A dam leakage detection method and system based on a patrol robot dog

Abstract

The application provides a dam leakage detection method and system based on a patrol robot dog. The patrol robot dog comprises an image acquisition module, a soil humidity detection module, an edge calculation module and a communication module. The soil humidity detection module is arranged at the end of the four legs of the patrol robot dog. The method comprises the following steps: generating an initial inspection path based on a regional plane map; identifying obstacles and dynamically avoiding obstacles through the image acquisition module during the inspection process; identifying the ground type by linking the image acquisition module and the soil humidity detection module; collecting the original data set of the measured area and sending it to the edge calculation module; calculating and cleaning the effective soil moisture content data; comparing the effective soil moisture content data with the preset soil moisture content safety threshold; and transmitting the soil moisture content data exceeding the soil moisture content safety threshold to the inspector in the form of a warning. The application has the effect of real-time and efficient guarantee of safe operation of the dam.

Description

Technical Field

[0001] This invention belongs to the field of dam inspection technology, specifically relating to a dam leakage detection method and system based on an inspection robot dog. Background Technology

[0002] Dikes are indispensable and crucial facilities in water conservancy projects, playing a multifaceted and vital role. They effectively constrain water flow, prevent flooding, and provide reliable protection for industrial and agricultural production and the safety of people's lives and property on both banks. Therefore, dike inspections are essential because, over long-term operation, dikes are affected by various factors such as water erosion, geological changes, and biological erosion, making them prone to safety hazards such as cracks, seepage, and deformation. Especially during the main flood season, when heavy rainfall is concentrated, the impact and soaking of floodwaters on dikes intensifies, making these hazards more apparent. If these hazards are not detected and addressed promptly, small problems can develop into major accidents, seriously threatening dike safety, endangering the lives and property of people along the banks, and disrupting normal production and living order.

[0003] Existing dam inspection technologies are gradually evolving towards intelligent systems, with intelligent inspection robots replacing traditional manual inspections in some scenarios. However, current mainstream intelligent inspection robots primarily rely on single image acquisition modules such as high-definition cameras and infrared thermal imagers as their core sensing methods, resulting in significant limitations in detection. The effectiveness of image acquisition modules depends on lighting conditions, and in complex environments such as heavy rain, fog, and nighttime, images are prone to blurring and feature loss, leading to missed detections or misjudgments. Furthermore, vegetation cover on the dam surface and water reflections can interfere with image feature extraction, further reducing detection reliability. Moreover, image acquisition can only reflect visual features such as cracks and collapses on the dam surface, while seepage problems often initially manifest as abnormal soil moisture. For example, water seepage inside the dam can increase the surface soil moisture content, and the corresponding latent features of seepage problems cannot be directly identified through images, making it difficult to detect seepage risks in the early stages. In addition, single visual features are insufficient to quantify the degree of seepage, only qualitatively determining the presence of anomalies, and cannot provide accurate data support for subsequent treatment. Therefore, how to overcome the technical bottleneck of single image acquisition and combine multi-dimensional sensor data to achieve accurate and real-time detection of dam seepage has become a key direction for upgrading existing intelligent inspection technology to in-depth applications. Summary of the Invention

[0004] This invention provides a method and system for detecting seepage in dams based on a patrol robot dog, in order to solve the problems of weak anti-interference ability, low recognition accuracy, low detection reliability, and difficulty in achieving rapid real-time patrol when using a single image acquisition module for dam inspection.

[0005] In a first aspect, the present invention provides a method for detecting seepage in dams based on a patrol robot dog, which is applied to the patrol robot dog. The patrol robot dog includes an image acquisition module, a soil moisture detection module, an edge computing module, and a communication module. The soil moisture detection module is respectively disposed at the ends of the four legs of the patrol robot dog. The image acquisition module and the soil moisture detection module are both connected to the edge computing module, and the edge computing module is connected to the communication module.

[0006] The method includes the following steps:

[0007] Obtain a regional planar map of the target inspection area, and generate an initial inspection path for the inspection robot dog based on the regional planar map and a path planning algorithm.

[0008] During the inspection process of the inspection robot dog following the initial inspection path, the image acquisition module acquires environmental images in front of the inspection robot dog in real time, identifies target obstacles in front of the inspection robot dog based on the environmental images, and realizes dynamic obstacle avoidance of the inspection robot dog for the target obstacles.

[0009] The image acquisition module and the soil moisture detection module work together. The image acquisition module identifies the surface type of the area to be measured in front of the inspection robot dog, controls the inspection robot dog to move to the area to be measured, and collects the raw dataset of the ground in the area to be measured through the soil moisture detection module, and sends the raw dataset to the edge computing module.

[0010] The edge computing module decomposes the original dataset into soil moisture content data and performs abnormal data cleaning on the soil moisture content data to obtain effective soil moisture content data.

[0011] The effective soil moisture content data is compared with the preset soil moisture content safety threshold, and the effective soil moisture content data exceeding the soil moisture content safety threshold is transmitted to the inspection personnel in the target inspection area in the form of an early warning through the communication module.

[0012] Optionally, the image acquisition module acquires real-time environmental images in front of the inspection robot dog, identifies target obstacles in front of the inspection robot dog based on the environmental images, and performs dynamic obstacle avoidance for the inspection robot dog against the target obstacles, including the following steps:

[0013] The image acquisition module continuously acquires environmental images in front of the inspection robot dog, and performs real-time preprocessing on the environmental images to obtain preprocessed environmental images.

[0014] The preprocessed environmental image is input into a pre-trained target detection model, which identifies the target obstacle in front of the inspection robot dog in the preprocessed environmental image and calculates the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints.

[0015] The target obstacle is projected onto the area plane map where the initial inspection path is located by combining the multi-view geometric constraints and the three-dimensional position coordinates. The shortest distance between the projection of the target obstacle and the initial inspection path is calculated in the area plane map. If the shortest distance is less than or equal to a preset distance safety threshold, a path conflict is determined to have occurred.

[0016] When the path conflict occurs, the projection center point of the target obstacle that caused the path conflict is taken as the path conflict point.

[0017] Based on the preset spatial range of the inspection robot dog's movement radius, the path planning algorithm is invoked to replan the optimal temporary path for bypassing the target obstacle within the preset spatial range and with the path conflict point as the center. The starting point of the optimal temporary path is the real-time position of the inspection robot dog when the path conflict occurs, and the ending point of the optimal temporary path coincides with the initial inspection path.

[0018] Optionally, the image acquisition module includes a first camera and a second camera, which are respectively positioned at the left and right eyes of the inspection robot dog's head. The preprocessed environmental image includes a first environmental image acquired by the first camera and a second environmental image acquired by the second camera. The multi-view geometric constraints include the bounding box of the target obstacle in the preprocessed environmental image and the disparity map of the target obstacle.

[0019] The step of inputting the preprocessed environment image into a pre-trained target detection model, identifying target obstacles in front of the inspection robot dog in the preprocessed environment image through the target detection model, and calculating the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints includes the following steps:

[0020] The target detection model identifies the bounding boxes of target obstacles in the preprocessed environmental image;

[0021] The parallax map of the target obstacle located in front of the inspection robot dog is calculated by processing the first environmental image and the second environmental image into a target obstacle based on the correspondence between the first environmental image and the second environmental image.

[0022] The three-dimensional position coordinates of the target obstacle relative to the inspection robot dog are calculated by combining the bounding box, the disparity map, and the camera parameters of the first and second cameras in the image acquisition module.

[0023] Optionally, the method further includes the following optimization step for the initial inspection path:

[0024] The preprocessed environmental image is input into the target detection model. The target detection model is trained on a dataset containing obstacles in a dam scene. The feature extraction network in the target detection model extracts feature maps of the texture and contour features of the obstacles in the preprocessed environmental image. Based on the feature maps, the target obstacles are classified and identified, and the obstacle category of the target obstacles is output.

[0025] The actual size data of the target obstacle is obtained by combining the camera parameters of the first camera and the second camera in the image acquisition module with the pixel size of the bounding box;

[0026] The three-dimensional position coordinates of the target obstacle, the obstacle category, the actual size data, and the regional planar map of the target inspection area are fused to construct an optimized regional planar map of the target inspection area. The optimized regional planar map is used to optimize the initial inspection path in the next inspection cycle.

[0027] Optionally, the soil moisture detection module is a near-infrared spectral sensor. The steps of identifying the surface type of the area to be measured in front of the inspection robot dog through the image acquisition module, controlling the inspection robot dog to move to the area to be measured, acquiring the raw dataset of the ground surface in the area to be measured through the near-infrared spectral sensor, and sending the raw dataset to the edge computing module include the following steps:

[0028] The image acquisition module inputs the environmental image of the area to be measured in front of the inspection robot dog into a surface classification model trained with typical surface samples of the dam scene. The surface classification model identifies the surface type of the environmental image, which includes soil, rock, vegetation and water.

[0029] When the surface type is identified as non-soil, the data collection for the corresponding area to be measured is abandoned.

[0030] When the surface type is identified as soil, the corresponding area to be measured is marked as the target area to be measured, and the inspection robot dog is guided to move to the target area to be measured. After the inspection robot dog moves to the target area to be measured, the ground of the target area to be measured is detected by the near-infrared spectral sensor.

[0031] The near-infrared spectral sensor emits near-infrared light into the soil of the target area to be measured, and the photoelectric detection unit of the near-infrared spectral sensor receives the reflected light signal formed after the beam of near-infrared light is reflected by the soil.

[0032] The reflected light signal is converted into digital spectral data by the photoelectric detection unit.

[0033] The image acquisition module synchronously records the position coordinates of the inspection robot dog equipped with the near-infrared spectral sensor, and stores the digital spectral data in association with the position coordinates to form the original dataset.

