An intelligent method and system for determining the most dangerous sliding surface of a slope
By combining the SlopeNet model and the UNet segmentation neural network with residual convolutional blocks and coordinate attention mechanisms, the problems of low computational efficiency and insufficient accuracy in identifying the most dangerous slip surface of slopes are solved, and efficient and accurate judgment of slope engineering safety assessment is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LIAO NING GONG CHENG JI SHU DA XUE E ER DUO SI YAN JIU YUAN
- Filing Date
- 2026-05-11
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies suffer from low computational efficiency and difficulty in guaranteeing accuracy when identifying the most dangerous slip surfaces on slopes, especially under complex geological conditions where they are difficult to identify quickly and accurately. Furthermore, machine learning methods have shortcomings in sample construction, model generalization ability, and multi-source data fusion.
Using the SlopeNet model, a 7-channel normalized input tensor is constructed. Combined with residual convolutional blocks and coordinate attention mechanisms, a mean squared error loss function with mask and class imbalance weighting is designed to construct a UNet segmentation neural network for end-to-end slip surface determination. The most dangerous slip surface is output by processing the Gaussian heatmap skeleton.
It enables efficient, accurate, and intelligent identification of the most dangerous slip surfaces on slopes, improving the accuracy and speed of slip surface identification and providing a reliable technical means for slope engineering safety assessment.
Smart Images
Figure CN122389631A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent determination of the most dangerous slip surface of a slope, and specifically relates to an intelligent determination method and system for the most dangerous slip surface of a slope. Background Technology
[0002] Accurate identification of the most dangerous slip surface is a core prerequisite for slope stability assessment, disaster early warning, and reinforcement. Traditional slope stability analysis methods mainly rely on limit equilibrium methods, including the Swedish slice method, Bishop's method, and Janbu's method. These methods calculate the safety factor by sliced potential sliding masses and establishing static equilibrium equations. However, traditional methods typically rely on mesh enumeration or manual calculations when searching for the most dangerous slip surface, which is not only computationally inefficient but also prone to missing the true most dangerous slip surface when dealing with complex geological conditions, non-circular slip surfaces, and multiple weak interlayers, making it difficult to guarantee search accuracy. While the finite element strength reduction method can automatically track slip surfaces, its high computational cost makes it difficult to meet the rapid analysis needs of engineering sites. In recent years, heuristic algorithms such as genetic algorithms and particle swarm optimization have been introduced into slip surface search, improving global search capabilities to some extent, but problems such as slow convergence speed, strong hyperparameter dependence, and insufficient generalization ability still exist. With the rapid development of artificial intelligence technology, machine learning methods provide a new technical path for the intelligent determination of the most dangerous slip surface of slopes. By training on a large number of slope condition samples, machine learning models can automatically extract the nonlinear mapping relationship between multi-dimensional features such as slope geometry and physical and mechanical parameters of soil and rock mass and the location of the most dangerous slip surface, thereby achieving rapid prediction and intelligent determination of unknown dangerous slip surfaces on slopes. However, existing research still has significant shortcomings in training sample construction, model generalization ability, multi-source data fusion, and the engineering interpretability of prediction results. A systematic and engineering-practical machine learning intelligent determination method and system for the most dangerous slip surface on slopes has not yet been formed. Therefore, it is necessary to develop an intelligent determination method and system for the most dangerous slip surface on slopes to achieve efficient, accurate, and automated identification of dangerous slip surfaces under complex geological conditions, providing a reliable technical means for slope engineering safety evaluation. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this application proposes an intelligent method and system for determining the most dangerous slip surface of a slope, thereby enabling intelligent determination of the most dangerous slip surface of a slope.
[0004] In a first aspect, the present invention provides an intelligent method for determining the most dangerous slip surface of a slope, comprising: Obtain the data of the slope to be measured; The slope data to be measured is preprocessed by normalization to obtain a standardized tensor; The SlopeNet model, after training, inputs a standardized tensor into the model. The output of the SlopeNet model is processed by a sigmoid activation function to obtain a slip surface probability heatmap. The SlopeNet model includes an encoder, a bottleneck layer, and a decoder. The encoder, bottleneck layer, and decoder are all composed of residual convolutional blocks, which include two layers of convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. A multi-scale skip connection is set between the encoder and decoder. The encoder features input to the skip connection are weighted and enhanced by the coordinate attention mechanism module before being fused with the decoder features. The training process includes: using the standardized tensor after normalization of historical slope data as the input of the SlopeNet model; constructing a Gaussian heatmap supervision label based on the pixel labels of the most dangerous slip surface in the historical slope data; extracting multi-scale spatial features by the encoder, bottleneck layer, and decoder and outputting a single-channel slip surface prediction value; processing the single-channel slip surface prediction value by a sigmoid activation function to obtain a predicted slip surface heatmap; and using a weighted mean square error loss function under slope mask constraints to supervise the training of the SlopeNet model. Gaussian heatmap skeletonization is performed on the slip surface probability heatmap to obtain the most dangerous slip surface curve with a single pixel width; Convert the pixel coordinates of the most dangerous slip surface curve with a single pixel width into actual engineering physical coordinates, and output the most dangerous slip surface.
[0005] The step of normalizing the slope sample to be tested to obtain a standardized tensor includes: The cohesion, internal friction angle and unit weight of the slope sample to be tested were normalized respectively. The normalized results were multiplied element by element with the slope mask to obtain the first channel data, the second channel data and the third channel data. The stepped slope data of the slope sample to be tested is used as the fourth channel data, and the flat plate data of the slope sample to be tested is used as the fifth channel data. The coordinate codes of the two-dimensional grid size of the slope sample to be measured are generated. The column direction coordinates are divided by the grid width to obtain the normalized horizontal coordinate code, which is used as the sixth channel data. The row direction coordinates are divided by the grid height to obtain the normalized vertical coordinate code, which is used as the seventh channel data. The data from the first, second, third, fourth, fifth, sixth, and seventh channels are concatenated along the channel dimension to obtain a standardized tensor.
