A railway tunnel construction leakage risk identification and disposal method and device

Abstract

The present application provides a kind of railway tunnel construction leakage risk identification disposal method and device, and the image to be analyzed is reconstructed after being transmitted by being stored after data block, block compression, relationship mapping and coding by preset link;Enhancement and denoising processing are carried out on the image using median filtering and homomorphic filtering;A railway tunnel construction seepage risk identification model based on multiple encoders and decoders is constructed to identify the location and risk level of seepage, the encoder uses multi-layer convolution and hollow pyramid pooling operation to extract image features, the output of the corresponding stage encoder is input into the decoder through jump connection, and the jump connection integrates a self-attention mechanism module.The present application can efficiently and accurately identify the location and risk level of railway tunnel construction seepage, effectively improve the monitoring efficiency and accuracy, realize the overall coverage of the whole operation surface, and is not affected by the construction environment, which provides a strong guarantee for the safety of railway tunnel construction.

Description

Technical Field

[0001] This invention relates to the field of railway engineering risk detection technology, and in particular to a method and device for identifying and handling water leakage risks during railway tunnel construction. Background Technology

[0002] Water leakage is a significant factor affecting tunnel structural stability and construction safety during railway tunnel construction. Current methods for monitoring and identifying water leakage risks in railway tunnel construction have several limitations. On the one hand, while manual inspections can visually detect leakage, they are inefficient, the results are heavily influenced by personnel experience and subjective factors, and they lack comprehensive and detailed detection, easily overlooking potential leakage areas. On the other hand, while physical sensor-based monitoring methods can provide localized data, they can only monitor specific areas and cannot fully cover the entire work surface. Installation and maintenance costs are high, and in complex construction environments, the accuracy and reliability of sensors are easily affected, leading to inaccurate monitoring results. Furthermore, traditional image recognition methods suffer from poor adaptability and low accuracy in identifying water leakage risks in railway tunnel construction. They struggle to cope with complex lighting conditions and background interference within tunnels, and the manual feature extraction process is cumbersome and inaccurate, prone to misjudgments and omissions. Therefore, there is an urgent need for an efficient, accurate, and adaptable image recognition method and device for identifying water leakage risks in railway tunnel construction to overcome the shortcomings of existing technologies and meet practical engineering needs. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and apparatus for identifying and handling water leakage risks in railway tunnel construction, so as to eliminate or improve one or more defects existing in the prior art, and solve the problems of low efficiency, poor accuracy, difficulty in fully covering the work surface, and susceptibility to interference from the construction environment in existing railway tunnel construction water leakage risk monitoring and identification methods.

[0004] One aspect of the present invention provides a method for identifying and handling water leakage risks during railway tunnel construction, the method comprising the following steps:

[0005] Images to be analyzed, collected at the railway tunnel construction site, are acquired through a preset link. These images are obtained by transmitting and reconstructing data after being segmented into blocks, compressed within blocks, mapped onto relationships, and encoded for storage.

[0006] Median filtering and homomorphic filtering are applied to the image to be analyzed to enhance the image and denoise it, thereby obtaining an intermediate image.

[0007] The intermediate image is input into a pre-trained railway tunnel construction seepage risk identification model to output the identification results of seepage location and seepage risk level; the railway tunnel construction seepage risk identification model is based on downsampling by multiple encoders, the encoders use multi-layer convolution and hollow pyramid pooling operations to extract image features, and upsampling is performed by multiple decoders. The input of the decoder is connected to the output of the encoder at the corresponding level, and a self-attention mechanism module is integrated at the connection point.

[0008] The hollow pyramid pooling includes three branches: the first branch contains a 1×1 convolutional layer, the second branch contains multiple dilated convolutional layers with varying dilation rates, and the third branch contains a 1×1 pooling layer, a 1×1 convolutional layer, and an upsampling layer. The outputs of the first, second, and third branches are concatenated along the channel dimension and then fused and dimensionality-reduced by a 1×1 convolutional layer to output the image features. The self-attention mechanism module downsamples the first intermediate feature of the multi-layer convolution in the encoder and the output of the upsampling layer in the hollow pyramid pooling and performs an XOR operation to obtain a second intermediate feature. After performing self-attention and resampling on the second intermediate feature, it is multiplied with the first intermediate feature to obtain a third intermediate feature. The upsampled image features, the outputs of the multi-layer convolution in the encoder, and the third intermediate feature are aggregated, upsampled, and then input into the corresponding level of the decoder to achieve skip connections.

[0009] In some embodiments, the processing steps during the transmission of the image to be analyzed include:

[0010] The image to be analyzed is divided into multiple non-overlapping pixel sub-blocks according to a set size;

[0011] A discrete cosine transform is performed on each pixel sub-block to convert the data from the spatial domain to the frequency domain. The converted data is then quantized and encoded to remove redundant information and achieve compression.

[0012] For each pixel sub-block, find similar pixel block pairs that meet the set similarity criteria within the entire range of the image to be analyzed, and establish a mapping relationship between the similar pixel block pairs. The mapping relationship includes scaling ratio, rotation angle, and translation amount.

[0013] After each pixel sub-block is compressed and encoded based on the mapping relationship, it is then transmitted using the Real-Time Streaming Protocol (RTSP) for decompression and reconstruction of the image to be analyzed.

[0014] In some embodiments, median filtering and homomorphic filtering are performed on the image to be analyzed to enhance and denoise the image and obtain an intermediate image, including:

[0015] A filter window of a set size is used to scan the image to be analyzed row by row and column by column; gray values ​​are extracted from the pixels extracted by the filter window each time, and arranged in ascending or descending order; the value of the center pixel in the filter window is replaced by the middle value after the arrangement.

[0016] The image to be analyzed after median filtering is represented as the product of the illumination component and the reflection component, and logarithmic and Fourier transforms are performed. A homomorphic filter is created to filter the image based on the frequency characteristics of the illumination component and the reflection component. The filtered frequency domain image is converted back to the spatial domain through inverse Fourier transform, and then the enhanced image is obtained through exponential transform.

