Method and device for rapidly and intelligently identifying tunnel diseases and medium
Through the self-supervised DRCNN model extracting features on labelless data and combining ROI Align Pooling to optimize feature extraction, the problem of limited sample data and low accuracy in tunnel lining disease detection is solved, and efficient and automated disease identification is achieved.
Patent Information
- Application Number
- CN202411216663.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2025-01-14
AI Technical Summary
The existing deep learning-based methods have problems such as limited sample data and uneven data quality in tunnel lining disease detection, resulting in low detection accuracy. The traditional methods rely on a large amount of manual annotation and experience, making it difficult to ensure efficiency and objectiveness.
Using a DRCNN model based on self-supervised learning network, extract the features of ground-penetrating radar images through self-supervised comparative learning network, build a disease detection network, use label-free data for pre-training, and optimize feature extraction with ROI Align Pooling to reduce manual labeling costs and improve Detection accuracy.
Effectively utilize a large amount of unlabeled radar data, improve the accuracy of detection of cracks, air leakage and other diseases, enhance the robustness and adaptability of the model, and realize the automation and accurate identification of tunnel lining diseases.
Smart Images
Figure CN119313936A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the fields of target classification, machine vision and intelligent recognition, and in particular to a method, device and medium for rapid intelligent recognition of tunnel defects. Background Art
[0002] Tunnel lining, as the core component of tunnel structure, is mainly constructed of concrete, supplemented by steel bars and steel arches to enhance its stability. However, this type of structure often faces typical defects such as voids, insufficient density, crack expansion and water leakage, which seriously threaten the safety and durability of the tunnel. Therefore, continuous monitoring of tunnel lining quality from the construction period to the operation period is particularly critical. Given the non-destructive advantages of non-destructive testing technology, it occupies an important position in tunnel lining quality inspection. Among them, ground penetrating radar (GPR) technology has become the mainstream choice due to its efficient and accurate scanning capabilities. GPR captures and analyzes the electromagnetic wave reflection signals inside the lining to draw detailed radar images to intuitively reflect the health of the lining.
[0003] However, faced with the massive amount of GPR data generated by long tunnels, traditional manual interpretation methods are not only time-consuming and laborious, but also highly dependent on the experience and skills of experts, making it difficult to ensure efficiency and objectivity. To overcome this bottleneck, automated and intelligent disease identification technologies have emerged, especially those based on deep learning, which have shown great potential.
[0004] In recent years, although machine learning technologies such as neural networks, support vector machines and clustering algorithms have been applied, they are still insufficient in processing complex GPR data. In contrast, deep learning, especially convolutional neural networks (CNN), stands out in GPR tunnel lining disease detection with its powerful feature extraction capabilities. It is completely feasible to use deep learning methods to detect tunnel lining diseases with GPR data. However, there are still some problems in practical work: (1) Most of the ground penetrating radar inspection data of tunnel linings do not contain defects.
[0005] (2) Labeling disease samples requires a lot of time.
[0006] (3) Directly migrating the parameters of a network (e.g., weights pre-trained using the ImageNet dataset) to a radar dataset does not improve the detection capability of the neural network. Summary of the invention
[0007] The purpose of the present invention is to provide a method, device and medium for rapid intelligent identification of tunnel diseases in order to solve the problem that the existing tunnel lining disease detection technology based on deep learning methods often has low detection accuracy of the model due to limited sample data or uneven data quality when dealing with disease identification in complex environments.
[0008] The above-mentioned purpose of the present application is achieved through the following technical solutions: S1: Obtain unlabeled ground penetrating radar images and perform preprocessing to build a sample library; S2: constructing a DRCNN model based on self-supervision; the DRCNN model includes: a self-supervised contrastive learning network and a disease detection network; training the self-supervised contrastive learning network through a sample library, extracting tunnel ground penetrating radar wave features of the ground penetrating radar images in the sample library, and obtaining feature extraction network parameters; S3: Migrate the feature extraction network parameters to the backbone feature module of the disease detection network; S4: Screen and mark the damage radar response data in the sample library to construct a tunnel damage ground penetrating radar image set; The defect detection network is trained based on the tunnel defect ground penetrating radar image set; S5: Obtain the ground penetrating radar image of the tunnel disease to be detected and input it into the trained disease detection network to identify the tunnel disease results in the ground penetrating radar image of the tunnel disease to be detected.
