Subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization
Through Hilbert curve point cloud serialization and deep learning network optimization, the efficiency and accuracy of leakage detection in subway shield tunnels are solved, and efficient leakage detection is achieved, which is suitable for practical applications of subway tunnels.
Patent Information
- Application Number
- CN202510350350.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-24
- Publication Date
- 2025-08-12
AI Technical Summary
Traditional leakage detection methods are inefficient and have low accuracy in subway shield tunnels, making it difficult to effectively use point cloud data for efficient analysis.
The Hilbert curve point cloud serialization technology is used to preprocess and data enhancement of point cloud data. Combined with the Point Transformer deep learning network and Lovasz Hinge loss function, the model training process is optimized and detection accuracy and efficiency are improved.
It significantly improves the accuracy and efficiency of leakage detection, is suitable for leakage monitoring and maintenance of actual subway tunnels, and reduces calculation complexity and time cost.
Smart Images
Figure CN120471824A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an operation and maintenance technology for a subway shield tunnel, and in particular to a subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization. Background Art
[0002] During the operation of subway shield tunnels, leakage problems can have a serious impact on tunnel structure and operational safety. Traditional leakage detection methods rely on manual inspections and simple sensor data analysis, which are inefficient and inaccurate. With the development of lidar technology, the use of point cloud data for tunnel leakage detection has become an effective method. However, the high dimensionality and large-scale nature of point cloud data place higher demands on data processing and analysis. The Hilbert curve, as an efficient spatial serialization method, can maintain the spatial locality of point cloud data, helping deep learning models better learn leakage characteristics. Summary of the Invention
[0003] Purpose of the invention: The present invention provides a subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization, aiming to improve the accuracy and efficiency of leakage detection.
[0004] Technical solution: The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization of the present invention includes the following steps:
[0005] S1, using a mobile LiDAR system to obtain point cloud data of a subway shield tunnel;
[0006] S2, preprocessing of point cloud data, including removal of tunnel ancillary facilities and point cloud denoising;
[0007] S3, manually annotate the leakage instance labels, divide the point cloud data into leakage class and background class, build the leakage detection dataset, and divide the dataset into training set and validation set;
[0008] S4, perform data augmentation on the dataset;
[0009] S5, serializes the point cloud data in the leakage detection dataset based on the Hilbert space curve to optimize the spatial continuity of the point cloud and improve the input efficiency of the subsequent deep learning model;
[0010] S6, input the serialized point cloud data and the corresponding point cloud intensity values into the Point Transformer deep learning network for model training, construct the Lovasz Hinge loss function, and obtain the best model by optimizing the loss function;
[0011] S7, using the optimal model to perform leakage detection on the tunnel point cloud data to be inspected, obtain high-precision detection results, and visualize the detection results.
[0012] Furthermore, the removal of tunnel ancillary facilities in step S2 is specifically as follows:
[0013] The tunnel point cloud data is sliced along the tunnel direction at intervals of l rows (l = 0.5m) to form a slice set For each slice S j Perform cylinder fitting and determine the center coordinates C of the fitted elliptical cylinder j and radius R; fit the distance to the center of the elliptical cylinder C j Smaller than the tunnel design radius R d The points are identified as tunnel ancillary facility points and removed to obtain the point cloud data P' after removing the ancillary facilities.
[0014] Furthermore, the point cloud denoising in step S2 adopts a spatially distributed denoising algorithm, specifically:
[0015] Search for each point p in the point cloud i The k nearest neighboring points of the local neighborhood point set N(p i ); calculate the neighborhood point set N(p i ) in the Euclidean distance of k neighboring points:
[0016] d ij =||p i -p j ||
[0017] Calculate the mean μ and variance σ of these distances, define the distance range [μ-σ,μ+σ], filter and remove abnormal points outside this range, and obtain the denoised point cloud data.
[0018] Furthermore, the implementation process of step S3 is as follows:
[0019] Tunnel point cloud data was rendered using point cloud intensity values. Combined with leak images captured during data collection, leak instances were manually annotated in Cloud Compare, a point cloud data processing software.
[0020] After the leakage instance samples were labeled, they were divided into six areas according to the S3DIS dataset format. Area 5 was used as the test set, and the remaining areas were used as the training set. Within each area, the tunnel direction was divided into several blocks at intervals of d, and each block was input into the deep learning network as a batch.
