Artificial intelligence-based embankment slope stability assessment method

Abstract

The invention relates to an embankment slope stability assessment method based on artificial intelligence, and belongs to the technical field of embankment slope monitoring and assessment. The method comprises the following steps: acquiring and marking embankment slope stress sensing data; dividing the data into a plurality of spatio-temporal data blocks; constructing a stability evaluation model; extracting multi-scale convolution features by adopting multi-scale cavity convolution of a high-frequency channel and pooling-deconvolution operation of a low-frequency channel to obtain a fused multi-scale feature matrix; a hidden state sequence is obtained through a space attention mechanism and a double-door-setting mechanism; calculating time interval saliency based on the hidden state vector, then calculating a weighted feature vector, further obtaining a weighted feature matrix, and processing through deep convolution and point-by-point convolution to obtain a pooling feature vector; carrying out stability evaluation grade classification through learnable category prototype and gating feature transformation; and dynamically adjusting sample weight and constraint attention distribution by adopting a total loss function. According to the method, the progressive instability identification capability can be improved.

Description

Technical Field

[0001] This invention belongs to the field of embankment slope monitoring and assessment technology, specifically relating to an artificial intelligence-based method for assessing the stability of embankment slopes. Background Technology

[0002] With the rapid development of highways, railways, and other infrastructure, the stability of embankment slopes has received increasing attention. Embankment slope stability directly affects traffic safety and the long-term service life of engineering structures. Therefore, accurately predicting slope stability, especially under critical instability conditions, has become a crucial issue in the engineering field. Traditional slope stability assessment methods mainly rely on civil engineering theories and experience, such as geological surveys, mechanical models, or finite element analysis. However, these methods often have limitations, especially in practical applications, failing to reflect real-time dynamic changes in slopes and struggling to cope with complex environmental factors.

[0003] In recent years, with the development of sensor technology, more and more studies have begun to use real-time monitoring equipment such as stress sensors to monitor slopes. Stress sensors can collect stress data of slopes, but due to the asynchronous nature of this data in time and space, traditional data processing methods often cannot fully capture the propagation characteristics of stress waves in slope materials. At the same time, slope instability is often a gradual process, with slow changes in the early stages, and signs of critical instability are often overlooked. Therefore, accurately identifying the critical state of slopes using this data and issuing early warnings has become a key challenge in this field. Existing technologies still have many shortcomings in practical applications: conventional methods usually use fixed-length time windows for data segmentation, ignoring spatiotemporal correlations and making it difficult to capture the complete characteristics of stress wave propagation; conventional convolutional neural networks or gated recurrent unit models cannot effectively process high-frequency impact signals and low-frequency gradual features simultaneously, easily missing key instability features; existing methods do not fully consider the spatial importance differences between sensors, treating all sensor features equally, making it difficult for the model to focus on anomalous signals in key areas; conventional stability assessment methods lack sensitivity to critical states, easily exhibiting a delayed response to precursors of slope instability, making it difficult to provide timely warnings of instability risks. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides an artificial intelligence-based method for assessing the stability of embankment slopes, comprising the following steps: Embankment slope stress sensing data acquisition and annotation; based on sensor mutual information values, embankment slope stress sensing data is divided into multiple spatiotemporal data blocks; A stability assessment model is constructed. Multi-scale convolutional features in spatiotemporal data blocks are extracted using multi-scale dilated convolution in high-frequency paths and pooling-deconvolution operations in low-frequency paths to obtain a fused multi-scale feature matrix. Sensor weights are dynamically allocated through a spatial attention mechanism, and the dependence on historical information is adaptively adjusted using a dual-reset gate mechanism to obtain the hidden state sequence of each spatiotemporal data block. The time-segment saliency is calculated based on the hidden state vector at each time step in the hidden state sequence, and a weighted feature vector is obtained based on the time-segment saliency score, thus obtaining the weighted feature matrix of each spatiotemporal data block. The weighted feature matrix is ​​processed through depthwise convolution and pointwise convolution to obtain the pooled feature vector of each spatiotemporal data block. Stability assessment level classification is performed through learnable category prototypes and gated feature transformations.

[0005] During training, the sample weights and constrained attention distribution are dynamically adjusted by using a total loss function that combines the weighted focus loss and the attention smoothing loss.

[0006] Furthermore, stress sensors are deployed at key locations on the embankment slope to collect stress values ​​at a fixed sampling frequency, generating time-series data to obtain embankment slope stress sensing data; based on the actual stability state of the embankment slope, the labeled categories are divided into stable state, critical state, and unstable state.

[0007] Furthermore, this invention calculates the mutual information values ​​of all sensor pairs within a preset time delay range and dynamically determines the set of segmented boundary points based on the curvature change of the sum of the maximum mutual information values ​​of all sensor pairs. This divides the original stress data into multiple spatiotemporal data blocks while preserving the complete spatiotemporal correlation of stress wave propagation. The specific steps are as follows: Histogram statistical sample frequency estimation is used to estimate the joint empirical probability distribution and marginal empirical probability distribution of embankment slope stress sensing data, obtaining the mutual information value of sensor pairs, reflecting the spatiotemporal correlation of stress wave propagation; based on the curvature change of the sum of the maximum mutual information values ​​of all sensor pairs, the key moment points of stress wave propagation mode abrupt change are identified by calculating the second derivative at time points and comparing curvature thresholds, and the set of segmented boundary points is dynamically determined; according to the set of segmented boundary points, the embankment slope stress sensing data is divided into multiple spatiotemporal data blocks, preserving the complete spatiotemporal correlation of stress wave propagation.

[0008] Furthermore, this invention captures transient anomalies through multi-scale dilated convolution in high-frequency channels, extracts gradual trends by combining pooling-deconvolution operations in low-frequency channels, and performs feature fusion and normalization. Simultaneously, it extracts high-frequency transient components and low-frequency trend components from stress data. The specific steps are as follows: Multi-branch dilated convolution is applied to the input spatiotemporal data block, with each branch using a convolutional kernel weight matrix with a different dilation rate. Combined with a modified linear unit activation function, the features of each branch of the high-frequency pathway are output to capture transient shocks and high-frequency anomalous components. Max pooling is applied to the input spatiotemporal data block to compress the temporal dimension, and deconvolution is used to restore the temporal resolution, outputting low-frequency pathway features and extracting low-frequency trend components. The features of each branch of the high-frequency pathway are concatenated with the features of the low-frequency pathway in the channel dimension, and layer normalization is applied to output a fused multi-scale feature matrix to stabilize the feature distribution and enhance training stability.

