Tunnel lining leakage infrared-millimeter wave fusion detection method and system

Abstract

The present invention relates to an infrared-millimeter wave fusion detection method and system for tunnel lining leakage, and belongs to the technical field of civil engineering detection. The method comprises: acquiring high-density millimeter wave radar data and infrared thermal imager data; performing data preprocessing to achieve spatial alignment and time synchronization; extracting topological features, performing manifold learning mapping on the preprocessed millimeter wave radar data, and constructing a topological signature matrix; performing multi-phase spectrum analysis, performing nonlinear phase space reconstruction and multi-scale entropy feature analysis on the preprocessed infrared thermal imager data, and obtaining an entropy feature map; performing probability fusion and progressive reasoning to generate a leakage area probability map through a conditional random field model and a progressive belief propagation algorithm; and finally determining the leakage area location and leakage rate. The method overcomes the limitations of a single sensor detection method, achieves high-precision detection and quantitative evaluation of tunnel lining leakage, and has technical effects such as high detection accuracy, strong environmental adaptability, and a high degree of automation.

Description

Technical Field

[0001] The present invention relates to the technical field of civil engineering detection, and in particular to a tunnel lining leakage infrared-millimeter wave fusion detection method and a system thereof. Background Art

[0002] As critical transportation infrastructure, tunnel safety is directly linked to the reliability of transportation and the safety of people's lives and property. Lining leakage is a common problem during tunnel operation, often caused by factors such as groundwater pressure and construction quality. In severe cases, this can degrade tunnel structural performance and even lead to accidents.

[0003] Currently, tunnel lining leakage detection relies primarily on single methods, including manual inspections, infrared thermal imaging, or radar. Manual inspections are inefficient and subjective, making it difficult to quantitatively analyze tunnel lining leakage. While infrared thermal imaging alone can detect areas of abnormal surface temperature, it struggles to determine the internal structure. Radar detection alone, while capable of detecting internal structures, has limited ability to identify subtle surface leaks.

[0004] Existing technologies lack an efficient method for simultaneously detecting surface leakage characteristics and internal structural conditions in tunnel linings, and also lack the technical means to effectively fuse multi-source heterogeneous data. Therefore, a method and system for high-precision tunnel lining leakage detection is urgently needed that can combine infrared thermal imaging and millimeter-wave radar technology. Summary of the Invention

[0005] The purpose of the present invention is to provide an infrared-millimeter wave fusion detection method and system for tunnel lining leakage. By integrating infrared thermal imaging technology and millimeter wave radar technology, combined with advanced theories such as topological mathematics, nonlinear dynamics and probabilistic graphical models, high-precision detection and quantitative evaluation of tunnel lining leakage can be achieved.

[0006] The present invention proposes a tunnel lining leakage infrared-millimeter wave fusion detection method, comprising:

[0007] Acquire high-density millimeter-wave radar data and infrared thermal imager data;

[0008] Performing data preprocessing to spatially align and time-synchronize the high-density millimeter-wave radar data and the infrared thermal imager data;

[0009] Extract topological features, perform manifold learning mapping on the pre-processed high-density millimeter-wave radar data, and construct a topological signature matrix;

[0010] Perform multi-phase spectrum analysis, perform nonlinear phase space reconstruction and multi-scale entropy feature analysis on the pre-processed infrared thermal imager data, and obtain the entropy feature map;

[0011] Performing probability fusion and progressive reasoning, inputting the topological signature matrix and the entropy feature map into a conditional random field model, and generating a leakage area probability map through a progressive belief propagation algorithm;

[0012] Based on the leakage area probability map, the location of the tunnel lining leakage area and the leakage rate are determined.

[0013] Preferably, the data preprocessing includes:

[0014] Performing frequency-varying sorting on the high-density millimeter-wave radar data, extracting one-dimensional signal and two-dimensional signal features, and constructing point cloud data;

[0015] performing non-uniformity correction and distortion correction on the infrared thermal imager data to generate a standardized thermodynamic data matrix;

[0016] Achieving spatial alignment of the point cloud data with the standardized thermodynamic data matrix through feature point matching and coordinate transformation;

[0017] Time synchronization between the point cloud data and the standardized thermodynamic data matrix is ​​achieved through timestamp alignment and interpolation resampling.

[0018] Preferably, the manifold learning mapping includes:

[0019] Construct a K-nearest neighbor neighborhood graph, where the value of K ranges from 10 to 20;

[0020] Calculating geodesic distances of the point cloud surface based on the neighborhood graph;

[0021] Learning the Riemannian metric tensor based on local geometric properties;

[0022] A dimensionality reduction mapping is performed under the constraint of maintaining local distance relationships to generate a low-dimensional manifold embedding.

[0023] Preferably, the construction of the topological signature matrix includes:

[0024] constructing a Vietoris-Rips complex based on the low-dimensional manifold embedding;

[0025] Calculate the Betti number under different radius parameters and generate persistence diagrams;

[0026] Converting the persistence graph into a vector representation;

[0027] Construct a matrix containing multi-scale topological information;

[0028] Key topological features are enhanced through eigenvalue decomposition and matrix operations to form a topological signature matrix.

[0029] Preferably, the nonlinear phase space reconstruction comprises:

[0030] Determine the optimal delay time through mutual information analysis;

[0031] Determine embedding dimensions through pseudo-nearest neighbor analysis;

[0032] Construct phase space trajectories based on delayed coordinates;

[0033] Identify strange attractor structures in phase space;

[0034] Compute the trajectory density distribution in phase space and identify the dynamic characteristics caused by leakage.

