Large water resource allocation project high and steep slope deformation prediction method based on Beidou data

Abstract

The invention discloses a big water resource allocation project high and steep slope deformation prediction method based on Beidou data, and the method comprises the following steps: 1) setting a water resource allocation project slope deformation detection point, and obtaining the surface displacement data of the detection point in a T time period; 2) preprocessing the surface displacement data; 3) constructing a high and steep slope deformation prediction model with associated Patch and composite features; and 4) inputting the preprocessed surface displacement data into the high and steep slope deformation prediction model, and predicting slope deformation in a future t period to obtain a slope deformation prediction result. The method can be applied to slope deformation prediction of a large-scale engineering construction site, and a certain basis is provided for construction safety.

Description

Technical Field

[0001] The present invention relates to the field of high and steep slope deformation prediction in large-scale water conservancy construction projects, and specifically to a high and steep slope deformation prediction method for large-scale water resource allocation projects based on Beidou data. Background Art

[0002] With the increasing requirements for water resources management, it is particularly important to conduct scientific analysis, accurate prediction and effective early warning of high and steep dangerous slopes based on the monitoring data of the Beidou satellite navigation system. Improving the accuracy and reliability of high and steep slope displacement prediction from the perspective of the overall monitoring system not only ensures the safety and reliability of the project, but also protects the surrounding environment from damage, and also makes positive contributions to the safety and sustainability of water resources allocation engineering.

[0003] For the prediction of high and steep slope deformation, neural network models such as LSTM and BP neural network are static models that only rely on the data at the current time point and ignore the correlation between the previous and next data, resulting in limited prediction results. At present, new models such as TFT and MSCNet generally adopt fixed patch ideas or channel independence strategies. These methods divide the time series into patches and extract relevant features. Although the model based on the patch idea has proved its good performance, the fixed patch size limits the adaptability of the model to different time series, and only focuses on local correlation, failing to effectively capture global information and channel dependency information. This limitation causes the model to perform poorly when processing complex data and cannot fully reveal the potential pattern. In addition, although the channel independence strategy has achieved certain results in time series prediction, it ignores the correlation between channels, further restricting the model performance. Some studies have adopted a mixed channel method in the hope of further improving the model performance, which often requires relying on a complex attention mechanism, otherwise it is difficult to achieve the best performance, and it is difficult to ensure the high robustness of the model and adapt to changing environments and data. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data, comprising the following steps:

[0005] 1) Obtain the surface displacement data of multiple slope deformation detection points of water resource allocation projects within a T period;

[0006] 2) Preprocessing the surface displacement data;

[0007] 3) Construct a high and steep slope deformation prediction model with associated patches and composite features;

[0008] 4) The preprocessed surface displacement data is input into the high steep slope deformation prediction model to predict the slope deformation in the future period t and obtain the slope deformation prediction result.

[0009] Further, in step 1), the surface displacement data are surface displacement data in three directions in the Northeastern Universe (ENU) coordinate system, which are recorded as (E_shift, N_shift, U_shift).

[0010] Furthermore, the surface displacement data are obtained through the BeiDou satellite navigation system.

[0011] Furthermore, the preprocessing includes outlier processing, missing value filling and data resampling.

[0012] Further, in step 2), the step of preprocessing the surface displacement data includes:

[0013] 2.1) Using the IQR method to identify outliers in the surface displacement data, and replacing the outliers with the rolling window average;

[0014] 2.2) Use the rolling window average to replace missing values ​​in the surface displacement data;

[0015] 2.3) The sampling frequency of the surface displacement data is adjusted using resampling technology, and the gaps in the surface displacement data are filled using linear interpolation.

[0016] Furthermore, the pre-processed surface displacement data is input into the high steep slope deformation prediction model, and the steps of predicting the slope deformation in the future period t include:

[0017] 4.1) Surface displacement data X ACFM ∈R C×L×d Segmentation is performed to obtain P patches, denoted as (X 1 ,X 2 ,...,X P ), L is the length of the surface displacement data, C is the number of surface displacement data input to the model, and d is the feature dimension; Stride is the sliding step; S is the time step;

[0018] 4.2) Modify the values ​​of the sliding stride Stride and the time step S, and return to step 4.1) until n patches of different sizes and partial overlap are generated to form a multi-scale associated segmentation view;

[0019] 4.3) Perform feature embedding on each patch to obtain the embedded time series representation of the Intra-Attention mechanism where d m is the embedding dimension;

[0020] 4.4) Based on the embedded time series representation, the attention score of each patch is calculated, and the attention scores of all patches are concatenated to generate a local detail representation for describing the time steps contained in the time series.

[0021] 4.5) Surface displacement data X 1:P ∈R C×P×S×d Perform feature embedding to obtain embedded time series representation And the embedded time series is represented as Rearrange so that the embedded time series representation The size of the input is adjusted to d′ m =S×d m , thus obtaining the time series representation of the Inter-Attention mechanism And calculate the time series representation X Inter The attention score, The attention score Used to describe the global details of the time steps contained in the time series; X 1:P ∈R C×P×S×d Refers to the sequence (X 1 ,X 2 ,...,X P );

[0022] 4.6) Surface displacement data X ACFM ∈R C×L×d Perform feature embedding to obtain embedded time series representation

[0023] 4.7) Using STAR module to embed time series representation X Channel Processing is performed to obtain the attention weight o between channels, that is:

[0024] o=Stoch_Pool(MLP 1 (X Channel ))(1)

