Artificial intelligence-based embankment slope stability assessment method

Abstract

The application relates to an embankment slope stability evaluation method based on artificial intelligence and belongs to the technical field of embankment slope monitoring and evaluation. The method comprises the following steps: collecting and labeling embankment slope stress sensing data; dividing the data into multiple spatiotemporal 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-anti-convolution operation of a low-frequency channel, obtaining a multi-scale feature matrix; obtaining an implicit state sequence through a spatial attention mechanism and a double reset gate mechanism; calculating a time period saliency based on an implicit state vector, then calculating a weighted feature vector, and further obtaining a weighted feature matrix, which is processed through deep convolution and point-by-point convolution to obtain a pooling feature vector; performing stability evaluation grade classification through a learnable category prototype and a gated feature transformation; and dynamically adjusting sample weights and constraint attention distribution by adopting a total loss function. The application can improve the recognition ability for progressive instability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of embankment slope monitoring and evaluation, and particularly relates to an embankment slope stability evaluation method based on artificial intelligence. BACKGROUND

[0002] With the rapid development of highways, railways and other infrastructures, the stability of embankment slopes has attracted increasing attention. The stability of embankment slopes directly affects traffic safety and the long-term use of engineering structures. Therefore, how to accurately predict the stability of slopes, especially in the critical instability state, has become an important problem in the engineering field. Traditional methods for evaluating the stability of slopes mainly rely on the theory and experience of civil engineering, such as geological exploration, mechanical models or finite element analysis. However, these methods often have limitations, especially in practical operation, they cannot reflect the dynamic changes of slopes in real time, and it is difficult to cope with complex environmental factors.

[0003] In recent years, with the development of sensor technology, more and more research has begun to monitor slopes with real-time monitoring equipment such as stress sensors. Stress sensors can collect stress data of slopes, but due to the asynchronous problem of these data in space and time, 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, and the signs of critical instability are often overlooked. Therefore, how to accurately identify the critical state of the slope using these data and make early warnings has become a key challenge in this field, and existing technologies still have many shortcomings in practical application: conventional methods usually use fixed-length time windows for data segmentation, ignoring the spatio-temporal correlation, and it is 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 change characteristics at the same time, and key instability features are easily missed; existing methods do not fully consider the spatial importance differences between sensors, and all sensor features are treated equally, making it difficult for the model to focus on abnormal signals in key areas; conventional stability evaluation methods lack sensitivity to critical states, and are prone to lag response to signs of slope instability, making it difficult to timely warn of instability risks. SUMMARY

[0004] To achieve the above-mentioned purpose, the following technical solutions are implemented:

[0005] The application provides an embankment slope stability evaluation method based on artificial intelligence, comprising the following steps:

[0006] Embodiment 1: The method comprises the following steps:

[0007] The stability evaluation model is constructed; multi-scale convolution features in the space-time data block are extracted by using multi-scale cavity convolution of the high-frequency channel and pooling-deconvolution operation of the low-frequency channel, and a fusion multi-scale feature matrix is obtained; sensor weights are dynamically allocated through a spatial attention mechanism, and the dependence degree of historical information is adaptively adjusted by using a double reset gate mechanism, and a hidden state sequence of each space-time data block is obtained; the period significance is calculated based on the hidden state vector of each time step in the hidden state sequence, the weighted feature vector is obtained based on the period significance score, and then the weighted feature matrix of each space-time data block is obtained; the weighted feature matrix is processed through deep convolution and point-by-point convolution, and the pooling feature vector of each space-time data block is obtained; and the stability evaluation grade classification is performed through the learnable class prototype and the gated feature transformation.

[0008] In the training process, the total loss function obtained by adding the weighted focal loss and the attention smoothing loss is used to dynamically adjust the sample weight and constrain the attention distribution.

[0009] Further, stress sensors are deployed at key positions of the embankment slope to collect stress values at a fixed sampling frequency, generate time series data, and obtain embankment slope stress sensor data; based on the actual stability state of the embankment slope, the labeled categories are divided into stable state, critical state and unstable state.

[0010] Further, the present application divides the original stress data into multiple space-time data blocks by calculating the mutual information value of all sensor pairs within a preset time lag range and dynamically determining the segmentation boundary point set based on the curvature change of the sum of the maximum mutual information values of all sensor pairs, thereby retaining the complete space-time correlation of stress wave propagation. The specific steps are as follows:

[0011] The histogram is used to estimate the joint empirical probability distribution and the marginal empirical probability distribution of the embankment slope stress sensor data, and the mutual information value of the sensor pair is obtained, which reflects the space-time correlation of stress wave propagation; based on the curvature change of the sum of the maximum mutual information values of all sensor pairs, the second derivative at the time point is calculated and compared with the curvature threshold to identify the key time points of the mutation of the stress wave propagation mode, and the segmentation boundary point set is dynamically determined; the embankment slope stress sensor data is divided into multiple space-time data blocks according to the segmentation boundary point set, and the complete space-time correlation of stress wave propagation is retained.

