Expansive soil landslide displacement prediction method, system, equipment and medium

Abstract

The invention provides an expansive soil landslide displacement prediction method, system and device and a medium, and belongs to the technical field of geological disaster monitoring and prediction.The method comprises the following steps that real-time time sequence data of a plurality of landslide influence factors of a to-be-predicted area are obtained; the method comprises the following steps: taking a neural network model Transform as a basic model, and replacing an embedded layer of the Transform by using a time sequence convolutional network TCN to obtain a hybrid prediction model TCN-Transform; a physical loss function is designed on the basis of the expansive soil constitutive equation, and TCN-Transformer is trained; and inputting real-time time sequence data into the trained TCN-Transformer, splicing features at corresponding moments by using the TCN, carrying out feature mixing through a full connection layer, and carrying out linear projection on the mixed features to obtain the landslide displacement. The landslide displacement prediction precision can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of geological disaster monitoring and prediction, and particularly relates to a method, system, device and medium for predicting the displacement of expansive soil landslides. Background Art

[0002] Landslide geological disasters are major challenges faced by China's infrastructure construction and national security. According to the data of the 2022 annual geological disaster bulletin of the Ministry of Natural Resources, landslide disasters in China account for 72.6% of the total number of geological disasters. Among them, due to the special engineering geological characteristics of expansive soil landslides, the annual direct economic loss caused in key engineering sections such as the Yangtze River Basin and the southwestern region exceeds 2 billion yuan. The unique montmorillonite mineral composition of expansive soil endows it with significant water sensitivity. The crack expansion-healing mechanism generated under the action of wet-dry cycles leads to the displacement evolution showing multi-stage jump characteristics, which poses a severe challenge to traditional prediction theories.

[0003] With the improvement of the "space-air-ground" integrated monitoring system, the new intelligent sensor network can collect multi-dimensional characteristic parameters (displacement, earth pressure, water content, etc.) up to the TB level per hour. At present, in combination with the collected data, the machine learning model LSTM network or the pure data-driven model Transformer architecture is used to predict the displacement of expansive soil landslides. However, both of them have certain defects, which are specifically as follows:

[0004] (1) Defects of machine learning models: Although the LSTM network can handle time series dependencies, its gating mechanism shows gradient dispersion after more than 50 time steps, and physical constraints such as the Mohr-Coulomb yield criterion of expansive soil are not considered;

[0005] (2) Distortion of pure data-driven models: When the existing Transformer architecture is directly applied in landslide prediction, its global attention mechanism will over-smooth local mutation features.

[0006] In summary, the defects existing in the above two methods result in low prediction accuracy of expansive soil displacement. Summary of the Invention

[0007] In order to overcome the deficiencies of the above-mentioned existing technologies, the present invention provides a method for predicting the displacement of expansive soil landslides, including the following steps:

[0008] Obtain the real-time time series data of multiple landslide influencing factors in the area to be predicted;

[0009] Based on the neural network model Transformer, use the temporal convolutional network TCN to replace the embedding layer of Transformer to obtain the hybrid prediction model TCN-Transformer;

[0010] Design a physical loss function based on the constitutive equation of expansive soil, and use the physical constraint loss function to train the TCN-Transformer;

[0011] Input the real-time time series data of multiple landslide influencing factors into the trained TCN-Transformer. Use the TCN to splice the features at the corresponding moments, perform feature mixing through the fully connected layer, and then input the mixed features into the self-attention layer of the Transformer for linear projection and combined output to obtain the landslide displacement.

[0012] Preferably, before inputting the real-time time series data of multiple landslide influencing factors into the trained TCN-Transformer, it also includes preprocessing the real-time time series data, specifically including the following steps:

[0013] Resample and perform wavelet transform on the real-time time series data of multiple landslide influencing factors at different time intervals to remove the noise data in the real-time time series data;

[0014] Use the Grey Relational Analysis (GRA) to calculate the correlation degree between multiple landslide influencing factors and displacement, construct dynamic lag features according to the correlation degree between multiple landslide influencing factors and displacement, and input the dynamic lag features into the trained TCN-Transformer for processing.