[0009] Furthermore, this invention dynamically allocates sensor weights through a spatial attention mechanism and adaptively adjusts the dependence of historical information using a dual reset gate mechanism, focusing on key sensor signals and adaptively controlling the retention strength of historical information. The specific steps are as follows: Based on the input feature vector of each sensor at each time step, an attention score is calculated using a multilayer perceptron. The attention score is then normalized into spatial attention weights using a softmax function, enabling dynamic focusing of key sensors. Based on historical hidden states and the current integrated feature vector, a basic reset gate output vector is calculated using a sigmoid activation function. Based on the basic reset gate output vector and feature changes, a dynamic reset gate output vector is calculated using the sigmoid activation function. This dynamic reset gate output vector controls the intensity of historical information retention, coupling the feature change magnitude with the gating mechanism. Building upon the standard gated recurrent unit update mechanism, the calculation method for the dynamic reset gate output replaces the conventional reset gate output calculation method, ultimately yielding the hidden state sequence.

[0010] Furthermore, conventional attention mechanisms treat all time steps equally, but there are often critical triggering periods before slope instability, making it difficult for the model to focus on key period features that are highly correlated with the instability mode, thereby reducing the evaluation accuracy. This invention calculates the significance of each time period and performs adaptive weighting based on the significance to dynamically strengthen the features of time periods that are highly correlated with the instability mode, thereby focusing on key period features. The specific steps are as follows: Based on the hidden state vector at each time step in the hidden state sequence, a projection vector is calculated using a weight matrix and a hyperbolic tangent activation function. The projection vector is then multiplied by the reference vector and normalized to obtain a time-segment saliency score, reflecting the correlation between the time segment and the instability mode. Based on the time-segment saliency score, adaptive weights are calculated using a softmax function and a focus intensity factor to weight the hidden state vector at each time step, resulting in a weighted feature vector that dynamically strengthens the key time-segment features highly correlated with the instability mode. The weighted feature vectors of all time steps are stacked in chronological order to obtain the weighted feature matrix for each spatiotemporal data block.

[0011] Furthermore, this invention achieves feature fusion and dimensionality reduction by using a two-step operation of depthwise convolution and pointwise convolution, which significantly reduces the number of parameters while preserving the spatial structure information of the feature map. The specific steps are as follows: For each channel of the weighted feature matrix of each spatiotemporal data block, a one-dimensional convolution operation is performed using a deep convolution kernel. The output of the deep convolution feature vector is combined with a bias term and a modified linear unit activation function, preserving the spatial structure information within each channel. Based on the concatenated feature matrix output from all channels of the deep convolution, a one-dimensional convolution operation is performed using a pointwise convolution weight matrix. The output of the compressed feature matrix is ​​combined with a bias term, achieving feature fusion and dimensionality reduction. Global average pooling is performed on each channel of the compressed feature matrix in the time dimension to obtain a pooled feature vector representing the global statistical features of each spatiotemporal data block, eliminating the influence of variable time steps.

[0012] Furthermore, this invention enhances the interpretability and robustness of classification decisions by using learnable category prototypes and gated feature transformations, achieving stability assessment-based hierarchical classification. The specific steps are as follows: Based on the pooled feature vector of each spatiotemporal data block, a gate value is calculated using a gate weight matrix and a sigmoid activation function. Simultaneously, a transform value is calculated using a feature transformation weight matrix and an exponential linear unit activation function. The gate value and the transform value are multiplied element-wise to obtain the gated transform feature vector of each spatiotemporal data block. Based on the gated transform feature vector and the prototype vector of each category, a cosine similarity is calculated. Combined with the temperature parameter, a probability distribution is output to obtain the probability of the category to which each spatiotemporal data block belongs.

[0013] Furthermore, this invention improves the model's evaluation performance for gradual instability by dynamically adjusting sample weights and constraining attention distribution through a total loss function that sums the weighted focus loss and the attention smoothing loss. The specific steps are as follows: Based on the output probability and the true label, a weighted focus loss is calculated by combining the category weight and focus parameter; based on the spatial attention weight and the time-segment saliency score, the sum of squared differences between adjacent time steps is calculated, and the attention smoothing loss is output by combining the smoothing coefficient.

[0014] The advantages of this invention are: This invention employs a spatiotemporal segmentation method based on mutual information to address the spatiotemporal asynchrony of sensor data while preserving the complete spatiotemporal correlation of stress wave propagation. Multi-scale dilated convolution and pooling-deconvolution operations effectively extract high-frequency transient components and low-frequency gradual trends, improving the accuracy of feature extraction. By combining spatial attention and a dual reset gate mechanism, the weights of the sensors are dynamically adjusted, and the retention intensity of historical information is adaptively adjusted, enhancing the focus on key location and time-period features. Stability assessment level classification based on prototype similarity, combined with gated feature transformation, enhances the model's interpretability and robustness, and improves the ability to identify gradual instability. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a comparison of the distinguishability of feature distributions under different stability states of the present invention; Figure 3 This is a comparison of the time-series embankment slope stability assessment results of the present invention; Figure 4 This invention provides a comparison of the detection performance of different models in the critical state. Figure 5 This is a comparison of the changes in the loss function during the training process of this invention; Figure 6 This is a comparison of the impact of the number of sensors on the stability evaluation performance of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 In this embodiment, as Figure 1 As shown, this invention provides an artificial intelligence-based method for assessing the stability of embankment slopes, the specific steps of which include: S1. Data Acquisition and Labeling of Embankment Slope Stress Sensing Data Multiple stress sensors are deployed at key locations on the embankment slope. In one embodiment, five sensors are installed at different depths on the slope surface and inside to form a sensor network for continuous monitoring of the stress response of the slope under external loads and environmental factors. The sensor collects stress values ​​at a fixed sampling frequency and generates time series data. Each sensor outputs a continuous time step stress sequence with a time step index from 1 to 4095, covering the complete cycle of the slope from stability to potential instability. The collected raw stress data includes time series of stress values ​​from sensors at different locations, reflecting the propagation process and distribution characteristics of stress waves in the slope material.