[0012] Further, the present application captures transient anomalies through multi-scale cavity convolution of the high-frequency channel, extracts gradual trends in combination with pooling-deconvolution operation of the low-frequency channel, and performs feature fusion and normalization, while extracting high-frequency transient components and low-frequency trend components in the stress data, and the specific steps are as follows:

[0013] The multi-branch hole convolution operation is applied to the input spatio-temporal data block, each branch uses a convolution kernel weight matrix with different hole rates to perform convolution, and the features of each branch of the high-frequency channel are output by combining a rectified linear unit activation function to capture transient impact and high-frequency abnormal components; the maximum pooling operation is applied to the input spatio-temporal data block to compress the time dimension, and the time resolution is restored through the deconvolution operation to output the low-frequency channel features and extract low-frequency trend components; the features of each branch of the high-frequency channel and the low-frequency channel features are spliced in the channel dimension, and a layer normalization operation is applied to output a fused multi-scale feature matrix, which is used to stabilize the feature distribution and enhance the training stability.

[0014] Further, the application dynamically allocates sensor weights through a spatial attention mechanism, and adaptively adjusts the dependence degree of historical information by using a double reset gate mechanism, focuses on key sensor signals and adaptively controls the retention intensity of historical information, and the specific steps are as follows:

[0015] Based on the input feature vector of each sensor at each time step, the attention score is calculated through a multi-layer perception, and the attention score is normalized to a spatial attention weight using a softmax function to realize dynamic focusing of key sensors; based on the historical hidden state and the current integrated feature vector, the basic reset gate output vector is calculated through a Sigmoid activation function, and based on the basic reset gate output vector and the feature change amount, the dynamic reset gate output vector is calculated through a Sigmoid activation function, which is used to control the retention intensity of historical information and realize the coupling of feature change amplitude and gating mechanism; based on the standard gating 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.

[0016] Further, the conventional attention mechanism equally processes all time steps, but there is often a key triggering period for slope instability, which makes it difficult for the model to focus on the key period features highly related to the instability mode, thereby reducing the evaluation accuracy, the application dynamically enhances the period features highly related to the instability mode by calculating the saliency of each period and adaptively weighting based on the saliency, and then focuses on the key period features, and the specific steps are as follows:

[0017] Based on the hidden state vector of each time step in the hidden state sequence, the projection vector is calculated through a weight matrix and a hyperbolic tangent activation function, the projection vector is subjected to inner product operation with a reference vector and is normalized to obtain a period saliency score, which reflects the relevance of the period to the instability mode; based on the period saliency score, the adaptive weight is calculated through a softmax function and a focus intensity factor, and the hidden state vector of each time step is weighted to obtain a weighted feature vector, which dynamically enhances the key period features highly related to the instability mode; the weighted feature vectors of all time steps are stacked in time sequence to obtain a weighted feature matrix of each spatio-temporal data block.

[0018] Further, the application realizes feature fusion and dimension reduction by two-step operation of deep convolution and point-by-point convolution, while greatly reducing the number of parameters and preserving the spatial structure information of the feature map, and the specific steps are as follows:

[0019] The deep convolution kernel is applied to each channel of the weighted feature matrix of each spatiotemporal data block to perform one-dimensional convolution operation, and the bias term and the rectified linear unit activation function are combined to output the deep convolution feature vector, thereby preserving the spatial structure information within the channel; based on the spliced feature matrix output by all channels of the deep convolution, a point-by-point convolution weight matrix is applied to perform one-dimensional convolution operation, and a bias term is combined to output a compressed feature matrix, thereby realizing feature fusion and dimension reduction; the global average pooling is performed on each channel of the compressed feature matrix in the time dimension to obtain a pooling feature vector representing the global statistical features of each spatiotemporal data block, thereby eliminating the influence of the variable time step.

[0020] Further, the application enhances the explainability and robustness of classification decision by learning class prototypes and gated feature transformation, realizes stable evaluation grade classification based on prototype similarity, and the specific steps are as follows:

[0021] The gating value is calculated based on the pooling feature vector of each spatiotemporal data block through the gating weight matrix and the Sigmoid activation function, and the transformation value is calculated through the feature transformation weight matrix and the exponential linear unit activation function, and the gating 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 transformation feature vector and each class prototype vector, and the probability distribution is output combined with the temperature parameter to obtain the probability of the class to which each spatiotemporal data block belongs.

[0022] Further, the application dynamically adjusts the sample weight and constrains the attention distribution by adding the total loss function of the weighted focal loss and the attention smoothing loss, thereby improving the evaluation performance of the model on the gradual instability, and the specific steps are as follows:

[0023] Based on the output probability and the true label, the weighted focal loss is calculated combined with the class weight and the focus parameter; based on the spatial attention weight and the time period significance score, the difference square sum of adjacent time steps is calculated, and the attention smoothing loss is output combined with the smoothing coefficient.