[0017] In some embodiments, the pre-training steps of the railway tunnel construction seepage risk identification model include:

[0018] Multiple sample images were acquired for railway tunnel construction scenarios. Each sample image was processed by median filtering and homomorphic filtering, and the location of the water seepage area and the corresponding risk level were marked as labels to construct a training sample set.

[0019] The training sample set is used to train the railway tunnel construction seepage risk identification model in its initial state. The sample images processed by median filtering and homomorphic filtering are input into the railway tunnel construction seepage risk identification model in its initial state, and the model outputs predicted values ​​for the location and risk level of the seepage area. A loss function is constructed based on the deviation between the predicted values ​​and the labels, and the parameters of the railway tunnel construction seepage risk identification model are updated.

[0020] The method employs the Adam optimizer to train the railway tunnel construction seepage risk identification model, with the learning rate decreasing via cosine annealing.

[0021] In some embodiments, the backbone network of the railway tunnel construction seepage risk identification model adopts the MobileNetV3_small network to achieve lightweight deployment.

[0022] In some embodiments, the method further includes: performing morphological operations on the identification results of water seepage locations in the image to be analyzed, filling small holes and narrow discontinuities in the image by dilation, and eliminating bumps and noise in the image by erosion.

[0023] In some embodiments, the method further includes: updating the acquired image to be analyzed in real time at a set frequency and updating the identification results of the seepage location and seepage risk level; generating early warning information and providing audio-visual prompts for the identification results where the seepage risk level is higher than a set standard.

[0024] On the other hand, the present invention also provides a device for identifying and handling water leakage risks during railway tunnel construction, the device comprising:

[0025] Multiple image acquisition and transmission modules are deployed at predetermined locations in the railway tunnel construction site. The image acquisition and transmission modules acquire images of the designated construction locations according to a predetermined format, and perform data segmentation, intra-block compression, relation mapping, and encoding storage on the acquired images to be analyzed, and transmit them through a preset link.

[0026] An image preprocessing module is used to receive the image to be analyzed and perform median filtering and homomorphic filtering on the image to be analyzed to enhance the image and remove noise to obtain an intermediate image;

[0027] The image recognition module is used to execute the above-mentioned method for identifying and handling water leakage risks in railway tunnel construction, and output the identification results of water leakage location and water leakage risk level;

[0028] The early warning display module is used to generate early warning information and provide audio-visual prompts when the identification result of the water seepage risk level being higher than the set standard.

[0029] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0030] On the other hand, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0031] The method and apparatus for identifying and handling water leakage risks in railway tunnel construction as described in this invention acquires and reconstructs images to be analyzed after data segmentation, intra-block compression, relation mapping, and encoding storage via a preset link. Median filtering and homomorphic filtering are used to enhance and denoise the images. A railway tunnel construction water leakage risk identification model based on multiple encoders and decoders is constructed to identify the location and risk level of water leakage. The encoders use multi-layer convolution and hollow pyramid pooling operations to extract image features. The decoder input skips the output of the corresponding level encoder, and a self-attention mechanism module is integrated at the skip connection. This invention can efficiently and accurately identify the location and risk level of water leakage in railway tunnel construction, effectively improving monitoring efficiency and accuracy, achieving comprehensive coverage of the entire work surface, and is unaffected by the construction environment, providing strong protection for the safety of railway tunnel construction.

[0032] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0033] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0035] Figure 1 This is a flowchart illustrating the method for identifying and handling water leakage risks during railway tunnel construction according to an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the working process of the railway tunnel construction water leakage risk identification and handling device according to an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the segmented compression technology used in the railway tunnel construction water leakage risk identification and treatment device according to an embodiment of the present invention.

[0038] Figure 4 This is a schematic diagram of the filtering process used in the railway tunnel construction water leakage risk identification and handling device according to an embodiment of the present invention.

[0039] Figure 5 This is a schematic diagram of the RT-Unet model used in the railway tunnel construction water leakage risk identification and treatment device according to an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0041] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0042] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0043] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0044] To address the problems existing in the prior art, the present invention provides a method for identifying and handling water leakage risks during railway tunnel construction, the method comprising the following steps S101~S103:

[0045] Step S101: Obtain the images to be analyzed collected at the railway tunnel construction site through a preset link. The images to be analyzed are obtained by transmitting and reconstructing after data segmentation, intra-block compression, relation mapping and encoding storage.

[0046] Step S102: Perform median filtering and homomorphic filtering on the image to be analyzed to enhance the image and denoise it to obtain an intermediate image.

[0047] Step S103: Input the intermediate image into the pre-trained railway tunnel construction seepage risk identification model to output the identification results of seepage location and seepage risk level; the railway tunnel construction seepage risk identification model is based on downsampling by multiple encoders. The encoders use multi-layer convolution and hollow pyramid pooling operations to extract image features, and perform upsampling through multiple decoders. The input of the decoder is connected to the output of the corresponding layer encoder, and the self-attention mechanism module is integrated at the jump connection.

[0048] The hollow pyramid pooling method comprises three branches: the first branch contains a 1×1 convolutional layer, the second branch contains multiple dilated convolutional layers with varying dilation rates, and the third branch contains a 1×1 pooling layer, a 1×1 convolutional layer, and an upsampling layer. The outputs of the first, second, and third branches are concatenated along the channel dimension and then fused and reduced in dimensionality by a 1×1 convolutional layer to output image features. The self-attention mechanism module downsamples the first intermediate features of the multi-layer convolutions in each level of the encoder and the output of the upsampling layer in the hollow pyramid pooling method, and performs an XOR operation to obtain the second intermediate feature. The second intermediate feature is then multiplied by the first intermediate feature after performing the self-attention mechanism and resampling to obtain the third intermediate feature. The upsampled image features, the outputs of the multi-layer convolutions in the encoder, and the third intermediate feature are aggregated, upsampled, and then input into the corresponding level of the decoder to achieve skip connections.

[0049] In step S101, images of the railway tunnel construction site to be analyzed are acquired through a preset link. These images are transmitted after data segmentation, intra-block compression, relation mapping, and encoding storage, and are finally reconstructed. This processing method can effectively reduce the bandwidth pressure during the transmission of high-definition images and ensure reliable transmission of image data in the network.