[0034] The original dataset is sent to the edge computing module.

[0035] Optionally, the edge computing module resolves the original dataset into soil moisture content data, and performs anomaly cleaning on the soil moisture content data to obtain valid soil moisture content data, including the following steps:

[0036] The edge computing module performs real-time preprocessing on the original dataset, and then inputs the preprocessed original dataset into the pre-trained solution model to obtain soil moisture content data.

[0037] The isolated forest algorithm was used to construct a random decision tree to identify isolated points in the soil moisture content data that deviated from the normal distribution as potential outliers.

[0038] The LOF algorithm is used to calculate the local outlier factor of the soil moisture content data. When the value of the local outlier factor exceeds a preset threshold, the corresponding local outlier factor is marked as outlier data.

[0039] The intersection of the potential abnormal data and the outlier data is taken as the abnormal data, and the abnormal data is removed to obtain the effective soil moisture content data.

[0040] Optionally, the method further includes an environmental self-calibration step for solving the model:

[0041] The inspection robot dog is guided to one or more preset reference areas via a remote control terminal. The reference area includes at least one dry standard soil area and at least one saturated standard soil area, and the soil type of the reference area is consistent with the soil type of the target inspection area.

[0042] The near-infrared spectral sensor collects reference spectral data of the soil in the reference area, and the reference soil moisture content data in the reference area is calculated based on the reference spectral data and the edge computing module, and a mapping relationship between the reference spectral data and the reference soil moisture content data is established.

[0043] Based on the mapping relationship, a mapping relationship model for the target inspection area is generated, and the environmental self-calibration of the solution model is completed through the mapping relationship model.

[0044] Optionally, comparing the effective soil moisture content data with a preset safety threshold and transmitting the effective soil moisture content data exceeding the safety threshold to the inspection personnel in the target inspection area in the form of an early warning via the communication module includes the following steps:

[0045] The safety threshold is set based on the soil type in the target inspection area, and the safety threshold is pre-stored through the edge computing module;

[0046] Effective soil moisture content data are divided into zones according to the location coordinates of the target inspection area;

[0047] For any of the aforementioned partitions, the effective soil moisture content data is compared with the safety threshold. When the effective soil moisture content data exceeds the safety threshold, the abnormality level is classified based on the proportion of the effective soil moisture content data exceeding the safety threshold, and the corresponding partition is marked as a warning partition and associated with the location coordinates and measurement time of the warning partition.

[0048] The soil moisture content data, location coordinates, measurement time, and anomaly level of the warning zone are packaged into a standardized warning frame. The standardized warning frame is sent to the inspection terminal held by the inspection personnel in the target inspection area through the communication module. After receiving the standardized warning frame, the inspection terminal triggers an audio-visual prompt and marks the location of the warning zone on the environmental map displayed on the inspection terminal.

[0049] In a second aspect, the present invention also provides a dam leakage detection system based on an inspection robot dog, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the dam leakage detection method based on an inspection robot dog as described in any one of the first aspects.

[0050] Thirdly, the present invention also provides a computer-readable storage medium storing instructions, characterized in that, when executed by a processor, the instructions cause the processor to be configured to perform the dam leakage detection method based on the inspection robot dog according to any one of the first aspects.

[0051] The beneficial effects of this invention are:

[0052] By integrating image acquisition, soil moisture detection, edge computing, and communication modules, and incorporating path planning algorithms, the inspection robot achieves real-time, accurate perception and efficient early warning of dam seepage risks. This transforms passive inspections relying on a single image acquisition module into an intelligent and proactive monitoring mode. An initial inspection path is generated based on a regional planar map, and the image acquisition module identifies obstacles in real time and dynamically avoids them. While ensuring comprehensive inspection coverage, this effectively improves the smoothness and environmental adaptability of the inspection process, balancing systematicness and flexibility. The robot also works in conjunction with the image acquisition module to identify surface types, guiding the soil moisture detection module to target and collect raw data from the area to be measured. After being processed and cleaned by the edge computing module, effective soil moisture content data is obtained, ensuring the relevance and authenticity of the detection data. Optimization of the data processing improves the accuracy of seepage risk assessment, making the detection results more valuable. Finally, the communication module quickly relays abnormal data exceeding safety thresholds to inspection personnel as early warnings, constructing a complete link from data acquisition and analysis to risk transmission. This transforms the traditional passive response to seepage problems into a proactive early warning system, significantly accelerating risk response speed. More importantly, through the coordinated operation and process design of each module, the system can achieve autonomous inspection, dynamic adjustment, accurate detection, and rapid early warning, continuously optimizing the efficiency and reliability of dam leakage detection, and ultimately forming an intelligent monitoring system adapted to the dam environment, providing comprehensive protection for the safe operation of the dam. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the specific structure of the inspection robot dog in one embodiment of this application.

[0054] Figure 2 This is a flowchart illustrating a method for detecting dam seepage based on an inspection robot dog in one embodiment of this application. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0056] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0057] This invention discloses a method for detecting dam seepage based on an inspection robot dog, and the corresponding method is applied to the inspection robot dog, such as... Figure 1 As shown, the inspection robot dog includes an image acquisition module, a soil moisture detection module, an edge computing module, and a communication module. The soil moisture detection module is located at the ends of the four legs of the inspection robot dog, the image acquisition module is located at the head of the inspection robot dog, the edge computing module can be located at the abdomen of the inspection robot dog, and the communication module can be located at the back of the inspection robot dog. The image acquisition module and the soil moisture detection module are both connected to the edge computing module, and the edge computing module is connected to the communication module.

[0058] Figure 2 This is a flowchart illustrating a dam seepage detection method based on an inspection robot dog in one embodiment. It should be understood that, although... Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps. For example Figure 2 As shown, the dam leakage detection method based on an inspection robot dog disclosed in this invention specifically includes the following steps:

[0059] S101. Obtain a regional planar map of the target inspection area, and generate the initial inspection path for the inspection robot dog based on the regional planar map and using a path planning algorithm.

[0060] The regional planar map is obtained using a GIS (Geographic Information System). A new network connection is created within the GIS, and the URL of the satellite map corresponding to the geographical scope of the target inspection area is entered. While connected to the network, the satellite map covering the corresponding area is obtained. The corresponding satellite map refers to a satellite map that matches the geographical scope of the target inspection area and whose coverage area includes the target inspection area. Subsequently, based on the pre-defined target inspection area, the obtained satellite map is precisely cropped to obtain a regional planar map containing only the target inspection area, ensuring that the coverage area is completely consistent with the target inspection area. Next, based on the obtained regional planar map, an initial inspection path is generated using a path planning algorithm. Before planning, the starting point and target point need to be clearly defined. The starting point is the starting position of the inspection robot, and the target point is the destination that the inspection robot needs to reach within the target inspection area. The path planning algorithm can employ A... The algorithm inputs the cropped area planar map into A. The algorithm generates a collision-free path node sequence from the starting point to the target point. This node sequence is a discrete set of nodes containing travel parameters. Then, based on this node sequence, a continuous path processing method is applied. An interpolation algorithm generates smooth transition path segments between adjacent nodes, transforming the discrete node sequence into a continuous path curve to ensure the robot dog's motion is consistent. The generated path is the robot dog's initial inspection path.

[0061] S102. During the inspection process of the inspection robot dog following the initial inspection path, the image acquisition module acquires environmental images in front of the inspection robot dog in real time, identifies target obstacles in front of the inspection robot dog based on the environmental images, and realizes dynamic obstacle avoidance of the inspection robot dog for the target obstacles.

[0062] When the inspection robot dog patrols along its initial patrol path, the image acquisition module includes a first camera and a second camera, mounted on the robot dog's left and right eyes, respectively. The lenses of both cameras face directly forward. The image acquisition module also includes a high-definition industrial camera and an adaptive fill light. The high-definition camera has a resolution of 1920×1080, a frame rate of 30fps, and a field of view of 60 degrees. The fill light intensity is adjustable. The image acquisition module captures light from the front through the camera lens, converts the light signal into an electrical signal through an image sensor, and then processes it into a digital environmental image through analog-to-digital conversion. The fill light automatically turns on in low-light environments to ensure image brightness. The environmental image refers to a real-time image of the ground, vegetation, rocks, water, and other scenery along the robot dog's patrol path. The environmental image undergoes preprocessing, including distortion correction, grayscale conversion, noise reduction, and edge enhancement. Then, a deep learning-based target detection algorithm is used to extract and analyze features from the image, identifying objects significantly different from the background as target obstacles, and determining the distance, orientation, size, and shape of these obstacles relative to the robot dog. The target detection algorithm can be an image detection model based on the YOLOv5 framework. Based on the obstacle information of the identified target obstacles, the inspection robot dog adjusts its own motion parameters, including turning angle, stride length, and movement speed, to avoid obstacles by changing its inspection trajectory, thus achieving dynamic obstacle avoidance and ensuring the continuous progress of the inspection process.

[0063] S103. Link the image acquisition module and the soil moisture detection module. The image acquisition module identifies the surface type of the area to be measured in front of the inspection robot dog, controls the inspection robot dog to move to the area to be measured, and collects the raw data set of the ground in the area to be measured through the soil moisture detection module, and sends the raw data set to the edge computing module.