[0024] The application has the following advantages:

[0025] The application adopts a space-time segmentation method based on mutual information, solves the space-time asynchrony problem of sensor data, and retains the complete space-time correlation of stress wave propagation; adopts multi-scale hollow convolution and pooling-anti-convolution operation, can effectively extract high-frequency transient components and low-frequency gradual trends, and improves the accuracy of feature extraction; combines the spatial attention mechanism and the double reset gate mechanism, dynamically adjusts the weight of the sensor, and adaptively adjusts the retention strength of the historical information, enhances the attention to the key position and period characteristics; based on the prototype similarity, the stability evaluation grade classification is carried out, combined with the gated feature transformation, the model's explainability and robustness are enhanced, and the identification ability to the progressive instability is improved. BRIEF DESCRIPTION OF DRAWINGS

[0026] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application.

[0027] Figure 1 The step flow chart of the method of the application is shown in the figure;

[0028] Figure 2 The distinguishability contrast of feature distribution under different stability states of the application is shown in the figure;

[0029] Figure 3 The time series of the application is shown in the figure;

[0030] Figure 4 The detection performance of different models in the critical state of the application is shown in the figure;

[0031] Figure 5 The loss function change in the training process of the application is shown in the figure;

[0032] Figure 6 The influence of the number of sensors on the stability evaluation performance of the application is shown in the figure. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0034] Embodiment 1

[0035] In this embodiment, as shown in the figure, Figure 1 The application provides a road embankment slope stability evaluation method based on artificial intelligence, and the specific steps include:

[0036] S1, embankment slope stress sensor data collection and labeling

[0037] A plurality of stress sensors are deployed at key positions of the embankment slope. In an embodiment, 5 sensors are installed at different depths on the surface and inside the slope to form a sensor network for continuously monitoring the stress response of the slope under external load and environmental factors;

[0038] The sensors collect stress values at a fixed sampling frequency to generate time series data. Each sensor outputs a continuous time step stress sequence with a time step index from 1 to 4095, covering the entire period from stability to potential instability of the slope;

[0039] The collected raw stress data includes time series of stress values from different position sensors, reflecting the propagation process and distribution characteristics of stress waves in the slope material.

[0040] Further, the collected data is labeled. The data labeling is based on the actual stability state of the slope. The stability level of each time period is determined through field monitoring, historical data recording and expert evaluation;

[0041] In an embodiment, the labeling categories are divided into stable state, critical state and unstable state, corresponding to normal operation, near instability and sliding or collapse of the slope;

[0042] The labeling process associates the stress data sequence of each time period with the corresponding stability label to form a labeled data set 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 corresponds accurately to the true stability state, thereby providing a supervised learning basis for the model.

[0043] S2, stress data spatiotemporal segmentation based on mutual information

[0044] The embankment slope stress sensor monitoring data has a spatiotemporal asynchrony problem. Multiple sensors collect time series of stress values in the spatial dimension, but due to the delay of stress wave propagation, there is a phase difference in the response of different position sensors. The conventional processing method uses a sliding window segmentation method, but this method destroys the spatiotemporal correlation of stress data, making it difficult for the model to capture the propagation characteristics of stress waves before critical instability, thereby missing the progressive instability signs.

[0045] The present application divides the original stress data into multiple spatiotemporal data blocks by calculating the mutual information value of all sensor pairs within a preset time lag range and dynamically determining the segmentation boundary point set based on the curvature change of the sum of all sensor pair mutual information maximum values, thereby preserving the complete spatiotemporal correlation of stress wave propagation. The specific steps are as follows:

[0046] 1) Calculate the time-delay mutual information between sensor pairs

[0047] 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:

[0048] ,

[0049] 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 values ​​of the sensors 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.

[0050] In practical implementation, the empirical probability distribution is obtained through histogram statistical estimation. Item by 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 by 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.

[0051] 2) Dynamically determine segment boundaries

[0052] 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:

[0053] ,

[0054] 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 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.

[0055] 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.

[0056] 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... With equivalent sensor pairs, reduce the computational complexity, and focus on the strongest spatiotemporal correlation.

[0057] 3) According to the segmented boundary point set The original stress data is divided into multiple spatiotemporal data blocks, the complete spatiotemporal correlation of stress wave propagation is retained, and the ith spatiotemporal data block is defined as , is the data block index, corresponding to the data segment between the segmented boundary point and the segmented boundary point , containing stress values of 5 sensors and time steps, with a dimension of , is the time step of the kth segment, and the calculation method is represented as .

[0058] It should be noted that The segmented boundary is dynamically set, which is an adaptive setting, reflecting the interval of the mutation of the stress wave propagation mode.