[0035] Preferably, the multi-scale entropy feature analysis includes:

[0036] Generate multi-scale time series through sliding average;

[0037] Calculate sample entropy and fuzzy entropy at each scale;

[0038] Calculate cross entropy between different regions to evaluate entropy differences;

[0039] Analyze the law of entropy value changing with scale;

[0040] Map the entropy features back to the spatial domain and construct an entropy feature map;

[0041] The spatial gradient of entropy value is calculated to highlight the boundary features of leakage area.

[0042] Preferably, the conditional random field model includes:

[0043] The point cloud data points and thermodynamic data points are used as nodes;

[0044] Establish edge connections based on spatial proximity and feature similarity;

[0045] Construct a two-layer graph structure consisting of a point cloud layer and a thermodynamic layer;

[0046] Design node potential functions, edge potential functions and inter-layer potential functions to describe the relationship between features and labels and the interaction between nodes.

[0047] Preferably, the progressive belief propagation algorithm includes:

[0048] Initialize messages based on node potential function;

[0049] Belief propagation is first performed within each layer, and then belief propagation is performed between layers;

[0050] Guide the propagation direction based on areas with high confidence;

[0051] Adaptively terminate propagation based on the magnitude of belief changes;

[0052] Calculate the class edge probability of each node;

[0053] The class with the maximum posterior probability is selected as the outcome;

[0054] Assess confidence and uncertainty in decisions.

[0055] Preferably, determining the location and leakage rate of the tunnel lining leakage area includes:

[0056] Extract the precise boundaries of the leakage area;

[0057] Estimate the depth distribution of the leakage area;

[0058] Establish leakage model based on thermodynamics and fluid mechanics principles;

[0059] Estimate model parameters from observed data;

[0060] Calculate the leakage per unit time and per unit area;

[0061] Analyze the time trend of leakage parameters;

[0062] Build short-term and long-term forecast models;

[0063] Set warning thresholds at different levels.

[0064] The tunnel lining leakage infrared-millimeter wave fusion detection system is characterized by comprising:

[0065] A sensor subsystem, including a 77GHz millimeter-wave radar and a passive infrared thermal imager, is used to acquire high-density millimeter-wave radar data and infrared thermal imager data;

[0066] The mobile platform subsystem includes a track-type mobile chassis and a sensor mounting structure to support stable movement and precise positioning of the sensor;

[0067] A data processing subsystem, including an embedded computing unit and a high-performance computing server, is used to perform data preprocessing, topological feature extraction, multi-phase spectrum analysis, and probabilistic fusion and inference;

[0068] The data processing subsystem includes:

[0069] A data preprocessing module, configured to spatially align and temporally synchronize the high-density millimeter-wave radar data and the infrared thermal imager data;

[0070] The topological feature extraction module is used to perform manifold learning mapping on the pre-processed high-density millimeter-wave radar data and construct a topological signature matrix;

[0071] The multi-phase spectrum analysis module is used to perform nonlinear phase space reconstruction and multi-scale entropy feature analysis on the pre-processed infrared thermal imager data to obtain the entropy feature map;

[0072] A probability fusion and reasoning module, configured to input the topological signature matrix and the entropy feature map into a conditional random field model, and generate a leakage area probability map through a progressive belief propagation algorithm;

[0073] The result output and evaluation module is used to determine the location and leakage rate of the tunnel lining leakage area based on the leakage area probability map.

[0074] The beneficial effects of the present invention include:

[0075] 1. Through multi-sensor fusion technology, the limitations of single-sensor detection methods are overcome, and the accuracy and reliability of detection are improved. The detection accuracy rate is increased by more than 50%, and the false positive rate is reduced by 70%.

[0076] 2. Introducing a topology-aware point cloud processing method to make the detection results robust to changes in tunnel geometry and adapt to tunnel surfaces with different curvatures.

[0077] 3. Using multi-phase spectrum analysis technology, it can effectively distinguish between thermal patterns and environmental fluctuations caused by leakage, and improve the detection ability of small area leakage.

[0078] 4. Through probability fusion and progressive reasoning mechanism, the system handles data uncertainty, integrates static and dynamic features, and achieves precise positioning of the leakage area.

[0079] 5. A leakage rate calculation model was established to achieve quantitative assessment of the degree of leakage, providing a scientific basis for tunnel maintenance and repair.

[0080] 6. The system's modular design and standardized interfaces ensure good scalability and adaptability, and can meet the detection needs of different tunnel types and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 Schematic diagram of the overall architecture of the infrared-millimeter wave fusion detection system for tunnel lining leakage of the present invention;

[0082] Figure 2 Schematic diagram of the process of the infrared-millimeter wave fusion detection method for tunnel lining leakage of the present invention;

[0083] Figure 3 Schematic diagram of the flow of the data preprocessing module of the present invention;

[0084] Figure 4 Schematic diagram of the topological feature extraction module of the present invention;

[0085] Figure 5 Schematic diagram of the process of the multi-phase spectrum analysis module of the present invention;

[0086] Figure 6 Schematic diagram of the flow of the probability fusion and progressive reasoning module of the present invention. DETAILED DESCRIPTION

[0087] Please refer to Figure 1 - Figure 6 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0088] Reference Figure 1 The tunnel lining leakage infrared-millimeter wave fusion detection system provided by the present invention includes a sensor subsystem 10, a mobile platform subsystem 20, a data processing subsystem 30 and a power supply and communication subsystem 40.

[0089] The sensor subsystem 10 includes a 77GHz millimeter-wave radar 11 and a passive infrared thermal imager 12, which are used to acquire high-density millimeter-wave radar and infrared thermal imager data. Preferably, the 77GHz millimeter-wave radar 11 has a range resolution of no greater than 2mm, an angular resolution of no greater than 0.1°, and a scanning frequency of 2-5Hz. The passive infrared thermal imager 12 has a temperature resolution of no greater than 0.05°C, a spatial resolution of no less than 640×480 pixels, and an acquisition frequency synchronized with the millimeter-wave radar. This parameter setting ensures the acquisition of sufficiently detailed data to capture subtle changes in the tunnel lining surface and interior.

