Shield segment defect processing method and system with negative pressure cooperating with multi-source fusion

Abstract

The invention provides a negative pressure cooperative multi-source fusion shield segment defect treatment method and system, and relates to the technical field of tunnel engineering components, and the method comprises the following steps: hoisting a shield segment into a closed negative pressure treatment cabin for positioning and interference elimination; identifying original defect features according to the quality detection model; inputting the defect original features into the D-S evidence theory optimized by the multi-source fusion network to quantify and grade a defect grading result; when the defect grading result is the first level, defect integrated processing is conducted on the shield segment through a reinforcement learning strategy network, and a first-level repairing result is obtained; when the defect grading result is a second level, planning a repairing scheme for the shield segment in real time through a greedy algorithm to obtain a second-level repairing result; and performing quality scoring on the repair result, and when the quality score does not reach the standard, generating an adjustment parameter through a reinforcement learning strategy network until the quality score reaches the standard, thereby forming a defect processing scheme. The problems of out-of-control quality of the duct piece and potential safety hazards of long-term service of the tunnel are solved.

Description

A method and system for handling defects in tunnel lining segments using negative pressure collaborative multi-source fusion Technical Field

[0001] This invention relates to the field of tunnel engineering component technology, and more specifically, to a method and system for handling defects in shield tunnel segments using negative pressure collaborative multi-source fusion. Background Technology

[0002] In existing tunnel engineering component technologies, traditional tunnel segments require six discrete processes after demolding: hoisting, transportation, temporary storage, inspection, grinding, and spraying. Each process is completed by independent equipment. This operational mode has several drawbacks: First, the risk of segment collision damage increases significantly during hoisting and transportation; simultaneously, waiting time during hoisting and process switching can easily lead to data link breaks and disconnections. Second, multiple transportation processes can cause dust and laitance to adhere to the segment surface, adversely affecting the effectiveness of subsequent processing steps. Third, the inspection, grinding, and spraying of segment damage rely heavily on manual operation, making it difficult to meet the technical requirements of curved surface treatment and precision protection, leading to uncontrolled segment quality and potential safety hazards in the long-term service of the tunnel.

[0003] Therefore, there is an urgent need for a method and system for handling tunnel segment defects using negative pressure collaborative multi-source fusion, which can solve the problems of uncontrolled segment quality and safety hazards in long-term tunnel service. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for handling defects in tunnel lining segments using negative pressure-coordinated multi-source fusion, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides a method for handling defects in tunnel lining segments using negative pressure collaborative multi-source fusion, including:

[0006] The tunnel segments are hoisted into a sealed negative pressure treatment chamber for positioning and interference elimination, resulting in pre-treated segments.

[0007] Defects are identified in the pre-processed segments according to a preset quality inspection model to obtain the original characteristics of the defects.

[0008] The original features of the defects are input into the DS evidence theory optimized by the multi-source fusion network for quantitative classification, and the defect classification results are obtained.

[0009] When the defect classification result is Level 1, the shield tunnel segment is processed by a reinforcement learning strategy network to obtain a Level 1 repair result.

[0010] When the defect classification result is level two, a repair scheme for the shield tunnel segment is planned in real time and the pressure parameters are dynamically adjusted using a greedy algorithm to obtain a level two repair result;

[0011] The quality scores of the first-level repair results and the second-level repair results are evaluated. If the quality score does not meet the standard, adjustment parameters are generated through a reinforcement learning policy network until the quality score meets the standard, thus forming a defect handling plan.

[0012] Secondly, this application also provides a negative pressure collaborative multi-source fusion shield tunnel segment defect treatment system, including:

[0013] The pre-processing module is used to hoist the tunnel segments into a sealed negative pressure treatment chamber for positioning and interference elimination, resulting in pre-processed segments.

[0014] The identification module is used to identify defects in the pre-processed pipe segments according to a preset quality inspection model, and obtain the original features of the defects.

[0015] The grading module is used to input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network for quantitative grading, so as to obtain the defect grading result.

[0016] The Level 1 Repair Module is used to perform integrated defect processing on the shield tunnel segments through a reinforcement learning policy network when the defect classification result is Level 1, thereby obtaining a Level 1 repair result.

[0017] The secondary repair module is used to plan a repair scheme for the shield tunnel segment in real time and dynamically adjust the pressure parameters when the defect classification result is secondary.

[0018] The scoring module is used to score the quality of the first-level repair results and the second-level repair results. When the quality score does not meet the standard, the reinforcement learning policy network generates adjustment parameters until the quality score meets the standard, thus forming a defect handling plan.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention utilizes a sealed negative pressure treatment chamber for positioning and interference elimination, effectively isolating external vibrations, dust, and other interference factors to ensure that pre-treated tunnel segments are in a stable and standard inspection state. It combines a quality inspection model to extract original defect features, and then uses DS evidence theory optimized by a multi-source fusion network for quantitative grading, overcoming the limitations of single-algorithm grading and achieving multi-dimensional fusion and objective quantification of defect features, thus improving the accuracy of defect grading results. For Level 1 defects, a reinforcement learning strategy network is used for integrated processing, leveraging the autonomous learning and optimization capabilities of reinforcement learning. For Level 2 defects, a greedy algorithm is used to plan repair schemes in real time and dynamically adjust pressure parameters, avoiding resource waste from one-size-fits-all repairs while ensuring repair effectiveness. By scoring the repair results for quality, and using the reinforcement learning strategy network to generate adjustment parameters until the quality meets the standard when the score is below standard, a closed-loop control system of "inspection-grading-repair-acceptance-optimization" is formed, effectively avoiding rework waste caused by substandard repairs. In summary, this invention solves the problems of uncontrolled tunnel segment quality and long-term tunnel safety hazards.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 is a schematic diagram of the shield tunnel segment defect handling method of pressure-coordinated multi-source fusion described in the embodiment of the present invention;

[0024] Figure 2 is a schematic diagram of the negative pressure treatment system in the sealed negative pressure treatment chamber described in an embodiment of the present invention;

[0025] Figure 3 is a schematic diagram of the support and adjustment system inside the sealed negative pressure treatment chamber described in an embodiment of the present invention;

[0026] Figure 4 is a schematic diagram of the grinding monitoring system in the sealed negative pressure treatment chamber described in the embodiment of the present invention;

[0027] Figure 5 is a schematic diagram of the shield tunnel segment defect treatment equipment with pressure-coordinated multi-source fusion described in the embodiment of the present invention.

