Tunnel leakage water disease automatic identification method, device and equipment and medium thereof

Abstract

The invention relates to a tunnel leakage water disease automatic identification method, device and equipment and a medium thereof. The method comprises the following steps: acquiring multi-source original data in a tunnel; preprocessing the multi-source original data to obtain multi-sensor data; the multi-source original data comprises visible light images, thermal imaging and point cloud data; dividing a tunnel section along the center line of the tunnel based on the multi-sensor data to obtain a tunnel section sequence; the tunnel section sequence comprises each tunnel section and the corresponding multi-sensor data; based on the tunnel section sequence, performing binary segmentation on the tunnel section to obtain a section binary segmentation mask; and based on the section binary segmentation mask, identifying the leakage water disease type of each tunnel section, and obtaining a disease classification result. By adopting the method, a unified section reference coordinate system can be established to track a disease development track, and further disease development can be accurately warned.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering technology, and in particular relates to an automatic identification method, device, equipment and medium for tunnel water leakage defects. Background Technology

[0002] With the continuous development of tunnel engineering inspection technology, intelligent monitoring methods based on multi-sensor fusion have emerged. This technology features comprehensive data collection and high analytical accuracy, which in turn has spurred the development of automated identification and assessment methods for modern tunnel seepage defects. Traditional techniques primarily rely on manual inspections combined with single sensors for defect detection. Inspectors rely on visual observation and experience to judge, using simple tools to mark the location of leaks and estimate their severity, then compiling the results into paper reports for subsequent analysis. However, this traditional approach has significant limitations: the subjectivity of manual judgment leads to poor consistency in assessment results; the long inspection cycle makes it difficult to detect defect development in a timely manner; the limited dimensions of data from a single inspection cannot fully reflect the characteristics of the defect, and there is a lack of effective time-series data analysis methods, making it difficult to accurately predict and warn of defect development; data collected at different times are difficult to correspond precisely in space, resulting in inaccurate tracking of defect development trajectories, and the inability to establish a unified cross-sectional reference coordinate system. Summary of the Invention

[0003] Therefore, it is necessary to provide methods, devices, equipment, and media for automatic identification of tunnel seepage defects that can accurately predict and warn of defect development, track defect development trajectory, and establish a unified cross-sectional reference coordinate system, in response to the above-mentioned technical problems.

[0004] Firstly, this application provides an automatic identification method for tunnel water leakage defects, including:

[0005] Acquire multi-source raw data from the tunnel; and preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging, and point cloud data;

[0006] Based on multi-sensor data, tunnel sections are divided along the centerline of the tunnel to obtain a tunnel section sequence; the tunnel section sequence includes each tunnel section and the corresponding multi-sensor data.

[0007] Based on the tunnel cross-section sequence, the tunnel cross-section is binary segmented to obtain the cross-section binary segmentation mask;

[0008] Based on the cross-sectional binary segmentation mask, the type of water leakage disease in each tunnel cross section is identified, and the disease classification results are obtained.

[0009] Furthermore, based on multi-sensor data, the tunnel cross-section is divided along the tunnel's centerline to obtain a tunnel cross-section sequence, including:

[0010] Based on point cloud data from multi-sensor data, the three-dimensional centerline of the tunnel is extracted to obtain the tunnel centerline;

[0011] Based on a preset interval, equidistant cross-sectional positions are generated on the tunnel centerline; and the central tangent direction is calculated as the cross-sectional normal vector based on the equidistant cross-sectional positions to obtain a preliminary cross-sectional position sequence.

[0012] Based on the preliminary cross-section position sequence, the cross-section position and the real structural features are aligned according to the visible light image and thermal imaging to obtain the accurate cross-section position sequence;

[0013] Based on the location of each cross section in the precise cross section location sequence, data slices corresponding to each cross section are extracted from multi-sensor data to obtain the tunnel cross section sequence.

[0014] Furthermore, based on the tunnel cross-section sequence, binary segmentation is performed on the tunnel cross-sections to obtain a cross-section binary segmentation mask, including:

[0015] Based on data slices of visible light images, a semantic segmentation network is used to identify wet areas and seepage textures, and region segmentation is performed based on the wet areas and seepage textures to obtain visible light segmentation results.

[0016] Based on thermal imaging data slicing, abnormally low temperature regions with temperatures below a preset temperature threshold are identified, and temperature segmentation results are obtained.

[0017] Based on point cloud data slicing, surface wet areas are detected according to point cloud reflection intensity information, and geometrically abnormal areas are identified to obtain point cloud segmentation results;

[0018] By integrating the visible light segmentation results, temperature segmentation results, and point cloud segmentation results, the segmentation results of the water seepage area are obtained.

[0019] Morphological optimization is performed on the segmentation results of the leakage area to obtain the cross-sectional binary segmentation mask.

[0020] Furthermore, by integrating the visible light segmentation results, temperature segmentation results, and point cloud segmentation results, the segmentation results of the seepage area are obtained, including:

[0021] Spatial alignment is performed on the visible light segmentation results, temperature segmentation results, and point cloud segmentation results to obtain a multimodal segmentation result matrix;

[0022] Calculate the confidence score of the multimodal segmentation result matrix to obtain the confidence score for each modality.

[0023] Based on prior knowledge of modalities under different environmental conditions, a weight coefficient is assigned to each modality; and based on the confidence score and the weight coefficient, the multimodal confidence distribution of each pixel is calculated to obtain the multimodal segmentation result.

[0024] Based on the multimodal segmentation results, classification is performed according to the classification threshold to obtain preliminary fusion segmentation results;

[0025] Based on the area and shape characteristics of the leakage area from the preliminary fusion and segmentation results, the abnormal segmentation results are filtered out to obtain the leakage area segmentation results.

[0026] Furthermore, based on the cross-sectional binary segmentation mask, the type of water leakage defect in each tunnel cross-section is identified, and the defect classification results are obtained, including:

[0027] Based on the cross-sectional binary segmentation mask, the features of each seepage area are extracted to obtain a multi-dimensional feature vector; the features include at least one of geometric features, texture features, temperature features, morphological features and distribution features;

[0028] The importance of the multidimensional feature vectors is evaluated, and the subset of features that contributes the most to the classification task is selected to obtain the region feature vectors.

[0029] The region feature vectors are input into a pre-trained attention classification model and a pre-trained machine learning classifier, respectively, to obtain multiple preliminary classification results.

[0030] A weighted vote is performed on all preliminary classification results to obtain the integrated leakage type.

[0031] Based on domain knowledge, consistency verification is performed on the integrated leakage types to obtain the disease classification results.

[0032] Furthermore, based on the cross-sectional binary segmentation mask, the type of water leakage defect is identified for each tunnel cross-section. After obtaining the defect classification results, the following steps are also included:

[0033] Based on the disease classification results, the severity of the disease is assessed to obtain a disease assessment report for the corresponding tunnel section; and all disease assessment reports are integrated to obtain a tunnel assessment report.

[0034] Based on multiple tunnel assessment reports, the evolution pattern of defects was analyzed, and time series data of defect development were obtained;

[0035] Based on the time series data of disease development, the development trend of disease is predicted, and a tunnel disease development trend report is obtained.

[0036] Based on the disease development trend report, the urgency of maintenance is determined, and a predictive maintenance report decision is obtained; the predictive maintenance report decision is used to assist in the maintenance decision of the tunnel.

[0037] Furthermore, based on multiple tunnel assessment reports, the evolution patterns of the defects were analyzed, and time series data on defect development were obtained, including:

[0038] Based on the tunnel assessment report, cross-sectional data from different periods are matched to obtain the cross-sectional matching structure;

[0039] Based on the cross-section matching results, the disease development rate is calculated using the following formula to obtain the cross-section dataset:

[0040]

[0041] in, For the rate of development, , These are the weighting coefficients. Let t be the area of ​​disease at the current time t. This represents the area of ​​disease at the previous time t−1. Score the severity of the disease at the current time t. The severity score of the disease at the previous time t−1. The time interval between adjacent detections;

[0042] Based on the cross-sectional dataset, the development patterns of diseases are classified using a clustering algorithm to obtain the disease evolution classification results;

[0043] Missing values ​​were filled in the disease evolution classification results to obtain disease development time series data.

