Liquid leakage detection method based on laser point cloud

Abstract

The application discloses a liquid leakage detection method based on a laser point cloud, and comprises the following steps: adopting a multi-scale adaptive graph convolution network to encode noise-guided features obtained by splicing a point cloud data noise intensity map and point cloud features, and outputting global features; processing point cloud geometry and attribute features respectively to obtain an enhanced global descriptor and estimate uncertainty, and obtaining an uncertainty prediction result; processing the enhanced global descriptor by a liquid type predictor to output an initial liquid type and a probability; fusing the enhanced global descriptor according to type gating weights determined according to the liquid type probability and adaptive gating weights determined according to the enhanced global descriptor, combining the uncertainty prediction result and the initial liquid type, and obtaining a final liquid type; performing asymptotic detection based on the global features to obtain an initial liquid leakage area segmentation result and a leakage probability, refining the leakage probability by learning a correlation graph, and outputting a final liquid leakage area segmentation result. The application can improve detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of tunnel liquid leakage detection technology, and in particular to a liquid leakage detection method based on laser point clouds. Background Technology

[0002] Liquid leakage in infrastructure such as tunnels, underground utility tunnels, and industrial pipelines seriously threatens structural safety and operational efficiency. With the development of lidar technology, point cloud data has become an effective method for leakage detection. Point cloud data contains both geometric information (spatial coordinates X, Y, Z) and attribute information (reflection intensity I, color RGB). Effectively fusing these two types of information is key to improving detection accuracy.

[0003] Existing point cloud leakage detection methods mainly suffer from the following problems:

[0004] It is sensitive to noise, and its detection accuracy drops significantly in real-world environments. Its performance deteriorates drastically in high-noise environments, and its robustness is insufficient. The fusion of geometric and attribute information is inadequate. Existing methods typically simply splice geometric coordinates and attribute features, failing to fully explore the complementary relationship between the two. It cannot effectively capture leakage features at different scales, and its detection capability for small droplets and large-area diffusion regions is insufficient. It does not adequately consider the specificity of liquid types: different liquid types (water, oil, chemical liquids) have different degrees of dependence on geometric and attribute features, and existing methods use a uniform fusion strategy.

[0005] Therefore, there is an urgent need for an intelligent leakage detection method that can adaptively integrate geometric and property features, quantify and predict uncertainties, and adopt differentiated strategies for different liquid types. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and defects of the prior art and provide a liquid leakage detection method based on laser point clouds. Specifically, it is a point cloud leakage detection and liquid classification method based on uncertainty-guided multi-scale graph networks, which can achieve accurate detection of leakage points and accurate identification of leakage liquid types. It aims to overcome the problems of weak noise resistance, insufficient multi-scale feature extraction, and inability to identify liquid types in existing liquid leakage detection methods.

[0007] A liquid leakage detection method based on laser point clouds includes the following steps:

[0008] A noise estimator is used to estimate the noise intensity of the point cloud data of the area to be detected, resulting in a noise intensity map representing the noise level of each point in the point cloud data. A noise-guided feature encoder is then used to fuse and encode the noise intensity map with the original point cloud features, resulting in a noise-guided feature code.

[0009] Noise-guided feature encoding is input into a multi-scale adaptive graph convolutional network for hierarchical feature extraction and adaptive scale fusion to obtain global features that characterize the spatial structure information of the leakage area.

[0010] Using an uncertain geometry-attribute feature fusion module, geometric features and attribute features in point cloud data are encoded separately. Then, the two encoded features are weighted and pooled using the attention weights corresponding to the global features to obtain a geometric global descriptor and an attribute global descriptor. The geometric global descriptor and the attribute global descriptor are bidirectionally enhanced based on a multi-head attention mechanism, and uncertainty prediction is performed to obtain the enhanced geometric global descriptor and the attribute global descriptor, as well as the uncertainty prediction result of the fusion of geometric and attribute uncertainty prediction.

[0011] The enhanced geometric global descriptor and attribute global descriptor are concatenated and input into the liquid type predictor and dynamic gating network. The liquid type predictor outputs the initial liquid type prediction result and liquid type probability. The type gating weight determined based on the liquid type probability and type gating parameters is fused with the dynamic gating weight determined by the dynamic gating network to obtain the adaptive gating weight. The gating fusion feature obtained by gating the enhanced geometric descriptor and attribute descriptor according to the adaptive gating weight, the uncertainty prediction result, and the initial liquid type prediction result are fused and classified to obtain the final liquid type prediction result.

[0012] The global features are input into the progressive detector, and the leakage area is sequentially subjected to coarse detection, fine detection and boundary-aware refinement detection to obtain the initial liquid leakage area segmentation prediction results and leakage probability.

[0013] The leakage probability and noise intensity map are input into a learnable noise leakage correlation module to refine the leakage probability and output the final liquid leakage area segmentation prediction result.

[0014] The multi-scale adaptive graph convolutional network consists of at least three layers of graph convolutional modules; the input features of the subsequent graph convolutional modules are the output features of the preceding graph convolutional modules.

[0015] Each graph convolution module is based on three preset scales. It determines the neighborhood index of each point in the input features according to the K-nearest neighbor search algorithm. After forming edge features based on each point and its neighbors, it performs edge convolution and then performs max pooling operation. It aggregates along the neighborhood dimension to obtain fused features at different scales. The aggregated features at different scales are weighted and fused based on the scale importance weights before output.

[0016] The three scales are: fine-grained scale, used to capture droplet and edge details; medium-grained scale, used to capture the shape of the leak area; and coarse-grained scale, used to capture liquid diffusion patterns.

[0017] Specifically, the geometric global descriptor and the attribute global descriptor are bidirectionally enhanced based on a multi-head attention mechanism. This includes generating a query vector from the geometric global descriptor, generating a key vector and a value vector from the attribute global descriptor, and obtaining an enhanced geometric global descriptor through multi-head attention computation; and generating a query vector from the attribute global descriptor, generating a key vector and a value vector from the geometric global descriptor, and obtaining an enhanced attribute global descriptor through multi-head attention computation.