[0006] The encoder includes a stem module, a first downsampling module, a second downsampling module, and a third downsampling module, which are connected sequentially. A normalized tensor is input into the stem module to obtain initial features. These initial features are then input into the first downsampling module to obtain first downsampling features. The first downsampling features are then input into the second downsampling module to obtain second downsampling features. Finally, the second downsampling features are input into the third downsampling module to obtain third downsampling features. The stem module is a residual convolutional block. The first, second, and third downsampling modules each include a max-pooling layer and a residual convolutional block. The residual convolutional block includes two convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. The initial feature, the first downsampled feature, and the second downsampled feature are used as multi-scale skip connection features input to the decoder, and the third downsampled feature is input to the bottleneck layer. The bottleneck layer includes residual convolutional blocks with dilated convolutions, and a coordinate attention mechanism module is embedded in the residual convolutional blocks. The third downsampled feature is input to the bottleneck layer, and the receptive field of the feature is expanded by dilated convolutions. High-level spatial semantic information is extracted in the low-resolution feature space to obtain the bottleneck feature. The decoder includes a third upsampling module, a second upsampling module, and a first upsampling module, which are connected sequentially from deep features to shallow features. The third upsampling module is skipped to the second downsampling module, the second upsampling module is skipped to the first downsampling module, and the first upsampling module is skipped to the STEM module. Each upsampling module includes a transposed convolutional upsampling layer, a skip-connection feature fusion layer, and a residual convolutional block. The residual convolutional block includes two convolutional normalization units, a residual bypass connection, and a coordinate attention mechanism module. The encoder features input via the skip connection are weighted and enhanced by the coordinate attention mechanism module before fusion. The bottleneck features output from the bottleneck layer are input into the third upsampling module and fused with the second downsampling features to obtain the third upsampling features. The third upsampling features are input into the second upsampling module and fused with the first downsampling features to obtain the second upsampling features. The second upsampling features are input into the first upsampling module and fused with the initial features output from the stem module to obtain the first upsampling features. The first upsampling features are input into the output head to obtain the single-channel slip surface prediction value, which is then processed by the sigmoid activation function to obtain the slip surface probability heatmap.
[0007] The coordinate attention mechanism module includes: Obtain the input feature map; The input feature map is subjected to global average pooling along the height and width directions respectively to obtain height-oriented pooling features and width-oriented pooling features; After transposing the width-direction pooling features, they are concatenated with the height-direction pooling features in the spatial dimension to obtain joint coordinate features; The joint coordinate features are input into a shared 1×1 convolutional layer, a batch normalization layer, and a nonlinear activation function to obtain compressed coordinate embedding features. The coordinate embedding features are split along the height and width directions to obtain the height-direction embedding features and the width-direction embedding features, respectively. The height direction attention weights are obtained by embedding the features into a 1×1 convolutional layer in the height direction and processing them with the sigmoid function; the width direction attention weights are obtained by embedding the features into a 1×1 convolutional layer in the width direction and processing them with the sigmoid function. The height and width attention weights are multiplied element-wise with the input feature map to obtain the output of the coordinate attention mechanism module.
[0008] The weighted mean square error loss function with masking is expressed as follows: ; in, To use a masked weighted mean square error loss function, The Sigmoid activation function is used. i For the SlopeNet model to the first i Output of 1 pixel, y i For the first i The heatmap soft label corresponding to each pixel m i For the first i The entity region mask value of 1 pixel, w + Ω represents the class imbalance weight, 1{·} is the indicator function, which takes the value of 1 when the condition in parentheses is true, and takes the value of 0 otherwise, and Ω is the set of spatial indices of all pixels in the grid. ε To prevent numerical stability terms from being divided by zero.
[0009] The process of obtaining the heatmap soft label includes: The most dangerous images in the historical slope data are binarized to obtain a binary image of the slip surface; Perform an Euclidean distance transformation on the binary image, take a Gaussian decay function on the Euclidean distance transformation result, and generate a heatmap soft label.
[0010] The Euclidean distance transformation result is then subjected to a Gaussian decay function to generate soft labels for the heatmap. The calculation formula is as follows: ; in, For heatmap soft labels, σ The spatial diffusion width parameter of the Gaussian kernel. x This represents the distance value corresponding to the current pixel in the Euclidean distance transformation result. It is a Gaussian decay function.
[0011] The process of obtaining the class imbalance weights includes: Only the physical field mean and standard deviation within the slope area from historical slope data are statistically analyzed; Count the number of pixels in positive samples and the number of pixels in negative samples; The class imbalance weight is calculated based on the number of positive and negative sample pixels, and the class imbalance weight is restricted to the range of [5, 50].
[0012] The class imbalance weight is calculated based on the number of positive sample pixels and the number of negative sample pixels, as follows: ; in, p w For class imbalance weights, p x The number of positive sample pixels. n x The number of sample pixels, This is a numerically stable term (to prevent the denominator from being 0).