[0009] Optionally, step S1 includes: S11: preprocessing the ground penetrating radar image, including: removing direct waves, filtering and gain; S12: The preprocessed ground penetrating radar image is cropped into data blocks of a preset size through a sliding window to build a sample library.
[0010] Optionally, the self-supervised contrastive learning network includes: a backbone feature extraction module, a first global projection head, a second global projection head, a first dense projection head and a second dense projection head; the backbone feature extraction module adopts a residual network structure ResNet50, including: a first backbone feature extraction branch and a second backbone feature extraction branch.
[0011] Optionally, the first global projection head and the first dense projection head are connected in parallel to the first trunk feature extraction branch; The second global projection head and the second dense projection head are connected in parallel to the second trunk feature extraction branch; The first global projection head and the second global projection head are set as fully connected layers, and the first dense projection head and the second dense projection head are set as 1×1 convolutional layers.
[0012] Optionally, the loss function of the self-supervised contrastive learning network includes: a global contrastive loss function and a dense contrastive loss function; the global contrastive loss function is used to calculate the output value loss of the first global projection head and the second global projection head; the dense contrastive loss function is used to calculate the output value loss of the first dense projection head and the second dense projection head.
[0013] Optionally, step S3 includes: The disease detection network includes: a backbone feature module, a candidate frame generation module PRN, a ROI Align Pooling module, a fully connected module FC1, a fully connected module FC2, a fully connected module FC3, and a fully connected module FC4; The backbone feature module adopts the residual network structure ResNet50; The candidate frame generation module PRN adopts the Region Proposal Network network structure; The ROI Align Pooling module is connected to the candidate frame generation module PRN; the fully connected module FC1 is connected to the ROI Align Pooling module; the fully connected module FC1 is connected to the fully connected module FC2; the fully connected module FC3 and the fully connected module FC4 are connected to the fully connected module FC2; The fully connected module FC3 outputs the classification results of the diseases and the probability of correct classification; The fully connected module FC4 outputs the defect positioning result; the tunnel defect result includes: tunnel defect positioning result, classification result and classification correctness probability.
[0014] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes a method for rapid intelligent identification of tunnel diseases.
[0015] A computer-readable storage medium stores instructions, and when the instructions are executed, a method for quickly and intelligently identifying tunnel diseases is executed.
[0016] The beneficial effects of the technical solution provided by this application are: The construction of a self-supervised pre-training model (DRCNN model) based on deep learning technology can effectively utilize a large amount of unlabeled radar data, reduce labor costs and improve the accuracy of various defects detection such as cracks, voids, looseness, etc. A self-supervised pre-training deep learning model (DenseCL Faster-RCNN, DRCNN) is proposed, which consists of two parts: a self-supervised contrastive learning network and a defect detection network. The network model has the following characteristics: (1) Through self-supervised pre-training, the network model can learn effective feature representations of tunnel linings and defects on a large number of unlabeled GPR images, thereby improving the recognition accuracy of tunnel defects; (2) After the initial training, the network model is further trained with GPR images of labeled defect samples, which enables it to quickly identify defects and easily adapt to the recognition of new defect types, thereby enhancing the robustness of the model; (3) The network model optimizes the feature extraction method of the target area and replaces the traditional ROI Pooling with ROI Align Pooling, making the feature extraction process more accurate and reducing quantization errors, thereby further improving the model's defect detection accuracy and realizing the automatic and accurate recognition of tunnel lining defects. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The present application will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a step diagram in an embodiment of the present application; Figure 2 is a structural diagram of a DenseCL network model in an embodiment of the present application; Figure 3 It is a structural diagram of the FasterRCNN network model in the embodiment of the present application; Figure 4 is a tunnel disease identification result diagram of the DRCNN method in the embodiment of the present application; Figure 5 is an example diagram of a sample of a ground penetrating radar image of a tunnel disease in an embodiment of the present application; Figure 6 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION
[0018] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.
[0019] The embodiments of the present application provide a method for rapid and intelligent identification of tunnel defects.