[0028] The diagram is labeled as follows: 800, pressure-coordinated multi-source fusion shield tunnel segment defect processing equipment; 801, processor; 802, memory; 803, multimedia component; 804, I / O interface; 805, communication component. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0030] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] Example 1:

[0032] This embodiment provides a method for handling defects in tunnel lining segments using pressure-coordinated multi-source fusion.

[0033] Referring to Figures 1 to 4, the figures illustrate steps S1 to S6 of this method, including:

[0034] S1: The shield tunnel segments are hoisted into a sealed negative pressure treatment chamber for positioning and interference elimination, resulting in pre-treated segments;

[0035] In this step, the tunnel segment is hoisted into a sealed negative pressure treatment chamber. An expandable insertable support device and a horizontally adjustable hydraulic positioning system are used to inflate and lock the handholes, and the segments are then three-dimensionally leveled. Air is blown onto the surface of the segments to remove dust. Simultaneously, the negative pressure dust removal and hot air drying units within the sealed negative pressure treatment chamber are activated to ensure that the dust concentration inside the chamber is ≤5mg / m³. 3 With a surface moisture content of ≤3%, dust, water vapor, and vibration interference are eliminated, resulting in pre-treated tube segments with a positioning accuracy of ≤0.2mm. This effectively isolates external vibration, dust, and other interference factors, ensuring that the pre-treated tube segments are in a stable and standard testing state.

[0036] The sealed negative pressure treatment chamber includes a segment positioning and support adjustment system for fixing segments; a segment quality intelligent assessment unit based on multi-source sensor fusion; an automatic repair unit with a retractable hollow drilling and injection integrated design; a grinding monitoring system, a negative pressure treatment system, and a dust removal unit that integrates quality feedback learning; a surface cleaning and intelligent air drying unit; and an automatic coating, curing, and quality monitoring unit based on a dual learning mechanism.

[0037] S2: Based on a preset quality inspection model, defects are identified in the pre-processed pipe segments to obtain the original characteristics of the defects;

[0038] To clarify the specific method for obtaining the original features of the defect, step S2 includes S21 to S26, specifically:

[0039] S21: Based on the binocular structured light three-dimensional imaging model in the quality inspection model, a sinusoidal stripe pattern is projected onto the preprocessed tube segment, and a multi-frame structured light original image sequence is acquired.

[0040] In this step, the binocular structured light 3D imaging model sequentially projects a series of sinusoidal stripe patterns of different frequencies onto the surface of the tube segment, while the left and right industrial cameras in the quality inspection model synchronously acquire a multi-frame sequence of original structured light images that are deformed by the surface morphology and reflection characteristics of the tube segment.

[0041] The binocular structured light 3D imaging model is a high-precision digital light processing projector.

[0042] S22: Train the U-shaped depth segmentation model using the layered sliding window transform encoder to obtain the layered depth segmentation model;

[0043] The hierarchical sliding window transform encoder is a Swing Transform encoder, and the hierarchical deep segmentation model is a U-Net deep learning model with a Swing Transform encoder.

[0044] To clarify the specific method for obtaining the hierarchical depth segmentation model, step S22 includes S221 to S224, specifically:

[0045] S221: Select shield tunnel segment specimens based on the hierarchical sliding window transform encoder, and obtain the dataset by using a preset multi-frequency heterodyne sinusoidal fringe structured light imaging sequence as training samples and selecting standard tunnel segments as negative samples.

[0046] In this step, multiple sets of shield tunnel segment specimens with surface defects such as pitting, cracks, bubbles, and pores are prefabricated on the test bench as positive samples. Multi-frequency heterodyne sinusoidal fringe structured light imaging sequences are collected as training samples, while standard tunnel segments are collected as negative samples. All shield tunnel segment specimens are labeled at the pixel level according to defect type and defect level. Finally, a special dataset of tunnel segment structured light images, defect categories, and defect levels is constructed, and the dataset is divided into training, validation, and test sets in a ratio of 7:2:1.

[0047] S222: Extract features from the dataset based on a hierarchical sliding window transform encoder, and obtain pre-trained weights by iteratively training and optimizing the extracted features;

[0048] To clarify the specific method for obtaining the pre-training weights, step S222 includes S2221 to S2225, which are as follows:

[0049] S2221: Based on the hierarchical sliding window transform encoder, multi-channel input tensors are concatenated on the multi-frame image sequence in the dataset. Attention is calculated on the multi-channel input tensors through a multi-head self-attention mechanism to obtain window attention features.

[0050] In this step, a hierarchical sliding window transform encoder is used to concatenate multiple consecutively acquired image sequences from the dataset into an image patch sequence tensor in the channel dimension. ,in, Let R be the image patch sequence tensor, N be the number of patches in the window, and C be the feature dimension. Then, it is divided into non-overlapping local patches in a fixed window of 7×7 pixels. Each patch is flattened into a 49-dimensional vector and linearly mapped to obtain a C-dimensional embedding. Within each Swing Trans former, multi-head self-attention is calculated on a window-by-window basis to obtain window attention features.

[0051] The expression for the component tensor is:

[0052] (1);

[0053] In the above formula (1), For the first The output feature matrix of each attention head To scale the dot product attention, For image patch sequence tensors, For the first Each attention head queries the projection weight matrix. For the first The key weight matrix of each attention head, For the first The values ​​of each attention head are projected onto the weight matrix. Scaling factor

[0054] The expression for the multi-head self-attention is:

[0055] (2);

[0056] In the above formula (2), For the sake of the bulls' self-attention, For the first The output feature matrix of each attention head It is a learnable linear projection matrix.

[0057] S2222: The window attention feature and the preset shifted window attention are processed alternately to obtain multi-level features;

[0058] In this step, an alternating strategy of multi-head self-attention of windows and multi-head self-attention of shifted windows is adopted to enhance cross-window information interaction capabilities while maintaining local modeling capabilities. Let... For the first After the feature representation in the layer is divided by window, the calculation process of a complete Swing Trans former Block can be represented by the following five steps:

[0059] The expression for the window attention is:

[0060] (3);

[0061] In the above formula (3), For the first An intermediate feature tensor after window partitioning For window-based multi-head self-attention, LN is used for layer normalization. For the first The output feature tensor of the layer;

[0062] The expression for the feedforward network processing (first layer) is:

[0063] (4);

[0064] In the above formula (4), Let be the final output feature tensor of the l-th layer Swing Trans former Block, MLP be the multilayer perceptron mechanism, and LN be the layer normalization mechanism. For the first An intermediate feature tensor after windowing;

[0065] The expression for the shift window attention is:

[0066] (5);

[0067] In the above formula (5), For the first The output feature tensor of the layer For multi-head self-attention based on shift windows, LN is used for layer normalization. For the first An intermediate feature tensor after windowing;

[0068] The expression for the feedforward network processing (second layer) is:

[0069] (6);

[0070] In the above formula (6), For the first The output feature tensor of the layer; MLP is a multilayer perceptron mechanism; LN is layer normalization. For the first An intermediate feature tensor after windowing.