[0013] Secondly, the present invention also provides an intelligent system for determining the most dangerous slip surface of a slope, comprising: The data acquisition module is used to acquire data of the slope to be measured. The data normalization module is used to perform normalization preprocessing on the slope data to be measured to obtain a standardized tensor; The heatmap output module is used to input the normalized tensor into the trained SlopeNet model. The output of the SlopeNet model is processed by the sigmoid activation function to obtain a slip surface probability heatmap. The SlopeNet model includes an encoder, a bottleneck layer, and a decoder. The encoder, bottleneck layer, and decoder are all composed of residual convolutional blocks. The residual convolutional block includes two layers of convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. A multi-scale skip connection is set between the encoder and the decoder. The encoder features input by the skip connection are weighted and enhanced by the coordinate attention mechanism module before being fused with the decoder features. The training process includes: using the normalized tensor of historical slope data as the input of the SlopeNet model; constructing Gaussian heatmap supervision labels based on the pixel labels of the most dangerous slip surface in the historical slope data; extracting multi-scale spatial features by the encoder, bottleneck layer, and decoder and outputting single-channel slip surface prediction values; processing the single-channel slip surface prediction values by the sigmoid activation function to obtain the predicted slip surface heatmap; and using a weighted mean square error loss function under slope mask constraints for supervised training. The skeleton processing module is used to perform Gaussian heatmap skeleton processing on the slip surface probability heatmap to obtain the most dangerous slip surface curve with a single pixel width. The slip surface output module is used to convert the pixel coordinates of the most dangerous slip surface curve with a single pixel width into actual engineering physical coordinates and output the most dangerous slip surface.
[0014] Beneficial effects: This application proposes an intelligent method and system for determining the most dangerous slip surface on a slope. It constructs a 7-channel standardized input tensor, fusing physical field parameters, geometric features, and spatial location information of the soil and rock mass. Euclidean distance transformation is used to generate soft labels for the slip surface heatmap, overcoming the sparsity problem of traditional binary labels. A mean squared error loss function with entity masking and class imbalance weighting is designed to effectively supervise extremely rare positive samples. A UNet segmentation neural network, SlopeNet, based on an encoder-decoder architecture, is constructed, combined with residual convolutional blocks, coordinate attention mechanisms, and encoder-decoder skip connection feature weighting to enhance the spatial positioning accuracy of slender slip surface structures. After model inference, Gaussian heatmap skeleton post-processing is performed. Through adaptive orientation judgment, weighted quadratic polynomial fitting, and mask constraint discretization, the diffuse heatmap is converged into a single-pixel accurate slip surface skeleton line, outputting the most dangerous slip surface of the slope. This invention achieves end-to-end intelligent determination from slope geological parameter input to the output of the most dangerous slip surface, providing an efficient and accurate technical means for slope engineering safety evaluation. Attached Figure Description
[0015] Figure 1 Flowchart of an intelligent method for determining the most dangerous slip surface on a slope according to an embodiment of the present invention; Figure 2 A schematic diagram of the standardized input tensor of the slope physical field channel in an embodiment of the invention; wherein, (a) is a schematic diagram of the internal friction angle, (b) is a schematic diagram of the cohesion, and (c) is a schematic diagram of the unit weight; Figure 3 A schematic diagram of the input tensor for the slope geometric features in an embodiment of the invention; wherein, (a) is a schematic diagram of the stepped slope surface, and (b) is a schematic diagram of the flat plate; Figure 4 This is a schematic diagram illustrating the principle of generating soft tags for Gaussian distance transformation heatmaps in an embodiment of the invention. Figure 5 A schematic diagram of the network structure of the UNet segmentation model SlopeNet, as described in an embodiment of the invention; Figure 6 This is a schematic diagram of the coordinate attention mechanism in an embodiment of the invention; Figure 7 This is a diagram showing the results of determining the most dangerous slip surface of the slope in an embodiment of the invention. Figure 8 This is a block diagram of an intelligent system for determining the most dangerous slip surface on a slope, as an embodiment of the invention. Detailed Implementation
[0016] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.
[0017] Example 1: This embodiment provides an intelligent method for determining the most dangerous slip surface of a slope, such as... Figure 1 As shown, it includes: Step S1: Obtain the data of the slope to be measured; In this embodiment, the slope data to be measured includes: physical field grid data, slope mask, slip surface annotation map, and geometric bounding box. The physical field grid data includes: an input grid array of shape (H, W, C), containing physical and mechanical parameters of the soil and rock mass (cohesion c, internal friction angle φ, unit weight γ) and geometric features (slope morphology information). H is the number of pixel rows in the slope image, corresponding to the height direction of the slope section; W is the number of pixel columns in the slope image, corresponding to the width direction of the slope section; and C represents the number of channels. The slope mask includes: a solid region mask of shape (H, W), with a value of 0 (air) or 1 (soil and rock mass). The geometric bounding box contains the slope geometric bounding box coordinates (left boundary, lower boundary, right boundary, upper boundary), which are used to calculate the pixel physical resolution from the bounding box physical size and grid width during the inference stage, converting the curve pixel coordinates into actual engineering physical coordinates.
[0018] Step S2: Normalize the slope data to be measured to obtain the standardized tensor, including: Step S2.1: Normalize the cohesion, internal friction angle and unit weight of the slope sample to be tested, and multiply the normalized results element by element with the slope mask to obtain the first channel data, the second channel data and the third channel data. In this embodiment, the first, second, and third channel data represent the physical field: z-score normalization is performed on the first three channels of the physical field mesh data, followed by multiplication by a slope mask to force the pixels in the air region to zero, eliminating invalid value interference. Figure 2 As shown.
[0019] Step S2.2: Use the stepped slope data of the slope sample to be tested as the fourth channel data, and use the flat plate data of the slope sample to be tested as the fifth channel data; In this embodiment, the fourth and fifth channel data are geometric features: the fourth and fifth channels of the physical field mesh data are directly extracted as the geometric features of the slope morphology, such as... Figure 3 As shown.