[0050] In some embodiments, the processing steps during the transmission of the image to be analyzed include steps S1011 to S1014:

[0051] Step S1011: Divide the image to be analyzed into multiple non-overlapping pixel sub-blocks according to the set size.

[0052] Step S1012: Perform discrete cosine transform on each pixel sub-block to convert the data from the spatial domain to the frequency domain, and quantize and encode the converted data to remove redundant information and achieve compression.

[0053] Step S1013: For each pixel sub-block, find similar pixel block pairs that meet the set similarity criteria within the entire range of the image to be analyzed, and establish a mapping relationship between similar pixel block pairs. The mapping relationship includes scaling ratio, rotation angle and translation amount.

[0054] Step S1014: After compressing and encoding each pixel sub-block based on the mapping relationship, the image to be analyzed is decompressed and reconstructed after transmission based on the Real-Time Streaming Protocol (RTSP).

[0055] During the image acquisition phase, multiple high-definition image acquisition units, each equipped with high-resolution and wide-angle lenses, are strategically deployed on the working face within the tunnel. These units utilize built-in photosensors to monitor the ambient light intensity in real time, automatically calculating and adjusting the optimal exposure time and focus distance based on the light intensity. This ensures clear, high-quality images are acquired even in various complex environments. The images are of two types: still images and videos. Still images are stored in .jpg format, and videos in .mp4 format.

[0056] To alleviate bandwidth pressure caused by large data volumes during high-definition image transmission, an image block compression technique based on fractal theory is proposed. This technique mainly involves data block division, intra-block compression, relational mapping, and encoded storage. The specific operation process is as follows:

[0057] 1) Data segmentation: Divide the image to be compressed into several non-overlapping pixel sub-blocks. These sub-blocks can be set to a fixed size, such as 2×2, 4×4, 8×8, etc., to facilitate finding self-similarity within a local area. Pixel sub-blocks It can be defined as:

[0058] ; (1)

[0059] in, represent An identity matrix of order 1, satisfying , To initialize the matrix , To initialize the matrix, the row and column transpose matrix, To initialize the diagonal transpose of the matrix, the number of rows and columns in the matrix is... One, and .

[0060] 2) Intra-block compression: For each pixel sub-block The Discrete Cosine Transform (DCT) is performed to transform the data from the spatial domain to the frequency domain, so that the energy of the data is mainly concentrated on a few low-frequency coefficients. Then, these coefficients are quantized and encoded to remove redundant information in the data, thereby achieving the purpose of compression.

[0061] 3) Relationship mapping: For each pixel sub-block To find other blocks that are similar to it within the entire image or a specific search range. Record the mapping relationship between them. This includes parameters such as scaling ratio, rotation angle, and translation amount.

[0062] 4) Encoding and storage: Mapping the relationship of each pixel sub-block It also encodes and stores some necessary auxiliary information (such as the location of the block, whether it is a boundary block, etc.).

[0063] Then, the encoded and compressed image data is transmitted to the edge computing device via the Real-Time Streaming Protocol (RTSP) to ensure reliable transmission of image data over the network. Upon receiving the encoded and compressed image data, the edge computing device, during decompression, reconstructs the original image by iteratively rebuilding individual sub-blocks based on stored mapping relationships and parameters, thereby reconstructing the entire image.

[0064] In step S102, the acquired image to be analyzed is subjected to median filtering and homomorphic filtering to achieve image enhancement and denoising, resulting in an intermediate image. Median filtering can remove noise from the image, while homomorphic filtering can enhance the feature information of the image, improve image quality, and lay the foundation for the subsequent recognition process.

[0065] In some embodiments, median filtering and homomorphic filtering are performed on the image to be analyzed to enhance and denoise the image and obtain an intermediate image, including steps S1021~S1022:

[0066] Step S1021: Scan the image to be analyzed row by row and column by column using a filter window of a set size; extract the gray value of each pixel extracted by the filter window and arrange them in ascending or descending order, and replace the value of the center pixel in the filter window with the middle value after the arrangement.

[0067] Step S1022: Represent the image to be analyzed after median filtering as the product of the illumination component and the reflection component, and perform logarithmic transformation and Fourier transformation. Create a homomorphic filter and perform filtering based on the frequency characteristics of the illumination component and the reflection component. Convert the filtered frequency domain image back to the spatial domain through inverse Fourier transform, and then obtain the enhanced image through exponential transformation.

[0068] Considering the complex lighting conditions inside the tunnel, the acquired images need to be preprocessed. A hybrid algorithm combining median filtering and homomorphic filtering is used for image enhancement and denoising, thereby enhancing the feature information of the seepage area. Specifically, the steps include:

[0069] 1) Perform median filtering on the original image, first using a... A sliding window of pixels scans the image row by row and column by column, and the pixel values ​​scanned within the window are denoted as a set. ,but It can be defined as:

[0070] ; (2)

[0071] 2) Extraction The grayscale values ​​of each pixel are denoted as a set. ,but It can be defined as:

[0072] ; (3)

[0073] 3) Sort the grayscale values ​​by size and use the median value after sorting. Replace the pixel value of the center pixel of the window to complete the filtering operation at that point. Continue moving the window and performing the above operation on each pixel of the image until the entire image has undergone median filtering.

[0074] 4) Apply homomorphic filtering to the median-filtered image, representing the image as the product of the illumination and reflection components, which can be defined as:

[0075] ; (4)

[0076] in, This is the image after median filtering. It is the light component. It is the reflection component.

[0077] 5) Input image Perform a logarithmic transformation to obtain .

[0078] 6) Transforming from the spatial domain to the frequency domain, we obtain Through Fourier transform, we obtain .in This indicates the Fourier transform operation.

[0079] 7) Design homomorphic filters Based on the frequency characteristics of illumination and reflection components, for After filtering, the result is obtained. .

[0080] 8) Filter the frequency domain image Transform back into the spatial domain to obtain ,in This indicates the inverse Fourier transform operation.

[0081] 9) To Perform an exponential transform to obtain the enhanced image. .

[0082] By following the steps above, noise in the image can be removed to the greatest extent while preserving edge information, further improving image quality and helping to improve the accuracy of subsequent recognition.