[0044] Secondly, this application also provides an automatic identification device for tunnel water leakage defects, comprising:

[0045] The preprocessing module is used to acquire multi-source raw data in the tunnel and preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging and point cloud data;

[0046] The cross-section module is used to divide the tunnel cross-section along the centerline of the tunnel based on multi-sensor data, and obtain the tunnel cross-section sequence; the tunnel cross-section sequence includes each tunnel cross-section and the corresponding multi-sensor data.

[0047] The leakage module is used to perform binary segmentation of the tunnel cross-section based on the tunnel cross-section sequence to obtain the cross-section binary segmentation mask;

[0048] The disease module is used to identify the type of water leakage disease in each tunnel section based on the cross-sectional binary segmentation mask, and obtain the disease classification results.

[0049] Thirdly, this application also provides a computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any step of the method provided in the first aspect of this application.

[0050] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.

[0051] The aforementioned method, device, equipment, and medium for automatic identification of tunnel seepage defects acquire multi-source raw data from the tunnel; preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging, and point cloud data; based on the multi-sensor data, tunnel sections are divided along the tunnel's centerline to obtain a tunnel section sequence; the tunnel section sequence includes each tunnel section and its corresponding multi-sensor data; based on the tunnel section sequence, binary segmentation is performed on the tunnel sections to obtain a section binary segmentation mask; based on the section binary segmentation mask, the type of seepage defect for each tunnel section is identified to obtain the defect classification result. By accurately mapping data collected at different periods to spatially corresponding sections, the development trajectory of defects can be accurately tracked, a unified section reference coordinate system can be established, and the comparability and trend analysis value of long-term monitoring data can be improved. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of the process of an automatic identification method for tunnel water leakage defects provided in an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the structure of an automatic identification device for tunnel water leakage defects provided in an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] In one embodiment, such as Figure 1 As shown, an automatic identification method for tunnel water leakage defects is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and furthermore, to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0057] Step 101: Acquire multi-source raw data from the tunnel; and preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging, and point cloud data.

[0058] The multi-source raw data refers to unprocessed data about the tunnel interior directly collected from sensors based on different physical principles. It contains at least three types of information: visible light images, captured by ordinary cameras, reflecting visual information such as color, texture, and stains on the tunnel surface; thermal images, captured by infrared thermal imagers, recording the temperature distribution on the tunnel surface and representing temperature differences in image form; and point cloud data, acquired by laser scanners, which is a dataset consisting of millions of three-dimensional coordinate points. Each point contains its precise spatial location and records the intensity of laser reflection, collectively depicting the three-dimensional geometry of the tunnel wall. Multi-sensor data refers to a standardized dataset aligned in time and space after standardization, denoising, and registration of the multi-source raw data. This serves as the foundation for subsequent analysis, ensuring that data from different sources can be fused and analyzed within the same coordinate system. The terminal utilizes a mobile measurement system equipped with a visible light camera, an infrared thermal imager, and a laser scanner. Moving within the tunnel, it simultaneously triggers sensors to collect three types of raw data. The collected data undergoes preliminary cleaning, removing lens distortion from the visible light images, filtering noise from the point cloud, and performing data registration. Through algorithmic calculations, data collected by different sensors at different times is precisely mapped to the same unified three-dimensional coordinate system. This ensures that every pixel in the visible light image and every temperature point in the thermal image can find its corresponding spatial location in the three-dimensional point cloud. Time synchronization ensures that all data was collected at the same moment, avoiding positional errors caused by vehicle movement. This transforms the chaotic raw data into a spatiotemporally aligned and clean multi-sensor data set, laying a reliable data foundation for subsequent precise geometric division and feature analysis of the tunnel as a whole.

[0059] Step 102: Based on multi-sensor data, divide the tunnel into sections along the centerline of the tunnel to obtain a tunnel section sequence; the tunnel section sequence includes each tunnel section and the corresponding multi-sensor data.

[0060] Specifically, the tunnel centerline is a three-dimensional spatial curve representing the overall orientation of the tunnel, located at the geometric center of the tunnel body, similar to the tunnel's spine. A tunnel cross-section is an imaginary plane perpendicular to the tunnel centerline, forming the cross-sectional profile where this plane intersects the tunnel's inner wall. A tunnel cross-section sequence is an ordered set of tunnel cross-sections generated at regular intervals along the tunnel centerline. Each element in the sequence contains not only the geometric location information of the cross-section itself but also data from all multi-sensor data located near that cross-section. Based on the preprocessed point cloud data, the terminal uses an algorithm to automatically fit the three-dimensional centerline of the tunnel. Along the centerline, the cross-section position is set at a preset fixed distance. At each position point, the tangent direction of the centerline at that point is calculated, and a plane perpendicular to the tangent is generated. This is the preliminary tunnel cross-section. In order to improve the alignment between the cross-section and the real tunnel structure, the position of the preliminary cross-section is fine-tuned by combining obvious structural features in visible light images and thermal imaging, so that it falls precisely on the real structural boundary. For each refined cross-section position, a small spatial range of data near the cross-section is cut out from the registered multi-sensor data to form a data slice exclusive to that cross-section.

[0061] Step 103: Based on the tunnel cross-section sequence, perform binary segmentation on the tunnel cross-section to obtain the cross-section binary segmentation mask.

[0062] Specifically, a cross-sectional binary segmentation mask is an image representation of a single tunnel cross-section. It has the same size as the visible light image of the cross-section, but each pixel in the image has only two values: for example, a pixel can be white (value 1) representing a leaking area, or black (value 0) representing a normal background area, clearly indicating the precise pixel location of the defect. The terminal uses a trained semantic segmentation AI model to identify leak-related textures such as wetting marks, water flow paths, and calcifications in the image. A temperature threshold is set to identify abnormally low-temperature areas where the surface temperature is lower than the surrounding normal area. The reflection intensity and geometric information of the point cloud are analyzed, and these three analysis results are fused. Considering the reliability of different sensors under different environments, appropriate weights are assigned to each result to calculate a comprehensive and more reliable region segmentation result. The fused result undergoes image morphological processing, including closing operations to fill small holes inside the region and opening operations to remove small burrs at the edges, making the segmented region boundaries smoother and more complete, thus obtaining the cross-sectional binary segmentation mask.

[0063] Step 104: Based on the cross-sectional binary segmentation mask, identify the type of water leakage disease in each tunnel cross-section and obtain the disease classification result.

[0064] The disease classification result refers to the determination of the specific leakage type for each individual identified leakage area. Common types include seepage, dripping, flowing water, and mud leakage, and include the label, location, and classification confidence score for each disease area. For each leakage area identified in the mask, the terminal extracts a series of numerical values ​​describing its characteristics from multi-sensor data to form a feature vector. It evaluates which features are most important for classification, selects a subset of features, and inputs these features into a pre-trained classification model. The classification opinions from multiple models are combined through a voting-like mechanism to obtain a final integrated leakage type. Domain knowledge is then used for rationality verification to ensure the result conforms to common engineering sense.

[0065] The automatic identification method for tunnel seepage defects provided in this embodiment acquires multi-source raw data from the tunnel; preprocesses the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging, and point cloud data; based on the multi-sensor data, the tunnel is divided into sections along the centerline to obtain a tunnel section sequence; the tunnel section sequence includes each tunnel section and its corresponding multi-sensor data; based on the tunnel section sequence, the tunnel sections are binary segmented to obtain a section binary segmentation mask; based on the section binary segmentation mask, the seepage defect type of each tunnel section is identified to obtain a defect classification result; based on the defect classification result, the severity of the defect is assessed to obtain a defect assessment report for the corresponding tunnel section; and all defect assessment reports are integrated to obtain a tunnel assessment report. By segmenting the tunnel into multiple sections, data collected at different periods can be accurately correlated spatially, the development trajectory of defects can be accurately tracked, a unified section reference coordinate system can be established, and the comparability and trend analysis value of long-term monitoring data can be improved.