[0020] Step S2.3: Generate coordinate codes for the two-dimensional grid size of the slope sample to be measured. Divide the column direction coordinates by the grid width to obtain the normalized horizontal coordinate codes, which are used as the sixth channel data. Divide the row direction coordinates by the grid height to obtain the normalized vertical coordinate codes, which are used as the seventh channel data. In this embodiment, the sixth and seventh channel data are coordinate encoded: a normalized pixel coordinate grid is generated, and the column direction coordinate encoding value is... The row direction coordinate encoding value is ,in, i The row index of the current pixel, with a value ranging from 0 to... H 1; j The column index of the current pixel, with a value ranging from 0 to... W 1. Coordinate encoding provides the network with explicit spatial location priors, helping the model establish the correspondence between features and spatial locations.
[0021] Step S2.4: Concatenate the data from the first channel, second channel, third channel, fourth channel, fifth channel, sixth channel, and seventh channel according to the channel dimension to obtain a standardized tensor.
[0022] Step S3: Input the normalized tensor into the trained SlopeNet model. The output of the SlopeNet model is processed by the sigmoid activation function to obtain a slip surface probability heatmap. The SlopeNet model includes an encoder, a bottleneck layer, and a decoder. The encoder, bottleneck layer, and decoder are all composed of residual convolutional blocks. The residual convolutional block includes two layers of convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. A multi-scale skip connection is set between the encoder and the decoder. The encoder features input by the skip connection are weighted and enhanced by the coordinate attention mechanism module before being fused with the decoder features. The training process includes: using the normalized tensor of historical slope data as the input of the SlopeNet model; constructing a Gaussian heatmap supervision label based on the pixel label of the most dangerous slip surface in the historical slope data; extracting multi-scale spatial features by the encoder, bottleneck layer, and decoder and outputting a single-channel slip surface prediction value; processing the single-channel slip surface prediction value by the sigmoid activation function to obtain the predicted slip surface heatmap; and using a weighted mean square error loss function under slope mask constraints to supervise the training of the SlopeNet model. In this embodiment, the historical slope data includes: physical field grid data, slope mask, slip surface annotation map, and geometric bounding box. The physical field grid data includes an input grid array of shape (H, W, C), containing the physical and mechanical parameters of the soil and rock mass (cohesion c, internal friction angle φ, unit weight γ) and geometric features (slope morphology information). H represents the number of pixel rows in the slope image, corresponding to the height direction of the slope section; W represents the number of pixel columns in the slope image, corresponding to the width direction of the slope section; and C represents the number of channels. The slope mask includes a solid region mask of shape (H, W), with values of 0 (air) or 1 (soil and rock mass). The slip surface annotation map includes slip surface annotation data of shape (H, W), with values of 0 or 1, marking the spatial location of the most dangerous slip surface. The geometric bounding box includes the slope geometric bounding box coordinates (x...). min y min x max y max ), where x min and x max Let y represent the coordinates of the left and right boundaries of the slope's geometric extent, respectively. min and y max These represent the lower and upper boundary coordinates of the slope's geometric extent, respectively. During the training phase, only the physical field mean and standard deviation within the slope's physical region in the historical slope data are calculated to normalize the historical slope data, resulting in a standardized tensor of the historical data, which serves as the input to the SlopeNet model.
[0023] Extract the slope width Δx = x from the geometric bounding box. max x minWith height Δy = y max y min The pixel physical resolution is calculated by combining the grid width W and grid height H, which is used to convert the pixel coordinates of the inferred smooth curve into actual engineering physical coordinates; where x max x min These are the coordinates of the right and left boundaries of the bounding box, y and y', respectively. max y min These are the coordinates of the upper and lower boundaries of the geometric bounding box, respectively.
[0024] In this embodiment, the UNet segmentation neural network (U-shaped Convolutional Neural Network) SlopeNet (Slope is taken from a slope, Net is taken from a neural network, collectively referred to as SlopeNet) designed in this invention is composed of the following core components, such as... Figure 5 As shown, this embodiment employs a hierarchical 5-fold cross-validation strategy to divide the training and validation sets. Both the training and validation sets are obtained from historical slope data, and the normalized statistics of each fold of the training set are calculated only from the corresponding training set.
[0025] In this embodiment, the 7-channel input tensor from the training set is input into the UNet segmentation model SlopeNet. Multi-scale spatial features are extracted through an encoder-bottleneck layer-decoder structure, and a single-channel slip surface position prediction value corresponding to the input spatial size is output. The single-channel prediction value is activated by sigmoid to obtain a predicted heatmap, and a masked weighted mean square error loss function is used for supervised training. The loss function is calculated only within the entity region specified by the slope mask, and class imbalance weights are applied to the slip surface neighborhood region with a heatmap response value greater than 0.2. The encoder includes a stem module, a first downsampling module, a second downsampling module, and a third downsampling module, which are connected sequentially. A normalized tensor is input into the stem module to obtain initial features; the initial features are input into the first downsampling module to obtain first downsampling features; the first downsampling features are input into the second downsampling module to obtain second downsampling features; and the second downsampling features are input into the third downsampling module to obtain third downsampling features. The stem module is a residual convolutional block. The first, second, and third downsampling modules each include a max-pooling layer and a residual convolutional block. The residual convolutional block includes two convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. The initial features, the first downsampling features, and the second downsampling features are input into the decoder as multi-scale skip connection features, and the third downsampling features are input into the bottleneck layer. In this embodiment, the encoder consists of a 4-level feature extraction module: the stem module performs initial feature extraction on the 7-channel input and outputs 32 channels; then, the feature map is downsampled by 2 times through three-level downsampling modules down1 (i.e., the first downsampling module), down2 (i.e., the second downsampling module), and down3 (i.e., the third downsampling module), expanding the number of channels to 64, 128, and 256 respectively, extracting multi-scale feature representations from local texture to global structure layer by layer.