[0083] In step S103, the preprocessed intermediate image is input into a pre-trained railway tunnel construction seepage risk identification model. This model is built based on multiple encoders and decoders and can output the identification results of seepage location and seepage risk level.

[0084] In some embodiments, the pre-training steps of the railway tunnel construction seepage risk identification model include steps S1031~S1032:

[0085] Step S1031: Acquire multiple sample images for the railway tunnel construction scenario. Each sample image is processed by median filtering and homomorphic filtering, and the location of the water seepage area and the corresponding risk level are marked as labels to construct a training sample set.

[0086] Step S1032: Train the railway tunnel construction seepage risk identification model in the initial state using the training sample set. Input the sample images processed by median filtering and homomorphic filtering into the railway tunnel construction seepage risk identification model in the initial state and output the predicted values ​​of the seepage area location and risk level. Construct a loss function based on the deviation between the predicted value and the label, and update the parameters of the railway tunnel construction seepage risk identification model.

[0087] The method employs the Adam optimizer to train the railway tunnel construction seepage risk identification model, with the learning rate decreasing via cosine annealing. In some embodiments, the backbone network of the railway tunnel construction seepage risk identification model uses a MobileNetV3_small network for lightweight deployment.

[0088] In the image recognition inference part, this invention constructs an RT-Unet model that has been pre-trained and optimized with a large amount of real tunnel construction image data. This model is a specific deep learning semantic segmentation model that is implemented by integrating a self-attention module (SAM) on the basis of the traditional convolutional neural network Unet architecture.

[0089] The RT-Unet model architecture consists of an encoding part and a decoding part.

[0090] The encoding part progressively extracts image features through multiple convolutional and pooling operations. As the number of layers increases, the number of feature maps gradually increases, while the spatial resolution gradually decreases. Each convolutional layer contains multiple convolutional kernels, which perform convolution operations on the image using a sliding window to extract different features such as edges and textures. After feature extraction at each convolutional layer, pooling layers are used for downsampling to reduce the image size while increasing the feature dimension. The pooling method employs Dilated Pyramid Pooling (ASPP) to reduce the resolution of the feature maps, decrease the amount of data and computation, and retain the main features.

[0091] First, a convolution operation is performed on the input feature map, and the size of the output feature map is calculated using the following formula:

[0092] ; (5)

[0093] in, and These are the row and column pixel values ​​of the input feature map, respectively. It's the fill size. It is the kernel size. It is the convolution stride.

[0094] Then, pooling is performed on the feature map output by the convolution. The formula for calculating the size of the pooled feature map is as follows:

[0095] ; (6)

[0096] in, and It is the size of the feature map after convolution. It is the pooling window size. It is the pooling step size.

[0097] The decoding part uses deconvolution and skip connections, combined with the self-attention mechanism module SAM, to restore the feature map to the original image size, achieving pixel-level image segmentation.

[0098] The principle of the attention mechanism is to assign different weights to different regions of the input feature map, thereby focusing more on key features in the image and improving the model's sensitivity to important information. The Self-Attention (SAM) module calculates the correlation between pixels in the input sequence, assigns attention weights to each pixel, and generates the output through weighted summation, thus focusing on the important features. The calculation process is as follows:

[0099] 1) Input processing: The deconvolutional layer and skip connection layer of the decoding part of the RT-Unet model are used as input sequences, and vector expressions are generated through the embedding layer, such as vector A=[a1,a2,a3,…,an].

[0100] 2) Generate Q, K, and V vectors. Each input vector in the above vector expression is transformed into a query, key, and value vector through three linear transformations, as shown in the formula:

[0101] , , ; (7)

[0102] in, It is a trainable parameter matrix.

[0103] 3) Calculate the attention score by using the dot product to calculate the similarity between the query and the key, and then divide by... The weights are obtained by scaling with a stable gradient and then normalizing using Softmax.

[0104] ; (8)

[0105] Finally, the output values ​​are weighted and summed to generate the output, reflecting the correlation between elements.

[0106] During model pre-training, data diversity was increased by adjusting hyperparameters such as the learning rate and optimizer, and by using data augmentation techniques such as flipping, rotating, and scaling. The learning rate was adjusted based on the changes in the loss function during training, with the maximum learning rate set to 1×10⁻⁴ and the minimum learning rate set to 0.01 of the maximum. The Adam optimizer was used, with cosine annealing as the learning rate descent method. Different gradient calculation and parameter update strategies were employed to optimize the model and improve its generalization ability.

[0107] For the output of the model inference, a multi-threshold judgment and post-processing algorithm is employed. By setting multiple different thresholds, the probability map output by the model is segmented and classified to accurately identify the leakage areas. The threshold settings are based on the optimal range derived from a large amount of experimental data and statistical analysis, including the following three intervals:

[0108] A. Low threshold (0.1~0.2): Used for initial screening of areas that may have water leakage. A lower threshold allows the model to capture as many potential water leakage areas as possible and avoid omissions, but it may introduce more false positives, identifying some non-leakage areas as water leakage areas.

[0109] B. Medium Threshold (0.3~0.5): Based on the initial screening, further identify the more likely areas of leakage. This threshold range can filter out some areas that are obviously not leaking, while retaining most of the real leaking areas, thus balancing recall and precision.

[0110] C. High threshold (0.6~0.9): Used to definitively identify leaks with high certainty. A high threshold ensures high accuracy in identifying leaks, but may miss some edge or less obvious leaks.

[0111] Meanwhile, the segmentation results are optimized by incorporating morphological operations such as dilation and erosion after threshold judgment. Dilation can fill small holes and narrow discontinuities in the image, while erosion can eliminate small bumps and noise, removing isolated noise points and small regions, making the recognition results more accurate and reliable.

[0112] The RT-Unet model is deployed on customized edge computing hardware. The backbone network uses MobileNetV3_small, which has been trained for hundreds of rounds. This greatly reduces the number of model parameters and computational load, while facilitating lightweight deployment on resource-constrained edge computing devices. It improves the inference speed of the model while ensuring a certain level of segmentation accuracy, thus meeting the high real-time requirements of field application scenarios.

[0113] In some embodiments, the method further includes: updating the acquired image to be analyzed in real time at a set frequency and updating the identification results of the seepage location and seepage risk level; generating early warning information and providing audio-visual prompts for the identification results where the seepage risk level is higher than a set standard.

