A Defect Identification Method for Subway Tunnel Structures Based on Lightweight Neural Networks

By deploying a lightweight neural network model in the edge computing unit, the problems of real-time performance and accuracy in identifying structural defects in subway tunnels were solved, enabling automatic identification of surface defects in tunnel linings and automation of operation and maintenance processes.

CN123118604APending Publication Date: 2026-11-17JILIN JIANZHU UNIVERSITY
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611493153.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-23
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for identifying structural defects in subway tunnels rely on high-performance servers and stable communication links, making it difficult to achieve real-time and accurate defect identification. Furthermore, traditional inspection methods are limited by the tunnel environment, making it difficult to continuously and stably record the location and type of defects.

Method used

A lightweight neural network model, including depthwise separable convolution, channel shuffling layer, and defect classification output layer, is deployed on the edge computing unit to locate and identify defect areas in tunnel lining surface image data, and generate inspection records, early warning information, and repair early warning information.

Benefits of technology

It enables real-time and accurate identification of structural defects in subway tunnels, supports defect area location, type identification, and severity classification, drives the automation of operation and maintenance processes, and improves inspection efficiency and the feasibility of operation and maintenance tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118604A_ABST
    Figure CN123118604A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of artificial intelligence technology, specifically relating to a method for identifying structural defects in subway tunnels based on lightweight neural networks. It utilizes an inspection robot to collect image data of the tunnel lining surface, and a lightweight neural network defect identification model deployed on an edge computing unit for defect identification. The model employs a cascaded structure of depthwise separable convolutional blocks, inverted residual blocks, and a channel washing layer to achieve multi-scale feature extraction and defect area localization. This invention can locate, classify, and grade the severity of defects such as cracks, water leakage, spalling, spalling, misalignment, and exposed rebar. It automatically generates identification conclusions based on a pre-set defect judgment table, classifying defects as normal, warning defects, or repair defects. This drives the maintenance process, including inspection record generation, warning information output, and maintenance work order dispatch, achieving fully automated processing from defect identification to maintenance task dispatch.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method for identifying structural defects in subway tunnels based on lightweight neural networks. Background Technology

[0002] As a crucial infrastructure for urban rail transit, the structural safety of subway tunnels directly impacts operational and public safety. During long-term service, tunnel lining surfaces are susceptible to structural defects such as cracks, water leakage, spalling, spalling, misalignment, and exposed rebar due to factors including geological conditions, groundwater, train vibrations, and surrounding construction. Traditional inspection methods rely primarily on manual visual inspection, handheld testing instruments, or inspection vehicles. However, these methods are limited by tunnel lighting conditions, obstructions, dust, and specific work time windows, making it difficult to continuously, stably, and objectively record the location, type, and severity of defects. Furthermore, there is a lack of close integration between inspection results and subsequent maintenance tasks.

[0003] With the development of machine vision and deep learning technologies, image-based automatic defect identification solutions for subway tunnel structures have gradually gained attention. However, many existing solutions rely on high-performance servers and stable communication links, typically requiring images collected by inspection equipment to be uploaded to a remote platform for processing. This results in problems such as high transmission latency, high bandwidth pressure, and difficulty in real-time output when communication is limited. General-purpose deep neural networks have a large number of parameters, making it difficult to balance recognition accuracy and operational efficiency when directly deployed in edge computing units such as inspection robots. Furthermore, surface defects in tunnel linings often exhibit low contrast, multi-scale variations, and strong background interference. Existing methods still have shortcomings in defect area localization, defect type identification, and early warning information generation, making it difficult to meet the practical needs of rapid on-site inspection and maintenance response.

[0004] Therefore, a method for identifying structural defects in subway tunnels based on lightweight neural networks is needed. This method deploys lightweight structures such as depthwise separable convolution and channel washing on edge computing units to perform defect region localization and defect type discrimination on tunnel lining surface image data collected by inspection equipment. It outputs defect identification results containing defect region location information, defect type information, and defect severity information. Based on the identification conclusions, it distinguishes between normal structural identification results, warning structural defect identification results, and repairable structural defect identification results, and generates inspection records, structural defect warning information, and structural repair warning information respectively, so as to improve the real-time performance, accuracy, and executability of operation and maintenance tasks for identifying structural defects in subway tunnels. Summary of the Invention

[0005] This invention provides a method for identifying structural defects in subway tunnels based on lightweight neural networks, comprising: S1, acquire image data of the tunnel lining surface collected by the subway tunnel inspection equipment at the preset detection position; S2, input the tunnel lining surface image data into a lightweight neural network defect recognition model deployed on an edge computing unit. The lightweight neural network defect recognition model includes a depth-separable convolutional feature extraction layer, a channel shuffling feature enhancement layer, and a defect classification output layer connected in sequence. S3, using the lightweight neural network defect recognition model, perform defect region localization and defect type discrimination on the tunnel lining surface image data, and output defect recognition results containing defect region location information, defect type information and defect degree information; S4. Determine the identification conclusion corresponding to the current tunnel structure state based on the defect identification results; S5, when the identification conclusion is a normal structural identification result, an inspection record that does not require manual review is generated; when the identification conclusion is a warning structural defect identification result, a structural defect warning message is generated; when the identification conclusion is a repairable structural defect identification result, a structural repair warning message is generated.

[0006] Further, S1 includes: The subway tunnel inspection equipment is an inspection robot that moves along the subway tunnel track. The inspection robot is equipped with the edge computing unit, image acquisition unit, supplementary lighting unit and local storage unit. The inspection robot is controlled to move along the subway tunnel track, and the image acquisition unit and the supplementary lighting unit are triggered when it reaches the preset detection position. The image acquisition unit acquires image data of the tunnel lining surface towards the tunnel lining surface, and the edge computing unit records the detection position identifier, acquisition time information, and inspection equipment identifier corresponding to the preset detection position; The tunnel lining surface image data is associated with the detection location identifier, the acquisition time information and the inspection equipment identifier and stored in the local storage unit, and the tunnel lining surface image data is input into the lightweight neural network defect recognition model in the edge computing unit.

[0007] Furthermore, the lightweight neural network defect recognition model is deployed on the edge computing unit, and the lightweight neural network defect recognition model includes an input layer, a depthwise separable convolutional block, an inverted residual block, a channel shuffling layer, a global average pooling layer, and a defect classification output layer; The input layer receives the image data of the tunnel lining surface; The depth-separable convolutional block performs channel-wise convolution and point-wise convolution on the tunnel lining surface image data to obtain shallow features of the tunnel lining surface. The inverted residual block performs channel expansion, depth convolution, and channel compression on the shallow features of the tunnel lining surface to obtain the deep features of the tunnel lining surface; the channel mixing layer performs cross-channel feature exchange on the deep features of the tunnel lining surface to obtain the enhanced features of the tunnel lining surface. The global average pooling layer performs pooling on the enhanced features of the tunnel lining surface to obtain the feature vector of the tunnel lining surface; The defect classification output layer outputs the defect identification result based on the feature vector of the tunnel lining surface, and sends the defect identification result to the defect judgment table calling module in the edge computing unit.

[0008] Further, S3 includes: The lightweight neural network defect recognition model uses a depthwise separable convolutional block, an inverted residual block, and a channel washing layer to sequentially extract features from the tunnel lining surface image data, resulting in tunnel lining surface feature maps at multiple scales. The defect region suggestion sublayer in the lightweight neural network defect identification model generates candidate defect regions on the tunnel lining surface feature maps at multiple scales. The candidate region filtering sublayer in the lightweight neural network defect identification model determines the target defect region from the candidate defect regions based on the overlap relationship between the candidate defect regions and the response intensity of the candidate defect regions in the feature map of the tunnel lining surface. The defect type discrimination sublayer in the lightweight neural network defect recognition model performs defect category discrimination on the target defect region to obtain defect type information; The defect severity discrimination sublayer in the lightweight neural network defect recognition model performs defect severity discrimination on the target defect region to obtain defect severity information; The location output sublayer in the lightweight neural network defect recognition model generates defect region location information based on the image coordinates of the target defect region in the tunnel lining surface image data and the detection location identifier. The defect type information, the defect severity information, and the defect area location information are combined to form the defect identification result, and the defect identification result is associated with the detection location identifier, the acquisition time information, and the inspection equipment identifier.

[0009] Further, S4 includes: The defect determination table pre-installed in the edge computing unit is invoked. The defect determination table records the correspondence between defect type information, defect severity information and identification conclusion. The defect type and defect severity information in the defect identification results are matched with the defect determination table; When the defect identification result does not contain defect type information, the identification conclusion is determined to be a structurally normal identification result; When the defect type information in the defect identification result is water leakage or crack and the defect severity information is mild, the identification conclusion is determined to be a warning structural defect identification result. When the defect type information in the defect identification result is spalling, spalling, misalignment, or exposed rebar, or when the defect type information is crack and the defect severity information is moderate or severe, the identification conclusion is determined to be a reparative structural defect identification result. The identification conclusion is associated with and saved along with the defect identification result, the detection location identifier, the acquisition time information, and the inspection equipment identifier.

[0010] Furthermore, when the identification conclusion is a structurally normal identification result, an inspection record that does not require manual review is generated, including: Generate a normal structural inspection record, which includes the detection location identifier, the acquisition time information, the inspection equipment identifier, the image identifier corresponding to the tunnel lining surface image data, the identification conclusion, and the model version identifier of the lightweight neural network defect identification model; Write the normal inspection record of the structure into the local storage unit, and when the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, upload the normal inspection record of the structure to the tunnel operation and maintenance database in the operation and maintenance server. In the inspection display interface of the subway tunnel inspection equipment, the tunnel lining section corresponding to the detection location is marked as being in normal structural condition; the subway tunnel inspection equipment is controlled to continue moving to the next preset detection location, and the next round of tunnel lining surface image data acquisition is triggered when the next preset detection location is reached.