[0064] The soil moisture detection module can be a near-infrared spectral sensor. A synchronization signal is triggered by the real-time clock (RTC) of the edge computing module, causing the image acquisition module and the near-infrared spectral sensor to synchronize at 10ms intervals, entering a real-time collaborative state. For each frame of environmental image acquired by the image acquisition module (30fps), the near-infrared spectral sensor synchronously updates its standby state. While the image acquisition module continuously acquires environmental images of the ground in front of the inspection robot dog, the near-infrared spectral sensor is synchronously ready to collect data at any time. Simultaneously, a deep learning-based surface classification model (which can be a ResNet-50 model in specific implementations) processes the real-time acquired environmental images. The surface classification model is specifically trained on a dataset containing surface type image samples including soil, rocks, vegetation, and water bodies. The model uses surface type as a specific target, extracting texture, color, and roughness features of the ground area in the environmental images, matching them with the trained surface type features, and outputting the surface type of the area in front in real time. If the surface classification model determines the surface type to be soil, the near-infrared spectral sensor is activated to emit near-infrared light to illuminate the ground. The reflected near-infrared light is received, and the internal photoelectric conversion element converts the light signal into an electrical signal, acquiring a raw dataset of the corresponding area in real time. This raw dataset includes data reflecting soil moisture status, such as near-infrared spectral reflectance and light intensity attenuation values ​​at different wavelengths. This raw dataset is then sent to the edge computing module in real time via a data transmission interface connected to the edge computing module, providing foundational data for subsequent data processing. Simultaneously, the image acquisition module continuously tracks the area ahead, repeating the above identification and acquisition process when moving to a new area. If the surface classification model determines the surface type is not soil, the near-infrared spectral sensor is not activated for data acquisition. The robot continues to patrol along the initial patrol path, and the image acquisition module continues to identify the environmental image and surface type of the new area ahead in real time.

[0065] S104. The original dataset is solved into soil moisture content data through the edge computing module, and the soil moisture content data is cleaned of abnormal data to obtain effective soil moisture content data.

[0066] The process involves using an edge computing module to call a pre-defined solution model, which can be a Partial Least Squares Regression (PLSR) model. The PLSR model is based on the statistical relationship between spectral data and soil moisture content, effectively addressing multicollinearity issues in spectral data. Near-infrared spectral reflectance and light intensity attenuation values ​​from the original dataset are input into the PLSR model as independent variables. Utilizing a pre-trained mapping relationship between spectral data and soil moisture content, the model extracts, analyzes, and transforms the spectral data to calculate the corresponding soil moisture content data. After obtaining the soil moisture content data, the edge computing module initiates an anomaly cleaning process. First, the Isolation Forest algorithm is run to construct multiple isolated trees, performing isolated analysis on the soil moisture content data. Anomalies that deviate significantly from the majority of data in the feature space and are easily isolated are identified and marked as potential anomalies. Next, the LOF algorithm is run through the edge computing module. By comparing the density differences between each soil moisture content data point and its surrounding neighboring data points, local anomalies are further identified and marked as local outliers. Finally, the intersection of potential abnormal data and local outliers is taken as abnormal data through the edge computing module. The abnormal data is then removed from the soil moisture content data to obtain effective soil moisture content data, which provides reliable data support for subsequent safety threshold judgment and early warning.

[0067] S105. Compare the effective soil moisture content data with the preset safety threshold, and transmit the effective soil moisture content data exceeding the safety threshold to the inspection personnel in the target inspection area in the form of an early warning through the communication module.

[0068] The process involves obtaining effective soil moisture content data through the edge computing module, then invoking a preset safety threshold parameter. This threshold parameter is set according to the soil safety moisture content standard for the soil type in the target inspection area, and includes an upper limit for soil moisture content. The edge computing module compares the effective soil moisture content data of each data point with the safety threshold. If the effective soil moisture content data exceeds the upper limit of the safety threshold, it is determined to be abnormal data; otherwise, it is considered normal data. For abnormal data, the abnormality level is determined based on the proportion of effective soil moisture content exceeding the safety threshold; normal data is not processed. For effective soil moisture content data determined to be abnormal, the edge computing module generates a corresponding standardized warning frame. This standardized warning frame includes the corresponding soil moisture content data, corresponding location coordinates, corresponding measurement time, and corresponding abnormality level. Subsequently, the standardized warning frame is sent to the communication module via the edge computing module. The communication module can use 4G wireless communication, based on the communication protocol of the receiving terminal held by the inspection personnel in the target inspection area, to encode the standardized warning frame and send it out via wireless signal. After the inspection personnel's receiving terminal receives the wireless signal, it decodes the standardized warning frame and restores it to recognizable warning content, reminding the inspection personnel in the form of sound prompts, text pop-ups, or vibrations, thus completing the warning transmission.

[0069] In one embodiment, the image acquisition module acquires real-time environmental images in front of the inspection robot dog, identifies target obstacles in front of the inspection robot dog based on the environmental images, and performs dynamic obstacle avoidance for the inspection robot dog against the target obstacles, including the following steps:

[0070] The image acquisition module continuously acquires environmental images in front of the inspection robot dog, and performs real-time preprocessing on the environmental images to obtain preprocessed environmental images.

[0071] The preprocessed environmental image is input into the pre-trained target detection model. The target detection model identifies the target obstacle in front of the inspection robot dog in the preprocessed environmental image and calculates the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints.

[0072] By combining multi-view geometric constraints and three-dimensional position coordinates, the target obstacle is projected onto the area plane map where the initial inspection path is located. The shortest distance between the projection of the target obstacle and the initial inspection path is calculated in the area plane map. If the shortest distance is less than or equal to a preset safety threshold, a path conflict is determined to have occurred.

[0073] When the path conflict occurs, the projection center point of the target obstacle that caused the path conflict is taken as the path conflict point.

[0074] Based on the preset spatial range of the inspection robot dog's movement radius, the path planning algorithm is invoked to replan the optimal temporary path for bypassing the target obstacle within the preset spatial range and with the path conflict point as the center. The starting point of the optimal temporary path is the real-time position of the inspection robot dog when the path conflict occurs, and the ending point of the optimal temporary path coincides with the initial inspection path.

[0075] In this embodiment, the image acquisition module on the inspection robot dog acquires real-time images of the environment in front of the robot dog. The process of identifying target obstacles and dynamically avoiding them based on these environmental images is as follows: The image acquisition module is activated and enters continuous operation. The module includes a high-definition industrial camera and an adaptive fill light. The camera resolution is set to 1920×1080, the frame rate is maintained at 30fps, and the field of view covers a 75-degree range. The fill light automatically adjusts its brightness within the range of 300-1500 lumens according to the ambient light. The image acquisition module continuously captures environmental images in front of the inspection robot dog. These images include objects that may constitute obstacles, such as rocks, tree trunks, and puddles, as well as the surrounding terrain background. The acquired environmental images are preprocessed in real-time. Gaussian filtering is used to remove noise caused by sensor noise or light fluctuations. Histogram equalization is used to adjust image contrast and brightness to enhance the clarity of object outlines. Simultaneously, the image size is uniformly scaled to a fixed specification of 640×480 pixels, resulting in a preprocessed environmental image.

[0076] The preprocessed environment image is input into the pre-trained target detection model, which includes an obstacle recognition model based on the YOLOv5 framework. The obstacle recognition model is trained with image samples containing various types of obstacles and can identify target obstacles in front of the inspection robot dog in the preprocessed environment image. It also calculates the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints.

[0077] Subsequently, based on multi-view geometric constraints and three-dimensional position coordinates, the spatial position of the target obstacle is projected onto the regional planar map where the initial inspection path is located. The regional planar map is a pre-constructed two-dimensional planar model that includes the inspection route and the surrounding environment. The specific steps for projecting the spatial position of the target obstacle onto the regional planar map of the initial inspection path are as follows: First, determine the coordinate system of the regional planar map. Establish a two-dimensional rectangular coordinate system with the origin at the starting point of the initial inspection path, the X-axis along the initial inspection path, and the Y-axis perpendicular to the initial inspection path and pointing towards the inside of the region. Obtain the real-time coordinates (X0, Y0) of the inspection robot dog in the corresponding coordinate system. These coordinates are obtained by combining the 1m positioning accuracy GPS positioning module on the inspection robot dog with HMM (Hidden Markov Model) map matching technology. By probabilistically matching the GPS real-time coordinates with the road network nodes of the regional planar map, positioning drift is corrected, ensuring that the coordinate error is ≤0.5m. Convert the three-dimensional position coordinates (Δx, Δy, Δz) of the target obstacle relative to the robot dog into planar coordinates (Δx, Δy) relative to the robot dog, ignoring the height information Δz, and ensuring that the directions of Δx and Δy are consistent with the coordinate system of the regional planar map. Through coordinate superposition calculation, obtain the absolute coordinates (X0+Δx, Y0+Δy) of the target obstacle in the regional planar map, thus forming the projection of the target obstacle onto the regional planar map. The shortest distance between the target obstacle projection and the initial inspection path is calculated on the regional planar map. If the shortest distance is less than or equal to a preset distance safety threshold, a path conflict is determined to have occurred. The distance safety threshold is set according to the size and mobility of the inspection robot. For example, for a robot that is 50cm long and 30cm wide, the distance safety threshold is set to 30cm, meaning that a path conflict is determined when the shortest distance between the obstacle and the path is ≤30cm.

[0078] When a path conflict is detected, the geometric center of the area projected onto the regional planar map of the target obstacle is taken as the path conflict point, which represents the core location of the conflict between the obstacle and the path. Based on the movement radius of the inspection robot dog, a pre-set area of ​​space that it can detour is defined. The movement radius refers to the maximum radius of movement of the robot dog when it turns or moves around itself. The pre-set area is a circular region centered on the path conflict point and bounded by the movement radius.