[0071] S2223: Input the multi-level features into the U-shaped deep segmentation model, restore the spatial resolution through upsampling and jump-connect with the shallow features of the encoder to output the defect segmentation map;

[0072] In this step, multi-level features are input into a symmetric four-stage U-shaped segmentation network: each level of the decoder first upsamples the low-resolution features by a factor of 2 using a 2× transposed convolution, concatenates them with the features of the encoder at the same level in the channel dimension, and then fuses them through a 3×3 depthwise separable convolution, batch normalization, and ReLU activation function to restore them to 1 / 4 of the original image size step by step; the top layer uses a 1×1 convolution to map the number of channels to the number of defect categories K and interpolates back to the original resolution using bilinear interpolation to output a defect segmentation map classified pixel by pixel; skip connections continuously supplement high-resolution edge details and suppress over-scratching and missed detection.

[0073] S2224: Normalize the defect category scores of the defect segmentation map to obtain the defect probability distribution;

[0074] In this step, the expression normalized by the Soft max function is:

[0075] (7)

[0076] In the above formula (7), The position belongs to the first Probability of class defects As a category, For pixel position, This represents the score output by the model at image location (x, y) for defect type i. For all Summing is performed on each defect category.

[0077] S2225: Construct a multi-channel image dataset based on the defect probability distribution and preset defect labels, and iteratively train and optimize the parameters of the U-shaped depth segmentation model using the stochastic gradient descent algorithm to obtain pre-trained weights.

[0078] In this step, based on the defect probability distribution, the original image sequence of the pipe segment surface in the quality inspection model and the defect labels (including categories such as pitting, cracks, bubbles, and laitance) are subjected to data enhancement such as rotation, brightness, and dust, and pixel-level alignment is completed to construct a multi-channel image dataset. Then, the multi-channel image dataset is input into the U-shaped depth segmentation model for forward inference to obtain the defect probability distribution, and the frequency and severity of each category are statistically analyzed. Dynamically weighted cross-entropy loss is used to alleviate category imbalance. Finally, with the dynamically weighted cross-entropy loss as the target, stochastic gradient descent iteration is performed under the control of cosine annealing learning rate and early stopping strategy. After convergence, the pre-trained weights are output.

[0079] S223: Perform normal initialization on the U-shaped depth segmentation model according to the pre-trained weights and the light image sequence of the standard tube segment to obtain the initial depth segmentation model;

[0080] In this step, the pre-trained weights are loaded into the U-shaped deep segmentation model starting from the basic backbone network of the hierarchical deep segmentation model (Swin-Transformer), which has been pre-trained on a large-scale dataset with 22,000 categories and over 14 million images, while retaining window attention and hierarchical feature representation. Based on the pixel distribution statistical characteristics of the four types of defects on the pipe segment surface—roughness, cracks, bubbles, and laitance—the upsampling branches of the classification head and decoder are reinitialized with a Gaussian distribution of zero mean and 0.02 standard deviation. Channel-spatial parallel convolutional block attention bridging layer (CBAM) is inserted at the encoder-decoder skip connection. Then, the first forward propagation is performed using the optical image sequence of the standard pipe segment, which outputs an initial defect probability map matching the actual curvature of the pipe segment, avoiding overly smooth edges during cold start and obtaining the initial deep segmentation model.

[0081] S224: Perform multi-stage iterative optimization and verification on the initial depth segmentation model to obtain a hierarchical depth segmentation model.

[0082] In this step, the initial depth segmentation model is iteratively optimized and validated in multiple stages, using a three-stage training strategy of "coarse-fine-refined": ① Coarse stage in the basic feature learning stage (1-20 epochs): the model learns the texture features and defect contour features of the structured light image of the tube segment; ② Fine stage in the defect category differentiation stage (21-50 epochs): the model's ability to distinguish features between rough surfaces and laitance layers, and between microbubbles and small cracks is enhanced; ③ Fine stage in the grade matching optimization stage (51-80 epochs): the model parameters are adjusted with the apparent defect grade standard as the optimization target, so that the output defect probability distribution is accurately matched with the defect grade judgment rule.

[0083] S23: Based on the hierarchical depth segmentation model, perform appearance defect identification on the multi-frame structured light original image sequence to obtain an appearance defect segmentation map;

[0084] In this step, the multi-frame structured light original image sequence is stitched together in the channel dimension and input into the hierarchical depth segmentation model. The 1 / 4 resolution branch is used to quickly locate the suspected defect ROI, and then the 1 / 8 and 1 / 16 branches are cascaded to enhance the details inside the ROI. After Softmax, the probability of each pixel belonging to the four types of defects is output. Noise is filtered according to the dual criteria of maximum probability and area threshold of 50 pixels. Finally, an apparent defect segmentation map with defect category, confidence level and pixel area is generated to realize the world coordinate positioning of defects.

[0085] S24: Based on the acoustic wave detection model in the quality inspection model, the pre-treated pipe segment is subjected to acoustic wave damage sensitive treatment to obtain the internal porosity index;

[0086] In this step, a robotic arm based on the acoustic wave detection model in the quality inspection model couples an excitation hammer and a microphone array to the inner arc surface of the pre-processed segment. The excitation hammer generates a 1–10 kHz swept-frequency elastic wave, and the microphone array synchronously acquires the transmitted acoustic wave signal. The measured wave velocity, peak frequency, and amplitude are extracted for each measurement point and compared with a defect-free benchmark. The wave velocity change rate, main frequency offset, and amplitude attenuation coefficient are calculated sequentially. Then, the three parameters are input into a pre-trained porosity regression network, which outputs the internal porosity index of the measurement point. By traversing all measurement points, the internal defect identification of the segment is completed.