[0015] Preferably, the total loss function of the TCN-Transformer consists of the weighted mean square error loss function and the physical constraint loss function as follows:

[0016]

[0017] where λ is the weight;

[0018] In the formula, the weighted mean square error loss function is as follows:

[0019]

[0020] where is the predicted displacement on the i-th day, is the actual displacement on the i-th day, l is the prediction duration, and w i is the dynamic allocation weight of the prediction length;

[0021] The physical constraint loss function is as follows:

[0022]

[0023] where ΔD pred$(t)$ is the predicted displacement increment, $\epsilon(t)$ is the theoretical increment at the $i$-th step calculated based on historical humidity and soil pressure, and $l$ is the predicted time duration.

[0024] Preferably, a physical weight bias is introduced into the self-attention layer of the Transformer and obtained through the physical weight bias.

[0025] Preferably, the real-time time-series data of the multiple landslide influencing factors include rainfall and soil pressure.

[0026] The present invention also provides an expansive soil landslide displacement prediction system, including:

[0027] A data acquisition module for acquiring the real-time time-series data of multiple landslide influencing factors in the area to be predicted;

[0028] A model construction module for using the time series convolutional network TCN to replace the embedding layer of the Transformer with the neural network model Transformer as the basic model to obtain the hybrid prediction model TCN-Transformer;

[0029] A model training module for designing a physical loss function based on the expansive soil constitutive equation and training the TCN-Transformer using the physical constraint loss function;

[0030] A landslide displacement prediction module for inputting the real-time time-series data of multiple landslide influencing factors into the trained TCN-Transformer, using the TCN to splice the features at the corresponding moments, performing feature mixing through a fully connected layer, and then inputting the mixed features into the self-attention layer of the Transformer for linear projection and combined output to obtain the landslide displacement amount.

[0031] The present invention also provides a computer device, including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the expansive soil landslide displacement prediction method.

[0032] The present invention also provides a computer-readable storage medium, which stores a computer program, and the computer program is suitable for being loaded by a processor to execute the expansive soil landslide displacement prediction method.

[0033] The expansive soil landslide displacement prediction method provided by the present invention has the following beneficial effects:

[0034] By using the constitutive equation of expansive soil to design a physical loss function and using the physical constraint loss function to train the TCN-Transformer, the present invention can ensure that the prediction results conform to the mechanical laws and improve the prediction accuracy. By using the temporal convolutional network (TCN), the features at the corresponding moments of the input real-time temporal data can be spliced, and feature mixing can be performed through a fully connected layer. The TCN extracts local temporal patterns, reduces noise sensitivity, and improves the prediction accuracy. By inputting the mixed features into the self-attention layer of the Transformer, the features at key time points can be focused on, realizing high-precision and high-confidence prediction of the short-term displacement of expansive soil landslides, and providing an interpretable decision basis for disaster warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more clearly illustrate the embodiments of the present invention and their design schemes, the accompanying drawings required for the present embodiments will be briefly introduced below. The accompanying drawings in the following description are only partial embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a flowchart of the method for predicting the displacement of expansive soil landslides according to the embodiments of the present invention.

[0037] Figure 2 It is the wavelet threshold denoising process;

[0038] Figure 3 It is the comparison between Fourier transform and wavelet transform;

[0039] Figure 4 It is the threshold quantization of the high-frequency coefficients of wavelet decomposition; where Figure 4 (a) of is the hard threshold quantization; Figure 4 (b) of is the soft threshold quantization;

[0040] Figure 5 It is the variant structure of TCN;

[0041] Figure 6 It is the TCN-Transformer model structure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] In order to enable those skilled in the art to better understand the technical solutions of the present invention and implement them, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0043] Embodiment

[0044] The present invention provides a method for predicting the displacement of expansive soil landslides, specifically as Figure 1As shown, it includes the following steps:

[0045] Step 1: Obtain the real-time time-series data of multiple landslide influencing factors in the area to be predicted.

[0046] First, it is necessary to preprocess the real-time time-series data, which specifically includes the following steps:

[0047] (1) Resampling and interpolation.