[0019] Furthermore, the collected data is labeled based on the actual stability state of the slope. The stability level for each time period is determined through on-site monitoring, historical data recording, and evaluation by engineering experts. In one implementation, the labeling categories are divided into three types: stable state, critical state, and unstable state, which correspond to the normal operation of the slope, the imminent signs of instability, and the occurrence of slippage or collapse, respectively. The labeling process associates the stress data sequence for each time period with the corresponding stability label to form a labeled dataset for subsequent model training. Specifically, for each sensor sequence, the time interval is divided according to the overall behavior of the slope and a consistent label is assigned to ensure that the data accurately corresponds to the real stability state, thereby providing a supervised learning basis for the model.

[0020] S2. Spatiotemporal segmentation of stress data based on mutual information The monitoring data of stress sensors on embankment slopes suffers from spatiotemporal asynchrony. That is, multiple sensors collect stress value time series in the spatial dimension, but due to the stress wave propagation delay, there is a phase difference in the response of sensors at different locations. The conventional processing method uses the sliding window segmentation method, but it will destroy the spatiotemporal correlation of stress data, making it difficult for the model to capture the propagation characteristics of stress waves before critical instability, thus missing the signs of gradual instability.

[0021] This invention divides the original stress data into multiple spatiotemporal data blocks by calculating the mutual information values ​​of all sensor pairs within a preset time delay range and dynamically determining the set of segmented boundary points based on the curvature change of the sum of the maximum mutual information values ​​of all sensor pairs, thereby preserving the complete spatiotemporal correlation of stress wave propagation. The specific steps are as follows: 1) Calculate the time-delay mutual information between sensor pairs The mutual information values ​​of all sensor pairs within a preset time delay range are calculated through empirical probability distribution estimation. Specifically, this involves using histogram statistical sample frequencies to estimate the joint empirical probability distribution and marginal empirical probability distribution of the sensor stress sequences, thereby obtaining the mutual information value used to quantify the statistical dependence of stress responses among sensors. This value reflects the spatiotemporal correlation of stress wave propagation and is expressed as: , In the formula, It represents the mutual information value between the p-th sensor and the q-th sensor under a time delay of τ, which is used to measure the statistical dependence of the stress sequences of the two sensors and reflect the spatiotemporal correlation of stress wave propagation. For time step index, the range of values ​​is... p is the sensor index. ;q is the sensor index that distinguishes it from p. ; Indicates the first The original stress value of each sensor at time t; Indicates the first The original stress value of each sensor at time t+τ; This represents the time delay parameter, with a value range of [value range missing]. This is used to cover possible stress wave propagation delay; This represents the empirical probability distribution function, which is estimated statistically using a histogram. express and The joint empirical probability distribution; express Marginal empirical probability distribution; express Marginal empirical probability distribution; This represents a logarithmic function, with the default base being the natural constant.

[0022] In practical implementation, the empirical probability distribution is obtained through histogram statistical estimation. Item will pass and The range of values ​​is divided into a preset number of equal-width intervals, such as 20 intervals. The frequency of the statistical sample falling into each two-dimensional interval is used as the joint probability estimate. Item will pass The value range is divided into a preset number of equal-width intervals, such as 20 intervals. The frequency of the sample occurrence within each interval is counted as a probability estimate. Item will pass The range of values ​​is divided into a preset number of equal-width intervals, such as 20 intervals, and the frequency of the sample in each interval is counted as the probability estimate.

[0023] 2) Dynamically determine segment boundaries Based on the curvature change of the sum of the maximum mutual information values ​​from all sensors, the key moment points of abrupt changes in the stress wave propagation mode are identified by calculating the second derivative at time points and comparing curvature thresholds. This dynamically determines the set of segmented boundary points, outputs data blocks of adaptive length, and preserves the complete spatiotemporal correlation of stress wave propagation, as expressed in: , In the formula, It represents the set of segment boundary points, which is the set of moments when the stress wave propagation mode changes abruptly, such as stress wave reflection, material fracture, and other abrupt changes. Indicates the first Each segment boundary point corresponds to a key turning point in the stress data; For the index of the segment boundary points, ; This is the total number of segment boundary points; The second derivative operator with respect to a point in time is used to calculate the curvature change and is approximated by discrete difference. Indicates the time delay parameter The maximum value within the range is used to capture the strongest spatiotemporal correlation; Indicates sensor pair The mutual information value at the time delay of τ; This represents the curvature threshold, preferably 0.25. When the curvature exceeds this threshold, it indicates a sudden change in the stress propagation mode.

[0024] In practical implementation, the total number K of segment boundary points is automatically obtained through an algorithm that dynamically determines the segment boundaries. Specifically, this is achieved when calculating the mutual information values ​​of all sensor pairs. After determining the curvature of the term, scan all terms belonging to the interval. At the point in time, the statistical curvature exceeds the curvature threshold. The number of time points is the K value. The magnitude of the K value depends on the frequency of abrupt changes in the stress wave propagation mode. It is the result of the algorithm's operation rather than a preset parameter.

[0025] It should be noted that, Item to sensor index Greater than Summation is performed to avoid redundant calculations of sensor pairs, and then... Item to sensor index Summing from 1 to 5 avoids double-counting sensor pairs, such as... and Equivalent sensor pairs reduce computational complexity and focus on the strongest spatiotemporal correlations.

[0026] 3) Based on the set of segmented boundary points The original stress data is divided into multiple spatiotemporal data blocks to preserve the complete spatiotemporal correlation of stress wave propagation. The i-th spatiotemporal data block is defined as... , For data block index, corresponding to the first Each segment boundary point and the Each segment boundary point The data segments between them contain The stress values ​​at one time step and from five sensors, with dimensions of [missing information]. , Let k be the time step of the kth segment, calculated as follows: .

[0027] It should be noted that, The segment boundary is dynamically set and is adaptive, reflecting the interval of abrupt changes in the stress wave propagation mode.

[0028] S3. Construct a stability assessment model and perform model training. S301, Multi-scale Convolution Feature Extraction The segmented stress data contains a mixture of high-frequency noise and low-frequency trends. Conventional single-scale convolutional neural networks have difficulty capturing transient impact signals and gradual instability features simultaneously. Conventional processing methods use a single convolutional kernel, but due to the fixed receptive field, they cannot adapt to the feature patterns of different frequency bands during stress wave propagation, resulting in insufficient extraction of key features of the critical state of the slope.