[0021] Furthermore, the implementation process of step S4 is as follows:
[0022] The feature normalization of the point cloud data in the leakage detection dataset is performed to standardize the feature value range of each dimension and normalize the feature. As shown below:
[0023]
[0024] where f max ,f min are the maximum and minimum values of the features, respectively;
[0025] Randomly scale the point cloud data with a scaling factor s, satisfying s∈[s min ,s max ];
[0026] Randomly flip the point cloud data, that is, perform a mirror flip on any coordinate axis.
[0027] Furthermore, the implementation process of step S5 is as follows:
[0028] For each point p = (x, y, z), convert the three-dimensional coordinates into Gray code g x :
[0029]
[0030] in, Represents a bitwise exclusive OR (XOR) operation; It means shifting x right by one position, which is the integer part of x divided by 2;
[0031] Assume that each Gray code consists of n bits. For each bit position k, convert the three-dimensional Gray code g x ,g y ,g z Interleave in bit-level order to form a bit sequence of length 3n:
[0032] H 3k =g x (k) ,H 3k+1 =g y (k) ,H 3k+2 =g z (k)
[0033] where g x (k) ,g y (k) ,g z (k) Represents Gray code g x ,g y ,g z The k bits, from low to high, k = 0 is the lowest bit;
[0034] The final bit sequence after interleaving is:
[0035] H=H 3(n-1) H 3(n-1)+1 H 3(n -1)+2...H0H1H2
[0036] Based on the Skilling algorithm, the interleaved bit sequence H is transformed to obtain the transformed bit sequence H3' k :
[0037]
[0038] in, Represents bitwise exclusive OR operation, ∧ represents logical "and" operation; Represents the logical "not" operation;
[0039] Combine the transformed bit sequence H' into a single integer, which is the Hilbert index of the point:
[0040]
[0041] Among them, H k ' is the kth position in the Hilbert sequence after transformation, 2 k is the bit weight.
[0042] Furthermore, the PointTransformer deep learning network in step S6 includes four encoders and four decoders, the encoders are used to extract features of point cloud data, and the decoders are used to generate leakage detection results.
[0043] Furthermore, the LovaszHinge loss function in step S6 is specifically:
[0044] LovaszHinge loss function L LovaszHinge In the Hinge loss function L Hinge Based on the optimization, first define the matching set Indicates the classification error of positive and negative samples; for each pair of positive and negative samples, if their labels are consistent or inconsistent, they are included in the matching set:
[0045]
[0046] where y * and Represent the true label and predicted label respectively;
[0047] Use an indicator representing the IoU difference, which is a value in the {0,1} set, to determine the contribution of each sample to the final loss:
[0048]
[0049] Calculate the standard Hinge loss function:
[0050] L Hinge =max(1-F i (x)y * ,0)
[0051] Among them, L Hinge represents the Hinge loss function, F is the model output score of the i-th point;
[0052] Introducing the IoU difference into the loss function, we get the Lovász hinge loss function:
[0053]
[0054] in, It is a smooth extension of the IoU-based loss.
[0055] Furthermore, the model training in step S6 is specifically as follows:
[0056] Use LovaszHinge loss function L LovaszHinge Perform optimization and record the average intersection over union (mIoU) and average accuracy (mAcc) during training:
[0057]
[0058] Among them TP i , FP i , FN i ,TN i They represent the true positive, false positive, false negative, and true negative of the Nth category respectively; the model parameters with the highest mIoU and mAcc indicators on the validation set are selected as the final optimal model.
[0059] Furthermore, the implementation process of step S7 is as follows:
[0060] The point cloud data P of the tunnel to be detected test , serialized through the Hilbert space curve mapping function, and the serialized point cloud data P is obtained test s ; The serialized point cloud data P test s , input into the Point Transformer network; load the optimal model parameters and output the detection label L of each pointtest ; According to the detection label L test Visualize the leaking area.
[0061] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are as follows: the present invention combines deep learning with Hilbert curve point cloud serialization technology to achieve high-precision detection of subway shield tunnel leakage; the three-dimensional point cloud data is converted into a one-dimensional sequence through the Hilbert curve, which effectively maintains the spatial locality of the data and ensures the relative position relationship of adjacent points in the sequence, thereby maximizing the retention of the spatial structure information of the point cloud; this serialization method significantly reduces the dimension of the data, reduces the computational complexity of the deep learning model, and improves the training and inference efficiency of the model; in addition, the recursive fractal characteristics of the Hilbert curve make the data processing process more efficient and shorten the time cost of data preprocessing and model training; by optimizing the data processing process and utilizing the advanced PointTransformer deep learning model, the present invention not only improves the accuracy and reliability of leakage detection, but also greatly improves the overall detection efficiency, and is suitable for leakage monitoring and maintenance of actual subway tunnels, with significant technical advantages and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 is a flow chart of the present invention;
[0063] Figure 2 This is a schematic diagram of the subway shield tunnel leakage dataset collected by the present invention;
[0064] Figure 3 Schematic diagram of the serialization connection method of the Hilbert curve of the three-dimensional point cloud of the present invention;
[0065] Figure 4 This is a schematic diagram of the Hilbert curve serialization of the tunnel leakage point cloud of the present invention;
[0066] Figure 5 This is the leakage detection result before serialization of the Hilbert curve point cloud used in the present invention;
[0067] Figure 6 This is the leakage detection result after serialization of the Hilbert curve point cloud in the present invention. DETAILED DESCRIPTION
[0068] The present invention will be described in further detail below with reference to the accompanying drawings.
