A channel concrete lining disease identification method based on deep learning

Abstract

This application discloses a deep learning-based method for identifying defects in channel concrete lining, comprising: acquiring ground-penetrating radar (GPR) data of a target channel section; the GPR data including B-scan radar images acquired by multiple antenna channels along the same survey line; extracting candidate defect regions from the B-scan radar images, comparing the candidate defect regions with a pre-established defect morphology feature database to obtain defect type determination results for the candidate defect regions; inputting the candidate defect regions into a defect identification model to obtain defect judgment results for the B-scan radar images acquired by each antenna channel; and performing multi-channel fusion of the defect judgment results of each antenna channel at the same spatial location to obtain the defect identification result corresponding to the target channel section. The method of this application improves the accuracy and reliability of channel lining concrete defect identification by performing multi-channel identification and determination of the same spatial location of the channel concrete lining through a defect identification model.

Description

Technical Field

[0001] This application belongs to the field of engineering testing technology, and specifically relates to a method for identifying channel concrete lining defects based on deep learning. Background Technology

[0002] As a crucial component of water conservancy infrastructure for farmland irrigation, water diversion, flood control, and drainage, canal engineering is prone to internal or hidden defects such as cracks, voids, cavities, and loose areas in its concrete lining structure during long-term service due to factors such as hydraulic erosion, temperature changes, uneven foundation settlement, and material aging. If these defects are not detected and assessed in a timely and accurate manner, they often lead to increased leakage, reduced structural load-bearing capacity, and even canal instability and failure, seriously affecting the operational safety and service life of the project.

[0003] Currently, among non-destructive testing methods for internal defects in concrete linings, ground-penetrating radar (GPR) is widely used in the health monitoring of concrete structures such as channels, tunnels, and bridges due to its advantages of high detection efficiency, high resolution, and suitability for rapid continuous measurement. Several existing patents propose using B-scan images acquired by GPR for automatic or semi-automatic identification of internal defects in concrete structures. For example, patents CN113447536A and CN111781576A disclose defect identification through dielectric constant inversion combined with threshold discrimination, while patents CN119339063A and CN119471610A disclose the introduction of deep neural networks for end-to-end defect detection and classification of radar images to improve identification efficiency and automation. On the other hand, some existing technologies attempt to extract shape features of reflection anomalies from the geometric morphology of radar images and determine the type of defect based on template matching or shape similarity, thereby reducing reliance on manual interpretation, as exemplified by patent CN109544531A.

[0004] Existing ground-penetrating radar (GPR)-based technologies for identifying defects in concrete linings have made some progress in automation and intelligence. However, most solutions rely primarily on single-antenna channel analysis results, lacking effective fusion of redundant and complementary information between different channels. This leads to insufficient stability in defect type determination and depth positioning results. Therefore, there is an urgent need for a defect identification method specifically for canal concrete lining structures to better meet the practical needs for high-precision, automated defect detection in canal engineering operation and management. Summary of the Invention

[0005] The objective of this application is to provide a deep learning-based method for identifying defects in channel concrete lining, thereby achieving efficient and accurate identification of defects in channel concrete lining.

[0006] To achieve the above-mentioned technical effects, this application provides a deep learning-based method for identifying defects in channel concrete linings, comprising the following steps: The process involves acquiring ground-penetrating radar (GPR) detection data for the target channel section. The GPR detection data includes B-scan radar images collected by multiple antenna channels along the same survey line. Candidate disease areas are extracted from the B-scan radar images, and these candidate disease areas are compared with a pre-established disease morphology feature database to obtain disease type determination results for the candidate disease areas. The candidate disease areas are then input into a disease identification model to obtain disease judgment results for the B-scan radar images collected by each antenna channel. Finally, the disease judgment results from each antenna channel at the same spatial location are fused using a multi-channel method to obtain the disease identification results corresponding to the target channel section.

[0007] In this scheme, B-scan radar images collected by multiple antenna channels along the same survey line are acquired to conduct multi-channel observation of the same spatial location of the channel concrete lining. Before deep learning identification, candidate defect areas are extracted and pre-screened by combining them with a defect morphology feature database. This allows the deep learning model to determine defects based on the integration of prior engineering knowledge, thereby reducing the risk of invalid identification and misjudgment. At the same time, by fusing the defect judgment results from multiple channels and integrating the observation information from different channels, the stability and consistency of defect category and depth judgment results are improved, thus enhancing the overall accuracy, reliability, and engineering applicability of channel concrete lining defect identification.

[0008] As an improvement to the technical solution of this application, the step of extracting candidate disease regions from the B-scan radar image includes: Edge detection processing is performed on the B-scan radar image to filter the set of edge pixels whose reflection intensity differs from that of their neighbors by a preset threshold; a closing operation is performed on the set of edge pixels to aggregate discrete edge pixels into several connected regions; based on preset area thresholds, aspect ratio thresholds, or reflection intensity anomaly thresholds, image regions that meet the abnormal feature requirements are selected from the connected regions as candidate disease regions.

[0009] Furthermore, the step of comparing the candidate disease areas with a pre-established disease morphology feature database to obtain a preliminary disease type determination result for the candidate disease areas includes: Calculate the similarity between the candidate disease region and the disease morphology templates in the pre-established disease morphology feature library; when the maximum morphological similarity is greater than the first threshold, mark the candidate disease region as a suspected disease region that is given priority to enter deep learning inference; when the maximum morphological similarity is between the first threshold and the second threshold, mark the candidate disease region as a suspicious abnormal region; when the maximum morphological similarity is lower than the second threshold, mark the candidate disease region as a non-disease region and remove it.

[0010] Furthermore, the disease morphology feature database is pre-established through the following steps: Based on historical multi-channel B-scan radar images of different disease types, several standard sample areas are determined as disease morphology templates. For each disease morphology template, a set of morphological feature parameters is calculated and stored to characterize the geometric shape, radar reflection distribution, and statistical properties of the disease. Using disease type, template identifier, template sub-image, and morphological feature parameter set as indexes, the corresponding B-scan radar sub-image is stored to construct a disease morphology feature library.