[0029] This invention captures transient anomalies through multi-scale dilated convolution in high-frequency pathways, extracts gradual trends by combining pooling-deconvolution operations in low-frequency pathways, and performs feature fusion and normalization. Simultaneously, it extracts high-frequency transient components and low-frequency trend components from stress data. The specific steps are as follows: 1) High-frequency path multi-scale dilated convolution feature extraction A multi-branch dilated convolution operation is applied to the input spatiotemporal data block. Each branch is convolved using a convolution kernel weight matrix with a different dilation rate. Combined with a modified linear unit activation function, the features of each branch of the high-frequency pathway are output to capture transient shocks and high-frequency anomalous components, as shown below: , In the formula, Indicates the high-frequency path number The output features of each branch are used to preserve high-frequency transient components in the stress data; This is a branch index for high-frequency paths. ; This represents the number of branches in the multi-scale dilated convolution in the high-frequency path, with a preferred value of [value to be filled in]. This corresponds to three different void ratios; Indicates the first The convolutional kernel weight matrix of each branch is a trainable parameter; The bias term representing the high-frequency path is a trainable parameter; This indicates a modified linear unit activation function; This represents the dilated convolution operation; The dilation rate of the dilated convolution operation has the following value sequence: , used to control the receptive field size of the convolution kernel; In one implementation, the high-frequency path is a single-layer, multi-branch structure containing three branches corresponding to three different dilation rates, with each branch performing dilated convolution operations independently.

[0030] 2) Low-frequency path downsampling-upsampling feature extraction Max pooling is applied to the input spatiotemporal data block to compress the temporal dimension, and then deconvolution is used to restore the temporal resolution. The low-frequency path features are output, and the low-frequency trend components are extracted, represented as follows: , In the formula, This represents the output characteristics of the low-frequency path, used to extract the low-frequency trend components of stress data; This indicates a deconvolution operation, which restores the compressed feature map to its original temporal resolution. Indicates pooling size as The max pooling operation, i.e., downsampling by taking the maximum value of a local time window, takes the maximum value every four consecutive time steps, thus compressing the time dimension to the original size. .

[0031] 3) Multi-scale feature fusion and layer normalization The features of each branch of the high-frequency pathway are concatenated with the features of the low-frequency pathway along the channel dimension. Then, layer normalization is applied to output a fused multi-scale feature matrix, which is used to stabilize the feature distribution and enhance training stability. This is represented as: , In the formula, Indicates the first A fused multi-scale feature matrix of spatiotemporal data blocks, with dimensions of [missing information]. It integrates high-frequency transient anomalies and low-frequency gradual trends; The representation layer normalization operation standardizes the feature vector at each time step, accelerating model convergence. This indicates a channel-dimensional concatenation operation, which merges the output features of different pathways along the channel dimension. This represents the output characteristic of the first branch of the high-frequency path, corresponding to a porosity of 1; This represents the output characteristic of the first branch of the high-frequency path, corresponding to a porosity of 3; This represents the output characteristic of the second branch of the high-frequency path, corresponding to a void ratio of 5.

[0032] S302, Improved Gated Loop Unit Modeling Long-Term Dependency Conventional gated cyclic units are prone to gradient vanishing problems in long-sequence stress data modeling, and do not consider the differences in importance of sensors at different spatial locations. During slope instability, the response intensity of each sensor varies significantly. Conventional gated cyclic units treat all sensor features equally, making it difficult to focus on anomalous signals at key locations, which affects the accurate prediction of progressive instability.

[0033] This invention dynamically allocates sensor weights through a spatial attention mechanism and adaptively adjusts the dependence of historical information using a dual reset gate mechanism. It focuses on key sensor signals and adaptively controls the retention strength of historical information. The specific steps are as follows: 1) Spatial attention weight calculation Based on the input feature vector of each sensor at each time step, an attention score is calculated using a multilayer perceptron, and then normalized to spatial attention weights using a softmax function to achieve dynamic focusing of key sensors, as shown below: , In the formula, Indicates the first The attention scores of each sensor at time t are calculated using a multilayer perceptron, and the calculation method is expressed as follows: ; This represents a multilayer perceptron used to learn importance mappings of sensor features; Indicates the first The input feature vector of each sensor at time t is a 32-dimensional feature vector. The 32-dimensional feature vector corresponding to the t-th time step and the p-th sensor; Indicates the first The spatial attention weights of each sensor at time t are such that the larger the weight, the greater the contribution of the sensor's features to the instability symptoms at the current time. Indicates the first The attention score of each sensor at time t; To distinguish the sensor indexes from p and q, ; This represents the natural exponential function.

[0034] In one implementation, the multilayer perceptron includes an input layer, a hidden layer, and an output layer. The input layer has a dimension of 32, the hidden layer has a dimension of 64 and uses the ReLU activation function, and the output layer has a dimension of 1 and uses a linear activation function.

[0035] 2) Dual reset door mechanism design Based on the historical hidden states and the current integrated feature vector, the basic reset gate output vector is calculated using the Sigmoid activation function. Simultaneously, based on the feature change, the dynamic reset gate output vector is calculated using the Sigmoid activation function. This dynamic reset gate output vector controls the strength of historical information retention, thus coupling the feature change magnitude with the gating mechanism. This is expressed as: , , In the formula, The basic reset gate output vector at time t controls the degree of reset of the historical hidden states; The dynamic reset gate output vector at time t is used to ultimately control the strength of historical information retention; This indicates the Sigmoid activation function, which compresses the output to... interval; The weight matrix representing the basic reset gate is used to construct the feature vector after linear transformation and is a trainable parameter. The hidden state vector at time t-1 is... 3D eigenvectors, at t=1, It is the initial hidden state vector, which is defined as the zero vector; The weight matrix represents the dynamic reset gate, used to handle feature changes, and is a trainable parameter. The Euclidean distance between feature vectors at adjacent time points quantifies the magnitude of feature change, and is calculated as follows: ; Indicates the first The integrated feature vector at time t represents the set of all sensor features at time t, and is the... The fusion of multi-scale feature matrices from spatiotemporal data blocks The OK; Indicates the t-th The set of all sensor features at any given time; This indicates the calculation of the L2 norm, used to measure the degree of change in features.

[0036] It should be noted that the reset gate of a conventional gated loop unit relies solely on historical hidden states and the current input, making it difficult to adapt to the suddenness of stress wave propagation. This invention addresses this by employing... As a characteristic change quantity, the reset gate can adaptively respond to data mutations when When the value is large, such as when stress wave reflection causes drastic changes in the data, The output value approaches 1, strengthening the role of the basic reset gate and promptly ignoring irrelevant historical information. When the value is small, the dynamic reset gate reduces the reset intensity, retains the long-term gradual trend, and couples the feature change amplitude with the gating mechanism. This allows the model to capture both transient shocks and model gradual instability, rather than simply weighting historical information, which is difficult to achieve directly by conventional gating loop units.

