Intelligent evaluation method and system for slope stability

Abstract

The invention provides a slope stability intelligent evaluation method and system, and belongs to the technical field of slope stability intelligent evaluation, and the method comprises a form segmentation branch and an index quantification branch, the form segmentation branch is based on an improved ResUNet network, introduces an auxiliary regression module, and synchronously learns the slip plane form and global stability; the index quantification branch directly evaluates the stability coefficient by adopting a full convolutional regression network; in a model training stage, a slope geometric mask is taken as a physical constraint, a multi-task collaborative optimization target is constructed based on an asymmetric Focal-Tversky loss function and a mask weighted binary cross loss function, and intelligent evaluation of slope stability is realized. Accurate segmentation of the slope critical slip plane and reliable evaluation of the stability coefficient are achieved, and therefore the accuracy, reliability and overall efficiency of slope stability evaluation are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent slope stability evaluation technology, specifically relating to an intelligent slope stability evaluation method and system. Background Technology

[0002] Slope stability assessment is a core component of ensuring safe production in open-pit mines and preventing slope geological disasters. Traditional analysis methods rely on limit equilibrium slip surface search and trial calculations or complex numerical simulation iterative solutions, which are computationally time-consuming and difficult to meet real-time requirements, thus failing to meet the needs of rapid and automated engineering evaluation. In recent years, with the development of deep learning technology, data-driven intelligent evaluation methods have shown significant potential in automatic feature extraction and computational acceleration. However, when dealing with slope data involving multi-physics coupling and complex background interference, several key issues remain: Firstly, critical slip surfaces in mesh models typically appear as elongated, low-pixel band-shaped or arc-shaped structures, leading to a severe imbalance between positive and negative samples in semantic segmentation tasks. The model is easily dominated by background pixels, resulting in slip surface omissions, reconstruction breaks, or topological discontinuities, severely affecting the reliability of morphological segmentation. Secondly, most existing studies treat slip surface segmentation and stability coefficient evaluation as two independent stages, lacking a modeling mechanism that can collaboratively utilize local morphological details and global mechanical features within a unified framework, limiting the overall improvement of slip surface reconstruction accuracy and stability evaluation results. Therefore, constructing an intelligent evaluation method for slope stability based on dual-branch collaborative deep learning to achieve intelligent evaluation of slope stability has become a key issue that urgently needs to be addressed in the current intelligent transformation of slope engineering. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this application proposes an intelligent evaluation method and system for slope stability. Intelligent evaluation of slope stability is achieved by constructing a parallel architecture of morphological segmentation branches and index quantification branches.

[0004] In a first aspect, the present invention provides an intelligent evaluation method for slope stability, comprising:

[0005] Acquire the data of the slope to be evaluated, which includes: the geotechnical data of the slope to be evaluated, the step surface of the slope to be evaluated, the characteristics of the flat plate to be evaluated, the mask of the slope to be evaluated, and the geometric dimensions of the slope to be evaluated.

[0006] The geotechnical data of the slope to be evaluated, the slope step surface of the slope to be evaluated, and the features of the plate to be evaluated are transformed to obtain multiple channel features to be evaluated. The multiple channel features to be evaluated are used to form the tensor of the multiple channel input to be evaluated.

[0007] The features of multiple channels to be evaluated in the tensor of the input to be evaluated are input into the morphological segmentation branch to obtain the Logits feature map to be evaluated. The morphological segmentation branch includes a pre-trained improved ResUNet network and an auxiliary regression module. The features of multiple channels to be evaluated are input into the pre-trained improved ResUNet network to obtain the Logits feature map to be evaluated that represents the distribution of the critical slip surface.

[0008] The Logits feature map to be evaluated is processed by a Sigmoid nonlinear activation mapping to obtain the slip surface probability distribution field.

[0009] Perform a Hadamard volume physical purification operation on the probability distribution field of the slope mask and slip surface to be evaluated, filter out the background noise interference outside the mask, and obtain the purified probability distribution field.

[0010] Adaptive threshold binarization is performed on the purified probability distribution field to obtain the final critical slip surface geometry.

[0011] The features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated are input into the index quantization branch to obtain the first principal stability coefficient. The index quantization branch includes: a pre-trained fully convolutional network, a global average pooling layer and a multilayer perceptron. The pre-trained fully convolutional network performs continuous downsampling on the tensor of the multi-channel input to be evaluated to extract the first global spatial features. The first global spatial features are compressed into a one-dimensional feature vector to be evaluated through the global average pooling layer. According to the geometric dimensions of the slope to be evaluated, the first geometric bounding box is calculated. The first scale feature vector is obtained according to the first geometric bounding box. The one-dimensional feature vector to be evaluated is concatenated with the first scale feature vector. The concatenated fused features are input into the multilayer perceptron to obtain the first principal stability coefficient.

[0012] The first principal stability coefficient and the final critical slip surface geometry are used as the results of slope stability evaluation.

[0013] The pre-trained improved ResUNet network training process includes:

[0014] Historical slope data is acquired, including: historical slope geotechnical data, historical slope step surface, historical flat plate characteristics, historical slope mask, and historical slope geometric dimensions; the historical slope data is randomly divided into training set and validation set.

[0015] The historical slope geotechnical data, historical slope step surface and historical flat plate features in the training set are transformed to obtain historical multi-channel features. The historical multi-channel features and historical slope mask are used to form a tensor for historical multi-channel input.

[0016] By inputting historical features from multiple channels into the improved ResUNet network, a historical Logits feature map representing the distribution of the critical slip surface is obtained.

[0017] Simultaneously, the encoder bottleneck layer features and decoder intermediate layer features are extracted from the improved ResUNet network. The encoder bottleneck layer features and decoder intermediate layer features are then processed by global average pooling. The pooled features are then concatenated and fused with the normalized geometric scale feature vector. The fused result is then input into the auxiliary regression module to obtain the auxiliary stability coefficient.

[0018] The total loss function is constructed based on historical Logits feature maps, historical slope masks, and auxiliary stability coefficients. The parameters of the morphological segmentation branch are optimized and trained through gradient backpropagation.

[0019] During training, the validation set is used to evaluate the segmentation accuracy of the morphological segmentation branch and the regression error of the auxiliary stability coefficient. The network parameters corresponding to the minimum total loss function of the validation set are selected as the pre-trained improved ResUNet network.

[0020] The improved ResUNet network includes: convolutional layers, intermediate activation layers, group normalization layers, encoders, decoders, bottleneck blocks, and splicing blocks. The convolutional layers are connected to the intermediate activation layers, the intermediate activation layers are connected to the group normalization layers, the group normalization layers are connected to the encoders, the encoders and decoders are connected by skip connections, the encoders are connected to the bottleneck blocks, the bottleneck blocks and decoders are connected to the splicing blocks respectively, and the splicing blocks are connected to the auxiliary regression module.