[0059] S3, constructing a stability evaluation model and performing model training

[0060] S301, multi-scale convolution feature extraction

[0061] The segmented stress data contains mixed characteristics of high-frequency noise and low-frequency trends. The conventional single-scale convolutional neural network is difficult to capture both transient impact signals and gradual instability characteristics. The conventional processing method uses a single convolution kernel, but due to the fixed receptive field, it cannot adapt to different frequency band feature patterns in the stress wave propagation process, resulting in insufficient extraction of key features of the slope critical state.

[0062] The present application captures transient anomalies through multi-scale hollow convolution of high-frequency channels, extracts gradual trends through pooling-deconvolution operations of low-frequency channels, and performs feature fusion and normalization, while extracting high-frequency transient components and low-frequency trend components in stress data. The specific steps are as follows:

[0063] 1) Multi-scale hollow convolution feature extraction of high-frequency channel

[0064] Apply multi-branch hollow convolution operation to the input spatiotemporal data block, and use different hollow rate kernel weight matrices for convolution in each branch. Combine the modified linear unit activation function to output the features of each branch of the high-frequency channel, capture transient impact and high-frequency abnormal components, and represent as:

[0065] ,

[0066] In the formula, represents the output feature of the high-frequency path branch, used to retain the high-frequency transient component in the stress data; represents the output feature of the high-frequency path branch, used to retain the high-frequency transient component in the stress data; is the branch index of the high-frequency path, ; represents the number of branches of the multi-scale dilated convolution in the high-frequency path, preferably , corresponding to three different dilated rates; represents the convolution kernel weight matrix of the th branch, which is a trainable parameter; represents the bias term of the high-frequency path, which is a trainable parameter; represents the rectified linear unit activation function; represents the dilated convolution operation; is the dilated rate of the dilated convolution operation, and the value sequence is , used to control the receptive field size of the convolution kernel;

[0067] In an embodiment, the high-frequency path is a single-layer multi-branch structure, containing 3 branches corresponding to three dilated rates, and each branch independently performs a dilated convolution operation.

[0068] 2) Low-frequency path down-sampling and up-sampling feature extraction

[0069] The maximum pooling operation is applied to the input spatio-temporal data block to compress the time dimension, and then the time resolution is restored through the deconvolution operation, outputting the low-frequency path feature, extracting the low-frequency trend component, represented as:

[0070] ,

[0071] In the formula, represents the output feature of the low-frequency path, used to extract the low-frequency trend component of the stress data; represents the deconvolution operation, which maps the compressed feature back to the original time resolution; represents the maximum pooling operation with a pooling size of , that is, the down-sampling is realized by taking the maximum value in the local time window, and the maximum value is taken every 4 consecutive time steps, so that the time dimension is compressed to .

[0072] 3) Multi-scale feature fusion and layer normalization

[0073] The features of each branch of the high-frequency path are spliced with the low-frequency path feature in the channel dimension, and then the layer normalization operation is applied to output the fused multi-scale feature matrix, which is used to stabilize the feature distribution and enhance the training stability, represented as:

[0074] ,

[0075] 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.

[0076] S302, Improved Gated Loop Unit Modeling Long-Term Dependency

[0077] 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.

[0078] 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:

[0079] 1) Spatial attention weight calculation

[0080] 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:

[0081] ,

[0082] 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; The spatial attention weight of the p-th sensor at the t-th time point, the greater the weight, the greater the contribution of the feature of the sensor to the current time instability sign; The attention score of the p-th sensor at the t-th time point; To distinguish from the sensor index p and q, ; The natural exponential function is represented. In an implementation manner, the multi-layer perception machine includes an input layer, a hidden layer and an output layer, the dimension of the input layer is 32, the dimension of the hidden layer is 64, the ReLU activation function is adopted, the dimension of the output layer is 1-dimensional, and the linear activation function is adopted.

[0083] 2) Double reset gate mechanism design

[0084] 2) Double reset gate mechanism design

[0085] Based on the historical hidden state and the current integrated feature vector, the basic reset gate output vector is calculated through the Sigmoid activation function, and at the same time, based on the feature change amount, the dynamic reset gate output vector is calculated through the Sigmoid activation function, the dynamic reset gate output vector is used to control the historical information retention strength, realize the coupling of the feature change amplitude and the gating mechanism, and is represented as:

[0086] ,

[0087] ,

[0088] In the formula, The basic reset gate output vector at the t-th time point controls the reset degree of the historical hidden state; The dynamic reset gate output vector at the t-th time point is used to finally control the historical information retention strength; The Sigmoid activation function is represented, and the output is compressed to the interval ; The weight matrix of the basic reset gate is represented, which is used for linear transformation of the spliced feature vector, and is a trainable parameter; The hidden state vector at the t-1 time point is represented, which is a dimensional feature vector, when t=1, is the initial hidden state vector, and the initial hidden state vector is defined as a zero vector; The weight matrix of the dynamic reset gate is represented, which is used for processing the feature change amount, and is a trainable parameter; The Euclidean distance of the feature vectors of adjacent time points is represented, which quantifies the feature change amplitude, and the calculation method is represented as ; The hidden state vector at the t-1 time point is represented, which is a​ 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.