[0066] In one embodiment, based on multi-sensor data, tunnel sections are divided along the tunnel's centerline to obtain a tunnel section sequence, including:

[0067] Step 201: Based on the point cloud data in the multi-sensor data, extract the three-dimensional centerline of the tunnel to obtain the tunnel centerline.

[0068] Point cloud data consists of a large collection of three-dimensional spatial points acquired by a laser scanner. Each point contains precise coordinate information, collectively depicting the three-dimensional morphology of the tunnel's inner wall surface. The tunnel centerline is a three-dimensional spatial curve that runs through the entire tunnel and represents its macroscopic orientation; it is a skeleton line calculated after integrating the tunnel's global geometry. Based on the overall distribution of the point cloud, the terminal uses principal component analysis to find the longest extension direction in the point cloud data. This direction is the tunnel's initial axis. Along this initial axis, a series of cutting planes perpendicular to this direction are generated at close intervals. These planes slice the entire tunnel point cloud into many thin slices. The center point of each slice is calculated. For each cutting plane, all point clouds falling within a very thin thickness range in its vicinity are extracted to form a cross-sectional point cloud slice. The center of the two-dimensional slice point cloud is calculated. This two-dimensional point cloud slice is fitted into a circle or ellipse. The center of this fitted circle or ellipse is the center point of the cross section. This will result in a series of ordered points located approximately at the center of the tunnel. Due to the influence of point cloud noise and tunnel appendages, the directly calculated center point sequence may not be smooth enough and may have jitter. The initial center point sequence is smoothed and filtered. A three-dimensional spline curve fitting algorithm is used to fit the smoothed center point sequence. The algorithm will generate a continuous and smooth three-dimensional space curve that passes through all the center points. The fitted curve is the final tunnel centerline.

[0069] Step 202: Based on a preset interval, generate equidistant cross-sectional positions on the tunnel centerline; and calculate the central tangent direction as the cross-sectional normal vector based on the equidistant cross-sectional positions to obtain a preliminary cross-sectional position sequence.

[0070] Specifically, the preset interval is a fixed distance value set in advance, which can be 1 meter or 0.5 meters, determining the density of cross-sections generated along the centerline. The equidistant cross-section positions are a series of equally spaced three-dimensional spatial points marked at preset intervals along the tunnel centerline. The central tangent direction is defined as the direction of a straight line tangent to the curve at any point on the centerline. The cross-section normal vector is a three-dimensional direction vector perpendicular to the cross-section plane; in this embodiment, it is defined as the direction perpendicular to the central tangent at the cross-section position, determining the orientation of the generated cross-section plane. The preliminary cross-section position sequence is an ordered list, where each element contains the three-dimensional coordinates of an equidistant cross-section position and the corresponding cross-section normal vector, defining a series of preliminary, regularly distributed cross-section planes. The terminal, located on the three-dimensional centerline, starts from the starting point and performs equal-arc-length sampling at preset intervals. Along the length of the curve, a point is taken at fixed distances to obtain a series of equidistant cross-sectional position points. For each sampling point, the mathematical properties of curve fitting can be used to accurately calculate the tangent direction of the centerline at that point. The tangent direction represents the tunnel's forward direction at that point. The plane perpendicular to this tangent direction is the cross-section. Therefore, the tangent direction vector is defined as the cross-section normal vector. The three-dimensional coordinates and corresponding normal vectors of each sampling point are packaged together and arranged in order from the tunnel's starting point to its ending point to form a preliminary cross-sectional position sequence. The sequence is essentially a list containing position and direction information.

[0071] Step 203: Based on the preliminary cross-section position sequence, the cross-section position and the real structural features are aligned according to the visible light image and thermal imaging to obtain the accurate cross-section position sequence.

[0072] Specifically, actual structural features refer to the physical structures that actually exist on the tunnel wall and are clearly identifiable in visible light and thermal images. For example, the most common features are segment joints (the gaps between precast concrete segments), bolt holes, grouting holes, etc. These features typically appear as clear straight edges or regular patterns in the image. The precise cross-section position sequence is an adjusted and optimized sequence of cross-section positions. Based on the initial sequence, the position of each cross-section is fine-tuned to precisely align it with the nearest actual structural feature. The cross-section positions in the sequence are no longer strictly equidistant, but they better reflect engineering reality. The terminal utilizes edge detection and line detection algorithms from computer vision to automatically identify linear structural features such as pipe segment joints on a visible light image that has been registered with point cloud data. These features appear as distinct edges in the image. Each cross-section position in the initial cross-section position sequence is projected onto the corresponding two-dimensional visible light image. Within a small search range centered on the projection point, the nearest identified significant structural feature is searched. Once the nearest joint line is found, the current cross-section position is fine-tuned along the centerline so that its projection point on the image precisely falls on this joint line. This fine-tuning process is achieved through an optimization algorithm, aiming to minimize the alignment error between the cross-section position and the structural feature on the image. After all cross-section positions have undergone the above fine-tuning, a new sequence is formed, namely the precise cross-section position sequence. At this point, the cross-section positions are no longer strictly equidistant, but they are all aligned with the physical structure.

[0073] Step 204: Based on the position of each cross section in the precise cross section position sequence, extract the data slice corresponding to each cross section from the multi-sensor data to obtain the tunnel cross section sequence.

[0074] Specifically, a data slice is a local data block extracted from the corresponding registered multi-sensor data for each cross-section location in the precise cross-section location sequence. It represents all the information within a small segment of tunnel space near that cross-section location. For example, it includes: a visible light image slice (a local image near the cross-section); a thermal imaging slice (a local temperature distribution image near the cross-section); and a point cloud data slice (a three-dimensional point cloud within a certain thickness range near the cross-section). The tunnel cross-section sequence is an ordered set of all cross-section data slices. Each element in the sequence represents a complete tunnel cross-section, containing geometric location information and associated with all multi-sensor data slices at that cross-section, thus preparing the data for the next step of leakage identification. For each location in the precise cross-section location sequence, the terminal defines a three-dimensional thin-slice space of finite thickness centered on its cross-section plane. From the complete point cloud, all points falling within the three-dimensional thin-slice space are extracted to form a point cloud data slice for that cross-section. Since the image has been registered with the point cloud, it can be back-mapped onto the visible light image and thermal image based on the projection of the three-dimensional thin-slice space onto the point cloud, thereby extracting the corresponding image data slice. All types of data slices extracted from each cross-section location are packaged into a complete data unit, and then all data units are organized in sequence to obtain the tunnel cross-section sequence.

[0075] This embodiment transforms continuous tunnel data into discrete analysis units, resulting in a structured, information-rich sequence of tunnel cross-sections. This breaks down the lengthy tunnel data into smaller parts, allowing for independent and parallel refined identification and analysis of leakage for each standardized cross-section unit, thereby enabling automated detection of tunnel leakage defects. By aligning the geometrically defined cross-sections with the real structure in the physical world, a precise sequence of cross-section locations is obtained, ensuring that each cross-section is located at a consistent structural boundary. This makes it possible to accurately compare the same cross-section detected at different times, greatly facilitating long-term monitoring of defect development and structural health.

[0076] In one embodiment, based on the tunnel cross-section sequence, seepage areas are identified to obtain a cross-section binary segmentation mask, including:

[0077] Step 301: Based on the data slices of the visible light image, identify the wet areas and seepage textures through a semantic segmentation network, and perform region segmentation based on the wet areas and seepage textures to obtain the visible light segmentation results.