[0021] The features of multiple channels to be evaluated are input into the convolutional layer to obtain the output of the convolutional layer. The output of the convolutional layer is input into the intermediate activation layer to obtain the pre-normalized features. The pre-normalized features are input into the group normalization layer to obtain the normalized output.

[0022] Normalization is encoded by an encoder, and the output of the encoder is fed into the decoder through a jump connection. The output of the encoder is the Logits feature map to be evaluated.

[0023] The output of the decoder is used as the input of the bottleneck block. The output of the bottleneck block and the output of the encoder are globally pooled separately. The solution of the global pooling is then concatenated through the concatenation block.

[0024] The spliced ​​results are input into the auxiliary regression module to obtain the auxiliary stability coefficient.

[0025] The training process of the pre-trained fully convolutional network includes:

[0026] The characteristics of multiple historical channels are input into the index quantization branch to obtain the second principal stability coefficient; the mean squared error loss function between the second principal stability coefficient and the true stability coefficient is calculated, and the gradient backpropagation algorithm is used to optimize the parameters.

[0027] During training, the regression mean square error of the branches is quantified using the validation set evaluation index. The network parameters corresponding to the minimum validation set error are selected as the pre-trained fully convolutional network.

[0028] The process involves transforming the geotechnical data of the slope to be evaluated, the slope step surface, and the features of the flat plate to be evaluated to obtain multiple channel features to be evaluated. These multiple channel features, along with the slope mask, constitute a tensor for the input of the multiple channels to be evaluated. This includes:

[0029] The geotechnical data of the slope to be evaluated are standardized to obtain normalized physical field data;

[0030] Based on the features of the slope step surface and the flat plate to be evaluated, an initial geometric shape matrix is ​​extracted. The initial geometric shape matrix is ​​then transposed and numerically cleaned to construct the slope geometric field data.

[0031] Based on the spatial resolution of the slope geometric field data, a two-dimensional index grid is generated, and the grid coordinates are normalized to construct coordinate field data.

[0032] Based on the geometric dimensions of the slope to be evaluated, calculate the first geometric bounding box that can completely cover the slope area;

[0033] Based on the physical width and height of the first geometric bounding box, the first scale feature vector is calculated and normalized.

[0034] Based on the geometric dimensions of the slope to be evaluated, spatial rasterization is performed to generate a slope mask that identifies the effective area of ​​the slope.

[0035] The normalized physical field data, slope geometric field data, coordinate field data, and slope mask that identifies the effective area of ​​the slope are used to form the tensor of the multi-channel input to be evaluated.

[0036] The method of segmenting the feature input morphology of multiple channels in the tensor to be evaluated includes: normalized physical field data, slope geometric field data, and coordinate field data.

[0037] The method for quantizing the input index of multiple channel features in the tensor of the multi-channel input to be evaluated includes: normalized physical field data, slope geometric field data, coordinate field data, and the slope mask to be evaluated.

[0038] The improved ResUNet network consists of convolutional layers and group normalization layers. A loss function is constructed based on Logits feature maps, slope masks, and auxiliary stability coefficients. The morphology segmentation branch is trained and its parameters optimized through gradient backpropagation, including:

[0039] Based on the pixel distribution inside the slope mask in the training set, the global positive sample weights are statistically calculated.

[0040] Based on the true critical slip surface, we define true positives, false positives, and false negatives within the slope mask area and construct an asymmetric Focal-Tversky loss function.

[0041] Based on the slope mask and Logits feature map, a mask-weighted binary cross loss function is constructed;

[0042] We construct a segmentation loss function by weighted summation of the asymmetric Focal-Tversky loss function and the mask-weighted binary cross loss function.

[0043] Based on the auxiliary stability coefficient and the true stability coefficient, a regression loss function is constructed using mean squared error;

[0044] Based on the segmentation loss function and the regression loss function, the total loss function is obtained;

[0045] Based on the training samples and the total loss function, the ResUNet network is trained and optimized using gradient backpropagation to obtain the trained and optimized morphological segmentation model.

[0046] The step of calculating the mean squared error loss function between the second principal stability coefficient and the true stability coefficient, and then using the gradient backpropagation algorithm for parameter optimization training, includes:

[0047] High-dimensional global features are extracted using a fully convolutional network, and after global average pooling and scale vector concatenation, the second principal stability coefficient is obtained by regression.

[0048] Based on the second principal stability coefficient and the true stability coefficient, construct the mean squared error loss function;

[0049] Based on the training samples and the mean squared error loss function, the parameters of the fully convolutional network are optimized and trained using the gradient backpropagation algorithm to obtain the optimized index quantization model.

[0050] Secondly, the present invention also provides an intelligent evaluation system for slope stability, comprising:

[0051] The data acquisition module is used to acquire the data of the slope to be evaluated, which includes: the geotechnical mechanics data of the slope to be evaluated, the step surface of the slope to be evaluated, the characteristics of the flat plate to be evaluated, the mask of the slope to be evaluated, and the geometric dimensions of the slope to be evaluated.

[0052] The tensor construction module is used to transform the geotechnical mechanics data of the slope to be evaluated, the step surface of the slope to be evaluated, and the features of the flat plate to be evaluated, to obtain multiple channel features to be evaluated. The multiple channel features to be evaluated and the slope mask to be evaluated constitute the tensor of the multiple channel input to be evaluated.

[0053] The morphological segmentation module is used to input the features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated into the morphological segmentation branch to obtain the Logits feature map to be evaluated. The morphological segmentation branch includes a pre-trained improved ResUNet network and an auxiliary regression module, which inputs the features of multiple channels to be evaluated into the pre-trained improved ResUNet network to obtain the Logits feature map to be evaluated that represents the distribution of the critical slip surface.

[0054] The critical slip surface acquisition module is used to obtain the slip surface probability distribution field by passing the Logits feature map to be evaluated through the Sigmoid nonlinear activation mapping; to perform Hadamard product physical purification operation on the slope mask to be evaluated and the slip surface probability distribution field to obtain the purified probability distribution field; and to perform adaptive threshold binarization processing on the purified probability distribution field to obtain the final critical slip surface geometry.