[0011] Furthermore, when the identification conclusion is a warning structural defect identification result, structural defect warning information is generated, including: Based on the defect type information, defect severity information, and defect area location information in the defect identification results, structural defect early warning information is generated; The structural defect early warning information is output to the local early warning output unit of the subway tunnel inspection equipment, and the local early warning output unit performs audible and visual prompts. The structural defect early warning information is sent to the operation and maintenance terminal, which maps the location of the early warning defect area to the corresponding tunnel lining section in the tunnel lining unfolded diagram, and displays the early warning mark on the tunnel lining section; the structural defect early warning information is associated with the tunnel lining component identifier in the tunnel structural component ledger to generate a task to be reviewed; The task to be reviewed is written into the tunnel operation and maintenance database and displayed in the task list of the operation and maintenance terminal. At the same time, the task to be reviewed is associated with the tunnel lining section corresponding to the detection location identifier.

[0012] Furthermore, when the identification conclusion is a remedial structural defect identification result, structural repair early warning information is generated, including: Based on the defect type information, defect severity information, and defect area location information in the defect identification results, structural repair early warning information is generated; The structural repair early warning information is output to the maintenance work order system, which then generates a structural repair work order. The structural repair early warning information is associated with the tunnel lining component identifier in the tunnel structural component ledger, and the component status corresponding to the tunnel lining component identifier is marked as pending repair. The structural repair work order is sent to the operation and maintenance terminal, which then displays the structural repair work order in the task list according to the repair priority. The structural repair early warning information and the structural repair work order are written into the tunnel operation and maintenance database, and the structural repair work order is associated with and saved with the tunnel lining section corresponding to the detection location identifier.

[0013] Furthermore, the edge computing unit is deployed in the subway tunnel inspection equipment, and the lightweight neural network defect recognition model runs in the edge computing unit; When the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, the edge computing unit uploads the defect identification results, the structural defect early warning information and the structural repair early warning information to the operation and maintenance server. When the communication connection is unavailable, the edge computing unit saves the defect identification result, the structural defect early warning information and the structural repair early warning information to the local storage unit, and uploads the defect identification result, the structural defect early warning information and the structural repair early warning information saved in the local storage unit to the operation and maintenance server after the communication connection is restored; During the upload process, the detection location identifier, the acquisition time information, and the inspection equipment identifier are sent together with the defect identification result, the structural defect early warning information, and the structural repair early warning information. The operation and maintenance server then writes the uploaded information into the tunnel operation and maintenance database.

[0014] Furthermore, the edge computing unit receives a pre-configured lightweight neural network defect identification model parameter package from the operation and maintenance server, loads the lightweight neural network defect identification model parameter package into the lightweight neural network defect identification model, and records the model version identifier after loading. When the subway tunnel inspection equipment reaches the subsequent preset detection position, the edge computing unit uses the lightweight neural network defect recognition model loaded with the parameter package of the lightweight neural network defect recognition model to perform defect area localization and defect type discrimination on the subsequently collected tunnel lining surface image data, output the subsequent defect recognition results, and determine the subsequent recognition conclusion based on the subsequent defect recognition results; When the subsequent identification conclusion is a normal structural identification result, an inspection record is generated that does not require subsequent manual review. When the subsequent identification conclusion is a warning structural defect identification result, a subsequent structural defect warning information is generated; When the subsequent identification conclusion is a corrective structural defect identification result, a subsequent structural repair early warning information is generated, and the subsequent defect identification result, the subsequent structural defect early warning information, and the subsequent structural repair early warning information are associated and saved with the detection location identifier, collection time information, and inspection equipment identifier corresponding to the subsequent preset detection location.

[0015] This invention belongs to the field of artificial intelligence technology, specifically relating to a method for identifying structural defects in subway tunnels based on lightweight neural networks. It utilizes an inspection robot to collect image data of the tunnel lining surface, and a lightweight neural network defect identification model deployed on an edge computing unit for defect identification. The model employs a cascaded structure of depthwise separable convolutional blocks, inverted residual blocks, and a channel washing layer to achieve multi-scale feature extraction and defect area localization. This invention can locate, classify, and grade the severity of defects such as cracks, water leakage, spalling, spalling, misalignment, and exposed rebar. It automatically generates identification conclusions based on a pre-set defect judgment table, classifying defects as normal, warning defects, or repair defects. This drives the maintenance process, including inspection record generation, warning information output, and maintenance work order dispatch, achieving fully automated processing from defect identification to maintenance task dispatch. Attached Figure Description

[0016] Figure 1 This is a flowchart of the process of a method for identifying structural defects in subway tunnels based on a lightweight neural network, as claimed in the embodiments of the present invention. Figure 2 This is a second flowchart of the proposed method for identifying structural defects in subway tunnels based on a lightweight neural network, as claimed in the embodiments of the present invention. Figure 3 This is the third flowchart of the proposed method for identifying structural defects in subway tunnels based on a lightweight neural network, as claimed in this embodiment of the invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] The technical solutions disclosed in the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] According to the first embodiment of the present invention, referring to Figure 1 This invention provides a method for identifying structural defects in subway tunnels based on lightweight neural networks, comprising: S1, acquire image data of the tunnel lining surface collected by the subway tunnel inspection equipment at the preset detection position; S2, input the tunnel lining surface image data into a lightweight neural network defect recognition model deployed on an edge computing unit. The lightweight neural network defect recognition model includes a depth-separable convolutional feature extraction layer, a channel shuffling feature enhancement layer, and a defect classification output layer connected in sequence. S3, using the lightweight neural network defect recognition model, perform defect region localization and defect type discrimination on the tunnel lining surface image data, and output defect recognition results containing defect region location information, defect type information and defect degree information; S4. Determine the identification conclusion corresponding to the current tunnel structure state based on the defect identification results; S5, when the identification conclusion is a normal structural identification result, an inspection record that does not require manual review is generated; when the identification conclusion is a warning structural defect identification result, a structural defect warning message is generated; when the identification conclusion is a repairable structural defect identification result, a structural repair warning message is generated.

[0020] Further, S1 includes: The subway tunnel inspection equipment is an inspection robot that moves along the subway tunnel track. The inspection robot is equipped with the edge computing unit, image acquisition unit, supplementary lighting unit and local storage unit. The inspection robot is controlled to move along the subway tunnel track, and the image acquisition unit and the supplementary lighting unit are triggered when it reaches the preset detection position. The image acquisition unit acquires image data of the tunnel lining surface towards the tunnel lining surface, and the edge computing unit records the detection position identifier, acquisition time information, and inspection equipment identifier corresponding to the preset detection position; The tunnel lining surface image data is associated with the detection location identifier, the acquisition time information and the inspection equipment identifier and stored in the local storage unit, and the tunnel lining surface image data is input into the lightweight neural network defect recognition model in the edge computing unit.

[0021] In this embodiment, step 1 involves acquiring image data of the tunnel lining surface collected by a subway tunnel inspection device at a preset detection location. The subway tunnel inspection device is implemented using an inspection robot that moves along the subway tunnel track. This inspection robot is equipped with an edge computing unit, an image acquisition unit, a supplementary lighting unit, and a local storage unit. The edge computing unit uses an embedded high-performance processor, equipped with 4GB of RAM and 32GB of storage memory, with a processor clock speed of no less than 1.5GHz, and supports a neural network inference acceleration instruction set. The image acquisition unit uses an industrial-grade digital camera with a CMOS sensor, an effective pixel count of no less than 20 million, and an autofocus lens with a focal length range of 8mm to 50mm and a field of view coverage of [missing information]. The range is 60 to 120 degrees, the image resolution is set to 4096×3000 pixels, and the image acquisition frame rate is dynamically adjusted according to the moving speed of the inspection robot, with a typical value of 2 to 5 frames per second; the supplementary lighting unit adopts a multi-spectral LED supplementary lighting group, including visible light LED beads and near-infrared LED beads. The color temperature of the visible light LED beads is 5500K to 6500K, and the wavelength of the near-infrared LED beads is 850nm. The maximum illuminance of the supplementary lighting group is not less than 5000Lux, and the illumination coverage is matched with the field of view of the image acquisition unit; the local storage unit adopts a solid-state storage device with a capacity of not less than 512GB, supports high-speed data writing, and the write speed is not less than 500MB per second.

[0022] The inspection robot moves along the subway tunnel track using a track-wheel drive system. The drive motor is a DC brushless motor with a rated power of 500W, a maximum speed of 5 km / h, and a positioning accuracy of ±10cm. The preset inspection positions are determined based on the length of the subway tunnel section and the number of lining blocks. The spacing between adjacent preset inspection positions is set to 5m to 20m, with the specific spacing adjusted according to the tunnel grade and maintenance requirements. The inspection robot's control system controls the motor driver according to the preset inspection path planning data, causing the inspection robot to move along the predetermined trajectory. When the inspection robot reaches the preset inspection position, the control system triggers the image acquisition unit and the supplementary lighting unit. Specifically, the system sends an acquisition command to the image acquisition unit and a lighting command to the supplementary lighting unit simultaneously. The preheating time for the supplementary lighting unit is 200ms, and the stable illumination time after the supplementary lighting unit is turned on is 100ms.