[0018] In this study, a multilayer perceptron was used for uncertainty prediction, and the multilayer perceptron employed Softplus activation. The uncertainty prediction results are presented as follows:

[0019] L_uncertainty = w_geo_norm×L_geo + w_intensity_norm×L_intensity;

[0020] Where L_uncertainty represents the uncertainty prediction result, w_geo_norm represents the geometric uncertainty prediction normalization weight, L_geo represents the geometric uncertainty prediction result, w_intensity_norm represents the attribute uncertainty prediction normalization weight, and L_intensity represents the attribute uncertainty prediction result.

[0021] The adaptive gating weights are determined through the following steps:

[0022] Normalize a set of learnable gate parameters corresponding to each liquid type to obtain a type gate matrix; each set of learnable gate parameters includes geometric feature weights and attribute feature weights.

[0023] The type gating matrix is ​​weighted and summed based on the liquid type probability to obtain the type gating weight;

[0024] The enhanced geometry global descriptor and attribute global descriptor are concatenated and input into the liquid type adaptive gating network to output dynamic gating weights.

[0025] The type-based gating weights and dynamic gating weights are weighted and fused to obtain adaptive gating weights.

[0026] Among them, the coarse detection uses a two-layer one-dimensional convolutional network to process global features and output the coarse detection results and their probabilities;

[0027] Fine detection employs a three-layer one-dimensional convolutional network to process the feature map concatenated with global features and coarse detection probabilities, and outputs fine detection results and their probabilities.

[0028] Boundary-aware refined detection employs a two-layer one-dimensional convolutional network to process the concatenated feature map of global features and refined detection probabilities, performing boundary detection to obtain a boundary probability map. The global features and boundary probability map are then concatenated and the boundary enhancement features are extracted and output through a two-layer one-dimensional convolutional network. The global features and boundary enhancement features are then concatenated and a gate value is calculated using one-dimensional convolution and a sigmoid activation function. Based on the gate value and through residual connections, refined features are obtained and output.

[0029] The process involves inputting the leakage probability and noise intensity map into a learnable noise-leakage correlation module to refine the leakage probability, including:

[0030] The noise intensity map and leakage probability are concatenated and input into the noise-leakage correlation module. This module learns the correlation between the noise intensity map and the leakage probability using predefined learnable correlation parameters and outputs a correlation map. Based on the correlation map, the learnable correlation parameters, and the noise intensity map, the leakage probability is refined to determine the refined leakage probability, including:

[0031] P_leak_refined = P_leak×(1 +α×C×σ);

[0032] Where P_leak_refined represents the refining leakage probability, P_leak represents the leakage probability, α represents the learnable correlation parameter, C represents the correlation plot, and σ represents the noise intensity plot.

[0033] The final predicted liquid leakage area segmentation result is obtained by concatenating the refined leakage probability and the non-leakage probability, including:

[0034] S_output = [1-P_leak_refined, P_leak_refined];

[0035] Wherein, S_output represents the final liquid leakage area segmentation prediction result, and P_leak_refined represents the refining leakage probability.

[0036] Among them, the attribute features are the reflection features of the point cloud or the combination of the reflection features and color features of the point cloud.

[0037] The method of this invention employs a deep fusion strategy of geometric and attribute features, innovatively separating and encoding the geometric information (X, Y, Z coordinates) and attribute information (reflection intensity, color) of the point cloud. It achieves bidirectional enhancement through a cross-feature attention mechanism, capturing the spatial morphology of the leak through geometric features and the optical properties of the liquid through attribute features; the two complement each other to improve detection accuracy. Compared to simple stitching methods, mIoU is improved by 5-8%.

[0038] The method of this invention employs an uncertainty-guided adaptive fusion strategy to estimate the prediction uncertainty for geometric features and attribute features respectively, and dynamically adjusts the fusion weights based on the prediction uncertainty, so that branches with low uncertainty receive higher weights, thereby improving the reliability of fusion. At the same time, the uncertainty estimation can also serve as an indicator of the credibility of the detection results, providing a basis for downstream decision-making.

[0039] The method of this invention performs type detection based on a liquid type adaptive strategy. It learns differentiated fusion strategies for different liquid types; for example, water leakage relies more on geometric features (flow morphology), oil leakage relies more on attribute features (reflectivity, color), and mixtures require a balance between the two. Through a type-adaptive gating mechanism, the model can dynamically adjust feature weights according to the liquid type, improving classification accuracy by 3-5%.

[0040] The method of this invention employs multi-scale adaptive feature extraction, performs three-scale (k=10, 20, 40) parallel graph convolution, and independently performs adaptive scale fusion in each layer. It uses fine-grained scale to capture small droplets, medium-scale to capture the main leakage, and coarse-grained scale to capture diffusion patterns. It also uses an attention mechanism to adaptively learn the importance of each scale, which improves performance by 8-12% compared to single-scale methods.

[0041] The method of this invention introduces a boundary-aware refinement mechanism in the process of regional boundary detection. By introducing boundary detection branches and residual gating mechanisms, the segmentation accuracy of leakage area boundaries is significantly improved, and the boundary mIoU is improved by 10-15%, which is of great significance for leakage area estimation and severity assessment.

[0042] The method of this invention models the noise-leakage correlation during area detection, innovatively transforming noise information into an auxiliary signal for detection. Through learnable correlation coefficients, the detection weight of high-noise areas (potential leakage areas) is enhanced, improving robustness in complex environments.

[0043] The method of this invention adopts a progressive detection strategy in region detection. It improves the region segmentation accuracy step by step through a three-stage progressive strategy of coarse detection, fine detection, and boundary awareness refinement. Progressive supervision ensures that each stage is effectively trained, resulting in a 3-5% improvement in mIoU compared to single-stage methods.

[0044] The method of this invention has high computational efficiency, with a model parameter count of approximately 2-3M, a GPU inference speed of ≤30ms / sample (1024 points), and supports real-time detection applications. Attached Figure Description

[0045] Figure 1 This is a network architecture diagram of the liquid leakage detection based on laser point cloud according to the present invention.

[0046] Figure 2 This is a network structure diagram of the noise estimator and the noise-guided encoder of the present invention.