[0087] The automatic detection instrument consists of an excitation hammer that generates elastic waves of a specific frequency, and a high-sensitivity microphone array that receives the acoustic signals propagating in the concrete of the tunnel segment.

[0088] Furthermore, the wave velocity change rate is used to reflect the degree of change in the elastic wave velocity relative to a reference value, and the formula for calculating the wave velocity change rate is:

[0089] (8);

[0090] In the above formula (8), The rate of change of wave velocity, The reference wave velocity under defect-free conditions. This is the measured wave velocity;

[0091] The main frequency offset is used to characterize the degree of deviation of the signal's main frequency relative to the reference main frequency. The formula for calculating the main frequency offset is as follows:

[0092] (9);

[0093] In the above formula (9), This is the main frequency offset. The reference frequency under defect-free conditions. This represents the peak frequency (dominant frequency) of the measured signal.

[0094] The amplitude attenuation coefficient is used to describe the degree of amplitude attenuation during the propagation of elastic waves (unit: The formula for calculating the amplitude attenuation coefficient is as follows:

[0095] (10);

[0096] In the above formula (10), The amplitude attenuation coefficient is... The reference amplitude under defect-free conditions. This represents the measured amplitude.

[0097] S25: The pre-treated tube segment is inspected according to the piezoelectric ceramic excitation model in the quality inspection model to obtain the damage quantification index;

[0098] In this step, the piezoelectric ceramic excitation model is an integrated piezoelectric ceramic excitation-receiving system. A robotic arm of the integrated piezoelectric ceramic excitation-receiving system is used to automatically detect the pre-processed tube segment. The energy moment is extracted by continuous wavelet transform of the echo signal through sinusoidal frequency sweep-pulse echo dual-mode excitation, and the damage index, wave impedance change rate, and energy attenuation coefficient are calculated. The damage index, wave impedance change rate, and energy attenuation coefficient are fused into 0-1 using an extreme gradient boosting regressor (XG Boost) pre-trained on a standard test block to obtain the damage quantification index.

[0099] The wave impedance change rate is used to reflect the degree of change of wave impedance relative to a reference value, and the formula for calculating the wave impedance change rate is as follows:

[0100] (11);

[0101] In the above formula (11), The rate of change of wave impedance, The reference wave impedance under defect-free conditions. This is the measured wave impedance.

[0102] The energy attenuation coefficient is used to describe the proportion of energy loss during elastic wave propagation, and the formula for calculating the energy attenuation coefficient is as follows:

[0103] (12);

[0104] In the above formula (12), The energy decay coefficient, Energy at the incident end, This refers to the energy received at the receiving end.

[0105] S26: The original features of the defect are obtained by integrating the damage quantification index, the internal porosity index and the apparent defect segmentation map.

[0106] S3: Input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network for quantitative classification to obtain the defect classification result;

[0107] By extracting the original features of defects through the quality inspection model described above, and using the DS evidence theory optimized by the multi-source fusion network in this step for quantitative classification, the limitations of single algorithm classification are overcome, and multi-dimensional fusion and objective quantification of defect features are realized, thereby improving the accuracy of defect classification results.

[0108] To clarify the specific method for obtaining the defect classification results, step S3 includes S31 to S34, specifically:

[0109] S31: Input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network to calculate the conflict factor, and calculate the joint probability by using the conflict factor and the preset defect category set to obtain the joint probability value;

[0110] In this step, the original features of the defects are used as independent sources of evidence to construct a defect category set, which includes first-level defects, second-level defects, and third-level defects. Basic probability assignments are performed on these independent sources of evidence based on historical operating conditions, statistical priors, and measured reliability, resulting in initial reliability distributions for structured light, acoustic, and piezoelectric sensors. Subsequently, the weight coefficients learned offline by the multi-source fusion network (Meta-DSF-Net) are used to weight and correct the basic probability assignments, suppressing misjudgment weights caused by high-conflict sensors. Finally, the conflict factor is calculated according to the DS combination rule in DS evidence theory. ,like If the output is "cannot be determined" and a manual review is requested, then... Then continue to synthesize the joint probability value.

[0111] (13);

[0112] In the above formula (13), As a conflict factor, For a set of defect categories, This represents the basic probability assignment of the defect category set to the structured light sensor. This represents the basic probability assignment of the acoustic sensor to the set of defect categories. This represents the basic probability assignment of the defect category set for piezoelectric sensors.

[0113] S32: Based on the dynamic weight vector of the multi-source fusion network, the original features of the defect are enhanced by cross-attention and fused with contextual features to output a dynamic confidence threshold;

[0114] In this step, the current segment model, environmental dust concentration, material strength dispersion, and other working condition contexts are extracted in real time based on the dynamic weight vector of the multi-source fusion network to generate a dynamic weight vector. Cross-attention enhancement is performed on the structured light point cloud features, acoustic features, and piezoelectric features in the dynamic weight vector to obtain a context feature vector. Finally, a new network branch mapping for generating the threshold is used and a dynamic confidence threshold is output after sigmoid limiting. This enables the threshold to be adaptively adjusted according to the working conditions, ensuring reliable classification even when there is high dust, weak reflection, or high internal material dispersion.

[0115] The expression for the context feature vector is:

[0116] (14);

[0117] In the above formula (14), The context feature vector is represented by Encoder(·), which is the encoder. Features of structured light Characteristics of sound waves It exhibits piezoelectric characteristics;

[0118] The expression for the dynamic confidence threshold is:

[0119] (15);

[0120] In the above formula (15), The reliability threshold is dynamically generated. It is the Sigmoid activation function. This is a new network branch specifically designed for generating thresholds. For context feature vectors, The maximum threshold boundary, This represents the minimum threshold boundary.

[0121] S33: Based on the dynamic confidence threshold, the joint probability value is filtered by maximum probability to obtain the defect category;

[0122] In this step, the joint probability value is filtered by maximum probability based on the dynamic confidence threshold. When the maximum probability is not less than the dynamic confidence threshold, the maximum probability is the final defect category. When the maximum probability is less than the dynamic confidence threshold, a suspected defect category is output to trigger manual review.

[0123] S34: Based on the DS evidence theory, the basic probability allocation of the defect categories is fused and quantified for classification, and the defect classification results are output.

[0124] In this step, the defect categories are mapped back to the space of the original basic probability assignment and the fusion confidence and residual uncertainty of the defect categories are calculated. The quantization classification is performed using the strategy of maximizing the fusion confidence and residual uncertainty, and the defect classification results are output.