[0023] The image acquisition unit collects image data of the tunnel lining surface, with the angle between the acquisition direction and the normal direction of the tunnel lining surface not exceeding 15 degrees. The acquisition distance is determined based on the lens focal length and field of view, typically ranging from 1.5m to 3m. During acquisition, the supplementary lighting unit projects uniform illumination onto the tunnel lining surface, with an illumination uniformity of no less than 90%, to ensure consistent brightness and contrast of the acquired images. The edge computing unit simultaneously records the detection position identifier corresponding to the preset detection position, acquisition time information, and inspection equipment identifier. The encoding format for the detection location markers is "interval number-mileage marker". The interval number uses a 4-digit code, and the mileage marker uses a 6-digit code, accurate to the decimeter. For example, "0101-001245" represents the 01 interval of Line 1, and the mileage marker is 0012.45 meters. The time information collected uses the UTC timestamp format, with a time accuracy of milliseconds. The inspection equipment identifier uses the equipment serial number format, which is "model code-production batch-equipment serial number". For example, "TR-2024-001" represents the first inspection robot produced in 2024.

[0024] The association and storage of tunnel lining surface image data with detection location identifiers, acquisition time information, and inspection equipment identifiers is achieved through data encapsulation. The data encapsulation format is a JSON object containing fields image_id, position_id, timestamp, device_id, image_data, and metadata. The image_id is automatically generated by the edge computing unit in UUID format. The image_data field stores the Base64 encoded string of the original image data, and the metadata field stores image acquisition parameters, including exposure time, gain value, focusing distance, and ambient temperature. The encapsulated data is written to the inspection data directory of the local storage unit, and the data file is named in the format "{position_id}_{timestamp}.json". After local storage is completed, the edge computing unit inputs the tunnel lining surface image data into a lightweight neural network defect recognition model for defect recognition processing.

[0025] Furthermore, the lightweight neural network defect recognition model is deployed on the edge computing unit, and the lightweight neural network defect recognition model includes an input layer, a depthwise separable convolutional block, an inverted residual block, a channel shuffling layer, a global average pooling layer, and a defect classification output layer; The input layer receives the image data of the tunnel lining surface; The depth-separable convolutional block performs channel-wise convolution and point-wise convolution on the tunnel lining surface image data to obtain shallow features of the tunnel lining surface. The inverted residual block performs channel expansion, depth convolution, and channel compression on the shallow features of the tunnel lining surface to obtain the deep features of the tunnel lining surface; the channel mixing layer performs cross-channel feature exchange on the deep features of the tunnel lining surface to obtain the enhanced features of the tunnel lining surface. The global average pooling layer performs pooling on the enhanced features of the tunnel lining surface to obtain the feature vector of the tunnel lining surface; The defect classification output layer outputs the defect identification result based on the feature vector of the tunnel lining surface, and sends the defect identification result to the defect judgment table calling module in the edge computing unit.

[0026] In this embodiment, step 2 involves inputting tunnel lining surface image data into a lightweight neural network defect recognition model deployed on an edge computing unit. The lightweight neural network defect recognition model includes a depthwise separable convolutional feature extraction layer, a channel shuffling feature enhancement layer, and a defect classification output layer connected in sequence. The input layer receives tunnel lining surface image data, with the input image size normalized to 224×224×3. The normalization process uses channel-level normalization, linearly mapping pixel values ​​from 0 to 255 to 0 to 1, and performing mean subtraction. The mean vectors are 0.485, 0.456, and 0.406, and the standard deviation vectors are 0.229, 0.224, and 0.225.

[0027] Depthwise separable convolutional blocks perform channel-wise and pointwise convolution on tunnel lining surface image data to obtain shallow features of the tunnel lining surface. The depthwise separable convolution operation decomposes standard convolution into two stages: the first stage is channel-wise convolution with a 3×3 kernel size, where the number of input channels equals the number of output channels, and the kernel slides independently on each input channel while maintaining the same number of channels in the output feature map; the second stage is pointwise convolution with a 1×1 kernel size, used for linear combination along the channel dimension to achieve information fusion of features from different channels. The stack size of the depthwise separable convolutional blocks is set to 4. Each convolutional block is followed by a batch normalization layer and a ReLU activation function in sequence. The first depthwise separable convolutional block has 3 input channels, 32 output channels, and a stride of 2. The second depthwise separable convolutional block has 32 input channels, 64 output channels, and a stride of 2. The third depthwise separable convolutional block has 64 input channels, 128 output channels, and a stride of 1. The fourth depthwise separable convolutional block has 128 input channels, 128 output channels, and a stride of 1. After processing by the depthwise separable convolutional blocks, the size of the output feature map is 28×28×128.

[0028] The inverted residual block performs channel expansion, depth convolution, and channel compression on the shallow features of the tunnel lining surface to obtain the deep features of the tunnel lining surface. The structure of the inverted residual block is a sequential connection of input, channel expansion, depth convolution, and pointwise convolution compression. In the channel expansion stage, the number of channels of the input feature map is expanded to 6 times the number of input channels, and the expanded number of channels is used to increase the dimension of feature representation. In the depth convolution stage, a 3×3 depth convolution kernel is used to extract spatial dimension features, and the number of depth convolution kernels is equal to the number of expanded channels, with independent convolution for each channel. In the pointwise convolution compression stage, a 1×1 convolution kernel is used to compress the channels. The input channels are compressed back to their original number, and residual connections are added before compression. The inputs of the residual connections are fused with the compressed outputs through element-wise addition. There are 6 inverse residual blocks, divided into 3 stages, with expansion factors of 6, 6, and 6 for each stage. The first stage has 128 input channels and 192 output channels; the second stage has 192 input channels and 384 output channels; and the third stage has 384 input channels and 512 output channels. Each inverse residual block uses the ReLU6 activation function, with an upper limit of 6 for the output of the ReLU6 function.

[0029] The channel shuffling layer performs cross-channel feature exchange on the deep features of the tunnel lining surface to obtain enhanced features. The channel shuffling operation is implemented as follows: First, the channel dimension of the input feature map is divided into multiple groups, with the number of groups set to 4. Then, the channel indices within each group are randomly shuffled using a uniform permutation algorithm to ensure that the features of each group are evenly distributed to other groups. Finally, the shuffled features are reassembled into a complete feature map. The channel shuffling layer is placed between two adjacent inverted residual block groups, and a channel shuffling layer is inserted between the inverted residual blocks in each stage. The channel shuffling layer does not change the size and number of channels of the feature map, but only changes the feature distribution along the channel dimension. The channel shuffling operation enables subsequent network layers to learn richer feature combinations from different feature groups, enhancing the model's ability to express multi-scale defect features.

[0030] The global average pooling layer performs pooling on the enhanced features of the tunnel lining surface, resulting in a feature vector of the tunnel lining surface. The global average pooling is calculated by independently calculating the average value of the spatial dimension for each channel of the feature map, and the dimension of the output vector is the same as the number of channels of the feature map. For a feature map of size 7×7×512, the output vector after global average pooling has a dimension of 512. After the global average pooling layer, a Dropout layer and a fully connected layer are connected in sequence. The dropout ratio of the Dropout layer is set to 0.2, and the output dimension of the fully connected layer is 256, with ReLU as the activation function. The output of the fully connected layer is connected to the defect classification output layer.

[0031] The defect classification output layer outputs defect identification results based on the surface feature vector of the tunnel lining and sends these results to the defect judgment table calling module in the edge computing unit. The defect classification output layer contains two fully connected sublayers. The first sublayer has an input dimension of 256 and an output dimension of 128, with ReLU activation. The second sublayer has an input dimension of 128 and an output dimension equal to the number of defect categories plus one. The defect categories include six types: cracks, water leakage, spalling, spalling, misalignment, and exposed rebar, plus a background category, for a total of seven output categories. Softmax activation is used. For the defect type discrimination task, each element of the output vector represents the defect type. For each category, the category with the highest predicted probability is taken as the defect type prediction result. For the defect severity discrimination task, an independent defect severity regression branch is set up with an input dimension of 256 and an output dimension of 3, corresponding to three defect severity levels: mild, moderate, and severe. The output uses the Sigmoid activation function, with an output value range of 0 to 1. The defect severity level is divided according to preset thresholds, which are set to 0.33 and 0.66. The defect classification output layer also outputs the location regression result of the defect area. The location regression branch has an input dimension of 256 and an output dimension of 4, representing the coordinate offsets of the upper left and lower right corners of the target area, respectively.

[0032] Further, S3 includes: The lightweight neural network defect recognition model uses a depthwise separable convolutional block, an inverted residual block, and a channel washing layer to sequentially extract features from the tunnel lining surface image data, resulting in tunnel lining surface feature maps at multiple scales. The defect region suggestion sublayer in the lightweight neural network defect identification model generates candidate defect regions on the tunnel lining surface feature maps at multiple scales. The candidate region filtering sublayer in the lightweight neural network defect identification model determines the target defect region from the candidate defect regions based on the overlap relationship between the candidate defect regions and the response intensity of the candidate defect regions in the feature map of the tunnel lining surface. The defect type discrimination sublayer in the lightweight neural network defect recognition model performs defect category discrimination on the target defect region to obtain defect type information; The defect severity discrimination sublayer in the lightweight neural network defect recognition model performs defect severity discrimination on the target defect region to obtain defect severity information; The location output sublayer in the lightweight neural network defect recognition model generates defect region location information based on the image coordinates of the target defect region in the tunnel lining surface image data and the detection location identifier. The defect type information, the defect severity information, and the defect area location information are combined to form the defect identification result, and the defect identification result is associated with the detection location identifier, the acquisition time information, and the inspection equipment identifier.