[0055] The index quantization module is used to input the features of multiple channels to be evaluated from the tensor of the multi-channel input to be evaluated into the index quantization branch to obtain the first principal stability coefficient. The index quantization branch includes: a pre-trained fully convolutional network, a global average pooling layer, and a multilayer perceptron. The pre-trained fully convolutional network performs continuous downsampling on the tensor of the multi-channel input to be evaluated to extract the first global spatial features. The first global spatial features are compressed into a one-dimensional feature vector to be evaluated by the global average pooling layer. According to the geometric dimensions of the slope to be evaluated, the first geometric bounding box is calculated. The first scale feature vector is obtained according to the first geometric bounding box. The one-dimensional feature vector to be evaluated is concatenated with the first scale feature vector. The concatenated fused features are input into the multilayer perceptron to obtain the first principal stability coefficient.

[0056] The results output module is used to take the first principal stability coefficient and the final critical slip surface geometry as the slope stability evaluation results.

[0057] Thirdly, this application proposes an electronic device, comprising: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to execute the aforementioned intelligent slope stability evaluation method.

[0058] Fourthly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned intelligent evaluation method for slope stability.

[0059] Fifthly, this application proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned intelligent evaluation method for slope stability.

[0060] Beneficial effects:

[0061] This application proposes an intelligent evaluation method and system for slope stability. Based on the collaborative architecture of a dual-branch neural network, the method achieves accurate segmentation of the critical slip surface of the slope and reliable evaluation of the stability coefficient, thereby significantly improving the accuracy, reliability and overall efficiency of slope stability evaluation. Attached Figure Description

[0062] Figure 1 Flowchart of an intelligent slope stability evaluation method according to an embodiment of the present invention;

[0063] Figure 2 Thermal diagram of internal friction angle according to an embodiment of the present invention;

[0064] Figure 3 Cohesion thermogram of an embodiment of the present invention;

[0065] Figure 4 Thermodynamic diagram of bulk density in an embodiment of the present invention;

[0066] Figure 5 Binary mask image of the stepped slope surface according to an embodiment of the present invention;

[0067] Figure 6 Flat binary mask diagram of an embodiment of the present invention;

[0068] Figure 7 Plots showing the true critical slip surface and minimum stability coefficient in embodiments of the present invention;

[0069] Figure 8 A schematic diagram of a slope mask as a spatial constraint condition in an embodiment of the present invention;

[0070] Figure 9 Flowchart of convolutional layers and group normalization in an embodiment of the present invention;

[0071] Figure 10 A structural diagram of the morphological segmentation branch in an embodiment of the present invention;

[0072] Figure 11 The slope stability evaluation results of the embodiments of the present invention;

[0073] Figure 12 A schematic diagram of the principle of an intelligent slope stability evaluation system according to an embodiment of the present invention. Detailed Implementation

[0074] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0075] This application proposes an intelligent evaluation method and system for slope stability. The method constructs a dual-branch collaborative deep learning framework, including a morphology segmentation branch and an index quantization branch. The morphology segmentation branch, based on an improved ResUNet architecture, introduces an auxiliary regression module to simultaneously learn the slip surface morphology and global stability. The index quantization branch employs a fully convolutional regression network to directly evaluate the stability coefficient. During model training, a slope geometric mask is used as the physical constraint, and asymmetric loss based on the Tversky exponent and smooth regression loss are employed to construct a multi-task collaborative optimization objective. Ultimately, the critical slip surface morphology and slope stability coefficient are simultaneously output, achieving intelligent evaluation of slope stability. This provides a reliable intelligent tool for slope engineering stability evaluation and has significant theoretical and practical value for promoting the digital transformation and intelligent evaluation of slope engineering.

[0076] Example 1:

[0077] This embodiment provides an intelligent evaluation method for slope stability, such as... Figure 1 As shown, it includes:

[0078] Step S1: Obtain the data of the slope to be evaluated, which includes: the geotechnical data of the slope to be evaluated, the step surface of the slope to be evaluated, the characteristics of the flat plate to be evaluated, the mask of the slope to be evaluated, and the geometric dimensions of the slope to be evaluated.

[0079] In this embodiment, the geotechnical parameters of the slope to be evaluated are obtained, such as... Figure 2 , Figure 3 , Figure 4 As shown, this includes cohesion c, internal friction angle φ, unit weight γ, the slope surface of the step to be evaluated, and the characteristics of the flat plate to be evaluated (such as...). Figure 5 , Figure 6 (As shown) The true critical slip surface and the true minimum stability coefficient (as shown) Figure 7 (As shown).

[0080] It is understandable that the evaluation process uses data of the slope to be evaluated, while the model training process uses historical slope data. The only difference between the two is the time at which the slope data was acquired; the data types and structures are completely identical. It is worth noting that the training set data needs to have noise added, while the evaluation data and validation set data do not have noise added.

[0081] Step S2: Transform the geotechnical data of the slope to be evaluated, the slope step surface of the slope to be evaluated, and the features of the flat plate to be evaluated to obtain multiple channel features to be evaluated. Combine the multiple channel features to be evaluated and the slope mask to be evaluated to form a tensor for the input of the multiple channels to be evaluated, including:

[0082] Step S2.1: Standardize the geotechnical data of the slope to be evaluated to obtain normalized physical field data.

[0083] Step S2.2: Based on the features of the slope step surface and the flat plate to be evaluated, extract the initial geometric shape matrix, perform dimension transpose and numerical cleaning on the initial geometric shape matrix, and construct the slope geometric field data;

[0084] In this embodiment, the original input data is assumed to be a tensor. In this model, channels 1, 2, and 3 represent cohesion c, internal friction angle φ, and unit weight γ, respectively, while channels 4 and 5 represent the features of the stepped slope and the flat plate, respectively. Geometric morphology matrix is ​​extracted. Then, dimensional transpose and data cleaning are performed to obtain geometric field data. :

[0085] ;

[0086] Where H is the height of the tensor and W is the width of the tensor. For any pixel in the image Geometric field data for the c-th channel, For any pixel in the image The tensor of the (c+3)th channel.

[0087] Step S2.3: Based on the spatial resolution of the slope geometric field data, generate a two-dimensional index grid, and normalize the grid coordinates to construct coordinate field data;

[0088] In this embodiment, for any pixel in the image Its normalized coordinate field data The calculation is as follows:

[0089] ;

[0090] ;

[0091] in, For any pixel Normalize the x-axis. For any pixel Normalized ordinate, .

[0092] Step S2.4: Calculate the first geometric bounding box that can completely cover the slope area based on the geometric dimensions of the slope to be evaluated;

[0093] In this embodiment, S is set as the set of pixel coordinates of the effective geometric region of the slope to be evaluated (i.e., the geometric dimensions of the slope to be evaluated), and the first geometric bounding box B is:

[0094] ;

[0095] ;

[0096] ;

[0097] Width of the first geometric bounding box With height They are respectively:

[0098] ;

[0099] ;

[0100] Where B is the first geometric bounding box. The minimum lateral value of the slope to be evaluated. This represents the maximum lateral value of the slope to be evaluated. The minimum longitudinal value of the slope to be evaluated. This represents the maximum longitudinal value of the slope to be evaluated. Let be the width of the first geometric bounding box. The height of the first geometric bounding box.