[0069] like Figure 1 As shown, the present invention provides a method for detecting leakage in a subway shield tunnel based on Hilbert curve point cloud serialization, comprising the following steps:
[0070] Step 1: Use a mobile lidar scanning system to obtain the point cloud data P(x, y, z) of the subway shield tunnel.
[0071] Step 2: Preprocess the acquired point cloud data, which includes removing tunnel ancillary facilities and point cloud denoising.
[0072] The tunnel point cloud data is sliced along the tunnel direction at intervals of l rows (l = 0.5m) to form a slice set For each slice S j Use the RANSAC algorithm to fit the cylinder and determine the center coordinates C of the fitted elliptical cylinder. j and radius R; fit the distance to the center of the elliptical cylinder C j Smaller than the tunnel design radius R d The points are identified as tunnel ancillary facilities and removed to obtain the point cloud data P' after removing the ancillary facilities. Taking the conventional subway shield tunnel as an example, the tunnel design radius R d =2.75m.
[0073] For the point cloud data P' after removing the auxiliary facilities, search for each point p in P' i The k nearest neighboring points of the local neighborhood point set N(p i ); calculate the neighborhood point set N(p i ) in the Euclidean distance of k neighboring points:
[0074] d ij =‖p i -p j ||
[0075] Calculate the mean μ and variance σ of these distances, define a distance range: [μ-σ,μ+σ], filter and remove abnormal points outside this range, and obtain denoised point cloud data.
[0076] Step 3: Manually label the leakage instance labels, divide the point cloud data into leakage class and background class, build a leakage detection dataset, and divide the dataset into training set and validation set.
[0077] The tunnel point cloud data was rendered using point cloud intensity values, and combined with the leakage images taken during data collection, manual instance annotation of leakage was performed in the point cloud data processing software Cloud Compare.
[0078] After labeling the leakage samples, we divided the labeled tunnel leakage samples into six areas according to the S3DIS dataset format. Area 5 served as the test set, and the remaining areas served as the training set. Within each area, we divided the samples into several blocks at 15-meter intervals along the tunnel's direction. Each block was then fed into the deep learning network as a batch.
[0079] Step 4: Perform data augmentation on the dataset.
[0080] The feature normalization of the point cloud data in the leakage detection dataset is performed to standardize the feature value range of each dimension and normalize the feature. As shown below:
[0081]
[0082] where f max ,f min are the maximum and minimum values of the feature, respectively.
[0083] Randomly scale the point cloud data with a scaling factor s, satisfying s∈[s min ,s max ]; Randomly flip the point cloud data, that is, mirror flip on any coordinate axis. Finally, the tunnel leakage dataset used for network training is obtained. Some tunnel leakage datasets are as follows Figure 2 shown.
[0084] Step 5: Use Hilbert curve to serialize point cloud. The serialization connection method of Hilbert in three-dimensional space is as follows: Figure 3 The specific implementation is as follows: for each point p = (x, y, z) in the tunnel leakage dataset, the three-dimensional coordinate representation is converted into Gray code:
[0085]
[0086] in, Represents a bitwise exclusive OR (XOR) operation. It means shifting x right by one position, which is the integer part of x divided by 2.
[0087] Assume that each Gray code consists of n bits. For each bit position k, convert the three-dimensional Gray code g x ,g y ,g z Interleave in bit-level order to form a bit sequence of length 3n:
[0088] H 3k =g x (k) ,H 3k+1 =g y (k) ,H 3k+2 =g z (k)
[0089] where g x (k) ,g y(k) ,g z (k) Represents Gray code g x ,g y ,g z k bits (from low to high, k=0 is the lowest bit)
[0090] The final bit sequence after interleaving is:
[0091] H=H 3(n-1) H 3(n-1)+1 H 3(n -1)+2...H0H1H2
[0092] Based on the Skilling algorithm, the interleaved bit sequence H is transformed to obtain the transformed bit sequence H3' k :
[0093]
[0094] in, Represents a bitwise exclusive OR (XOR) operation, and ∧ represents a logical AND operation (AND). Represents the logical NOT operation (NOT).