[0047] Figure 3 This is a structural diagram of the graph convolution module of the multi-scale adaptive graph convolutional network of the present invention.

[0048] Figure 4 This is a structural diagram of the multi-scale adaptive graph convolutional network of the present invention.

[0049] Figure 5 This is a schematic diagram of the structural principle of the uncertainty geometric-attribute feature fusion module of the present invention.

[0050] Figure 6 This is a schematic diagram of the adaptive gating-based liquid type prediction network structure of the present invention. Detailed Implementation

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] See Figure 1 As shown in the exemplary embodiment of this application, the liquid leakage detection method based on laser point clouds includes:

[0053] S1. A noise estimator is used to estimate the noise intensity of the point cloud data of the area to be detected, and a noise intensity map representing the noise level of each point in the point cloud data is obtained. A noise-guided feature encoder is used to fuse and encode the noise intensity map with the original point cloud features to obtain the noise-guided feature code.

[0054] S2. Input the noise-guided feature encoding into a multi-scale adaptive graph convolutional network, perform hierarchical feature extraction and adaptive scale fusion, and obtain global features that characterize the spatial structure information of the leakage area.

[0055] S3. Using the uncertain geometry-attribute feature fusion module, the geometric features and attribute features in the point cloud data are encoded separately. Then, the two encoded features are weighted and pooled using the attention weights corresponding to the global features to obtain the geometric global descriptor and the attribute global descriptor. The geometric global descriptor and the attribute global descriptor are bidirectionally enhanced based on the multi-head attention mechanism module, and uncertainty prediction is performed to obtain the enhanced geometric global descriptor and the attribute global descriptor, as well as the uncertainty prediction result of the fusion of geometric and attribute uncertainty prediction.

[0056] S4. The enhanced geometric global descriptor and attribute global descriptor are concatenated and input into the liquid type predictor and dynamic gating network; the liquid type predictor outputs the initial liquid type prediction result and liquid type probability; the type gating weight determined based on the liquid type probability and type gating parameters is fused with the dynamic gating weight determined by the dynamic gating network to obtain the adaptive gating weight; the gating fusion feature obtained by gating the enhanced geometric descriptor and attribute descriptor according to the adaptive gating weight, the uncertainty prediction result, and the initial liquid type prediction result are fused and classified to obtain the final liquid type prediction result;

[0057] S5. Input the global features into the progressive detector, and perform coarse detection, fine detection and boundary-aware refinement detection of the leakage area in sequence to obtain the initial liquid leakage area segmentation prediction results and leakage probability.

[0058] S6. Input the leakage probability and noise intensity map into the learnable noise leakage correlation module to refine the leakage probability and output the final liquid leakage area segmentation prediction result.

[0059] The point cloud data described in this application utilizes a LiDAR scanning system to acquire point cloud data P(x,y,z,I) or P(x,y,z,I,R,G,B) of the area to be detected. This data can be composed of the spatial coordinates x,y,z and reflection intensity I of the point cloud, or it can be composed of the spatial coordinates x,y,z, reflection intensity I and color information R,G,B. This adds true multimodal input (RGB image, infrared thermal image) to the point cloud, further improving detection accuracy. Specifically, during data acquisition, a LiDAR can be used to perform a three-dimensional scan of the area to be detected, recording the spatial coordinates (X,Y,Z) and reflection intensity (I) of each point. An RGB camera can be optionally equipped to simultaneously acquire color information (R,G,B). The obtained point cloud data format is [N,C], where N is the number of points (typically 1024-4096 points / frame) and C is the feature dimension (4 or 7).

[0060] Specifically, in one embodiment, when performing noise estimation on the acquired point cloud data, the noise estimator used is composed of a local geometry analysis module and a multilayer perceptron encoder cascaded together, such as... Figure 2As shown. This noise estimator extracts the X, Y, and Z spatial coordinate information of the point cloud; through the local geometric analysis module, it uses the K-nearest neighbor algorithm (setting the number of nearest neighbors K=20) to perform topological retrieval on the local space, calculates the local distance variance, curvature change, and point density change of each point, and fuses them to generate local geometric features; then, the local geometric features are input to the multilayer perceptron encoder (in specific implementation, a three-layer one-dimensional convolutional network can be used). Among them, the first layer has 64 output channels, a convolutional kernel size of 1, followed by a batch normalization layer and a ReLU activation function; the second layer has 32 output channels, a convolutional kernel size of 1, followed by a batch normalization layer and a ReLU activation function; the third layer has 1 output channel, a convolutional kernel size of 1, followed by a Sigmoid activation function to normalize the output value to the range [0,1]; finally, the noise intensity map σ is output, with a dimension of B×1×N, and the noise intensity σᵢ of each point is ∈ [0,1].

[0061] In this embodiment of the application, the noise-guided feature encoding is obtained by fusing point cloud features with a noise intensity map, such as... Figure 2 As shown, specifically, a noise-guided feature encoder is used. The original point cloud features F (dimension B×C×N) and the noise intensity map σ (dimension B×1×N) are concatenated along the feature dimension to obtain a combined feature with a dimension of B×(C+1)×N. Then, the combined feature is input into the noise-guided feature encoding network structure and processed using a two-layer one-dimensional convolutional network to obtain the noise-guided feature encoding. The first layer has C+1 input channels, 64 output channels, a convolutional kernel size of 1, followed by a batch normalization layer and a LeakyReLU activation function (negative slope 0.2). The second layer has 64 input channels, 128 output channels, a convolutional kernel size of 1, followed by a batch normalization layer and a LeakyReLU activation function (negative slope 0.2). The output noise-guided feature is F_guided with a dimension of B×128×N.

[0062] In some embodiments, the multi-scale adaptive graph convolutional network consists of at least three layers of graph convolutional modules. The input features of the subsequent graph convolutional modules are the output features of the preceding graph convolutional modules. Each layer of the graph convolutional module determines the neighborhood index of each point in the input features based on three preset scales using the K-nearest neighbor search algorithm. After constructing edge features based on each point and its neighbors, edge convolution is performed, followed by max pooling. The features are aggregated along the neighborhood dimension to obtain fused features at different scales. The aggregated features at different scales are weighted and fused based on scale importance weights before output. Among the three scales, the fine-grained scale is used to capture small droplets and edge details, the medium scale is used to capture the shape of the main leakage body, and the coarse-grained scale is used to capture the overall diffusion pattern. Each layer learns the importance weights of each scale through an attention network for adaptive fusion.