[0125] S4: When the defect classification result is Level 1, the shield tunnel segment is processed by a reinforcement learning strategy network to obtain a Level 1 repair result.

[0126] In this step, when the defect classification result is Level 1, it is regarded as a minor surface defect (roughness, slurry, color difference with a depth <0.3 mm or microbubbles with a diameter <2 mm), and the reinforcement learning policy network is immediately activated. Using the actual point cloud acquired by structured light as the state, and taking grinding pressure, feed speed, and tool path curvature as the first action, the mean square error between the pass probability of the ground area after rescanning and the measured probability is used as the immediate reward to update the strategy and value network online. After the network outputs the optimal action, the central control system parses it into specific process parameters (grinding wheel grit size P 120, pressure range 0.12–0.18 MPa, speed 80–120 mm / s, overlap rate 30%), driving the six-axis grinding robot arm to complete fully automatic grinding under the condition of negative pressure dust removal and synchronous operation. Immediately after grinding, the same structured light unit is used for re-inspection. If the pass probability is ≥0.95 and the surface roughness is ≤1.2μm, it is judged as a first-level repair qualified. Otherwise, it continues to iterate using new error signals until the indicators are met, and then outputs the first-level repair result and flows to the next process. The autonomous learning and optimization capabilities of the learning are enhanced through reinforcement learning strategy.

[0127] S5: When the defect classification result is level two, a repair scheme for the shield tunnel segment is planned in real time and the pressure parameters are dynamically adjusted using a greedy algorithm to obtain a level two repair result;

[0128] In this step, a greedy algorithm is used to plan the repair scheme in real time and dynamically adjust the pressure parameters, so as to avoid the waste of resources caused by a one-size-fits-all approach to repair while ensuring the repair effect.

[0129] To clarify the specific method for obtaining the secondary repair results, step S5 includes S51 to S54, specifically:

[0130] S51: When the defect classification result is level two, the optimal drilling and injection sequence and path are iteratively searched under the resource comprehensive efficiency weight by a greedy algorithm, and the defect area volume is solved to obtain the repair area volume.

[0131] In this step, when the defect classification result is level two, it is considered a crack with a width of 0.2 mm to 0.5 mm, or a bubble or hole with a diameter of 2 mm to 10 mm and a depth less than the protective layer thickness. Using a 26-neighborhood greedy growth strategy, the growth starts from the seed voxel with the highest confidence and expands outwards. At each step, the optimal drilling sequence is searched by selecting the direction with the largest confidence gain of the maximum defect that can be covered by a unit volume of material, until the confidence of the boundary voxel is less than 0.35. The voxel volume is accumulated simultaneously to obtain the repair area volume, and the repair area volume is used as a constraint for the reverse pruning path to ensure that the repair head moves the shortest distance and has the fewest turns.

[0132] The expression for the volume of the repaired area is:

[0133] (16);

[0134] In the above formula (16), For the volume of the repair area, For the reconstructed defective spatial region, For voxels in Physical resolution of direction, For voxels in Physical resolution of direction, For voxels in Physical resolution of direction.

[0135] S52: The repair pressure value is dynamically adjusted based on the defect depth and the repair pressure coefficient optimized by machine learning to obtain the repair pressure value;

[0136] In this step, the maximum value of the defect depth, the real-time temperature, and the dynamic viscosity of the material are input into a lightweight extreme gradient boosting model (XG Boost) that has been trained online with 27,000 sets. The model outputs the repair pressure coefficient optimized by machine learning and calculates the repair pressure value in real time. The model is updated incrementally every 10 closed-loop feedback data to ensure that the pressure prediction error is ≤3%.

[0137] The expression for the repair pressure value is:

[0138] (17);

[0139] In the above formula (17), To repair the dynamic adjustment of pressure, For defect depth, For the dynamic viscosity of the material, The pressure coefficient was optimized using machine learning to correct it.

[0140] S53: Calculate and optimize the number of repair layers for the tunnel segment based on the repair area volume and the preset repair head coverage area to obtain the repair layer value;

[0141] In this step, based on the volume of the repair area and the coverage area of ​​the repair head called by the system (expanded umbrella-shaped nozzle diameter 30 mm, S≈706 mm),... 2 The optimal thickness of a single layer is determined by the experiment, and the repair layer value is calculated. When the repair layer value is >4, the layering, cross, and staggered joint strategy is automatically triggered. The adjacent layer paths are rotated 90° and a 2-second interval is inserted between the layers to obtain the repair layer value, so as to ensure interlayer venting and temperature rise control.

[0142] The expression for the numerical value of the repair layer is:

[0143] (18);

[0144] In the above formula (18), For the number of repair layers, To repair the coverage area of ​​the head, For single-layer materials, the optimal coating thickness is... This is the floor function.

[0145] S54: Based on the repair area volume, the repair pressure value, and the repair layer value, the pressure parameters are dynamically adjusted using a retractable hollow drilling and injection integrated repair head to obtain a secondary repair result.

[0146] In this step, the retractable hollow drilling and injection integrated repair head dynamically generates closed-loop control commands based on the repair area volume, the repair pressure value, and the repair layer value: First, it drills at a low speed to the bottom of the defect. When the energy attenuation coefficient on the other side of the piezoelectric array is ≤0.05, it determines that it has reached the bottom and then switches to grouting mode. When dealing with surface defects such as surface cracks and pitting, the drill bit remains retracted, and the pressurized self-repairing material is sprayed out through the central grouting channel for uniform surface grouting. When encountering internal defects such as air bubbles and pores, the electric telescopic mechanism drives the hollow drill bit to extend and drill into the location of the internal defect. Then, high-pressure injection is performed through the central grouting channel to press the self-repairing material from the bottom of the defect upwards, ensuring that the filling is full, dense, and free of air bubbles. Finally, the secondary repair result is output. The secondary repair result includes the actual filling volume, the predicted value of interlayer bonding strength, the piezoelectric monitoring convergence curve, and the repair point cloud with a surface roughness ≤12 µm that can be directly used in the next process, realizing 100% automatic closed-loop repair of secondary defects.

[0147] The expression for the real-time correction amount of the dynamic correction is:

[0148] (19);

[0149] In the above formula (19), For real-time correction, For real-time traffic, For theoretical flow rate, The value is obtained after machine learning optimization of the flow-pressure feedback coefficient.