[0026] In this embodiment, each convolutional block in the residual convolutional block consists of two 3×3 convolutional layers, GroupNorm normalization, and ReLU activation function, and residual bypass connections are set to alleviate gradient vanishing; Dropout2d regularization layer and coordinate attention mechanism can be configured, and the Dropout2d rate is set to 0.2 in the training example; when the number of input channels is inconsistent with the number of output channels, the residual bypass connections use 1×1 convolutions for channel alignment.
[0027] The bottleneck layer includes a residual convolution block with dilated convolution, in which a coordinate attention mechanism module is embedded; the third downsampled feature is input into the bottleneck layer, the feature receptive field is expanded by dilated convolution, and high-level spatial semantic information is extracted in the low-resolution feature space to obtain the bottleneck feature; In this embodiment, the bottleneck layer uses a dilated convolutional residual block with an inflation rate of 2, which expands the effective receptive field to twice that of ordinary convolution without increasing the number of parameters. This helps the model capture long-range spatial dependencies with large spans at both ends of the smooth surface.
[0028] The decoder includes a third upsampling module, a second upsampling module, and a first upsampling module, which are connected sequentially from deep features to shallow features. The third upsampling module is skipped to the second downsampling module, the second upsampling module is skipped to the first downsampling module, and the first upsampling module is skipped to the STEM module. Each upsampling module includes a transposed convolutional upsampling layer, a skip-connection feature fusion layer, and a residual convolutional block. The residual convolutional block includes two convolutional normalization units, a residual bypass connection, and a coordinate attention mechanism module. The encoder features input via the skip connection are weighted and enhanced by the coordinate attention mechanism module before fusion. The bottleneck features output from the bottleneck layer are input into the third upsampling module and fused with the second downsampling features to obtain the third upsampling features. The third upsampling features are input into the second upsampling module and fused with the first downsampling features to obtain the second upsampling features. The second upsampling features are input into the first upsampling module and fused with the initial features output from the stem module to obtain the first upsampling features. The first upsampling features are input into the output head to obtain the single-channel slip surface prediction value, which is then processed by the sigmoid activation function to obtain the slip surface probability heatmap.
[0029] In this embodiment, the decoder consists of three upsampling modules: a third upsampling module, a second upsampling module, and a first upsampling module. Each stage restores the feature map size to double through transposed convolution. If the upsampled size does not perfectly match the corresponding encoder feature map, bilinear interpolation is used for precise alignment. Then, it is concatenated with the encoder feature map of the corresponding scale along the channel dimension (skip connection) to fuse low-level details and high-level semantics. Finally, the features are refined through residual convolution blocks. The decoder's final output features are mapped to single-channel predicted values through 1×1 convolution, and after sigmoid activation, the probability response of each pixel belonging to the most dangerous slippery neighborhood is obtained.
[0030] The coordinate attention mechanism module includes: Obtain the input feature map; The input feature map is subjected to global average pooling along the height and width directions respectively to obtain height-oriented pooling features and width-oriented pooling features; After transposing the width-direction pooling features, they are concatenated with the height-direction pooling features in the spatial dimension to obtain joint coordinate features; The joint coordinate features are input into a shared 1×1 convolutional layer, a batch normalization layer, and a nonlinear activation function to obtain compressed coordinate embedding features. The coordinate embedding features are split along the height and width directions to obtain the height-direction embedding features and the width-direction embedding features, respectively. The height direction attention weights are obtained by embedding the features into a 1×1 convolutional layer in the height direction and processing them with the sigmoid function; the width direction attention weights are obtained by embedding the features into a 1×1 convolutional layer in the width direction and processing them with the sigmoid function. The height and width attention weights are multiplied element-wise with the input feature map to obtain the output of the coordinate attention mechanism module.
[0031] In this embodiment, as Figure 6 As shown, a coordinate attention mechanism module is added at each residual convolutional block and the skip connection of the decoder. This module first extracts the global context in two directions through adaptive average pooling in the height direction and adaptive average pooling in the width direction, respectively; after transposing the features of the width direction pooling, it concatenates them with the features of the height direction pooling in the spatial dimension, and performs feature fusion through shared 1×1 convolution, BN normalization, and non-linear activation; then it splits the features according to the height and width directions, and generates the height direction attention weights through two independent 1×1 convolutions and a sigmoid function, respectively. a h Attention weights in the width direction a w Finally, it is multiplied element-wise with the original input feature map to achieve direction-aware channel-space joint attention correction, which helps to locate slender arc-shaped smooth surface structures.
[0032] This invention addresses the extreme imbalance between positive and negative samples in slope slip surface detection tasks by designing the following masked weighted mean square error loss function, expressed as follows: ; in, To use a masked weighted mean square error loss function, Use the Sigmoid activation function; i For the model to the first i Output of 1 pixel; y i For the first i The heatmap soft label corresponding to each pixel; m i For the first i The entity region mask value for each pixel, with a value of 0 (air) or 1 (rock and soil). w + Ω represents the class imbalance weight; 1{·} is an indicator function, which takes a value of 1 if the condition in parentheses is true, and a value of 0 otherwise; Ω is the set of spatial indices of all pixels in the grid. ε To prevent numerical stability terms from being divided by zero.
[0033] The process of obtaining the heatmap soft label includes: The most dangerous images in the historical slope data are binarized to obtain a binary image of the slip surface; Perform an Euclidean distance transformation on the binary image, take a Gaussian decay function on the Euclidean distance transformation result, and generate a heatmap soft label.