[0063] In specific implementation, the multi-scale adaptive graph convolutional network includes a first-layer multi-scale graph convolutional module, a second-layer multi-scale graph convolutional module, and a third-layer multi-scale graph convolutional module. Each multi-scale graph convolutional module includes a KNN-Graph module, a multilayer perceptron (MLP), a max-pooling layer, and an adaptive scale fusion module, as follows: Figure 3 As shown: The input to the first-layer multi-scale graph convolutional module is the noise-guided feature F_guided, and the following operations are performed in parallel at three scales k=10, 20, and 40:

[0064] The KNN-Graph module is used, based on the K-nearest neighbor search algorithm, to search for k nearest neighbors for each point of the input noise-guided feature, thus obtaining the neighborhood index. For a point pᵢ and its neighbor pⱼ, edge features are constructed, defined as Eᵢⱼ=[pⱼ-pᵢ,pᵢ], which is the concatenation of the relative coordinates of the neighbor points and the coordinates of the center point, with dimensions of B×2C×N×k. The edge features are processed by a shared multilayer perceptron (in specific implementation, a 2D convolution with a kernel size of 1×1)*, mapping its input channels 2C to output channels 64, followed by batch normalization layers and LeakyReLU activation function, to output edge convolution features. Max pooling is used to aggregate the edge convolution features along the neighborhood dimension, with an output dimension of B×64×N. In this application, the range of K values ​​can be adjusted across multiple scales, such as [5,15,30] for small scenes or [15,30,60] for large scenes, and is not limited to three scales: k=10, 20, and 40.

[0065] The adaptive scale fusion module concatenates features from three scales along the channel dimension to obtain features of dimension B×192×N, and then learns the importance weights w for each scale through an attention network. s Based on importance weights, multi-scale features are fused to obtain the first-layer multi-scale fused feature F_scale_fused = Σ s w s ×F s Where s∈{k=10, k=20, k=40}. The input to the attention network is a concatenated feature map of three scales K. The first layer of the attention network has 192 input channels and 64 output channels, with a convolution kernel size of 1, followed by a batch normalization layer and a ReLU activation function; the second layer has 64 input channels and 3 output channels, with a convolution kernel size of 1, followed by a Softmax function; the output is a set of scale importance weights. , dimension B×3×N.

[0066] Among them, such as Figure 4As shown, the input of the second-layer multi-scale graph convolution module is the first-layer multi-scale fusion feature, with a dimension of B×64×N; the output is the second-layer multi-scale fusion feature, with a dimension of B×128×N. The input of the third-layer multi-scale graph convolution module is the second-layer multi-scale fusion feature, with a dimension of B×128×N; the output is the third-layer multi-scale fusion feature, with a dimension of B×256×N. After obtaining the three-layer multi-scale fusion features, the three-layer multi-scale fusion features are concatenated in the channel dimension, with a dimension of B×448×N. A global convolution operation is performed through the global convolution module (input channel 448, output channel 512, followed by a batch normalization layer and the LeakyReLU activation function), outputting the global feature F_global (dimension of B×512×N).

[0067] In this application, a dual-branch independent encoding structure is adopted to achieve cross-modal alignment of geometric and attribute feature information. The uncertain geometric-attribute feature fusion module receives the input of the original point cloud data and the attention weights generated by the backbone network. The geometric feature branch extracts the spatial coordinate information of the first three channels, i.e., the X, Y, and Z coordinates of the point cloud, with a dimension of B×3×N. The attribute feature branch independently encodes the attribute information of the fourth channel and beyond, such as reflectance intensity or color information, with a dimension of B×(C-3)×N. Subsequently, the attention weights generated by the backbone network are used to perform weighted pooling on the two encoded features to obtain global descriptors for the geometric and attribute features. During independent encoding, a geometric feature encoder and an attribute feature encoder are used respectively.

[0068] The geometric feature encoder employs a two-layer one-dimensional convolutional structure. The first layer has 3 input channels and 64 output channels; the second layer has 64 input channels and 128 output channels. Each layer is followed by a batch normalization layer and a ReLU activation function. The output geometric feature is F_geo, with dimensions B×128×N. The attribute feature encoder uses the same structure: the first layer has C-3 input channels and 64 output channels; the second layer has 64 input channels and 128 output channels. The output attribute feature is F_intensity, with dimensions B×128×N.

[0069] like Figure 5 As shown, in this application, a feature separator is used to separate point cloud features into geometric features and attribute features, which are then independently encoded using a geometric feature encoder and an attribute feature encoder, respectively. The point-level geometric features and attribute features formed after feature separation and independent encoding are then weighted and pooled using an attention weight pooling layer to obtain the corresponding global descriptors, including:

[0070] First, the point-level weights are normalized to obtain the point-level normalized weights w_norm:

[0071] w_norm = w pt / (Σw pt +ε), where ε is a small constant to prevent division by zero, w pt Point-level weights;

[0072] The point-level geometric features and attribute features are weighted and pooled using point-level normalized weights to obtain the corresponding global descriptor.

[0073] Geometric global descriptor: D_geo = Σᵢ(w_normᵢ × F_geoᵢ), with dimensions B×128, where F_geoᵢ is a point-level geometric feature;

[0074] Global attribute descriptor: D_intensity = Σᵢ(w_normᵢ × F_intensityᵢ), with a dimension of B×128, where F_intensityᵢ is a point-level attribute feature.

[0075] In the embodiments of this application, such as Figure 5 As shown, the bidirectional enhancement module enhances the geometric global descriptor and the attribute global descriptor based on a multi-head attention mechanism. This involves mutual bidirectional enhancement through querying the attribute global descriptor from the geometric global descriptor and vice versa. This includes generating a query vector from the geometric descriptor, generating key and value vectors from the attribute descriptor, and then calculating the enhanced geometric global descriptor through multi-head attention; and generating a query vector from the attribute descriptor, generating key and value vectors from the geometric descriptor, and then calculating the enhanced attribute global descriptor through multi-head attention.