[0114] Furthermore, when constructing the training sample set to set the seepage risk level, the risk data output by the image recognition module and relevant data such as the tunnel's geographical information and construction progress are effectively integrated. The LEC quantitative evaluation method for operational condition hazards is adopted, as shown in Tables 1 and 2. A four-color grading standard for risk level is formulated. Based on the seepage area identified by the image recognition module, the water accumulation area and water flow size are determined, and a mapping relationship of risk level is established.

[0115] Table 1

[0116]

[0117] Table 2

[0118]

[0119] Then, after reasonable processing and analysis, the location of water leakage is marked on the two-dimensional or three-dimensional model of the tunnel using easy-to-understand icons, colors, and text. Presented through a graphical interface, this allows construction and management personnel to clearly see the specific location of the leakage, facilitating targeted treatment and monitoring. Simultaneously, the module receives the latest data from the image recognition module in real time, updating early warning information promptly to ensure that the displayed risk situation is synchronized with the actual construction status, enabling timely measures to address changes in risk. Furthermore, the module can also be linked through interfaces to issue early warning prompts via sound, flashing lights, and other methods, allowing personnel in different positions to quickly obtain the necessary warning information and make timely decisions and take appropriate measures.

[0120] On the other hand, the present invention also provides a device for identifying and handling water leakage risks during railway tunnel construction, the device comprising:

[0121] Multiple image acquisition and transmission modules are deployed at predetermined locations in the railway tunnel construction site. The image acquisition and transmission modules acquire images of the designated construction locations according to a predetermined format, and perform data segmentation, intra-block compression, relation mapping, and encoding storage on the acquired images to be analyzed, and transmit them through a preset link.

[0122] The image preprocessing module is used to accept the image to be analyzed and perform median filtering and homomorphic filtering on the image to enhance the image and remove noise to obtain an intermediate image.

[0123] The image recognition module is used to execute the railway tunnel construction water leakage risk identification and handling method described in steps S101 to S103 above, and output the identification results of water leakage location and water leakage risk level.

[0124] The early warning display module is used to generate early warning information and provide audio-visual prompts when the identification result of the water seepage risk level is higher than the set standard.

[0125] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0126] On the other hand, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0127] The present invention will now be described with reference to a specific embodiment:

[0128] This embodiment provides a device for identifying and handling water leakage risks during railway tunnel construction, including: an image acquisition and transmission module, an image preprocessing module, an image recognition module, and an early warning display module. The image preprocessing module and the image recognition module constitute an edge computing device. Figure 2 As shown, the specific implementation method is as follows:

[0129] 1. Image acquisition and transmission module

[0130] 1.1 Image Acquisition: At the working face inside the railway tunnel, multiple high-definition image acquisition units, consisting of high-resolution and wide-angle lenses, are rationally deployed based on the tunnel's actual structure, dimensions, and construction characteristics. For straight sections of the tunnel, equipment is installed every 50 meters at the tunnel top and sidewalls; in curved sections or areas with complex geology, the installation interval is appropriately shortened to 20 meters. Special attention is paid to areas prone to water leakage, such as tunnel junctions and areas with well-developed rock fissures.

[0131] Initialize the network parameters of the image acquisition unit device, including IP address, subnet mask, gateway, etc., to ensure a smooth network connection with the edge computing device. Set the image resolution to 1920×1080 and the frame rate to 60 frames per second.

[0132] Meanwhile, based on the continuous real-time monitoring of ambient light intensity in the tunnel by the photosensors within each acquisition unit, the exposure time, white balance, and other parameters of the acquisition unit are adjusted to adapt to the lighting conditions inside the tunnel. For example, when the sensor detects low light, it automatically calculates and adjusts the exposure time of the acquisition unit to a longer value, such as 1 / 30 second, while precisely adjusting the focus distance to ensure that the acquired images are clear and of high quality. The acquired images include .jpg format pictures and .mp4 format videos to comprehensively record the tunnel construction scene.

[0133] 1.2 Image compression: such as Figure 3 As shown, image block compression technology based on fractal theory is used to reduce the bandwidth pressure of high-definition image transmission.

[0134] 1.2.1 Data Blocking: Dividing the image to be compressed into several non-overlapping pixel sub-blocks. The size of the sub-block can be set to 2×2, 4×4, or 8×8, etc., according to actual needs. Taking an 8×8 sub-block as an example, according to formula (1), when When the value is 2, the pixel sub-block .

[0135] 1.2.2 Intra-block compression: For each pixel sub-block A Discrete Cosine Transform (DCT) is performed to convert the spatial domain data to the frequency domain, concentrating the data energy in a few low-frequency coefficients. These coefficients are then quantized and encoded to remove redundant information and achieve compression. For example, during quantization, low-frequency coefficients are quantized according to a pre-defined quantization table.

[0136] 1.2.3 Relationship Mapping: For each pixel sub-block Find similar blocks in the entire image or within a specific search area. Record the mapping relationships such as scaling ratio, rotation angle, and translation amount. For example, similar blocks and their mapping parameters can be determined by calculating the similarity metric between sub-blocks.

[0137] 1.2.4 Encoding and Storage: Mapping relationship for each pixel sub-block Encode and store necessary auxiliary information (such as block location and whether it is a boundary block).

[0138] 1.2.5 Image Transmission: Compressed image data is transmitted to the edge computing device via high-speed Ethernet or wireless network using the Real-time Streaming Protocol (RTSP) to ensure reliable transmission in the network and achieve efficient flow of image data from the acquisition end to the processing end.

[0139] 1.2.6 Image Restoration: During decompression, the edge computing device restores the original image sub-blocks by iterative reconstruction based on the stored information, and reconstructs the entire image.

[0140] 2. Image preprocessing module

[0141] Given the complex lighting conditions inside the tunnel, such as Figure 4 As shown, a hybrid algorithm of median filtering and homomorphic filtering is used to preprocess the acquired images.

[0142] 2.1 Median Filtering

[0143] 2.1.1 Window Scanning: Using a A sliding window of the pixel matrix scans the original image row by row and column by column, and records the pixel values ​​scanned within the window as a set. ,Right now (Formula 2). For example, when At that time, the set of pixel values ​​within the window .