[0048] Data resampling refers to resampling time-series data at different time intervals. By data acquisition, the time interval of the received data points can be modified to only obtain key information, eliminate noise data, and improve data visualization effects. At the same time, resampling also helps to align multiple data sources from different time intervals to ensure consistency when creating visualizations or conducting analyses. According to requirements, upsample or downsample the frequency of the time-series data.

[0049] Upsampling is to increase the number of minority class samples to balance the classes in the dataset. Upsampling maintains the information of the majority class samples, increases the model's recognition ability for the minority class, but may lead to overfitting. Common methods include:

[0050] Repeated sampling: Randomly copy minority class samples to increase their number to be comparable to that of the majority samples.

[0051] Synthetic minority over-sampling technique (SMOTE): Generate new samples between minority samples through interpolation to increase the number of minority class samples.

[0052] Downsampling is to reduce the number of majority class samples to balance the classes in the dataset. Downsampling reduces the amount of data, making the training speed faster and also balancing the classes in the dataset, but it may lose valuable information and may lead to a decrease in the model's recognition ability for the majority class. Common methods include:

[0053] Random undersampling: Randomly remove majority class samples to reduce their number to be comparable to that of the minority class samples.

[0054] Ensemble method: Use a combination of multiple classifiers, and each classifier is trained on different downsampled datasets to improve the generalization ability of the model.

[0055] For the missing values in the original data, the method of spline interpolation is usually used for processing. For the function f(x) to be interpolated, given the function values at the known nodes x0, x1,..., x n at, segment the adjacent two nodes to obtain n interpolation subintervals, and use a k-degree polynomial S i (x) for interpolation in each interval to make it satisfy the interpolation conditions and k - 1 order smoothness:

[0056] S i (x) = f(x i ), S i (x i+1 ) = f(x i+1 ), 0 ≤ i ≤ n - 1;

[0057]

[0058] Among them, S i (x) is the function to be interpolated at node x, f(x i ) is the interpolation polynomial at node x i , S i (x i+1 ) is the interpolation polynomial at node x i+1 , is the j - th interpolation polynomial of node x i+1 in the i - th interpolation interval, is the j - th interpolation polynomial of node x i+1 in the (i + 1) - th interpolation interval.

[0059] After combining the interpolation results of each segment, it is spline interpolation. The commonly used one is cubic spline interpolation. Assume S i (x) is a cubic polynomial:

[0060] S i (x) = a i + b i x + c i x 2 + d i x 3 ;

[0061] In the formula, a i , b i , c i and d i are all polynomial parameters.

[0062] It can be seen from the above formula that a cubic spline segment contains 4 unknowns. Then, the cubic splines corresponding to nodes x0, x1,..., x n contain 4n unknowns, while the conditions satisfied by cubic spline interpolation itself can only form 4n - 2 equations, which is less than the number of unknown parameters 4n. Additional conditions must be added, such as:

[0063] Natural boundary: S″(x0) = S″(x n ) = 0;

[0064] Fixed boundary: S′(x0) = c1, S′(r n ) = c2;

[0065] Periodic spline: S′(x0) = S′(xn ), S″(x0) = S″(x n );

[0066] Not-a-knot boundary: S″′(x0) = S″′(x1), S″′(x n-1 ) = S″′(x n ).

[0067] Where, S′(x) is the first derivative of the function to be interpolated, S″(x) is the second derivative of the function to be interpolated, S″′(x) is the third derivative of the function to be interpolated, and x is x0, x n-1 or x n ; c1 and c2 are constants.

[0068] (2) Wavelet threshold denoising.

[0069] Wavelet Threshold Denoising (WTD) realizes the function of removing noise components and leaving key information through decomposition + selective reconstruction. Its process is as Figure 2 shown.