[0025] In the formula, MLP 1 is a multi-layer perceptron; Stoch_Pool is a random pooling mechanism;

[0026] 4.8) Combine the attention weight o with the embedded time series representation X Channel Fusion and use multi-layer perceptron MLP 2 Restore the fused features and get the output Right now:

[0027] F = Repaeat_Contact(X Channel,o) (2)

[0028] STAR Channel =MLP 2 (F)+X Channel (3)

[0029] In the formula, F is the fused feature;

[0030] 4.9) Output of Intra-Attention Perform linearization, expand the Patch size dimension from 1 to S, and rearrange it to obtain a new local detail representation

[0031] 4.10) Use linear layers to represent local details Global detail representation The dimension S, d m Merge into dimension L, denoted as Attn Inter ∈R C×P×L ;

[0032] 4.11) Attn Intra ∈R C×P×L 、Attn Inter ∈R C×P×L , Add together to generate the final output of the multi-attention mechanism

[0033] 4.12) Use discrete Fourier transform to transform the time domain surface displacement data X∈R C×L×d Convert to frequency domain and keep the K with the largest amplitude f Then, the retained components are restored to the time domain by inverse discrete Fourier transform to obtain the seasonal component X of the surface displacement data. sea ,Right now:

[0034]

[0035] Where A and θ represent the phase and amplitude parameters of X in the frequency domain in Fourier transform, Indicates that the top K f The frequency of the amplitude, IDFT stands for Inverse Discrete Fourier Transform;

[0036] 4.13) Use the convolution kernel mean pooling method to decompose the remaining part X after seasonality rem =XX sea Extract and get the trend component X trend ,Right now:

[0037]

[0038] In the formula, is the mean pooling function of the ith core, N is the total number of cores, and Softmax(L(·)) is used to control the weights of the results from different cores;

[0039] 4.14) The surface displacement data X∈R C×L×d 、Trend component X trend , Seasonal Component X sea Add and perform linear mapping Linear(·) to obtain the input X for the path weight generation part Weights ,Right now:

[0040] X Weights =Linear(X+X sea +X trend ) (6)

[0041] 4.15) Based on input X Weights , generate the weight sequence R(X Weights ),Right now:

[0042] R(X Weights )=Soft max(X Weights W r +ε·Softplus(X Weights W noise )), ε~N(0,1) (7)

[0043] Where R(·) represents the entire routing function, W r , W noise ∈R d×P is the learnable parameter generated by the weight, d represents X Weights The feature dimension of , P represents the number of patch sizes;

[0044] 4.16) Retain the weight sequence R(X) Weights ) and reset the remaining weights to 0 to get the path weight Right now:

[0045]

[0046] 4.17) Using the dimension transformation function T i () Dimensionally align the final output of the multiple attention mechanisms and use path weights and indicator functions Perform weighted aggregation on the selected topK outputs to obtain the slope deformation prediction result X out ,Right now:

[0047]

[0048]

[0049]

[0050] In the formula, is the ith output.

[0051] Further, in step 4.3), the i-th Patch X i ∈R C×S×d The embedded time series representation is denoted as Right now:

[0052]

[0053] In the formula, Embedding represents the feature embedding operation; d m Represents the feature space dimension.

[0054] The local details of the i-th Patch are shown as follows:

[0055]

[0056]

[0057] In the formula, Respectively represent the Key and Value matrices in the attention mechanism; represents the query matrix.

[0058] Further, in step 4.5), the time series representation is embedded Input X of Inter-Attention Mechanism Inter They are as follows:

[0059]

[0060] Where Rearrange represents the reordering operation.

[0061] The output corresponding to the Inter-Attention mechanism As shown below:

[0062]

[0063] In the formula, They represent the query matrix, Key matrix and Value matrix in the attention mechanism respectively.

[0064] Furthermore, embedding time series representation As shown below:

[0065] X Channel =Embedding(X)(17).

[0066] The technical effect of the present invention is unquestionable. The present invention proposes a multi-scale feature extraction method based on associated patch partitioning, which enables the model to effectively retain the correlation between patches and fully capture the local and global dependencies in the time series.

[0067] This paper proposes a composite feature modeling architecture that integrates local, global and channel correlation. The channel correlation part adopts a star-shaped centralized structure and an efficient MLP module, which greatly reduces the computational complexity and improves the model efficiency while enhancing the feature extraction capability.

[0068] The present invention integrates time resolution, time distance and channel dependence, and performs adaptive path exploration. It can be used in slope deformation prediction at large-scale engineering construction sites, providing a certain basis for construction safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is the overall framework of the present invention. DETAILED DESCRIPTION

[0070] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0071] Embodiment 1:

[0072] See also Figure 1 , a method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data, comprising the following steps:

[0073] 1) Obtaining surface displacement data of multiple slope deformation detection points of water resource allocation projects within a period of T (set to 171 days in this embodiment); the slope deformation detection points of the water resource allocation project are fixed slope deformation monitoring points set up in actual water conservancy projects. When acquiring data, select the surface displacement data of all or part of the fixed slope deformation monitoring points.

[0074] 2) Preprocessing the surface displacement data;

[0075] 3) Construct a high and steep slope deformation prediction model with associated patches and composite features;

[0076] 4) The pre-processed surface displacement data is input into the high-steep slope deformation prediction model to predict the slope deformation in the future t period (set as 12 days in this embodiment) to obtain the slope deformation prediction result.