[0076] Specifically, a multi-head attention mechanism (e.g., 4 attention heads, 128 hidden dimensions) is used to achieve bidirectional enhancement of the geometric global descriptor and the attribute global descriptor. When the geometric global descriptor queries the attribute global descriptor, a query vector Q_geo is generated from the geometric global descriptor, and a key vector K_intensity and a value vector V_intensity are generated from the attribute global descriptor. The enhanced geometric global descriptor is obtained through multi-head attention: D_geo_enhanced = D_geo + Attention(Q_geo, K_intensity, V_intensity). When the attribute global descriptor queries the geometric global descriptor: a query vector Q_intensity is generated from the attribute global descriptor, and a key vector K_geo and a value vector V_geo are generated from the geometric global descriptor. The enhanced attribute global descriptor D_intensity_enhanced = D_intensity + Attention(Q_intensity, K_geo, V_geo) is obtained through multi-head attention.

[0077] In the embodiments of this application, such as Figure 5 As shown, based on an uncertainty estimator constructed using a multilayer perceptron (MLP), two branches are employed: geometry and attributes. Uncertainty prediction for geometry and attributes is performed separately based on the enhanced global descriptors of geometry and attributes. Specifically, the MLP uses Softplus activation for geometry and attribute uncertainty prediction. During fusion, the weights are dynamically adjusted based on the geometry and attribute uncertainty predictions, with features exhibiting lower uncertainty predictions receiving higher fusion weights. This includes:

[0078] The geometric uncertainty is estimated by u_geo = Softplus(MLP(D_geo_enhanced)), with a dimension of B×1, where MLP is a multilayer perceptron, and the Softplus function ensures that the output is positive; the attribute uncertainty is estimated by u_intensity = Softplus(MLP(D_intensity_enhanced)), with a dimension of B×1; the geometric uncertainty prediction L_geo, with a dimension of B×6, and the attribute uncertainty prediction L_intensity, with a dimension of B×6, are obtained through fully connected layers.

[0079] Furthermore, in this application, after obtaining the uncertainty predictions of geometry and attributes, a normalized weight is calculated based on the dynamically adjusted fusion weights of the geometry and attribute uncertainty predictions. Then, based on the calculated normalized weights, the two uncertainty predictions are weighted and fused to obtain the uncertainty prediction result, including:

[0080] First, the fusion weights w_geo and w_intensity are dynamically adjusted based on the uncertainty predictions of geometry and attributes; where the lower the uncertainty prediction, the higher the weight.

[0081] w_geo = 1 / (u_geo + ε);

[0082] w_intensity = 1 / (u_intensity + ε);

[0083] After normalizing the weights, we obtain the corresponding normalized weights w_geo_norm and w_intensity_norm for geometry and attributes:

[0084] w_geo_norm=w_geo / (w_geo+w_intensity);

[0085] w_intensity_norm = w_intensity / (w_geo + w_intensity);

[0086] Based on normalized weights, a weighted fusion prediction is performed on the uncertainty predictions of geometry and attributes to obtain the uncertainty prediction result L_uncertainty:

[0087] L_uncertainty = w_geo_norm×L_geo + w_intensity_norm×L_intensity.

[0088] In this application, such as Figure 6 As shown, further, a liquid type predictor is used to concatenate the enhanced global descriptor features of the geometric global descriptor and the attribute global descriptor to obtain a combined feature (dimension B×256), and output the initial liquid type prediction result and the prediction result of the liquid type probability, wherein the liquid type probability is obtained by normalizing the initial liquid type prediction result.

[0089] The liquid type predictor in this application adopts a three-layer fully connected network: the first layer has an input of 256 and an output of 128, followed by ReLU activation and Dropout (dropout rate 0.2); the second layer has an input of 128 and an output of 64, followed by ReLU activation; the third layer has an input of 64 and an output of 6 (corresponding to 6 liquid types); the output is the initial liquid type prediction result L_type (dimension B×6) and the liquid type probability P_type = Softmax(L_type) (dimension B×6).

[0090] like Figure 6As shown, in this application, the adaptive gating weight is obtained by the collaborative processing of the type prior path and the dynamic gating path. The determination of the adaptive gating weight includes the following steps:

[0091] In the type prior path, a set of learnable gating parameters is set for each liquid type; in this embodiment, a total of 6 sets of parameters are set, each containing 2 weight values, corresponding to geometric feature weights and attribute feature weights, respectively. Softmax normalization is performed on each set of gating parameters to obtain a type gating matrix G with a dimension of 6×2; then, the type gating matrix G is weighted and summed row-wise according to the liquid type probability P_type to obtain the type gating weight G_type = P_type×G, with a dimension of B×2.

[0092] In some embodiments, the setting of the type gating matrix G can be guided by prior relationships between liquid type and feature dependencies. For example, clean water (category 0) can rely more on geometric features to characterize its shape and fluidity; salt water (category 1) can give relatively balanced attention to geometric and property features; oily liquids (category 2) can rely more on property features to characterize its reflectivity and color differences; chemical solutions (category 3) can rely more on property features to characterize its color and corrosion-related features; mixtures (category 4) can take into account both geometric and property features; and leak-free liquids (category 5) can be relatively dominated by geometric features. The above prior relationships can serve as initial guidance for the type prior path, and in specific implementations, the number of liquid type categories can be increased or the feature dependencies corresponding to each category can be adjusted according to actual application needs.

[0093] In the dynamic gating path, the combined feature F_comb, obtained by concatenating the enhanced geometric descriptor and the enhanced attribute descriptor, is used as the input to the liquid type dynamic gating network. This network employs a two-layer fully connected structure: the first layer has an input of 256 and an output of 64, followed by a ReLU activation function; the second layer has an input of 64 and an output of 2, followed by Softmax normalization, outputting dynamic gating weights G_dynamic, with a dimension of B×2. The two components of the dynamic gating weights G_dynamic correspond to the geometric feature weights and attribute feature weights of the current sample, respectively.