[0079] Call A The algorithm replans the optimal temporary path to bypass the target obstacle within a preset space and centered on the path conflict point. The specific steps are as follows: First, perform A... The algorithm is initialized with the real-time location of the inspection robot dog at the time of the path conflict as the starting point and the point on the initial inspection path that connects with the temporary path as the ending point, with a preset spatial range of A. The algorithm's search boundaries are defined by constructing an open set and a closed set. The open set initially contains only the starting point, while the closed set is initially empty. For each node, values ​​g, h, and f are assigned. The g value represents the actual movement cost from the starting point to the current node, with the g value for the starting point set to 0. The h value uses Euclidean distance as a heuristic function. , and , and , respectively, are the x and y coordinates of the node on the regional planar map; f is the sum of g and h values.

[0080] The process begins with a loop search. The node with the smallest f-value from the open set is selected as the current node and moved from the open set to the closed set. If the current node is the destination, the search ends and the path is backtracked. Otherwise, the process iterates through its eight neighboring nodes in the surrounding eight directions. These nodes must be within a preset spatial range and not be obstacle nodes. For each neighboring node, the g-value (actual movement cost) of reaching that node from the starting point via the current node is calculated. To simplify the calculation, the map can be gridded, with the grid side length representing the actual distance (e.g., 0.1m). The straight-line movement cost (moving one grid) is 10, and the diagonal movement cost (moving one diagonal grid) is 14. The h-value is the Euclidean distance from the current node to the destination, i.e., the straight-line physical distance. If a neighboring node is not in the open set, the corresponding node is added to the open set, its parent node is recorded as the current node, and the g-value and f-value are updated. If the node is already in the open set and the new g-value is smaller, the parent node, g-value, and f-value are updated.

[0081] Repeat the loop until the endpoint is found. Backtrack from the endpoint to the starting point through the parent node. The resulting node sequence is the optimal temporary path. The starting point of the optimal temporary path is the real-time position of the inspection robot dog when a path conflict occurs, and the endpoint coincides with the initial inspection path.

[0082] In one embodiment, the preprocessed environmental image is input into a pre-trained target detection model, which identifies target obstacles in the preprocessed environmental image located in front of the inspection robot dog. The calculation of the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints includes the following steps:

[0083] The bounding boxes of target obstacles in the preprocessed environmental image are identified using an object detection model.

[0084] The first and second environmental images are processed into a disparity map of the target obstacle located in front of the inspection robot dog by calculating the correspondence between the first and second environmental images.

[0085] The three-dimensional position coordinates of the target obstacle relative to the inspection robot dog are calculated by combining the bounding box, disparity map, and camera parameters of the first and second cameras in the image acquisition module.

[0086] In this embodiment, the preprocessed environment image is processed by an object detection model to identify the bounding boxes of target obstacles. The object detection model also includes a scene recognition model based on the YOLOv5 framework. The training dataset contains 5000 images of dam scenes, including 10 types of obstacles such as rocks and vegetation. The Adam optimizer is used with an initial learning rate of 0.001, 100 epochs, and an IoU threshold of 0.5. The YOLOv5 model achieves an mAP of 92% on the validation set. First, the CSPDarknet53 backbone network in the scene recognition model extracts features from the preprocessed environmental image, generating feature maps at different scales. These feature maps contain information such as the edges, textures, and contours of objects in the image. Next, the PANet neck network in the scene recognition model fuses these multi-scale feature maps, enhancing the feature representation of obstacles of different sizes. Then, the head network in the scene recognition model generates a large number of candidate boxes based on the fused feature maps. Each candidate box contains the pixel coordinate range of the area where the obstacle may exist. A non-maximum suppression algorithm is used to filter the candidate boxes, removing those with overlap exceeding a preset threshold, and retaining the candidate box with the highest confidence as the bounding box of the target obstacle. The bounding box is represented by the top-left corner of the preprocessed environmental image as the origin, using the top-left corner pixel coordinates (x1, y1) and the bottom-right corner pixel coordinates (x2, y2), accurately defining the position and range of the target obstacle in the preprocessed environmental image.

[0087] The image acquisition module includes a first camera and a second camera, which are placed horizontally parallel to each other with a fixed distance between them equal to the baseline. The first and second cameras simultaneously acquire environmental images in front of the inspection robot dog; the image acquired by the first camera is the first environmental image, and the image acquired by the second camera is the second environmental image. The first and second environmental images undergo preprocessing, which is consistent with the preprocessing of the environmental images, including Gaussian filtering for noise reduction, histogram equalization, and scaling to 640×480 pixels. The ORB feature extraction algorithm is used to extract feature points from the preprocessed first and second environmental images and perform corresponding matching, obtaining 500-1000 pairs of matching points. Then, the fundamental matrix is ​​calculated based on the matching points to eliminate image distortion. Finally, the SGM semi-global matching algorithm is used to aggregate matching costs in eight directions, generating a disparity map with a resolution of 640×480 and a disparity range of 0-128 pixels.

[0088] The 3D position coordinates of the target obstacle relative to the inspection robot dog are calculated by combining the bounding box, disparity map, and camera parameters of the first and second cameras. Camera parameters include intrinsic and extrinsic parameters. Intrinsic parameters include focal length and principal point coordinates. The focal length is the distance from the optical center of the camera to the imaging plane, and the principal point coordinates are the pixel coordinates of the center of the imaging plane in the image. Extrinsic parameters include the baseline distance, i.e., the horizontal distance between the optical centers of the first and second cameras. Based on the top-left pixel coordinates (x1, y1) and bottom-right pixel coordinates (x2, y2) of the bounding box, the pixel region of the target obstacle in the first environmental image is determined. The disparity values ​​of all pixels within the corresponding pixel region are extracted from the disparity map, and the average disparity value is calculated as the effective disparity value of the target obstacle. Based on the pinhole camera model, the 3D position coordinates are calculated using the following formulas: the 3D X-axis value is the depth Z multiplied by (pixel region center x-coordinate minus principal point x-coordinate) divided by the focal length; the 3D Y-axis value is the depth Z multiplied by (pixel region center y-coordinate minus principal point y-coordinate) divided by the focal length; the 3D Z-axis value is the baseline distance multiplied by the focal length divided by the effective disparity value. The x-coordinate of the center of the pixel region is (x1+x2) / 2, and the y-coordinate of the center of the pixel region is (y1+y2) / 2. The calculated X, Y, and Z axis values ​​together constitute the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog.

[0089] In one embodiment, a dam leakage detection method based on a patrol robot dog further includes the following initial patrol path optimization step:

[0090] The preprocessed environmental image is input into the target detection model. The target detection model is trained on a dataset containing obstacles in the dam scene. The feature extraction network in the target detection model extracts feature maps of texture and contour features of obstacles in the preprocessed environmental image containing dam scene obstacles. Based on the feature maps, the target obstacles are classified and identified, and the obstacle category of the target obstacle is output.

[0091] The actual size data of the target obstacle is obtained by combining the camera parameters of the first and second cameras in the image acquisition module and the pixel size of the bounding box;

[0092] The three-dimensional position coordinates of the target obstacle, the obstacle category, the actual size data, and the regional planar map of the target inspection area are fused to construct an optimized regional planar map of the target inspection area. The optimized regional planar map is used to optimize the initial inspection path in the next inspection cycle.

[0093] In this embodiment, the preprocessed environmental image is input into a scene recognition model based on the YOLOv5 framework. The scene recognition model is trained on a dataset containing obstacles in a dam scene. The CSPDarknet53 feature extraction network in the scene recognition model is activated. The corresponding feature extraction network consists of 13 convolutional layers. The first 5 layers of the 13 convolutional layers are shallow convolutions, using sliding convolution operations with 3×3 convolutional kernels to extract low-level features such as edges and colors from the preprocessed environmental image. The last 8 layers of the 13 convolutional layers are deep convolutions, using 5×5 convolutional kernels combined with a residual connection structure to aggregate low-level features layer by layer, generating a feature map containing texture and contour features of obstacles in the dam scene. Texture features include the direction and density distribution of the texture on the obstacle surface, and contour features include the shape and corner morphology of the obstacle. The PANet neck network is used to perform multi-scale fusion of the feature map to enhance the expressive ability of obstacles of different sizes.

[0094] Based on the fused feature map, the SoftMax activation function is called through the classification branch of the head network in the scene recognition model. First, the fused feature map is converted into a fixed-dimensional feature vector through global average pooling. The dimension of the feature vector corresponds to the number of obstacle categories in the training set. The obstacle category feature library in the training set consists of standard feature vectors for each obstacle category. Each standard feature vector is obtained by performing the same global average pooling operation on the feature map of the corresponding category sample image. The extracted feature vector is then compared with each standard feature vector in the obstacle category feature library using cosine similarity calculation to obtain a similarity value between the feature vector and each standard feature vector. A higher similarity value indicates that the features of the two are more similar. All similarity values ​​are then input into the SoftMax activation function. The calculation process involves performing an exponential operation on each similarity value and dividing it by the sum of the exponential results of all similarity values ​​to obtain the probability value for each category. The sum of the probability values ​​is 1. The category with the highest output probability value is taken as the obstacle category of the target obstacle. The obstacle categories cover common obstacle types in dam scenes, such as rocks, tree trunks, puddles, cracks, and weeds.

[0095] By combining the parameters of the first and second cameras in the image acquisition module with the pixel dimensions of the obstacle's bounding box, the actual physical dimensions of the target obstacle are calculated. The camera parameters required for the calculation include the lens focal length and the pixel dimensions of the imaging chip; the required bounding box dimensions are obtained through image coordinate measurements, specifically the number of pixels in width and height of the target obstacle in the preprocessed environmental image.