[0011] As an improvement to the technical solution of this application, the step of inputting the candidate disease area into the disease identification model to obtain the disease judgment result of the B-scan radar image collected by each antenna channel includes: The candidate disease areas are input into the disease identification model. A first identification subtask coarsely classifies the candidate disease areas, obtaining three categories: defective, non-defective, and suspicious. For candidate disease areas identified as defective or suspicious, a second identification subtask is executed to further subdivide them into multiple categories, obtaining the disease type classification and confidence level. The disease type includes at least one of cracks, cavities, voids, and looseness. For candidate disease areas identified as defective, a depth positioning subtask is executed to obtain the starting and ending depths of the disease in the longitudinal direction of the B-scan radar image. The results from the first identification subtask, the second identification subtask, and the depth positioning subtask are integrated to output the disease judgment results of the B-scan radar images acquired by each antenna channel.

[0012] Furthermore, the depth localization subtask includes: For candidate disease areas at the same location along the path, B-scan radar images from multiple antenna channels are sampled along the depth direction to obtain multi-channel reflection intensity sets at each depth layer. The multi-channel reflection intensity sets at each depth layer are aggregated to form a multi-channel reflection depth curve. The start and end positions of abnormal changes in reflection energy on the multi-channel reflection depth curve are identified, and the depths corresponding to the start and end positions are taken as the start and end depths of the disease. The start and end depths of the disease on each single-channel reflection depth curve are statistically analyzed, and the multi-channel depth estimation results are fused to obtain a comprehensive estimate of the disease depth range.

[0013] Furthermore, the disease identification model is trained through the following steps: Multi-channel B-scan radar image samples from multiple channel projects were collected. Through on-site drilling and manual interpretation, the samples were annotated at multiple levels, indicating the presence, type, depth, and boundaries of diseased areas. The annotated samples were divided into training, validation, and test sets to prevent data from the same physical survey line from being used across different sets. The training set was used to train the first identification subtask, the second identification subtask, and the depth positioning subtask individually or jointly until a preset accuracy or precision index was achieved. The validation set was used to monitor the training process and adjust training parameters. The test set was used to comprehensively evaluate the trained disease identification model, confirming its suitability for automatic identification of channel concrete lining diseases.

[0014] As an improvement to the technical solution of this application, the multi-channel fusion of the defect judgment results of each antenna channel at the same spatial location includes: For diseases located at the same spatial position, the corresponding area of ​​the disease is determined in the B-scan radar images of each channel; the contrast between the reflected signal and the background signal in the disease area of ​​each channel is calculated to obtain the observation signal-to-noise ratio of each channel for the disease; based on the observation signal-to-noise ratio and the channel visibility information obtained in advance for different disease types, channel weighting coefficients are assigned to each channel; using the channel weighting coefficients, the disease category determination results and disease depth estimation results of each channel are weighted and fused to obtain the disease category and disease depth range after multi-channel fusion.

[0015] As an improvement to the technical solution of this application, after acquiring the ground-penetrating radar detection data of the target channel section, the method further includes: preprocessing the B-scan radar images of each channel. The preprocessing includes the following steps: time alignment of the raw B-scan radar data for each channel; amplitude normalization of the time-aligned B-scan radar image to ensure that the reflection intensity of each channel is within a uniform amplitude range; and uniform processing of the size and grayscale range of the B-scan radar image according to the input format requirements of the disease identification model.

[0016] Furthermore, the preprocessing also includes noise suppression and contrast enhancement.

[0017] The beneficial effects of this application are as follows: The solution provided in this application acquires B-scan radar images collected by multiple antenna channels along the same survey line, enabling multi-channel observation of the same spatial location of the channel concrete lining. Before deep learning identification, candidate defect areas are extracted and pre-screened using a defect morphology feature database. This allows the deep learning model to determine defects based on prior engineering knowledge, thereby reducing the risk of invalid identification and misjudgment. Simultaneously, by fusing the defect judgment results from multiple channels and integrating observation information from different channels, the stability and consistency of defect category and depth determination results are improved, thus enhancing the overall accuracy, reliability, and engineering applicability of channel concrete lining defect identification. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a deep learning-based method for identifying defects in channel concrete lining, as described in one embodiment of this application. Detailed Implementation

[0020] The embodiments of the technical solution of this application will be described in detail below. The following embodiments are only used to more clearly illustrate the technical solution of this application, and are therefore merely examples and should not be used to limit the scope of protection of this application. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0021] The device parameters, threshold settings, and model structures involved in the embodiments of this application are merely illustrative examples, and those skilled in the art can adjust them according to actual engineering conditions.

[0022] Example 1 Please see Figure 1 The following is a flowchart illustrating a deep learning-based method for identifying defects in channel concrete lining, as provided in one embodiment of this application. The method may include the following steps: S1. Obtain ground-penetrating radar detection data for the target canal section; Specifically, the ground-penetrating radar detection data includes B-scan radar images acquired along the same survey line by multiple antenna channels. The horizontal axis of the B-scan radar image represents the position moved along the survey line, the vertical axis represents the electromagnetic wave two-way travel time or the converted depth, and the pixel values ​​in the image represent the amplitude of the reflection intensity.

[0023] For example, after selecting the detection area in the target channel section, several parallel survey lines are arranged longitudinally along the channel, with the spacing between the survey lines determined according to the detection accuracy requirements. Operators operate a multi-channel ground-penetrating radar (GPR) device, moving it at a constant speed along each survey line. The device automatically records GPR data along the route. During data acquisition, multiple antenna channels operate synchronously, each independently acquiring electromagnetic wave reflection signals and generating corresponding B-scan radar images. Taking a specific survey line as an example, after acquisition, six B-scan radar images are obtained, corresponding to channels A to F respectively. Each B-scan radar image can be represented as a two-dimensional matrix, with the number of horizontal and vertical sampling points determined by the radar device's sampling rate, time window settings, and data acquisition parameters, corresponding to the survey line position and depth range within a certain length.