[0090] The mobile platform subsystem 20 comprises a track-mounted mobile chassis 21 and a sensor-mounted mounting structure 22, designed to support stable movement and precise positioning of the sensors. The sensors are mounted on the mobile chassis in a rack-mounted layout, ensuring sufficient overlap between the two sensors' fields of view while maintaining a fixed distance and angle from the tunnel lining. In practice, the movement speed is typically controlled between 0.5 and 2 m / s to ensure continuous and complete data collection.

[0091] The data processing subsystem 30 is the core of the present invention and includes an embedded computing unit 31 and a high-performance computing server 32, as well as a series of functional modules: a data preprocessing module 33, a topological feature extraction module 34, a multi-phase spectrum analysis module 35, a probability fusion and inference module 36, and a result output and evaluation module 37. These modules work together to complete the entire process from data acquisition to leak detection result output.

[0092] The power and communication subsystem 40 includes a long-lasting power supply system 41 and wired / wireless communication modules 42, ensuring long-term system operation and data transmission. Preferably, the power system uses a high-capacity lithium battery capable of supporting more than eight hours of continuous operation. The communication module supports multiple communication methods, including 4G, 5G, and WiFi, ensuring real-time data transmission to a remote monitoring center.

[0093] Reference Figure 2 The tunnel lining leakage infrared-millimeter wave fusion detection method of the present invention comprises the following steps:

[0094] This step first obtains high-density millimeter-wave radar data and infrared thermal imager data through a 77GHz millimeter-wave radar and a passive infrared thermal imager, respectively. Then, data preprocessing is performed to spatially align and temporally synchronize the two heterogeneous data.

[0095] Reference Figure 3 , data preprocessing specifically includes the following:

[0096] Frequency-variant sorting is performed on high-density millimeter-wave radar data to extract one-dimensional and two-dimensional signal features and construct point cloud data. Specifically, the raw radar data is first de-noised using an adaptive filter to remove ambient noise. A frequency-variant sorting algorithm is then used to extract one-dimensional and two-dimensional signal features from the radar data. Finally, three-dimensional point cloud data is reconstructed based on these signal features.

[0097] Infrared thermal imager data is corrected for non-uniformity and distortion to generate a standardized thermodynamic data matrix. Specifically, non-uniformity correction is first performed to eliminate non-uniformity in sensor response; distortion correction is then performed to correct for image distortion caused by the optical system; temperature calibration is then performed to convert grayscale values ​​into actual temperature values; and finally, a time-series-based thermodynamic data matrix is ​​constructed.

[0098] The spatial alignment of the point cloud data with the standardized thermodynamic data matrix is ​​achieved through feature point matching and coordinate transformation. Specifically, the common feature points in the two data sets are first identified; the rotation matrix and translation vector are then calculated to achieve rigid body transformation. For nonlinear deformation, non-rigid registration methods are used.

[0099] Time synchronization between point cloud data and the standardized thermodynamic data matrix is ​​achieved through timestamp alignment and interpolation resampling. Specifically, data alignment is first performed based on acquisition timestamps; then, time domain interpolation is performed on the asynchronous data; and finally, cross-correlation analysis is used to verify the synchronization quality.

[0100] This step ensures the data quality and consistency for subsequent processing and lays the foundation for feature extraction and fusion.

[0101] In this step, the preprocessed high-density millimeter-wave radar data is subjected to manifold learning mapping to construct a topological signature matrix.

[0102] Reference Figure 4 , topological feature extraction specifically includes the following contents:

[0103] First, a manifold learning mapping is performed, including constructing a K-nearest neighbor neighborhood graph, where K ranges from 10 to 20; calculating the geodesic distance of the point cloud surface based on the neighborhood graph; learning the Riemannian metric tensor based on local geometric properties; and performing dimensionality reduction mapping under the constraint of maintaining local distance relationships to generate a low-dimensional manifold embedding.

[0104] The geodesic distance calculation here uses the metric tensor in Riemannian geometry and can be expressed as:

[0105] ,

[0106] in: for point and point The geodesic distance between For connection and The curve, Points on the curve The metric tensor at , is the tangent vector of the curve, is the curve parameter, .

[0107] The Riemannian metric tensor learning process optimizes the following objective function:

[0108] ,

[0109] in: is the metric tensor to be learned, for point and point The weight between them is usually set to 1 when point i is the lk nearest neighbor of point i, otherwise it is 0. Based on the metric tensor The calculated distance, is the observation distance, is a regularization parameter used to control the complexity of the model. is the Frobenius norm of the matrix, calculated as the square root of the sum of the squares of all elements of the matrix. The value is usually set between 0.01 and 0.1 to balance the fitting accuracy and model complexity. For tunnel structures, we recommend using 0.05 as the initial value and then fine-tuning it based on the actual effect.

[0110] Then, a topological signature matrix is ​​constructed, including constructing a Vietoris-Rips complex based on low-dimensional manifold embedding; calculating the Betti number under different radius parameters to generate a persistence graph; converting the persistence graph into a vector representation; constructing a matrix containing multi-scale topological information; and enhancing key topological features through eigenvalue decomposition and matrix operations to form a topological signature matrix.

[0111] The Vietoris-Rips complex construction process can be expressed as:

[0112] ,

[0113] in: The parameters are Vietoris-Rips complex, is a point set containing all point cloud data points, is a simplex, representing a subset of the point set, for point and point The distance between is the distance threshold parameter, which controls which pairs of points form connections. In practical applications, we usually start from a smaller The value starts at a small value (such as 0.5 times the average nearest neighbor distance of the point cloud) and then gradually increases to a larger value (such as 0.5 times the diameter of the point cloud).