[0033] In this embodiment, step 3 involves using a lightweight neural network defect recognition model to locate and identify defect types in the tunnel lining surface image data, outputting defect recognition results that include defect location information, defect type information, and defect severity information. Depth-separable convolutional blocks, inverted residual blocks, and channel washing layers sequentially extract features from the tunnel lining surface image data to obtain multi-scale tunnel lining surface feature maps. The feature map extraction process starts from the input image, sequentially passing through four depth-separable convolutional blocks and six inverted residual blocks. Each convolutional block and inverted residual block outputs an intermediate feature map. Finally, multi-scale feature maps are extracted after the inverted residual blocks for defect location. The multi-scale feature maps are obtained by setting different layers in the network. The feature pyramid structure uses lower-level feature maps with larger sizes and fewer channels, while higher-level feature maps have smaller sizes and more channels. The specific settings are as follows: the first-scale feature map is taken from the second depthwise separable convolutional block, with a size of 56×56×64; the second-scale feature map is taken from the fourth depthwise separable convolutional block, with a size of 28×28×128; the third-scale feature map is taken from the second inverse residual block, with a size of 14×14×192; the fourth-scale feature map is taken from the fourth inverse residual block, with a size of 7×7×384; and the fifth-scale feature map is taken from the sixth inverse residual block, with a size of 4×4×512. Feature maps at each scale are unified to the same resolution through upsampling or downsampling operations and then fused using element-wise addition.

[0034] The defect region proposal sublayer in the lightweight neural network defect recognition model generates candidate defect regions on the surface feature maps of the tunnel lining at multiple scales. The defect region proposal sublayer is implemented using a region proposal network, which contains two parallel convolutional branches: one for defect region confidence prediction and the other for defect region coordinate regression. The confidence prediction branch has a 1×1 kernel size and 18 output channels, corresponding to 9 candidate region anchors and 2 confidence values ​​(background / foreground). The coordinate regression branch has a 1×1 kernel size and 36 output channels, corresponding to 9 candidate region anchors and 4 coordinate offsets. The candidate region anchors are set using a multi-scale, multi-aspect-ratio strategy, with the anchor scale set to 8. 16, 32, 64, 128, with aspect ratios set to 0.5, 1.0, 2.0; candidate defect areas include at least one of the following: crack candidate area, water leakage candidate area, spalling candidate area, spalling candidate area, misalignment candidate area, and exposed rebar candidate area; the shape of the crack candidate area is a narrow rectangle, with an aspect ratio generally greater than 2; the shape of the water leakage candidate area is an irregular polygon with blurred edges; the shape of the spalling candidate area is approximately circular or elliptical with clear edges; the shape of the spalling candidate area is an irregular concave shape with serrated edges; the shape of the misalignment candidate area is linearly discontinuous with height differences; the shape of the exposed rebar candidate area is a regular line or grid shape with rust marks.

[0035] The candidate region filtering sublayer in the lightweight neural network defect recognition model determines the target defect region from the candidate defect regions based on the overlap relationship between them and the response intensity of the candidate defect regions in the feature map of the tunnel lining surface. The candidate region filtering employs a non-maximum suppression algorithm, whose inputs are the confidence scores and bounding box coordinates of all candidate defect regions. First, the candidate regions are sorted in descending order according to their confidence scores. Then, the sorted candidate region list is traversed, and for each candidate region, the intersection-union ratio (IUU) with the selected candidate regions is calculated. The IUU is calculated using the following formula:

[0036] in, This represents the intersection area of ​​the bounding boxes of two candidate regions. This represents the area of ​​the union of the bounding boxes of two candidate regions. The preset intersection-union ratio (IU) threshold is 0.7. When the IU of a candidate region and the selected region is greater than the threshold, the candidate region is suppressed; otherwise, the candidate region is retained. After traversal, the retained candidate regions are used as target defect regions. The response intensity is determined by the confidence score of the candidate region in the defect region suggestion sub-layer. The score threshold is set to 0.5, and only candidate regions with confidence scores greater than the threshold are retained.

[0037] The defect type discrimination sublayer in the lightweight neural network defect recognition model performs defect category discrimination on the target defect region to obtain defect type information. The defect type information includes at least one of the following: crack type, water leakage type, spalling type, spalling type, misalignment type, and exposed rebar type. The defect type discrimination sublayer performs classification operations on the feature vectors extracted from the target defect region. The classification operation is implemented using a fully connected layer with 7 neurons in the output layer, corresponding to 7 categories. The classification result is the category index corresponding to the maximum probability in the output vector, and the category index is mapped to the corresponding defect type name. The crack type discrimination is based on the geometric and texture features of the target region. The characteristics include length, width, and direction; texture features include the sharpness of crack edges and the linearity of internal texture; the criteria for identifying water leakage types are color distribution and wetting characteristics, with color distribution showing dark colors or water stains, and wetting characteristics showing enhanced surface reflectivity; the criteria for identifying spalling types are the clarity of area edges and internal material characteristics, with edges being arc-shaped or circular and internal concrete spalling marks present; the criteria for identifying flaking types are the shape and depth of the recessed area, with the shape being an irregular concave shape and the depth exceeding a certain threshold; the criteria for identifying misaligned types are the height difference between the two surfaces and linear discontinuity features; the criteria for identifying exposed rebar types are regular linear or grid-like features and rust marks.

[0038] The defect severity sublayer in the lightweight neural network defect recognition model performs defect severity assessment on the target defect region to obtain defect severity information. The defect severity information includes one of three levels: mild, moderate, and severe. The defect severity assessment employs a multi-index comprehensive evaluation method, with evaluation indices including defect size, defect area ratio, defect clarity, and defect density. The defect size index is calculated based on the width and height of the defect region's bounding box; for crack-type defects, length is used as the primary index, while for other types of defects, area is used as the primary index. The defect area ratio index is the ratio of the defect region area to the total image area. The defect clarity index is calculated based on the mean gradient of the defect region's edges. A higher value indicates a clearer edge; the defect density index is the ratio of the number of defect areas to the image area; each evaluation index is normalized and then weighted and summed, with the weight coefficients dynamically adjusted according to the defect type; for crack-type defects, the length weight is set to 0.4, the area ratio weight is set to 0.2, the clarity weight is set to 0.3, and the density weight is set to 0.1; for other types of defects, the area ratio weight is set to 0.4, the length weight is set to 0.1, the clarity weight is set to 0.3, and the density weight is set to 0.2; after the comprehensive score is mapped by Sigmoid, it is divided into three levels: mild, moderate, and severe based on thresholds of 0.33 and 0.66.

[0039] The location output sublayer in the lightweight neural network defect recognition model generates defect region location information based on the image coordinates of the target defect region in the tunnel lining surface image data and the detection location marker. This defect region location information includes the pixel coordinates of the defect region in the image and the corresponding actual coordinates of the tunnel lining. The pixel coordinates are represented by the bounding box of the target defect region in the format x_min, y_min, x_max, y_max, with the top-left corner of the image as the origin. The actual coordinates are transformed based on the odometer information in the detection location marker and the attitude angle during image acquisition. The attitude angle is obtained through the inertial measurement unit (IMU) mounted on the inspection robot, which includes a three-axis accelerometer and a three-axis gyroscope, with an angle measurement accuracy of 0.1 degrees. The rotation matrix and translation vector between the image coordinate system and the tunnel lining coordinate system are calculated based on the attitude angle. The pixel coordinates are then projected onto the tunnel lining surface to obtain the actual coordinates of the defect region on the unfolded diagram of the tunnel lining surface. The actual coordinates are represented in the format of odometer, ring number, and offset within the ring. The ring number is determined based on the number of tunnel lining blocks, and each ring is 1m wide.

[0040] The defect type information, defect severity information, and defect location information are combined into a defect identification result, which is then associated with the detection location identifier, data acquisition time information, and inspection equipment identifier. The data structure of the defect identification result includes the fields `defect_type`, `defect_level`, `defect_position`, `confidence_score`, `bounding_box`, and `feature_vector`. `defect_type` is the defect type string, `defect_level` is an enumeration of defect severity levels (0 for mild, 1 for moderate, 2 for severe), and `defect_position` is the defect location index. The defect identification database contains an array of actual coordinates of the defect area, a confidence score (ranging from 0 to 1) for defect type identification, an array of pixel bounding box coordinates of the defect area, and a feature vector for the defect area, used for subsequent similar defect retrieval and cluster analysis. The defect identification results, along with the detection location identifier, acquisition time information, and inspection equipment identifier, are stored in the defect database in the local storage unit via foreign key association. The database uses SQLite format and contains the tables defect_records and inspection_metadata, which are linked by the inspection_id field.

[0041] Furthermore, referring to Figure 2 S4 includes: The defect determination table pre-installed in the edge computing unit is invoked. The defect determination table records the correspondence between defect type information, defect severity information and identification conclusion. The defect type and defect severity information in the defect identification results are matched with the defect determination table; When the defect identification result does not contain defect type information, the identification conclusion is determined to be a structurally normal identification result; When the defect type information in the defect identification result is water leakage or crack and the defect severity information is mild, the identification conclusion is determined to be a warning structural defect identification result. When the defect type information in the defect identification result is spalling, spalling, misalignment, or exposed rebar, or when the defect type information is crack and the defect severity information is moderate or severe, the identification conclusion is determined to be a reparative structural defect identification result. The identification conclusion is associated with and saved along with the defect identification result, the detection location identifier, the acquisition time information, and the inspection equipment identifier.

[0042] In this embodiment, step 4 involves determining the identification conclusion corresponding to the current tunnel structural state based on the defect identification results; calling the defect judgment table pre-installed in the edge computing unit, which records the correspondence between defect type information, defect severity information, and identification conclusion; the defect judgment table uses a key-value pair mapping structure for storage, where the key is a combination of defect type code and defect severity level, and the value is an identification conclusion code; the defect type code uses numerical encoding, where 0 represents cracks, 1 represents water leakage, 2 represents spalling, 3 represents spalling, 4 represents misalignment, and 5 represents exposed rebar; the defect severity level code uses numerical encoding, where 0 represents mild, 1 represents moderate, and 2 represents severe; and the identification conclusion code uses numerical encoding, where 0 represents normal structure, 1 represents warning structural defect, and 2 represents repairable structural defect.