[0024] To eliminate equipment differences and environmental noise, in one implementation, after this step, preprocessing of the B-scan radar images of each channel is also included. The preprocessing includes the following steps: a. Time-align the raw B-scan radar data for each channel; b. Normalize the amplitude of the time-aligned B-scan radar image to ensure that the reflection intensity of each channel is within a uniform amplitude range; c. Based on the input format requirements of the disease identification model, the B-scan radar image is processed to unify its size and grayscale range.

[0025] Specifically, due to the physical spatial spacing of the six antenna channels, there is a slight time difference in their response to the same reflecting target. The zero-time point of each channel is determined by detecting the position of the direct wave at the air-concrete interface, and the data from all channels are aligned to the same time reference. The alignment accuracy is typically on the order of 0.1-0.5 nanoseconds. Amplitude normalization aims to eliminate differences in reflection intensity between channels caused by variations in receiver gain or concrete moisture content, facilitating subsequent unified processing. This processing can optionally be achieved by statistically analyzing the reflection intensity values ​​of all pixels in each image and determining the minimum value. and maximum value ,,according to and Linearly normalize each pixel value I in the image so that all pixel values ​​in the image are uniformly mapped to a standard grayscale range, such as 0-255.

[0026] Size and grayscale uniformity can be achieved optionally by scaling or cropping B-scan images of different resolutions to a standard size, ensuring that grayscale values ​​are distributed within the range of 0-255. If the image resolution is lower than the target size, bilinear interpolation is used for upsampling; if it is higher than the target size, average pooling is used for downsampling.

[0027] Preferably, the preprocessing further includes noise suppression and contrast enhancement. Noise suppression includes background component subtraction and adaptive filtering. Background component subtraction calculates the average reflection intensity of each depth layer along the transverse direction (survey line direction) as the background component, representing the stable layer boundary reflection characteristics within the concrete. Subtracting this background component from the original image effectively suppresses surface reflections and strong reflections from the upper and lower surfaces of the concrete, thereby highlighting anomalous reflections associated with defects. Adaptive filtering is used to reduce high-frequency random noise and periodic interference. Wiener filtering is applied in areas where defects may exist, while medium-range filtering is applied in background areas. The filtering parameters are adaptively adjusted based on local signal statistical characteristics to suppress noise while preserving useful signals.

[0028] Contrast enhancement employs an adaptive histogram equalization (CLAHE) algorithm for local contrast enhancement. The specific steps are as follows: the image is divided into multiple local regions, for example, 8×8 sub-blocks, each sub-block being 64×64 pixels in size. Histogram equalization is performed on each sub-block to limit the magnitude of contrast enhancement. Bilinear interpolation is then used to smooth the boundaries of each sub-block, resulting in an image with overall enhanced contrast while avoiding artifacts caused by over-enhancement.

[0029] S2. Extract candidate disease regions from the B-scan radar image, compare the candidate disease regions with a pre-established disease morphology feature library, and obtain the disease type determination result of the candidate disease regions. Specifically, this step combines traditional image processing with prior engineering knowledge to quickly identify regions in B-scan images that may contain defects, and then performs preliminary classification using a morphological feature library, providing guidance and verification for subsequent deep learning recognition.

[0030] In one implementation, this step includes the following sub-steps: S21. Perform edge detection processing on the B-scan radar image to filter the set of edge pixels whose reflection intensity and neighborhood difference are greater than a preset threshold. Specifically, Canny edge detection or the Sobel operator is used to calculate gradients in the B-scan image to detect locations where adjacent pixels have significant gray-level differences. A gradient magnitude threshold is set (e.g., 30-50, depending on the image's gray-level range), and pixels with gradient magnitudes exceeding the threshold are marked as edge pixels. These edge pixels typically correspond to concrete delamination interfaces, crack boundaries, or the boundaries of internal defects.

[0031] S22. Perform a closing operation on the set of edge pixels to aggregate discrete edge pixels into several connected regions; Specifically, the binarized edge image is first subjected to dilation (e.g., using a 5×5 or 7×7 structuring element) to connect adjacent discrete edge pixels into continuous regions; then, erosion is performed to smooth the boundaries and remove small noise. After the dilation-erosion cycle, the originally scattered edge pixels are aggregated into several independent connected regions.

[0032] S23. Based on preset area threshold, aspect ratio threshold or abnormal reflection intensity threshold, select image regions that meet the abnormal feature requirements from the connected regions as candidate disease regions.

[0033] Specifically, for each connected region Calculate the following characteristic parameters: area Used to represent the number of pixels and aspect ratio of a connected region. Used to represent the length-to-width ratio of the circumscribed rectangle of a connected region, and the anomaly of reflection intensity. Used to represent the difference between the average reflectance of a pixel within a connected region and the average reflectance of the surrounding background region.

[0034] The settings should be set within a reasonable range based on the project being tested and the actual conditions of the ground-penetrating radar. A lower bound (e.g., 50 pixels, to remove small noise points) and an upper bound (e.g., 50,000 pixels, to remove background interference) should be set. For different types of defects, Set the appropriate aspect ratio range. For example, cracks typically have a larger aspect ratio (e.g., greater than 2), while cavities have an aspect ratio close to 1. If the difference exceeds a preset threshold (e.g., 50-100), it is identified as an abnormal area. After the above screening, several candidate defect areas are obtained. These candidate areas may contain both real defects and false anomalies (such as steel bar reflection, special aggregates, etc.), and further comparison with the morphological feature library is required.

[0035] In one implementation, this step also includes the following sub-steps: S24. Calculate the similarity between the candidate disease area and the disease morphology template in the pre-established disease morphology feature library; Specifically, morphological feature parameters identical to those in the template are extracted from each candidate disease area to form a feature vector for the candidate area. The morphological feature parameters preferably include geometric feature parameters (width W, depth D, aspect ratio LWR, morphological coefficient C), signal feature parameters (boundary sharpness G, internal uniformity U, signal-to-noise ratio SNR), and statistical feature parameters (center pixel value). Gradient rate of change Entropy value H).