[0077] In step 1), the surface displacement data are surface displacement data in three directions in the Northeastern Universe (ENU) coordinate system, denoted as (E_shift, N_shift, U_shift).

[0078] The surface displacement data are obtained through the BeiDou satellite navigation system.

[0079] The preprocessing includes outlier processing, missing value filling and data resampling.

[0080] In step 2), the step of preprocessing the surface displacement data includes:

[0081] 2.1) Using the IQR method to identify outliers in the surface displacement data, and replacing the outliers with the rolling window average;

[0082] 2.2) Use the rolling window average to replace missing values ​​in the surface displacement data;

[0083] 2.3) The sampling frequency of the surface displacement data is adjusted using resampling technology, and the gaps in the surface displacement data are filled using linear interpolation.

[0084] The pre-processed surface displacement data is input into the high and steep slope deformation prediction model. The steps of predicting the slope deformation in the future period t include:

[0085] 4.1) Surface displacement data X ACFM ∈R C×L×d Segmentation is performed to obtain P patches, denoted as (X 1 ,X 2 ,...,X P ), L is the length of the surface displacement data, C is the number of surface displacement data input to the model, and d is the feature dimension; Stride is the sliding step; S is the time step;

[0086] 4.2) Modify the values ​​of the sliding stride Stride and the time step S, and return to step 4.1) until n patches of different sizes and partial overlap are generated to form a multi-scale associated segmentation view;

[0087] 4.3) Perform feature embedding on each patch to obtain the embedded time series representation of the Intra-Attention mechanism where d m is the embedding dimension;

[0088] 4.4) Based on the embedded time series representation, the attention score of each patch is calculated, and the attention scores of all patches are concatenated to generate a local detail representation for describing the time steps contained in the time series.

[0089] 4.5) Surface displacement data Perform feature embedding to obtain embedded time series representation And the embedded time series is represented as Rearrange so that the embedded time series representation The size of the input is adjusted to d′ m =S×d m , thus obtaining the time series representation of the Inter-Attention mechanism And calculate the time series representation X Inter The attention score, The attention score Used to describe the global details of the time steps contained in the time series; X 1:P ∈R C×P×S×d Refers to the sequence (X 1 ,X 2 ,...,X P );

[0090] 4.6) Surface displacement data X ACFM ∈R C×L×d Perform feature embedding to obtain embedded time series representation

[0091] 4.7) Using STAR module to embed time series representation X Channel Processing is performed to obtain the attention weight o between channels, that is:

[0092] o=Stoch_Pool(MLP 1 (X Channel ))(1)

[0093] In the formula, MLP 1 is a multi-layer perceptron; Stoch_Pool is a random pooling mechanism;

[0094] 4.8) Combine the attention weight o with the embedded time series representation X Channel Fusion and use multi-layer perceptron MLP 2 Restore the fused features and get the output Right now:

[0095] F = Repaeat_Contact(X Channel ,o) (2)

[0096] STAR Channel =MLP 2 (F)+X Channel (3)

[0097] In the formula, F is the fused feature;

[0098] 4.9) Output of Intra-Attention Perform linearization, expand the Patch size dimension from 1 to S, and rearrange it to obtain a new local detail representation

[0099] 4.10) Use linear layers to represent local details Global detail representation The dimension S, d m Merge into dimension L, denoted as Attn Intra ∈R C×P×L 、Attn Inter ∈R C×P×L ;

[0100] 4.11) Attn Intra ∈R C×P×L 、Attn Inter ∈R C×P×L , Add together to generate the final output of the multi-attention mechanism Attn Intra 、Attn Inter Using the attention mechanism for calculation, Attn Intra Extract the correlation within the patch (do local details), Attn Inter Extract correlations between patches (global details), STAR Channel is the channel correlation.

[0101] 4.12) Use discrete Fourier transform to transform the time domain surface displacement data X∈R C×L×d Convert to frequency domain and keep the K with the largest amplitude f Then, the retained components are restored to the time domain by inverse discrete Fourier transform to obtain the seasonal component X of the surface displacement data. sea ,Right now:

[0102]

[0103] Where A and θ represent the phase and amplitude parameters of X in the frequency domain in Fourier transform, Indicates that the top K f The frequency of the amplitude, IDFT stands for Inverse Discrete Fourier Transform;

[0104] 4.13) Use the convolution kernel mean pooling method to decompose the remaining part X after seasonality rem =XX sea Extract and get the trend component X trend ,Right now:

[0105]

[0106] In the formula, is the mean pooling function of the ith core, N is the total number of cores, and Softmax(L(·)) is used to control the weights of the results from different cores;

[0107] 4.14) The surface displacement data X∈R C×L×d 、Trend component X trend , Seasonal Component X sea Add and perform linear mapping Linear(·) to obtain the input X for the path weight generation part Weights ,Right now:

[0108] X Weights =Linear(X+X sea +X trend ) (6)

[0109] 4.15) Based on input X Weights , generate the weight sequence R(X Weights ),Right now:

[0110] R(X Weights )=Soft max(X Weights W r +ε·Softplus(X Weights W noise )),ε~N(0,1) (7)

[0111] Where R(·) represents the entire routing function, W r ,W noise ∈R d×P is the learnable parameter generated by the weight, d represents X Weights The feature dimension of , P represents the number of patch sizes;

[0112] 4.16) Retain the weight sequence R(X) Weights ) and reset the remaining weights to 0 to get the path weight Right now:

[0113]

[0114] 4.17) Using the dimension transformation function T i () Dimensionally align the final output of the multiple attention mechanisms and use path weights and indicator functions Perform weighted aggregation on the selected topK outputs to obtain the slope deformation prediction result X out ,Right now:

[0115]

[0116]

[0117]

[0118] In the formula, is the ith output.