[0020] Please refer to Figure 1 , Figure 1: is a step diagram of a method for rapid intelligent identification of tunnel diseases in an embodiment of the present application, comprising: S1: Obtain unlabeled ground penetrating radar images and perform preprocessing to build a sample library; S2: constructing a DRCNN model based on self-supervision; the DRCNN model includes: a self-supervised contrastive learning network and a disease detection network; training the self-supervised contrastive learning network through a sample library, extracting tunnel ground penetrating radar wave features of the ground penetrating radar images in the sample library, and obtaining feature extraction network parameters; S3: Migrate the feature extraction network parameters to the backbone feature module of the disease detection network; S4: Screen and mark the damage radar response data in the sample library to construct a tunnel damage ground penetrating radar image set; The defect detection network is trained based on the tunnel defect ground penetrating radar image set; S5: Obtain the ground penetrating radar image of the tunnel disease to be detected and input it into the trained disease detection network to identify the tunnel disease results in the ground penetrating radar image of the tunnel disease to be detected.
[0021] Specifically, the tunnel disease GPR image set needs to be obtained through manual interpretation of GPR images. The interpretation results should include two important information: the type and location of the disease, such as Figure 5 The type of tunnel disease is determined by experts. Common tunnel disease types include voids, looseness, cracks, and voids, which are coded with numbers 1, 2, 3, etc. The location of the tunnel disease is defined by a rectangular box, and the coordinates of the upper left corner of the rectangular box are defined as (x1, y1), and the lower right corner is defined as (x2, y2).
[0022] The specific implementation process of the present invention is as follows: (1) Collect a large number of ground penetrating radar images and perform de-direction, filtering, and gain processing on the ground penetrating radar images.
[0023] (2) Use a 512×800 window to slide along the profile of the GPR image to divide the GPR image into a large number of 512×800 data slices; the distance of each window sliding is 1 / 4 of the window length, that is, 128 (or 1 / 2 of the window length = 256); the data obtained by sliding the window is composed of an unlabeled sample library.
[0024] (3) According to Figure 2 The structure shown creates a self-supervised DenseCL neural network model, i.e., a self-supervised contrastive learning network; (4) Train a self-supervised contrastive learning network using an unlabeled sample library; (5) According to Figure 3The structure shown creates a self-supervised FasterRCNN neural network model, i.e., a disease detection network; (6) The results (feature extraction network parameters) of the “backbone feature 1” extraction network module (ResNet50) obtained by the DenseCL neural network model are transferred to the “backbone feature 1” module in the disease detection network (FasterRCNN) (e.g. Figure 3 shown); (7) Manually select the defect radar response data from the unlabeled data set and label them to construct a tunnel defect ground penetrating radar image set. The defect labeling method is as follows: (1) determine the defect type, for example, void is 1, looseness is 2, crack is 3, etc.; (2) determine the location and range of the defect on the ground penetrating radar data ( Figure 5 ). The location and range of the disease can be marked by the coordinates of the upper left corner (x1, y1) and the lower right corner (x2, y2) of a rectangular box.
[0025] (8) Set the training parameters of the FasterRCNN neural network model, including the number of training times, learning rate, and batch size; use the labeled training set data to train the improved FasterRCNN network.
[0026] (9) Reuse the test set data to test the performance of the network model; adjust the hyperparameters of the network model based on the test results so that the model has the best performance and the training is completed.
[0027] (10) Using the ground penetrating radar image of the tunnel disease to be detected and inputting it into the trained network model, we can obtain the following Figure 4 Results shown.
[0028] The DRCNN model of this application consists of two core parts: the feature extraction network DenseCL and the disease detection network Faster-RCNN. The training process is divided into two stages: first, using the self-supervised learning ability of DenseCL, a large amount of unlabeled GPR data (ground penetrating radar images) is pre-trained to extract and optimize image features; then, the pre-trained feature extraction network parameters are transferred to the Faster-RCNN network, and the network is fine-tuned based on a small amount of labeled data to achieve accurate detection of diseases. On the basis of Faster-RCNN, DRCNN further optimizes the network structure and introduces the ROIAlign pooling method. Compared with the traditional ROI Pooling, it can more accurately retain the spatial position information in the feature map, thereby improving the accuracy of detection. In addition, by fusing the underlying features, DRCNN can more comprehensively capture the subtle changes in GPR images and provide more abundant clues for disease identification. With its unique self-supervised pre-training strategy and optimized network structure, the DRCNN model can effectively solve the data labeling problems and detection accuracy problems in tunnel lining disease detection, and provide strong technical support for the safety management of tunnel engineering. Through practical verification, DRCNN has demonstrated excellent performance in tunnel lining disease detection and can open up new paths for the development of future intelligent detection technology.