[0070] Wavelet transform represents a signal through a series of scalable and translatable basis functions, which are called wavelet functions. Compared with the traditional Fourier transform, wavelet transform replaces the infinitely long trigonometric function basis with a finitely long decaying wavelet basis, as Figure 3 shown:

[0071]

[0072] Where, f(t) is the signal with unknown spectrum, F(ω) is the spectrum function after Fourier transform, its independent variable is the ω frequency, e -jωt is a complex signal, (α, τ) is an arbitrary pair of real numbers, WT(α, τ) is the frequency-time function after wavelet transform, the parameter α is the scale factor (i.e., the frequency ω in FT), which is responsible for adjusting the stretching degree of the wavelet function and corresponds to the frequency information, t is the integration variable, the parameter τ is the translation factor or time center parameter (i.e., the time in Fourier transform), which is responsible for adjusting the translation degree of the wavelet function and corresponds to the time information, and ψ is the wavelet function.

[0073] Wavelet transform includes wavelet decomposition and wavelet reconstruction. It can decompose a relatively complex signal into several simple signals for processing, and the several simple signals after processing can be combined into a signal that is more complete than the original signal.

[0074] The wavelet coefficients obtained by wavelet transform of the signal contain important time-frequency information. The wavelet coefficients of real information are larger, and the wavelet coefficients of noise are smaller. Applying a threshold rule to the wavelet coefficients, as Figure 4As shown, where Figure 4 (a) of Figure 4 is hard threshold quantization; Figure 4 (b) of Figure 4 is soft threshold quantization. Coefficients less than a certain threshold are regarded as noise and set to zero or their values are reduced.

[0075] Among them, the hard threshold is:

[0076]

[0077] The soft threshold is:

[0078]

[0079] Among them, W ij is the wavelet coefficient obtained by decomposition, λ is the threshold, is the processed wavelet coefficient, and sgn(·) represents the sign function.

[0080] Reconstruct the processed signal to obtain the denoised signal.

[0081] (3) Dynamic lag feature construction.

[0082] The nonlinear relationship between the moisture diffusion coefficient of expansive soil and the displacement response exhibits time-varying lag characteristics. It is difficult to accurately establish a dynamic correlation model for multi-source heterogeneous data using conventional cross-correlation analysis methods. Therefore, this invention uses grey relational analysis (GRA) to calculate the correlation degree between each factor and the displacement, specifically through the following formula:

[0083]

[0084] Among them, γ(X, Y) is the correlation degree, Δ i is the absolute difference between the corresponding elements of the comparison sequence (each factor) and the reference sequence (displacement), n is the number of evaluated objects, ρ is the resolution coefficient, usually taking 0.5, select the lag time (such as rainfall lagging 12 days, displacement autocorrelation lagging 30 days), X(i) is the comparison sequence, and Y(i) is the reference sequence. Input feature: Construct multi-dimensional time series input X∈R T×d .

[0085] Step 2: Based on the neural network model Transformer as the basic model, use the temporal convolutional network TCN to replace the embedding layer of Transformer to obtain the hybrid prediction model TCN-Transformer.

[0086] Transformer is a neural network architecture that is completely based on the attention mechanism (Self-Attention), abandoning the traditional recurrent (RNN) and convolutional (CNN) structures. It can capture long-range dependencies and avoid the sequential computation of RNNs, significantly improving the training speed. As one of the most important breakthroughs in the field of deep learning in recent years, it has almost become the "standard" model, especially in natural language processing (NLP).

[0087] Transformer is stacked by an encoder and a decoder, and can well handle sequences with variable lengths for both input and output. However, in landslide prediction, both the input and output are usually short sequences with fixed lengths, and the encoder-decoder architecture does not bring significant advantages but increases the computational cost. Therefore, this invention only refers to the structure of its encoder and directly performs a fully connected mapping at the end to obtain the result. The following is a detailed introduction to the composition of the model result in this invention.

[0088] (1) TCN encoder.

[0089] Landslide displacement prediction is essentially a regression problem, and the embedding layer in the traditional Transformer encoder is not suitable for feature extraction in regression problems. Therefore, the original model needs to be modified. Considering that the samples at a single moment cannot effectively describe the movement state of the landslide at the current moment, the temporal convolutional network (TCN) is used to replace the embedding layer.

