utility tunnel. The communication module supports multiple communication methods such as 5G, Wi-Fi or fiber optic to ensure the real-time performance and reliability of data transmission.

[0054] Preferably, step S6 specifically includes the following steps:

[0055] S61. Calculation of the area of ​​the leaking zone:

[0056] The program uses a model to detect images transmitted from the bionic inspection robot in the cloud. If no leakage is detected, the output is "no leakage". If leakage is detected, the program performs instance loop processing. In the single-instance processing flow, the program first obtains the mask and bounding box information after the SimAM-Seg model is run. Then, it converts the PyTorch tensor into a NumPy array and transfers it to the CPU to convert it into uint8 type. Since the model's fixed recognition window is 640×640, the image scaling ratio needs to be calculated. If the original image is not square, the image is filled. Subsequently, the effective area is cropped from the mask and the mask is restored to the original image size. Nearest neighbor interpolation is used to maintain the binary property. The total number of pixels with a value of 1 in the mask is counted, which is the pixel area of ​​the leaking water image in the segmented region. In the visualization operation, a copy of the original image is created as an overlay and the mask area is colored purple. Then, the original image and the overlay are blended using transparency to obtain the final output image.

[0057] S62. Calculation of crack width:

[0058] The algorithm is used in the cloud to find the central axis of the crack in the image, and then the normal vector of each point on the central axis is calculated. The kd-tree algorithm is used to find the set of nearest points on the central axis of any point, and then the normal vector of the point is obtained by SVD decomposition. This is used as the y-axis. The hband function is used to define the distance, and the points on the crack edge line within this distance are retained. Then the vband function is used to detect noise. If noise exists, the mean y-coordinate of each point on the crack edge line in the four quadrants divided by the coordinate axis is calculated first, and the points with y-coordinates greater than the mean are retained. The points closest to the y-axis in the four regions are calculated respectively, and denoted as A1, A2, B1, and B2. Connecting A1A2 and B1B2, the intersection points A and B with the y-axis are obtained. The distance between points A and B is the crack width at point C.

[0059] Preferably, in step S7, the structured inspection report includes statistical charts of disease classification, comprehensive health score of the utility tunnel, risk level distribution map, maintenance recommendations and priority ranking; the report is directly connected to the utility tunnel asset management system to provide data support for preventive maintenance and overhaul decisions.

[0060] The present invention achieves the following technical effects compared to the prior art:

[0061] (1) The bionic robot designed in this invention is free from the dependence of traditional wheeled or tracked robots on the flatness of the ground; whether it crosses obstacles, passes through waterlogged areas, or moves on the inclined or damaged surface of the corridor, it can maintain stability and achieve blind spot-free and full-coverage detection of the corridor space.

[0062] (2) By introducing the parameter-free SimAM attention module, the model has gained a powerful feature selection capability. The model is based on the energy function and can automatically identify and enhance the key neurons related to the disease in the feature map, while suppressing the interference of complex backgrounds (such as pipelines, supports, shadows, and water stains). This allows the model to concentrate limited computing resources on the most suspicious area, just like the human eye focuses, so as to accurately "lock" the target in the complex pipe gallery environment.

[0063] (3) Traditional standard convolution kernels are rigid when faced with geometric deformations (such as bending cracks, irregular corrosion, and perspective distortion) that are common in pipe corridors. The DCNv4 module of this invention introduces learnable offsets, which allows the convolution kernel to adaptively adjust the shape and position of its receptive field according to the content of the input features, thereby accurately matching the actual contour of the disease and greatly improving the generalization of the model to different manifestations of the same disease and the positioning accuracy of the detection box.

[0064] (4) The fully automated process of this invention frees manpower from dangerous, heavy and repetitive inspection work. The robot can perform routine and high-frequency autonomous inspections and can operate continuously, which greatly reduces the investment of professional inspection manpower and its safety risk costs, and significantly reduces the comprehensive operation and maintenance cost of the entire life cycle of the utility tunnel.

[0065] (5) The overall process of this invention, from data collection and analysis to report generation and trend prediction, forms a closed-loop management of pipe gallery defects, realizing the transformation of the operation and maintenance mode from "passive response" to "proactive early warning". At the same time, the reports automatically generated by the system include both current status statistics and trend prediction analysis of defects, providing a strong scientific basis for the formulation of maintenance plans and the optimization of resource allocation. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the detection process of an embodiment of the present invention;

[0067] Figure 2 The image shows a physical representation of the bionic robot used in the embodiments of this invention.