[0144] 2.1.2 Extracting grayscale values: from the set Extract the grayscale values ​​of each pixel and denote them as the set. ,Right now (Formula 3). when At that time, the set of grayscale values ​​of pixels within the window .

[0145] 2.1.3 Sorting and Replacing Gray Values: Sorting and replacing gray value sets The grayscale values ​​in the data are sorted by size, and the median value of the sorted data is used. Replace the pixel value of the center pixel of the window to complete the filtering operation at that point. Continue moving the window and performing the above operation on each pixel of the image until the entire image has undergone median filtering.

[0146] 2.2 Homomorphic Filtering

[0147] 2.2.1 Image Decomposition: The image after median filtering is represented as the product of the illumination component and the reflection component, denoted as... (Formula 4).

[0148] 2.2.2 Logarithmic Transformation: For the input image Perform a logarithmic transformation to obtain

[0149] 2.2.3 Fourier Transform: [This section appears to be incomplete and requires further context.] Transforming from the spatial domain to the frequency domain, we obtain Through Fourier transform, we obtain .

[0150] 2.2.4 Design and Application of Homomorphic Filters: Based on the frequency characteristics of illumination and reflection components, homomorphic filters are designed for... After filtering, the result is obtained. .

[0151] 2.2.5 Inverse Fourier Transform: This transforms the filtered frequency domain image... Transform back into the spatial domain to obtain .

[0152] 2.2.6 Exponential Transformation: For Perform an exponential transform to obtain the enhanced image. .

[0153] This process removes noise to the greatest extent possible while preserving edge information, improving image quality and laying the foundation for subsequent recognition.

[0154] 3. Image recognition module

[0155] 3.1 Model Construction and Architecture: such as Figure 5As shown, an RT-Unet model was constructed, pre-trained and optimized using a large amount of real tunnel construction image data. This model integrates a self-attention mechanism module (SAM) on the basis of the traditional Unet architecture. The model consists of two parts: encoding and decoding.

[0156] 3.2 Encoding and Downsampling: The encoding part extracts image features through multi-layer convolution and pooling operations. Each convolution layer uses multiple convolution kernels to slide and extract features such as edges and textures, such as using 3×3×3 convolution kernels. The input feature map is convolved according to formula (5).

[0157] After convolution, dilated pyramid pooling (ASPP) is used for downsampling to reduce the feature map resolution and decrease the amount of data and computation. Pooling is performed on the feature map output by convolution according to formula (6).

[0158] Comparison model structure diagram Figure 5 The main working principle of the hollow pyramid pooling ASPP part is as follows:

[0159] a. ASPP structure: ASPP consists of a convolutional layer Conv1×1, dilated convolution (Conv3×3, rate=6, 12, 18) and ASPPooling (including 1×1 pooling, Conv1×1 and bilinear interpolation upsampling).

[0160] b. Convolutional layer Conv1×1: First, the input feature map is preliminarily extracted and the number of channels is adjusted, mapping the input feature map to a new feature space to prepare for subsequent processing.

[0161] c. Dilated Convolution: Parallel sampling is performed using dilated convolution with different dilation rates. The dilation rate controls the padding expansion during convolution. By using different padding and dilation, receptive fields of different scales can be obtained, and multi-scale information can be extracted, thereby capturing the features of objects of different sizes in the image.

[0162] d.ASPP Pooling: First, a 1×1 pooling layer is used to compress the feature maps of each channel to 1×1, thereby extracting the features of each channel and obtaining the global features; then, a Conv1×1 convolutional layer is used to further extract the features obtained in the previous step and reduce the dimensionality; finally, an upsampling operation is performed based on the bilinear interpolation algorithm to restore the feature map size.

[0163] e. Feature Fusion and Output: The outputs of Conv1×1 convolution, dilated convolution, and ASPP Pooling are concatenated along the channel dimension to obtain a feature representation containing multi-scale and global information. Finally, a Conv1×1 convolutional layer is used to fuse and reduce the dimensionality of the concatenated features to obtain the final output feature map, which can be used for subsequent image segmentation or other tasks.

[0164] 3.3 Decoding and Upsampling: The decoding part restores the feature map to the original image size through deconvolution and skip connections, combined with the self-attention mechanism module SAM, to achieve pixel-level image segmentation.

[0165] 3.4 Self-Attention Mechanism (SAM): The Self-Attention Mechanism (SAM) module assigns different weights to different regions of the input feature map, enabling it to focus more on key features in the image and improve the model's sensitivity to important information. It calculates the correlation between pixels in the input sequence, assigns attention weights to each pixel, and generates the output through weighted summation, thereby focusing on the important features.

[0166] 3.4.1 Input Process

[0167] 3.4.1.1 Input downsampling process feature map , dimension After a 1×1×1 convolution This yields a new feature map A.

[0168] 3.4.1.2 Extracting a coarser-scale upsampled signal from the ASPP Pooling upsampling process , dimension After a 1×1×1 convolution This yields a new feature map B.

[0169] 3.4.2 Calculation process

[0170] 3.4.2.1 XOR Operation: Perform an XOR operation on feature maps A and B to obtain feature map C, with dimension 1. .

[0171] 3.4.2.2 Activation Function ReLU Feature map C is activated by the ReLU function. The feature map D is obtained through calculation.

[0172] 3.4.2.3 Conv1×1×1 Convolution: The feature map D is subjected to a 1×1×1 convolution operation. The feature map E is obtained, and its dimension becomes .

[0173] 3.4.2.4 Activation function Sigmoid: Feature map E is activated by the Sigmoid function. The feature map F is calculated, and the output value is compressed to between 0 and 1 to generate the attention coefficient.

[0174] 3.4.3 Output Process

[0175] 3.4.3.1 Resampling: The feature map F is resampled using trilinear interpolation to obtain a result similar to the input feature map. Same spatial dimension Attention coefficient .

[0176] 3.4.3.2 Weighted Operation: Attention Coefficient With input feature map Perform element-wise multiplication to obtain the weighted output feature map. .