[0034] The Euclidean distance transformation result is then subjected to a Gaussian decay function to generate soft labels for the heatmap. The calculation formula is as follows: ; in, For heatmap soft labels, σ The spatial diffusion width parameter of the Gaussian kernel. x This represents the distance value corresponding to the current pixel in the Euclidean distance transformation result. It is a Gaussian decay function.
[0035] The process of obtaining the class imbalance weights includes: Only the physical field mean and standard deviation within the slope area from historical slope data are statistically analyzed; Count the number of pixels in positive samples and the number of pixels in negative samples; The class imbalance weight is calculated based on the number of positive and negative sample pixels, and the class imbalance weight is restricted to the range of [5, 50].
[0036] The class imbalance weight is calculated based on the number of positive sample pixels and the number of negative sample pixels, as follows: ; in, p w For class imbalance weights, p x The number of positive sample pixels. n x The number of sample pixels, This is a numerically stable term (to prevent the denominator from being 0). It can be understood that pixels with smooth surfaces are positive samples, and pixels without smooth surfaces are negative samples.
[0037] In this embodiment, pixels with values greater than 0.5 in the smooth surface annotation image are taken as smooth surface pixels; when there are positive sample pixels in the binary smooth surface image, an Euclidean distance transformation is performed to calculate the Euclidean distance from each pixel to the nearest smooth surface pixel; a heatmap is generated based on the Gaussian decay function, such as... Figure 4 As shown.
[0038] To improve the model's generalization ability, the following data augmentation operations are dynamically applied to each sample during the training phase: Horizontal Flip: With a 50% probability, the 7-channel input tensor and its corresponding mask and heatmap label are simultaneously flipped horizontally to simulate the change of left and right slope. Physical field noise: superimposed on physical field channels (channels 1-3) with a 30% probability, following the rules of physical field noise. N Random Gaussian noise in the range of (0, 0.05²) applies only to the area inside the slope mask.
[0039] Step S4: Perform Gaussian heatmap skeletonization on the slip surface probability heatmap to obtain the most dangerous slip surface curve with a single pixel width; The specific process for processing the Gaussian heatmap skeleton is as follows: Step S4.1: Effective Region Extraction: Extract the set of effective pixels with a predicted heatmap probability > 0.5 that are located within the slope area mask. V m Obtain the set of row and column coordinates of valid pixels. y i , x i The corresponding heatmap probability values are used as fitting weights. If the total number of valid pixels is less than 10, it is determined that there is no identifiable slip surface in the current slope section. Step S4.2: Sliding surface direction determination: Calculate the vertical and horizontal spans of the effective pixels. If the vertical span is greater than the horizontal span, the sliding surface is determined to be vertically dominant. (Using row coordinates...) y Independent variable, column coordinates x Fit the dependent variable; otherwise, determine it as a horizontally dominant type and use column axes. x For independent variable, row coordinate y Fit the dependent variable; this judgment ensures that the polynomial fit expands along the main extension direction of the sliding surface, avoiding the vertical sliding surface being truncated by the horizontal regression; Step S4.3: Weighted quadratic polynomial fitting: Using the heatmap probability value as the sample weight, perform polynomial fitting on the effective pixel coordinates.
[0040] In this embodiment, the heatmap probability value is used as the sample weight to perform a weighted quadratic polynomial fitting on the effective pixel coordinates. For vertically dominant sliding surfaces, the row coordinates are used... y Independent variable, column coordinates x As the dependent variable, fit x = a y ² + b y + c; For transversely dominant sliding surfaces, use column coordinates x For independent variable, row coordinate y As the dependent variable, fit y = a x ² + b x + c. Among them... y The row coordinates of the pixels (vertical pixel index, ranging from 0 to H) 1); xThe column coordinates of the pixels (horizontal pixel index, ranging from 0 to W) 1) a, b, and c are the three regression coefficients obtained by weighted least squares quadratic polynomial fitting, where a is the quadratic coefficient, which controls the curvature of the curve; b is the linear coefficient, which controls the overall tilt of the curve; and c is the constant term, which controls the intercept position of the curve.
[0041] Step S4.4: Skeleton Curve Discretization: Generate a uniform discrete point sequence along the independent variable interval according to the effective pixel span, and calculate the corresponding dependent variable coordinates by fitting a polynomial to obtain the continuous coordinate point set of the smooth curve. p x , p y ); Remove coordinate points that exceed the grid boundary range, and further remove coordinate points that are not within the mask in the slope area, finally generating the most dangerous slip surface curve with a single pixel width; Step S5: Convert the pixel coordinates of the most dangerous slip surface curve with a single pixel width into actual engineering physical coordinates, and output the most dangerous slip surface.
[0042] In this embodiment, the most dangerous slip surface determination output is as follows: using a single-pixel curve as the spatial location of the most dangerous slip surface of the slope, combining the slope's geometric bounding box and grid size to calculate the pixel physical resolution, converting the curve's pixel coordinates into actual engineering physical coordinates, and outputting the most dangerous slip surface. The most dangerous slip surface is shown below. Figure 7 As shown.