[0095] Combine the transformed bit sequence H' into a single integer, which is the Hilbert index of the point:
[0096]
[0097] Among them H k ' is the kth bit in the Hilbert sequence after transformation (starting from 0, k=0 is the lowest bit), 2 k is the bit weight. The tunnel leakage after point cloud serialization using Hilbert curve is as follows Figure 4 shown.
[0098] Step 6: Input the obtained serialized point cloud data and its corresponding point cloud intensity values into the PointTransformer deep learning network for model training.
[0099] The PointTransformer deep learning network includes four encoders and four decoders. The encoders are used to extract features of point cloud data, and the decoders are used to generate leakage detection results.
[0100] PointTransformer deep learning network uses LovaszHinge loss function L LovaszHinge Optimize to solve the problem of sample imbalance in the dataset.
[0101] LovaszHinge loss function L LovaszHinge In the Hinge loss function L Hinge First, define the matching set Indicates the classification error of positive and negative samples. For each pair of positive and negative samples, if their labels are consistent or inconsistent, they are included in the matching set:
[0102]
[0103] Among them, y * and represent the true label and the predicted label respectively.
[0104] Use an indicator representing the IoU difference, which is a value in the {0,1} set, to determine the contribution of each sample to the final loss:
[0105]
[0106] Calculate the standard Hinge loss function:
[0107] L Hinge =max(1-F i (x)y * ,0)
[0108] Among them L Hinge Represents the Hinge loss function, and F is the model output score of the i-th point.
[0109] Introducing the IoU difference into the loss function, we get the Lovászhinge loss function:
[0110]
[0111] in, It is a smooth extension of the IoU-based loss.
[0112] During the training process, record the mIoU (mean intersection over union) and mAcc (mean accuracy) metrics:
[0113]
[0114] Among them TP i , FP i , FN i , TN i They represent the true positive, false positive, false negative, and true negative of class N, respectively. The model parameters with the highest mIoU and mAcc indicators on the validation set are selected as the final optimal model.
[0115] Step 7: Use the trained optimal model to perform leakage detection on the tunnel point cloud data to be inspected, and display the detection results in a visual form.
[0116] The point cloud data P of the tunnel to be detected test , serialized through the Hilbert space curve mapping function, and the serialized point cloud data P is obtained test s ; The serialized point cloud data P test s , input into the Point Transformer network; load the optimal model parameters and output the detection label L of each point test ; According to the detection label L test Visualize the leaking area.
[0117] Before using Hilbert point cloud serialization, the leakage detection results obtained by training the Point Transformer network and predicting the test data are as follows: Figure 5 The leakage detection results obtained by training the PointTransformer network after serializing the Hilbert point cloud and predicting the test data are shown in Figure 6 The results show that the accuracy of tunnel leakage detection is significantly improved after using Hilbert point cloud serialization.
[0118] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization, characterized in that: The following steps are involved: S1, using a mobile LiDAR system to obtain point cloud data of a subway shield tunnel; S2, preprocessing of point cloud data, including removal of tunnel ancillary facilities and point cloud denoising; S3, manually annotate the leakage instance labels, divide the point cloud data into leakage class and background class, build the leakage detection dataset, and divide the dataset into training set and validation set; S4, perform data augmentation on the dataset; S5, serializes the point cloud data in the leakage detection dataset based on the Hilbert space curve to optimize the spatial continuity of the point cloud and improve the input efficiency of the subsequent deep learning model; S6, input the serialized point cloud data and the corresponding point cloud intensity values into the Point Transformer deep learning network for model training, construct the Lovasz Hinge loss function, and obtain the best model by optimizing the loss function; S7, using the optimal model to perform leakage detection on the tunnel point cloud data to be inspected, obtain high-precision detection results, and visualize the detection results.
2. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The removal of tunnel ancillary facilities in step S2 is specifically as follows: The tunnel point cloud data is sliced along the tunnel direction at intervals of l rows (l = 0.5m) to form a slice set For each slice S j Perform cylinder fitting and determine the center coordinates C of the fitted elliptical cylinder j and radius R; fit the distance to the center of the elliptical cylinder C j Smaller than the tunnel design radius R d The points are identified as tunnel ancillary facility points and removed to obtain the point cloud data P' after removing the ancillary facilities.
3. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The point cloud denoising in step S2 adopts a spatial distribution denoising algorithm, specifically: Search for each point p in the point cloud i The k nearest neighboring points of the local neighborhood point set N(p i ); calculate the neighborhood point set N(p i ) in the Euclidean distance of k neighboring points: d ij =||p i -p j || Calculate the mean μ and variance σ of these distances, define the distance range [μ-σ,μ+σ], filter and remove abnormal points outside this range, and obtain the denoised point cloud data.
4. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The implementation process of step S3 is as follows: Tunnel point cloud data was rendered using point cloud intensity values. Combined with leak images captured during data collection, leak instances were manually annotated in Cloud Compare, a point cloud data processing software. After the leakage instance samples were labeled, they were divided into six areas according to the S3DIS dataset format. Area 5 was used as the test set, and the remaining areas were used as the training set. Within each area, the tunnel direction was divided into several blocks at intervals of d, and each block was input into the deep learning network as a batch.
5. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The implementation process of step S4 is as follows: The feature normalization of the point cloud data in the leakage detection dataset is performed to standardize the feature value range of each dimension and normalize the feature. As shown below: where f max ,f min are the maximum and minimum values of the features, respectively; Randomly scale the point cloud data with a scaling factor s, satisfying s∈[s min ,s max ]; Randomly flip the point cloud data, that is, perform a mirror flip on any coordinate axis.
6. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The implementation process of step S5 is as follows: For each point p = (x, y, z), convert the three-dimensional coordinates into Gray code g x : in, Represents a bitwise exclusive OR (XOR) operation; It means shifting x right by one position, which is the integer part of x divided by 2; Assume that each Gray code consists of n bits. For each bit position k, convert the three-dimensional Gray code g x ,g y ,g z Interleave in bit-level order to form a bit sequence of length 3n: H 3k =g x (k) ,H 3k+1 =g y (k) ,H 3k+2 =g z (k) where g x (k) ,g y (k) ,g z (k) Represents Gray code g x ,g y ,g z The k bits, from low to high, k = 0 is the lowest bit; The final bit sequence after interleaving is: H=H 3(n-1) H 3(n-1)+1 H 3(n -1)+2...H0H1H2 Based on the Skilling algorithm, the interleaved bit sequence H is transformed to obtain the transformed bit sequence H3' k : in, Represents bitwise exclusive OR operation, ∧ represents logical "and" operation; Represents the logical "not" operation; Combine the transformed bit sequence H' into a single integer, which is the Hilbert index of the point: Among them, H k ' is the kth position in the Hilbert sequence after transformation, 2 k is the bit weight.
7. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The Point Transformer deep learning network in step S6 includes four encoders and four decoders. The encoders are used to extract features of point cloud data, and the decoders are used to generate leakage detection results.
8. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The Lovasz Hinge loss function in step S6 is specifically: Lovasz Hinge loss function L LovaszHinge In the Hinge loss function L Hinge Based on the optimization, first define the matching set Indicates the classification error of positive and negative samples; for each pair of positive and negative samples, if their labels are consistent or inconsistent, they are included in the matching set: where y * and Represent the true label and predicted label respectively; Use an indicator representing the IoU difference, which is a value in the {0,1} set, to determine the contribution of each sample to the final loss: Calculate the standard Hinge loss function: L Hinge =max(1-F i (x)y * ,0) Among them, L Hinge represents the Hinge loss function, F is the model output score of the i-th point; Introducing the IoU difference into the loss function, we get the Lovász hinge loss function: in, It is a smooth extension of the IoU-based loss.
9. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1 is characterized in that: The model training described in step S6 is specifically as follows: Use LovaszHinge loss function L LovaszHinge Perform optimization and record the average intersection over union (mIoU) and average accuracy (mAcc) during training: Among them TP i , FP i , FN i ,TN i They represent the true positive, false positive, false negative and true negative of the Nth category respectively; The model parameters with the highest mIoU and mAcc indicators on the validation set are selected as the final best model.
10. The subway shield tunnel leakage detection method based on Hilbert curve point cloud serialization according to claim 1, characterized in that: The implementation process of step S7 is as follows: The point cloud data P of the tunnel to be detected test , serialized through the Hilbert space curve mapping function, and the serialized point cloud data P is obtained test s ; The serialized point cloud data P test s , input into the Point Transformer network; load the optimal model parameters and output the detection label L of each point test ; According to the detection label L test Visualize the leaking area.