[0114] The persistence graph generation process is calculated differently Betti number under the value , , which represent the number of connected components, the number of loops and the number of cavities respectively. Follow Increase and decrease, and The leakage region usually appears as a special pattern in the persistence graph, such as Areas with higher values ​​tend to correspond to seepage channels.

[0115] Topological signature matrix The construction is carried out by converting the persistence graph information into a matrix form and performing feature enhancement:

[0116] ,

[0117] in: is an eigenvector matrix whose column vectors are the eigenvectors of the matrix, is an eigenvalue diagonal matrix, and the diagonal elements are the corresponding eigenvalues, express The transpose of . Feature enhancement is achieved by adjusting the eigenvalues:

[0118] ,

[0119] in: is the enhancement function. In this embodiment, an exponential enhancement form is adopted:

[0120] ,

[0121] in: For the eigenvalues, is the enhancement factor, usually ranging from 1.5 to 2.5, used to amplify the impact of important features. is the threshold used to filter important eigenvalues, usually the median of all eigenvalues. For tunnel lining leakage detection, we usually choose As a default value, this parameter can be adjusted according to the significance of the actual leakage characteristics.

[0122] The enhanced topological signature matrix is:

[0123] ,

[0124] This step effectively extracts the topological features in the point cloud data, especially with good recognition capabilities for structures such as cavities and cracks in leakage areas.

[0125] In this step, the preprocessed infrared thermal imager data is subjected to nonlinear phase space reconstruction and multi-scale entropy feature analysis to obtain an entropy feature map.

[0126] Reference Figure 5 , multi-phase spectrum analysis specifically includes the following contents:

[0127] First, a nonlinear phase space reconstruction is performed, including determining the optimal delay time through mutual information analysis; determining the embedding dimension through false neighbor analysis; constructing phase space trajectories based on the delay coordinates; identifying the strange attractor structure in the phase space; calculating the trajectory density distribution in the phase space, and identifying the dynamic characteristics caused by leakage.

[0128] The phase space reconstruction of time series is based on Takens embedding theorem. , the phase space is reconstructed as:

[0129] ,

[0130] in: is the reconstructed phase space vector, which represents the time status, The original time series at time The value of is the delay time, which represents the time interval between adjacent coordinates. is the embedding dimension, which represents the dimension of the reconstruction space, Indicates the length of the time window extracted from the time series.

[0131] Optimal delay time Determine through mutual information analysis, select the first local minimum point of the mutual information function:

[0132] ,

[0133] in: The delay time is The mutual information value when For time series in time The value is The probability of For time series in time The value is The probability of is the joint probability, indicating that the time series is in time The value is And in time The value is The probability of is the logarithm with base 2. In practical applications, the mutual information function is usually The first local minimum is reached when the number of sampling points is between 5 and 15, and for tunnel thermal imaging data, it is usually around 10 sampling points.

[0134] Embedding Dimension Through false neighbor analysis, it is determined that the dimension with a false neighbor ratio lower than the threshold is the appropriate embedding dimension:

[0135] ,

[0136] in: Dimension The proportion of false neighbors when dimensional space but appear to be adjacent The ratio of points far away in dimensional space, is the total number of points in the time series, is a Heaviside step function that outputs 1 when the input is greater than 0, otherwise it outputs 0. for The first dimension in the reconstruction space Points, for The index of the nearest neighbor of represents the Euclidean distance, is a threshold used to determine whether the distance change is large enough, and is usually set to 10 to 15. In this embodiment, the embedding dimension is usually set to 3 to 5, which is determined according to the complexity of the tunnel thermal imaging data.

[0137] Then, a multi-scale entropy feature analysis is performed, including generating a multi-scale time series through sliding average; calculating sample entropy and fuzzy entropy at each scale; calculating the cross entropy between different regions to evaluate the entropy difference; analyzing the law of entropy value change with scale; mapping the entropy feature back to the spatial domain and constructing an entropy feature map; calculating the spatial gradient of the entropy value to highlight the boundary characteristics of the leakage area.

[0138] The multi-scale time series generation process is:

[0139] ,

[0140] in: The scale factor is The first elements, is the first elements, is the scale factor, which indicates the window size for coarsening the original sequence. is the original sequence length, is the length of the coarse-grained sequence. In this embodiment, the scale factor The range is 1 to 20 to cover different time scales from fast changes to slow changes.

[0141] The sample entropy calculation formula is:

[0142] ,

[0143] in: The parameters are The sample entropy of represents the complexity of the time series. is the pattern length, indicating the dimension of the comparison vector, is the similarity threshold, which represents the maximum distance between two vectors that are considered similar. is the sequence length, For distance less than of dimensional vector logarithm, For distance less than of dimensional vector logarithm. In practical applications, Usually take 1 or 2, Take 0.1 to 0.25 times the standard deviation of the time series. For tunnel thermal imaging data, we usually choose Standard deviation.

[0144] Cross entropy calculation is used to evaluate the entropy difference between different regions:

[0145] ,

[0146] in: is the cross-sample entropy, which represents the similarity between two time series, From the sequence of dimensional vector and from the sequence of The distance between the vectors is less than The number of vector pairs, From the sequence of dimensional vector and from the sequence of The distance between the vectors is less than The cross-entropy between leaky and normal regions is typically significantly higher than the cross-entropy between normal regions, making it an important feature for identifying leaks. In practice, we set the cross-entropy threshold to 1.5 times the average cross-entropy between normal regions.

[0147] The entropy feature map is constructed by mapping the multi-scale entropy of each spatial point back to the original space:

[0148] ,

[0149] in: for point The entropy eigenvector at , represents the entropy value of the point at different scales, The scale factor is The multi-scale entropy value when is the maximum scale factor, which is 20 in this embodiment. In this way, each spatial point is associated with a The entropy eigenvector is associated with .