[0036] Calculate the similarity between the candidate region and each template in the morphological feature library. Normalized Euclidean distance is used as the similarity metric, and the calculation formula is as follows: ,in Let be the standard deviation of the i-th feature parameter across all templates. The similarity S is defined as: S = 1 / (1+d), where S ranges from (0,1], with S closer to 1 indicating higher similarity. For each candidate region, calculate its similarity to all templates in the disease morphology feature library.

[0037] S25. When the maximum morphological similarity is greater than the first threshold, the candidate disease region is marked as a suspected disease region that is given priority to enter deep learning inference. When the maximum morphological similarity exceeds a first threshold (e.g., 0.75), the candidate disease region is marked as a suspected disease region that is prioritized for deep learning inference. At this point, the region has a high matching degree with a certain defect template in the morphological feature library and has strong defect features. Such regions are preferentially input into the subsequent deep learning disease identification model.

[0038] S26. When the maximum morphological similarity is between the first threshold and the second threshold, the candidate disease area is marked as a suspicious abnormal area. When the maximum morphological similarity falls between the first and second thresholds (e.g., between 0.5 and 0.75), the candidate defect region is marked as a suspicious anomaly region. At this point, the region neither perfectly matches any standard defect template nor can it be determined as normal background. Such regions require manual review or further inference through a deep learning model for identification.

[0039] S27. When the maximum morphological similarity is lower than the second threshold, the candidate disease area is marked as a non-disease area and removed.

[0040] When the maximum morphological similarity is below the second threshold (e.g., 0.5), the candidate defect area is marked as a non-defect area and removed. At this point, the features of this area have low similarity to all known defect templates and are judged as normal concrete background or equipment interference signals, which can be deleted from the candidate set to reduce the amount of data for subsequent processing.

[0041] Before implementing this step, a disease morphology feature database needs to be established in advance. This database is constructed based on historical multi-channel B-scan radar images and field verification data. In one implementation, the aforementioned disease morphology feature database is established in advance using the following method: Based on historical multi-channel B-scan radar images of different disease types, several standard sample areas were identified as disease morphology templates. For each of the disease morphology templates, calculate and store a set of morphological feature parameters to characterize the geometric morphology, radar reflection distribution, and statistical characteristics of the disease; Using disease type, template identifier, template sub-image, and morphological feature parameter set as indexes, store the corresponding B-scan radar sub-images to construct a disease morphological feature library.

[0042] Specifically, typical defect samples, verified by on-site drilling or other methods, were selected from historical testing data of multiple channel projects and concrete test blocks. For each defect type, 3-5 standard samples were selected as templates. For each sample, a rectangular sub-region containing the defect was extracted from the B-scan radar image as the defect morphology template.

[0043] In this embodiment, for each disease morphology template, the following morphological feature parameters are extracted and calculated as an example: Geometric feature parameters: width W, depth D, aspect ratio LWR, and shape coefficient C used to characterize the degree of curvature or straightness; Signal characteristic parameters: boundary sharpness G (average gradient intensity at the boundary of the template region), internal uniformity U (reciprocal of the standard deviation of the reflection intensity of pixels inside the template region), signal-to-noise ratio SNR, and attenuation characteristics (reflection attenuation curve in the depth direction). Statistical feature parameter: center pixel value Gradient rate of change Entropy value H (information entropy of pixel grayscale distribution in template region).

[0044] These parameters and their tolerance ranges are stored in feature vectors, forming disease morphology templates. The disease morphology feature library is stored in JSON format or a relational database for easy retrieval and updating. Below is an example of a database record: Template ID: Disease type_serial number (e.g., "crack_001") Type of damage: Cracks Template sub-image: 512×512 pixel grayscale image Set of morphological feature parameters: the vector form of the above morphological feature parameters Source information: Project name, survey line number, station number, depth Verification information: Description of the results of field drilling or other verification methods.

[0045] S3. Input the candidate disease area into the disease identification model to obtain the disease judgment results of the B-scan radar images collected by each antenna channel; Specifically, after the initial screening in the S2 stage, the candidate disease areas are input into a multi-level disease identification model, which makes a comprehensive judgment on the disease through three progressive identification sub-tasks.

[0046] In one implementation, this step includes the following sub-steps: S31. Input the candidate disease area into the disease identification model, and perform coarse classification of the candidate disease area through the first identification sub-task to obtain three categories of judgment results: defective, non-defective, and suspicious. Specifically, this subtask employs a lightweight classification network based on ResNet50. The input is a standardized B-scan sub-image (e.g., 512×512 pixels), which, after passing through the output of a shared feature extraction network (such as ResNet50), is followed by a global average pooling layer and a fully connected layer, outputting a 3D vector. After softmax activation, the probability distributions for "defect," "non-defect," and "suspicious" are obtained. The purpose of this subtask is to quickly determine whether the input region is a true defect. Regions judged as "non-defect" can be directly marked as disease-free and will not proceed to subsequent subtasks, thereby improving overall processing efficiency.

[0047] S32. For candidate disease areas that are identified as defects or suspected, perform the second identification subtask to perform multi-category subdivision identification and obtain the classification results and confidence levels of disease types; the disease types include at least one of cracks, cavities, voids, and loose areas; Specifically, for candidate regions identified as "defective" or "suspicious," a fine-classification subtask is performed. This fine-classification subtask connects the output of the shared feature extraction network to a morphological feature library fusion module and a classification head. The morphological feature library fusion module concatenates the similarity vector between the candidate region and the template in the morphological feature library with the deep features to form a fused feature vector. After passing through a fully connected layer, the fused feature vector outputs a multi-dimensional vector, which, after Softmax activation, yields the probability distribution of each disease type.

[0048] The types of defects include at least one of the following: cracks, voids, cavitation, and loose areas. Cracks include through cracks and non-through cracks; voids represent localized defects within the concrete, which may contain water, air, or debris; cavitation represents separation between the lining and the base layer; and loose areas represent diffuse regions where the concrete is not sufficiently dense. The defect type with the highest probability is selected as the sub-classification result, and this probability value is used as the confidence level.