[0089] 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.

[0090] 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.

[0091] 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.

[0092] S303, Temporal Self-Attention Feature Enhancement

[0093] Conventional attention mechanism treats all time steps equally, but there is often a key triggering period for slope instability, making it difficult for the model to focus on key period features highly related to the instability mode, thereby reducing the evaluation accuracy. The present application dynamically enhances the period features with high correlation to the instability mode by calculating the saliency of each period and adaptive weighting based on the saliency, and then focuses on the key period features, represented as:

[0094] 1) Calculate period saliency

[0095] Based on the hidden state vector of each time step, the projection vector is calculated through the weight matrix and the hyperbolic tangent activation function, then the inner product operation is performed with the reference vector and normalized to obtain the period saliency score, reflecting the correlation between the period and the instability mode, represented as:

[0096] ,

[0097] In the formula, represents the saliency score of the th period, reflecting the correlation between the period and the instability mode; represents the projection vector of the th period feature, which enhances the representation ability through nonlinear transformation, and the calculation method is represented as ; represents the hyperbolic tangent activation function, which constrains the projection value to the interval; is the period saliency weight matrix, which is used for linear transformation of period features, and is a trainable parameter; represents the hidden state vector of the th time step of the th spatiotemporal data block, which is a 128-dimensional feature vector, and is the hidden state vector of the th time step of , i.e., the th row of ; is the period index, ; represents the number of time domain segments, and for a single data block, each time step corresponds to a period, so the number of time domain segments is equal to the time step length of each data block; represents the vector inner product operation; is the reference vector, which characterizes the typical instability mode, and is a trainable parameter, whose value is optimized through training.

[0098] 2) Adaptive weighting

[0099] 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:

[0100] ,

[0101] 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, .

[0102] 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.

[0103] S304, Depthwise Separable Convolutional Feature Compression

[0104] 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.

[0105] This invention achieves feature fusion and dimensionality reduction by using two steps: depthwise convolution and pointwise convolution. This significantly reduces the number of parameters while preserving the spatial structure information of the feature map. The specific steps are as follows:

[0106] 1) Depthwise convolution processing

[0107] 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:

[0108] ,

[0109] In the formula, is the output feature vector of the d-th channel of the deep convolution, with dimension is the output feature vector of the d-th channel of the deep convolution, with dimension is the channel index; is the deep convolution kernel of the d-th channel of the deep convolution, responsible for extracting spatial features of a single channel; is the bias term of the d-th channel of the deep convolution, which is a trainable parameter; is the one-dimensional convolution operation; is the feature vector of the c-th channel of the weighted feature matrix of the d-th spatio-temporal data block, i.e., the d-th column vector of is a 128-dimensional feature vector containing all the feature values at the d-th channel of all time steps.

[0110] 2) Point-wise convolution compression

[0111] Based on the spliced feature matrix output by all channels of the deep convolution, a point-wise convolution weight matrix is applied for one-dimensional convolution operation, combined with a bias term to output a compressed feature matrix, realizing feature fusion and dimension reduction, represented as:

[0112] ,

[0113] In the formula, is the compressed feature matrix of the d-th channel of the point-wise convolution, with dimension Point-wise convolution is the second step of depth separable convolution, which uses a convolution kernel to convolve the output of the deep convolution; is the point-wise convolution weight matrix with convolution kernel size of 1 and output channel number of 32, responsible for feature fusion and dimension reduction between channels, which is a trainable parameter; is the point-wise convolution bias term, which is a trainable parameter; is the spliced feature matrix output by all channels of the deep convolution, i.e., with dimension

[0114] 3) Global average pooling

[0115] Global average pooling is performed on each channel of the compressed feature matrix in the time dimension, outputting a pooling feature vector representing the global statistical features of the entire data block, eliminating the influence of variable time steps, represented as:​​​​​​​​​​​​

[0116] ,

[0117] wherein, is the compressed feature vector of the th time step, i.e. is the th row vector of the 32-dimensional feature vector, representing the feature value set of all channels at the th time step; is the pooled feature vector of the th spatio-temporal data block, which is a 32-dimensional feature vector representing the global statistical features of the entire data block, fusing multi-scale spatio-temporal information and eliminating the influence of variable time steps.

[0118] S305, stability evaluation grade classification

[0119] After feature extraction and compression, the segmented stress data obtains the pooled feature vector, but the conventional classifier directly uses the full connection layer for classification, ignoring the internal structure information of the feature vector and the correlation between classes, which easily leads to insufficient evaluation accuracy for progressive instability.