[0177] 3.5 Model Training: During model pre-training, hyperparameters such as the learning rate and optimizer are adjusted, and data augmentation techniques such as flipping, rotating, and scaling are used to increase data diversity. The maximum learning rate is set to 1×10⁻⁶. -4 The minimum learning rate is 0.01 times the maximum learning rate. The Adam optimizer is used, and the learning rate descent method is cosine annealing (cosine annealing) to optimize the model and improve its generalization ability.

[0178] 3.6 Inference and Post-processing: The model inference output employs a multi-threshold judgment and post-processing algorithm. A low threshold (0.1~0.2) is set to initially screen potential leakage areas, a medium threshold (0.3~0.5) to further determine possible areas, and a high threshold (0.6~0.9) to determine highly certain areas. Morphological dilation and erosion operations are combined to optimize the segmentation results. Dilation fills small holes and discontinuities, while erosion eliminates small protrusions and noise, removing isolated noise points and small regions. The RT-Unet model is deployed on customized edge computing hardware. The backbone network uses MobileNetV3_small, trained for hundreds of rounds, reducing the number of model parameters and computational load. Lightweight deployment on edge computing devices improves inference speed and meets real-time requirements.

[0179] 4. Early Warning Display Module

[0180] 4.1 Interface Development: The visual interface was developed using the Vue.js framework in JavaScript. The interface includes modules such as tunnel project overview, surrounding rock measurement, advanced geological prediction, safety risk type overview, risk visualization "one map", machine vision, machinery operation status, personnel positioning, and emergency plan.

[0181] 4.2 Risk Assessment and Classification: The early warning display module integrates risk data output from the image recognition module with relevant data such as tunnel geographic information and construction progress. It employs the LEC quantitative evaluation method for operational condition hazards and establishes a four-color risk level classification standard. Based on the leakage areas identified in the images, it determines the water accumulation area and flow rate, establishing a risk level mapping relationship.

[0182] 4.3 Visualization: Risk data is processed and analyzed, and leakage locations are marked on the tunnel's 2D or 3D model using icons, colors, and text. Presented through a graphical interface, this allows construction and management personnel to clearly understand the leakage locations, facilitating targeted treatment and monitoring. The module receives the latest data in real time and updates early warning information promptly, ensuring synchronization with the actual construction situation.

[0183] 4.4 Early Warning Prompts: Through interface linkage, early warning prompts are issued in various ways such as sound and flashing lights, enabling personnel in different positions to quickly obtain early warning information, make timely decisions, and take measures to deal with changes in risks.

[0184] The tunnel construction water leakage risk image recognition technology based on the RT-Unet semantic segmentation model in this embodiment has many significant beneficial effects.

[0185] First, compared to traditional manual inspections and physical monitoring methods, this embodiment offers significant advantages in terms of detection efficiency and accuracy. Traditional manual inspections rely on personnel experience and subjective judgment, resulting in low efficiency, difficulty in guaranteeing accuracy, and a high risk of missed or false detections. This embodiment employs advanced image acquisition equipment and deep learning models. Leveraging the powerful feature extraction and segmentation capabilities of deep learning models, it can quickly and automatically process and analyze large amounts of image data, unaffected by human factors, thus significantly improving detection efficiency. Simultaneously, traditional physical monitoring equipment often only monitors localized areas, incurs high installation and maintenance costs, and is easily affected by the construction environment. This embodiment, however, through real-time image acquisition and intelligent analysis of the construction site, can cover the entire work area, reducing costs and maintenance work. Furthermore, it is unaffected by environmental factors such as electromagnetic interference, resulting in more stable and reliable monitoring results.

[0186] Secondly, in terms of real-time data processing, this embodiment far surpasses traditional methods in terms of timeliness. Traditional monitoring methods have long data collection and analysis cycles, making it impossible to provide timely risk warnings. In contrast, the model in this embodiment is deployed on a high-efficiency edge computing device, enabling real-time image processing and analysis, and providing leakage risk prediction results immediately. This allows relevant personnel to take immediate action, greatly shortening response time and effectively preventing the accident from escalating.

[0187] Furthermore, this technology innovates and enhances existing safety management models. Traditional safety management relies primarily on periodic inspections, experience-based judgment, and limited monitoring data, resulting in significant uncertainty and lag. This embodiment employs advanced deep learning and image recognition technologies to effectively address the complex and variable lighting conditions and background interference within tunnels. Through multi-threshold judgment and post-processing algorithms, it accurately extracts the characteristics of water leakage, improving the precision and accuracy of identification. This enables automated risk identification and monitoring, transforming safety management from a passive response to proactive prevention. Simultaneously, real-time data transmission and remote monitoring capabilities allow for the immediate acquisition and processing of monitoring information, breaking down time and space limitations and improving the response speed and decision-making efficiency of safety management.

[0188] Furthermore, from a cost perspective, although the initial training and development of the model requires a certain investment, in the long run, compared with traditional manual inspections that require a large amount of manpower, as well as high-cost, easily damaged and complex physical monitoring equipment, this embodiment can significantly reduce manpower and material costs.

[0189] Finally, this embodiment provides intuitive and clear decision-making support for safety management through a visual interface design. Managers can quickly understand the tunnel's water leakage situation through intuitive images and data displays, conduct accurate analysis and judgment, formulate scientific and reasonable construction strategies and safety measures, ensure the project progresses smoothly as planned, avoid personnel casualties, mechanical equipment damage, and project shutdowns caused by water leakage accidents, and improve overall construction efficiency.

[0190] In summary, the railway tunnel construction water leakage risk identification and handling method and device of the present invention acquires the image to be analyzed after data segmentation, intra-block compression, relation mapping, and encoding storage through a preset link; enhances and denoises the image using median filtering and homomorphic filtering; and constructs a railway tunnel construction water leakage risk identification model based on multiple encoders and decoders to identify the location and risk level of water leakage. The encoders use multi-layer convolution and hollow pyramid pooling operations to extract image features, and the decoder input skips the output of the corresponding level encoder, with a self-attention mechanism module integrated at the skip connection. This invention can efficiently and accurately identify the location and risk level of water leakage in railway tunnel construction, effectively improving monitoring efficiency and accuracy, achieving comprehensive coverage of the entire work surface, and is unaffected by the construction environment, providing strong protection for railway tunnel construction safety.