[0078] Specifically, visible light image data slices refer to local visible light images extracted from a specific tunnel cross-section, clearly showing the surface condition of the tunnel at that cross-section. Semantic segmentation networks are trained deep learning models tasked with pixel-level classification, determining the predefined category for each pixel in the input image. Wet areas and seepage textures are visual traces left by seeping water in the image. Wet areas appear as dark patches that are darker than the surrounding dry areas; seepage textures include specific patterns such as water flow marks, calcification, white exudates, and mold. The visible light segmentation result is a binary image of the same size as the input image. The value of each pixel in the image represents the confidence level of the model identifying that location as a leak, or is simply the binarized result, intuitively showing suspicious areas judged solely from visual appearance. The terminal inputs data slices of the visible light image into a pre-trained semantic segmentation network. The multi-layer convolutional structure inside the network automatically extracts low-level and high-level features of the image and outputs a segmentation map, in which pixel regions considered to have wet areas and seepage texture features are highlighted. The segmentation map directly output by the model is the visible light segmentation result, which is a probability map representing the probability that each pixel has wet areas and seepage texture features. It is then converted into a binary map through thresholding.

[0079] Step 302: Based on the data slices from thermal imaging, identify abnormally low temperature regions with temperatures below a preset temperature threshold to obtain temperature segmentation results.

[0080] The thermal imaging data slice refers to a local thermal image corresponding to the same tunnel cross-section. This image visually displays the temperature distribution of the object's surface in a color-coded format. The preset temperature threshold is a pre-set temperature value based on the ambient temperature and the normal temperature range of the tunnel surface. This threshold is lower than the average temperature of a dry surface and is used to define anomalies. Abnormally low-temperature areas are caused by the absorption of heat during moisture evaporation, resulting in a lower temperature on the wet tunnel surface compared to the surrounding dry areas. Areas with significantly lower temperatures are abnormally low-temperature areas and are important indicators of leakage. The temperature segmentation result is a binary image, where white pixels represent points with temperatures below the preset threshold, thus identified as abnormally low-temperature areas; black pixels represent normal temperatures. The terminal reads the temperature value of each pixel in the thermal imaging data slice and compares the temperature of each pixel with the preset threshold. If the pixel temperature is below the threshold, it is marked as 1, indicating an anomaly; otherwise, it is marked as 0, indicating normal. The judgment results of all pixels are combined into an image, i.e., the temperature segmentation result.

[0081] Step 303: Based on the data slices of the point cloud data, detect the wet areas on the surface according to the point cloud reflection intensity information, and identify geometrically abnormal areas to obtain the point cloud segmentation results.

[0082] Specifically, a point cloud data slice is a local 3D point cloud data corresponding to the same cross-section, containing spatial coordinates and reflection intensity information for each point. Point cloud reflection intensity is the signal intensity returned after a laser beam emitted by a lidar strikes the surface of an object. Wet surfaces absorb more light energy or produce diffuse reflection, resulting in a lower returned reflection intensity value than dry, smooth surfaces. Wet areas are potential leakage zones identified by analyzing regions in the point cloud where reflection intensity is significantly reduced. Geometric anomaly areas are unevenness phenomena such as bulges, depressions, and peeling that may occur on the surface due to long-term erosion by seeping water. These will appear as deviations from the normal tunnel outline in the 3D coordinates of the point cloud. The point cloud segmentation result is a binary map that identifies suspected leakage areas analyzed using the point cloud's reflection intensity and geometric information. The terminal calculates the distribution of reflection intensity in the point cloud slice, sets an intensity threshold, clusters points with low reflection intensity into regions, identifies them as wet surface regions, compares the point cloud contour of the current section with the point cloud of the standard tunnel design model or adjacent normal regions, detects geometrically abnormal regions with obvious concavity and convexity changes, merges the regions identified based on reflection intensity and the regions identified based on geometrically abnormalities, and generates point cloud segmentation results.

[0083] Step 304: Integrate the visible light segmentation results, temperature segmentation results, and point cloud segmentation results to obtain the segmentation results of the water leakage area.

[0084] The leakage area segmentation result is a more reliable comprehensive binary or probabilistic map obtained by fusing segmentation results from visible light, thermal imaging, and point cloud. It comprehensively considers multiple pieces of evidence and is more accurate and complete than any single-modality result. The terminal ensures that the three segmentation maps obtained in the first three steps are perfectly aligned in pixel coordinates, assigns a weight to the reliability of each modality under different environments, calculates the weighted average probability of the three modal results for each pixel position, and sets a final judgment threshold based on the fused probability value. If the fused probability of a pixel exceeds the threshold, it is marked as a leakage area in the leakage area segmentation result.

[0085] Step 305: Morphological optimization is performed on the segmentation results of the leakage area to obtain the cross-sectional binary segmentation mask.

[0086] Morphological optimization is an image processing technique used to improve the shape quality of regions in binary images. Basic operations include dilation, erosion, opening, and closing operations. The cross-sectional binary segmentation mask is a higher-quality segmentation result of the leaking water area after morphological optimization. In the image, white areas (value 1) represent the finally confirmed leaking water area, and black areas (value 0) represent the background. The terminal uses opening operations to eliminate isolated small points or spiky false areas in the leaking water area segmentation result, and uses closing operations to fill small holes inside the leaking area caused by incomplete detection, making the area more complete. Through a series of dilation and erosion operations, the edges of the area are smoothed to better reflect the actual morphology of the lesion.

[0087] This embodiment utilizes a machine learning model to mimic the human eye's judgment of image features, initially identifying suspicious leakage areas from their appearance. Its advantages include the ability to capture complex texture patterns; detection based on the principle of evaporative heat absorption, unaffected by color and texture under visible light, effectively identifying wet areas on the surface; providing evidence of optical properties and structural dimensions, enhancing the reliability of detection, especially in detecting early or hidden leaks that are difficult to detect under visible light; and multi-sensor information fusion, effectively overcoming the limitations of a single sensor, significantly improving the accuracy and recall rate of leakage area identification.

[0088] In one embodiment, the visible light segmentation results, temperature segmentation results, and point cloud segmentation results are integrated to obtain the leakage area segmentation results, including:

[0089] Step 401: Spatial alignment is performed on the visible light segmentation results, temperature segmentation results, and point cloud segmentation results to obtain a multimodal segmentation result matrix.

[0090] Spatial alignment is an image registration process designed to ensure that the segmentation results from different sensors completely correspond in terms of the physical regions and pixel coordinates covered on the same tunnel cross-section. Although the data has undergone preliminary registration, slight scale or positional deviations may be introduced during the generation of the segmentation map. The multimodal segmentation result matrix is ​​a three-dimensional data structure that can be imagined as a stack of images of identical size, with each layer representing the segmentation result of one modality. For any pixel location in the image, this matrix can simultaneously provide the segmentation value of that point in all three modalities. The terminal uses sensor calibration and registration parameters to map the three independently generated binary or probabilistic images onto the same precise two-dimensional coordinate system. This coordinate system is generated based on high-precision point cloud slice projection. If the image resolutions of different modalities are inconsistent, the lower-resolution image is resampled to have the same pixel size and number as the highest-resolution image. The aligned three images are then combined into a three-dimensional matrix, namely the multimodal segmentation result matrix, which serves as the common basis for subsequent fusion calculations.

[0091] Step 402: Calculate the confidence score of the multimodal segmentation result matrix to obtain the confidence score for each modality.

[0092] Specifically, confidence calculation is an evaluation process used to determine whether the segmentation results for each modality are reliable under its own conditions. The confidence score is a quantified value representing the overall reliability of the segmentation results for the corresponding modality; each modality calculates its own global confidence score. For the thermal imaging modality, its confidence score is positively correlated with the ambient temperature difference within the tunnel. If the temperature difference is small, the detection capability of thermal imaging will decrease, and a lower confidence score should be assigned. For the visible light modality, the confidence score is related to the gradient intensity of the segmented region's edges. If the boundaries of the humid region are very blurry, the confidence score will be low. For the point cloud modality, the confidence score is related to the signal-to-noise ratio of the reflection intensity value. If the signal is cluttered, the confidence score will be low.