[0037] It should also be noted that, "Larger" and "smaller" are relative concepts, determined by the distribution of the training data and the weight matrix of the dynamically reset gate. In one implementation method, it is jointly decided that... The magnitude is mapped to the Sigmoid activation function. Value, when When the change exceeds the average change of the training set. It can be considered "relatively large", and similarly, It can be considered "small", but the specific threshold is adaptively learned by the model.

[0038] 3) Based on the standard gated loop unit update mechanism, use dynamic reset gate output. The calculation method replaces the conventional reset gate output calculation method, controls the degree of retention of historical information, and the final output dimension is... Hidden state sequence ,definition Let be the hidden state sequence of the i-th spatiotemporal data block, which is the hidden state sequence processed by the improved gated cyclic unit. , dimension , where 128 is the dimension of the hidden state.

[0039] S303, Temporal Self-Attention Feature Enhancement Conventional attention mechanisms treat all time steps equally, but there are often critical triggering periods before slope instability, making it difficult for the model to focus on key period features highly correlated with the instability mode, thus reducing evaluation accuracy. This invention calculates the significance of each time step and performs adaptive weighting based on the significance to dynamically strengthen the features of time steps highly correlated with the instability mode, thereby focusing on key period features, expressed as: 1) Calculate the significance of the time period Based on the hidden state vector at each time step, a projection vector is calculated using the weight matrix and the hyperbolic tangent activation function. This projection vector is then multiplied by the reference vector and normalized to obtain a time-segment significance score, reflecting the correlation between that time segment and the instability mode. This score is expressed as: , In the formula, Indicates the first The significance score for a time period reflects the correlation between that time period and the instability pattern; Indicates the first The projection vector of time-period features is enhanced with nonlinear transformation to improve its representational ability. The calculation method is expressed as follows: ; This represents the hyperbolic tangent activation function, which constrains the projected values ​​to... interval; is a time-period significant weight matrix used for linear transformation of time-period features; it is a trainable parameter. Indicates the first The first spatiotemporal data block The hidden state vector at each time step is a 128-dimensional feature vector. The The hidden state vector at each time step, i.e. The OK; For time period index, ; This indicates the number of time-domain segments. For a single data block, each time step corresponds to a time period, so the number of time-domain segments is equal to the time step size of each data block. This represents the vector dot product operation; It is a reference vector that represents a typical instability mode, and it is a trainable parameter whose value can be optimized through training.

[0040] 2) Adaptive weighting Based on the time-segment saliency score, adaptive weights are calculated using the softmax function and a focus intensity factor. Then, the hidden state vector at each time step is weighted to output a weighted feature vector, dynamically strengthening key time-segment features highly correlated with instability modes, as shown below: , In the formula, Indicates the first The first spatiotemporal data block The weighted feature vector at each time step is a 128-dimensional feature vector. This represents the focus intensity factor, which controls the sharpness of the weighted distribution of the significance score; the optimal value is... ; Indicates the first Significance score for the time period; To distinguish from Time period index, .

[0041] 3) Stack the weighted feature vectors of all time steps in chronological order to form a dimension of The weighted characteristic matrix is ​​defined as the first... The weighted feature matrix of each spatiotemporal data block is , , This indicates a transpose operation, ensuring that the matrix rows correspond to time steps and the columns correspond to feature channels, thereby dynamically strengthening the key time period features related to instability.

[0042] S304, Depthwise Separable Convolutional Feature Compression Fully connected layers directly classify features by ignoring the spatial structure information of the feature maps, and the large number of parameters can easily lead to overfitting. Conventional standard convolutions cannot effectively preserve the independence between channels when compressing features, thus reducing the feature representation ability.

[0043] This invention achieves feature fusion and dimensionality reduction by using two steps: depthwise convolution and pointwise convolution, which significantly reduces the number of parameters while preserving the spatial structure information of the feature map. The specific steps are as follows: 1) Depthwise convolution processing One-dimensional convolution is performed independently on the feature vector of each channel using a depthwise convolution kernel. The output, combining a bias term and a modified linear unit activation function, is a depthwise convolution feature vector that preserves the spatial structure information within each channel. This is represented as: , In the formula, Represents the depthwise convolution. The output feature vector of each channel has a dimension of . Retain the first The spatial structure information of each channel; depthwise convolution is the first step of depthwise separable convolution, which performs convolution independently on each input channel. For channel indexing; Represents the depthwise convolution. Each channel has a depthwise convolutional kernel, which is responsible for extracting the spatial features of a single channel. Represents the depthwise convolution. The bias terms for each channel are trainable parameters; This represents a one-dimensional convolution operation; Indicates the first Weighted feature matrix of spatiotemporal data blocks The feature vector of the c-th channel, i.e. The The column vector is a 128-dimensional feature vector containing all time steps at the 1st dimension. Feature values ​​on each channel.

[0044] 2) Pointwise convolution compression Based on the concatenated feature matrix output from all channels of the depthwise convolution, a one-dimensional convolution operation is performed using the pointwise convolution weight matrix. This, combined with the bias term, outputs a compressed feature matrix, achieving feature fusion and dimensionality reduction, as shown below: , In the formula, Represents the pointwise convolution. The compressed feature matrix of each channel has a dimension of . Pointwise convolution is the second step of depthwise separable convolution, and it uses the output of depthwise convolution to... Convolution is performed using convolution kernels; This represents a pointwise convolution weight matrix with a kernel size of 1 and 32 output channels. It is responsible for feature fusion and dimensionality reduction between channels and is a trainable parameter. This represents the pointwise convolution bias term, which is a trainable parameter; This represents the concatenated feature matrix of all channels output by the depthwise convolution, i.e. , dimension .

[0045] 3) Global average pooling Global average pooling is performed on each channel of the compressed feature matrix along the time dimension, outputting a pooled feature vector representing the global statistical features of the entire data block, thus eliminating the influence of variable time steps, as shown below: , In the formula, For the first The compressed feature vector at each time step, i.e. The A row vector is a 32-dimensional feature vector that represents the row vector at the 3rd row. The set of eigenvalues ​​for all channels at each time step; For the first The pooled feature vector of each spatiotemporal data block is a 32-dimensional feature vector that represents the global statistical features of the entire data block, integrates multi-scale spatiotemporal information, and eliminates the influence of variable time steps.