[0150] S6: Assess the quality of the first-level repair results and the second-level repair results. If the quality score does not meet the standard, generate adjustment parameters through a reinforcement learning strategy network until the quality score meets the standard, thus forming a defect handling plan.

[0151] In this step, when the score fails to meet the standard, a reinforcement learning strategy network is used to generate and adjust parameters until the quality meets the standard. This forms a closed-loop control system for the entire process of "detection-grading-repair-acceptance-optimization", which effectively avoids the problem of rework waste caused by substandard repair.

[0152] To clarify the specific method for obtaining the defect handling plan, step S6 includes S61 to S64, specifically:

[0153] S61: The first-level repair results and the second-level repair results are compared and analyzed in real time using a logistic regression model, and the preset scanning data and polishing acceptance standards are compared and analyzed to obtain the quality inspection results;

[0154] In this step, the structured light 3D scanner in the quality inspection model is activated to perform secondary point sampling of the repair area based on the first-level repair results and the second-level repair results. After obtaining the point cloud, it is imported into the L2 regularized online logistic regression model. The point cloud is compared point by point with the preset grinding acceptance standards (depth ≤ 0.3 mm, width ≤ 0.5 mm, continuous bubble diameter ≤ 2 mm, no exposed rebar) to obtain the quality inspection results. The coordinates of the unqualified areas are marked. The quality inspection results include a defect probability map.

[0155] S62: Based on the quality inspection results, the non-conforming areas are processed in an integrated manner. The integrated processing area is inspected through an online learning mechanism for quality feedback parameters to obtain a quality score.

[0156] In this step, after marking the unqualified areas based on the quality inspection results, the integrated repair process in the sealed negative pressure chamber is initiated: First, the grinding unit, which integrates the online learning module in the sealed negative pressure treatment chamber, performs rapid rough repair at the target quality threshold of 0.95 and simultaneously removes dust; then, the retractable hollow drilling and repair head in the sealed negative pressure treatment chamber quantitatively injects into the defective micro-area; finally, the coating head in the sealed negative pressure treatment chamber completes local touch-up coating and curing, maintaining negative pressure and dust control throughout the process; after the repair is completed, the surface data is scanned again to obtain the real-time flow score using a regression model, and the deviation from the theoretical value is used as an online learning signal. The model weights are dynamically updated through stochastic gradient descent, and the final quality score is output, realizing a closed-loop control of "detection-repair-re-detection-model self-optimization".

[0157] S63: The quality score is judged according to the preset quality compliance threshold. When the quality score fails to meet the standard, the non-conforming area feature analysis and parameter optimization are performed through the reinforcement learning strategy network to generate polishing adjustment parameters.

[0158] In this step, the quality score is judged according to the preset quality compliance threshold. When the quality score is < 0.95, the reinforcement learning strategy network is triggered. The reinforcement learning strategy network takes the defect feature vector extracted by the PointNet++ model of the unqualified area as the state, and takes the grinding pressure increment, speed ratio, recoating thickness and nozzle dwell time as the action space. It uses a greedy strategy to search in the action space, with the goal of maximizing the comprehensive reward of defect area reduction rate, surface roughness reduction rate and material consumption. After 200 iterations, the optimal grinding adjustment parameters are output.

[0159] S64: Based on the grinding adjustment parameters, perform integrated processing on the unqualified areas and determine whether the quality score meets the standard. When the quality score meets the standard, a defect handling plan is formed.

[0160] In this step, the non-conforming areas are rescheduled for a second integrated processing based on the grinding adjustment parameters. After processing, the S61-S62 scanning-scoring process is executed again until the quality score is ≥0.95 and the score fluctuation is <0.01 for two consecutive rounds. The quality is then judged to meet the standard. At this time, the central controller packages the parameter set of final grinding pressure, speed, coating thickness, curing energy, point cloud acceptance report and material consumption into the database to form a traceable and reproducible defect handling solution.

[0161] Example 2:

[0162] This embodiment provides a pressure-coordinated multi-source fusion shield tunnel segment defect processing system, the system comprising:

[0163] The pre-processing module is used to hoist the tunnel segments into a sealed negative pressure treatment chamber for positioning and interference elimination, resulting in pre-processed segments.

[0164] The identification module is used to identify defects in the pre-processed pipe segments according to a preset quality inspection model, and obtain the original features of the defects.

[0165] To clarify the specific methods for obtaining the identification module, the following are included:

[0166] The projection unit is used to project sinusoidal stripe patterns onto the preprocessed tube segment based on the binocular structured light three-dimensional imaging model in the quality inspection model, and to acquire a multi-frame structured light original image sequence.

[0167] The training unit is used to train the U-shaped depth segmentation model based on the hierarchical sliding window transform encoder to obtain the hierarchical depth segmentation model.

[0168] To clarify the specific methods for obtaining training units, the following are included:

[0169] A sub-unit is selected for selecting shield tunnel segment specimens based on the hierarchical sliding window transform encoder. A dataset is obtained by using a preset multi-frequency heterodyne sinusoidal fringe structured light imaging sequence as training samples and selecting standard tunnel segments as negative samples.

[0170] Extraction sub-units are used to extract features from the dataset based on the hierarchical sliding window transform encoder, and pre-trained weights are obtained by iteratively training and optimizing the extracted features.

[0171] The initial subunit is used to perform normal initialization of the U-shaped depth segmentation model based on the pre-trained weights and the light image sequence of the standard tube segment to obtain the initial depth segmentation model.

[0172] The iterative subunit is used to perform multi-stage iterative optimization and verification on the initial depth segmentation model to obtain a hierarchical depth segmentation model.

[0173] The appearance recognition unit is used to perform appearance defect recognition on the multi-frame structured light original image sequence based on the hierarchical depth segmentation model to obtain an appearance defect segmentation map.

[0174] The sensitive processing unit is used to perform acoustic damage sensitive processing on the pre-treated pipe segment based on the acoustic detection model in the quality inspection model to obtain the internal porosity index.

[0175] The detection unit is used to detect the pretreated tube segment according to the piezoelectric ceramic excitation model in the quality detection model to obtain damage quantification index;

[0176] The integration unit is used to integrate the damage quantification index, the internal porosity index and the apparent defect segmentation map to obtain the original defect features.

[0177] The grading module is used to input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network for quantitative grading, so as to obtain the defect grading result.

[0178] To clarify the specific methods for obtaining the hierarchical module, the following are included:

[0179] The calculation unit is used to input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network to calculate the conflict factor, and calculate the joint probability by using the conflict factor and the preset defect category set to obtain the joint probability value.