[0090] TCN is a convolutional neural network (CNN) architecture specifically designed for processing time series data. It improves the traditional convolutional operation to make it more suitable for sequence modeling tasks (such as time series prediction, speech recognition, natural language processing, etc.). The core of TCN is the combination of causal convolution and dilated convolution, and the introduction of residual connections to capture long-term dependencies in time series and improve the model effect. Using TCN can introduce length-controllable historical information into the features of the current moment while keeping the input form unchanged, and retain the parallelizability of the Transformer model training, taking into account both the actual effect of the model and ensuring the model training speed.

[0091] Considering that the long-term memory part in the prediction model is completed by the self-attention mechanism and TCN does not need to consider historical information over too long a time, dilated convolution is not introduced. The TCN structure that retains causal convolution and residual connections is used in the model, as shown in Figure 5 shown. The detailed steps are as follows.

[0092] ① Causal convolution.

[0093] In TCN, causal convolution aligns the output of the convolution kernel with the input in the time dimension, ensuring that when the model predicts at time step t, it only depends on historical information (t′≤t) and avoids leakage of future information, making it suitable for time series autoregressive scenarios. Let the input sequence be The output is The convolution kernel is where k is the size of the convolution kernel. The calculation of the t-th time step of the output Y of the causal convolution is:

[0094]

[0095] In the formula, Y t is the output at time step t, W i is the i-th convolution kernel, X t-i+(k-1) is the input corresponding to time step t - i+(k + 1). To ensure the causal relationship, (k - 1) zeros need to be padded on the left side of the input so that the convolution kernel only covers the current and past time steps.

[0096] ② Weight normalization.

[0097] In TCN, to accelerate the training convergence and replace batch normalization (BatchNorm), the convolution kernel weight W is decomposed into a direction vector v and a scalar gain g, and the normalized weight is:

[0098]

[0099] where ||v|| is the L2 norm of v.

[0100] ③ ReLU activation function.

[0101] Introduce non-linearity, the formula is:

[0102] Y ReLU = ReLU(Y)= max(0, Y);

[0103] ④ Random dropout.

[0104] During training, some neurons are randomly set to zero with probability p to prevent overfitting:

[0105] Y aropout = Dropout(Y ReLU , p)

[0106] where Y dropout represents the output after random dropout.

[0107] ⑤ Residual connection and 1x1 convolution.

[0108] To solve the problem of deep network training and enhance the modeling ability for long-term dependencies in time series, residual connections are introduced. Let the input be X and the output of the residual path be F(X), then the final output is:

[0109] Y = X + F(X);

[0110] If the dimension needs to be adjusted, a 1x1 convolution is applied to the input X, and the formula is:

[0111] Y = Conv1x1(X) + F(X).

[0112] Here, F(X) is the output Y after dropout dropout .

[0113] (2) Transformer encoder.

[0114] The attention mechanism combines the query (autonomous prompt) and the key (non-autonomous prompt) through attention pooling to achieve a selective tendency for the value (sensory input). The overall structure of the TCN-Transformer model used in this model is as Figure 6 shown. After the sequence data such as landslide displacement and rainfall pass through the TCN respectively, the features at the corresponding time steps are concatenated, mixed through a fully connected layer, and then input into the self-attention layer of the Transformer after adding positional encoding to output the hidden state at the current time step, and finally the landslide displacement at the next time step is output through a fully connected layer.

[0115] ① Feature concatenation.

[0116] The sequence data such as landslide displacement and rainfall are processed through independent TCN layers respectively, and then the features are concatenated in the time step dimension and the features are mixed through a fully connected layer.

[0117] 1. Independent TCN processing.

[0118] Each feature such as displacement and rainfall is separately passed through the TCN to extract local temporal patterns.

[0119] Output feature dimension: Assume that each TCN output is where T is the time step and d i is the number of channels for each feature.

[0120] 2. Time step concatenation.

[0121] The TCN outputs of different features are concatenated in the channel dimension:

[0122]

[0123] In the formula, X feature1 , X feature2 , …, Xfeature n respectively represent each feature tensor, X concat represents the tensor after feature concatenation.

[0124] 3. Fully connected feature mixing.