[0093] Step 403: Based on prior knowledge of modalities under different environmental conditions, assign weight coefficients to each modality; and based on the confidence score and weight coefficients, calculate the multimodal confidence distribution of each pixel to obtain the multimodal segmentation result.

[0094] Specifically, prior knowledge consists of empirical rules summarizing the general performance of different sensors under various environmental conditions based on extensive practical experience. Examples include: thermal imaging detection is highly reliable during seasons with large diurnal temperature variations; visible light image analysis is more reliable when lighting is uniform and there is no shadow interference. Weighting coefficients are fixed weight values ​​pre-assigned to each modality based on prior knowledge, reflecting the inherent reliability of the sensor under typical ideal conditions. The multimodal confidence distribution is a new probability value calculated for each pixel in the image, integrating information from the three modalities; it is a weighted fusion of the joint judgment. The multimodal segmentation result is a new probability map, where the value of each pixel is the multimodal confidence distribution, representing the fused, preliminary leakage probability. The terminal assigns weight coefficients to each modality based on prior knowledge. For each pixel in the multimodal segmentation result matrix, its segmentation probability value in the three modalities is multiplied by the weight coefficient and confidence score of the corresponding modality, and then summed to obtain the multimodal confidence distribution of the pixel. This process is repeated for all pixels to generate a brand new probability map, i.e., the multimodal segmentation result.

[0095] Step 404: Based on the multimodal segmentation results, classify according to the classification threshold to obtain preliminary fusion segmentation results.

[0096] Specifically, the initial fusion segmentation result is a binary segmentation map, obtained by binarizing the multimodal segmentation results by setting a classification threshold. Pixels with a fusion probability greater than the threshold are classified as "leakage"; otherwise, they are considered background. This initial segmentation map is directly generated after the fusion of multimodal data. The terminal selects an appropriate probability threshold, for example, 0.5, and iterates through each pixel in the probability map of the multimodal segmentation result. If its probability value is greater than or equal to the threshold, the pixel is marked as 1 in the initial fusion segmentation result; otherwise, it is marked as 0.

[0097] Step 405: Based on the area and shape characteristics of the leakage area in the preliminary fusion segmentation results, filter out abnormal segmentation results to obtain the leakage area segmentation results.

[0098] Area and shape features are quantitative indicators extracted from the binary image to describe connected regions. Area refers to the total number of pixels contained within a region; shape features include the region's perimeter, roundness, and aspect ratio. Abnormal segmentation results refer to those segmented regions in the preliminary fusion segmentation results that clearly do not conform to the physical characteristics of the actual leakage area. For example, an area that is too small may be noise; an area that is too narrow and elongated may be a crack shadow. The leakage area segmentation result is a high-quality binary image of the leakage area that is more in line with engineering common sense after filtering by area and shape features. In the preliminary fusion segmentation result, the terminal finds all white, connected pixel regions and assigns a unique label to each region. It calculates the area and shape features of each connected region, sets filtering rules according to a preset reasonable range, and deletes regions with too small an area or abnormal shape from the result, that is, changes their pixel value from 1 to 0. All the leakage areas retained after filtering constitute the final leakage area segmentation result.

[0099] This embodiment considers the inherent reliability of the sensor and the adaptability to the specific environment to achieve dynamic weighted fusion, making disease identification more reliable. It transforms continuous probability values ​​into explicit binary segmentation maps, generating direct region identification results based on multi-source information fusion. By utilizing simple geometric common sense, it effectively filters out unreasonable noise and obvious misjudgments that may be generated during the fusion process, greatly improving the credibility and practicality of automatic disease identification.

[0100] In one embodiment, based on a cross-sectional binary segmentation mask, the type of water leakage defect in each tunnel cross-section is identified to obtain the defect classification results, including:

[0101] Step 501: Based on the cross-sectional binary segmentation mask, extract the features of each leakage area to obtain a multi-dimensional feature vector; the features include at least one of geometric features, texture features, temperature features, morphological features and distribution features.

[0102] The cross-sectional binary segmentation mask is a black and white image where white pixels identify areas suspected of water leakage. A leakage area is an independent connected region formed by all interconnected white pixels in the cross-sectional binary segmentation mask, representing an independent suspected defect point. The multidimensional feature vector is a numerical array used to mathematically describe a leakage area. Each dimension of the vector represents a specific quantitative feature of the area, transforming an image region into a set of numbers that can be processed by machine learning models. The terminal performs connected component analysis on the cross-sectional binary segmentation mask to identify all independent leakage areas and assigns a unique label to each area. For each identified area, various types of features are extracted from the registered multi-sensor data. All calculated feature values ​​are combined in a fixed order into a one-dimensional array, which is the multidimensional feature vector of that leakage area. For example, the calculations include geometric features calculated from the binary mask itself, such as the area (number of pixels), perimeter, and bounding box of the region; texture features calculated from the interior of the region in the corresponding visible light image slice, such as contrast, correlation, and entropy extracted using the gray-level co-occurrence matrix, used to quantify the texture of water stains and crystals; temperature features calculated from the interior of the region in the corresponding thermal imaging slice, such as the average temperature, minimum temperature, and temperature variance of the region; morphological features describing the shape of the region, such as roundness, aspect ratio, and density; and distribution features used to describe the position of the region on the tunnel cross-section, such as the radial distance and angle of the region's centroid relative to the center of the cross-section.

[0103] Step 502: Evaluate the feature importance of the multidimensional feature vectors, select the feature subset that contributes the most to the classification task, and obtain the region feature vector.

[0104] Specifically, feature importance assessment is a statistical analysis process used to measure the contribution of each feature dimension in a multidimensional feature vector to distinguish different leakage types. Unimportant features may be redundant or noisy. A feature subset is a smaller, more refined set of features selected from all the original features. The regional feature vector is a new feature vector composed of the selected feature subsets, retaining the most discriminative information while being lower in dimensionality and more efficient. The terminal uses feature importance ranking based on random forest, mutual information, or recursive feature elimination to analyze a large number of labeled training samples, calculates the importance score of each feature, selects the top-ranked features based on the importance scores to form a feature subset, and for each leakage area, only the feature values ​​corresponding to its feature subset are retained and recombined to form the regional feature vector.

[0105] Step 503: Input the region feature vectors into the pre-trained attention classification model and the pre-trained machine learning classifier respectively to obtain multiple preliminary classification results.

[0106] Specifically, the pre-trained attention classification model is an advanced deep learning model, which can be a Transformer-based model or a convolutional neural network with an attention mechanism. It has been trained using a large amount of labeled leakage data and is characterized by its ability to automatically focus on and amplify the most critical features in the region feature vector for classification, mimicking the ability of experts to identify key points. The pre-trained machine learning classifier refers to some traditional but effective machine learning models, including support vector machines, random forests, or gradient boosting trees, which have also been trained. The preliminary type classification result is the classification conclusion independently output by each classification model after receiving a region feature vector, including the predicted disease type and a confidence score. The terminal simultaneously inputs the same region feature vector into both the pre-trained attention classification model and the pre-trained machine learning classifier. Multiple models independently perform inference calculations based on their own learning mechanisms, and each model outputs its own preliminary type classification result.

[0107] Step 504: Perform a weighted vote on all preliminary type classification results to obtain the integrated leakage type.

[0108] Weighted voting is an ensemble learning strategy that assigns a weight to the vote of each classification model. The weight is typically based on the model's historical accuracy on the validation set; higher accuracy results in a larger weight. The ensemble leakage type is the final collective decision type obtained by combining the preliminary classification results of all models through the weighted voting mechanism. The terminal assigns a weight to each participating classification model. For each possible leakage type, the sum of the weights of all models predicting that type is accumulated, and the leakage type with the highest total weight score is selected as the ensemble leakage type.