[0180] The enhancement unit is used to perform cross-attention enhancement and context feature fusion on the original features of the defect based on the dynamic weight vector of the multi-source fusion network, and output a dynamic confidence threshold.

[0181] A filtering unit is used to perform maximum probability filtering on the joint probability value based on the dynamic confidence threshold to obtain the defect category;

[0182] The grading unit is used to fuse and quantify the basic probability allocation of the defect categories according to the DS evidence theory, and output the defect grading result.

[0183] The Level 1 Repair Module is used to perform integrated defect processing on the shield tunnel segments through a reinforcement learning policy network when the defect classification result is Level 1, thereby obtaining a Level 1 repair result.

[0184] The secondary repair module is used to plan a repair scheme for the shield tunnel segment in real time and dynamically adjust the pressure parameters when the defect classification result is secondary.

[0185] To clarify the specific methods for obtaining the Level 2 repair module, the following are provided:

[0186] The solution unit is used to iteratively search for the optimal drilling and injection sequence and path and solve for the volume of the defect area when the defect classification result is level two, using a greedy algorithm under the weight of comprehensive resource efficiency, to obtain the volume of the repair area.

[0187] An adjustment unit is used to dynamically adjust the pressure value of the shield tunnel segment based on the defect depth and a repair pressure coefficient optimized by machine learning, so as to obtain the repair pressure value.

[0188] The layer calculation unit is used to calculate and optimize the number of layers of the shield tunnel segment based on the volume of the repair area and the preset coverage area of ​​the repair head, so as to obtain the repair layer value;

[0189] The correction unit is used to dynamically correct the pressure parameters based on the repair area volume, the repair pressure value, and the repair layer value, using a retractable hollow drilling and injection integrated repair head, to obtain a secondary repair result.

[0190] The scoring module is used to score the quality of the first-level repair results and the second-level repair results. When the quality score does not meet the standard, the reinforcement learning policy network generates adjustment parameters until the quality score meets the standard, thus forming a defect handling plan.

[0191] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0192] Example 3:

[0193] Corresponding to the above method embodiments, this embodiment also provides a shield tunnel segment defect treatment device based on pressure-coordinated multi-source fusion. The shield tunnel segment defect treatment device based on pressure-coordinated multi-source fusion described below and the shield tunnel segment defect treatment method based on pressure-coordinated multi-source fusion described above can be referred to in correspondence.

[0194] Figure 5 is a block diagram illustrating a pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800 according to an exemplary embodiment. As shown in Figure 5, the pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800 may include a processor 801 and a memory 802. The pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800 may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0195] The processor 801 controls the overall operation of the pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800 to complete all or part of the steps in the aforementioned pressure-coordinated multi-source fusion shield tunnel segment defect processing method. The memory 802 stores various types of data to support the operation of the pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800. This data may include, for example, instructions for any application or method operating on the pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the pressure-coordinated multi-source fusion shield tunnel segment defect processing equipment 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0196] In an exemplary embodiment, the pressure-coordinated multi-source fusion shield tunnel segment defect processing device 800 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the aforementioned pressure-coordinated multi-source fusion shield tunnel segment defect processing method.

[0197] Example 4:

[0198] Corresponding to the above method embodiments, this embodiment also provides a medium. The medium described below can be referred to in relation to the pressure-coordinated multi-source fusion shield tunnel segment defect treatment method described above.

[0199] A medium storing a computer program, which, when executed by a processor, implements the steps of the shield tunnel segment defect handling method of pressure-coordinated multi-source fusion described in the above method embodiments.

[0200] The medium can specifically be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0201] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0202] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for handling defects in tunnel lining segments using negative pressure collaborative multi-source fusion, characterized in that, include: The tunnel segments are hoisted into a sealed negative pressure treatment chamber for positioning and interference elimination, resulting in pre-treated segments. Defects are identified in the pre-processed segments according to a preset quality inspection model to obtain the original characteristics of the defects. The original features of the defects are input into the DS evidence theory optimized by the multi-source fusion network for quantitative classification to obtain the defect classification result. When the defect classification result is Level 1, the shield tunnel segment is processed by the reinforcement learning policy network to obtain the Level 1 repair result. When the defect classification result is Level 2, the shield tunnel segment is planned in real time and the pressure parameters are dynamically adjusted by the greedy algorithm to obtain the Level 2 repair result. The quality scores of the first-level repair results and the second-level repair results are evaluated. If the quality score does not meet the standard, adjustment parameters are generated through a reinforcement learning policy network until the quality score meets the standard, thus forming a defect handling plan.

2. The shield tunnel segment defect treatment method based on negative pressure collaborative multi-source fusion according to claim 1, characterized in that, The pre-processed pipe segment is subjected to defect identification based on a preset quality inspection model to obtain the original defect features. This includes: projecting a sinusoidal fringe pattern onto the pre-processed pipe segment using a binocular structured light 3D imaging model in the quality inspection model, and acquiring a multi-frame structured light original image sequence; training a U-shaped depth segmentation model using a layered sliding window transform encoder to obtain a layered depth segmentation model; identifying apparent defects using the layered depth segmentation model on the multi-frame structured light original image sequence to obtain an apparent defect segmentation map; performing acoustic damage sensitive processing on the pre-processed pipe segment using an acoustic detection model in the quality inspection model to obtain an internal porosity index; detecting the pre-processed pipe segment using a piezoelectric ceramic excitation model in the quality inspection model to obtain a damage quantification index; and integrating the damage quantification index, the internal porosity index, and the apparent defect segmentation map to obtain the original defect features.

3. The shield tunnel segment defect treatment method based on negative pressure collaborative multi-source fusion according to claim 2, characterized in that, The U-shaped depth segmentation model is trained using a layered sliding window transform encoder to obtain a layered depth segmentation model. This process includes: selecting tunnel segment specimens using the layered sliding window transform encoder; using a pre-set multi-frequency heterodyne sinusoidal fringe structured light imaging sequence as training samples and selecting standard tunnel segments as negative samples to obtain a dataset; extracting features from the dataset using the layered sliding window transform encoder; iteratively training and optimizing the extracted features to obtain pre-trained weights; performing normal initialization on the U-shaped depth segmentation model using the pre-trained weights and the light image sequence of the standard tunnel segments to obtain an initial depth segmentation model; and performing multi-stage iterative optimization and verification on the initial depth segmentation model to obtain the layered depth segmentation model.