[0101] Step S2.5: Calculate and normalize the first scale feature vector based on the physical width and height of the first geometric bounding box;

[0102] In this embodiment, the maximum value of the width and height of the bounding box is selected as the normalization factor. ;

[0103] According to the normalization factor The first-scale eigenvector is calculated using the following formula:

[0104] ;

[0105] in, The first-scale feature vector, This is the normalization factor.

[0106] Step S2.6: Perform spatial rasterization processing based on the geometric dimensions of the slope to be evaluated to generate a slope mask that identifies the effective area of ​​the slope (e.g., ...). Figure 8 (as shown)

[0107] In this embodiment, the geometric bounding box is converted into a slope physical domain. Rasterization processing is performed based on the slope physical domain to generate a slope mask that identifies the effective area of ​​the slope. The calculation formula is as follows:

[0108] ;

[0109] in, In discrete pixel coordinate system The lower slope area is covered by a mask. This refers to the physical domain of the slope.

[0110] Step S2.7: Combine the normalized physical field data, slope geometric field data, coordinate field data, and slope mask that identifies the effective area of ​​the slope to form the tensor of the multi-channel input to be evaluated.

[0111] It is understandable that steps S2.1 to S2.7 are also applicable to constructing tensors for historical multi-channel inputs, except that all input data are historical slope data. In addition, Gaussian noise needs to be added to the historical slope data input.

[0112] If the constructed historical multi-channel input tensor requires input morphological segmentation branches, then Gaussian noise is additionally added to the normalized physical field data, including:

[0113] Step S2.1.1: Generate a Gaussian noise matrix that conforms to a standard normal distribution;

[0114] Step S2.1.2: Spatial filtering of the Gaussian noise matrix is ​​performed using a historical slope mask, retaining only the noise components inside the geometric boundaries of the slope;

[0115] Step S2.1.3: Superimpose the normalized physical field data with the filtered noise component to obtain normalized physical field data with superimposed Gaussian noise.

[0116] In this embodiment, a historical slope mask is used. As a spatial constraint, random noise injection is performed during the training phase. First, a Gaussian noise matrix conforming to a standard normal distribution is generated, and this Gaussian noise matrix is ​​spatially filtered using a historical slope mask, retaining only the noise components inside the slope's geometric boundaries. Subsequently, the processed noise is superimposed onto a normalized physical parameter field to generate an enhanced physical field (i.e., normalized physical field data superimposed with Gaussian noise).

[0117] For physical fields For each channel, the generated standard Gaussian noise matrix is: Enhanced physical field The calculation formula is:

[0118] ;

[0119] in, For physical fields The c-th physical channel data, To superimpose Gaussian noise onto the normalized physical field data of the c-th physical channel, This is the Hadamard product, which is the product of elements one by one. Noise intensity coefficient; This indicates that the noise superposition operation is triggered randomly with a 50% probability.

[0120] Step S3: Input the features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated into the morphological segmentation branch to obtain the Logits feature map to be evaluated. The morphological segmentation branch includes a pre-trained improved ResUNet network and an auxiliary regression module. Input the features of multiple channels to be evaluated into the pre-trained improved ResUNet network to obtain the Logits feature map to be evaluated that represents the distribution of the critical slip surface.

[0121] In this embodiment, the ResUNet network is used as the backbone architecture to construct the morphological segmentation branches, such as... Figure 9 , Figure 10 As shown, a design combining convolutional layers and group normalization layers is adopted, including: convolutional layers, intermediate activation layers, group normalization layers, encoders, decoders, bottleneck blocks, and splicing blocks. The convolutional layers are connected to the intermediate activation layers, the intermediate activation layers are connected to the group normalization layers, the group normalization layers are connected to the encoders, the encoders and decoders are connected by skip connections, the encoders are connected to the bottleneck blocks, the bottleneck blocks and decoders are connected to the splicing blocks respectively, and the splicing blocks are connected to the auxiliary regression module.

[0122] The features of multiple channels to be evaluated are input into the convolutional layer to obtain the output of the convolutional layer. The output of the convolutional layer is input into the intermediate activation layer to obtain the pre-normalized features. The pre-normalized features are input into the group normalization layer to obtain the normalized output.

[0123] The normalization is encoded by an encoder, and the output of the encoder is connected to the input of the decoder via a jump connection.

[0124] The output of the decoder is used as the input of the bottleneck block. The output of the bottleneck block and the output of the encoder are globally pooled separately. The solution of the global pooling is then concatenated through the concatenation block.

[0125] The spliced ​​results are input into the auxiliary regression module to obtain the auxiliary stability coefficient.

[0126] In this embodiment, the auxiliary regression module is a multilayer perceptron composed of fully connected layers. It calculates the regression error and participates in the gradient backpropagation of the total loss function, achieving synchronous training and parameter optimization with the backbone network. The bottleneck block is the deepest residual transition module connecting the encoder and decoder. It is responsible for extracting the most semantically rich high-dimensional abstract features at the minimum spatial resolution and simultaneously serves as the key input information source for the decoder feature recovery and auxiliary regression modules. The ResUNet network's encoder extracts multi-level abstract features through downsampling and gradually extracts these features through a cascaded structure of max pooling and residual convolution. The decoder gradually recovers the spatial resolution of the feature map using skip connections between the pre-encoders. Furthermore, an auxiliary regression module is introduced between the bottleneck block and the decoder. Global average pooling is performed on both the bottleneck block features and the decoder features, and then they are concatenated with the scale vector along the channel dimension. The concatenated fused features are input into the auxiliary regression module to obtain auxiliary stability coefficients.

[0127] Step S100: Obtain historical slope data, which includes: historical slope geotechnical data, historical slope step surface, historical flat plate features, historical slope mask, and historical slope geometric dimensions; randomly divide the historical slope data into training set and validation set;

[0128] Step S101: Transform the historical slope geotechnical data, historical slope step surface and historical flat plate features in the training set to obtain historical multi-channel features, and construct the historical multi-channel input tensor with the historical slope mask.

[0129] Step S102: Input the historical features from multiple channels into the improved ResUNet network to obtain the historical Logits feature map representing the distribution of the critical slip surface;

[0130] Step S103: Simultaneously extract the encoder bottleneck layer features and decoder intermediate layer features from the improved ResUNet network. Perform global average pooling on the encoder bottleneck layer features and decoder intermediate layer features. Concatenate and fuse the pooled features with the normalized geometric scale feature vector. Input the fused result into the auxiliary regression module to obtain the auxiliary stability coefficient.