[0109] Step 505: Based on domain knowledge, perform consistency verification on the integrated leakage types to obtain the disease classification results.

[0110] Domain knowledge comprises the physical laws and common sense about water leakage defects summarized from tunnel engineering practice. For example, at the tunnel arch, due to gravity, dripping is more likely than seepage; point-like areas typically have high circularity; and the area of ​​a certain type of defect is usually within a reasonable range. Consistency verification involves logically comparing the integrated leakage type with the domain knowledge rules extracted from the region's characteristics to check its rationality. The defect classification result is a highly reliable defect type determination after dual assurance of model integration and knowledge verification. The terminal compares the integrated leakage type with the actual characteristics of the region. If a significant contradiction is found, verification is triggered. Results that pass verification are directly adopted as the defect classification result; questionable results are marked as pending review or forcibly corrected to the most probable type according to rules.

[0111] This embodiment effectively reduces the bias or random errors that may be generated by a single model through ensemble learning, resulting in more stable and reliable classification results. By verifying theoretical knowledge, it combines data-driven model results with the experience and knowledge of human experts, greatly improving the rationality and practicality of disease classification results and avoiding absurd errors that violate common sense in physics.

[0112] In one embodiment, after identifying the type of seepage defects in each tunnel section based on a cross-sectional binary segmentation mask and obtaining the defect classification results, the method further includes:

[0113] Step 601: Based on the disease classification results, assess the severity of the disease to obtain a disease assessment report for the corresponding tunnel section; and integrate all disease assessment reports to obtain a tunnel assessment report.

[0114] The defect assessment report is a summary of the assessment for a single tunnel section, including the section's location number, the number of defects found, the type, area, and location of each defect, and a comprehensive severity rating. The tunnel assessment report is formed by integrating defect assessment reports from all sections, and includes: an overview of the overall tunnel condition, a statistical map of defect distribution throughout the tunnel, a list of key defect sections, and an overall safety status rating. For each section, the terminal calculates its severity level based on defect classification results and a binary segmentation mask, combined with predefined assessment criteria, generating a section-level defect assessment report. All section assessment reports are then compiled in tunnel mileage order. Through data analysis and visualization techniques, charts and statistical information reflecting the overall distribution patterns of defects are generated, ultimately compiling into a complete tunnel assessment report.

[0115] Step 602: Based on multiple tunnel assessment reports, analyze the evolution pattern of the defects to obtain time series data of defect development.

[0116] Specifically, multiple tunnel assessment reports refer to several complete assessment reports generated after inspecting the same tunnel at different times, forming the historical data basis for time-series analysis. The evolution pattern of defects refers to the changing pattern of specific leakage defects within a tunnel over time, including whether the area expands at a uniform or accelerated rate, and how the severity evolves. Defect development time-series data is a structured dataset that records key indicators of specific defect points or tunnel sections in different inspection cycles, arranged chronologically. It can be a time-sorted data table, with each row representing one inspection and columns including fields such as defect area, severity score, and development rate. The terminal accurately matches detection data from different periods. Based on the precise cross-sectional location sequence aligned with the actual structure, it associates data from different periods on the same physical cross-section. Within the cross-section, it also matches leakage disease areas detected at the same location in different periods to determine if they are the same disease point. For each successfully matched disease point, it calculates the change in its value between two adjacent detections. The core is to calculate the disease development rate. The area, severity, and calculated development rate of each disease point at each detection time point are organized in chronological order to form the disease development time series data.

[0117] Step 603: Based on the time series data of disease development, predict the development trend of the disease and obtain a tunnel disease development trend report.

[0118] Among them, development trend prediction is based on historical time series data and uses mathematical models to infer the most likely development of defects in the future, falling under the category of time series forecasting. The tunnel defect development trend report is a predictive document whose core content is a forecast of the future health of the tunnel, indicating which defect points are likely to expand significantly, whose severity will escalate, and which sections require special attention. The terminal selects an appropriate prediction model for the defect development time series data for each defect point. For points with long historical data, complex time series prediction models such as autoregressive integral moving average models or long short-term memory networks may be used. For points with limited data, simple linear or exponential regression may be used for preliminary prediction. Using a trained model and inputting historical data, the system predicts key indicators for the defect point at one or more future time points. The prediction results for all defect points are summarized, visualized, and analyzed to generate the tunnel defect development trend report, which describes the overall trend and displays potential high-risk points.

[0119] Step 604: Based on the disease development trend report, determine the urgency of maintenance and obtain a predictive maintenance report decision; the predictive maintenance report decision is used to assist in making maintenance decisions for the tunnel.

[0120] Maintenance urgency is a comprehensive assessment indicator used to determine the level of urgency required to intervene in a particular defect. It comprehensively considers the current severity of the defect, its predicted rate of development, and its potential impact on tunnel structural and operational safety. Predictive maintenance report decisions are decision support recommendations directly geared towards operational actions. Their core is to provide a prediction-based, scientific maintenance plan. The terminal has a pre-set rule base that maps prediction results to specific action recommendations. These rules are typically based on a risk matrix, considering two dimensions: severity and urgency. The prediction result for each defect is matched with the decision rules, assigning it a maintenance urgency level and generating specific maintenance decision recommendations. The decision recommendations, urgency ranking, and prediction basis for all defect points are integrated to form the final predictive maintenance report decision. The report clearly recommends which points require immediate attention and which can be addressed later, thereby optimizing the allocation of limited maintenance resources. For example, a rule could be: if the current severity is high or the predicted area growth rate exceeds a threshold and a higher risk level is predicted before the next inspection, then the maintenance urgency is high, and the decision is to arrange maintenance immediately. If the current severity is moderate and the predicted development is slow, then the maintenance urgency is moderate, and the decision is to include it in the maintenance plan within 6 months.

[0121] This embodiment transforms predictive data into direct action guidelines, and the resulting predictive maintenance report decisions can help tunnel management units upgrade from preventive maintenance to more advanced predictive maintenance, thereby maximizing cost reduction, efficiency improvement, and safety assurance.

[0122] In one embodiment, based on multiple tunnel assessment reports, the evolution pattern of defects is analyzed to obtain time series data on defect development, including:

[0123] Step 701: Based on the tunnel assessment report, match the cross-sectional data from different periods to obtain the cross-sectional matching structure.

[0124] The tunnel assessment report is a summary report generated after a complete inspection of the tunnel at a specific point in time. Its basic data consists of the assessment results of a series of cross-sections along the tunnel axis. The cross-section matching structure is a data structure or mapping table whose core function is to establish the correspondence between inspection data from different periods. It can clearly indicate that the i-th cross-section in the t-1 period inspection corresponds to the j-th cross-section in the t period inspection. Furthermore, a specific defect area X on this cross-section corresponds to a defect area Y in the next period inspection, ensuring that the comparison is between the same physical location and the same physical defect in the tunnel. The terminal has generated a precise sequence of cross-sectional locations aligned with the actual tunnel structure. Using inherent structural features as anchor points, and by comparing the spatial distribution of features such as segment joints detected at different times, cross-sections from different periods can be associated one-to-one with great precision. After successful cross-section matching, specific defect areas need to be matched within the cross-section. The location, basic shape, and characteristics of all leakage areas in the same cross-section in two adjacent periods are compared. By calculating indicators such as the distance between the centroids of the areas and the similarity of the area shapes, it is determined whether the two areas in the two periods represent the development and evolution of the same defect.

[0125] Step 702: Based on the cross-section matching results, calculate the disease development rate using the following formula to obtain the cross-section dataset:

[0126]

[0127] in, For the rate of development, , These are the weighting coefficients. Let t be the area of ​​disease at the current time t. This represents the area of ​​disease at the previous time t−1. Score the severity of the disease at the current time t. The severity score of the disease at the previous time t−1. This represents the time interval between adjacent detections.