[0094] In the gating fusion stage, the type gating weight G_type and the dynamic gating weight G_dynamic are weighted and fused according to a preset ratio to obtain the adaptive gating weight G_adaptive:

[0095] G_adaptive=0.6×G_type+0.4×G_dynamic;

[0096] In G_adaptive, the first component is the geometric feature weight, the second component is the attribute feature weight, and the sum of the two components is 1.

[0097] Based on the adaptive gating weights G_adaptive, the enhanced geometric global descriptor D_geo_enhanced and the enhanced attribute global descriptor D_intensity_enhanced are weighted using a gating fusion module to obtain the gating fusion representation F_fused.

[0098] F_fused=G_adaptive[0]×D_geo_enhanced+G_adaptive[1]×D_intensity_enhanced.

[0099] In this embodiment, the gated fusion representation F_fused is input into the fusion classifier, which employs a three-layer fully connected network: the first layer has an input of 256 and an output of 256, followed by batch normalization, ReLU activation, and Dropout (dropout rate 0.3); the second layer has an input of 256 and an output of 128, followed by batch normalization, ReLU activation, and Dropout (dropout rate 0.2); the third layer has an input of 128 and an output of 6, outputting the fusion classification prediction result L_fused.

[0100] Finally, the multi-prediction ensemble is achieved by the final detection classifier as follows:

[0101] L_final = 0.4×L_fused + 0.3×L_uncertainty + 0.3×L_type.

[0102] In this application, the coarse detection uses a two-layer one-dimensional convolutional network to process global features and outputs coarse detection results and their probabilities. The fine detection uses a three-layer one-dimensional convolutional network to process the feature map concatenated with the global features and the coarse detection probabilities and outputs fine detection results and their probabilities. The boundary-aware refined detection includes a boundary detection module, which receives the concatenated feature map of the global features and the fine detection probabilities as input, performs boundary detection using a two-layer one-dimensional convolutional network, and obtains a boundary probability map. It also includes a boundary enhancement module, which concatenates the global features and the boundary probability map and extracts boundary enhancement features through a two-layer one-dimensional convolutional network. It concatenates the global features and the boundary enhancement features and calculates a gate value through one-dimensional convolution and a sigmoid activation function. Based on the gate value and through residual connections, it obtains refined features and outputs them. That is, based on the boundary probability map, the boundary enhancement feature representation is refined and output based on the residual gating mechanism.

[0103] In this application, when using a progressive detector for leakage region segmentation, the specific steps include: In the coarse detection stage, the global feature F_global is input, with dimensions B×512×N; the coarse detection head uses a two-layer one-dimensional convolution: the first layer has 512 input channels and 128 output channels, followed by a batch normalization layer and a ReLU activation function; the second layer has 128 input channels and 2 output channels (corresponding to the two classes of leakage / non-leakage); the output coarse detection result L_coarse (dimension B×2×N) is output, and the coarse detection probability P_coarse is obtained through the Softmax function;

[0104] In the fine detection stage, the input is a feature formed by concatenating global features with coarse detection probabilities, with dimensions B×514×N. The fine detection head uses three layers of one-dimensional convolution: the first layer has 514 input channels and 128 output channels; the second layer has 128 input channels and 64 output channels; the third layer has 64 input channels and 2 output channels. Each layer is followed by a batch normalization layer and a ReLU activation function. The output is the fine detection result L_fine (dimension B×2×N), and the fine detection probability P_fine is obtained through the Softmax function.

[0105] In the boundary-aware refinement stage, the process is divided into two branches: boundary detection and boundary enhancement. The boundary detection branch receives the concatenated features of the global features and the probability map output from the refinement stage as input. It uses a three-layer one-dimensional convolution to extract boundary-aware features. The first layer has 512 input channels and 64 output channels; the second layer has 64 input channels and 32 output channels; and the third layer has 32 input channels and 1 output channel, followed by a Sigmoid activation function. The output boundary probability map M_boundary has dimensions of B×1×N.

[0106] The boundary feature enhancement involves concatenating the global feature with the boundary probability map M_boundary (dimension B×513×N), extracting the enhanced feature through a two-layer one-dimensional convolution. The first layer has 513 input channels and 128 output channels; the second layer has 128 input channels and 512 output channels. The output enhanced feature F_enhanced has a dimension of B×512×N. After concatenating the global feature F_global with the enhanced feature F_enhanced (dimension B×1024×N), the gate value G_gate is calculated through a one-dimensional convolution (input channel 1024, output channel 512) and a sigmoid activation function. The refined feature F_refined = F_global + G_gate × F_enhanced is obtained through residual connections.

[0107] The refined feature F_refined, the fine detection probability P_fine, and the boundary probability map M_boundary are concatenated with dimensions B×515×N. The final detection head then processes the data using a two-layer one-dimensional convolutional module (e.g., the first layer has 515 input channels and 64 output channels, followed by a batch normalization layer and a ReLU activation function; the second layer has 64 input channels and 2 output channels) to output the initial leakage region segmentation result L_refined with dimensions B×2×N, and the leakage probability P_leak is obtained.

[0108] In this application, the leakage probability P_leak and the noise intensity map σ are input into a learnable noise leakage correlation module to refine the leakage probability. This includes: concatenating the noise intensity map and the leakage probability and inputting them into the noise leakage correlation module; wherein, in this embodiment, the leakage probability is obtained by probability normalization from the output result of the final boundary-aware refined detection segmentation of the progressive detector. The noise leakage correlation module learns the correlation between the noise intensity map and the leakage probability through predefined learnable correlation parameters and outputs a correlation map; based on the correlation map, the learnable correlation parameters, and the noise intensity map, the leakage probability is refined to determine the refined leakage probability. Specifically, the learningable noise leakage correlation module process includes: defining a learningable correlation parameter α, with an initial value of 0.3, constrained to the range of [0,1] by the Sigmoid function; using the correlation network, the learningable noise leakage correlation module learns the correlation between the noise map and the leakage probability, enhances the leakage probability of high noise and high correlation regions, and outputs a correlation map C with dimensions B×1×N for further refinement of the leakage probability.