[0131] Step S104: Construct a total loss function based on historical Logits feature maps, historical slope masks, and auxiliary stability coefficients, and perform parameter optimization training on the morphological segmentation branch through gradient backpropagation;

[0132] Step S105: During training, the validation set is used to evaluate the segmentation accuracy of the morphological segmentation branch and the regression error of the auxiliary stability coefficient. The network parameters corresponding to the minimum total loss function of the validation set are selected as the pre-trained improved ResUNet network.

[0133] The step of constructing a total loss function based on historical Logits feature maps, historical slope masks, and auxiliary stability coefficients, and then optimizing the parameters of the morphological segmentation branch through gradient backpropagation, includes:

[0134] Step S104.1: Calculate the global positive sample weights based on the pixel distribution inside the historical slope mask in the training set;

[0135] The global positive sample weight is calculated as follows:

[0136] ;

[0137] Clip() is the truncation function; As the lower limit of the weight, This is the upper limit of the weight. This is a set of real sliding surface elements within the historical slope cover. It is the set of non-slip surface pixels within the historical slope mask.

[0138] In this embodiment, during the training iteration process, the weights of positive samples are calculated in real time based on the pixel distribution inside the slope mask. Let the actual sliding surface within the mask (e.g., Figure 7 The pixel set shown is The set of non-slip surface pixels is And set a numerical upper limit to truncate, ensuring that the model focuses on the pixels of the slip surface during the optimization process.

[0139] Step S104.2: Based on the actual critical slip surface, define true positives, false positives, and false negatives within the historical slope mask area, and construct the asymmetric Focal-Tversky loss function;

[0140] Based on the actual critical slip surface, true positives, false positives, and false negatives are defined within the slope mask area, and an asymmetric Focal-Tversky loss function is constructed, calculated as follows:

[0141] Let L be the historical Logits feature map output by the model, and let L be the probability map obtained after Sigmoid activation. ,in, The activation function is Sigmoid. Classification indices are calculated only under the constraint of the historical slope mask.

[0142] ;

[0143] ;

[0144] ;

[0145] Constructing the asymmetric Focal-Tversky loss :

[0146] ;

[0147] ;

[0148] in, Let (i, j) be the probability of a slip surface occurring at position (i, j). Let be the probability of a true slip surface occurring at position (i, j). For the historical slope mask at location (i, j), T P For true positives within the historical slope embankment area, F P For false positives within the historical slope embankment area, F N H represents the false negatives within the historical slope masking area, H is the height of the tensor of the historical multi-channel input, and W is the width of the tensor of the historical multi-channel input. As a focusing factor, First preset constant, It is the second preset constant, and > , The loss function is asymmetric Focal-Tversky. For Tversky similarity coefficient, For smoothing terms, .

[0149] Step S104.3: Construct a mask-weighted binary cross loss function based on the historical slope mask and historical Logits feature map;

[0150] In this embodiment, the calculated positive sample weights are used. The positive samples are weighted, and only the error of the area covered by mask M is accumulated:

[0151] ;

[0152] ;

[0153] in, Here, M is the mask-weighted binary cross-loss function, where M is the historical slope mask. For position Historical slope masking values ​​at the location; For position The weighted binary cross-entropy loss term at the location, For global positive sample weights, For position The actual sliding surface label at the location, For position The output is the predicted slip surface probability value.

[0154] Step S104.4: Perform a weighted summation of the asymmetric Focal-Tversky loss function and the mask-weighted binary crossover loss function to construct the segmentation loss function;

[0155] The asymmetric Focal-Tversky loss function and the mask-weighted binary cross loss function are weighted and summed to construct the segmentation loss function, and the calculation formula is as follows:

[0156] ;

[0157] in, For the cut loss function, As the first weight, As the second weight, .

[0158] Step S104.5: Based on the auxiliary stability coefficient and the true stability coefficient, construct the regression loss function using the mean squared error;

[0159] Based on the auxiliary stability coefficient and the true stability coefficient, a regression loss function is constructed using the mean squared error, and the calculation formula is as follows:

[0160] ;

[0161] Where B is the number of samples input into the network during one training session; The auxiliary stability coefficients output by the auxiliary regression module; The true stability coefficient, Regression loss function.

[0162] Step S104.6: Based on the segmentation loss function and the regression loss function, the total loss function is obtained, and the calculation formula is as follows:

[0163] ;

[0164] in, For the total loss function, To segment the weights that have suffered losses, For the weights of the regression loss, ;

[0165] Step S104.7: Based on the training samples and the total loss function, perform gradient backpropagation to train and optimize the ResUNet network, respectively, to obtain the trained and optimized morphological segmentation model.

[0166] In this embodiment, the AdamW optimizer is used to update the network parameters. :

[0167] ;

[0168] in, The learning rate is the initial value. Furthermore, a cosine annealing mechanism is introduced to achieve adaptive dynamic decay of the learning rate during training. Let be the network parameters at time t. These are the network parameters at time t+1.

[0169] Step S4: Pass the Logits feature map to be evaluated through a Sigmoid nonlinear activation mapping to obtain the slip surface probability distribution field;

[0170] In this embodiment, let X be the input tensor of the slope data to be evaluated, and let the Logits feature map output by the morphological segmentation branch be the following: For any pixel in the feature map Calculate the original probability that it belongs to the slip surface. :

[0171] ;

[0172] Where e is a constant, For the probability distribution field of the slip surface, For any pixel in the Logits feature map to be evaluated .

[0173] Step S5: Perform a Hadamard volume physical purification operation on the probability distribution field of the slope mask and slip surface to be evaluated to obtain the purified probability distribution field;

[0174] In this embodiment, the input slope mask to be evaluated is called. For the probability distribution field of the slip surface Performing the Hadamard product operation, the purified probability distribution field for:

[0175] ;

[0176] Its scalar form is:

[0177] ;

[0178] in, The slope cover is to be evaluated.

[0179] Step S6: Analyze the purified probability distribution field Adaptive threshold binarization is performed, and the final extracted binary image of the critical slip surface is: for:

[0180] ;

[0181] in, The final critical slip surface geometry is all The set, For any point on the critical slip surface .

[0182] In this embodiment, this operation is forced. The probability of the region appearing is reduced to zero, thus eliminating background noise.