[0119] In step 4.3), the i-th Patch X i ∈R C×S×d The embedded time series representation is denoted as Right now:

[0120]

[0121] In the formula, Embedding represents the feature embedding operation; d m Represents the feature space dimension.

[0122] The local details of the i-th Patch are shown as follows:

[0123]

[0124]

[0125] In the formula, Respectively represent the Key and Value matrices in the attention mechanism; represents the query matrix.

[0126] In step 4.5), embed the time series representation The input X of the Inter-Attention mechanism Inter They are as follows:

[0127]

[0128] Where Rearrange represents the reordering operation.

[0129] The output corresponding to the Inter-Attention mechanism As shown below:

[0130]

[0131] In the formula, They represent the query matrix, Key matrix and Value matrix in the attention mechanism respectively.

[0132] Embedded time series representation As shown below:

[0133] XChannel =Embedding(X) (17).

[0134] Embodiment 2:

[0135] A method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data includes the following steps:

[0136] 1) Obtain the surface displacement data of multiple slope deformation detection points of water resource allocation projects within a T period;

[0137] 2) Preprocessing the surface displacement data;

[0138] 3) Construct a high and steep slope deformation prediction model with associated patches and composite features;

[0139] 4) The preprocessed surface displacement data is input into the high steep slope deformation prediction model to predict the slope deformation in the future period t and obtain the slope deformation prediction result.

[0140] Embodiment 3:

[0141] A method for predicting deformation of steep slopes of large-scale water resource allocation projects based on Beidou data, the technical content of which is the same as that of Example 2. Furthermore, in step 1), the surface displacement data are surface displacement data in three directions in the Northeastern Universe (ENU) coordinate system, recorded as (E_shift, N_shift, U_shift).

[0142] Embodiment 4:

[0143] A method for predicting deformation of high and steep slopes of large-scale water resource allocation projects based on Beidou data, the technical content of which is the same as any one of Examples 2-3, and further, the surface displacement data is obtained through the Beidou satellite navigation system.

[0144] Embodiment 5:

[0145] A method for predicting deformation of high and steep slopes of large-scale water resource allocation projects based on Beidou data, the technical content of which is the same as any one of Examples 2-4. Furthermore, the preprocessing includes outlier processing, missing value filling and data resampling.

[0146] Embodiment 6:

[0147] A method for predicting deformation of steep slopes in large-scale water resource allocation projects based on Beidou data, the technical content of which is the same as any one of Embodiments 2-5, further, in step 2), the step of preprocessing the surface displacement data comprises:

[0148] 2.1) Using the IQR method to identify outliers in the surface displacement data, and replacing the outliers with the rolling window average;

[0149] 2.2) Use the rolling window average to replace missing values ​​in the surface displacement data;

[0150] 2.3) The sampling frequency of the surface displacement data is adjusted using resampling technology, and the gaps in the surface displacement data are filled using linear interpolation.

[0151] Embodiment 7:

[0152] A method for predicting deformation of a high-steep slope of a large-scale water resource allocation project based on Beidou data, the technical content of which is the same as any one of Embodiments 2-6. Further, the pre-processed surface displacement data is input into a high-steep slope deformation prediction model, and the step of predicting the slope deformation in the future t period includes:

[0153] 4.1) Surface displacement data X ACFM ∈R C×L×d Segmentation is performed to obtain P patches, denoted as (X 1 ,X 2 ,...,X P ), L is the length of the surface displacement data, C is the number of surface displacement data input to the model, and d is the feature dimension; Stride is the sliding step; S is the time step;

[0154] 4.2) Modify the values ​​of the sliding stride Stride and the time step S, and return to step 4.1) until n patches of different sizes and partial overlap are generated to form a multi-scale associated segmentation view;

[0155] 4.3) Perform feature embedding on each patch to obtain the embedded time series representation of the Intra-Attention mechanism where d m is the embedding dimension;

[0156] 4.4) Based on the embedded time series representation, the attention score of each patch is calculated, and the attention scores of all patches are concatenated to generate a local detail representation for describing the time steps contained in the time series.

[0157] 4.5) Surface displacement data X 1:P ∈R C×P×S×d Perform feature embedding to obtain embedded time series representation And the embedded time series is represented as Rearrange so that the embedded time series representation The size of the input is adjusted to d′ m =S×d m , thus obtaining the time series representation of the Inter-Attention mechanism And calculate the time series representation X Inter The attention score, The attention score Used to describe the global details of the time steps contained in the time series; X 1:P ∈R C×P×S×d Refers to the sequence (X 1 ,X 2 ,...,X P );

[0158] 4.6) Surface displacement data X ACFM ∈R C×L×d Perform feature embedding to obtain embedded time series representation

[0159] 4.7) Using STAR module to embed time series representation X Channel Processing is performed to obtain the attention weight o between channels, that is:

[0160] o=Stoch_Pool(MLP 1 (X Channel ))(1)

[0161] In the formula, MLP 1 is a multi-layer perceptron; Stoch_Pool is a random pooling mechanism;

[0162] 4.8) Combine the attention weight o with the embedded time series representation X Channel Fusion and use multi-layer perceptron MLP 2 Restore the fused features and get the output Right now:

[0163] F = Repaeat_Contact(X Channel ,o) (2)

[0164] STAR Channel =MLP 2 (F)+X Channel (3)

[0165] 4.9) Output of Intra-Attention Perform linearization, expand the Patch size dimension from 1 to S, and rearrange it to obtain a new local detail representation

[0166] 4.10) Use linear layers to represent local details Global detail representation The dimension S, d m Merge into dimension L, denoted as AttnIntra ∈R C×P×L 、Attn Inter ∈R C×P×L ;

[0167] 4.11) Attn Intra ∈R C×P×L 、Attn Inter ∈R C×P×L , Add together to generate the final output of the multi-attention mechanism

[0168] 4.12) Use discrete Fourier transform to transform the time domain surface displacement data X∈R C×L×d Convert to frequency domain and keep the K with the largest amplitude f Then, the retained components are restored to the time domain by inverse discrete Fourier transform to obtain the seasonal component X of the surface displacement data. sea ,Right now:

[0169]

[0170] Where A and θ represent the phase and amplitude parameters of X in the frequency domain in Fourier transform, Indicates that the top K f The frequency of the amplitude, IDFT stands for Inverse Discrete Fourier Transform;

[0171] 4.13) Use the convolution kernel mean pooling method to decompose the remaining part X after seasonality rem =XX sea Extract and get the trend component X trend ,Right now:

[0172]

[0173] In the formula, is the mean pooling function of the ith core, N is the total number of cores, and Softmax(L(·)) is used to control the weights of the results from different cores;

[0174] 4.14) The surface displacement data X∈R C×L×d 、Trend component X trend , Seasonal Component X sea Add and perform linear mapping Linear(·) to obtain the input X for the path weight generation part Weights ,Right now:

[0175] X Weights =Linea(X+X sea +X trend ) (6)

[0176] 4.15) Based on input X Weights , generate the weight sequence R(X Weights ),Right now:

[0177] R(X Weights )=Soft max(X Weights W r +ε·Softplus(X Weights W noise )), ε~N(0,1) (7)

[0178] Where R(·) represents the entire routing function, W r , W noise ∈R d×P is the learnable parameter generated by the weight, d represents X Weights The feature dimension of , P represents the number of patch sizes;

[0179] 4.16) Retain the weight sequence R(X) Weights ) and reset the remaining weights to 0 to get the path weight Right now:

[0180]

[0181] 4.17) Use the dimension transformation function T(·) to align the dimensions of the final output of the multi-attention mechanism, and use the path weight and indicator function Perform weighted aggregation on the selected topK outputs to obtain the slope deformation prediction result X out ,Right now:

[0182]

[0183]

[0184]

[0185] Embodiment 8:

[0186] A method for predicting deformation of steep slopes in large-scale water resource allocation projects based on Beidou data, the technical content of which is the same as any one of Embodiments 2-7, further, in step 4.3), the i-th Patch X i∈R C×S×d The embedded time series representation is denoted as Right now:

[0187]

[0188] In the formula, Embedding represents the feature embedding operation; d m Represents the feature space dimension.

[0189] The local details of the i-th Patch are shown as follows:

[0190]

[0191]

[0192] In the formula, Respectively represent the Key and Value matrices in the attention mechanism; represents the query matrix.

[0193] Embodiment 9:

[0194] A method for predicting deformation of steep slopes in large-scale water resource allocation projects based on Beidou data, the technical content of which is the same as any one of embodiments 2-8, further, in step 4.5), the time series representation is embedded The input X of the Inter-Attention mechanism Inter They are as follows:

[0195]

[0196] Where Rearrange represents the reordering operation.

[0197] The output corresponding to the Inter-Attention mechanism As shown below:

[0198]

[0199] In the formula, They represent the query matrix, Key matrix and Value matrix in the attention mechanism respectively.

[0200] Embodiment 10:

[0201] A method for predicting deformation of steep slopes in large-scale water resource allocation projects based on Beidou data, the technical content is the same as any one of embodiments 2-9, further, embedded time series representation As shown below:

[0202] X Channel =Embedding(X)(17).

[0203] Embodiment 11:

[0204] A method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data is as follows:

[0205] This example uses the Beidou satellite navigation (BDS) measurement system to monitor the surface displacement data of high and steep slopes in the Chongqing Yuxi large-scale water resources allocation project as an example for empirical analysis. Deformation prediction based on the APEX-CPM algorithm is performed to provide more accurate and comprehensive slope safety warning support for slope warning managers. The specific steps include:

[0206] S100: Prediction target definition. The E_shift, N_shift and U_shift surface displacement data of the three detection points BDS001, BDS002 and BDS003 in the project were obtained in the 171-day ENU coordinate system. We will predict the slope deformation in the next 12 days.

[0207] S200: Data preprocessing. After outlier processing, missing value filling and data resampling, the next step is performed.

[0208] S300: Model construction. We constructed an APEX-CPM method with associated patches and composite feature modeling. The entire network structure consists of reversible instance normalization, stacking of ACFM blocks (Adaptive Composite Feature Modeling Block) and linear prediction layers. The ACFM block, as a core component, integrates data feature extraction mechanisms from three perspectives. It also includes a CFT module and an adaptive path. Based on the associated patch partitioning, the CFT module uses a composite feature modeling mechanism that can simultaneously capture local correlation, global correlation and channel correlation to extract composite features. The adaptive path consists of two main components: the composite feature router (CFR) and the composite feature aggregator (CFA). The composite feature aggregator is responsible for weighted integration of the features selected by the router, and outputs a comprehensive temporal feature representation through weighted aggregation of composite features, and finally generates the output result of the ACFM block.