[0096] First, the dimensions of the target obstacle on the camera's imaging chip are calculated, i.e., the imaging size. The imaging width is calculated by multiplying the width of the bounding box in pixels by the side length of a single pixel, and then dividing the product by one thousand to convert the unit from micrometers to millimeters. The imaging height is calculated in the same way. Next, depth information, i.e., the straight-line distance between the target obstacle and the camera, is extracted using the obtained 3D position coordinates of the target obstacle. Finally, the actual dimensions of the obstacle are calculated based on the principle of similar triangles. The actual width is calculated by multiplying the previously calculated imaging width by the depth information, dividing the product by the lens's focal length, and then multiplying by one thousand to align the units. The actual height is calculated in exactly the same way. Through this process, the actual width and actual height data of the target obstacle can be obtained.

[0097] Next, the 3D position coordinates, obstacle category, and calculated actual size data of the target obstacle are fused with the regional planar map of the target inspection area to construct an optimized regional planar map. The initial state of the regional planar map is a 2D grid map with a grid precision of 0.5 meters by 0.5 meters. Coordinate transformation converts the 3D position coordinates of the target obstacle into the grid coordinates of the regional planar map. The transformation method for the grid X coordinate is: subtract the X coordinate of the map origin from the 3D X coordinate of the target obstacle, and then divide the difference by 0.5. The transformation method for the grid Y coordinate is the same. After the coordinate transformation is completed, the obstacle category information is labeled in the corresponding grid, and the actual width and height of the obstacle are also labeled numerically. This fusion operation is repeated for all identified target obstacles, ultimately forming an optimized regional planar map containing accurate obstacle distribution, category, and size information.

[0098] The corresponding optimized area plan map will be used for the initial inspection path optimization in the next inspection cycle. At the start of the next inspection task, the optimized area plan map is read, and obstacle impact assessment is performed on each grid in the corresponding map. When the actual width or height of an obstacle in a grid exceeds 0.3 meters, this grid is marked as a high-impact area. When three or more consecutive high-impact areas appear, this area is determined to be a dense obstacle area. Finally, through A... The path planning algorithm adjusts the initial inspection path based on the evaluation results, so that the new path can avoid all high-impact areas and areas with dense obstacles, and prioritizes areas within the grid with no obstacles or obstacles with an actual size of less than 0.1 meters, thereby forming a more efficient and safer optimized inspection path.

[0099] In one embodiment, the soil moisture detection module is a near-infrared spectral sensor. The probe of the near-infrared spectral sensor is embedded in the footpads at the ends of the four legs of the inspection robot dog, with the probe end face 0.5mm-1mm higher than the bottom surface of the footpad, ensuring that a sealed detection cavity is formed upon contact with the ground and directly acquiring soil spectral signals. The process involves identifying the surface type of the area to be measured in front of the inspection robot dog through an image acquisition module, controlling the inspection robot dog to move to the area to be measured, acquiring the raw dataset of the ground surface in the area to be measured through the near-infrared spectral sensor, and sending the raw dataset to the edge computing module, including the following steps:

[0100] The image acquisition module inputs the environmental image of the area to be measured in front of the inspection robot dog into a surface classification model trained with typical surface samples of the dam scene. The surface classification model identifies the surface type of the environmental image, which includes soil, rock, vegetation and water.

[0101] If the surface type is identified as non-soil, the data collection for the corresponding area to be measured is abandoned.

[0102] When the surface type is identified as soil, the corresponding area to be measured is marked as the target area to be measured, and the inspection robot dog is guided to move to the target area to be measured. After the inspection robot dog moves to the target area to be measured, the ground of the target area to be measured is detected by the near-infrared spectral sensor.

[0103] Near-infrared light is emitted into the soil of the target area by a near-infrared spectral sensor, and the reflected light signal formed after the beam of near-infrared light is reflected by the soil is received by the photoelectric detection unit of the near-infrared spectral sensor.

[0104] The reflected light signal is converted into digital spectral data through a photoelectric detection unit;

[0105] The location coordinates of the inspection robot dog equipped with a near-infrared spectral sensor are recorded synchronously by the image acquisition module. The digital spectral data is then associated with the location coordinates and stored to form the original dataset.

[0106] The original dataset is sent to the edge computing module.

[0107] In this embodiment, the image acquisition module is activated. The module includes a high-definition industrial camera and an adaptive fill light. The camera has a resolution of 1920×1080, a frame rate of 30fps, and a field of view of 60 degrees. The fill light supports brightness adjustment from 300 to 1500 lumens. The image acquisition module captures environmental images of the area to be measured in front of the inspection robot dog and inputs these images into the surface classification model. The surface classification model can be a ResNet-50 deep learning model, trained on typical surface samples from dam scenarios. The training samples cover image data of common surface types such as soil, rock, vegetation, and water, with at least 5000 samples for each type.

[0108] The ResNet-50 model consists of 50 convolutional layers, each composed of multiple residual blocks. Each residual block uses skip connections to address the vanishing gradient problem during deep network training. When processing input environmental images, the ResNet-50 model first extracts features from the image through an initial convolutional layer, obtaining a primary feature map. Then, it processes the image layer by layer through multiple residual blocks, continuously extracting higher-level features. Finally, a global average pooling layer transforms the feature map into a fixed-dimensional feature vector. During feature extraction, the extracted features include texture, color, and morphological features. Texture features include the density and direction of the surface texture; color features include the numerical distribution of the RGB channels; and morphological features include the undulations and overall contour of the surface.

[0109] Next, the extracted feature vectors are analyzed through the fully connected layers of the model. The first hidden layer of the fully connected layer receives the feature vector output by the global average pooling layer, corresponding to a feature vector dimension of 2048. The first hidden layer contains 1024 neurons, each neuron establishing connections with all dimensions of the input feature vector. The calculation process involves matrix multiplication of the input feature vector with the weight matrix of this layer, followed by the addition of bias terms, resulting in a 1024-dimensional intermediate vector. The weight matrix has a dimension of 1024×2048, with each element being a pre-trained weight parameter. The bias term is a 1024-dimensional vector, with each element representing the bias parameter of the corresponding neuron. Then, the ReLU activation function is applied to the intermediate vector. This involves calculating the value of each element in the intermediate vector; if the element value is greater than 0, the original value is retained; if it is less than or equal to 0, 0 is output, resulting in the activated 1024-dimensional vector.

[0110] The output vector of the first hidden layer is fed into the second hidden layer, which contains 512 neurons and is also fully connected to the input 1024-dimensional vector. The calculation process involves matrix multiplication of the 1024-dimensional vector with the layer's weight matrix (dimension 512×1024), followed by the addition of a 512-dimensional bias term to obtain a 512-dimensional intermediate vector. This intermediate vector is then processed by the ReLU activation function, where each element retains its original value if it is greater than 0, and outputs 0 otherwise, resulting in a 512-dimensional feature processing vector.

[0111] The output vector of the second hidden layer is then fed into the output layer. The output layer contains four neurons, each corresponding to one of the four possible land surface types. In each layer, a matrix multiplication operation is first performed: a 4x512 weight matrix is ​​multiplied by the 512-dimensional input vector, and a 4-dimensional bias term is added to obtain a 4-dimensional original output vector. Then, the SoftMax activation function is applied to the original output vector to calculate the probability value for each category. Specifically, the calculation process is as follows: first, the exponent value corresponding to each element in the original output vector is calculated; then, these four exponent values ​​are summed to obtain a total; finally, the individual exponent value of each element is divided by this total to obtain the probability of the category corresponding to that element. This process ultimately generates a probability distribution vector consisting of four probability values, where each probability value corresponds to a land surface type.

[0112] Simultaneously, the system calls pre-stored feature templates for various land surface types. Each template is itself a standard feature vector for the corresponding type, with a dimension of 512, consistent with the output vector dimension of the second hidden layer. The system uses the 512-dimensional feature processing vector output from the second hidden layer as the feature vector to be detected, and calculates the Euclidean distance between it and each standard feature vector. The calculation steps for the Euclidean distance are as follows: First, calculate the difference between the element at each corresponding position of the feature vector to be detected and the standard feature vector; second, square each calculated difference; then, sum all these squared differences to obtain a final sum; finally, calculate the square root of this sum to obtain the Euclidean distance between the two. By comparing the Euclidean distances between the feature vector to be detected and the four standard feature vectors, the land surface type with the smallest distance value is the type with the highest matching degree. By combining the probability distribution vector output by SoftMax, when the category with the highest probability coincides with the category with the smallest Euclidean distance, the corresponding category is output by the land surface classification model as the land surface type corresponding to the environmental image. The land surface type includes soil, rock, vegetation and water.

[0113] When the surface type is identified as non-soil, the inspection robot's control system receives the identification result, issues a stop command, abandons the data collection work for the corresponding area to be measured, and switches to the next area to be measured.

[0114] When the surface type is identified as soil, the control system marks the corresponding area to be measured as the target area and generates a movement path. The movement path starts from the current position of the inspection robot dog and ends at the center of the target area to be measured. Path planning uses A... An algorithm ensures the inspection robot avoids obstacles along its path. The control system drives the robot's drive wheels to move along a planned path. During movement, the GPS positioning module obtains the robot's location in real time. When the deviation between the robot's location and the center of the target measurement area is less than 0.1m, the control system issues a stop command, and the inspection robot stops moving.

[0115] After the inspection robot dog moves to the target measurement area, the near-infrared spectral sensor is activated. The wavelength range of the near-infrared spectral sensor is set to 900-1700 nm, the spectral resolution is 10 nm, and the sampling interval is 1 nm. The near-infrared spectral sensor emits near-infrared light into the soil of the target measurement area, with the beam perpendicularly illuminating the soil surface and the illumination diameter being 5 cm. The photoelectric detection unit of the near-infrared spectral sensor receives the reflected light signal formed by the near-infrared light after reflection by the soil. The photoelectric detection unit uses an InGaAs detector with a response time of 10 μs and can convert the optical signal into an electrical signal.