[0150] The entropy gradient calculation is used to highlight the boundaries of leaky areas:

[0151] ,

[0152] in: for point The entropy gradient vector at , represents the direction and rate of change of the entropy feature at this point, is the entropy characteristic along The partial derivative in the direction, is the entropy characteristic along The boundary of the leakage area is usually manifested as an area with a large entropy gradient amplitude. When the gradient amplitude exceeds 3 times the average value of the background gradient, it can be identified as a potential leakage boundary.

[0153] This step effectively extracts the dynamic features in the thermodynamic data, especially has a good ability to identify heat conduction anomalies caused by leakage.

[0154] In this step, the topological signature matrix and entropy feature map are input into the conditional random field model, and the leakage area probability map is generated through the progressive belief propagation algorithm.

[0155] Reference Figure 6 ,Probability fusion and progressive reasoning specifically include the following contents:

[0156] First, a conditional random field model is constructed, which includes taking point cloud data points and thermodynamic data points as nodes; establishing edge connections based on spatial proximity and feature similarity; constructing a two-layer graph structure consisting of a point cloud layer and a thermodynamic layer; and designing node potential functions, edge potential functions, and inter-layer potential functions to describe the relationship between features and labels and the interaction between nodes.

[0157] The probability distribution of the conditional random field model can be expressed as:

[0158] ,in: For a given observation Next label The conditional probability distribution of Represents the label set of all nodes, represents the set of observation features of all nodes, is a normalization factor to ensure that the sum of probabilities is 1, is the node potential function, indicating the node Tags and observation characteristics relationship, is the edge potential function, representing the adjacent nodes and Tags and The relationship between is the inter-layer potential function, indicating the corresponding nodes between different layers and Tags and The relationship between and is the index of adjacent nodes in the same layer, and It is the corresponding node index between different layers.

[0159] The node potential function is defined as:

[0160] ,

[0161] in: is the characteristic function, representing the node Tags Its observation characteristics The relationship between is the weight parameter, indicating the The importance of the feature function is obtained through learning, In this embodiment, the characteristic function includes topological features and entropy features, and through these characteristic functions, the information of the topological signature matrix and the entropy feature map is integrated into the model.

[0162] The edge potential function is defined as:

[0163] ,

[0164] in: is an indicator function, which is 1 when the condition is true and 0 otherwise. It indicates whether the labels of nodes i and j are different. is a smoothing parameter that controls the strength of label consistency, is the spatial distance between nodes i and j, is the distance attenuation parameter, which controls the influence of distance on the smoothing effect. Usually 0.5 to 2, Take 0.1 to 0.5. For tunnel lining leakage detection, we recommend the initial value and ,These parameters can be adjusted according to the actual tunnel environment and leakage characteristics.

[0165] The interlayer potential function is defined as:

[0166] ,

[0167] in: is the inter-layer association strength parameter, which controls the strength of label consistency between different layers. is the feature similarity between nodes i and k, with a value range of [0,1], where a larger value indicates a greater similarity. Usually takes 1 to 3, similarity Calculated by cosine similarity of feature vectors:

[0168] ,

[0169] in: and are the feature vectors of nodes i and k respectively, represents the vector dot product, Represents a vector The Euclidean norm of .

[0170] Next, we perform variational Bayesian parameter learning to learn the potential function parameters of the conditional random field based on the training data. The variational Bayesian method estimates parameters by minimizing the evidence lower bound (ELBO):

[0171] ,

[0172] in: is the posterior distribution The variational approximation of is used to approximate the posterior distribution that is difficult to solve directly. are model parameters, including , and wait, Represents the distribution By alternating optimization and To maximize ELBO, we can obtain the optimal parameter estimation.

[0173] Then, a progressive belief propagation algorithm is executed, including initializing messages based on node potential functions; first propagating beliefs within each layer, and then propagating beliefs between layers; guiding the propagation direction according to areas with high confidence; adaptively terminating propagation according to the magnitude of belief changes; calculating the category edge probability of each node; selecting the category with the maximum posterior probability as the result; and evaluating the confidence and uncertainty of the decision.

[0174] The message update rule in the belief propagation process is:

[0175] ,

[0176] in: For nodes Send to node About Tags Message, indicating the node For Node Should have label "Faith", is a normalization constant to ensure that the sum of the messages is 1, Indicates the node The sum of all possible labels, For nodes The neighbor set of node All directly connected nodes, Representation node Exclude nodes from the neighbor set All nodes of Indicates that all other neighbors send to the node The product of the messages.

[0177] The progressive strategy design reduces computational complexity and improves convergence speed by first propagating within each layer and then between layers. The propagation order is guided by the confidence level, starting with the area with high confidence:

[0178] ,

[0179] Where: Confidence(i) is the confidence of node i, which indicates the certainty of its label prediction. is the most likely label probability of node i, is the probability of the second most likely label. The higher the confidence, the more certain the prediction is, and the information propagation of these nodes should be given priority.

[0180] The adaptive termination condition is based on the magnitude of the belief change:

[0181] ,

[0182] Where: ΔBelief is the total amount of belief change, indicating the degree of change in the beliefs of all nodes, is the label of node i after the tth iteration The belief that Probability estimate For the The belief of the iterations. The propagation is terminated when ΔBelief < ε or the maximum number of iterations is reached. ε is usually set to 0.001 to 0.01, and the maximum number of iterations is usually set to 100.

[0183] Finally, the class marginal probability of each node is calculated as:

[0184] ,

[0185] in: is the label of node i The marginal probability of , which indicates the probability that the node belongs to the leakage area, is a normalization constant that ensures that the sum of the probabilities of all possible labels is 1.