[0128] Specifically, the cross-sectional dataset is a structured collection of data. Based on the cross-sectional matching structure, it calculates a quantitative development index for each successfully matched disease point. It can be viewed as a data table, where each row represents the change record of a disease point between two monitoring periods, and the columns include the unique identifier of the disease point, its location, and the calculated index. The composite development rate is a comprehensive quantitative index used to describe the overall development speed of a disease point per unit time. It considers changes in both disease area and severity, avoiding the potential bias of a single index. The terminal extracts a successfully matched pair of disease points from the cross-sectional matching structure, obtains the area and severity scores at time t-1 and time t, and simultaneously obtains the time interval between the two detections. The obtained values ​​are substituted into the provided formula for calculation. The first part of the formula calculates the relative change rate of area, and the second part calculates the relative change rate of severity. The weighting coefficients are determined by expert experience or model training and are used to adjust the importance ratio of area change and severity change in the total score, reflecting the flexibility of the system to be adjusted according to different maintenance strategies. The calculated values ​​and other information of the disease points are recorded in the cross-sectional dataset, quantifying the dynamic development of the disease into a comparable numerical composite development rate.

[0129] Step 703: Based on the cross-sectional dataset, the development patterns of diseases are classified using a clustering algorithm to obtain the disease evolution classification results.

[0130] Specifically, clustering algorithms are unsupervised machine learning methods used to automatically group data based on their inherent similarity without pre-labeling. Disease evolution classification results refer to several typical patterns obtained after summarizing the development behavior of all disease points in the dataset through cluster analysis. Each disease point is assigned a pattern label. For example, patterns include rapid deterioration, slow growth, stable, and intermittent outbreaks. The terminal constructs a feature vector for each disease point to describe its development behavior, including multi-period area sequences, severity sequences, and location information. The feature vector is input into the clustering algorithm, which automatically analyzes all vectors, grouping disease points with similar development behaviors into the same group, analyzing the common characteristics of disease points within each cluster, and defining a physically meaningful label for each group—the disease evolution classification result.

[0131] Step 704: Fill in missing values ​​in the disease evolution classification results to obtain disease development time series data.

[0132] Specifically, missing value imputation is a data preprocessing technique. In time series data, due to omissions in a particular detection, data quality issues, or the intermittent nature of disease occurrence, data for a disease point at a certain time point may be missing. Missing value imputation uses reasonable estimates to fill these data gaps, forming a complete and continuous data sequence. Disease development time series data is a high-quality, complete, and well-organized collection of time series data. For each disease point, it contains a complete data record arranged chronologically, including the area, severity, development rate, and development pattern category of each detection period. The terminal inspection section dataset identifies which disease points have missing data at which time points. Based on the cause and context of the missing data, the most appropriate imputation strategy is selected. After imputing all missing values, a complete, continuous, and high-quality disease development time series data is obtained. For example, imputation strategies include: interpolation: if only a single point is missing, linear interpolation or spline interpolation can be used to estimate the missing value using data from adjacent time points; imputation based on clustering results: if a disease point is classified as "stable" but its data for a certain period is missing, the missing value can be estimated using the average development rate of other disease points in that category or the average performance of that point itself in other periods; marking as not detected: if it is believed that the disease may have temporarily disappeared, a specific value representing zero or negligible can be filled in.

[0133] This embodiment helps to understand the typical disease development patterns in tunnels through classification, providing an important basis for prediction and differentiated maintenance strategies. Data completion ensures that the data input to the subsequent trend prediction model is complete and reliable. The obtained disease development time series data is a solid data foundation for accurate predictive maintenance, thereby improving the reliability of identification and maintenance.

[0134] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0135] Based on the same inventive concept, this application also provides an automatic tunnel leakage defect identification device for implementing the above-mentioned automatic tunnel leakage defect identification method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the automatic tunnel leakage defect identification device provided below can be found in the limitations of the automatic tunnel leakage defect identification method above, and will not be repeated here.

[0136] In one exemplary embodiment, such as Figure 2 As shown, an automatic identification device 800 for tunnel water leakage defects is provided, comprising:

[0137] The preprocessing module 801 is used to acquire multi-source raw data in the tunnel and preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging and point cloud data;

[0138] The cross-section module 802 is used to divide the tunnel cross-section along the centerline of the tunnel based on multi-sensor data to obtain a tunnel cross-section sequence; the tunnel cross-section sequence includes each tunnel cross-section and the corresponding multi-sensor data.

[0139] The leakage module 803 is used to perform binary segmentation of the tunnel cross section based on the tunnel cross section sequence to obtain the cross section binary segmentation mask;

[0140] The disease module 804 is used to identify the type of water leakage disease in each tunnel section based on the cross-sectional binary segmentation mask, and obtain the disease classification results.

[0141] Furthermore, the section module 802 is also used for:

[0142] Based on point cloud data from multi-sensor data, the three-dimensional centerline of the tunnel is extracted to obtain the tunnel centerline;

[0143] Based on a preset interval, equidistant cross-sectional positions are generated on the tunnel centerline; and the central tangent direction is calculated as the cross-sectional normal vector based on the equidistant cross-sectional positions to obtain a preliminary cross-sectional position sequence.

[0144] Based on the preliminary cross-section position sequence, the cross-section position and the real structural features are aligned according to the visible light image and thermal imaging to obtain the accurate cross-section position sequence;

[0145] Based on the location of each cross section in the precise cross section location sequence, data slices corresponding to each cross section are extracted from multi-sensor data to obtain the tunnel cross section sequence.

[0146] Furthermore, the leakage module 803 is also used for:

[0147] Based on data slices of visible light images, a semantic segmentation network is used to identify wet areas and seepage textures, and region segmentation is performed based on the wet areas and seepage textures to obtain visible light segmentation results.

[0148] Based on thermal imaging data slicing, abnormally low temperature regions with temperatures below a preset temperature threshold are identified, and temperature segmentation results are obtained.

[0149] Based on point cloud data slicing, surface wet areas are detected according to point cloud reflection intensity information, and geometrically abnormal areas are identified to obtain point cloud segmentation results;

[0150] By integrating the visible light segmentation results, temperature segmentation results, and point cloud segmentation results, the segmentation results of the water seepage area are obtained.

[0151] Morphological optimization is performed on the segmentation results of the leakage area to obtain the cross-sectional binary segmentation mask.

[0152] Furthermore, the leakage module 803 is also used for:

[0153] Spatial alignment is performed on the visible light segmentation results, temperature segmentation results, and point cloud segmentation results to obtain a multimodal segmentation result matrix;

[0154] Calculate the confidence score of the multimodal segmentation result matrix to obtain the confidence score for each modality.

[0155] Based on prior knowledge of modalities under different environmental conditions, a weight coefficient is assigned to each modality; and based on the confidence score and the weight coefficient, the multimodal confidence distribution of each pixel is calculated to obtain the multimodal segmentation result.

[0156] Based on the multimodal segmentation results, classification is performed according to the classification threshold to obtain preliminary fusion segmentation results;

[0157] Based on the area and shape characteristics of the leakage area from the preliminary fusion and segmentation results, the abnormal segmentation results are filtered out to obtain the leakage area segmentation results.

[0158] Furthermore, the disease module 804 is also used for:

[0159] Based on the cross-sectional binary segmentation mask, the features of each seepage area are extracted to obtain a multi-dimensional feature vector; the features include at least one of geometric features, texture features, temperature features, morphological features and distribution features;

[0160] The importance of the multidimensional feature vectors is evaluated, and the subset of features that contributes the most to the classification task is selected to obtain the region feature vectors.

[0161] The region feature vectors are input into a pre-trained attention classification model and a pre-trained machine learning classifier, respectively, to obtain multiple preliminary classification results.

[0162] A weighted vote is performed on all preliminary classification results to obtain the integrated leakage type.

[0163] Based on domain knowledge, consistency verification is performed on the integrated leakage types to obtain the disease classification results.