[0043] The defect type and severity information in the defect identification results are matched with the defect judgment table. The matching process first extracts the defect type code and severity level code from the defect identification results, and then constructs a query key for searching. If the search is successful, the corresponding identification conclusion code is returned; if the search fails, the default identification conclusion of normal structure is returned. The matching logic also includes compound condition judgment. When the defect identification results contain multiple defects of different types or different levels, the identification conclusion is determined according to the principle of choosing the highest level, that is, the identification conclusion with the highest level among all defects is taken as the final identification conclusion.

[0044] When the defect identification result does not include defect type information, the identification conclusion is determined to be a normal structural identification result. The absence of defect type information includes situations where the defect type field is empty, the defect type field value is "background," or the defect type confidence score is below 0.3. When the defect type information in the defect identification result is water leakage or cracking and the defect severity information is "mild," the identification conclusion is determined to be a warning structural defect identification result. Water leakage and cracking at the "mild" level are defined as warning levels, indicating that the defect currently does not affect the tunnel structure safety but requires continuous monitoring and verification during subsequent inspections. When the defect type information in the defect identification result is spalling, spalling, misalignment, or exposed rebar, or when the defect type information is cracking and the defect severity information is "moderate" or "severe," the identification conclusion is determined to be a repairable structural defect identification result. Repairable structural defects are defined as defects requiring immediate repair work. Spalling, spalling, misalignment, and exposed rebar defects, regardless of severity, are directly determined to be repairable.

[0045] The identification conclusion is associated with and saved along with the defect identification result, detection location identifier, acquisition time information, and inspection equipment identifier. The associated data record contains the fields conclusion_id, inspection_id, defect_record_id, conclusion_type, conclusion_timestamp, position_id, device_id, and metadata. Among them, conclusion_id is the unique identifier of the identification conclusion record, inspection_id is the inspection record identifier, defect_record_id is the defect record identifier, conclusion_type is the identification conclusion type code, and conclusion_timestamp is the identification conclusion generation time. The associated data is written to the conclusion database in the local storage unit. The database adopts the same SQLite format as the defect database and contains the table conclusion_records. The table structure is associated with the defect record table through the inspection_id field.

[0046] Furthermore, when the identification conclusion is a structurally normal identification result, an inspection record that does not require manual review is generated, including: Generate a normal structural inspection record, which includes the detection location identifier, the acquisition time information, the inspection equipment identifier, the image identifier corresponding to the tunnel lining surface image data, the identification conclusion, and the model version identifier of the lightweight neural network defect identification model; Write the normal inspection record of the structure into the local storage unit, and when the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, upload the normal inspection record of the structure to the tunnel operation and maintenance database in the operation and maintenance server. In the inspection display interface of the subway tunnel inspection equipment, the tunnel lining section corresponding to the detection location is marked as being in normal structural condition; the subway tunnel inspection equipment is controlled to continue moving to the next preset detection location, and the next round of tunnel lining surface image data acquisition is triggered when the next preset detection location is reached.

[0047] Furthermore, when the identification conclusion is a warning structural defect identification result, structural defect warning information is generated, including: Based on the defect type information, defect severity information, and defect area location information in the defect identification results, structural defect early warning information is generated; The structural defect early warning information is output to the local early warning output unit of the subway tunnel inspection equipment, and the local early warning output unit performs audible and visual prompts. The structural defect early warning information is sent to the operation and maintenance terminal, which maps the location of the early warning defect area to the corresponding tunnel lining section in the tunnel lining unfolded diagram, and displays the early warning mark on the tunnel lining section; the structural defect early warning information is associated with the tunnel lining component identifier in the tunnel structural component ledger to generate a task to be reviewed; The task to be reviewed is written into the tunnel operation and maintenance database and displayed in the task list of the operation and maintenance terminal. At the same time, the task to be reviewed is associated with the tunnel lining section corresponding to the detection location identifier.

[0048] Furthermore, referring to Figure 3 When the identification conclusion is a remedial structural defect identification result, structural repair early warning information is generated, including: Based on the defect type information, defect severity information, and defect area location information in the defect identification results, structural repair early warning information is generated; The structural repair early warning information is output to the maintenance work order system, which then generates a structural repair work order. The structural repair early warning information is associated with the tunnel lining component identifier in the tunnel structural component ledger, and the component status corresponding to the tunnel lining component identifier is marked as pending repair. The structural repair work order is sent to the operation and maintenance terminal, which then displays the structural repair work order in the task list according to the repair priority. The structural repair early warning information and the structural repair work order are written into the tunnel operation and maintenance database, and the structural repair work order is associated with and saved with the tunnel lining section corresponding to the detection location identifier.

[0049] In this embodiment, step 5 involves generating corresponding output information based on the identification conclusion. When the identification conclusion is "structurally normal," an inspection record that does not require manual review is generated. The specific process for generating the structurally normal inspection record is as follows: an inspection record data object is created, which includes a detection location identifier, acquisition time information, inspection equipment identifier, image identifier corresponding to the tunnel lining surface image data, identification conclusion, and model version identifier of the lightweight neural network defect identification model. The detection location identifier is the location code of the current inspection location; the acquisition time information is the UTC timestamp when the image was acquired; the inspection equipment identifier is the serial number of the equipment performing this inspection; the image identifier is the unique identifier of the image acquired this time; the identification conclusion is "structurally normal"; and the model version identifier is the currently loaded lightweight neural network defect identification model. The version number of the neural network defect identification model is specified in the format "major version number; minor version number; revision number", for example, "2.1.3". An example of the JSON format for the inspection record is: {"inspection_id":"INS-2024-1001","position_id":"0101-001245","timestamp":"2024-01-15T08:30:45.123Z","device_id":"TR-2024-001","image_id":"IMG-a1b2c3d4","conclusion":"NORMAL","model_version":"2.1.3"}.

[0050] Normal structural inspection records are written to the inspection record database in the local storage unit. The database table is named inspection_records and contains fields such as inspection_id, position_id, timestamp, device_id, image_id, conclusion, and model_version. After writing, the communication connection status between the subway tunnel inspection equipment and the operation and maintenance server is checked. The communication connection status is determined by a heartbeat detection mechanism with a heartbeat interval of 30 seconds and a timeout threshold of 90 seconds. When the communication connection is available, the edge computing unit uploads the normal structural inspection records to the tunnel operation and maintenance database interface of the operation and maintenance server via HTTP protocol. The upload interface address is " / api / v1 / inspection / upload", the request method is POST, the request body format is JSON, and a response status code of 200 indicates successful upload. When the communication connection is unavailable, the edge computing unit caches the normal structural inspection records in the upload queue of the local storage unit and waits for the communication connection to be restored before performing batch upload.

[0051] In the inspection display interface of the subway tunnel inspection equipment, the tunnel lining section corresponding to the detection location is marked as being in a normal structural state. The inspection display interface uses a graphical method to display the unfolded diagram of the tunnel lining. The unfolded diagram is pre-drawn according to the tunnel design parameters, with ring number and mileage markings located on both sides of the unfolded diagram. The normal structural state is marked by filling the corresponding section with a green mark, with the transparency of the green fill set to 0.3 and the mark continuously displayed for 5 seconds. The subway tunnel inspection equipment is controlled to continue moving to the next preset detection position at a speed of 2 km / h. During the movement, the supplementary lighting unit is kept off to save power. When the next preset detection position is reached, the next round of tunnel lining surface image data acquisition is triggered, and the processing flow of the next detection cycle begins.

[0052] When the identification conclusion is a warning structural defect, a structural defect warning message is generated. Based on the defect type, defect severity, and defect location information in the defect identification result, the structural defect warning message is generated. The structural defect warning message includes the warning defect type, warning defect severity, warning defect location, detection location identifier, acquisition time information, inspection equipment identifier, image identifier, and model version identifier. The warning defect type is the defect type field value in the defect identification result; the warning defect severity is the defect severity field value in the defect identification result; the warning defect location is the defect location information in the defect identification result, including pixel coordinates and actual coordinates. The detection location identifier, acquisition time information, inspection equipment identifier, image identifier, and model version identifier originate from the same source as the normal structural inspection record. An example of the JSON format for the structural defect warning message is: {"warning_id": "WRN-2024-5001", "defect_type": "CRACK", "defect_level": 0, "defect_position": {"pixel": 120, 350, 280, 480, "actual":} 001245, 13, 0.5}, "position_id": "0101-001245", "timestamp": "2024-01-15T08:30:45.123Z", "device_id": "TR-2024-001", "image_id": "IMG-a1b2c3d4", "model_version": "2.1.3"}.

[0053] The structural defect early warning information is output to the local early warning output unit of the subway tunnel inspection equipment, which provides both audible and visual alerts. The local early warning output unit includes a speaker and an LED indicator. The speaker outputs a pulsed alert tone with an audio frequency of 800Hz to 1000Hz, a pulse interval of 500ms, and a duration of 3 seconds. The LED indicator uses a yellow light source, flashes at a frequency of 1Hz, and has a duty cycle of 50%. The structural defect early warning information is then sent to the maintenance terminal, which is a remote monitoring computer deployed in the operation control center. Communication is via a 4G or 5G wireless network, using the TCP protocol and port number 8888. Upon receiving the early warning information, the maintenance terminal maps the location of the early warning defect area to the corresponding tunnel lining section in the tunnel lining unfolded diagram. The mapping method calculates the section number and ring number based on the actual coordinates in the early warning defect area location information and draws an early warning mark at the corresponding position in the unfolded diagram. The early warning mark is a yellow circular icon with a diameter of 20 pixels, and the icon displays the abbreviation of the defect type.