[0046] S305, Stability Assessment Level Classification After segmentation, the stress data is processed by feature extraction and compression to obtain pooled feature vectors. However, conventional classifiers directly use fully connected layers for classification, ignoring the structural information inside the feature vectors and the correlation between categories, which can easily lead to insufficient accuracy in evaluating progressive instability.

[0047] This invention enhances the interpretability and robustness of classification decisions by using learnable category prototypes and gated feature transformations, achieving stability assessment-based hierarchical classification. The specific steps are as follows: 1) Gated Feature Transformation Based on pooled feature vectors, a gate value is calculated using a gate weight matrix and a sigmoid activation function, while a transform value is calculated using a feature transformation weight matrix and an exponential linear unit activation function. Then, element-wise multiplication is performed to output a gated transform feature vector, enhancing feature representation capabilities. This is expressed as: , In the formula, Indicates the first The gated transform feature vector of each spatiotemporal data block, with dimension . This is used to enhance feature representation capabilities; This represents the exponential linear unit activation function, which handles negative input values. This represents the gate weight matrix, with dimension 1. , are trainable parameters; Represents the gated bias vector, with dimension . , are trainable parameters; This represents the feature transformation weight matrix, with dimension 1. , are trainable parameters; This represents the feature transformation bias vector, with dimension . , are trainable parameters; This indicates element-wise multiplication.

[0048] 2) Prototype Similarity Classification The cosine similarity is calculated between the gated transformation feature vector and the prototype vector of each category. Then, the probability distribution is output by combining the temperature parameter for stability assessment level classification, as shown below: , In the formula, Indicates the first The spatiotemporal data block belongs to the first The probability of each category; For category indexing, ; Indicates the number of stability assessment levels, such as, It corresponds to three states: stable, critical, and unstable. This represents a temperature parameter that controls the sharpness of the probability distribution, with an optimal value. ; Represents the cosine similarity function; Indicates the first The prototype vectors of each category, with dimensions of... , are trainable parameters; To distinguish from Category index, .

[0049] It should be noted that prototype similarity classification utilizes learnable prototype vectors. The typical feature patterns that characterize each stability level enhance the interpretability of classification decisions, and the cosine similarity is insensitive to feature scale, thus improving the robustness of the model.

[0050] S306, Loss Function Calculation Conventional cross-entropy loss functions tend to favor the majority class in cases of class imbalance and do not consider the consistency constraints of the attention mechanism mentioned earlier, which can easily lead to insufficient ability of the model to identify critical unstable states.

[0051] This invention utilizes a multi-task loss function that combines weighted focus loss and attention smoothing loss to dynamically adjust sample weights and constrain attention distribution, thereby improving the model's evaluation performance for gradual instability. The specific steps are as follows: 1) Calculation of weighted focus loss Based on the output probability and the true label, a weighted focus loss is calculated by combining class weights and focus parameters. This dynamically reduces the loss contribution of easily classified samples and focuses on difficult-to-classify samples, as shown below: , In the formula, This represents the weighted focus loss, which makes the model pay more attention to samples that are difficult to classify. Indicates the first The weights of each category are used to balance category frequencies, and are calculated as follows: ; Indicates the total number of training samples; Indicates the first Number of samples in each category; Indicates the first The actual label of the data block is in oneHot encoding format, when the actual category is the first... When there are categories Otherwise ; This indicates the focusing parameter and the preferred value. This is used to adjust the weights of easy and difficult samples.

[0052] 2) Calculation of attention smoothing loss Based on spatial attention weights and time-segment saliency scores, the sum of squared differences between adjacent time steps is calculated. Then, an attention smoothing loss is output by combining a smoothing coefficient to encourage smooth changes in attention distribution, expressed as: , In the formula, This represents the attention smoothing loss, constrains the temporal consistency of attention weights, makes the model more stable in modeling the spatiotemporal characteristics of stress wave propagation, and reduces misjudgments caused by abnormal fluctuations. The spatial attention smoothing coefficient is represented by a preferred value. ; The value represents the significance smoothing coefficient over a given time period, with a preferred value. ; Indicates the first Spatial attention weights of each sensor at time t; Indicates the first The sensor at time t Spatial attention weights at different times; Indicates the first Significance score for the time period; Indicates the first Significance score for the time period.

[0053] 3) Calculation of total loss function The total loss function is obtained by adding the weighted focus loss and the attention smoothing loss, which is used to optimize the model parameters.

[0054] S307, Execution Model Training Process The model training process is performed based on labeled stress data. First, the dataset is divided into a training set and a validation set. The training set is used to optimize model parameters, and the validation set is used to monitor the training process and prevent overfitting.

[0055] The training process employs stochastic gradient descent, aiming to minimize the total loss function through multiple iterations. In each iteration, batches of data are randomly sampled from the training set, input into the model for forward propagation, the output probability and total loss function are calculated, and then the gradient is calculated through the backpropagation algorithm, and all trainable parameters are updated. During training, the changes in training loss and validation loss are recorded. Training is stopped when the validation loss no longer decreases significantly within several consecutive iterations or reaches the preset maximum number of iterations, indicating that the model has converged or reached its optimal performance. In addition, using an early stopping strategy avoids overfitting and ensures the model's generalization ability; After training, the optimal model parameters are saved for subsequent embankment slope stability assessment.

[0056] S4. Embankment Slope Stability Assessment After the model training is completed, for the newly collected stress sensing data of the embankment slope, the stress data spatiotemporal segmentation method based on mutual information is applied to divide the new data into multiple spatiotemporal data blocks. Each data block contains multiple time steps and stress values ​​of all sensors to ensure that the complete spatiotemporal correlation of stress wave propagation is preserved. Then, each spatiotemporal data block is input into the trained stability evaluation model. The model outputs the probability that each data block belongs to a stable, critical, or unstable state through multi-scale convolutional feature extraction, improved gated recurrent unit modeling of long-term dependencies, temporal self-attention feature enhancement, depthwise separable convolutional feature compression, and stability evaluation level classification steps. Finally, the stability assessment level of the slope is determined based on the highest probability or a preset threshold, and the assessment results are output to provide a basis for engineering decisions, such as timely warning of potential instability risks or confirmation of the slope's safety status.