[0125] Use a fully connected layer to map the high-dimensional concatenated features to a low-dimensional space d is the number of feature channels after feature mixing:

[0126]

[0127] Among them, X fuesd represents the tensor after feature mixing, X concat represents the tensor after feature concatenation, W fused represents the feature mixing weight, b fused represents the feature mixing bias term, and ReLU(·) represents the ReLU activation function.

[0128] ② Learnable position encoding.

[0129] Landslide monitoring data may have non-uniform time intervals (such as missing due to equipment failures), and the fixed sin / cos encoding in the original model cannot adapt. Therefore, a learnable position encoding method is adopted, and its formula is as follows:

[0130] X embed = X fused + P,

[0131] Among them, P represents the position feature of each time step learned, which can be learned during training, X embed represents the feature tensor with learnable position encoding added, X fuesd represents the tensor after feature mixing.

[0132] ③ Multi-head self-attention mechanism.

[0133] 1. Single-head attention calculation;

[0134] This model uses the self-attention mechanism, which can make the distance between any two positions in the sequence be represented by a constant and has better parallelism.

[0135] Linear projection: Use self-attention in the Transformer encoder to map the input into query (Query), key (Key), and value (Value) matrices:

[0136]

[0137] Among them, are the learnable weight parameters of the query, key, and value of the i-th attention head respectively, dk = d / h, where h is the number of attention heads. In the original model, h = 4, Q i , K i and V i are the query matrix, key matrix, and value matrix respectively.

[0138] Attention score calculation: For Q i and K i obtained for each attention head i, perform attention operation to get head i . In the original Transformer, dot product scaling is adopted, and physical weights are introduced in this step to enhance the attention to key physical events.

[0139]

[0140] where β is a learnable parameter, W physics The calculation formula synthesizes the coupling effects of comprehensive humidity H(t), earth pressure P(t), and rainfall R(t) as follows:

[0141]

[0142] where R norm (t) represents the normalized rainfall, amplifying the impact of actual rainfall; H0 is the swelling humidity threshold, which needs to be measured according to the properties of local expansive soil; k is a learnable parameter; the Sigmoid function limits the physical weight to [0, 1] to avoid numerical instability.

[0143] 2. Multi-head attention combination.

[0144] Concatenate the outputs of h attention heads and then perform a linear projection:

[0145] X attention = MultiHead(Q, K, V) = Concat(head1,... head h )W O ;

[0146] where is a learnable weight parameter.

[0147] ④ Feed-forward network.

[0148] Perform a non-linear transformation on the self-attention output X attention to enhance the model's expressive ability, specifically through the following formula:

[0149] X FFC = FFN(X out ) = ReLU(X out W1 + b1)W2 + b2;

[0150] Among them, X out represents the output of attention aggregation after residual connection and layer normalization; are respectively learnable weight parameters, b1 and b2 are respectively learnable bias parameters.

[0151] ⑤Residual connection and layer normalization.

[0152] Add residual connections and layer normalization after the multi-head attention mechanism and feedforward network (the output is X out X final ), its purpose is to prevent the gradient from disappearing and accelerate the training convergence:

[0153] X out =LaycrNorm(X embed +X attention )

[0154] X final =LayerNorm(X out +X FFC );

[0155] In this step, the present invention transforms the original Transformer model and uses Post-LayerNorm (post-layer normalization) in sequence, that is, residual connection and layer normalization are applied after each sub-layer (self-attention and feed-forward network).

[0156] ⑥Output layer.

[0157] The output layer directly converts the global features extracted by the TCN-Transformer encoder into the predicted displacement value for the next t days through a simple fully connected mapping.

[0158] Y pred =X final W pred +b pred ;

[0159] Where W pred ∈R d×t , b pred ∈R t are the learnable weight parameters and bias parameters, and the output is Y pred ∈R t (Predict the displacement in the future t days).

[0160] Step 3: Design a physical loss function based on the constitutive equation of expansive soil and train the TCN-Transformer using the physical constraint loss function.

[0161] (1) Model training.