[0049] S33. For candidate defect areas identified as defects, perform a depth localization subtask to obtain the starting and ending depths of the defect in the longitudinal direction of the B-scan radar image. Preferably, the above-mentioned depth localization subtask includes the following execution steps: a) For candidate disease areas at the same location along the route, B-scan radar images from multiple antenna channels are sampled along the depth direction to obtain the multi-channel reflection intensity set for each depth layer; b) Aggregate the multi-channel reflection intensity sets of each depth layer to form a multi-channel reflection depth curve; c) Identify the starting and ending positions of the abnormal changes in reflected energy on the multi-channel reflection depth curve, and take the depths corresponding to the starting and ending positions as the starting and ending depths of the disease. d) Statistically calculate the starting and ending depths of the disease on each single-channel reflection depth curve, and fuse the multi-channel depth estimation results to obtain a comprehensive estimate of the disease depth range.

[0050] Specifically, this subtask samples B-scan radar images from multiple antenna channels along the depth direction for candidate defect areas at the same location along the path, i.e., iterates through the reflection intensity values ​​of each channel row by row to obtain a multi-channel reflection intensity set for each depth layer. The multi-channel reflection intensity sets for each depth layer are then aggregated, for example, by calculating the mean or median of the reflection intensity at each depth, forming a one-dimensional curve, i.e., the multi-channel reflection depth curve. This curve is then Gaussian smoothed, and its gradient and curvature are calculated. By identifying locations where the derivative changes significantly, the initial depth of the defect is determined. and termination depth The starting and ending depths of the disease on each single-channel reflection depth curve were statistically analyzed, and the multi-channel depth estimation results were fused. The median fusion method was preferred to avoid the influence of outliers, and a comprehensive estimate of the disease depth range was obtained.

[0051] S34. Integrate the results obtained from the first identification subtask, the second identification subtask, and the depth positioning subtask, and output the defect judgment results of the B-scan radar images collected by each antenna channel.

[0052] Before performing step S3, the disease identification model needs to be trained. In one implementation, the disease identification model is trained through the following steps: a. Collect multi-channel B-scan radar image samples from multiple channel projects, and through on-site drilling and manual interpretation, perform multi-level labeling on the samples for the presence or absence of disease, disease type, disease depth range, and disease area boundary; b. Divide the labeled samples into training set, validation set and test set to prevent data from the same physical survey line from being used in different sets; c. Use the training set to train the first recognition subtask, the second recognition subtask, and the depth localization subtask separately or jointly until the preset accuracy or precision index is reached. d. Monitor the training process and adjust the training parameters using the validation set, and use the test set to comprehensively evaluate the trained disease identification model to confirm that the model is suitable for the automatic identification of defects in canal concrete lining.

[0053] Specifically, in this embodiment, the disease identification model adopts a multi-task learning architecture, consisting of a shared feature extraction network and three parallel task branches. The overall structure of the model is as follows: (1) Input layer The model's input is a preprocessed B-scan radar image, with an image size of 512×512 pixels, a single-channel grayscale image, and pixel values ​​ranging from [0, 255]. The input image is normalized to map pixel values ​​to the range [0, 1].

[0054] (2) Shared feature extraction network The shared feature extraction network employs a variant of the ResNet50 architecture to extract general high-dimensional features from B-scan radar images. The specific layer settings are as follows: Initial layer: Contains a convolutional layer with a kernel size of 7×7 (64 kernels, stride 2, padding 3), followed by a max-pooling layer with a kernel size of 3×3 (stride 2, padding 1). The output feature map size is 128×128×64.

[0055] Residual block set: Group 1: Contains 3 residual blocks, each containing 3 convolutional layers, with 256 output channels and a feature map size of 128×128; Group 2: Contains 4 residual blocks, each containing 3 convolutional layers. The first residual block has a stride of 2, 512 output channels, and a feature map size of 64×64. Group 3: Contains 6 residual blocks, each containing 3 convolutional layers. The first residual block has a stride of 2, 1024 output channels, and a feature map size of 32×32. Group 4: Contains 3 residual blocks, each containing 3 convolutional layers. The first residual block has a stride of 2, 2048 output channels, and a feature map size of 16×16. Feature Output Layer: Connected to the global average pooling layer, this layer aggregates the 16×16 spatial features into a 2048-dimensional deep feature vector. This deep feature vector serves as a common input, which is then distributed to subsequent subtask branches.

[0056] (3) First recognition subtask Input: A 2048-dimensional feature vector output by the shared feature extraction network; Structure: Fully connected layer FC1 (2048→512, ReLU, Dropout=0.5) → Fully connected layer FC2 (512→3, Softmax); Output: A 3-dimensional probability vector, corresponding to the labels: [0: defect, 1: non-defect, 2: suspicious].

[0057] (4) Second recognition subtask Input A: A 2048-dimensional feature vector output by the shared feature extraction network; Input B: M-dimensional morphological similarity vector calculated in step S2 (M=32 in this embodiment, corresponding to 3-5 standard templates for 4 disease types). Operation: Concatenate A and B along the channel dimension to form a (2048+M) dimensional fusion feature vector; Classification layer: Fully connected layer FC1 (2048+M→1024, ReLU, Dropout=0.5) → Batch normalization layer → FC2 (1024→512, Dropout=0.3) → Output layer FC3 (512→8, Softmax); Output: A K-dimensional probability vector representing the probability that a candidate region belongs to each disease type; the maximum probability is used as the confidence level. (K is the number of disease types; in this example, K=4, corresponding to cracks, voids, cavitation, and looseness.) (5) Depth localization subtask Input: A 2048-dimensional feature vector output by the shared feature extraction network. Structure: Fully connected layer FC1 (2048→512, ReLU, Dropout=0.5) → FC2 (51→256, ReLU, Dropout=0.3) → Regression layer FC3 (256→2, Sigmoid).