[0186] The leakage area probability map is composed of the leakage category probability of each node, forming a two-dimensional probability distribution, which intuitively shows the location and probability of leakage areas on the tunnel lining surface.

[0187] This step achieves the effective fusion of topological features and entropy features, processes data uncertainty through the probabilistic graph model, and generates a reliable leakage area probability map.

[0188] This step determines the location and leakage rate of the tunnel lining leakage area based on the leakage area probability map.

[0189] Specifically include the following:

[0190] First, the precise boundary of the leakage area is extracted, and the probability is set to Figure 2 The probability threshold is then quantified and an edge detection algorithm is applied to extract the leakage area boundary. In practice, the probability threshold is typically set between 0.7 and 0.85. We recommend a default of 0.75, which can be adjusted based on the stringency of the detection requirements. Edge detection uses a gradient-based method, such as the Canny edge detector, with parameters set as a low threshold of 0.1, a high threshold of 0.3, and a Gaussian smoothing kernel size of 3×3.

[0191] The depth distribution of the leak area is then estimated using a deep regression model based on the point cloud data and the probability map. The depth estimation model fuses the geometric features of the point cloud with the probabilistic features to output a depth map of the leak area. The model uses a multi-layer perceptron architecture, taking as input a concatenated vector of point cloud features and probabilistic features, and outputting a normalized depth value (between 0 and 1). The actual depth is then obtained by multiplying it by the lining thickness.

[0192] Then, a leakage model is established based on the principles of thermodynamics and fluid mechanics. The model parameters are estimated from the observed data to calculate the leakage rate per unit time and per unit area. The leakage rate calculation is based on Darcy's law:

[0193] ,

[0194] in: is the leakage flow rate, the unit is m3 / s, which means the amount of water seepage per unit time. The permeability coefficient is in m / s, which indicates the water permeability of the material. is the head difference, in m, which represents the pressure difference that drives the seepage. is the lining thickness, in m, Permeability coefficient Determined by inversion analysis, typical values ​​range from 10 to 10 m / s, the specific value depends on the tunnel lining material and moisture content.

[0195] Finally, the temporal trends of leakage parameters are analyzed, and short-term and long-term prediction models are developed to set different levels of warning thresholds. These prediction models use time series analysis methods, such as the Autoregressive Integrated Moving Average (ARIMA) model, to predict trends in leakage parameters. Warning thresholds are typically set at three levels: minor leakage (leakage rate <0.1 L / min·m²), moderate leakage (0.1-1 L / min·m²), and severe leakage (>1 L / min·m²). These thresholds are determined based on engineering experience and tunnel safety standards and can be adjusted based on the specific tunnel type and usage requirements.

[0196] This step outputs the location, extent and development trend of tunnel lining leakage, providing a scientific basis for tunnel maintenance and repair.

[0197] The technical solution of the present invention is described in detail below with reference to specific embodiments.

[0198] Example 1: Highway tunnel lining leakage detection.

[0199] This embodiment is applied to leak detection in a highway tunnel lining. The tunnel is approximately 2 kilometers long and the lining is a reinforced concrete structure with a thickness of approximately 40 centimeters. During the inspection, the track system is first installed, and then the detection system of the present invention is deployed for automated testing.

[0200] During the data collection phase, the scanning frequency of the 77GHz millimeter-wave radar was set to 3Hz, the acquisition frequency of the passive infrared thermal imager was also set to 3Hz, and the moving speed was controlled at 1m / s to ensure the continuity and integrity of the data.

[0201] During data preprocessing, high-density millimeter-wave radar data undergoes frequency-dependent sorting, using an adaptive filter with a window size of 15×15 and a filter threshold of 0.2. Non-uniformity and distortion correction are performed on the thermal imager data, using pre-calibrated camera intrinsic parameters.

[0202] During the topological feature extraction phase, the K nearest neighbor parameter in the manifold learning mapping was set to 15, and the regularization parameter λ in the Riemannian metric learning was set to 0.05. During the topological signature matrix construction, the radius parameter of the Vietoris-Rips complex was set to a range of 0.1 to 2.0 meters in increments of 0.1 meters. The parameter α for feature enhancement was set to 2.0, and the threshold τ was set to the median of the eigenvalues.

[0203] In the multiphase spectrum analysis stage, the delay time τ for nonlinear phase space reconstruction was determined to be 10 sampling points through mutual information analysis, and the embedding dimension m was determined to be 4 through false nearest neighbor analysis. In the multiscale entropy feature analysis, the scale factor τ was set to range from 1 to 15, and the sample entropy calculation parameters were set to: m = 2, r = 0.2 × standard deviation.

[0204] During the probability fusion and progressive inference phase, the parameters of the conditional random field model were set as follows: node potential function weights were obtained through variational Bayesian learning, the edge potential function smoothing parameter μ was set to 1.0, the distance decay parameter β was set to 0.3, and the correlation strength parameter λ of the inter-layer potential function was set to 2.0. The termination condition for progressive belief propagation was set to the belief change threshold ε = 0.005 or the maximum number of iterations was 100.

[0205] The test results showed that the system successfully identified 23 leaks, 18 of which were minor, 4 were moderate, and 1 was severe. Compared to manual detection results, the detection accuracy reached 95.7%, with a false positive rate of only 4.3%. For the most severe leak, the system estimated the leakage rate to be 1.5 L / min·m², which was only 7.1% lower than the measured value of 1.4 L / min·m².

[0206] Example 2: Railway tunnel lining leakage detection.

[0207] This embodiment is applied to the lining leakage detection of a railway tunnel. The tunnel is located in a mountainous area and is greatly affected by groundwater. The lining structure is a composite lining with an inner layer of reinforced concrete and an outer layer of sprayed concrete.