[0162] 1. Loss function:

[0163] The total loss of the model consists of the weighted mean square error (MSE) and the physical constraint loss (L1):

[0164] Force the displacement increment predicted by the model to conform to the constitutive equation, ensuring that the constrained predicted displacement increment ΔD pred is consistent with the theoretical value ∈(t). The total loss function is as follows:

[0165]

[0166] where λ is an adjustable weight, generally taken as 0.2, and the weighted MSE loss function in the formula is as follows, and the weight w is dynamically allocated according to the prediction length i (such as higher weight for recent periods):

[0167]

[0168] where is the predicted displacement on the i-th day, is the actual displacement on the i-th day, and l is the prediction duration;

[0169] The physical loss function in the formula The calculation formula is as follows:

[0170]

[0171] where ΔD pred (t) is the predicted displacement increment, and ∈(t) is the theoretical increment at the i-th step calculated based on historical humidity and soil pressure, and its calculation formula is:

[0172]

[0173] In the formula, k is a learnable parameter, H(t) is the comprehensive humidity, H0 is the swelling humidity threshold, and P(t) is the soil pressure.

[0174] 2. Optimizer and learning rate scheduling.

[0175] Optimizer: AdamW, parameter update formula:

[0176]

[0177] where are the first-order and second-order moment estimates after bias correction respectively; ∈ is a very small parameter used to prevent division by zero; the initial learning rate η = 1×10 -4 , θ t+1 , θ t represent the parameters before and after optimization respectively.

[0178] Learning rate scheduling: Cosine Annealing:

[0179]

[0180] where η t is the learning rate at training batch t, and η max = 1×10 -4 , η min = 1×10 -5 , and T epoch is the training batch.

[0181] 3. Regularization and training techniques.

[0182] Dropout: Apply Dropout after the TCN layer and the Transformer layer, with a dropout rate p = 0.1. Formula: Output = Dropout(Output, p).

[0183] Gradient clipping: Limit the gradient norm not to exceed the threshold θ = 1.0.

[0184]

[0185] where g is the gradient and ||g|| is the L2 norm of the gradient.

[0186] Scheduled sampling: Gradually reduce the probability p of using the true label according to the training epoch teacher , to mitigate the accumulation of autoregressive errors:

[0187] p teacher = max((0.0, 1.0 - 0.005·epoch).

[0188] (2) Interpretability analysis.

[0189] ① Attention weight analysis:

[0190] Calculate the attention weights: The attention score matrix of the h-th head in the l-th layer of the Transformer reflects the influence of historical time steps on the current prediction.

[0191] Identification of key time points: Sum the attention weights for the lag period t rain + τ of the research metric with a duration of T:

[0192]

[0193] If S rain > the threshold, it indicates that the model pays attention to the lag effect of this metric.

[0194] Visualization: A heatmap is used to display the distribution of attention weights at each time step.

[0195] ② Physical consistency verification.

[0196] Plot the landslide displacement prediction increment ΔD pred and the scatter plot with the theoretical increment ∈(t), and calculate the correlation coefficient R 2

[0197]

[0198] where is the mean of the theoretical displacement increment. If R 2 > 0.2, it indicates that the prediction conforms to the physical law.

[0199] For the time period of sudden displacement increase, check whether it corresponds to a significant increase in ∈(t) (such as humidity exceeding H0 and soil pressure decreasing).

[0200] Step 4: Input the real-time time-series data of multiple landslide influencing factors into the trained TCN-Transformer. Use TCN to splice the features at the corresponding time, perform feature mixing through a fully connected layer, and then input the mixed features into the self-attention layer of Transformer for linear projection and combined output to obtain the landslide displacement.

[0201] The present invention also provides a swelling soil landslide displacement prediction system, including:

[0202] A data acquisition module for acquiring the real-time time-series data of multiple landslide influencing factors in the area to be predicted;

[0203] A model construction module for using the neural network model Transformer as the basic model and replacing the embedding layer of Transformer with the temporal convolutional network TCN to obtain the hybrid prediction model TCN-Transformer;

[0204] A model training module for designing a physical loss function based on the swelling soil constitutive equation and training the TCN-Transformer using the physical constraint loss function;

[0205] A landslide displacement prediction module for inputting the real-time time-series data of multiple landslide influencing factors into the trained TCN-Transformer, using TCN to splice the features at the corresponding time, performing feature mixing through a fully connected layer, and then inputting the mixed features into the self-attention layer of Transformer for linear projection and combined output to obtain the landslide displacement.