[0183] Step S7: Input the features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated into the index quantization branch to obtain the first principal stability coefficient. The index quantization branch includes: a pre-trained fully convolutional network, a global average pooling layer and a multilayer perceptron. The pre-trained fully convolutional network performs continuous downsampling on the tensor of the multi-channel input to be evaluated to extract the first global spatial features. The first global spatial features are compressed into a one-dimensional feature vector to be evaluated by the global average pooling layer. According to the geometric dimensions of the slope to be evaluated, the first geometric bounding box is calculated. The first scale feature vector is obtained according to the first geometric bounding box. The one-dimensional feature vector to be evaluated is concatenated with the first scale feature vector. The concatenated fused features are input into the multilayer perceptron to obtain the first principal stability coefficient.

[0184] The training process of the pre-trained fully convolutional network includes:

[0185] Step S200: Input the historical features of multiple channels into the index quantization branch to obtain the second principal stability coefficient; calculate the mean square error loss function between the second principal stability coefficient and the true stability coefficient, and use the gradient backpropagation algorithm to perform parameter optimization training.

[0186] The step of calculating the mean squared error loss function between the second principal stability coefficient and the true stability coefficient, and then using the gradient backpropagation algorithm for parameter optimization training, includes:

[0187] Step S200.1: Extract high-dimensional global features using a fully convolutional network, and then obtain the second principal stability coefficient by regressing after global average pooling and scale vector concatenation;

[0188] In this embodiment, the input data is an eight-channel regression input tensor. The normalized scale eigenvector is High-dimensional global features Z are extracted using a fully convolutional network:

[0189] ;

[0190] in, For high-dimensional global features, This represents a feature extraction network that includes convolution, batch normalization, and activation functions. Number of output channels The width of the tensor's eight channels (channels 1-3 are normalized physical field data, including: cohesion c, internal friction angle φ, unit weight γ; channels 4-5 are slope geometry field data, including: derived from slope step surface and plate transformation; channels 6-7 are coordinate field data: normalized coordinate field; and channel 8: slope mask). The height of the eight channels of the tensor data. For the tensor's first One channel.

[0191] Perform global average pooling on the feature map Z to compress spatial information, and then concatenate it with the scale vector:

[0192] ;

[0193] ;

[0194] in, This represents the global feature vector of the image. To fuse the joint eigenvectors of geometric scales, V s Let C be the scale vector, and C() be the concatenation operation. Let be the feature vector at the height h and width w of the feature map Z.

[0195] Finally, the predicted second principal stability coefficient is calculated using the multilayer perceptron regression head. :

[0196] ;

[0197] in, This represents a regression mapping function consisting of a fully connected layer, ReLU activation, and Dropout. This is the second principal stability coefficient.

[0198] Step S200.2: Construct the mean squared error loss function based on the second principal stability coefficient and the true stability coefficient;

[0199] Based on the second principal stability coefficient and the true stability coefficient, construct the mean squared error loss function. :

[0200] ;

[0201] in, This is the k-th data point for the second principal stability coefficient. Let B be the mean squared error loss function, and B be the number of samples input into the network in one training iteration. The true safety coefficient for the k-th sample;

[0202] Step S200.3: Based on the training samples and the mean squared error loss function, the fully convolutional network is trained using the gradient backpropagation algorithm to optimize the parameters and obtain the trained and optimized index quantization model.

[0203] Step S201: During training, the regression mean square error of the branches is quantified using the validation set evaluation index. The network parameters corresponding to the minimum validation set error are selected as the pre-trained fully convolutional network.

[0204] In this embodiment, the parameter set of the fully convolutional network is assumed to be... Iterative updates are performed using the AdamW optimizer:

[0205] ;

[0206] ;

[0207] in, This is the calculated value of the current gradient; For learning rate, Let be the parameters of the fully convolutional network at time t. Let be the parameters of the fully convolutional network at time t+1. For the AdamW optimizer.

[0208] Step S8: Use the first principal stability coefficient and the final critical slip surface geometry as the slope stability evaluation results.

[0209] In this embodiment, the mapping function of the fully convolutional network in the index quantization branch is assumed to be... The input is the eight-channel tensor X of the slope data to be evaluated. d and scale vector V s :

[0210] ;

[0211] in: It serves as the first primary stability coefficient, which, together with the finally extracted critical slip surface, constitutes the slope stability evaluation.

[0212] Finally, the slope stability coefficient F of the branch road is output as a quantitative indicator for evaluation. s By combining the binarized critical slip surface image, slope stability evaluation results are generated, such as... Figure 11 As shown.

[0213] In summary, this embodiment proposes an intelligent evaluation method and system for slope stability. The method constructs a dual-branch collaborative deep learning framework, including a morphology segmentation branch and an index quantization branch. The morphology segmentation branch, based on an improved ResUNet architecture, introduces an auxiliary regression module to simultaneously learn the slip surface morphology and global stability. The index quantization branch employs a fully convolutional regression network to directly evaluate the stability coefficient. During model training, a slope geometric mask is used as the physical constraint, and asymmetric loss based on the Tversky exponent and smooth regression loss are employed to construct a multi-task collaborative optimization objective. Ultimately, the critical slip surface morphology and slope stability coefficient are output simultaneously, achieving intelligent evaluation of slope stability. This provides a reliable intelligent tool for slope engineering stability evaluation and has significant theoretical and practical value for promoting the digital transformation and intelligent evaluation of slope engineering.

[0214] Example 2:

[0215] This embodiment also provides an intelligent slope stability evaluation system, such as Figure 12 As shown, it includes: a data acquisition module, a tensor construction module, a morphological segmentation module, an index quantization module, and a result output module. The data acquisition module is connected to the tensor construction module. The tensor construction module is connected to the morphological segmentation module and the index quantization module, respectively. The morphological segmentation module is connected to the critical slip surface acquisition module. The critical slip surface acquisition module and the index quantization module are respectively connected to the result output module.

[0216] The data acquisition module acquires the data of the slope to be evaluated, which includes: the geotechnical data of the slope to be evaluated, the step surface of the slope to be evaluated, the characteristics of the flat plate to be evaluated, the mask of the slope to be evaluated, and the geometric dimensions of the slope to be evaluated.

[0217] The tensor construction module is used to transform the geotechnical mechanics data of the slope to be evaluated, the step surface of the slope to be evaluated, and the features of the flat plate to be evaluated, to obtain multiple channel features to be evaluated. The multiple channel features to be evaluated and the slope mask to be evaluated constitute the tensor of the multiple channel input to be evaluated.

[0218] The morphological segmentation module is used to input the features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated into the morphological segmentation branch to obtain the Logits feature map to be evaluated. The morphological segmentation branch includes a pre-trained improved ResUNet network and an auxiliary regression module, which inputs the features of multiple channels to be evaluated into the pre-trained improved ResUNet network to obtain the Logits feature map to be evaluated that represents the distribution of the critical slip surface.