[0116] The reflected light signal is converted into an electrical signal by the photoelectric detection unit and transmitted to the A / D converter. The A / D converter has a sampling accuracy of 16 bits and a conversion rate of 1MHz. It converts the electrical signal into digital spectral data, which contains reflectance values ​​at different wavelengths. Each wavelength corresponds to one reflectance data point, and the number of data points is 801 (covering the wavelength range of 900-1700nm).

[0117] While the near-infrared spectral sensor collects data, the image acquisition module simultaneously records the location coordinates of the inspection robot dog equipped with the near-infrared spectral sensor. The location coordinates are provided by the GPS positioning module, with a positioning accuracy of 1 meter, and include longitude and latitude information. The digital spectral data and location coordinates are associated and stored according to the acquisition timestamp to form the raw dataset. The raw dataset is in CSV format, and each row of data contains wavelength, reflectance, longitude, latitude, and acquisition time information.

[0118] Finally, the original dataset is sent to the edge computing module via the data transmission interface (using RS485 protocol, transmission rate of 115200bps) connected to the near-infrared spectral sensor, and the edge computing module receives and stores the original dataset.

[0119] In one implementation, the process of converting the original dataset into soil moisture content data using an edge computing module, and then cleaning the soil moisture content data to remove outliers to obtain valid soil moisture content data includes the following steps:

[0120] The original dataset is preprocessed in real time by the edge computing module, and the preprocessed original dataset is input into the pre-trained solution model to obtain soil moisture content data.

[0121] The isolated forest algorithm was used to construct a random decision tree to identify isolated points in the soil moisture content data that deviated from the normal distribution as potential outliers.

[0122] The LOF algorithm is used to calculate the local outlier factor of the soil moisture content data. When the value of the local outlier factor exceeds a preset threshold, the corresponding local outlier factor is marked as outlier data.

[0123] The intersection of potential outlier data and outlier data is taken as the outlier data, and the outlier data is removed to obtain the effective soil moisture content data.

[0124] In this embodiment, after receiving the raw dataset, the edge computing module performs real-time preprocessing on it. The raw dataset is in CSV format and contains wavelength, reflectance, longitude, latitude, and acquisition time information. The preprocessing process includes: first, detecting missing values ​​in the reflectance data; if missing values ​​exist, linear interpolation is used to fill them in, i.e., the reflectance value at the missing position is calculated by linear fitting based on the two valid reflectance data adjacent to the missing value; second, Savitzky-Golay filtering is used for smoothing, with the filter window size set to 11 and the polynomial order set to 2 to eliminate the interference of random noise on the spectral data; finally, the reflectance data is standardized, converting the reflectance value corresponding to each wavelength to the 0-1 range. The conversion formula is: standardized reflectance = (original reflectance - minimum reflectance) / (maximum reflectance - minimum reflectance), where the minimum and maximum reflectance are the minimum and maximum values ​​of all reflectance values ​​in the raw dataset.

[0125] The preprocessed raw dataset is input into a pre-trained solution model. The solution model employs a PLSR (Partial Least Squares Regression) model, trained on a training set containing soil sample reflectance data and corresponding measured moisture content data. During training, standardized reflectance corresponding to wavelengths of 900-1700 nm is used as the input feature, and measured soil moisture content is used as the output label. Cross-validation is used to determine the optimal number of principal components (set to 10), establishing a mapping relationship between reflectance and moisture content. During solution processing, the preprocessed reflectance data is input into the PLSR model, which calculates the corresponding soil moisture content data using the established mapping relationship, with data precision retained to two decimal places.

[0126] The Isolation Forest algorithm is used to construct random decision trees to identify isolated points in soil moisture content data that deviate from the normal distribution as potential outliers. The Isolation Forest algorithm is initialized with 100 trees, each with a sample size of 256. The process of constructing the random decision trees is as follows: A sample size is randomly selected from the soil moisture content data as a subsample; a feature (here, the moisture content value) and a corresponding random split point are randomly selected; based on the split point, the subsample is divided into two parts: samples smaller than the split point are assigned to the left subtree, and samples greater than or equal to the split point are assigned to the right subtree; this splitting process is repeated until each subtree contains only one sample or a preset tree depth (set to 8) is reached. After each tree is constructed, the average path length of each soil moisture content data point across all trees is calculated. The path length is the number of edges traversed by the sample from the root to the leaf node. The shorter the average path length, the greater the likelihood of the sample being isolated. When the average path length is less than 1 / 3 of the average path length of all samples, the corresponding sample is marked as potential outlier.

[0127] The Loophole Factor (LOF) algorithm was used to calculate the local outlier factor of soil moisture content data. In the LOF algorithm, the number of nearest neighbors, k, was set to 20, meaning that the k-neighborhood of each sample contained the 20 closest samples around it. The calculation process was as follows: First, the Euclidean distance of each soil moisture content data point to all other data points was calculated, and the top 20 samples were selected as the k-neighborhood based on their distances from smallest to largest. Next, the reachability distance from the sample to each sample in the k-neighborhood was calculated. The reachability distance was the maximum of the actual distance between the sample and its neighboring samples and the k-distance (the distance from the neighboring sample to its k-th nearest neighbor). The local reachability density of the sample was calculated using the reachability distance, which is the reciprocal of the average reachability distance of all samples in the k-neighborhood. Finally, the local outlier factor was calculated, which is the average ratio of the local reachability density of all samples in the k-neighborhood to the local reachability density of the corresponding sample. A preset threshold of 1.5 was set; when the local outlier factor exceeded 1.5, the corresponding soil moisture content data was marked as outlier data.

[0128] To further improve the accuracy of data cleaning, one implementation method can also integrate the robot dog's own motion state information to clean abnormal soil moisture data. Specifically, while acquiring each frame of spectral data through a near-infrared spectral sensor, the robot dog's built-in attitude perception unit, including an IMU (Inertial Measurement Unit) and leg joint encoders, simultaneously acquires motion attitude data that strictly corresponds to the timestamp of the corresponding spectral data. The corresponding motion attitude data includes three-axis angular velocity, three-axis linear acceleration, and foot contact force estimated by the controller. Based on the motion attitude data synchronously acquired by the edge computing module, a real-time MSS (Motion Stability Score) is calculated for each calculated soil moisture data point. The MSS score is obtained by normalizing and weighting the instability indicators such as angular velocity modulus, linear acceleration modulus, and foot contact force, which can quantify the instantaneous motion stability of the robot dog. The motion stability value ranges from 0 to 1, with a higher score indicating more unstable motion of the robot dog.

[0129] Specifically, the edge computing module calculates the corresponding MSS motion stability score for each soil moisture content data point as follows:

[0130] Feature extraction is performed. For the three-axis angular velocity vector and three-axis linear acceleration vector provided by the IMU (Inertial Measurement Unit), the system calculates their Euclidean norms, i.e., moduli. The moduli are calculated by squaring the components of the vector along the x, y, and z axes, summing the three squared values, and then taking the square root of the sum. This result is used as a feature quantity characterizing the intensity of the motion. The foot contact force estimated by the controller is directly used as a feature quantity because it is a single numerical value (scalar).

[0131] The extracted features are normalized to map their values ​​to a dimensionless range of 0 to 1 for subsequent fusion calculations. For the angular velocity and linear acceleration moduli, the normalization method is as follows: the current value is divided by a preset maximum possible value based on the robot dog's kinematic characteristics. To ensure stability, if the calculated ratio exceeds 1, it is set to 1 as the final result. For the foot contact force, a minimum-maximum scaling method is used for normalization: a preset minimum contact force value is subtracted from the current contact force value, and the difference is divided by the preset difference between the maximum and minimum contact forces. Similarly, to address potential out-of-range readings from the sensors, the calculation result is strictly limited to between 0 and 1.

[0132] The normalized features are weighted and fused to calculate the final motion stability score. A key step is to conceptually reverse the influence of foot contact force, as a larger contact force represents stability, while larger angular velocities and accelerations represent instability. This reversal is achieved by subtracting the normalized foot contact force value from 1. The final calculation formula is described as follows: multiply the normalized angular velocity magnitude, the normalized linear acceleration magnitude, and the reversed contact force value by their respective preset weighting coefficients, and then add these three products together; the sum is the final motion stability score. Using this method, when the inspection robot dog is in a stationary standing state, the score will approach 0, representing a very stable motion state; conversely, during vigorous movement, the score will approach 1, representing a very unstable motion state. The motion stability score is then used to dynamically adjust the algorithm thresholds of subsequent data cleaning processes.

[0133] In the subsequent data cleaning process, the sensitivity of the anomaly detection algorithm can be dynamically adjusted using motion stability scores. When using the Isolation Forest algorithm to construct a random decision tree to identify potential anomalies, the original anomaly scores for each data point are corrected based on the motion stability scores corresponding to their original anomaly scores. This means that data collected in unstable states is assigned higher anomaly weights, making it easier to label them as potential anomalies. Similarly, when calculating the local outlier factor for each data point using the LOF algorithm, the preset threshold used to determine whether a data point is an outlier is no longer a fixed value but is dynamically adjusted based on the motion stability score of the corresponding data point. Specifically, the higher the motion stability score, the lower the corresponding outlier threshold. This means that data points collected when the robot is moving violently, even if their deviation from their neighborhood is not significant, are more likely to be identified as outliers.