[0206] The present invention also provides a computer device, including a memory and a processor; the memory stores a computer program, and the processor is configured to run the computer program in the memory to execute the expansive soil landslide displacement prediction method.

[0207] The present invention also provides a computer-readable storage medium storing a computer program, which is suitable for being loaded by a processor to execute the expansive soil landslide displacement prediction method.

[0208] The above-described embodiments are only preferred specific embodiments of the present invention, and the protection scope of the present invention is not limited thereto. Any simple variations or equivalent replacements of technical solutions that can be obviously obtained by those skilled in the art within the technical scope disclosed by the present invention all belong to the protection scope of the present invention.

Claims

1. A method for predicting the displacement of expansive soil landslides, characterized in that, The steps include: Obtain real-time time series data of multiple landslide influencing factors in the area to be predicted; Taking the neural network model Transformer as the basic model, the temporal convolutional network TCN is used to replace the embedding layer of Transformer to obtain the hybrid prediction model TCN-Transformer; Design a physical loss function based on the constitutive equation of expansive soil and use the physical constraint loss function to train TCN-Transformer; The real-time time series data of multiple landslide influencing factors are input into the trained TCN-Transformer. The features at the corresponding moments are spliced using TCN, and the features are mixed through the fully connected layer. The mixed features are then input into the self-attention layer of the Transformer for linear projection and merged output to obtain the landslide displacement.

2. The displacement prediction method for expansive soil landslides according to claim 1, wherein Before inputting the real-time time series data of multiple landslide influencing factors into the trained TCN-Transformer, the real-time time series data is preprocessed, specifically including the following steps: The real-time time series data of multiple landslide influencing factors are resampled and wavelet transformed at different time intervals to remove the noise data in the real-time time series data; Grey relational analysis (GRA) is used to calculate the correlation between multiple landslide influencing factors and displacement. Dynamic hysteresis features are constructed based on the correlation between multiple landslide influencing factors and displacement, and the dynamic hysteresis features are input into the trained TCN-Transformer for processing.

3. The method for predicting the displacement of expansive soil landslides according to claim 1, wherein, The total loss function of the TCN-Transformer consists of a weighted mean square error loss function and a physical constraint loss function and is specifically as follows: Among them, λ is the weight; In the formula, the weighted mean square error loss function is as follows: Among them, is the predicted displacement on the i-th day, is the actual displacement on the i-th day, l is the predicted duration, and w i is the dynamically allocated weight for the prediction length; The physical constraint loss function is as follows: where ΔD pred (t) is the predicted displacement increment, ∈(t) is the theoretical increment at the i-th step calculated based on historical humidity and soil pressure, and l is the predicted time duration.

4. The displacement prediction method for expansive soil landslides according to claim 1, characterized in that, The real-time time series data of the multiple landslide influencing factors include rainfall and soil pressure.

5. A displacement prediction system for expansive soil landslides, characterized in that, include: A data acquisition module is used to obtain real-time time series data of multiple landslide influencing factors in the area to be predicted; The model building module is used to replace the embedding layer of the Transformer neural network model with the temporal convolutional network (TCN) to obtain the hybrid prediction model TCN-Transformer. The model training module is used to design a physical loss function based on the constitutive equation of expansive soil and train the TCN-Transformer using the physical constraint loss function; The landslide displacement prediction module is used to input the real-time time series data of multiple landslide influencing factors into the trained TCN-Transformer, use TCN to splice the features at the corresponding moments, mix the features through the fully connected layer, and then input the mixed features into the self-attention layer of the Transformer for linear projection and merging output to obtain the landslide displacement.

6. A computer device, characterized in that, It comprises a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the method for predicting displacement of expansive soil landslide according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the method for predicting displacement of an expansive soil landslide according to any one of claims 1 to 4.

Images