[0043] Example 2: In this embodiment, an intelligent system for determining the most dangerous slip surface of a slope is also provided, such as... Figure 8 As shown, it includes: The data acquisition module is used to acquire data of the slope to be measured. The data normalization module is used to perform normalization preprocessing on the slope data to be measured to obtain a standardized tensor; The heatmap output module is used to input the normalized tensor into the trained SlopeNet model. The output of the SlopeNet model is processed by the sigmoid activation function to obtain a slip surface probability heatmap. The SlopeNet model includes an encoder, a bottleneck layer, and a decoder. The encoder, bottleneck layer, and decoder are all composed of residual convolutional blocks. The residual convolutional block includes two layers of convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. A multi-scale skip connection is set between the encoder and the decoder. The encoder features input by the skip connection are weighted and enhanced by the coordinate attention mechanism module before being fused with the decoder features. The training process includes: using the normalized tensor of historical slope data as the input of the SlopeNet model; constructing Gaussian heatmap supervision labels based on the pixel labels of the most dangerous slip surface in the historical slope data; extracting multi-scale spatial features by the encoder, bottleneck layer, and decoder and outputting single-channel slip surface prediction values; processing the single-channel slip surface prediction values by the sigmoid activation function to obtain the predicted slip surface heatmap; and using a weighted mean square error loss function under slope mask constraints for supervised training. The skeleton processing module is used to perform Gaussian heatmap skeleton processing on the slip surface probability heatmap to obtain the most dangerous slip surface curve with a single pixel width. The slip surface output module is used to convert the pixel coordinates of the most dangerous slip surface curve with a single pixel width into actual engineering physical coordinates and output the most dangerous slip surface.
[0044] Example 3: This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the intelligent method for determining the most dangerous slip surface of a slope.
[0045] The electronic device can be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements a method for intelligently determining the most dangerous slip surface of a slope as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.
[0046] The processor is used to execute all or part of the steps in the intelligent method for determining the most dangerous slip surface of a slope as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in an electronic device, as well as application-related data.
[0047] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the intelligent method for determining the most dangerous slip surface of a slope as described in the above embodiments.
[0048] Example 4: This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
[0049] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the intelligent method for determining the most dangerous slip surface of a slope as described in the various embodiments of this application.
[0050] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, etc.), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned intelligent method for determining the most dangerous slip surface of a slope.
[0051] Example 5: This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned intelligent method for determining the most dangerous slip surface of a slope.
[0052] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.
[0053] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0054] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of equivalent technology of this disclosure, then the intent of this disclosure also includes such modifications and variations.
Claims
1. A method for intelligently determining the most dangerous slip surface on a slope, characterized in that, include: Obtain the data of the slope to be measured; The slope data to be measured is preprocessed by normalization to obtain a standardized tensor; The SlopeNet model, after training, inputs a standardized tensor into the model. The output of the SlopeNet model is processed by a sigmoid activation function to obtain a slip surface probability heatmap. The SlopeNet model includes an encoder, a bottleneck layer, and a decoder. The encoder, bottleneck layer, and decoder are all composed of residual convolutional blocks, which include two layers of convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. A multi-scale skip connection is set between the encoder and decoder. The encoder features input to the skip connection are weighted and enhanced by the coordinate attention mechanism module before being fused with the decoder features. The training process includes: using the standardized tensor after normalization of historical slope data as the input of the SlopeNet model; constructing a Gaussian heatmap supervision label based on the pixel labels of the most dangerous slip surface in the historical slope data; extracting multi-scale spatial features by the encoder, bottleneck layer, and decoder and outputting a single-channel slip surface prediction value; processing the single-channel slip surface prediction value by a sigmoid activation function to obtain a predicted slip surface heatmap; and using a weighted mean square error loss function under slope mask constraints to supervise the training of the SlopeNet model. Gaussian heatmap skeletonization is performed on the slip surface probability heatmap to obtain the most dangerous slip surface curve with a single pixel width; Convert the pixel coordinates of the most dangerous slip surface curve with a single pixel width into actual engineering physical coordinates, and output the most dangerous slip surface.
2. The intelligent method for determining the most dangerous slip surface of a slope according to claim 1, characterized in that, The step of normalizing the slope sample to be tested to obtain a standardized tensor includes: The cohesion, internal friction angle and unit weight of the slope sample to be tested were normalized respectively. The normalized results were multiplied element by element with the slope mask to obtain the first channel data, the second channel data and the third channel data. The stepped slope data of the slope sample to be tested is used as the fourth channel data, and the flat plate data of the slope sample to be tested is used as the fifth channel data. The coordinate codes of the two-dimensional grid size of the slope sample to be measured are generated. The column direction coordinates are divided by the grid width to obtain the normalized horizontal coordinate code, which is used as the sixth channel data. The row direction coordinates are divided by the grid height to obtain the normalized vertical coordinate code, which is used as the seventh channel data. The data from the first, second, third, fourth, fifth, sixth, and seventh channels are concatenated along the channel dimension to obtain a standardized tensor.
3. The intelligent method for determining the most dangerous slip surface of a slope according to claim 1, characterized in that, The encoder includes a stem module, a first downsampling module, a second downsampling module, and a third downsampling module, which are connected sequentially. A normalized tensor is input into the stem module to obtain initial features; the initial features are input into the first downsampling module to obtain first downsampling features; the first downsampling features are input into the second downsampling module to obtain second downsampling features; and the second downsampling features are input into the third downsampling module to obtain third downsampling features. The stem module represents the residual. The convolutional blocks, including the first, second, and third downsampling modules, all comprise max pooling layers and residual convolutional blocks. Each residual convolutional block includes two convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. The initial features, the first downsampling features, and the second downsampling features are input to the decoder as multi-scale skip connection features. The third downsampling feature is input to the bottleneck layer. The bottleneck layer includes a residual convolutional block with dilated convolutions, and the coordinate attention mechanism module is embedded within this residual convolutional block. Inputting the third downsampling feature into the bottleneck layer yields the bottleneck feature. The decoder includes a third upsampling module, a second upsampling module, and a first upsampling module, which are connected sequentially from deep features to shallow features. The third upsampling module is skipped to the second downsampling module, the second upsampling module is skipped to the first downsampling module, and the first upsampling module is skipped to the STEM module. Each upsampling module includes a transposed convolutional upsampling layer, a skip-connected feature fusion layer, and a residual convolutional block. The residual convolutional block includes two convolutional normalization units, a residual bypass connection, and a coordinate attention mechanism module. The input encoder features are weighted and enhanced by the coordinate attention mechanism module before fusion. The bottleneck features output from the bottleneck layer are input into the third upsampling module and fused with the second downsampling features to obtain the third upsampling features. The third upsampling features are input into the second upsampling module and fused with the first downsampling features to obtain the second upsampling features. The second upsampling features are input into the first upsampling module and fused with the initial features output from the stem module to obtain the first upsampling features. The first upsampling features are input into the output head to obtain the single-channel slip surface prediction value, which is then processed by the sigmoid activation function to obtain the slip surface probability heatmap.