[0219] The critical slip surface acquisition module is used to obtain the slip surface probability distribution field by passing the Logits feature map to be evaluated through the Sigmoid nonlinear activation mapping; to perform Hadamard product physical purification operation on the slope mask to be evaluated and the slip surface probability distribution field to obtain the purified probability distribution field; and to perform adaptive threshold binarization processing on the purified probability distribution field to obtain the final critical slip surface geometry.

[0220] The index quantization module inputs the features of multiple channels to be evaluated from the tensor of the multi-channel input to be evaluated into the index quantization branch to obtain the first principal stability coefficient. The index quantization branch includes: a pre-trained fully convolutional network, a global average pooling layer, and a multilayer perceptron. The pre-trained fully convolutional network performs continuous downsampling on the tensor of the multi-channel input to be evaluated to extract the first global spatial features. The first global spatial features are compressed into a one-dimensional feature vector to be evaluated through the global average pooling layer. According to the geometric dimensions of the slope to be evaluated, the first geometric bounding box is calculated. The first scale feature vector is obtained according to the first geometric bounding box. The one-dimensional feature vector to be evaluated is concatenated with the first scale feature vector. The concatenated fused features are input into the multilayer perceptron to obtain the first principal stability coefficient.

[0221] The results output module is used to take the first principal stability coefficient and the final critical slip surface geometry as the slope stability evaluation results.

[0222] Example 3:

[0223] 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 aforementioned intelligent evaluation method for slope stability.

[0224] 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 slope stability intelligent evaluation method as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0225] The processor is used to execute all or part of the steps in the intelligent slope stability evaluation method 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.

[0226] 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 slope stability evaluation method described in the above embodiments.

[0227] Example 4:

[0228] 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.

[0229] 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 evaluation method for slope stability described in the various embodiments of this application.

[0230] 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, 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 slope stability evaluation method.

[0231] Example 5:

[0232] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned intelligent evaluation method for slope stability.

[0233] 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.

[0234] 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.

[0235] 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 smart evaluation method for slope stability, characterized in that, include: Acquire the data of the slope to be evaluated, which includes: the geotechnical data of the slope to be evaluated, the step surface of the slope to be evaluated, the characteristics of the flat plate to be evaluated, the mask of the slope to be evaluated, and the geometric dimensions of the slope to be evaluated. The geotechnical data of the slope to be evaluated, the slope step surface of the slope to be evaluated, and the features of the plate to be evaluated are transformed to obtain multiple channel features to be evaluated. The multiple channel features to be evaluated are used to form the tensor of the multiple channel input to be evaluated. The features of multiple channels to be evaluated in the tensor of the input to be evaluated are input into the morphological segmentation branch to obtain the Logits feature map to be evaluated. The morphological segmentation branch includes a pre-trained improved ResUNet network and an auxiliary regression module. The features of multiple channels to be evaluated are input into the pre-trained improved ResUNet network to obtain the Logits feature map to be evaluated that represents the distribution of the critical slip surface. The Logits feature map to be evaluated is processed by a Sigmoid nonlinear activation mapping to obtain the slip surface probability distribution field. Perform a Hadamard volume physical purification operation on the probability distribution field of the slope mask and slip surface to be evaluated, filter out the background noise interference outside the mask, and obtain the purified probability distribution field. Adaptive threshold binarization is performed on the purified probability distribution field to obtain the final critical slip surface geometry. The features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated are input into the index quantization branch to obtain the first principal stability coefficient. The index quantization branch includes: a pre-trained fully convolutional network, a global average pooling layer and a multilayer perceptron. The pre-trained fully convolutional network performs continuous downsampling on the tensor of the multi-channel input to be evaluated to extract the first global spatial features. The first global spatial features are compressed into a one-dimensional feature vector to be evaluated through the global average pooling layer. According to the geometric dimensions of the slope to be evaluated, the first geometric bounding box is calculated. The first scale feature vector is obtained according to the first geometric bounding box. The one-dimensional feature vector to be evaluated is concatenated with the first scale feature vector. The concatenated fused features are input into the multilayer perceptron to obtain the first principal stability coefficient. The first principal stability coefficient and the final critical slip surface geometry are used as the results of slope stability evaluation.

2. The intelligent evaluation method for slope stability according to claim 1, characterized in that, The pre-trained improved ResUNet network training process includes: Historical slope data is acquired, including: historical slope geotechnical data, historical slope step surface, historical flat plate characteristics, historical slope mask, and historical slope geometric dimensions; the historical slope data is randomly divided into training set and validation set. The historical slope geotechnical data, historical slope step surface and historical flat plate features in the training set are transformed to obtain historical multi-channel features. The historical multi-channel features and historical slope mask are used to form a tensor for historical multi-channel input. By inputting historical features from multiple channels into the improved ResUNet network, a historical Logits feature map representing the distribution of the critical slip surface is obtained. Simultaneously, the encoder bottleneck layer features and decoder intermediate layer features are extracted from the improved ResUNet network. The encoder bottleneck layer features and decoder intermediate layer features are then processed by global average pooling. The pooled features are then concatenated and fused with the normalized geometric scale feature vector. The fused result is input into the auxiliary return module to obtain the auxiliary stability coefficient. The total loss function is constructed based on historical Logits feature maps, historical slope masks, and auxiliary stability coefficients. The parameters of the morphological segmentation branch are optimized and trained through gradient backpropagation. During training, the validation set is used to evaluate the segmentation accuracy of the morphological segmentation branch and the regression error of the auxiliary stability coefficient. The network parameters corresponding to the minimum total loss function of the validation set are selected as the pre-trained improved ResUNet network.

3. The intelligent evaluation method for slope stability according to claim 2, characterized in that, The improved ResUNet network includes: convolutional layers, intermediate activation layers, group normalization layers, encoders, decoders, bottleneck blocks, and splicing blocks. The convolutional layers are connected to the intermediate activation layers, the intermediate activation layers are connected to the group normalization layers, the group normalization layers are connected to the encoders, the encoders and decoders are connected by skip connections, the encoders are connected to the bottleneck blocks, the bottleneck blocks and decoders are connected to the splicing blocks respectively, and the splicing blocks are connected to the auxiliary regression module. The features of multiple channels to be evaluated are input into the convolutional layer to obtain the output of the convolutional layer. The output of the convolutional layer is input into the intermediate activation layer to obtain the pre-normalized features. The pre-normalized features are input into the group normalization layer to obtain the normalized output. Normalization is encoded by an encoder, and the output of the encoder is fed into the decoder through a jump connection. The output of the encoder is the Logits feature map to be evaluated. The output of the decoder is used as the input of the bottleneck block. The output of the bottleneck block and the output of the encoder are globally pooled separately. The solution of the global pooling is then concatenated through the concatenation block. The spliced ​​results are input into the auxiliary regression module to obtain the auxiliary stability coefficient.