[0208] Taking into account the characteristics of the tunnel environment, the system parameters were appropriately adjusted: the scanning frequency of the 77GHz millimeter-wave radar was increased to 4Hz to obtain higher-density point cloud data; the passive infrared thermal imager adopted high-sensitivity mode, and the temperature resolution was increased to 0.03°C to capture more subtle temperature changes.

[0209] During the data preprocessing phase, noise filtering was enhanced to address the humid environment of the tunnel, with the filter threshold adjusted to 0.15. Furthermore, to accommodate the composite lining structure, the layered structure of the lining was considered in the coordinate transformation, introducing a layered registration method.

[0210] During the topological feature extraction phase, due to the complexity of the composite lining, the K-nearest neighbor parameter of the manifold learning mapping was increased to 18 to capture more complex local structures. In the construction of the topological signature matrix, the feature enhancement parameter α was adjusted to 2.2 to highlight key topological features.

[0211] In the multiphase spectrum analysis phase, the time series length was extended to capture more complex dynamic characteristics, taking into account the periodic changes in groundwater. The embedding dimension m was increased to 5. In the multiscale entropy feature analysis, the scale factor τ was expanded to 1 to 20 to cover a wider range of time scales.

[0212] In the probability fusion and progressive reasoning stage, the inter-layer potential function of the conditional random field model was adjusted to the characteristics of the composite lining, the inter-layer correlation strength was enhanced, and the λ parameter was adjusted to 2.5 to better fuse the information of different layers.

[0213] The test results showed that the system maintained high detection performance in this complex environment, successfully identifying 35 leakage areas, including many hidden leaks. Compared with traditional detection methods, the system also discovered 7 additional early-stage leaks that were not identified by traditional methods, providing important reference for tunnel maintenance.

[0214] Example 3: Adaptability test under different environmental conditions.

[0215] This example tests the adaptability of the system under different environmental conditions, including high temperature environment (ambient temperature > 30°C), low temperature environment (ambient temperature < 5°C), and high humidity environment (relative humidity > 90%).

[0216] For high-temperature environments, the system adjusted the infrared thermal imager parameters: the temperature range was expanded to 10-60°C, noise reduction was enhanced, and the filtering parameters were adjusted to a 20×20 window size and a filter threshold of 0.25. The parameters for topological feature extraction and multiphase spectrum analysis remained unchanged, but in the probabilistic fusion stage, the weight distribution of the node potential function was adjusted, and the weight of the entropy feature was increased to account for the changes in thermodynamic properties at high temperatures.

[0217] In low-temperature environments, the system added an infrared compensation mechanism and extended the warm-up time to 15 minutes to ensure stable sensor operation. During data preprocessing, the frequency-variable sorting parameters were adjusted to address the characteristics of point cloud data at low temperatures, improving signal extraction sensitivity. In multi-phase spectrum analysis, the parameters of the multiscale entropy feature analysis were adjusted to account for changes in thermal conductivity at low temperatures, with the r value adjusted from 0.2× the standard deviation to 0.15× the standard deviation.

[0218] In high-humidity environments, the system enhanced millimeter-wave radar signal processing and adjusted dielectric parameters to accommodate these conditions. During data preprocessing, a humidity compensation step was added, acquiring real-time humidity data from an ambient humidity sensor to calibrate the point cloud and thermodynamic data. During the probabilistic fusion phase, the parameters of the conditional random field model were adjusted to account for variations in leakage characteristics in high-humidity environments. This enhanced the smoothing effect of the edge potential function, with the μ parameter adjusted from 1.0 to 1.5.

[0219] Test results show that through parameter adjustment and processing strategy optimization, the system maintains high detection performance under various environmental conditions. Compared with the standard environment, the reduction in detection accuracy is controlled within 5%, demonstrating good environmental adaptability.

[0220] Through the detailed description of the above embodiments, the technical effects of the present invention can be summarized as follows:

[0221] 1. High-precision detection: By integrating infrared thermal imaging technology and millimeter-wave radar technology, combined with topological feature extraction and multi-phase spectrum analysis, high-precision detection of tunnel lining leakage is achieved, with a detection accuracy rate of over 95%, significantly superior to traditional single-sensor methods.

[0222] 2. Enhanced robustness: A topology-aware point cloud processing method is introduced to make the detection results robust to changes in tunnel geometry and adapt to tunnel surfaces of different curvatures and different types of lining structures.

[0223] 3. Quantitative assessment: A leakage rate calculation model was established to achieve a quantitative assessment of the degree of leakage. The leakage rate estimation error was controlled within 10%, providing a scientific basis for tunnel maintenance decisions.

[0224] 4. Environmental adaptability: Through adaptive parameter adjustment and processing strategy optimization, the system maintains good detection performance under different environmental conditions and demonstrates excellent environmental adaptability.

[0225] 5. High degree of automation: The system realizes full process automation from data collection to result output, greatly improving detection efficiency and reducing labor costs.

[0226] The tunnel lining leakage infrared-millimeter wave fusion detection method and system of the present invention solves the key technical bottlenecks in traditional detection methods by innovatively fusing multi-source heterogeneous data and combining advanced mathematical theories and algorithms, providing important technical support for tunnel safety monitoring and maintenance.