[0209] S400: Use the prediction model to predict the deformation of high and steep slopes. The Beidou slope deformation dataset processed by S200 is input into the prediction model constructed by S300, and ablation experiments and comparison experiments are carried out.

[0210] S201: Use the IQR method to identify outliers in the data and replace them with the rolling window mean.

[0211] S202: Use the rolling window average to replace missing values.

[0212] S203: The original data ranges from recording only one data point per day to once per hour. The data collection frequency is adjusted to one data point per day using resampling technology, and the data gaps are filled by linear interpolation.

[0213] S301: Associate Patch Division. For the input Beidou deformation data sequence X ACFM ∈R C×L×d , whose input length is L, C represents the number of variables, and the feature dimension is d. In each patch segmentation operation of size S, the input sequence X ACFM Divided into P patches, namely X ACFM =(X 1 ,X 2 ,...,X P ),in Stride is the sliding stride, X i ∈R C×S×d , (i=1,...,P), each patch contains S time steps. By setting different patch sizes S and strides Stride, multiple patches of different sizes and partial overlaps can be generated, thus forming a multi-scale associated segmentation view.

[0214] S302: Intra-Attention. For the i-th input Patch X i ∈R C×S×d Perform feature embedding. First, map it from d to d through the Embedding operation m The feature space of where d m is the embedding dimension.

[0215]

[0216] Then, for the embedded time series Perform a linear transformation to obtain the Key and Value matrices in the attention mechanism, which are recorded as Query Matrix It is obtained through training and is used to calculate the interaction between different time steps of the time series to generate attention scores

[0217]

[0218] Finally, the attention results of all patches are concatenated to generate a global time series context information representation Used to describe the local details of the time steps contained in a time series.

[0219]

[0220] S303: Inter-Attention. For the input Beidou deformation data time series X 1:P ∈RC×P×S×d , first map it from d to d through the Embedding operation m dimensional feature space, thus obtaining the embedded time series representation where d m is the embedding dimension. On this basis, we further Inter Rearrange it to make it smaller than d′ m =S×d m , and finally get the time series representation input to the Inter-Attention mechanism

[0221]

[0222] Afterwards, the self-attention mechanism is used to model the correlation between different associated patches of the input. Inter Apply linear transformation to obtain the Key, Value, and Query matrices for the attention mechanism, which are Then, the attention scores between associated patches are calculated based on these matrices.

[0223]

[0224] S304: Channel-Attention. For the input time series X∈R without patch division C×L×d Map it from d to d through the embedding operation m dimensional feature space, thus obtaining the embedded time series representation where d m is the embedding dimension.

[0225] X Channel =Embedding(X)

[0226] Next, the STAR module is used to model channel correlation. Specifically, the multi-layer perceptron MLP is used to 1 For the embedded feature d m Projection to core dimension d c , and apply the activation function GELU to activate it to enhance the feature representation. Then, through the random pooling (Stoch_Pool) mechanism, the global information of the transformed result is compressed and sampled to generate the attention weights between channels.

[0227] o=Stoch_Pool(MLP 1 (X Channel ))

[0228] On this basis, the obtained attention weights are fused with the original features and finally processed by the multi-layer perceptron MLP. 2 And apply the activation function GELU to further restore the fused features and generate the final output

[0229] F = Repaeat_Contact(X Channel ,o)

[0230] STAR Channel =MLP 2 (F)+X Channel

[0231] The output of Intra-Attention is linearly changed, its Patch size dimension is expanded from 1 to S, and it is rearranged as Then, the processed Attn Intra With Attn Inter After the linear layer, S and d m These two dimensions are merged into a new dimension L.

[0232] S305: Multi-attention mechanism. The correlation Attn extracted from S302, S303 and S304 Intra 、Attn Inter 、STAR Channel Add them together to generate the final output of the multi-attention mechanism.

[0233] S306: Composite feature router. The time domain input X∈R is transformed by discrete Fourier transform (DFT). C×L×d Convert to the frequency domain and select the K with the largest amplitude in the frequency domain f Then, the seasonal part X of the input data X is obtained by restoring it to the time domain through the inverse discrete Fourier transform (IDFT). sea .

[0234]

[0235] Where A and θ represent the phase and amplitude parameters of X in the frequency domain in Fourier transform, Indicates that the top K f The frequency of the amplitude.

[0236] The convolution kernel mean pooling method is used to decompose the remaining part X after seasonality rem =XX sea Further extract the trend component X trend .

[0237]

[0238] in is the mean pooling function of the ith core, N is the total number of cores, and Softmax(L(·)) is used to control the weights of the results from different cores.

[0239] After seasonal and trend decomposition of the input time series X, we add the seasonal component, trend component and original input, and then perform linear mapping Linear(·) to finally obtain the input X for path weight generation. Weights .

[0240] X Weights =X+X sea +X trend

[0241] Noise is added to the weight generation process to increase randomness. The process of generating a composite path is as follows:

[0242] R(X Weights )=Soft max(X Weights W r +ε·Softplus(X Weights W noise )), ε~N(0,1)

[0243] Where R(·) represents the entire routing function, W r , W noise ∈R d×P is the learnable parameter generated by the weight, d represents X Weights The feature dimension of , and P represents the number of patch sizes.