[0058] Output: 2D depth vector D = [ , This indicates the starting and ending depths of the disease.

[0059] This embodiment uses a multi-task joint loss function. The definition is as follows: in: coarse classification loss and fine classification loss Employing the cross-entropy loss function, depth localization loss The mean squared error loss is calculated using the following formula: In this embodiment, considering that the second identification subtask is the most difficult, the weight parameter is set as follows: .

[0060] The labeled samples were divided into training, validation, and test sets in a 7:2:1 ratio. During the partitioning, it was ensured that data from the same physical survey line were not used across different sets to avoid overfitting due to data leakage. An end-to-end joint training approach was adopted, simultaneously optimizing the loss functions of the three sub-tasks. The specific training process is as follows: S30-1: Initialize model parameters A shared feature extraction network (ResNet50) is loaded with pre-trained weights from the ImageNet dataset to obtain basic image feature extraction capabilities. The fully connected layers of each task branch use the Xavier initialization method.

[0061] S30-2: Phased Training Phase 1 (Freeze Training): Freeze the first 3 residual block groups (Group 1-3) of the shared network, and train only Group 4 and the three task branches.

[0062] Optimizer: Adam (β1=0.9, β2=0.999).

[0063] Learning rate: 1×10 −3 .

[0064] Epoch: 30.

[0065] Phase 2: Unfreeze all network layers and perform end-to-end joint training.

[0066] Learning rate: Reduced to 1×10 −4 Furthermore, a cosine annealing strategy is employed to reduce the learning rate to 1% as the training progresses.

[0067] Epoch: 50.

[0068] S30-3: Training Hyperparameter Configuration Batch size: Set to 16. Gradient accumulation technique is used, with an accumulation step of 4, resulting in an equivalent batch size of 64.

[0069] Regularization: Set the L2 weight decay factor to 1×10. −4 To prevent overfitting.

[0070] Data augmentation: During training, random horizontal flipping (probability 0.5), brightness adjustment (±20%), and Gaussian noise addition (σ=0.01) are performed online to simulate radar noise under real-world conditions.

[0071] S30-4: Score, a comprehensive evaluation metric on the monitoring validation set in, Coarse classification accuracy measures the ability to distinguish between defects / non-defects / suspects. The accuracy rate of detailed classification measures the accuracy of identifying four core disease types. The mean absolute error, This represents the maximum penetration depth of the device.

[0072] When the metric does not improve for 10 consecutive epochs, training is stopped and the optimal model weights are saved to confirm that the model is suitable for the automatic identification of channel concrete lining defects.

[0073] S4. Perform multi-channel fusion on the defect judgment results of each antenna channel at the same spatial location to obtain the defect identification result corresponding to the target channel section.

[0074] In one implementation, this step includes the following sub-steps: S41. For defects at the same spatial location, determine the corresponding area of ​​the defect in the B-scan radar images of each channel; Specifically, for defects identified as "defects" by S3, their spatial coordinates (location x and depth z along the path) in each channel are recorded. The corresponding area can be determined by registration of the relative position of the defect in each channel image, or by association based on the geographical coordinates of the defect.

[0075] S42. Calculate the contrast between the reflected signal and the background signal in the disease area of ​​each channel to obtain the observation signal-to-noise ratio of each channel for the disease. Specifically, the signal-to-noise ratio (SNR) is calculated, for example, as the ratio of the average reflectance of the affected area to the average reflectance of the adjacent normal background area. For example, for the i-th channel, it is defined as: in, Diseases, Diseases This represents the reflection intensity of the channel at the location of the lesion. This is the average background reflection intensity of the channel (usually calculated from the area far from the defect). The SNR value represents the standard deviation of the background reflection intensity. A higher SNR value indicates that the detection signal of the disease is clearer for this channel; a lower SNR value indicates that the disease signal may be masked by background noise.

[0076] S43. Based on the observed signal-to-noise ratio and the channel visibility information obtained in advance for different disease types, assign channel weighting coefficients to each channel; Specifically, different defect types exhibit a priori patterns in the visibility of different channels. For example, some cracks, due to their depth penetrating the entire concrete structure, are typically visible in all channels, with a priori visibility of approximately 1 for each channel. Deep cavities are usually only visible in channels close to the cavity, with low visibility in channels further away. Channel visibility information can be pre-statistically determined based on different defect types or through experimentation. For instance, a defect-channel visibility matrix V can be established through statistical analysis of historical annotated data. c is the prior visibility score for channel i. Combining real-time SNR and prior visibility pattern, the visibility index of channel i for defect type c is defined as follows: in This is the weight (typically 0.6). Then, the channel is calculated. Category Weighting coefficients: confidence The confidence level of the i-th channel in determining the defect is output by S3.

[0077] S44. Using the channel weighting coefficients, the disease category determination results and disease depth estimation results of each channel are weighted and fused to obtain the disease category and disease depth range after multi-channel fusion.

[0078] Specifically, the predicted probability of defect type c based on the Softmax output of the second identification task S3 is: We then perform weighted fusion as the confidence level to obtain the final defect type: The depth estimates of each channel are weighted and fused. First, channels with extremely low weights (e.g., channels with very low weights) are excluded. (This indicates very poor observation capability), and then the median fusion is performed on the depth estimates of the remaining channels: in The weighted threshold is 0.1 (typical value). Through a multi-channel fusion mechanism, the complementary information from multiple antenna channels can be effectively utilized to improve the accuracy and stability of disease identification.

[0079] Table 1 below shows the test results of the optimal model weights on the test set.

[0080] Table 1. Individual test results for each disease type in the test set. Example 2 To verify the non-destructive testing effect of Embodiment 1 of this application on different types of defects, this embodiment designed and fabricated plain concrete test blocks capable of independently and clearly presenting specific types of defects for verification experiments. The test block specifications were 1.5 m × 1.5 m × 0.35 m, and the concrete grade was C30. The prefabrication defect scheme of the test block is shown in Table 2 below.