[0057] Example 2 In this embodiment, as Figure 2 As shown, the advantages of the proposed method in feature learning are demonstrated through feature distribution visualization. High-dimensional features are displayed on a two-dimensional plane using dimensionality reduction techniques, intuitively presenting the distribution of stable, critical, and unstable states in the feature space. Good distinguishability of the feature distribution is the foundation for accurate model classification and an important basis for evaluating feature extraction capabilities. The scatter plot clearly shows that the three states form relatively independent clusters in the feature space. Feature points of the stable state are mainly concentrated in the lower left region of the plot (green); feature points of the critical state are distributed in the middle region (orange); and feature points of the unstable state are clustered in the upper right region (red). The distribution pattern matches the actual physical process of the embankment slope. The transition from stable to unstable state is a continuous process, and the critical state, as a transitional stage, does indeed partially overlap with the other two states in terms of feature distribution. Each point in the plot represents a data sample, and the position of the point is determined by two feature dimensions. The spatiotemporal segmentation method based on mutual information determines two feature dimensions. Feature dimension 1 corresponds to the curvature change feature of the sum of the maximum mutual information values, obtained by calculating the mutual information values ​​of all sensor pairs within a preset time delay range and analyzing their statistical characteristics. Feature dimension 2 comes from the latent state features extracted by multi-scale convolution and the output of gated recurrent units. These features, after being compressed by depthwise separable convolution, can effectively characterize the intensity changes of spatiotemporal correlation. Although there are some overlapping areas among the points of the three categories, they form a clear distribution trend overall, indicating that the features learned by the method of this invention have good class discrimination ability. In particular, the critical state, as the most important early warning stage, has a relatively concentrated distribution area, which is conducive to the model accurately identifying early signs of slope instability.

[0058] Example 3 In this embodiment, time series analysis is used to compare the dynamic performance of different methods in the actual monitoring process. For example... Figure 3As shown, the experiment simulates the stability change process of a complete embankment slope under ideal conditions, gradually transitioning from an initial stable state to a critical state, and finally entering an unstable state. Time series analysis can intuitively demonstrate the detection capability and response speed of each method at key points of state transition. Three main curves can be observed in the time series graph: the black dashed line represents the actual stability state change, the red curve represents the prediction result of the conventional method, and the blue curve represents the prediction result of the method of this invention. The three colored regions in the background clearly delineate the time intervals of the stable region (green), the critical region (orange), and the unstable region (red). The actual state shows that the first quarter of the time period is in a stable state, the middle third is in a critical state, and the last quarter enters an unstable state. The conventional method uses a fixed-length time window to segment the continuous stress data collected by the sensor, then uses a standard convolutional neural network for local feature extraction, and combines this with basic gated recurrent units to model the time dependency. Comparing the two prediction curves reveals that the conventional method has a significant detection delay at the state transition point, especially at the transition point from stable to critical state, where the response of the conventional method is significantly lagging. The method of this invention can capture state changes more promptly, and the predicted curve matches the actual state curve more closely. In the stable phase, both predicted curves are relatively stable; in the critical phase, the predicted values ​​of this invention fluctuate less, exhibiting better stability; in the unstable phase, this invention can quickly identify state changes. The vertical axis corresponds to the three states of stability, criticality, and instability, respectively, while the horizontal axis represents the continuous sequence of sampling points.

[0059] Example 4 In this embodiment, as Figure 4 As shown, a comparative analysis of critical state detection performance was conducted to verify the advantages of the proposed model in detecting the most challenging critical states (precursors to instability). The Long Short-Term Memory (LSTM) network model, the Convolutional Neural Network-Gated Recurrent Unit (CNN-Gated Recurrent Unit) combination model, the Transformer model, and the proposed model were compared. Evaluation metrics included accuracy, recall, and F1 score. The experimental configuration used the same 5-sensor dataset, focusing on critical state samples (15% of the dataset). The grouped bar charts show that the proposed model outperforms other models in recall, indicating the lowest false negative rate. This demonstrates that the spatial attention mechanism enables the model to focus on sensor signals at key locations, promptly capturing abrupt changes such as material fracture. Furthermore, the proposed model outperforms the compared algorithms in both accuracy and F1 score.

[0060] Example 5 In this embodiment, as Figure 5As shown, the trend of the model loss function during training is analyzed, and the convergence characteristics and training stability of the proposed method are compared with those of conventional Long Short-Term Memory (LSTM) networks. The experiment uses a line graph to illustrate the relationship between training epochs and loss values. The horizontal axis represents the number of training epochs, indicating the progress of iterative training; the vertical axis represents the dimensionless value of the loss value, reflecting the magnitude of the model's prediction error. Lower values ​​indicate better model performance. The graph shows that the loss curve of the proposed method decreases rapidly in the early stages of training, quickly reaching and maintaining a low level with relatively small fluctuations. In contrast, the loss curve of the conventional method decreases slowly, eventually remaining at a higher level with relatively large fluctuations. This indicates that the proposed method has faster convergence speed and better training stability, enabling more efficient learning of slope instability feature patterns. Experimental results show that the multi-scale convolutional feature extraction and depthwise separable convolutional feature compression techniques in the proposed method effectively improve the efficiency of feature extraction and the robustness of model training.

[0061] Example 6 In this embodiment, the impact of the number of sensors on performance evaluation is analyzed to verify the adaptability and practicality of the method of the present invention under different engineering configurations. Figure 6 As shown in the figure, the experiment uses a combination of line graphs and scatter plots to illustrate the trend of evaluation accuracy as the number of sensors increases from 2 to 7. The horizontal axis represents the number of sensors (in units), and the vertical axis represents the evaluation accuracy (in dimensionless decimals). The figure shows that as the number of sensors increases, the accuracy of both methods increases, but the method of this invention maintains a higher accuracy across all sensor configurations, especially when the number of sensors is small. When the number of sensors reaches 5 or more, the accuracy of the method of this invention tends to stabilize and remain at a high level, while the conventional method, even with an increase in the number of sensors, shows limited improvement in accuracy and a lower overall level. This demonstrates that the method of this invention has a lower dependence on the number of sensors and can still achieve reliable evaluation under limited sensor deployment conditions. This indicates that the spatiotemporal segmentation technology based on mutual information can fully utilize the spatiotemporal correlation in the sensor network, effectively capturing the key characteristics of stress wave propagation even with a small number of sensors.