4. The shield tunnel segment defect treatment method based on negative pressure collaborative multi-source fusion according to claim 1, characterized in that, The original features of the defects are input into the DS evidence theory optimized by the multi-source fusion network for quantitative classification to obtain the defect classification result. This includes inputting the original features of the defects into the DS evidence theory optimized by the multi-source fusion network to calculate the conflict factor, and calculating the joint probability by using the conflict factor and a preset defect category set to obtain the joint probability value. The original features of the defect are enhanced by cross-attention and fused with contextual features based on the dynamic weight vector of the multi-source fusion network, and a dynamic confidence threshold is output. Based on the dynamic confidence threshold, the joint probability value is filtered by maximum probability to obtain the defect category; Based on the DS evidence theory, the basic probability assignments of the defect categories are fused and quantified for classification, and the defect classification results are output.

5. The method for handling shield tunnel segment defects using negative pressure collaborative multi-source fusion according to claim 1, characterized in that, When the defect classification result is Level II, a greedy algorithm is used to plan a repair scheme for the tunnel segment in real time and dynamically adjust the pressure parameters to obtain a Level II repair result. This includes: when the defect classification result is Level II, using a greedy algorithm to iteratively search for the optimal drilling and injection sequence and path under the resource comprehensive efficiency weight and solving for the defect area volume to obtain the repair area volume; dynamically adjusting the pressure value of the tunnel segment based on the defect depth and a repair pressure coefficient optimized by machine learning to obtain the repair pressure value; calculating and optimizing the number of repair layers for the tunnel segment based on the repair area volume and the preset repair head coverage area to obtain the repair layer value; and dynamically correcting the pressure parameters using a scalable hollow drilling and injection integrated repair head according to the repair area volume, the repair pressure value, and the repair layer value to obtain the Level II repair result.

6. A shield tunnel segment defect treatment system based on negative pressure collaborative multi-source fusion, characterized in that, include: The pre-processing module is used to hoist the tunnel segments into a sealed negative pressure treatment chamber for positioning and interference elimination, resulting in pre-processed tunnel segments; The identification module is used to identify defects in the pre-processed pipe segments according to a preset quality inspection model, and obtain the original features of the defects. The grading module is used to input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network for quantitative grading, and to obtain the defect grading result. The Level 1 Repair Module is used to perform integrated defect processing on the shield tunnel segments through a reinforcement learning policy network when the defect classification result is Level 1, thereby obtaining a Level 1 repair result. The secondary repair module is used to plan a repair scheme for the shield tunnel segment in real time and dynamically adjust the pressure parameters when the defect classification result is secondary. The scoring module is used to score the quality of the first-level repair results and the second-level repair results. When the quality score does not meet the standard, the reinforcement learning policy network generates adjustment parameters until the quality score meets the standard, thus forming a defect handling plan.

7. The shield tunnel segment defect treatment system with negative pressure collaborative multi-source fusion according to claim 6, characterized in that, The identification module includes: a projection unit for projecting sinusoidal fringe patterns onto the pre-processed tube segment based on a binocular structured light 3D imaging model in the quality inspection model, and acquiring a multi-frame structured light original image sequence; a training unit for training a U-shaped depth segmentation model according to a layered sliding window transform encoder to obtain a layered depth segmentation model; an appearance identification unit for performing appearance defect identification on the multi-frame structured light original image sequence based on the layered depth segmentation model to obtain an appearance defect segmentation map; a sensitive processing unit for performing acoustic damage sensitive processing on the pre-processed tube segment based on an acoustic detection model in the quality inspection model to obtain an internal porosity index; a detection unit for detecting the pre-processed tube segment according to a piezoelectric ceramic excitation model in the quality inspection model to obtain a damage quantification index; and an integration unit for integrating the damage quantification index, the internal porosity index, and the appearance defect segmentation map to obtain the original defect features.

8. The shield tunnel segment defect treatment system based on negative pressure collaborative multi-source fusion according to claim 7, characterized in that, The training unit includes: a selection subunit, used to select shield tunnel segment specimens according to the layered sliding window transform encoder, using a preset multi-frequency heterodyne sinusoidal fringe structured light imaging sequence as training samples and selecting standard tunnel segments as negative samples to obtain a dataset; an extraction subunit, used to extract features from the dataset based on the layered sliding window transform encoder, and to obtain pre-trained weights by iteratively training and optimizing the extracted features; an initialization subunit, used to perform normal initialization of the U-shaped depth segmentation model according to the pre-trained weights and the light image sequence of the standard tunnel segments to obtain an initial depth segmentation model; and an iteration subunit, used to perform multi-stage iterative optimization and verification of the initial depth segmentation model to obtain a layered depth segmentation model.

9. The shield tunnel segment defect treatment system based on negative pressure collaborative multi-source fusion according to claim 6, characterized in that, The hierarchical module includes a calculation unit, which is used to input the original features of the defect into the DS evidence theory optimized by the multi-source fusion network to calculate the conflict factor, and calculate the joint probability by using the conflict factor and a preset defect category set to obtain the joint probability value. An enhancement unit is used to perform cross-attention enhancement and context feature fusion on the original features of the defect based on the dynamic weight vector of the multi-source fusion network, and output a dynamic confidence threshold; a filtering unit is used to perform maximum probability filtering on the joint probability value based on the dynamic confidence threshold to obtain the defect category; The grading unit is used to fuse and quantify the basic probability allocation of the defect categories according to the DS evidence theory, and output the defect grading result.

10. The shield tunnel segment defect treatment system based on negative pressure collaborative multi-source fusion according to claim 6, characterized in that, The secondary repair module includes: a solution unit, used to iteratively search for the optimal drilling and injection sequence and path and solve for the defect area volume using a greedy algorithm under the resource comprehensive efficiency weight when the defect classification result is secondary, to obtain the repair area volume; an adjustment unit, used to dynamically adjust the pressure value of the shield tunnel segment based on the defect depth and a repair pressure coefficient optimized by machine learning, to obtain the repair pressure value; a layer calculation unit, used to calculate and optimize the number of layers of the shield tunnel segment based on the repair area volume and the preset repair head coverage area, to obtain the repair layer value; and a correction unit, used to dynamically correct the pressure parameters using a scalable hollow drilling and injection integrated repair head according to the repair area volume, the repair pressure value, and the repair layer value, to obtain the secondary repair result.

Images