[0164] Furthermore, the device also includes a decision module for:

[0165] Based on the disease classification results, the severity of the disease is assessed to obtain a disease assessment report for the corresponding tunnel section; and all disease assessment reports are integrated to obtain a tunnel assessment report.

[0166] Based on multiple tunnel assessment reports, the evolution pattern of defects was analyzed, and time series data of defect development were obtained;

[0167] Based on the time series data of disease development, the development trend of disease is predicted, and a tunnel disease development trend report is obtained.

[0168] Based on the disease development trend report, the urgency of maintenance is determined, and a predictive maintenance report decision is obtained; the predictive maintenance report decision is used to assist in the maintenance decision of the tunnel.

[0169] Furthermore, the decision-making module is also used for:

[0170] Based on the tunnel assessment report, cross-sectional data from different periods are matched to obtain the cross-sectional matching structure;

[0171] Based on the cross-section matching results, the disease development rate is calculated using the following formula to obtain the cross-section dataset:

[0172]

[0173] in, For the rate of development, , These are the weighting coefficients. Let t be the area of ​​disease at the current time t. This represents the area of ​​disease at the previous time t−1. Score the severity of the disease at the current time t. The severity score of the disease at the previous time t−1. The time interval between adjacent detections;

[0174] Based on the cross-sectional dataset, the development patterns of diseases are classified using a clustering algorithm to obtain the disease evolution classification results;

[0175] Missing values ​​were filled in the disease evolution classification results to obtain disease development time series data.

[0176] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the aforementioned method for automatic identification of tunnel seepage defects.

[0177] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0178] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0179] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for automatic identification of tunnel water leakage defects, characterized in that, The method includes: Acquire multi-source raw data from the tunnel; and preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging, and point cloud data; Based on the multi-sensor data, the tunnel is divided into sections along the centerline of the tunnel to obtain a tunnel section sequence; the tunnel section sequence includes each tunnel section and the corresponding multi-sensor data. Based on the tunnel cross-section sequence, the tunnel cross-section is binary segmented to obtain a cross-section binary segmentation mask; Based on the cross-sectional binary segmentation mask, the type of water leakage disease for each tunnel cross-section is identified, and the disease classification result is obtained.

2. The method according to claim 1, characterized in that, The process of dividing the tunnel cross-section along the centerline based on the multi-sensor data to obtain a tunnel cross-section sequence includes: Based on the point cloud data in the multi-sensor data, the three-dimensional centerline of the tunnel is extracted to obtain the tunnel centerline; Based on a preset interval, equidistant cross-sectional positions are generated on the centerline of the tunnel; and the central tangent direction is calculated as the cross-sectional normal vector based on the equidistant cross-sectional positions to obtain a preliminary cross-sectional position sequence. Based on the preliminary cross-section position sequence, the cross-section position and the actual structural features are aligned according to the visible light image and the thermal imaging to obtain the precise cross-section position sequence. Based on each cross-section location in the precise cross-section location sequence, a data slice corresponding to each cross-section is extracted from the multi-sensor data to obtain the tunnel cross-section sequence.

3. The method according to claim 2, characterized in that, The step of performing binary segmentation on the tunnel cross-section based on the tunnel cross-section sequence to obtain a cross-section binary segmentation mask includes: Based on the data slices of the visible light image, a semantic segmentation network is used to identify wet areas and seepage textures, and region segmentation is performed based on the wet areas and seepage textures to obtain visible light segmentation results. Based on the data slices from the thermal imaging, abnormally low temperature regions with temperatures below a preset temperature threshold are identified, and temperature segmentation results are obtained. Based on the data slices of the point cloud data, wet areas on the surface are detected according to the point cloud reflection intensity information, and geometrically abnormal areas are identified to obtain point cloud segmentation results; By integrating the visible light segmentation results, the temperature segmentation results, and the point cloud segmentation results, the leakage area segmentation results are obtained. Morphological optimization is performed on the segmentation results of the leakage area to obtain the cross-sectional binary segmentation mask.

4. The method according to claim 3, characterized in that, The integration of the visible light segmentation result, the temperature segmentation result, and the point cloud segmentation result yields the leakage area segmentation result, including: Spatially align the visible light segmentation result, the temperature segmentation result, and the point cloud segmentation result to obtain a multimodal segmentation result matrix; The confidence score of each modality is obtained by calculating the confidence score of the multimodal segmentation result matrix. Based on prior knowledge of the modalities under different environmental conditions, a weight coefficient is assigned to each modality; and based on the confidence score and the weight coefficient, the multimodal confidence distribution of each pixel is calculated to obtain the multimodal segmentation result. Based on the multimodal segmentation results, classification is performed according to the classification threshold to obtain preliminary fusion segmentation results; Based on the area and shape characteristics of the leakage area in the preliminary fusion segmentation results, the abnormal segmentation results are filtered to obtain the leakage area segmentation results.

5. The method according to claim 1, characterized in that, The method of identifying the type of water leakage in each tunnel section based on the cross-sectional binary segmentation mask, and obtaining the defect classification result, includes: Based on the cross-sectional binary segmentation mask, features of each leakage area are extracted to obtain a multi-dimensional feature vector; the features include at least one of geometric features, texture features, temperature features, morphological features, and distribution features; The multidimensional feature vectors are evaluated for feature importance, and the subset of features that contributes the most to the classification task is selected to obtain the region feature vectors. The region feature vectors are input into a pre-trained attention classification model and a pre-trained machine learning classifier, respectively, to obtain multiple preliminary type classification results; A weighted vote is performed on all the preliminary classification results to obtain the integrated leakage type; Based on domain knowledge, consistency verification is performed on the integrated leakage types to obtain the disease classification results.

6. The method according to claim 1, characterized in that, After identifying the type of water leakage damage for each tunnel section based on the cross-sectional binary segmentation mask and obtaining the damage classification result, the method further includes: Based on the disease classification results, the severity of the disease is assessed to obtain a disease assessment report for the corresponding tunnel section; and all the disease assessment reports are integrated to obtain a tunnel assessment report. Based on multiple tunnel assessment reports, the evolution pattern of the defects was analyzed to obtain time series data on defect development. Based on the time series data of the disease development, the development trend of the disease is predicted, and a tunnel disease development trend report is obtained. Based on the disease development trend report, the urgency of maintenance is determined, and a predictive maintenance report decision is obtained; the predictive maintenance report decision is used to assist in making maintenance decisions for the tunnel.

7. The method according to claim 6, characterized in that, Based on multiple tunnel assessment reports, the evolution pattern of the defects is analyzed to obtain time series data on defect development, including: Based on the tunnel assessment report, cross-sectional data from different periods are matched to obtain a cross-sectional matching structure. Based on the cross-section matching results, the development rate of the disease is calculated using the following formula to obtain the cross-section dataset: in, For the rate of development, , These are the weighting coefficients. Let t be the area of ​​disease at the current time t. This represents the area of ​​disease at the previous time t−1. Score the severity of the disease at the current time t. The severity score of the disease at the previous time t−1. The time interval between adjacent detections; Based on the cross-sectional dataset, the development patterns of the disease are classified using a clustering algorithm to obtain disease evolution classification results; Missing values ​​are filled in the disease evolution classification results to obtain the disease development time series data.

8. An automatic identification device for tunnel water leakage defects, characterized in that, The device includes: The preprocessing module is used to acquire multi-source raw data in the tunnel and preprocess the multi-source raw data to obtain multi-sensor data; the multi-source raw data includes visible light images, thermal imaging, and point cloud data. The cross-section module is used to divide the tunnel cross-section along the centerline of the tunnel based on the multi-sensor data to obtain a tunnel cross-section sequence; the tunnel cross-section sequence includes each tunnel cross-section and the corresponding multi-sensor data. The leakage module is used to perform binary segmentation on the tunnel cross-section based on the tunnel cross-section sequence to obtain a cross-section binary segmentation mask; The disease module is used to identify the type of water leakage disease for each tunnel section based on the cross-sectional binary segmentation mask, and obtain the disease classification result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images