[0134] Finally, the Isolation Forest algorithm and the LOF algorithm, which incorporate a dynamic adjustment mechanism based on motion stability scoring, were used to identify potential outliers and isolated data, respectively. The intersection of the corresponding potential outliers and isolated data was then used by the edge calculation module to obtain the final outlier data, which was removed from the calculated soil moisture content dataset. The resulting effective soil moisture content data not only eliminated statistical outliers but also filtered out measurement noise introduced by the physical movement of the inspection robot platform itself. This significantly improved the reliability and confidence of the data in the real, complex dam environment, providing a more solid data foundation for subsequent soil moisture analysis and early warning.

[0135] In one embodiment, the method further includes an environmental self-calibration step for solving the model:

[0136] The inspection robot dog is guided to one or more preset reference areas via a remote control terminal. The reference area includes at least one dry standard soil area and at least one saturated standard soil area, and the soil type of the reference area is consistent with the soil type of the target inspection area.

[0137] The reference spectral data of the soil in the reference area is collected by a near-infrared spectral sensor. Based on the reference spectral data, the reference soil moisture content data in the reference area is calculated by an edge computing module, and a mapping relationship between the reference spectral data and the reference soil moisture content data is established.

[0138] Based on the mapping relationship, a mapping relationship model for the target inspection area is generated, and the environmental self-calibration of the solution model is completed through the mapping relationship model.

[0139] In this embodiment, before the formal inspection, a command is issued via a remote control terminal to guide the inspection robot dog to one or more preset benchmark areas. The benchmark areas are selected in advance through on-site surveys and include at least one dry standard soil area and at least one saturated standard soil area. The soil in the dry standard soil area has been naturally air-dried, with a soil moisture content below 5%; the soil in the saturated standard soil area has been thoroughly soaked and then drained of surface moisture, with a soil moisture content above 30%. The soil type in the benchmark areas is consistent with the soil type in the target inspection area, ensuring that the basic properties such as soil particle composition and organic matter content are the same.

[0140] After the inspection robot dog reaches the reference area, it activates the near-infrared spectral sensor. The near-infrared spectral sensor has a wavelength range of 900-1700 nm, a spectral resolution of 10 nm, and a sampling interval of 1 nm. In the dry standard soil area, the near-infrared spectral sensor performs multi-point detection on the soil surface, with no fewer than 5 detection points. Spectral data is collected 3 times at each detection point, and the average value is used as the reference spectral data for that point. Similarly, the same operation is performed in the saturated standard soil area to obtain the reference spectral data for that area. The reference spectral data includes reflectance values ​​at different wavelengths and is stored in CSV format, with each row containing wavelength, reflectance, and sampling location information.

[0141] The baseline spectral data is sent to the edge computing module, which preprocesses the data. The preprocessing steps include missing value detection and linear interpolation, Savitzky-Golay filtering and smoothing (window size 11, polynomial order 2), and standardization (converting reflectance to the 0-1 range). The preprocessed baseline spectral data is then input into a temporary computational model in the edge computing module. This model uses the same PLSR partial least squares regression structure as the solution model to calculate the baseline soil moisture content in the baseline region. During the calculation, the output of the temporary computational model is corrected by considering the known moisture content ranges of the dry and saturated standard soil regions to ensure that the baseline soil moisture content in the dry standard soil region falls within the 3%-5% range and the baseline soil moisture content in the saturated standard soil region falls within the 30%-35% range.

[0142] A mapping relationship was established between baseline spectral data and baseline soil moisture content data. The process involved using preprocessed baseline spectral data as input and baseline soil moisture content data as output, and fitting the functional relationship between the input and output using the least squares method. During the fitting process, the correlation coefficient between reflectance and moisture content for each wavelength was calculated. Wavelengths with an absolute correlation coefficient greater than 0.7 were selected as characteristic wavelengths, and their corresponding reflectance data were retained for mapping relationship construction. Multiple linear regression analysis was performed on the reflectance data of the characteristic wavelengths and the moisture content data to obtain the mapping relationship expression, which is in the form of: Moisture Content = a1×R1 + a2×R2 + ... + an×Rn + b, where R1 to Rn are the reflectance of the characteristic wavelengths, a1 to an are the regression coefficients, and b is a constant term.

[0143] A mapping relationship model for the target inspection area is generated based on the mapping relationship. This model includes parameters such as the selection criteria for characteristic wavelengths, regression coefficients, and constant terms. The regression coefficients (a1-an) and constant term (b) from the mapping relationship model are imported into the original PLSR solution model, replacing the corresponding parameters in the mapping formula between spectral reflectance and water content in the PLSR model. After calibration, the error rate of the solution model is reduced to within ±2%, completing the environmental self-calibration of the solution model. The calibrated solution model can better adapt to the soil environmental characteristics of the target inspection area, improving the accuracy of soil moisture content data measurement.

[0144] In one embodiment, comparing the effective soil moisture content data with a preset soil moisture content safety threshold, and transmitting the effective soil moisture content data exceeding the soil moisture content safety threshold in the form of an early warning to the inspection personnel in the target inspection area via a communication module includes the following steps:

[0145] Based on the soil type in the target inspection area, a safe threshold for soil moisture content is set, and the safe threshold for soil moisture content is pre-stored through the edge computing module.

[0146] Effective soil moisture content data are divided into zones according to the location coordinates of the target inspection area;

[0147] For any of the aforementioned zones, the effective soil moisture content data is compared with the safe soil moisture content threshold. When the effective soil moisture content data exceeds the safe soil moisture content threshold, the ratio of the amount of data exceeding the threshold within the zone to the total amount of data is calculated, where the total amount of data is the total number of samples of effective soil moisture content data within the zone. Based on the ratio, anomaly levels are classified, and the corresponding zones are marked as warning zones, and the location coordinates and measurement time of the warning zones are associated.

[0148] The soil moisture content data, location coordinates, measurement time, and anomaly level of the warning zone are packaged into a standardized warning frame. The standardized warning frame is sent to the inspection terminal held by the inspection personnel in the target inspection area through the communication module. After receiving the standardized warning frame, the inspection terminal triggers an audio-visual prompt and marks the location of the warning zone on the environmental map displayed on the inspection terminal.

[0149] In this embodiment, a safe threshold for soil moisture content is set based on the soil type in the target inspection area. Different soil types correspond to different safe threshold standards. The safe threshold for sandy soil is set to 25%, and the safe threshold for loamy soil is set to 30%. The corresponding safe threshold for soil moisture content is pre-stored in the storage unit of the edge computing module. The storage format is key-value pairs, where the key is the soil type name and the value is the corresponding safe threshold value. The edge computing module can quickly retrieve the corresponding safe threshold for different soil types.

[0150] The effective soil moisture content data is partitioned according to the location coordinates of the target inspection area. The location coordinates of the target inspection area are within a preset longitude and latitude range, and the corresponding area is divided into 10m×10m square grids as the partition unit. Each partition is assigned a unique partition number, with the numbering rule being to increase sequentially from west to east for longitude and from south to north for latitude. All effective soil moisture content data are traversed, and based on the location coordinates (longitude and latitude) contained in each data point, the partition to which the corresponding data belongs is determined, and the data is assigned to the dataset of the corresponding partition. The dataset of each partition is stored separately, containing all effective soil moisture content data within the corresponding partition and their associated measurement time.

[0151] For any given partition, the edge computing module calls the corresponding soil moisture content safety threshold for the soil type within that partition, comparing each valid soil moisture content data point within the partition with the safety threshold. The number of valid soil moisture content data points exceeding the safety threshold within the partition is counted, and the proportion of this number to the total number of data points within the partition is calculated. Anomalies are classified based on this proportion: a proportion between 10% and 30% is classified as Level 1; a proportion between 30% and 50% is classified as Level 2; and a proportion exceeding 50% is classified as Level 3. When valid soil moisture content data within a partition exceeds the safety threshold, the corresponding partition is marked as a warning partition. Simultaneously, the location coordinates of the warning partition (the longitude and latitude of the partition center) are associated with the measurement time (the earliest and latest measurement times for all data within the partition).

[0152] Soil moisture content data, location coordinates, measurement time, and anomaly level for each warning zone are packaged into a standardized warning frame. The standardized warning frame adopts JSON format, and the data structure includes seven fields: "Zone Number", "Center Longitude", "Center Latitude", "Measurement Start Time", "Measurement End Time", "Anomaly Level", and "List of Exceeding Data". The "List of Exceeding Data" contains all valid soil moisture content data exceeding the safe threshold for soil moisture content within the corresponding zone, along with the corresponding specific measurement time and location coordinates.

[0153] Data transmission is conducted via a 4G wireless network at a rate of 1Mbps. Standardized early warning frames are acquired from the edge computing module via the communication module and then wirelessly transmitted to the inspection terminals held by personnel in the target inspection area. Upon receiving the standardized early warning frames, the built-in receiving module of the inspection terminal triggers its audio-visual alert function: the speaker emits a continuous "beep" sound for 10 seconds, and the display backlight flashes three times. Simultaneously, on the environmental map displayed on the inspection terminal, red polygonal areas are marked according to the location coordinates (center longitude and latitude) of the early warning zone. These polygonal areas cover the entire early warning zone. Different anomaly levels correspond to different marking styles: Level 1 anomalies are marked with a yellow border and red fill, Level 2 anomalies with an orange border and red fill, and Level 3 anomalies with a red border and red fill. The anomaly level and measurement time range are displayed next to the markings.

[0154] The present invention also discloses a dam leakage detection system based on an inspection robot dog, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the dam leakage detection method based on the inspection robot dog as described in any of the above embodiments.

[0155] The computer program can be stored in a machine-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The machine-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the machine-readable medium includes, but is not limited to, the above-mentioned components.

[0156] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0157] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of internal storage units and external storage devices of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0158] The present invention also discloses a computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform a dam leakage detection method based on an inspection robot dog as described in any of the above embodiments.