[0120] The present application enhances the explainability and robustness of classification decision by learning class prototypes and gating feature transformation, and realizes stability evaluation grade classification based on prototype similarity, the specific steps are as follows:

[0121] 1) Gating feature transformation

[0122] Based on the pooled feature vector, the gating value is calculated by the gating weight matrix and the Sigmoid activation function, and the transformation value is calculated by the feature transformation weight matrix and the exponential linear unit activation function, then element-wise multiplication is performed, and the gating transformation feature vector is output, which enhances the feature representation ability, represented as:

[0123] ,

[0124] wherein, represents the gating transformation feature vector of the th spatio-temporal data block, with a dimension of , used to enhance the feature representation ability; represents the exponential linear unit activation function, processing negative input; represents the gating weight matrix, with a dimension of , which is a trainable parameter; represents the gating bias vector, with a dimension of , which is a trainable parameter; represents the feature transformation weight matrix, with a dimension of , which is a trainable parameter; This represents the feature transformation bias vector, with dimension . , are trainable parameters; This indicates element-wise multiplication.

[0125] 2) Prototype Similarity Classification

[0126] 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 follows:

[0127] ,

[0128] 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 dimension [missing information]. , are trainable parameters; To distinguish from Category index, .

[0129] 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.

[0130] S306, Loss Function Calculation

[0131] 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.

[0132] 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:

[0133] 1) Calculation of weighted focus loss

[0134] Based on the output probability and the true label, the weighted focal loss is calculated by combining the class weight and the focus parameter, which dynamically reduces the loss contribution of easy-to-classify samples and focuses on difficult-to-classify samples, and is expressed as:

[0135] ,

[0136] In the formula, the weighted focal loss is represented, which makes the model pay more attention to difficult-to-classify samples; represents the weight of the th class, which is used to balance the class frequency and is calculated as ; represents the total number of training samples; represents the number of samples in the th class; represents the true label of the th data block, which is in oneHot encoding format. When the true class is the th class, , otherwise ; represents the focus parameter, which is preferably set to , and is used to adjust the weight of easy and difficult samples.

[0137] 2) Attention smoothing loss calculation

[0138] Based on the spatial attention weight and the time period significance score, the difference square sum of adjacent time steps is calculated, and then the attention smoothing loss is output by combining the smoothing coefficient, which encourages the smooth change of attention distribution, and is expressed as:

[0139] ,

[0140] In the formula, the attention smoothing loss is represented, which constrains the temporal consistency of attention weights, making the model more stable in modeling the spatiotemporal characteristics of stress wave propagation and reducing misjudgment caused by abnormal fluctuations; represents the spatial attention smoothing coefficient, which is preferably set to ; represents the time period significance smoothing coefficient, which is preferably set to ; represents the spatial attention weight of the th sensor at the t time; represents the spatial attention weight of the th sensor at the t time; represents the significance score of the th time period; represents the significance score of the th time period.

[0141] 3) Total loss function calculation

[0142] The weighted focal loss and the attention smooth loss are added to obtain a total loss function for optimizing the model parameters.

[0143] S307, execute model training process

[0144] The model training process is executed based on the labeled stress data. First, the data set is divided into a training set and a validation set. The training set is used for model parameter optimization, and the validation set is used to monitor the training process and prevent overfitting.

[0145] The training process adopts the stochastic gradient descent method, aiming to minimize the total loss function, and iterates multiple rounds;

[0146] In each iteration, a batch of data is randomly sampled from the training set, input into the model for forward propagation, and the output probability and total loss function are calculated. Then, the gradient is calculated through the back propagation algorithm, and all trainable parameters are updated;

[0147] During the training process, the changes of training loss and validation loss are recorded. When the validation loss no longer significantly decreases in consecutive multiple iteration periods or reaches the preset maximum number of iterations, the training is stopped, indicating that the model has converged or reached the best performance;

[0148] In addition, the early stopping strategy is used to avoid overfitting and ensure the generalization ability of the model;

[0149] After training is completed, the optimal model parameters are saved for subsequent evaluation of embankment slope stability.

[0150] S4, embankment slope stability evaluation

[0151] After the model training is completed, for newly collected embankment slope stress sensor data, 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, ensuring the complete spatiotemporal correlation of stress wave propagation;

[0152] Then, each spatiotemporal data block is input into the trained stability evaluation model. The model outputs the probability of each data block belonging to stable, critical, or unstable state through steps such as multi-scale convolution feature extraction, improved gated recurrent unit modeling long-term dependence, time domain self-attention feature enhancement, deep separable convolution feature compression, and stability evaluation grade classification.

[0153] Finally, the highest probability or preset threshold is used to determine the stability evaluation grade of the slope, and the evaluation result is output to provide a basis for engineering decision-making, such as timely warning of potential instability risk or confirming the safety state of the slope.