[0244] Select the top K weights and keep their values, set the remaining weights to 0, and the final generated path weight is expressed as

[0245]

[0246] S307: Composite feature aggregator. After aligning the dimensions of the outputs of different CFTBlocks using the dimension transformation function T(·), the path weight and indicator function are used to The output results of the selected topK CFT Blocks are weighted and aggregated to obtain the final output X of the ACFM block. out :

[0247]

[0248]

[0249]

[0250] S401: The Beidou slope deformation dataset processed by S200 is input into the Pathformer, APEX-CPM (noPatch) and APEX-CPM models for ablation experiments. The mean absolute error (MAE) and mean square error (MSE) are used as evaluation indicators to analyze the effectiveness of the model.

[0251] S402: The Beidou slope deformation dataset processed by S200 was input into seven models, including APEX-CPM, Pathformer, SOFTS, Crossformer, DLinear, TimesNet and Autoformer, for comparative experiments. The mean absolute error (MAE) and mean square error (MSE) were used as evaluation indicators to analyze the prediction performance of the models.

[0252] The present invention uses the surface displacement data of the Northeastern Universe (ENU) coordinate system of the high and steep slopes in the large-scale water resources allocation project in western Chongqing for 171 days as a case for empirical analysis. The data comes from the big data platform of the Beidou satellite navigation system, which contains a total of 513 records, each of which contains displacements in three directions. The data in each direction remain continuous without interruption, showing good monitoring continuity. However, the complex dynamic nonlinear characteristics and multiple time and space influencing factors presented by the deformation data of the high and steep slopes make it very difficult to accurately predict the deformation of the slopes. Therefore, how to make full use of small sample data and efficiently extract data features to improve the accuracy and generalization ability of the prediction of the deformation of the high and steep slopes is also a technical difficulty that the present invention urgently needs to solve.

[0253] In order to ensure the validity of the data and the accuracy of the experimental results, the present invention performs data preprocessing and resampling on the high and steep slope displacement data, uses the last 20% of the data for testing, and uses the remaining points for training.

[0254] Based on the evaluation prediction performance of multiple evaluation indicators, we use mean absolute error (MAE) and mean square error (MSE) as evaluation indicators, and the calculation formula is as follows:

[0255]

[0256]

[0257] Where n represents the total sample size, y i Indicates the actual value of the high and steep slope deformation data, Represents the predicted value of the high and steep slope deformation data.

[0258] In order to demonstrate the effect of the model of the present invention on the prediction of high and steep slope deformation, the present invention uses Pathformer, SOFTS, Crossformer, DLinear, TimesNet, and Autoformer for comparative experiments. The experimental results show that the performance of the present invention in the two evaluation indicators of MAE and MSE is generally better than other models, highlighting its enhanced generalization ability and transferability. Taking the BDS001 measurement point data as an example, the MSE of the model reached 0.82084 and the MAE was 0.70027. Compared with the best-performing PatchTST, the MSE of APEX-CPM decreased by 0.03117, an increase of 3.66%; the MAE decreased by 0.01793, an increase of 2.49%. This shows that APEX-CPM can effectively provide accurate results for time series prediction.

[0259] Table 1. Evaluation metrics of different models. The best ones are highlighted in bold and the second best results are underlined.

[0260]

[0261]

[0262] In summary, the present invention proposes a multi-scale feature extraction method based on associated patch partitioning, which enables the model to effectively retain the correlation between patches and fully capture the local and global dependencies in the time series.

[0263] This paper proposes a composite feature modeling architecture that integrates local, global and channel correlation. The channel correlation part adopts a star-shaped centralized structure and an efficient MLP module, which greatly reduces the computational complexity and improves the model efficiency while enhancing the feature extraction capability.

[0264] The application of APEX-CPM in large-scale water conservancy projects shows that it has high accuracy and good generalization ability in the prediction of deformation of high and steep slopes. To a certain extent, it can provide safety guarantees for engineering construction and help improve the safety and stability of construction.

[0265] In short, the present invention proposes a composite feature modeling method APEX-CPM based on associated patches, which integrates time resolution, time distance and channel dependency, and performs adaptive path exploration. The performance of the present invention is better than other algorithms for monitoring deformation of high and steep slopes, and can be used in slope deformation prediction at large-scale engineering construction sites, providing a certain basis for construction safety.

Claims

1. A method for predicting high and steep slope deformation of large-scale water resource allocation projects based on Beidou data, characterized in that: The following steps are involved: 1) Obtain the surface displacement data of multiple slope deformation detection points of water resource allocation projects within a T period; 2) Preprocessing the surface displacement data; 3) Construct a high and steep slope deformation prediction model with associated patches and composite features; 4) The preprocessed surface displacement data is input into the high steep slope deformation prediction model to predict the slope deformation in the future period t and obtain the slope deformation prediction result.

2. According to claim 1, a method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data is characterized in that: In step 1), the surface displacement data are surface displacement data in three directions in the Northeastern Universe (ENU) coordinate system, denoted as (E_shift, N_shift, U_shift).

3. According to claim 1, a method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data is characterized in that: The surface displacement data are obtained through the BeiDou satellite navigation system.

4. According to claim 1, a method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data is characterized in that: The preprocessing includes outlier processing, missing value filling and data resampling.

5. The method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data according to claim 1 is characterized in that: In step 2), the step of preprocessing the surface displacement data includes: 2.1) Using the IQR method to identify outliers in the surface displacement data, and replacing the outliers with the rolling window average; 2.2) Use the rolling window average to replace missing values ​​in the surface displacement data; 2.3) The sampling frequency of the surface displacement data is adjusted using resampling technology, and the gaps in the surface displacement data are filled using linear interpolation.