[0054] The structural defect early warning information is associated with the tunnel lining component identifiers in the tunnel structural component ledger to generate a review task. The tunnel structural component ledger is stored in the database of the operation and maintenance server, containing a table named tunnel_components with fields including component_id, position_range, maintenance_level, status, and last_inspection_date. The association method is to query the corresponding component record based on the mileage range of the early warning defect area, establish the association, and then generate a review task record. The review task includes the early warning defect type, early warning defect level, early warning defect area location, image identifier, and detection location identifier. An example of the JSON format for the review task is: {"task_id": "TASK-2024-2001", "defect_type": "CRACK", "defect_level": 0, "position":001245, 13, 0.5, "image_id": "IMG-a1b2c3d4", "position_id": "0101-001245", "created_at":} "2024-01-15T08:30:46.000Z", "status": "PENDING"}; The task to be reviewed is written to the task_records table in the tunnel operation and maintenance database, and the task to be reviewed is displayed in the task list of the operation and maintenance terminal; The task list is displayed in reverse chronological order, and each task record displays the task number, defect type, defect severity, location information and status label, with the status label in yellow; After receiving the task to be reviewed, the operation and maintenance personnel can view the task details through the operation and maintenance terminal, and can choose to assign the task to the on-site inspection personnel for on-site review.

[0055] When the identification result is a remedial structural defect, a structural repair early warning message is generated. Based on the defect type, severity, and location information of the defect identification result, the structural repair early warning message is generated. This message includes the defect type, severity, location of the defect area, detection location identifier, acquisition time, inspection equipment identifier, image identifier, and model version identifier. The sources of the defect type and severity are the same as those in the early warning message. Repair priority is calculated comprehensively based on the defect type and severity. For the defect type weight array, cracks are 0.7, water leakage is 0.6, spalling is 0.9, spalling is 0.95, misalignment is 0.85, and exposed rebar is 0.8. For the defect severity weight array, mild is 0.3, moderate is 0.6, and severe is 0.95. Higher priority values ​​indicate greater urgency, ranging from 0 to 1.9.

[0056] The structural repair early warning information is output to the maintenance work order system, which generates a structural repair work order. The maintenance work order system is a business application system on the operation and maintenance server. After receiving the structural repair early warning information, it automatically creates a work order record. The structural repair work order includes the type of repair defect, the degree of repair defect, the location of the repair defect area, image identification, and inspection equipment identification. The work order status is initially set to pending dispatch, and the work order number is automatically generated in the format "WO-year-serial number", such as "WO-2024-0001". The structural repair early warning information is associated with the tunnel lining component identification in the tunnel structural component ledger. The association method is the same as that for early warning information processing, and the component status corresponding to the tunnel lining component identification is marked as pending repair. The component status field is updated to "REPAIR_PENDING", and the last maintenance time field is updated to the current timestamp.

[0057] Structural repair work orders are sent to the maintenance terminal, which then displays them in the task list according to their repair priority. Work orders in the task list are displayed in descending order of priority, with higher priority values ​​appearing earlier. Each work order record includes the work order number, defect type, defect severity, location information, priority value, and status label (red). The maintenance terminal supports work order assignment, allowing maintenance personnel to assign work orders to designated maintenance teams. Assignment information is sent to the maintenance team's mobile terminals via the work order system. Structural repair warning information and work orders are written to the tunnel maintenance database. Warning information is written to the `warning_records` table, and work orders are written to the `work_orders` table. Work order records are linked to the tunnel lining sections corresponding to the inspection location identifiers via foreign keys, and related queries support searching the work order list by section.

[0058] The edge computing unit is deployed on the subway tunnel inspection equipment, and the lightweight neural network defect identification model runs in the edge computing unit. When the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, the edge computing unit uploads the defect identification results, structural defect early warning information, and structural repair early warning information to the operation and maintenance server. The upload method is batch upload, with each upload containing 50 records and an upload interval of 60 seconds. The upload interface address is " / api / v1 / defect / sync", the request method is POST, and the request body format is a JSON array, with each array element containing all fields of one record. When the communication connection is unavailable, the edge computing unit saves the defect identification results, structural defect early warning information, and structural repair early warning information to the pending upload cache directory of the local storage unit. The cache directory adopts a circular buffer structure with a capacity of 10,000 records. When the cache is full, it automatically overwrites the earliest written record. After the communication connection is restored, the edge computing unit detects the network availability and immediately triggers a batch upload task, uploading all pending upload records in the cache directory in chronological order. During the upload process, a breakpoint resume mechanism is used. Records that have been successfully uploaded are marked as uploaded, and records that failed to upload are retained in the cache directory for re-upload.

[0059] The edge computing unit receives a pre-configured lightweight neural network defect recognition model parameter package from the operations and maintenance server, loads the parameter package into the lightweight neural network defect recognition model, and records the model version identifier after loading. The model parameter package is distributed after the inspection robot completes initialization or when the model version is updated. The distribution interface address is " / api / v1 / model / push", the request method is POST, and the parameter package format is TensorFlow Lite model format or ONNX model format, with a file size typically between 5MB and 20MB. After receiving the parameter package, the edge computing unit first performs integrity verification by calculating the MD5 hash value of the parameter package file and comparing it with the hash value provided by the server. If the verification passes, the parameter package is written to the model storage directory, replacing the old version model file. The loading process includes creating an inference engine instance, allocating a memory buffer, loading model weights, and performing model warm-up. During the warm-up process, three inferences are performed using preset test images to ensure consistency of inference results. The model version identifier is updated to the new version number, and the version switch timestamp is recorded.

[0060] When the subway tunnel inspection equipment reaches the subsequent preset inspection position, the edge computing unit uses the lightweight neural network defect recognition model loaded with the lightweight neural network defect recognition model parameter package to perform defect area localization and defect type discrimination on the subsequently collected tunnel lining surface image data, outputs the subsequent defect recognition results, and determines the subsequent recognition conclusion based on the subsequent defect recognition results; the subsequent detection process is completely consistent with the processing flow of steps 1 to 5 above, only the data identifier and model version identifier are updated to the current value; when the subsequent recognition conclusion is a normal structural recognition result, an inspection record that does not require manual review is generated; when the subsequent recognition conclusion is a warning structural defect recognition result, a subsequent structural defect warning information is generated; when the subsequent recognition conclusion is a repairable structural defect recognition result, a subsequent structural repair warning information is generated; the subsequent defect recognition results, subsequent structural defect warning information, and subsequent structural repair warning information are associated and saved with the detection position identifier, acquisition time information, and inspection equipment identifier corresponding to the subsequent preset inspection position, and the associated data structure is consistent with the aforementioned format.

[0061] In the lightweight neural network defect recognition model, the depthwise separable convolutional block employs depthwise separable convolution operations. This operation decomposes standard convolution into two stages: channel-wise convolution and pointwise convolution. In the channel-wise convolution stage, each channel of the input feature map is independently convolved with a 3×3 kernel, a stride of 1 or 2, and the same padding method. The number of channels in the output feature map is the same as the number of input channels. In the pointwise convolution stage, a 1×1 kernel is used to fuse the channel dimension information of the channel-wise convolution output. The number of kernels is equal to the number of target output channels. Compared with standard convolution, this structure can significantly reduce the number of parameters and computational complexity.

[0062] When the convolution kernel size is 3×3 and the number of output channels is 64, the number of parameters is reduced by about 8 to 9 times; the reduction in computational complexity is comparable to the reduction in the number of parameters, enabling the model to be deployed on the embedded processor of the edge computing unit to run in real time, and the inference time for a single image is controlled within 80ms.

[0063] The lightweight neural network defect recognition model employs a three-stage structure for the inverse residual block: channel expansion, depthwise convolution, and channel compression. The channel expansion stage expands the number of channels in the input feature map to six times the input channel count, using a 1×1 pointwise convolution with ReLU6 activation. The depthwise convolution stage uses a 3×3 kernel to extract spatial dimensions from the expanded feature map, with a stride of 1 or 2, same padding, and ReLU6 activation. The channel compression stage restores the number of channels in the feature map to the same or close to the input channel count, using a 1×1 pointwise convolution without setting an activation function. Residual connections are added before channel compression, using element-wise addition, requiring the input feature map size to be the same as the pre-compression feature map size. When the stride is 2, the input feature map is downsampled to the same size as the intermediate feature map through average pooling before residual connections are made. This structure effectively controls the number of model parameters while maintaining feature representation capabilities; the number of parameters in the inverse residual block is approximately one-third that of the standard residual block.

[0064] In the lightweight neural network defect recognition model, the channel shuffling layer performs cross-channel feature exchange operations on the input feature map. The channel dimension of the input feature map is divided into multiple groups, with the number of groups depending on the network depth, typically set to 4 or 8. The channel division is based on a uniform partition of the channel index, and the number of channels in each group after partitioning is the total number of channels divided by the number of groups. The channel features within each group are shuffled and rearranged using a deterministic permutation algorithm. For a group with index g, the rearranged group index is (g+1) mod the number of groups. This rearrangement operation ensures that the output of each group contains a mixture of input features from other groups, enhancing the information flow between channels. The channel shuffling operation can be implemented using transpose, recombination, or gather / scatter instructions, with consistent results from different implementations. This operation helps enhance the efficiency of information transfer between channels and improves the model's ability to extract multi-scale defect features. Experiments show that the channel shuffling operation can improve defect classification accuracy by 2% to 3%.