[0068] Figure 3 This is a schematic diagram of the path planning of the bionic robot in an embodiment of the present invention;

[0069] Figure 4 This is a schematic diagram of the SimAM principle in an embodiment of the present invention;

[0070] Figure 5 This is a schematic diagram of the segmentation mask for calculating the area of ​​water leakage defects in an embodiment of the present invention;

[0071] Where a and d are the original images of the water seepage; b and e are the segmentation mask images; and c and f are the overlay effect images.

[0072] Figure 6 This is a flowchart illustrating the calculation of the leakage area in an embodiment of the present invention.

[0073] Figure 7 This is a schematic diagram illustrating the calculation of crack width in an embodiment of the present invention; Detailed Implementation

[0074] 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.

[0075] like Figure 1 As shown, this invention discloses a fully automated method for detecting defects inside utility tunnels, comprising the following steps:

[0076] S0. Use a drone to manually fly and autonomously model the utility tunnel, and divide the utility tunnel area according to the model;

[0077] S1. Build a local area network in the designated area using drones, robot dogs, or manual deployment methods;

[0078] S2. Integrates a bionic robot, enabling it to perform functions such as path planning and target detection;

[0079] S3. Deploy bionic robots inside the utility tunnel for autonomous inspection;

[0080] S4. During the inspection process, the bionic robot collects image data, toxic gas, temperature and humidity, and mechanical impact data of the inside of the pipe gallery through various devices.

[0081] S5. The edge end can quickly locate and alarm for defects in the pipe gallery and transmit data back.

[0082] S6. The cloud performs refined quantitative analysis on the returned disease data;

[0083] S7. Automatically generate detailed test reports based on analysis results.

[0084] Specifically, in step S0, the modeling method is as follows: the drone is manually operated to fly inside the utility tunnel, and the entire utility tunnel is modeled by laser point cloud scanning. The tunnel is divided into three sections as a region, and road spikes are set as positioning points in each section.

[0085] Specifically, in step S1, the method for setting up a local area network is as follows: a drone or robot dog equipped with a local area communication network device is operated at the positioning point as a local area network base station to provide network services for the autonomous inspection bionic robot. If the environment of the utility tunnel is not suitable for the passage of the drone or robot dog, the local area network device is installed at the positioning point by manual deployment.

[0086] Specifically, S2 includes the following steps:

[0087] S21. Design a biomimetic crawling robot. Each mechanical leg consists of an upper section and a lower section. A steering motor, a second servo joint, and a cushioning airbag are installed at the connection between the two sections. The bottom of the lower mechanical leg is equipped with an anti-slip pad, which enables it to walk stably on uneven ground and has the ability to cross obstacles.

[0088] S22, the bionic robot is equipped with a path planning system, which plans the path according to the unmanned aerial vehicle (UAV) pipe gallery modeling environment to optimize the route during the inspection process;

[0089] S23. The bionic robot is equipped with a side-mounted computer as the control host, which is internally deployed with a target detection and recognition model to perform real-time detection of the inspection process.

[0090] Specifically, in step S3, the autonomous inspection method is as follows: the bionic robot automatically inspects the pipe gallery according to the planned path. When the environment inside the pipe gallery changes, that is, when the height inside the pipe gallery changes, the robot can automatically adjust the lifting range of each foot to overcome obstacles.

[0091] Specifically, step S4 includes the following steps:

[0092] S41. The bionic robot is equipped with high-power lighting equipment and binocular high-definition distortion-free cameras to collect image data inside the pipe gallery.

[0093] S42. The bionic robot is equipped with a toxic gas detection sensor. When toxic gas is present in the surrounding environment, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform via the local area network.

[0094] S43. The bionic robot is equipped with temperature and humidity sensors to record the temperature and humidity inside the pipe gallery area in real time. Once the temperature or humidity is too high, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform through the local area network.

[0095] S44. The bionic robot is equipped with a tapping detection device on its feet. As the robot moves, it continuously taps the concrete inside the pipe gallery to determine if there are any voids or cracks. Once a problem is detected, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform via the local area network.