[0062] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing the stability of embankment slopes based on artificial intelligence, characterized in that, Includes the following steps: Embankment slope stress sensing data acquisition and annotation; based on sensor mutual information values, embankment slope stress sensing data is divided into multiple spatiotemporal data blocks; A stability assessment model is constructed. Multi-scale convolutional features in spatiotemporal data blocks are extracted using multi-scale dilated convolution in high-frequency paths and pooling-deconvolution operations in low-frequency paths, resulting in a fused multi-scale feature matrix. Sensor weights are dynamically allocated using a spatial attention mechanism, and the dependence on historical information is adaptively adjusted using a dual-reset gate mechanism, yielding a hidden state sequence for each spatiotemporal data block. The time-segment saliency is calculated based on the hidden state vector at each time step in the hidden state sequence, and a weighted feature vector is obtained based on the time-segment saliency score, thus obtaining a weighted feature matrix for each spatiotemporal data block. The weighted feature matrix is ​​processed using depthwise convolution and pointwise convolution to obtain a pooled feature vector for each spatiotemporal data block. Stability assessment level classification is performed using learnable category prototypes and gated feature transformations. During training, the sample weights and constrained attention distribution are dynamically adjusted by using a total loss function that combines the weighted focus loss and the attention smoothing loss.

2. The artificial intelligence-based embankment slope stability assessment method according to claim 1, characterized in that, Embankment slope stress sensing data acquisition and annotation: Stress sensors are deployed at key locations on the embankment slope to collect stress values ​​at a fixed sampling frequency, generating time-series data to obtain embankment slope stress sensing data. Based on the actual stability state of the embankment slope, the labeled categories are divided into stable state, critical state, and unstable state.

3. The artificial intelligence-based embankment slope stability assessment method according to claim 1, characterized in that, Based on the mutual information values ​​of the sensors, the stress sensing data of the embankment slope is divided into multiple spatiotemporal data blocks: Histogram statistical sample frequency estimation is used to estimate the joint empirical probability distribution and marginal empirical probability distribution of embankment slope stress sensing data, and the mutual information value of sensor pairs is obtained. Based on the curvature change of the sum of the maximum mutual information of all sensor pairs, the key moment points of stress wave propagation mode change are identified by calculating the second derivative at time points and comparing curvature thresholds, and the set of segment boundary points is dynamically determined. Based on the set of segment boundary points, the embankment slope stress sensing data is divided into multiple spatiotemporal data blocks.

4. The artificial intelligence-based method for assessing embankment slope stability according to claim 1, characterized in that, The specific process for obtaining the fused multi-scale feature matrix is ​​as follows: Multi-branch dilated convolution is applied to the input spatiotemporal data block, with each branch using a convolution kernel weight matrix with a different dilation rate. Combined with a modified linear unit activation function, the features of each branch of the high-frequency path are output. Max pooling is applied to the input spatiotemporal data block to compress the temporal dimension, and deconvolution is used to restore the temporal resolution, outputting the features of the low-frequency path. The features of each branch of the high-frequency path are concatenated with the features of the low-frequency path in the channel dimension, and layer normalization is applied to output a fused multi-scale feature matrix.

5. The artificial intelligence-based method for assessing the stability of embankment slopes according to claim 1, characterized in that, The specific process for obtaining the hidden state sequence of each spatiotemporal data block is as follows: Based on the input feature vector of each sensor at each time step, the attention score is calculated by a multilayer perceptron, and the attention score is normalized into spatial attention weights using the softmax function. Based on the historical hidden states and the current integrated feature vector, the basic reset gate output vector is calculated using the Sigmoid activation function. Based on the basic reset gate output vector and the feature change, the dynamic reset gate output vector is calculated using the Sigmoid activation function. On the basis of the standard gated recurrent unit update mechanism, the calculation method of the dynamic reset gate output is used to replace the calculation method of the conventional reset gate output, and finally the hidden state sequence is obtained.

6. The artificial intelligence-based method for assessing embankment slope stability according to claim 1, characterized in that, The specific process for obtaining the weighted feature matrix of each spatiotemporal data block is as follows: Based on the hidden state vector at each time step in the hidden state sequence, the projection vector is calculated using the weight matrix and the hyperbolic tangent activation function. The projection vector is then multiplied by the reference vector and normalized to obtain the time-segment saliency score. Based on the time-segment saliency score, adaptive weights are calculated using the softmax function and focus intensity factor. The hidden state vectors of each time step are weighted to obtain a weighted feature vector. The weighted feature vectors of all time steps are stacked in chronological order to obtain the weighted feature matrix of each spatiotemporal data block.

7. The artificial intelligence-based method for assessing the stability of embankment slopes according to claim 1, characterized in that, The specific process for obtaining the pooled feature vector of each spatiotemporal data block is as follows: For each channel of the weighted feature matrix of each spatiotemporal data block, a one-dimensional convolution operation is performed using a deep convolution kernel. The output of the deep convolution feature vector is then combined with a bias term and a modified linear unit activation function. Based on the concatenated feature matrix output from all channels of the deep convolution, a one-dimensional convolution operation is performed using a pointwise convolution weight matrix. The output of the compressed feature matrix is ​​then combined with a bias term. Global average pooling is then performed on each channel of the compressed feature matrix in the time dimension to obtain the pooled feature vector of each spatiotemporal data block.

8. The artificial intelligence-based method for assessing the stability of embankment slopes according to claim 1, characterized in that, Based on the pooled feature vector of each spatiotemporal data block, the gate value is calculated using the gate weight matrix and the Sigmoid activation function, and the transformation value is calculated using the feature transformation weight matrix and the exponential linear unit activation function. The gate value and the transformation value are multiplied element by element to obtain the gated transformation feature vector of each spatiotemporal data block. The cosine similarity is calculated based on the gated transform feature vector and the prototype vector of each category. The probability distribution is then combined with the temperature parameter to obtain the probability of each spatiotemporal data block belonging to a category.

9. The artificial intelligence-based method for assessing the stability of embankment slopes according to claim 1, characterized in that, The weighted focus loss is calculated based on the output probability and the true label, combined with the category weight and focus parameter.

10. The artificial intelligence-based method for assessing embankment slope stability according to claim 1, characterized in that, Based on spatial attention weights and time-segment saliency scores, the sum of squared differences between adjacent time steps is calculated, and the attention smoothing loss is output by combining the smoothing coefficient.

Images