4. The intelligent method for determining the most dangerous slip surface of a slope according to claim 3, characterized in that, The coordinate attention mechanism module includes: Obtain the input feature map; The input feature map is subjected to global average pooling along the height and width directions respectively to obtain height-oriented pooling features and width-oriented pooling features; After transposing the width-direction pooling features, they are concatenated with the height-direction pooling features in the spatial dimension to obtain joint coordinate features; The joint coordinate features are input into a shared 1×1 convolutional layer, a batch normalization layer, and a nonlinear activation function to obtain compressed coordinate embedding features. The coordinate embedding features are split along the height and width directions to obtain the height-direction embedding features and the width-direction embedding features, respectively. The height direction attention weights are obtained by embedding the features into a 1×1 convolutional layer in the height direction and processing them with the sigmoid function; the width direction attention weights are obtained by embedding the features into a 1×1 convolutional layer in the width direction and processing them with the sigmoid function. The height and width attention weights are multiplied element-wise with the input feature map to obtain the output of the coordinate attention mechanism module.
5. The intelligent method for determining the most dangerous slip surface of a slope according to claim 1, characterized in that, The weighted mean square error loss function with masking is expressed as follows: ； in, To use a masked weighted mean square error loss function, For the Sigmoid activation function, i For the SlopeNet model to the first i Output of 1 pixel, y i For the first i The heatmap soft label corresponding to each pixel m i For the first i The entity region mask value of 1 pixel, w + Ω represents the class imbalance weight, 1{·} is the indicator function, which takes the value of 1 when the condition in parentheses is true, and takes the value of 0 otherwise, and Ω is the set of spatial indices of all pixels in the grid. ε To prevent numerical stability terms from being divided by zero.
6. The intelligent method for determining the most dangerous slip surface of a slope according to claim 5, characterized in that, The process of obtaining the heatmap soft label includes: The most dangerous images in the historical slope data are binarized to obtain a binary image of the slip surface; Perform an Euclidean distance transformation on the binary image, take a Gaussian decay function on the Euclidean distance transformation result, and generate a heatmap soft label.
7. The intelligent method for determining the most dangerous slip surface of a slope according to claim 6, characterized in that, The Euclidean distance transformation result is then subjected to a Gaussian decay function to generate soft labels for the heatmap. The calculation formula is as follows: ； in, For heatmap soft labels, σ The spatial diffusion width parameter of the Gaussian kernel. x This represents the distance value corresponding to the current pixel in the Euclidean distance transformation result. It is a Gaussian decay function.
8. The intelligent method for determining the most dangerous slip surface of a slope according to claim 5, characterized in that, The process of obtaining the class imbalance weights includes: Count the number of pixels in positive samples and the number of pixels in negative samples; The class imbalance weight is calculated based on the number of positive and negative sample pixels, and the class imbalance weight is restricted to the range of [5, 50].
9. The intelligent method for determining the most dangerous slip surface of a slope according to claim 5, characterized in that, The class imbalance weight is calculated based on the number of positive sample pixels and the number of negative sample pixels, as follows: ； in, p w For class imbalance weights, p x The number of positive sample pixels. n x The number of sample pixels, It is a numerically stable term.
10. A smart system for determining the most dangerous slip surface of a slope, implemented using the smart method for determining the most dangerous slip surface of a slope as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire data of the slope to be measured. The data normalization module is used to perform normalization preprocessing on the slope data to be measured to obtain a standardized tensor; The heatmap output module is used to input the normalized tensor into the trained SlopeNet model. The output of the SlopeNet model is processed by the sigmoid activation function to obtain a slip surface probability heatmap. The SlopeNet model includes an encoder, a bottleneck layer, and a decoder. The encoder, bottleneck layer, and decoder are all composed of residual convolutional blocks. The residual convolutional block includes two layers of convolutional normalization units, residual bypass connections, and a coordinate attention mechanism module. A multi-scale skip connection is set between the encoder and the decoder. The encoder features input by the skip connection are weighted and enhanced by the coordinate attention mechanism module before being fused with the decoder features. The training process includes: using the normalized tensor of historical slope data as the input of the SlopeNet model; constructing Gaussian heatmap supervision labels based on the pixel labels of the most dangerous slip surface in the historical slope data; extracting multi-scale spatial features by the encoder, bottleneck layer, and decoder and outputting single-channel slip surface prediction values; processing the single-channel slip surface prediction values by the sigmoid activation function to obtain the predicted slip surface heatmap; and using a weighted mean square error loss function under slope mask constraints for supervised training. The skeleton processing module is used to perform Gaussian heatmap skeleton processing on the slip surface probability heatmap to obtain the most dangerous slip surface curve with a single pixel width. The slip surface output module is used to convert the pixel coordinates of the most dangerous slip surface curve with a single pixel width into actual engineering physical coordinates and output the most dangerous slip surface.