[0096] Specifically, the S5 step includes the following steps:

[0097] S51. SimAM assesses the importance of a neuron by defining an energy function to measure the difference between it and its surrounding neurons. The energy function for a neuron is defined according to the following formula:

[0098]

[0099] in, The activation value of the target neuron. This represents the activation value of other neurons in the same channel. , Yes and linear transformation, , For the weights and biases of the linear transformation, Assign labels to the target neurons. The total number of neurons in each channel, and The energy function, which labels the target neuron and other neurons, is designed to measure the linear separability between the target neuron and other neurons. To prevent overfitting and simplify computation, a regularization term is introduced. The energy function is modified to satisfy the calculation formula:

[0100]

[0101] Specifically, SimAM derives an analytical solution for the energy function that satisfies the calculation formula:

[0102]

[0103]

[0104] in, This represents the mean value excluding the target neuron. The variance excluding the target neuron is calculated using the following formula:

[0105]

[0106]

[0107] Specifically, the SimAM attention mechanism assumes that all neurons in each channel follow the same distribution during computation, and uses the mean and variance of the entire channel for estimation, satisfying the calculation formula:

[0108]

[0109] in, For minimum energy, and These represent the mean and variance of the channels, respectively. The lower the energy of the target neuron, the greater its difference from other neurons, and therefore its higher importance. Thus, specific SimAM applications... As attention weights, SimAM calculates the attention-weighted feature map according to the following formula:

[0110]

[0111] in, The energy value for each neuron, Representing element-wise multiplication, the Sigmoid function is used to map attention weights to the (0,1) interval to prevent the influence of excessively large values.

[0112] S52, Variable Convolutional Network DCNv4, introduces learnable offsets into the convolutional network, introducing a learnable offset amount at each sampling point. At the same time, a modulation mechanism is introduced, and modulation coefficients are added. To control the influence of each sampling point, the importance of different sampling points is weighted, improving the model's ability to suppress redundant regions and further enhancing the model's performance in complex scenes. Lightweight group convolution and multi-scale fusion modules are used to improve efficiency. The generation of offsets no longer depends on a single convolutional layer, and the definition of deformable convolution satisfies the calculation formula:

[0113]

[0114] in, Learned by an additional convolutional layer network, The sampling points are typically floating-point numbers, and bilinear interpolation is used to sample the input feature map. , For a predefined, fixed sampling offset.

[0115] The introduction of S53, SimAM and DCNv4 has created a powerful synergistic effect. DCNv4 is responsible for capturing the defect features with geometric invariance, while SimAM is responsible for screening and amplifying the most discriminative parts of these features. Together, they solve the problems of deformation adaptation and background interference in the detection of defects in pipe gallery.

[0116] S54. Once the bionic robot identifies defects in the utility tunnel, it will quickly locate and alarm at the edge. It will alarm the three types of defects identified through three different voices and record the location of the detected defects in the utility tunnel section.

[0117] S55. The raw data, detection results and environmental parameters collected by the robot are transmitted back to the cloud processing platform through the established local area network communication module of the utility tunnel. The communication module supports multiple communication methods such as 5G, Wi-Fi or fiber optic to ensure the real-time and reliable transmission of data.

[0118] Specifically, step S6 includes the following steps:

[0119] S61. Calculation of the area of ​​the leaking water region: The image transmitted back by the bionic inspection robot in the cloud is used to detect leaks using a model. If no leak is detected, the output is "no leak." If a leak is detected, the program performs instance loop processing. In the single-instance processing flow, the program first obtains the mask and bounding box information after the SimAM-Seg model is run. Then, the PyTorch tensor is converted to a NumPy array and transferred to the CPU for conversion to uint8 type (0 or 1). Since the model's fixed recognition window is 640×640, the image scaling ratio needs to be calculated (using letterboxes to maintain the aspect ratio). If the original image is not square, the image is padded. Subsequently, the effective area is cropped from the mask, and the mask is restored to the original image size. Nearest neighbor interpolation is used to maintain the binary property, and the total number of pixels with a value of 1 in the mask is counted, which is the pixel area of ​​the segmented leaking water image. In the visualization operation, a copy of the original image is created as an overlay, and the mask area is colored purple. Then, the original image and the overlay are blended using transparency to obtain the final output image.

[0120] S62. Crack width calculation: In the cloud, the algorithm is used to find the central axis of the crack in the image, and then the normal vector of each point on the central axis is obtained. The kd-tree algorithm is used to find the set of nearest points on the central axis of any point, and then the normal vector of the point is obtained by SVD decomposition. This is used as the y-axis. The hband function is used to define the distance, and the points on the crack edge line within this distance are retained. Then the vband function (y-coordinate range limit) is used to detect noise. If noise exists, the mean y-coordinate of each point on the crack edge line in the four quadrants divided by the coordinate axis is calculated first, and the points with y-coordinates greater than the mean are retained. The points closest to the y-axis in the four regions are calculated respectively, and denoted as A1, A2, B1, and B2. Connect A1A2 and B1B2 to obtain the intersection points A and B with the y-axis. The distance between points A and B is the crack width at point C.