[0065] In the lightweight neural network defect recognition model, the defect region suggestion sublayer adopts a multi-scale candidate region generation strategy. Candidate defect regions are generated on feature maps at multiple different levels, with each level corresponding to a different receptive field size and semantic level. Low-level feature maps, such as the first and second scale feature maps, provide fine location information and texture features, suitable for small-sized defects and detailed feature extraction. Intermediate-level feature maps, such as the third and fourth scale feature maps, provide medium-scale semantic information, suitable for detecting defects of regular size. High-level feature maps, such as the fifth scale feature map, provide abstract semantic information, suitable for large-sized defects and overall structure evaluation. Multi-scale fusion is achieved through a feature pyramid network, with the fusion method being top-down feature transfer and addition of features at the same scale. Anchor points of different sizes are set on feature maps at different scales to ensure that each scale effectively covers the defect region within its corresponding size range.

[0066] In the lightweight neural network defect recognition model, the candidate region screening sublayer employs a non-maximum suppression algorithm (NMS) to screen candidate defect regions. The main parameters of the NMS algorithm include the intersection-union ratio (IU) threshold and the confidence threshold. The default values ​​for the IU threshold and the confidence threshold are 0.5 and 0.7 respectively. The IU is calculated based on the degree of overlap between candidate defect regions. When the overlap between two candidate regions exceeds the preset threshold, the candidate region with higher confidence is retained while the candidate region with lower confidence is suppressed. The screening process is iterative. In each iteration, the region with the highest confidence among the remaining candidate regions is selected as the output, and other candidate regions with an IU greater than the threshold are suppressed. This process is repeated until the candidate region list is empty. The number of target defect regions after screening is reduced, typically from hundreds of candidate regions to dozens of target regions, effectively reducing the resource consumption for subsequent classification and regression calculations while lowering the false detection rate.

[0067] The defect judgment table is implemented using a predefined rule engine. The table records the mapping relationship between defect type information, defect severity information, and identification conclusions. The rule engine is implemented using a decision tree structure, where the root node represents the defect type, intermediate nodes represent the defect severity, and leaf nodes represent the identification conclusions. The decision tree traversal starts from the root node, first selecting the corresponding branch based on the defect type, then selecting the next-level branch based on the defect severity, and finally reaching the leaf node to obtain the identification conclusion. The mapping rules are formulated based on subway tunnel structure maintenance standards and industry specifications, specifically including the "Subway Design Code," "Urban Rail Transit Structure Maintenance Technical Specification," and "Railway Tunnel Maintenance and Repair Rules." The rule engine supports dynamic updates. Updates are made by maintenance personnel through the configuration interface of the maintenance terminal to modify the judgment rules, and the modified rules are synchronized to the edge computing unit through the configuration distribution interface. Regularly, maintenance personnel adjust the rules based on actual detection results and expert review opinions to ensure that the judgment criteria for identification conclusions are consistent with the latest maintenance requirements.

[0068] The inspection record contains six data fields: detection location identifier, acquisition time information, inspection equipment identifier, image identifier, recognition conclusion, and model version identifier. Each field is serialized and stored according to a unified data format and encoding standard. The detection location identifier is a string with a length not exceeding 20 characters and encoded in UTF-8. The acquisition time information uses the ISO 8601 standard time format, accurate to milliseconds. The inspection equipment identifier is a string with a length not exceeding 30 characters. The image identifier uses the UUID format. The recognition conclusion uses an enumeration type, supporting three values: NORMAL, WARNING, and REPAIR. The model version identifier uses a semantic version number format. Each field is stored in an SQLite database table in the local storage unit, with field types and constraints corresponding to the data format. The model version identifier record provides a quantifiable basis for model performance traceability and effect evaluation. By querying the inspection records and defect recognition results corresponding to a specific model version, performance indicators such as recognition accuracy and false negative rate for that version can be statistically analyzed.

[0069] The tunnel lining sections corresponding to the tasks awaiting review and the inspection location markers are graphically associated in the task list on the maintenance terminal. The task list is displayed in tabular form, with columns including task number, defect type, defect severity, location information, creation time, and status. Each row corresponds to one task awaiting review. The graphical association is achieved by overlaying marker icons on the unfolded tunnel lining diagram. These marker icons are associated with task list records using unique identifiers. Clicking a marker icon on the unfolded diagram will bring up a task details panel, which displays complete task information and associated image thumbnails. Maintenance personnel can access this information through the maintenance terminal. The terminal allows direct viewing of the original image data, defect type, and defect severity information corresponding to the task to be reviewed. The original image is viewed by clicking the image identifier link, which points to the image storage path on the operation and maintenance server. After the image is loaded, it is displayed in the image viewer, with a bounding box marker overlaid on the defect area. After completing the on-site review, the operation and maintenance personnel can fill in the review results through the task processing interface of the operation and maintenance terminal. The review results include the review status, review remarks, and on-site photos. After the review is completed, the review results are backfilled to the tunnel operation and maintenance database through the data upload interface, realizing a closed loop between the defect identification process and the manual review process.

[0070] Furthermore, the edge computing unit is deployed in the subway tunnel inspection equipment, and the lightweight neural network defect recognition model runs in the edge computing unit; When the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, the edge computing unit uploads the defect identification results, the structural defect early warning information and the structural repair early warning information to the operation and maintenance server. When the communication connection is unavailable, the edge computing unit saves the defect identification result, the structural defect early warning information and the structural repair early warning information to the local storage unit, and uploads the defect identification result, the structural defect early warning information and the structural repair early warning information saved in the local storage unit to the operation and maintenance server after the communication connection is restored; During the upload process, the detection location identifier, the acquisition time information, and the inspection equipment identifier are sent together with the defect identification result, the structural defect early warning information, and the structural repair early warning information. The operation and maintenance server then writes the uploaded information into the tunnel operation and maintenance database.

[0071] Furthermore, the edge computing unit receives a pre-configured lightweight neural network defect identification model parameter package from the operation and maintenance server, loads the lightweight neural network defect identification model parameter package into the lightweight neural network defect identification model, and records the model version identifier after loading. When the subway tunnel inspection equipment reaches the subsequent preset detection position, the edge computing unit uses the lightweight neural network defect recognition model loaded with the parameter package of the lightweight neural network defect recognition model to perform defect area localization and defect type discrimination on the subsequently collected tunnel lining surface image data, output the subsequent defect recognition results, and determine the subsequent recognition conclusion based on the subsequent defect recognition results; When the subsequent identification conclusion is a normal structural identification result, an inspection record is generated that does not require subsequent manual review. When the subsequent identification conclusion is a warning structural defect identification result, a subsequent structural defect warning information is generated; When the subsequent identification conclusion is a corrective structural defect identification result, a subsequent structural repair early warning information is generated, and the subsequent defect identification result, the subsequent structural defect early warning information, and the subsequent structural repair early warning information are associated and saved with the detection location identifier, collection time information, and inspection equipment identifier corresponding to the subsequent preset detection location.

[0072] According to another embodiment of the present invention, when the subway tunnel inspection equipment is implemented using a drone platform, the detailed contents of the image acquisition and data processing flow are as follows: The drone platform adopts a multi-rotor drone, equipped with a lightweight image acquisition unit and an edge computing module; the flight speed of the drone platform is set to 3 km / h to 8 km / h, which can be adjusted according to the tunnel section conditions; the preset detection positions are set using an equidistant layout method, with the distance between adjacent detection positions set to 10m; the drone platform hovers at the preset detection position and adjusts its attitude, with the attitude adjustment time set to 3 seconds to 5 seconds to ensure that the viewing angle of the image acquisition unit is aligned with the target area; the image acquisition unit adopts a wide-angle lens with a focal length of 2.8mm, a field of view of 120 degrees, an acquisition distance set to 0.5m to 2m, and an image resolution of 3840×2160 pixels.

[0073] The edge computing module uses an embedded GPU module equipped with 4GB of video memory and 8GB of system memory, supporting CUDA acceleration and the TensorRT inference engine. The lightweight neural network defect identification model is optimized and compiled for the embedded GPU module, with optimization options including operator fusion, memory optimization, and quantization processing. The model quantization adopts the INT8 quantization method, and the quantization calibration dataset contains 1,000 representative tunnel lining images. After quantization, the model's inference speed is improved by approximately 2 to 3 times. Communication between the UAV platform and the operation and maintenance server is achieved through a wireless bridge. When the UAV platform is flying inside the tunnel, it establishes a network connection through a pre-deployed wireless access point. When the communication connection between the UAV platform and the operation and maintenance server is available, the defect identification results are uploaded in real time. When the communication connection is unavailable, the UAV platform caches the defect identification results in the onboard storage module and exports the data upon returning to the ground station after the flight mission.

[0074] When a rail-mounted inspection vehicle is used for subway tunnel inspection, the detailed process of image acquisition and data processing is as follows: The rail-mounted inspection vehicle travels along a fixed track laid on the maintenance access roads on both sides of the tunnel; the vehicle's speed is set to 10 km / h to 20 km / h, driven by a servo motor, with a positioning accuracy of ±5 cm; the image acquisition unit uses a linear scan camera with a line resolution of 4096 pixels, a line frequency of 15 kHz, and continuous scanning, with the scanning speed synchronized with the vehicle's movement speed; the supplementary lighting unit uses a linear LED light source with a length of 500 mm, covering the visible and near-infrared spectral range.

[0075] The edge computing unit of the track-mounted inspection vehicle uses a high-performance industrial computer equipped with an Intel Core i7 processor and an NVIDIA Jetson Xavier module. The lightweight neural network defect recognition model runs on the Jetson Xavier module, utilizing CUDA cores and Tensor Cores for accelerated computation. Model inference adopts a pipelined approach, with image acquisition, model inference, and result output executed in parallel. The pipeline depth is set to 3 frames to ensure real-time processing. The data storage module of the track-mounted inspection vehicle uses a RAID1 disk array with a capacity of 2TB, supporting data redundancy protection. Defect recognition results are uploaded to the operation and maintenance server in real time via a 5G communication module installed at the front of the vehicle. The communication module uses an M.2 interface 5G module, supporting both NSA and SA network modes, with a downlink peak rate of 2Gbps and an uplink peak rate of 1Gbps.