[0029] Step S1 includes: S11: preprocessing the ground penetrating radar image, including: removing direct waves, filtering and gain; S12: The preprocessed ground penetrating radar image is cropped into data blocks of a preset size through a sliding window to build a sample library.
[0030] The self-supervised contrastive learning network includes: a backbone feature extraction module, a first global projection head, a second global projection head, a first dense projection head and a second dense projection head; the backbone feature extraction module adopts a residual network structure ResNet50, including: a first backbone feature extraction branch and a second backbone feature extraction branch.
[0031] The first global projection head and the first dense projection head are connected in parallel to the first trunk feature extraction branch; The second global projection head and the second dense projection head are connected in parallel to the second trunk feature extraction branch; The first global projection head and the second global projection head are set as fully connected layers, and the first dense projection head and the second dense projection head are set as 1×1 convolutional layers.
[0032] Specifically, through the self-supervised contrastive learning network, it is transformed from the dense contrastive network module DenseCL, such as Figure 2 shown.
[0033] The loss function of the self-supervised contrastive learning network includes: a global contrastive loss function and a dense contrastive loss function; the global contrastive loss function is used to calculate the output value loss of the first global projection head and the second global projection head; the dense contrastive loss function is used to calculate the output value loss of the first dense projection head and the second dense projection head.
[0034] Specifically, the self-supervised contrastive learning network (DenseCL) does not require labeled sample data for training. The feature extraction network is updated through the global contrast loss function and the dense contrast loss function to achieve the same features being close to each other and different features being far away from each other.
[0035] Specifically, the color transformation method of the self-supervised contrastive learning network is pre-set, and the color transformation method includes: random horizontal flipping, random application of color jitter, random application of Gaussian blur, random grayscale, and standardization. Color transformation is mainly achieved through the public torchvision library function. After color transformation, a set of radar data is a positive sample of the same sample and a negative sample of other samples.
[0036] The global projection head is set to a fully connected layer with 128 neurons; the dense projection head is set to a 1×1 convolutional layer. The global feature obtained by the global projection head is ; Sample library The dense feature obtained by the dense projection head is ( ). The loss function is calculated using the following formula:
[0037] in, is the weight coefficient, which is taken as 0.5 in this application; Represents the global loss function; is a dense loss function. Satisfy the following formula:
[0038] in, Represents the current image features; Represents the positive sample pair image features of the current image; Represents the negative sample pair image features of the current image; is a hyperparameter and is set to 0.07.
[0039] Satisfy the following formula:
[0040] In dense feature comparison, the extracted feature vectors are passed through the dense projection head to obtain a set of dense feature vectors ( ). and Respectively represent the spatial size of the generated dense feature map, assuming , using a dense contrast loss function to compare the changes between dense feature vectors, where express The eigenvector at the sth spatial position among the eigenvectors, Representation and The corresponding positive sample pair feature vector, Representation and The corresponding negative sample pair feature vector.
[0041] Step S3 includes: The disease detection network includes: a backbone feature module, a candidate frame generation module PRN, a ROI Align Pooling module, a fully connected module FC1, a fully connected module FC2, a fully connected module FC3, and a fully connected module FC4; The backbone feature module adopts the residual network structure ResNet50; The candidate frame generation module PRN adopts the Region Proposal Network network structure; The ROI Align Pooling module is connected to the candidate frame generation module PRN; the fully connected module FC1 is connected to the ROI Align Pooling module; the fully connected module FC1 is connected to the fully connected module FC2; the fully connected module FC3 and the fully connected module FC4 are connected to the fully connected module FC2; The fully connected module FC3 outputs the classification results of the diseases and the probability of correct classification; The fully connected module FC4 outputs the defect positioning result; the tunnel defect result includes: tunnel defect positioning result, classification result and classification correctness probability.