[0109] The learnable noise leakage correlation module takes as input the concatenated features of the noise intensity map σ and the leakage probability P_leak, with dimensions B×2×N. The correlation network employs a three-layer one-dimensional convolution. The first layer has 2 input channels and 32 output channels, followed by a batch normalization layer and a ReLU activation function. The second layer has 32 input channels and 16 output channels, followed by a batch normalization layer and a ReLU activation function. The third layer has 16 input channels and 1 output channel, followed by a Sigmoid activation function. The refined expression for the leakage probability is as follows:

[0110] P_leak_refined = P_leak×(1 +α×C ×σ);

[0111] In particular, the probability of leakage is enhanced in high-noise areas and highly correlated points.

[0112] The output is the result of concatenating the refining leakage probability and the non-leakage probability:

[0113] S_output = [1 - P_leak_refined, P_leak_refined], the dimension is B×2×N.

[0114] In this application, when detecting liquid leakage, the prediction model is trained using a multi-task joint loss function, including main classification loss, branch classification loss, uncertainty regularization loss, and liquid type prediction loss. The uncertainty regularization loss reduces uncertainty at high confidence levels by constraining the model through the product of uncertainty and confidence. Examples of its components include:

[0115] I. Design of Uncertainty-Driven Loss Function:

[0116] (1) Main classification loss, calculate cross-entropy loss for the final fusion prediction;

[0117] Loss_main = CrossEntropy(L_final, y_cls), where y_cls represents the true class label;

[0118] (2) Geometric and attribute branch classification loss, with auxiliary supervision of geometric and attribute branches;

[0119] Loss_geo = CrossEntropy(L_geo, y_cls);

[0120] Loss_intensity = CrossEntropy(L_intensity, y_cls);

[0121] Loss_branch = (Loss_geo + Loss_intensity) / 2;

[0122] (3) Uncertainty regularization loss encourages the model to reduce uncertainty when the prediction is correct;

[0123] Prediction confidence for computational geometry and attribute branches:

[0124] c_geo = max(Softmax(L_geo)),

[0125] c_intensity = max(Softmax(L_intensity));

[0126] Loss_uncertainty_reg = mean(u_geo×c_geo)+mean(u_intensity×c_intensity); High confidence levels should have low uncertainty, and this loss encourages a negative correlation between the two.

[0127] (4) Liquid type prediction loss, assisting in supervising liquid type prediction:

[0128] Loss_type = CrossEntropy(L_type, y_cls);

[0129] II. Constructing the total loss function:

[0130] Loss_total = α1×Loss_main + α2×Loss_branch + α3×Loss_uncertainty_reg + α4×Loss_type;

[0131] Recommended weighting values: α1=1.0, α2=0.3, α3=0.1, α4=0.2;

[0132] During training, the Adam optimizer was used with an initial learning rate of 0.001 and a weight decay of 1×10⁻⁻⁻⁶. 4 The learning rate scheduling uses a cosine annealing strategy, and the learning rate decays according to the following formula: lr_t = lr_min + 0.5 × (lr_max - lr_min) × (1 + cos(πt / T)), where t is the current training epoch, T is the total number of training epochs, and lr_min is the minimum learning rate (1×10⁻⁻¹). 6 Batch size recommended: 16-32; Number of training rounds recommended: 100.

[0133] It should be noted that other attention mechanisms, such as Transformer and LinearAttention, can be used in this application, not limited to the aforementioned multi-head attention mechanism. Different uncertainty estimation methods, such as Monte Carlo Dropout and deep ensemble, can also be used to estimate the uncertainty of geometric features and attribute features.

[0134] The method of this invention was tested in multiple scenarios and achieved excellent results. For example, in a standard detection scenario, a subway tunnel was used to detect two leakage areas: Leakage point 1 was a water leakage, with a segmentation mIoU of 91.5% and a classification confidence score of 0.93; geometric uncertainty prediction was 0.12, attribute uncertainty prediction was 0.18, and the geometric feature weight under adaptive gating was 0.62, while the attribute feature weight was 0.38. Leakage point 2 was an oil-containing liquid leakage, with a segmentation mIoU of 88.7%, a classification confidence score of 0.87, a geometric uncertainty prediction of 0.21, and an attribute uncertainty prediction of 0.09; and the geometric feature weight under adaptive gating was 0.35, while the attribute feature weight was 0.65. This verifies the effectiveness of the adaptive gating based on liquid type: water relies more on geometric features, while oil relies more on attribute features.

[0135] In multi-scale feature extraction, for small droplets (diameter < 5 cm): the weight of the fine-grained scale (k=10) is 0.52.

[0136] For medium-scale (k=20) leakage, the weight is 0.31; for coarse-grained scale (k=40) leakage, the weight is 0.17; the detection IoU is 85.3%. For medium-sized leakage (diameter 5-20cm): the weight is 0.28 for fine-grained scale, 0.48 for medium-scale, and 0.24 for coarse-grained scale; the detection IoU is 92.1%. For large-area diffusion (diameter >20cm): the weight is 0.15 for fine-grained scale, 0.35 for medium-scale, and 0.50 for coarse-grained scale; the detection IoU is 89.8%. This verifies that adaptive scale fusion can dynamically adjust the weights of each scale according to the leakage scale.

[0137] In the unrefined refining test, the boundary IoU was 72.3%, while in the refining test with a boundary, the boundary IoU was 85.6%, resulting in a boundary IoU increase of 13.3%. Regarding the impact of boundary testing on leakage area estimation, the area estimation error (without refining) was 15.2%, while the area estimation error (with refining) was 6.8%.

[0138] In uncertainty estimation, the actual error rate of high uncertainty samples (u>0.3) was 28.5%, and the actual error rate of low uncertainty samples (u<0.1) was 5.2%. The correlation coefficient between uncertainty and error rate was 0.82, which verifies that uncertainty estimation can effectively reflect the reliability of prediction.