[0076] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for identifying structural defects in subway tunnels based on lightweight neural networks, characterized in that, include: S1, acquire image data of the tunnel lining surface collected by the subway tunnel inspection equipment at the preset detection position; S2, input the tunnel lining surface image data into a lightweight neural network defect recognition model deployed on an edge computing unit. The lightweight neural network defect recognition model includes a depth-separable convolutional feature extraction layer, a channel shuffling feature enhancement layer, and a defect classification output layer connected in sequence. S3, using the lightweight neural network defect recognition model, perform defect region localization and defect type discrimination on the tunnel lining surface image data, and output defect recognition results containing defect region location information, defect type information and defect degree information; S4. Determine the identification conclusion corresponding to the current tunnel structure state based on the defect identification results; S5, when the identification conclusion is a normal structural identification result, an inspection record that does not require manual review is generated; when the identification conclusion is a warning structural defect identification result, a structural defect warning message is generated; when the identification conclusion is a repairable structural defect identification result, a structural repair warning message is generated.

2. The method according to claim 1, characterized in that, S1 includes: The subway tunnel inspection equipment is an inspection robot that moves along the subway tunnel track. The inspection robot is equipped with the edge computing unit, image acquisition unit, supplementary lighting unit and local storage unit. The inspection robot is controlled to move along the subway tunnel track, and the image acquisition unit and the supplementary lighting unit are triggered when it reaches the preset detection position. The image acquisition unit acquires image data of the tunnel lining surface towards the tunnel lining surface, and the edge computing unit records the detection position identifier, acquisition time information, and inspection equipment identifier corresponding to the preset detection position; The tunnel lining surface image data is associated with the detection location identifier, the acquisition time information and the inspection equipment identifier and stored in the local storage unit, and the tunnel lining surface image data is input into the lightweight neural network defect recognition model in the edge computing unit.

3. The method according to claim 2, characterized in that, The lightweight neural network defect recognition model is deployed on the edge computing unit. The lightweight neural network defect recognition model includes an input layer, a depthwise separable convolutional block, an inverted residual block, a channel shuffling layer, a global average pooling layer, and a defect classification output layer. The input layer receives the image data of the tunnel lining surface; The depth-separable convolutional block performs channel-wise convolution and point-wise convolution on the tunnel lining surface image data to obtain shallow features of the tunnel lining surface. The inverted residual block performs channel expansion, depth convolution, and channel compression on the shallow features of the tunnel lining surface to obtain the deep features of the tunnel lining surface. The channel washing layer performs cross-channel feature exchange on the deep features of the tunnel lining surface to obtain enhanced features of the tunnel lining surface. The global average pooling layer performs pooling on the enhanced features of the tunnel lining surface to obtain the feature vector of the tunnel lining surface; The defect classification output layer outputs the defect identification result based on the feature vector of the tunnel lining surface, and sends the defect identification result to the defect judgment table calling module in the edge computing unit.

4. The method according to claim 3, characterized in that, S3 includes: The lightweight neural network defect recognition model uses a depthwise separable convolutional block, an inverted residual block, and a channel washing layer to sequentially extract features from the tunnel lining surface image data, resulting in tunnel lining surface feature maps at multiple scales. The defect region suggestion sublayer in the lightweight neural network defect identification model generates candidate defect regions on the tunnel lining surface feature maps at multiple scales. The candidate region filtering sublayer in the lightweight neural network defect identification model determines the target defect region from the candidate defect regions based on the overlap relationship between the candidate defect regions and the response intensity of the candidate defect regions in the feature map of the tunnel lining surface. The defect type discrimination sublayer in the lightweight neural network defect recognition model performs defect category discrimination on the target defect region to obtain defect type information; The defect severity discrimination sublayer in the lightweight neural network defect recognition model performs defect severity discrimination on the target defect region to obtain defect severity information; The location output sublayer in the lightweight neural network defect recognition model generates defect region location information based on the image coordinates of the target defect region in the tunnel lining surface image data and the detection location identifier. The defect type information, the defect severity information, and the defect area location information are combined to form the defect identification result, and the defect identification result is associated with the detection location identifier, the acquisition time information, and the inspection equipment identifier.

5. The method according to claim 4, characterized in that, S4 includes: The defect determination table pre-installed in the edge computing unit is invoked. The defect determination table records the correspondence between defect type information, defect severity information and identification conclusion. The defect type and defect severity information in the defect identification results are matched with the defect determination table; When the defect identification result does not contain defect type information, the identification conclusion is determined to be a structurally normal identification result; When the defect type information in the defect identification result is water leakage or crack and the defect severity information is mild, the identification conclusion is determined to be a warning structural defect identification result. When the defect type information in the defect identification result is spalling, spalling, misalignment, or exposed rebar, or when the defect type information is crack and the defect severity information is moderate or severe, the identification conclusion is determined to be a reparative structural defect identification result. The identification conclusion is associated with and saved along with the defect identification result, the detection location identifier, the acquisition time information, and the inspection equipment identifier.

6. The method according to claim 5, characterized in that, When the identification conclusion is a normal structural identification result, an inspection record is generated that does not require manual review, including: Generate a normal structural inspection record, which includes the detection location identifier, the acquisition time information, the inspection equipment identifier, the image identifier corresponding to the tunnel lining surface image data, the identification conclusion, and the model version identifier of the lightweight neural network defect identification model; Write the normal inspection record of the structure into the local storage unit, and when the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, upload the normal inspection record of the structure to the tunnel operation and maintenance database in the operation and maintenance server. In the inspection display interface of the subway tunnel inspection equipment, the tunnel lining section corresponding to the detection location is marked as being in normal structural condition; the subway tunnel inspection equipment is controlled to continue moving to the next preset detection location, and the next round of tunnel lining surface image data acquisition is triggered when the next preset detection location is reached.

7. The method according to claim 6, characterized in that, When the identification conclusion is a warning structural defect identification result, structural defect warning information is generated, including: Based on the defect type information, defect severity information, and defect area location information in the defect identification results, structural defect early warning information is generated; The structural defect early warning information is output to the local early warning output unit of the subway tunnel inspection equipment, and the local early warning output unit performs audible and visual prompts. The structural defect early warning information is sent to the operation and maintenance terminal, which maps the location of the early warning defect area to the corresponding tunnel lining section in the tunnel lining unfolded diagram, and displays the early warning mark on the tunnel lining section; the structural defect early warning information is associated with the tunnel lining component identifier in the tunnel structural component ledger to generate a task to be reviewed; The task to be reviewed is written into the tunnel operation and maintenance database and displayed in the task list of the operation and maintenance terminal. At the same time, the task to be reviewed is associated with the tunnel lining section corresponding to the detection location identifier.

8. The method according to claim 7, characterized in that, When the identification conclusion is a remedial structural defect identification result, structural repair early warning information is generated, including: Based on the defect type information, defect severity information, and defect area location information in the defect identification results, structural repair early warning information is generated; The structural repair early warning information is output to the maintenance work order system, which then generates a structural repair work order. The structural repair early warning information is associated with the tunnel lining component identifier in the tunnel structural component ledger, and the component status corresponding to the tunnel lining component identifier is marked as pending repair. The structural repair work order is sent to the operation and maintenance terminal, which then displays the structural repair work order in the task list according to the repair priority. The structural repair early warning information and the structural repair work order are written into the tunnel operation and maintenance database, and the structural repair work order is associated with and saved with the tunnel lining section corresponding to the detection location identifier.

9. The method according to claim 8, characterized in that, The edge computing unit is deployed in the subway tunnel inspection equipment, and the lightweight neural network defect recognition model runs in the edge computing unit. When the communication connection between the subway tunnel inspection equipment and the operation and maintenance server is available, the edge computing unit uploads the defect identification results, the structural defect early warning information and the structural repair early warning information to the operation and maintenance server. When the communication connection is unavailable, the edge computing unit saves the defect identification result, the structural defect early warning information and the structural repair early warning information to the local storage unit, and uploads the defect identification result, the structural defect early warning information and the structural repair early warning information saved in the local storage unit to the operation and maintenance server after the communication connection is restored; During the upload process, the detection location identifier, the acquisition time information, and the inspection equipment identifier are sent together with the defect identification result, the structural defect early warning information, and the structural repair early warning information. The operation and maintenance server then writes the uploaded information into the tunnel operation and maintenance database.

10. The method according to claim 9, characterized in that, The edge computing unit receives a pre-configured lightweight neural network defect identification model parameter package from the operation and maintenance server, loads the lightweight neural network defect identification model parameter package into the lightweight neural network defect identification model, and records the model version identifier after loading. When the subway tunnel inspection equipment reaches the subsequent preset detection position, the edge computing unit uses the lightweight neural network defect recognition model loaded with the parameter package of the lightweight neural network defect recognition model to perform defect area localization and defect type discrimination on the subsequently collected tunnel lining surface image data, output the subsequent defect recognition results, and determine the subsequent recognition conclusion based on the subsequent defect recognition results; When the subsequent identification conclusion is a normal structural identification result, an inspection record is generated that does not require subsequent manual review. When the subsequent identification conclusion is a warning structural defect identification result, a subsequent structural defect warning information is generated; When the subsequent identification conclusion is a corrective structural defect identification result, a subsequent structural repair early warning information is generated, and the subsequent defect identification result, the subsequent structural defect early warning information, and the subsequent structural repair early warning information are associated and saved with the detection location identifier, collection time information, and inspection equipment identifier corresponding to the subsequent preset detection location.

Citation Information

Patent Citations

  • A container

    WO2020240001A1