[0227] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. Tunnel lining leakage infrared-millimeter wave fusion detection method, characterized by: The following steps are involved: Acquire high-density millimeter-wave radar data and infrared thermal imager data; Performing data preprocessing to spatially align and time-synchronize the high-density millimeter-wave radar data and the infrared thermal imager data; Extract topological features, perform manifold learning mapping on the pre-processed high-density millimeter-wave radar data, and construct a topological signature matrix; Perform multi-phase spectrum analysis, perform nonlinear phase space reconstruction and multi-scale entropy feature analysis on the pre-processed infrared thermal imager data, and obtain the entropy feature map; Performing probability fusion and progressive reasoning, inputting the topological signature matrix and the entropy feature map into a conditional random field model, and generating a leakage area probability map through a progressive belief propagation algorithm; Determining the location and leakage rate of the tunnel lining leakage area based on the leakage area probability map; The progressive belief propagation algorithm includes: Initialize messages based on node potential function; Belief propagation is first performed within each layer, and then belief propagation is performed between layers; Guide the propagation direction based on areas with high confidence; Adaptively terminate propagation based on the magnitude of belief changes; Calculate the class edge probability of each node; The class with the maximum posterior probability is selected as the outcome; Assess confidence and uncertainty in decisions.

2. The infrared-millimeter wave fusion detection method for tunnel lining leakage according to claim 1 is characterized in that: The data preprocessing includes: Performing frequency-varying sorting on the high-density millimeter-wave radar data, extracting one-dimensional signal and two-dimensional signal features, and constructing point cloud data; performing non-uniformity correction and distortion correction on the infrared thermal imager data to generate a standardized thermodynamic data matrix; Achieving spatial alignment of the point cloud data with the standardized thermodynamic data matrix through feature point matching and coordinate transformation; Time synchronization between the point cloud data and the standardized thermodynamic data matrix is ​​achieved through timestamp alignment and interpolation resampling.

3. The infrared-millimeter wave fusion detection method for tunnel lining leakage according to claim 1 is characterized in that: The manifold learning mapping includes: Construct a K-nearest neighbor neighborhood graph, where the value of K ranges from 10 to 20; Calculating geodesic distances of the point cloud surface based on the neighborhood graph; Learning the Riemannian metric tensor based on local geometric properties; A dimensionality reduction mapping is performed under the constraint of maintaining local distance relationships to generate a low-dimensional manifold embedding.

4. The infrared-millimeter wave fusion detection method for tunnel lining leakage according to claim 3 is characterized in that: The construction of the topological signature matrix includes: constructing a Vietoris-Rips complex based on the low-dimensional manifold embedding; Calculate the Betti number under different radius parameters and generate persistence diagrams; Converting the persistence graph into a vector representation; Construct a matrix containing multi-scale topological information; Key topological features are enhanced through eigenvalue decomposition and matrix operations to form a topological signature matrix.

5. The infrared-millimeter wave fusion detection method for tunnel lining leakage according to claim 1 is characterized in that: The nonlinear phase space reconstruction comprises: Determine the optimal delay time through mutual information analysis; Determine embedding dimensions through pseudo-nearest neighbor analysis; Construct phase space trajectories based on delayed coordinates; Identify strange attractor structures in phase space; Compute the trajectory density distribution in phase space and identify the dynamic characteristics caused by leakage.

6. The method for detecting tunnel lining leakage by infrared-millimeter wave fusion according to claim 1, characterized in that: The multi-scale entropy feature analysis includes: Generate multi-scale time series through sliding average; Calculate sample entropy and fuzzy entropy at each scale; Calculate cross entropy between different regions to evaluate entropy differences; Analyze the law of entropy value changing with scale; Map the entropy features back to the spatial domain and construct an entropy feature map; The spatial gradient of entropy value is calculated to highlight the boundary features of leakage area.

7. The method for detecting tunnel lining leakage by infrared-millimeter wave fusion according to claim 1, characterized in that: The conditional random field model includes: The point cloud data points and thermodynamic data points are used as nodes; Establish edge connections based on spatial proximity and feature similarity; Construct a two-layer graph structure consisting of a point cloud layer and a thermodynamic layer; Design node potential functions, edge potential functions and inter-layer potential functions to describe the relationship between features and labels and the interaction between nodes.

8. The infrared-millimeter wave fusion detection method for tunnel lining leakage according to claim 1 is characterized in that: Determining the location and leakage rate of the tunnel lining leakage area includes: Extract the precise boundaries of the leakage area; Estimate the depth distribution of the leakage area; Establish leakage model based on thermodynamics and fluid mechanics principles; Estimate model parameters from observed data; Calculate the leakage per unit time and per unit area; Analyze the time trend of leakage parameters; Build short-term and long-term forecast models; Set warning thresholds at different levels.

9. A tunnel lining leakage infrared-millimeter wave fusion detection system, used to implement the tunnel lining leakage infrared-millimeter wave fusion detection method according to any one of claims 1 to 8, characterized in that: include: A sensor subsystem, including a 77GHz millimeter-wave radar and a passive infrared thermal imager, is used to acquire high-density millimeter-wave radar data and infrared thermal imager data; Mobile platform subsystem, including a track-type mobile chassis and a sensor rack-type mounting structure, used to support the movement and positioning of the sensor; A data processing subsystem, including an embedded computing unit and a high-performance computing server, is used to perform data preprocessing, topological feature extraction, multi-phase spectrum analysis, and probabilistic fusion and inference; The data processing subsystem includes: A data preprocessing module, configured to spatially align and temporally synchronize the high-density millimeter-wave radar data and the infrared thermal imager data; The topological feature extraction module is used to perform manifold learning mapping on the pre-processed high-density millimeter-wave radar data and construct a topological signature matrix; The multi-phase spectrum analysis module is used to perform nonlinear phase space reconstruction and multi-scale entropy feature analysis on the pre-processed infrared thermal imager data to obtain the entropy feature map; A probability fusion and reasoning module, configured to input the topological signature matrix and the entropy feature map into a conditional random field model, and generate a leakage area probability map through a progressive belief propagation algorithm; The result output and evaluation module is used to determine the location and leakage rate of the tunnel lining leakage area based on the leakage area probability map.

Images