[0121] Specifically, in step S7, the system automatically generates a structured inspection report. The report includes statistical charts of defects, a comprehensive health score for the utility tunnel, a risk level distribution map, maintenance recommendations, and priority ranking. The report can be directly integrated into the utility tunnel asset management system, providing data support for preventative maintenance and overhaul decisions.

[0122] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A fully automated method for detecting defects inside utility tunnels, characterized in that, Includes the following steps: S0. Use a drone to manually fly and autonomously model the utility tunnel, and divide the utility tunnel area according to the model; S1. Build a local area network in the designated area using drones, robot dogs, or manual deployment methods; S2, integrates a bionic robot, enabling it to have path planning and target detection functions; S3. Deploy bionic robots inside the utility tunnel for autonomous inspection; S4. During the inspection process, the bionic robot collects image data, toxic gas, temperature and humidity, and mechanical impact data of the inside of the pipe gallery through various devices. S5. The edge end can quickly locate and alarm for defects in the pipe gallery and transmit data back. S6. The cloud performs refined quantitative analysis on the returned disease data; S7. Automatically generate detailed test reports based on analysis results.

2. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, In step S0, the modeling method is as follows: the drone is manually operated to fly inside the utility tunnel, and the entire utility tunnel is modeled by laser point cloud scanning. The tunnel is divided into three sections as a region, and road spikes are set as positioning points in each section.

3. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, In step S1, the method for setting up a local area network is as follows: a drone or robot dog equipped with a local area communication network device is operated to serve as a local area network base station at the positioning point to provide network services for the autonomous inspection bionic robot. If the environment of the utility tunnel is not suitable for the drone or robot dog to pass through, the local area network device is installed at the positioning point by manual deployment.

4. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Design a biomimetic crawling robot. Each mechanical leg consists of an upper section and a lower section. A steering motor, a second servo joint, and a cushioning airbag are installed at the connection between the two sections. The bottom of the lower mechanical leg is equipped with an anti-slip pad, which enables it to walk stably on uneven ground and has the ability to cross obstacles. S22, the bionic robot is equipped with a path planning system, which plans the path according to the unmanned aerial vehicle (UAV) pipe gallery modeling environment to optimize the route during the inspection process; S23. The bionic robot is equipped with a side-mounted computer as the control host, which is internally deployed with a target detection and recognition model to perform real-time detection of the inspection process.

5. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, In step S3, the autonomous inspection method is as follows: the bionic robot automatically inspects the pipe gallery according to the planned path. When the environment inside the pipe gallery changes, the height inside the pipe gallery also changes, and the robot automatically adjusts the lifting range of each foot to overcome obstacles.

6. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, The S4 step specifically includes the following steps: S41. The bionic robot is equipped with high-power lighting equipment and binocular high-definition distortion-free cameras to collect image data inside the pipe gallery. S42. The bionic robot is equipped with a toxic gas detection sensor. When toxic gas is present in the surrounding environment, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform via the local area network. S43. The bionic robot is equipped with temperature and humidity sensors to record the temperature and humidity inside the pipe gallery area in real time. Once the temperature or humidity is too high, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform through the local area network. S44. The bionic robot is equipped with a tapping detection device on its feet. As the robot moves, it continuously taps the concrete inside the pipe gallery to determine whether there are voids or cracks. Once a problem is detected, the robot will immediately sound an alarm and send the alarm location to the cloud monitoring platform via the local area network.

7. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, The S5 step specifically includes the following steps: S51. SimAM assesses the importance of a neuron by defining an energy function to measure the difference between it and its surrounding neurons. The energy function for a neuron is defined according to the following formula: in, The activation value of the target neuron. This represents the activation value of other neurons in the same channel. , Yes and linear transformation, , For the weights and biases of the linear transformation, Assign labels to the target neurons. The total number of neurons in each channel, and Labels for the target neuron and other neurons; This energy function aims to measure the linear separability of the target neuron from other neurons. To prevent overfitting and simplify calculations, a regularization term is introduced. The energy function is modified to satisfy the calculation formula: Specifically, SimAM derives an analytical solution for the energy function that satisfies the calculation formula: in, This represents the mean value excluding the target neuron. The variance excluding the target neuron is calculated using the following formula: Specifically, the SimAM attention mechanism assumes that all neurons in each channel follow the same distribution during computation, and uses the mean and variance of the entire channel for estimation, satisfying the calculation formula: in, For minimum energy, and These are the mean and variance of the channel, respectively; The lower the energy of the target neuron, the greater its difference from other neurons, and therefore its higher importance. Thus, specific applications of SimAM... As attention weights, SimAM calculates the attention-weighted feature map according to the following formula: in, The energy value for each neuron, Representing element-wise multiplication, the Sigmoid function is used to map attention weights to the (0,1) interval to prevent the influence of excessively large values; S52, Variable Convolutional Network DCNv4, introduces learnable offsets into the convolutional network, introducing a learnable offset amount at each sampling point. At the same time, a modulation mechanism is introduced, and modulation coefficients are added. To control the influence of each sampling point, the importance of different sampling points is weighted, improving the model's ability to suppress redundant regions and further enhancing the model's performance in complex scenes. Lightweight group convolution and multi-scale fusion modules are used to improve efficiency. The generation of offsets no longer depends on a single convolutional layer, and the definition of deformable convolution satisfies the calculation formula: in, Learned by an additional convolutional layer network, The sampling points are typically floating-point numbers, and bilinear interpolation is used to sample the input feature map. , For a predefined, fixed sampling offset. The introduction of S53, SimAM and DCNv4 has formed a powerful synergistic effect. DCNv4 is responsible for capturing the defect features with geometric invariance, while SimAM is responsible for filtering and amplifying the most discriminative parts of these features. Together, they solve the problems of deformation adaptation and background interference in the detection of defects in pipe gallery. S54. Once the bionic robot identifies defects in the utility tunnel, it will quickly locate and alarm at the edge. It will alarm the three types of defects identified through three different voices and record the location of the detected defects in the utility tunnel section. S55. The raw data, detection results and environmental parameters collected by the robot are transmitted back to the cloud processing platform through the established local area network communication module of the utility tunnel. The communication module supports multiple communication methods such as 5G, Wi-Fi or fiber optic to ensure the real-time and reliable transmission of data.

8. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, Step S6 specifically includes the following steps: S61. Calculation of the area of ​​the leaking zone: The program uses a model to detect images transmitted from the bionic inspection robot in the cloud. If no leakage is detected, the output is "no leakage". If leakage is detected, the program performs instance loop processing. In the single-instance processing flow, the program first obtains the mask and bounding box information after the SimAM-Seg model is run. Then, it converts the PyTorch tensor into a NumPy array and transfers it to the CPU to convert it into uint8 type. Since the model's fixed recognition window is 640×640, the image scaling ratio needs to be calculated. If the original image is not square, the image is filled. Subsequently, the effective area is cropped from the mask and the mask is restored to the original image size. Nearest neighbor interpolation is used to maintain the binary property. The total number of pixels with a value of 1 in the mask is counted, which is the pixel area of ​​the leaking water image in the segmented region. In the visualization operation, a copy of the original image is created as an overlay and the mask area is colored purple. Then, the original image and the overlay are blended using transparency to obtain the final output image. S62. Calculation of crack width: The algorithm is used in the cloud to find the central axis of the crack in the image, and then the normal vector of each point on the central axis is calculated. The kd-tree algorithm is used to find the set of nearest points on the central axis of any point, and then the normal vector of the point is obtained by SVD decomposition. This is used as the y-axis. The hband function is used to define the distance, and the points on the crack edge line within this distance are retained. Then the vband function is used to detect noise. If noise exists, the mean y-coordinate of each point on the crack edge line in the four quadrants divided by the coordinate axis is calculated first, and the points with y-coordinates greater than the mean are retained. The points closest to the y-axis in the four regions are calculated respectively, and denoted as A1, A2, B1, and B2. Connecting A1A2 and B1B2, the intersection points A and B with the y-axis are obtained. The distance between points A and B is the crack width at point C.

9. The fully automated detection method for internal defects in pipe corridors according to claim 1, characterized in that, In step S7, the structured inspection report includes statistical charts of disease classification, comprehensive health score of the utility tunnel, risk level distribution map, maintenance recommendations and priority ranking; the report is directly connected to the utility tunnel asset management system to provide data support for preventive maintenance and overhaul decisions.

Images