4. The intelligent evaluation method for slope stability according to claim 1, characterized in that, The training process of the pre-trained fully convolutional network includes: The characteristics of multiple historical channels are input into the index quantization branch to obtain the second principal stability coefficient; the mean squared error loss function between the second principal stability coefficient and the true stability coefficient is calculated, and the gradient backpropagation algorithm is used to optimize the parameters. During training, the regression mean square error of the branches is quantified using the validation set evaluation index. The network parameters corresponding to the minimum validation set error are selected as the pre-trained fully convolutional network.

5. The intelligent evaluation method for slope stability according to claim 1, characterized in that, The process involves transforming the geotechnical data of the slope to be evaluated, the slope step surface, and the features of the flat plate to be evaluated to obtain multiple channel features to be evaluated. These multiple channel features, along with the slope mask, constitute a tensor for the input of the multiple channels to be evaluated. This includes: The geotechnical data of the slope to be evaluated are standardized to obtain normalized physical field data; Based on the features of the slope step surface and the flat plate to be evaluated, an initial geometric shape matrix is ​​extracted. The initial geometric shape matrix is ​​then transposed and numerically cleaned to construct the slope geometric field data. Based on the spatial resolution of the slope geometric field data, a two-dimensional index grid is generated, and the grid coordinates are normalized to construct coordinate field data. Based on the geometric dimensions of the slope to be evaluated, calculate the first geometric bounding box that can completely cover the slope area; Based on the physical width and height of the first geometric bounding box, the first scale feature vector is calculated and normalized. Based on the geometric dimensions of the slope to be evaluated, spatial rasterization is performed to generate a slope mask that identifies the effective area of ​​the slope. The normalized physical field data, slope geometric field data, coordinate field data, and slope mask that identifies the effective area of ​​the slope are used to form the tensor of the multi-channel input to be evaluated.

6. The intelligent evaluation method for slope stability according to claim 1, characterized in that, The method of segmenting the feature input morphology of multiple channels in the tensor to be evaluated includes: normalized physical field data, slope geometric field data, and coordinate field data. The method for quantizing the input index of multiple channel features in the tensor of the multi-channel input to be evaluated includes: normalized physical field data, slope geometric field data, coordinate field data, and the slope mask to be evaluated.

7. The intelligent evaluation method for slope stability according to claim 1, characterized in that, The improved ResUNet network constructs a loss function based on Logits feature maps, slope masks, and auxiliary stability coefficients. It then performs parameter optimization training on the morphology segmentation branch using gradient backpropagation, including: Based on the pixel distribution inside the slope mask in the training set, the global positive sample weights are statistically calculated. Based on the true critical slip surface, we define true positives, false positives, and false negatives within the slope mask area and construct an asymmetric Focal-Tversky loss function. Based on the slope mask and Logits feature map, a mask-weighted binary cross loss function is constructed; We construct a segmentation loss function by weighted summation of the asymmetric Focal-Tversky loss function and the mask-weighted binary cross loss function. Based on the auxiliary stability coefficient and the true stability coefficient, a regression loss function is constructed using mean squared error; Based on the segmentation loss function and the regression loss function, the total loss function is obtained; Based on the training samples and the total loss function, the ResUNet network is trained and optimized using gradient backpropagation to obtain the trained and optimized morphological segmentation model.

8. The intelligent evaluation method for slope stability according to claim 4, characterized in that, The step of calculating the mean squared error loss function between the second principal stability coefficient and the true stability coefficient, and then using the gradient backpropagation algorithm for parameter optimization training, includes: High-dimensional global features are extracted using a fully convolutional network, and after global average pooling and scale vector concatenation, the second principal stability coefficient is obtained by regression. Based on the second principal stability coefficient and the true stability coefficient, construct the mean squared error loss function; Based on the training samples and the mean squared error loss function, the parameters of the fully convolutional network are optimized and trained using the gradient backpropagation algorithm to obtain the optimized index quantization model.

9. A slope stability intelligent evaluation system, used to implement the slope stability intelligent evaluation method according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire the data of the slope to be evaluated, which includes: the geotechnical mechanics data of the slope to be evaluated, the step surface of the slope to be evaluated, the characteristics of the flat plate to be evaluated, the mask of the slope to be evaluated, and the geometric dimensions of the slope to be evaluated. The tensor construction module is used to transform the geotechnical mechanics data of the slope to be evaluated, the step surface of the slope to be evaluated, and the features of the flat plate to be evaluated, to obtain multiple channel features to be evaluated. The multiple channel features to be evaluated and the slope mask to be evaluated constitute the tensor of the multiple channel input to be evaluated. The morphological segmentation module is used to input the features of multiple channels to be evaluated in the tensor of the multi-channel input to be evaluated into the morphological segmentation branch to obtain the Logits feature map to be evaluated. The morphological segmentation branch includes a pre-trained improved ResUNet network and an auxiliary return module, which inputs the features of multiple channels to be evaluated into the pre-trained improved ResUNet network to obtain the Logits feature map to be evaluated that represents the distribution of the critical slip surface. The critical slip surface acquisition module is used to obtain the slip surface probability distribution field by passing the Logits feature map to be evaluated through the Sigmoid nonlinear activation mapping; to perform Hadamard product physical purification operation on the slope mask to be evaluated and the slip surface probability distribution field to obtain the purified probability distribution field; and to perform adaptive threshold binarization processing on the purified probability distribution field to obtain the final critical slip surface geometry. The index quantization module is used to input the features of multiple channels to be evaluated from the tensor of the multi-channel input to be evaluated into the index quantization branch to obtain the first principal stability coefficient. The index quantization branch includes: a pre-trained fully convolutional network, a global average pooling layer, and a multilayer perceptron. The pre-trained fully convolutional network performs continuous downsampling on the tensor of the multi-channel input to be evaluated to extract the first global spatial features. The first global spatial features are compressed into a one-dimensional feature vector to be evaluated by the global average pooling layer. According to the geometric dimensions of the slope to be evaluated, the first geometric bounding box is calculated. The first scale feature vector is obtained according to the first geometric bounding box. The one-dimensional feature vector to be evaluated is concatenated with the first scale feature vector. The concatenated fused features are input into the multilayer perceptron to obtain the first principal stability coefficient. The results output module is used to take the first principal stability coefficient and the final critical slip surface geometry as the slope stability evaluation results.

Images