6. The method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data according to claim 1 is characterized in that: The pre-processed surface displacement data is input into the high and steep slope deformation prediction model. The steps of predicting the slope deformation in the future period t include: 4.1) Surface displacement data X ACFM ∈R C×L×d Split and get P patches, denoted as (X1,X2,...,X P ), L is the length of the surface displacement data, C is the number of surface displacement data input to the model, and d is the feature dimension; Stride is the sliding step; S is the time step; 4.2) Modify the values ​​of the sliding stride Stride and the time step S, and return to step 4.1) until n patches of different sizes and partial overlap are generated to form a multi-scale associated segmentation view; 4.3) Perform feature embedding on each patch to obtain the embedded time series representation of the Intra-Attention mechanism where d m is the embedding dimension; 4.4) Based on the embedded time series representation, calculate the attention score of each patch, concatenate the attention scores of all patches, and generate a local detail representation for describing the time steps contained in the time series 4.5) Surface displacement data X 1:P ∈R C×P×S×d Perform feature embedding to obtain embedded time series representation And the embedded time series is represented as Rearrange so that the embedded time series representation The size of the input is adjusted to d′ m =S×d m , thus obtaining the time series representation of the Inter-Attention mechanism And calculate the time series representation X Inter The attention score, The attention score Used to describe the global details of the time steps contained in the time series; 4.6) Surface displacement data X ACFM ∈R C×L×d Perform feature embedding to obtain embedded time series representation 4.7) Using STAR module to embed time series representation X Channel Processing is performed to obtain the attention weight o between channels, that is: o=Stoch_Pool(MLP1(X Channel ))(1) Where MLP1 is a multi-layer perceptron; Stoch_Pool is a random pooling mechanism; 4.8) Combine the attention weight o with the embedded time series representation X Channel The fused features are restored using the multi-layer perceptron MLP2 to obtain the output Right now: F=Repaeat_Contact(X Channel ,o)(2) STAR Channel =MLP2(F)+X Channel (3) In the formula, F is the fused feature; 4.9) Output of Intra-Attention Perform linearization, expand the Patch size dimension from 1 to S, and rearrange it to obtain a new local detail representation 4.10) Use linear layers to represent local details Global detail representation The dimension S, d m Merge into dimension L, denoted as Attn Intra ∈R C×P×L 、Attn Inter ∈R c×P×L ; 4.11) Attn Inter ∈R C×P×L 、Attn Inter ∈R C×P×L , Add together to generate the final output of the multi-attention mechanism 4.12) Use discrete Fourier transform to transform the time domain surface displacement data X∈R C×L×d Convert to frequency domain and keep the K with the largest amplitude f Then, the retained components are restored to the time domain by inverse discrete Fourier transform to obtain the seasonal component X of the surface displacement data. sea ,Right now: Where A and θ represent the phase and amplitude parameters of X in the frequency domain in Fourier transform, Indicates that the top K f The frequency of the amplitude, IDFT stands for Inverse Discrete Fourier Transform; 4.13) Use the convolution kernel mean pooling method to decompose the remaining part X after seasonality rem =XX sea Extract and get the trend component X trend ,Right now: In the formula, is the mean pooling function of the ith core, N is the total number of cores, and Softmax(L(·)) is used to control the weights of the results from different cores; 4.14) The surface displacement data X∈R C×L×d 、Trend component X trend , Seasonal Component X sea Add and perform linear mapping Linear(·) to obtain the input X for the path weight generation part Weights ,Right now: X Weights =Linear(X+X sea +X trend ) (6) 4.15) Based on input X Weights , generate the weight sequence R(X Weights ),Right now: R(X Weights )= Soft max(X Weights W r +ε·Softplus(X Weights W noise )),ε~N(0,1) (7) Where R(·) represents the entire routing function, W r , W noise ∈R d×P is the learnable parameter generated by the weight, d represents X Weights The feature dimension of , P represents the number of patch sizes; 4.16) Retain the weight sequence R(X) Weights ) and reset the remaining weights to 0 to get the path weight Right now: 4.17) Using the dimension transformation function T i () Dimensionally align the final output of the multiple attention mechanisms and use path weights and indicator functions Perform weighted aggregation on the selected topK outputs to obtain the slope deformation prediction result X out ,Right now: In the formula, is the ith output.

7. The method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data according to claim 6 is characterized in that: In step 4.3), the i-th The embedded time series representation is denoted as Right now: In the formula, Embedding represents the feature embedding operation; d m Represents the feature space dimension. The local details of the i-th Patch are shown as follows: In the formula, Respectively represent the Key and Value matrices in the attention mechanism; represents the query matrix.

8. The method for predicting deformation of high and steep slopes in large-scale water resource allocation projects based on Beidou data according to claim 6 is characterized in that: In step 4.5), embed the time series representation The input X of the Inter-Attention mechanism Inter They are as follows: Where Rearrange represents the reordering operation. The output corresponding to the Inter-Attention mechanism As shown below: In the formula, They represent the query matrix, Key matrix and Value matrix in the attention mechanism respectively.

9. The method for predicting high and steep slope deformation of a large-scale water resource allocation project based on Beidou data according to claim 6 is characterized in that: Embedded time series representation As shown below: X Channel =Embedding(X)(17)。

Images