[0154] Embodiment 2

[0155] In this embodiment, as shown in Figure 2 The advantages of the method of the present application in feature learning are demonstrated by feature distribution visualization, which shows the high-dimensional features on a two-dimensional plane through dimension reduction techniques, and intuitively presents the distribution of stable, critical and unstable states in the feature space. The good distinguishability of feature distribution is the basis for accurate classification of the model and an important basis for evaluating feature extraction capability. From the scatter plot, it can be clearly observed that the three states form relatively independent aggregation areas in the feature space. The feature points of the stable state are mainly concentrated in the lower left area of the graph, with green color; the feature points of the critical state are distributed in the middle area, with orange color; the feature points of the unstable state are concentrated in the upper right area, with red color. The distribution pattern is consistent with the actual physical process of the embankment slope, and the stable state to the unstable state is a continuous change process, and the critical state as a transition stage, its feature distribution will indeed have partial overlap with the other two states. Each point in the figure 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, the feature dimension 1 corresponds to the curvature change feature of the sum of the maximum values of mutual information, by calculating the mutual information values of all sensor pairs within the preset time lag range and analyzing the statistical characteristics, and the feature dimension 2 is derived from the multi-scale convolution feature extraction and the hidden state feature of the output of the gated recurrent unit, which can effectively represent the intensity change of the spatiotemporal correlation after compression by deep separable convolution. Although the points of the three categories have certain overlapping areas, they form a clear distribution trend as a whole, indicating that the features learned by the method of the present application have good class distinguishability, especially the critical state as the most important early warning stage, its distribution area is relatively concentrated, which is conducive to the model to accurately identify the early signs of slope instability.

[0156] Example 3

[0157] In this embodiment, the dynamic performance of different methods in the actual monitoring process is compared through time series analysis. As shown in Figure 3As shown, the experimental simulation shows the stability change process of the complete embankment slope in an ideal state, which gradually transitions from the initial stable state to the critical state, and finally enters the unstable state. Time series analysis can directly show the detection ability and response speed of each method at the key point of state transition. From the time series graph, three main curves can be observed: the black dotted line represents the true 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. The background of the three colors clearly divides the time interval of the stable region (green), the critical region (orange), and the unstable region (red). The true state shows that the first quarter of the time period is in a stable state, the middle third of the time period is in a critical state, and the last quarter of the time period 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 to extract local features, and then combines a basic gated recurrent unit to model the time dependence. Comparing the two prediction curves, it can be found that the conventional method has a significant detection delay at the state transition point, especially at the transition point from stable to critical. The response of the conventional method is significantly delayed. The method of the present application can capture the state change more timely, and the prediction curve has a higher degree of agreement with the true state curve. In the stable stage, both prediction curves are relatively stable; in the critical stage, the prediction value of the method of the present application fluctuates less, showing better stability; in the unstable stage, the method of the present application can quickly identify the state change. The vertical coordinate corresponds to the stable, critical, and unstable states, and the horizontal coordinate time step represents the continuous sampling point sequence.

[0158] Example 4

[0159] In this embodiment, as shown in Figure 4 , the critical state detection performance comparison analysis is carried out to verify the advantages of the model of the present application in the most challenging critical state (instability precursor) detection. Comparing the long short-term memory network model, the convolutional neural network-gated recurrent unit combined model, the transformer model, and the model of the present application, the evaluation indicators include accuracy, recall rate, and F1 score. The same 5-sensor data set is used in the experiment configuration, and the critical state samples (accounting for 15% of the data set) are focused on. The grouped column chart shows that the model of the present application leads other models in the recall rate indicator, indicating that it has the lowest missed detection rate, and the spatial attention mechanism enables the model to focus on key position sensor signals and timely capture mutation features such as material fracture. In addition, in the comparison of accuracy and F1 score, the model of the present application is higher than the comparison algorithms.

[0160] Example 5

[0161] In this embodiment, as shown in Figure 5As shown in the figure, the trend of the change of the loss function of the model during the training process is analyzed, and the convergence characteristics and training stability of the method of the application and the conventional long short-term memory network method are compared. The experiment shows the relationship between the training round and the loss value through the broken line graph. The unit of the training round is times, indicating the progress of the model iterative training. The unit of the loss value is a dimensionless value, reflecting the size of the model prediction error. The lower the value is, the better the model performance is. As shown in the figure, the loss curve of the method of the application quickly decreases at the beginning of the training, quickly reaches and maintains at a lower level, and the whole curve fluctuates less. The loss curve of the conventional method decreases slowly and finally stays at a higher position, and the fluctuation is relatively large. It shows that the method of the application has faster convergence speed and better training stability, and can more efficiently learn the feature mode of slope instability. The experimental results show that the multi-scale convolution feature extraction and depth separable convolution feature compression technologies in the method of the application effectively improve the efficiency of feature extraction and the robustness of model training.