[0159] The above-described method for detecting dam leakage based on a patrol robot dog is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0160] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0161] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for detecting dam seepage based on a patrol robot dog, characterized in that, An inspection robot dog is used in this application. The inspection robot dog includes an image acquisition module, a soil moisture detection module, an edge computing module, and a communication module. The soil moisture detection module is respectively located at the ends of the four legs of the inspection robot dog. The image acquisition module and the soil moisture detection module are both connected to the edge computing module, and the edge computing module is connected to the communication module. The method includes the following steps: Obtain a regional planar map of the target inspection area, and generate an initial inspection path for the inspection robot dog based on the regional planar map and a path planning algorithm. During the inspection process of the inspection robot dog following the initial inspection path, the image acquisition module acquires environmental images in front of the inspection robot dog in real time, identifies target obstacles in front of the inspection robot dog based on the environmental images, and realizes dynamic obstacle avoidance of the inspection robot dog for the target obstacles. The image acquisition module and the soil moisture detection module work together. The image acquisition module identifies the surface type of the area to be measured in front of the inspection robot dog, controls the inspection robot dog to move to the area to be measured, and collects the raw dataset of the ground in the area to be measured through the soil moisture detection module, and sends the raw dataset to the edge computing module. The edge computing module decomposes the original dataset into soil moisture content data and performs abnormal data cleaning on the soil moisture content data to obtain effective soil moisture content data. The effective soil moisture content data is compared with the preset soil moisture content safety threshold, and the effective soil moisture content data exceeding the soil moisture content safety threshold is transmitted to the inspection personnel in the target inspection area in the form of an early warning through the communication module.

2. The dam leakage detection method based on an inspection robot dog according to claim 1, characterized in that, The image acquisition module acquires real-time environmental images in front of the inspection robot dog, identifies target obstacles in front of the robot dog based on these images, and performs dynamic obstacle avoidance for the robot dog against these target obstacles, including the following steps: The image acquisition module continuously acquires environmental images in front of the inspection robot dog, and performs real-time preprocessing on the environmental images to obtain preprocessed environmental images. The preprocessed environmental image is input into a pre-trained target detection model, which identifies the target obstacle in front of the inspection robot dog in the preprocessed environmental image and calculates the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints. The target obstacle is projected onto the area plane map where the initial inspection path is located by combining the multi-view geometric constraints and the three-dimensional position coordinates. The shortest distance between the projection of the target obstacle and the initial inspection path is calculated in the area plane map. If the shortest distance is less than or equal to a preset distance safety threshold, a path conflict is determined to have occurred. When the path conflict occurs, the projection center point of the target obstacle that caused the path conflict is taken as the path conflict point. Based on the preset spatial range of the inspection robot dog's movement radius, the path planning algorithm is invoked to replan the optimal temporary path for bypassing the target obstacle within the preset spatial range and with the path conflict point as the center. The starting point of the optimal temporary path is the real-time position of the inspection robot dog when the path conflict occurs, and the ending point of the optimal temporary path coincides with the initial inspection path.

3. The dam seepage detection method based on an inspection robot dog according to claim 2, characterized in that, The image acquisition module includes a first camera and a second camera, which are respectively positioned at the left and right eyes of the inspection robot dog's head. The preprocessed environmental image includes a first environmental image acquired by the first camera and a second environmental image acquired by the second camera. The multi-view geometric constraints include the bounding box of the target obstacle in the preprocessed environmental image and the disparity map of the target obstacle. The step of inputting the preprocessed environment image into a pre-trained target detection model, identifying target obstacles in front of the inspection robot dog in the preprocessed environment image through the target detection model, and calculating the three-dimensional position coordinates of the target obstacle relative to the inspection robot dog based on multi-view geometric constraints includes the following steps: The target detection model identifies the bounding boxes of target obstacles in the preprocessed environmental image; The parallax map of the target obstacle located in front of the inspection robot dog is calculated by processing the first environmental image and the second environmental image into a target obstacle based on the correspondence between the first environmental image and the second environmental image. The three-dimensional position coordinates of the target obstacle relative to the inspection robot dog are calculated by combining the bounding box, the disparity map, and the camera parameters of the first and second cameras in the image acquisition module.

4. The dam leakage detection method based on an inspection robot dog according to claim 2 or 3, characterized in that, The method also includes the following optimization steps for the initial inspection path: The preprocessed environmental image is input into the target detection model. The target detection model is trained on a dataset containing obstacles in a dam scene. The feature extraction network in the target detection model extracts feature maps of the texture and contour features of the obstacles in the preprocessed environmental image. Based on the feature maps, the target obstacles are classified and identified, and the obstacle category of the target obstacles is output. The actual size data of the target obstacle is obtained by combining the camera parameters of the first camera and the second camera in the image acquisition module with the pixel size of the bounding box; The three-dimensional position coordinates of the target obstacle, the obstacle category, the actual size data, and the regional planar map of the target inspection area are fused to construct an optimized regional planar map of the target inspection area. The optimized regional planar map is used to optimize the initial inspection path in the next inspection cycle.

5. The dam leakage detection method based on an inspection robot dog according to claim 1, characterized in that, The soil moisture detection module is a near-infrared spectral sensor. The steps of identifying the surface type of the area to be measured in front of the inspection robot dog through the image acquisition module, controlling the inspection robot dog to move to the area to be measured, acquiring the raw dataset of the ground in the area to be measured through the near-infrared spectral sensor, and sending the raw dataset to the edge computing module include the following steps: The image acquisition module inputs the environmental image of the area to be measured in front of the inspection robot dog into a surface classification model trained with typical surface samples of the dam scene. The surface classification model identifies the surface type of the environmental image, which includes soil, rock, vegetation and water. When the surface type is identified as non-soil, the data collection for the corresponding area to be measured is abandoned. When the surface type is identified as soil, the corresponding area to be measured is marked as the target area to be measured, and the inspection robot dog is guided to move to the target area to be measured. After the inspection robot dog moves to the target area to be measured, the ground of the target area to be measured is detected by the near-infrared spectral sensor. The near-infrared spectral sensor emits near-infrared light into the soil of the target area to be measured, and the photoelectric detection unit of the near-infrared spectral sensor receives the reflected light signal formed after the beam of near-infrared light is reflected by the soil. The reflected light signal is converted into digital spectral data by the photoelectric detection unit. The image acquisition module synchronously records the position coordinates of the inspection robot dog equipped with the near-infrared spectral sensor, and stores the digital spectral data in association with the position coordinates to form the original dataset. The original dataset is sent to the edge computing module.

6. The dam leakage detection method based on an inspection robot dog according to claim 1, characterized in that, The edge computing module processes the original dataset into soil moisture content data and performs anomaly cleaning on the soil moisture content data to obtain valid soil moisture content data, including the following steps: The edge computing module performs real-time preprocessing on the original dataset, and then inputs the preprocessed original dataset into the pre-trained solution model to obtain soil moisture content data. The isolated forest algorithm was used to construct a random decision tree to identify isolated points in the soil moisture content data that deviated from the normal distribution as potential outliers. The LOF algorithm is used to calculate the local outlier factor of the soil moisture content data. When the value of the local outlier factor exceeds a preset threshold, the corresponding local outlier factor is marked as outlier data. The intersection of the potential abnormal data and the outlier data is taken as the abnormal data, and the abnormal data is removed to obtain the effective soil moisture content data.

7. The dam leakage detection method based on an inspection robot dog according to claim 6, characterized in that, The method also includes an environmental self-calibration step for solving the model: The inspection robot dog is guided to one or more preset reference areas via a remote control terminal. The reference area includes at least one dry standard soil area and at least one saturated standard soil area, and the soil type of the reference area is consistent with the soil type of the target inspection area. The reference spectral data of the soil in the reference area is collected by a near-infrared spectral sensor. Based on the reference spectral data, the reference soil moisture content data in the reference area is calculated by the edge computing module, and a mapping relationship between the reference spectral data and the reference soil moisture content data is established. Based on the mapping relationship, a mapping relationship model for the target inspection area is generated, and the environmental self-calibration of the solution model is completed through the mapping relationship model.

8. The dam leakage detection method based on an inspection robot dog according to claim 1, characterized in that, The process of comparing the effective soil moisture content data with a preset safe soil moisture content threshold, and then transmitting the effective soil moisture content data exceeding the safe soil moisture content threshold to the inspection personnel in the target inspection area via the communication module in the form of an early warning, includes the following steps: The soil moisture content safety threshold is set based on the soil type in the target inspection area, and the soil moisture content safety threshold is pre-stored through the edge computing module; Effective soil moisture content data are divided into zones according to the location coordinates of the target inspection area; For any of the aforementioned partitions, the effective soil moisture content data is compared with the soil moisture content safety threshold. When the effective soil moisture content data exceeds the soil moisture content safety threshold, the abnormality level is classified based on the proportion of the effective soil moisture content data exceeding the soil moisture content safety threshold, and the corresponding partition is marked as a warning partition and associated with the location coordinates and measurement time of the warning partition. The soil moisture content data, location coordinates, measurement time, and anomaly level of the warning zone are packaged into a standardized warning frame. The standardized warning frame is sent to the inspection terminal held by the inspection personnel in the target inspection area through the communication module. After receiving the standardized warning frame, the inspection terminal triggers an audio-visual prompt and marks the location of the warning zone on the environmental map displayed on the inspection terminal.

9. A dam seepage detection system based on an inspection robot dog, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the dam leakage detection method based on the inspection robot dog as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the dam leakage detection method based on an inspection robot dog according to any one of claims 1 to 8.

Images