[0139] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A liquid leakage detection method based on laser point clouds, characterized in that, Includes the following steps: A noise estimator is used to estimate the noise intensity of the point cloud data of the area to be detected, resulting in a noise intensity map representing the noise level of each point in the point cloud data. A noise-guided feature encoder is then used to fuse and encode the noise intensity map with the original point cloud features, resulting in a noise-guided feature code. Noise-guided feature encoding is input into a multi-scale adaptive graph convolutional network for hierarchical feature extraction and adaptive scale fusion to obtain global features that characterize the spatial structure information of the leakage area. Using an uncertain geometry-attribute feature fusion module, geometric features and attribute features in point cloud data are encoded separately. Then, the two encoded features are weighted and pooled using the attention weights corresponding to the global features to obtain a geometric global descriptor and an attribute global descriptor. The geometric global descriptor and the attribute global descriptor are bidirectionally enhanced based on a multi-head attention mechanism, and uncertainty prediction is performed to obtain the enhanced geometric global descriptor and the attribute global descriptor, as well as the uncertainty prediction result of the fusion of geometric and attribute uncertainty prediction. The enhanced geometric global descriptor and attribute global descriptor are concatenated and input into the liquid type predictor and dynamic gating network. The liquid type predictor outputs the initial liquid type prediction result and liquid type probability. The type gating weight determined based on the liquid type probability and type gating parameters is fused with the dynamic gating weight determined by the dynamic gating network to obtain the adaptive gating weight. The gating fusion feature obtained by gating the enhanced geometric descriptor and attribute descriptor according to the adaptive gating weight, the uncertainty prediction result, and the initial liquid type prediction result are fused and classified to obtain the final liquid type prediction result. The global features are input into the progressive detector, and the leakage area is sequentially subjected to coarse detection, fine detection and boundary-aware refinement detection to obtain the initial liquid leakage area segmentation prediction results and leakage probability. The leakage probability and noise intensity map are input into a learnable noise leakage correlation module to refine the leakage probability and output the final liquid leakage area segmentation prediction result.

2. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, A multi-scale adaptive graph convolutional network consists of at least three graph convolutional modules; the input features of the subsequent graph convolutional modules are the output features of the preceding graph convolutional modules.

3. The liquid leakage detection method based on laser point clouds according to claim 2, characterized in that, Each graph convolution module is based on three scales. It determines the neighborhood index of each point in the input features according to the K-nearest neighbor search algorithm. After forming edge features based on each point and its neighbors, it performs edge convolution and then performs max pooling operation. It aggregates along the neighborhood dimension to obtain fused features at different scales. The aggregated features at different scales are weighted and fused based on the scale importance weights before output. The three scales are: fine-grained scale, used to capture droplet and edge details; medium-grained scale, used to capture the shape of the leakage area; and coarse-grained scale, used to capture liquid diffusion patterns.

4. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, The geometric global descriptor and the attribute global descriptor are bidirectionally enhanced based on a multi-head attention mechanism. This includes generating a query vector from the geometric global descriptor, generating key vectors and value vectors from the attribute global descriptor, and obtaining an enhanced geometric global descriptor through multi-head attention computation; and generating a query vector from the attribute global descriptor, generating key vectors and value vectors from the geometric global descriptor, and obtaining an enhanced attribute global descriptor through multi-head attention computation.

5. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, Uncertainty prediction is performed using a multilayer perceptron with Softplus activation. The uncertainty prediction results are as follows: L_uncertainty = w_geo_norm×L_geo + w_intensity_norm×L_intensity; Where L_uncertainty represents the uncertainty prediction result, w_geo_norm represents the geometric uncertainty prediction normalization weight, L_geo represents the geometric uncertainty prediction result, w_intensity_norm represents the attribute uncertainty prediction normalization weight, and L_intensity represents the attribute uncertainty prediction result.

6. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, The adaptive gating weights are determined through the following steps: Normalize a set of learnable gate parameters corresponding to each liquid type to obtain a type gate matrix; each set of learnable gate parameters includes geometric feature weights and attribute feature weights. The type gating matrix is ​​weighted and summed based on the liquid type probability to obtain the type gating weight; The enhanced geometry global descriptor and attribute global descriptor are concatenated and input into the liquid type adaptive gating network to output dynamic gating weights. The type-based gating weights and dynamic gating weights are weighted and fused to obtain adaptive gating weights.

7. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, The coarse detection uses a two-layer one-dimensional convolutional network to process global features and output the coarse detection results and their probabilities. The fine detection uses a three-layer one-dimensional convolutional network to process the feature map after concatenating the global features and the coarse detection probability, and outputs the fine detection results and their probabilities. Boundary-aware refined detection employs a two-layer one-dimensional convolutional network to process the concatenated feature map of global features and refined detection probabilities, performing boundary detection to obtain a boundary probability map. The global features and boundary probability map are then concatenated and the boundary enhancement features are extracted and output through a two-layer one-dimensional convolutional network. The global features and boundary enhancement features are then concatenated and a gate value is calculated using one-dimensional convolution and a sigmoid activation function. Based on the gate value and through residual connections, refined features are obtained and output.

8. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, The leakage probability and noise intensity map are input into a learnable noise leakage correlation module to perform final refinement of the leakage probability, including: The noise intensity map and leakage probability are concatenated and input into the noise-leakage correlation module. This module learns the correlation between the noise intensity map and the leakage probability using predefined learnable correlation parameters and outputs a correlation map. Based on the correlation map, the learnable correlation parameters, and the noise intensity map, the leakage probability is refined to determine the refined leakage probability, including: P_leak_refined = P_leak×(1 +α×C×σ); Where P_leak_refined represents the refining leakage probability, P_leak represents the leakage probability, α represents the learnable correlation parameter, C represents the correlation plot, and σ represents the noise intensity plot.

9. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, The final liquid leakage area segmentation prediction result is obtained by refining the leakage probability and the non-leakage probability and concatenating them, including: S_output = [1-P_leak_refined, P_leak_refined]; Wherein, S_output represents the final liquid leakage area segmentation prediction result, and P_leak_refined represents the refining leakage probability.

10. The liquid leakage detection method based on laser point clouds according to claim 1, characterized in that, The attribute features are either the reflection features of the point cloud or a combination of the reflection features and color features of the point cloud.

Images