[0081] Table 2 Defect Settings for Test Blocks The type, spatial location, size, and depth of all defects were designed and marked before the test blocks were poured. The test blocks were poured in layers, with each layer vibrated sequentially, and the concrete mix proportions and vibration intensity were strictly controlled. Concrete curing was carried out according to GB50007-2011 standard, and testing was conducted after 28 days of curing.

[0082] The surface of the test block was scanned using a GP8100 ground-penetrating radar in a grid pattern. The main technical parameters of this equipment are as follows: Radar technology: Stepped frequency continuous wave (SFCW); Modulation frequency range: 400-4000MHz; Number of antenna channels: 6, labeled as channels A, B, C, D, E, and F respectively; Antenna arrangement: linear array, channel spacing approximately 4cm; Penetration depth: Maximum 80cm; Detection width: 25cm coverage width per scan; Scanning method: Multi-channel synchronous acquisition.

[0083] Six parallel survey lines were arranged along the length direction, with a spacing of 20 cm, a frequency step of 20 MHz, and a repetition rate of 2.0. Following steps S1-S4 of Example 1, the acquired multi-channel B-scan radar images underwent preprocessing, candidate region extraction, morphological feature library comparison, disease identification model inference, and multi-channel fusion. The disease identification results are detailed in Table 3 below. Table 3 Results of disease identification in test blocks Verification conclusions The detection results showed that 12 out of 13 defects were correctly identified, achieving an accuracy rate of 92.3%. Among these, 5 crack-type defects were correctly identified, with 1 deemed suspicious, resulting in an accuracy rate of 83.3%. Crack identification characteristics included the highest identification rates for through-cracks and semi-through cracks, with confidence levels of 0.96 and 0.91 respectively, and depth errors within ±1.2 cm. These cracks exhibited distinct characteristics and high signal strength across multiple channels. For inclined cracks and non-through cracks, the observation capabilities of different channels varied. A multi-channel dynamic weighted fusion strategy was employed to align the channel weights with the visibility index, effectively improving the confidence level. Internal straight cracks, due to shallow blind zones and weak signals, had generally poor visibility across multiple channels. Multi-channel fusion increased the confidence level from an average of 68% per channel to 75% after fusion, thus improving the overall confidence level to some extent.

[0084] Of the void-type defects, all three air-filled defects were correctly identified, while the two water-filled defects were both considered suspicious anomalies. Voiding and non-compact defects were also correctly identified. The identification performance of water-containing voids was significantly reduced because the dielectric constant of water is approximately 81, close to that of concrete (4-8), resulting in extremely weak reflected signals and a signal-to-noise ratio below the 2 dB threshold. Although multi-channel fusion employed dynamic weighting, it could not overcome this physical constraint, leading to a very low confidence level for water-containing defects. Manual verification or supplementary validation using other detection methods is recommended.

[0085] The void defect is located at the bottom of the test block and has a horizontal layered structure. It exhibits relatively consistent signal characteristics in the identifiable antenna channels. The void identification confidence level is 0.91, and the depth error is ±1.1 cm, making it the best performing defect identification among the four defect types.

[0086] The non-dense defect manifests as a diffuse anomalous region with a gradually decaying gradient and no clear boundary. By using multi-channel fusion with uneven weight distribution, the contribution of high signal-to-noise ratio channels is emphasized, with a confidence level of 0.81 and a depth error of ±1.12.2 cm.

[0087] Test block experiments show that the technical solution of this application provides stable and reliable identification results for typical channel lining defects such as cracks, air voids, cavities, and looseness. For low-sensitivity targets such as water-bearing defects and shallow non-penetrating cracks, while multi-channel data joint analysis improves the stability of identification, its accuracy is still fundamentally limited by the constraints of physical signals. For these types of defects, it is recommended to use other non-destructive testing methods for supplementary verification, or mark them as suspected areas requiring manual review.

[0088] Example 3 To verify the applicability and effectiveness of the technology in this application under actual engineering conditions, a section of a canal in a canal renovation project in Northwest China, which was slated for lining removal and renovation, was selected for on-site testing. This section of the canal is a typical concrete-lined canal, with the lining structure made of C25W6F150 cast-in-place concrete, a slope thickness of 40cm, and obvious surface defects on the canal slope. Before the removal of the canal lining, a grid area was drawn based on typical crack defects on the right bank of the canal during the water outage period. A GP8100 structural radar was used for area scanning and detection, with six parallel survey lines arranged longitudinally along the canal, spaced 30cm apart, covering a width of approximately 1.8m of the canal slope. Defect identification was carried out according to the procedure in Example 1, and core sampling tests were conducted simultaneously.

[0089] The identification results output by this embodiment are compared with the manual interpretation results completed by experienced detection engineers based on the same radar data. The results are shown in Table 4 below.

[0090] Table 4 Comparison of Typical Disease Identification Results in Channels with Borehole Verification Verification conclusions Field testing of the canal concrete lining structure revealed that the system correctly identified 6 defects out of 8 selected typical verification points. Considering that F4 was a shallow micro-crack and F7 was a false anomaly caused by high moisture content, the accuracy rate for identifying major structural defects (through cracks, voids, and large cavities) remained at a high level. Compared with traditional manual interpretation, automated identification effectively reduced human error.

[0091] During on-site inspection, due to uneven aggregate distribution, single-channel images often exhibit random, strong reflection noise. In F2 detection, the confidence level of single-channel recognition fluctuates significantly (from 0.55 to 0.72), and there are instances of misclassification as non-defective. Through multi-channel weighted fusion in step S4, utilizing the spatial continuity features of adjacent channels, the final fusion confidence level is improved to 0.78. Core drilling verified that the depth estimation errors for the six correctly identified defects ranged from 0.8 cm to 2.2 cm, with an average error of 1.38 cm. Among them, the void defect (F5) had the most accurate depth positioning due to its obvious dielectric constant difference interface; the irregular cavity (F8) had a relatively large depth error due to its blurred boundaries.