[0162] Embodiment 6

[0163] In this embodiment, the influence of the number of sensors on the evaluation performance is analyzed to verify the adaptability and practicability of the method of the application under different engineering configurations. As shown in the figure, Figure 6 the experiment shows the trend of the change of the evaluation accuracy when the number of sensors changes from 2 to 7 through the broken line graph combined with the scatter plot. The horizontal coordinate represents the number of sensors, and the unit is the number. The vertical coordinate represents the evaluation accuracy, and the unit is a dimensionless decimal value. As can be seen from the figure, with the increase of the number of sensors, the accuracy of the two methods shows an upward trend, but the method of the application maintains a high accuracy under all sensor number configurations, especially when the number of sensors is small. When the number of sensors reaches 5 or more, the accuracy of the method of the application tends to be stable and maintains at a very high level, while the accuracy of the conventional method increases limitedly even if the number of sensors is increased, and the overall level is low, which reflects that the method of the application has low dependence on the number of sensors and can still realize reliable evaluation under the condition of limited sensor deployment, indicating that the spatio-temporal segmentation technology based on mutual information can fully utilize the spatio-temporal correlation in the sensor network and effectively capture the key features of stress wave propagation even if the number of sensors is small.

[0164] Finally, it should be noted that: the above only describes the preferred embodiments of the application and is not used to limit the application. Although the application 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 equivalently replace some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. An embankment slope stability evaluation method based on artificial intelligence, characterized by, The method comprises the following steps: Road embankment slope stress sensor data acquisition and labeling; based on the mutual information value of the sensor, the road embankment slope stress sensor data is divided into multiple spatio-temporal data blocks; A stability evaluation model is constructed; multi-scale convolution features in the spatio-temporal data block are extracted by using the multi-scale cavity convolution of the high-frequency channel and the pooling-anti-convolution operation of the low-frequency channel, and a fusion multi-scale feature matrix is obtained; Through the spatial attention mechanism, the sensor weight is dynamically allocated, and the double reset gate mechanism is used to adaptively adjust the dependence degree of historical information, and the hidden state sequence of each spatio-temporal data block is obtained; the specific steps are as follows: Based on the input feature vector of each sensor at each time step, the attention score is calculated through a multi-layer perception, and the attention score is normalized to a spatial attention weight by using a softmax function, so as to realize the dynamic focusing of the key sensor; Based on the historical hidden state and the current integrated feature vector, the basic reset gate output vector is calculated through a Sigmoid activation function, and based on the basic reset gate output vector and the feature change amount, the dynamic reset gate output vector is obtained through a Sigmoid activation function, which is used to control the retention intensity of historical information, realize the coupling of feature change amplitude and gating mechanism; based on the standard gating cycle 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; Based on the hidden state vector at each time step in the hidden state sequence, the period significance is calculated, the weighted feature vector is obtained based on the period significance score, and then the weighted feature matrix of each spatio-temporal data block is obtained; the weighted feature matrix is processed through deep convolution and point-by-point convolution to obtain the pooling feature vector of each spatio-temporal data block; the stability evaluation level classification is performed through the learnable class prototype and the gating feature transformation. In the training process, the total loss function obtained by adding the weighted focal loss and the attention smoothing loss is used to dynamically adjust the sample weight and constrain the attention distribution. 2.The method of claim 1, wherein, Road embankment slope stress sensor data acquisition and labeling: Stress sensors are deployed at key positions of the road embankment slope to collect stress values at a fixed sampling frequency, generate time series data, and obtain road embankment slope stress sensor data; based on the actual stability state of the road embankment slope, the labeled categories are divided into stable state, critical state and unstable state. 3.The method of claim 1, wherein, Based on the mutual information value of the sensor, the road embankment slope stress sensor data is divided into multiple spatio-temporal data blocks: The joint empirical probability distribution and the marginal empirical probability distribution of the road embankment slope stress sensor data are estimated by using histogram statistics sample frequency, and the mutual information value of the sensor pair is obtained; based on the curvature change of the sum of the mutual information maximum values of all sensor pairs, the second derivative at the time point is calculated and compared with the curvature threshold, the key time points of stress wave propagation mode mutation are identified, and the segmentation boundary point set is dynamically determined; the road embankment slope stress sensor data is divided into multiple spatio-temporal data blocks according to the segmentation boundary point set. 4.The method of claim 1, wherein, The specific process of obtaining the fusion 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 method of claim 1, wherein, 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 method of claim 1, wherein, 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 method of claim 1, wherein, 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 method of claim 1, wherein, 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 method of claim 1, wherein, 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 method of claim 1, wherein, Based on the spatial attention weight and the time period saliency score, the difference square sum of adjacent time steps is calculated, and the attention smoothing loss is output in combination with the smoothing coefficient.

Images