[0042] Specifically, the Faster-RCNN network is improved to form a disease detection network, such as Figure 3 shown.
[0043] Specifically, the disease detection network uses a non-maximum suppression algorithm; the candidate box generation module is set to output 256 rectangular boxes and connected to the pooling layer ROI Align Pooling module; and then connected in series to two fully connected layers, the number of neurons in the first fully connected layer FC1 and the second fully connected layer FC2 are both 1024, such as Figure 3 shown.
[0044] The present application also discloses an electronic device. Figure 6 , Figure 6 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .
[0045] The communication bus 502 is used to realize the connection and communication between these components.
[0046] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.
[0047] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).
[0048] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for loading by a processor to execute the above-mentioned method for rapid intelligent identification of tunnel diseases.
[0049] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.
[0050] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.
Claims
1. A fast and intelligent method for identifying tunnel defects, characterized in that: The method comprises the following steps: S1: Obtain unlabeled ground penetrating radar images and perform preprocessing to build a sample library; S2: Build a DRCNN model based on self-supervision; The DRCNN model includes: a self-supervised contrastive learning network and a disease detection network; the self-supervised contrastive learning network is trained through a sample library, tunnel ground penetrating radar wave features of the ground penetrating radar image in the sample library are extracted, and feature extraction network parameters are obtained; S3: Migrate the feature extraction network parameters to the backbone feature module of the disease detection network; S4: Screen and label the damage radar response data in the sample library to construct a tunnel damage ground penetrating radar image set; train the damage detection network based on the tunnel damage ground penetrating radar image set; S5: Obtain the ground penetrating radar image of the tunnel disease to be detected and input it into the trained disease detection network to identify the tunnel disease results in the ground penetrating radar image of the tunnel disease to be detected.
2. A method for rapid intelligent identification of tunnel defects as claimed in claim 1, characterized in that: Step S1 includes: S11: preprocessing the ground penetrating radar image, including: removing direct waves, filtering and gain; S12: The preprocessed ground penetrating radar image is cropped into data blocks of a preset size through a sliding window to build a sample library.
3. A fast and intelligent identification method for tunnel defects as claimed in claim 1, characterized in that: The self-supervised contrastive learning network includes: a backbone feature extraction module, a first global projection head, a second global projection head, a first dense projection head and a second dense projection head; the backbone feature extraction module adopts a residual network structure ResNet50, including: a first backbone feature extraction branch and a second backbone feature extraction branch.
4. A fast and intelligent identification method for tunnel diseases as claimed in claim 3, characterized in that: The first global projection head and the first dense projection head are connected in parallel to the first trunk feature extraction branch; The second global projection head and the second dense projection head are connected in parallel to the second trunk feature extraction branch; The first global projection head and the second global projection head are set as fully connected layers, and the first dense projection head and the second dense projection head are set as 1×1 convolutional layers.
5. A fast and intelligent method for identifying tunnel defects as claimed in claim 4, characterized in that: The loss function of the self-supervised contrastive learning network includes: a global contrastive loss function and a dense contrastive loss function; the global contrastive loss function is used to calculate the output value loss of the first global projection head and the second global projection head; the dense contrastive loss function is used to calculate the output value loss of the first dense projection head and the second dense projection head.
6. A fast and intelligent method for identifying tunnel defects as claimed in claim 1, characterized in that: Step S3 includes: The disease detection network includes: a backbone feature module, a candidate frame generation module PRN, a ROI Align Pooling module, a fully connected module FC1, a fully connected module FC2, a fully connected module FC3, and a fully connected module FC4; The backbone feature module adopts the residual network structure ResNet50; The candidate frame generation module PRN adopts the Region Proposal Network network structure; The ROI Align Pooling module is connected to the candidate frame generation module PRN; the fully connected module FC1 is connected to the ROIAlign Pooling module; the fully connected module FC1 is connected to the fully connected module FC2; the fully connected module FC3 and the fully connected module FC4 are connected to the fully connected module FC2; The fully connected module FC3 outputs the classification results of the diseases and the probability of correct classification; The fully connected module FC4 outputs the defect positioning result; the tunnel defect result includes: tunnel defect positioning result, classification result and classification correctness probability.
7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.