[0092] In summary, Example 3 demonstrates that the identification method of this application, under the complex background noise of actual channel engineering sites, can still effectively identify and accurately locate key defects such as cracks, voids, and cavities by relying on the multi-channel fusion mechanism.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying defects in concrete lining of channels based on deep learning, characterized in that, include: Obtain ground-penetrating radar detection data for the target canal section; The ground-penetrating radar detection data includes B-scan radar images collected along the same survey line by multiple antenna channels; Candidate disease regions are extracted from the B-scan radar image, and the candidate disease regions are compared with a pre-established disease morphology feature library to obtain the disease type determination result of the candidate disease regions. The candidate disease areas are input into the disease identification model to obtain the disease judgment results of the B-scan radar images collected by each antenna channel; The defect judgment results of each antenna channel at the same spatial location are fused to obtain the defect identification result corresponding to the target channel section.

2. The method as described in claim 1, characterized in that, The extraction of candidate disease regions from the B-scan radar image includes: Edge detection processing is performed on the B-scan radar image to filter out a set of edge pixels whose reflection intensity and neighborhood difference are greater than a preset threshold. Perform a closing operation on the set of edge pixels to aggregate the discrete edge pixels into several connected regions; Based on preset area thresholds, aspect ratio thresholds, or abnormal reflection intensity thresholds, image regions that meet the abnormal feature requirements are selected from connected regions as candidate disease regions.

3. The method as described in claim 2, characterized in that, The step of comparing the candidate disease areas with a pre-established disease morphology feature database to obtain preliminary disease type determination results for the candidate disease areas includes: Calculate the similarity between the candidate disease region and the disease morphology template in the pre-established disease morphology feature library; When the maximum morphological similarity is greater than the first threshold, the candidate disease region is marked as a suspected disease region that is given priority to enter deep learning inference. When the maximum morphological similarity is between the first threshold and the second threshold, the candidate disease area is marked as a suspicious abnormal area. When the maximum morphological similarity is lower than the second threshold, the candidate disease area is marked as a non-disease area and removed.

4. The method according to any one of claims 1 to 3, characterized in that, The disease morphology feature database is pre-established through the following steps: Based on historical multi-channel B-scan radar images of different disease types, several standard sample areas were identified as disease morphology templates. For each of the disease morphology templates, calculate and store a set of morphological feature parameters to characterize the geometric morphology, radar reflection distribution, and statistical characteristics of the disease; Using disease type, template identifier, template sub-image, and morphological feature parameter set as indexes, store the corresponding B-scan radar sub-images to construct a disease morphological feature library.

5. The method as described in claim 1, characterized in that, The step of inputting the candidate disease areas into the disease identification model to obtain the disease judgment results of the B-scan radar images collected by each antenna channel includes: The candidate disease areas are input into the disease identification model. Through the first identification subtask, the candidate disease areas are coarsely classified to obtain three categories of judgment results: defective, non-defective, and suspicious. For candidate disease areas identified as defects or suspected, a second identification subtask is performed to perform multi-category subdivision identification, and to obtain the classification results and confidence levels of the disease types; the disease types include at least one of cracks, cavities, voids, and looseness; For candidate defect areas identified as defects, a depth localization subtask is performed to obtain the starting and ending depths of the defect in the longitudinal direction of the B-scan radar image. The results obtained from the first identification subtask, the second identification subtask, and the depth positioning subtask are integrated to output the defect judgment results of the B-scan radar images collected by each antenna channel.

6. The method as described in claim 5, characterized in that, The depth localization subtask includes: For candidate disease areas at the same location along the route, B-scan radar images from multiple antenna channels are sampled along the depth direction to obtain multi-channel reflection intensity sets for each depth layer; The multi-channel reflection intensity sets of each depth layer are aggregated to form a multi-channel reflection depth curve; Identify the start and end positions of abnormal changes in reflected energy on the multi-channel reflection depth curve, and take the depths corresponding to the start and end positions as the start and end depths of the disease. The starting and ending depths of the disease on each single-channel reflection depth curve were statistically analyzed, and the multi-channel depth estimation results were fused to obtain a comprehensive estimate of the disease depth range.

7. The method as described in claim 5, characterized in that, The disease identification model is trained through the following steps: Multi-channel B-scan radar image samples from multiple channel projects were collected, and through on-site drilling and manual interpretation, the samples were labeled with multi-level annotations of the presence or absence of disease, disease type, disease depth range, and disease area boundary. The labeled samples are divided into training set, validation set and test set to prevent data from the same physical survey line from being used in different sets. The first recognition subtask, the second recognition subtask, and the depth localization subtask are trained separately or jointly using the training set until the preset accuracy or precision index is reached. The training process was monitored and training parameters were adjusted using the validation set, and the trained disease identification model was comprehensively evaluated using the test set, confirming that the model is suitable for the automatic identification of defects in canal concrete lining.

8. The method as described in claim 1, characterized in that, The multi-channel fusion of the defect assessment results of each antenna channel at the same spatial location includes: For defects located at the same spatial location, the corresponding area of ​​the defect is determined in the B-scan radar images of each channel; The contrast between the reflected signal and the background signal in the disease area of ​​each channel is calculated to obtain the signal-to-noise ratio of the observation of the disease in each channel; Based on the observed signal-to-noise ratio and the channel visibility information obtained in advance for different disease types, channel weighting coefficients are assigned to each channel; Using the channel weighting coefficients, the disease category determination results and disease depth estimation results of each channel are weighted and fused to obtain the disease category and disease depth range after multi-channel fusion.

9. The method as described in claim 1, characterized in that, After acquiring the ground-penetrating radar detection data of the target channel section, the method further includes: preprocessing the B-scan radar images of each channel. The preprocessing includes the following steps: Time alignment is performed on the raw B-scan radar data from each channel; Amplitude normalization is performed on the time-aligned B-scan radar image to ensure that the reflection intensity of each channel is within a uniform amplitude range. According to the input format requirements of the disease identification model, the B-scan radar image is processed to unify its size and grayscale range.

10. The method as described in claim 9, characterized in that, The preprocessing also includes noise suppression and contrast enhancement.

Images