[0191] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0192] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0193] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0194] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying and handling water leakage risks during railway tunnel construction, characterized in that, The method includes the following steps: Images to be analyzed, collected at the railway tunnel construction site, are acquired through a preset link. These images are obtained by transmitting and reconstructing data after being segmented into blocks, compressed within blocks, mapped onto relationships, and encoded for storage. Median filtering and homomorphic filtering are applied to the image to be analyzed to enhance the image and denoise it, thereby obtaining an intermediate image. The intermediate image is input into a pre-trained railway tunnel construction seepage risk identification model to output the identification results of seepage location and seepage risk level; the railway tunnel construction seepage risk identification model is based on downsampling by multiple encoders, the encoders use multi-layer convolution and hollow pyramid pooling operations to extract image features, and upsampling is performed by multiple decoders. The input of the decoder is connected to the output of the encoder at the corresponding level, and a self-attention mechanism module is integrated at the connection point. The hollow pyramid pooling includes three branches: the first branch contains a 1×1 convolutional layer, the second branch contains multiple dilated convolutional layers with varying dilation rates, and the third branch contains a 1×1 pooling layer, a 1×1 convolutional layer, and an upsampling layer. The outputs of the first, second, and third branches are concatenated along the channel dimension and then fused and dimensionality-reduced by a 1×1 convolutional layer to output the image features. The self-attention mechanism module downsamples the first intermediate feature of the multi-layer convolution in the encoder and the output of the upsampling layer in the hollow pyramid pooling and performs an XOR operation to obtain a second intermediate feature. After performing self-attention and resampling on the second intermediate feature, it is multiplied with the first intermediate feature to obtain a third intermediate feature. The upsampled image features, the outputs of the multi-layer convolution in the encoder, and the third intermediate feature are aggregated, upsampled, and then input into the corresponding level of the decoder to achieve skip connections.

2. The method for identifying and handling water leakage risks during railway tunnel construction according to claim 1, characterized in that, The processing steps during the transmission of the image to be analyzed include: The image to be analyzed is divided into multiple non-overlapping pixel sub-blocks according to a set size; A discrete cosine transform is performed on each pixel sub-block to convert the data from the spatial domain to the frequency domain. The converted data is then quantized and encoded to remove redundant information and achieve compression. For each pixel sub-block, find similar pixel block pairs that meet the set similarity criteria within the entire range of the image to be analyzed, and establish a mapping relationship between the similar pixel block pairs. The mapping relationship includes scaling ratio, rotation angle, and translation amount. After each pixel sub-block is compressed and encoded based on the mapping relationship, it is then transmitted using the Real-Time Streaming Protocol (RTSP) for decompression and reconstruction of the image to be analyzed.

3. The method for identifying and handling water leakage risks during railway tunnel construction according to claim 1, characterized in that, The intermediate image is obtained by performing median filtering and homomorphic filtering on the image to be analyzed for image enhancement and denoising, including: A filter window of a set size is used to scan the image to be analyzed row by row and column by column; gray values ​​are extracted from the pixels extracted by the filter window each time, and arranged in ascending or descending order; the value of the center pixel in the filter window is replaced by the middle value after the arrangement. The image to be analyzed after median filtering is represented as the product of the illumination component and the reflection component, and logarithmic and Fourier transforms are performed. A homomorphic filter is created to filter the image based on the frequency characteristics of the illumination component and the reflection component. The filtered frequency domain image is converted back to the spatial domain through inverse Fourier transform, and then the enhanced image is obtained through exponential transform.

4. The method for identifying and handling water leakage risks during railway tunnel construction according to claim 1, characterized in that, The pre-training steps of the railway tunnel construction seepage risk identification model include: Multiple sample images were acquired for railway tunnel construction scenarios. Each sample image was processed by median filtering and homomorphic filtering, and the location of the water seepage area and the corresponding risk level were marked as labels to construct a training sample set. The training sample set is used to train the railway tunnel construction seepage risk identification model in its initial state. The sample images processed by median filtering and homomorphic filtering are input into the railway tunnel construction seepage risk identification model in its initial state, and the model outputs predicted values ​​for the location and risk level of the seepage area. A loss function is constructed based on the deviation between the predicted values ​​and the labels, and the parameters of the railway tunnel construction seepage risk identification model are updated. The method employs the Adam optimizer to train the railway tunnel construction seepage risk identification model, with the learning rate decreasing via cosine annealing.

5. The method for identifying and handling water leakage risks during railway tunnel construction according to claim 4, characterized in that, The backbone network of the railway tunnel construction seepage risk identification model adopts the MobileNetV3_small network to achieve lightweight deployment.

6. The method for identifying and handling water leakage risks during railway tunnel construction according to claim 1, characterized in that, The method further includes: Morphological operations are performed on the identification results of water seepage locations in the image to be analyzed. The dilation operation fills small holes and narrow discontinuities in the image, and the erosion operation removes bumps and noise in the image.

7. The method for identifying and handling water leakage risks during railway tunnel construction according to claim 1, characterized in that, The method further includes: The images to be analyzed are collected and updated in real time according to a set frequency, and the identification results of seepage location and seepage risk level are updated. For the identification results of seepage risk level higher than the set standard, early warning information is generated and audio-visual prompts are given.

8. A device for identifying and handling water leakage risks during railway tunnel construction, characterized in that, The device includes: Multiple image acquisition and transmission modules are deployed at predetermined locations in the railway tunnel construction site. The image acquisition and transmission modules acquire images of the designated construction locations according to a predetermined format, and perform data segmentation, intra-block compression, relation mapping, and encoding storage on the acquired images to be analyzed, and transmit them through a preset link. An image preprocessing module is used to receive the image to be analyzed and perform median filtering and homomorphic filtering on the image to be analyzed to enhance the image and remove noise to obtain an intermediate image; The image recognition module is used to execute the railway tunnel construction water leakage risk identification and handling method as described in any one of claims 1 to 7, and output the identification results of the water leakage location and water leakage risk level; The early warning display module is used to generate early warning information and provide audio-visual prompts when the identification result of the water seepage risk level being higher than the set standard.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images