A neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering

Abstract

The present invention discloses a method for predicting surrounding surface settlement caused by foundation pit dewatering based on a neural network, which belongs to the field of surface settlement prediction technology, and includes: obtaining settlement data of a target area, calculating the contribution of the settlement data of the target area, generating a contribution value, and constructing a settlement data set of the target area based on the contribution value; preprocessing the settlement data set of the target area to obtain a preprocessed data set; constructing a neural network model, training the neural network model based on a sample set, and obtaining a surface settlement prediction model; inputting the settlement data set of the target area into the surface settlement prediction model for calculation, and generating a surface settlement prediction result. The present invention can effectively capture time series data caused by foundation pit dewatering, such as the change of precipitation over time, and its influence on surrounding surface settlement.

Description

Technical Field

[0001] The invention belongs to the technical field of surface settlement prediction, and in particular relates to a method for predicting surrounding surface settlement caused by foundation pit dewatering based on a neural network. Background Art

[0002] With the development of my country's economy and technology, high-rise buildings and urban subways are increasing, and deep foundation pit projects are also growing by leaps and bounds. In cities with abundant groundwater, the rich groundwater resources and high groundwater levels have a great impact on the construction of foundation pit projects, and foundation pit dewatering has become an indispensable measure in underground construction. As the dewatering process continues to advance, problems such as settlement and lateral inclination of surrounding terrain, underground pipelines, and adjacent buildings have come one after another, and the damage caused by this cannot be underestimated. Recharge technology is often used in projects to reduce the settlement caused by foundation pit dewatering, but the construction period is bound to be extended and the cost will increase. Therefore, it is of great significance to accurately predict the surrounding surface settlement caused by foundation pit dewatering, judge the effective influence radius of foundation pit dewatering, and thus determine the optimal location of the water collection well to ensure the safety of construction projects.

[0003] Surface subsidence is affected by multiple factors, including hydrogeological conditions, water-stopping structures, and water level drop. These factors have highly nonlinear relationships and the actual values ​​cannot be accurately calculated. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method for predicting surrounding surface settlement caused by foundation pit dewatering based on a neural network to solve the problems existing in the above prior art.

[0005] To achieve the above object, the present invention provides a method for predicting surrounding surface settlement caused by foundation pit dewatering based on a neural network, comprising:

[0006] Acquire settlement data of a target area, calculate the contribution of the settlement data of the target area, generate a contribution value, and construct a settlement data set of the target area based on the contribution value;

[0007] Preprocessing the target area settlement data set to obtain a preprocessed data set;

[0008] Constructing a Transformer model, and training the Transformer model based on the sample set to obtain a surface subsidence prediction model;

[0009] The target area settlement data set is input into the surface settlement prediction model for calculation to generate a surface settlement prediction result.

[0010] Preferably, the process of constructing a target area settlement data set based on the contribution value includes:

[0011] Calculating the characteristic values ​​of the settlement data of the target area by PCA to generate data characteristic values;

[0012] Acquire a sedimentation contribution based on the sum of the data characteristic values ​​and the data characteristic values, and generate a contribution curve based on the sedimentation contribution;

[0013] Setting a contribution curve threshold, screening the contribution curve based on the contribution curve threshold, and obtaining a screening characteristic value;

[0014] The target area sedimentation data set is constructed based on the screening feature values.

[0015] Preferably, the target area settlement data set is preprocessed to obtain the preprocessed data set, comprising:

[0016] Obtaining time generation information of data in the target area settlement data set;

[0017] Cleaning abnormal data of the target area settlement data set, and then generating labels for the target area settlement data set based on the time generation information to obtain a label data set;

[0018] The label data set is converted into vector information to generate the preprocessed data set.

[0019] Preferably, the process of converting the label data set into vector information includes:

[0020] Converting the label data set into data vector information;

[0021] Constructing an attention mechanism module, inputting the data vector information into the attention mechanism module for processing, and obtaining a feature map of the target settlement area in time and space;

[0022] Processing the spatial and temporal feature map of the target subsidence area based on the SaveReLU activation function to obtain a time series weight vector;

[0023] The timing weight vector is updated to generate the vector information.

[0024] Preferably, the process of obtaining the surface subsidence prediction model includes:

[0025] Constructing a time processing module and an encoder-decoder architecture, introducing the time processing module into the encoder-decoder architecture, and generating the Transformer model;

[0026] The Transformer model is trained using a public data sample set to obtain the surface subsidence model.

[0027] Preferably, the time processing module is used to segment the data according to the time axis of the surrounding surface settlement caused by foundation pit dewatering. When the time interval is greater than the time threshold, the time processing module segments the data of the target area settlement data set and uses the segmented data as new data for calculation. At the attention mechanism level, the weight of the data of the target area settlement data set whose interval time exceeds the threshold is assigned to 0.

[0028] Preferably, the process of training the Transformer model using a public data sample set includes:

[0029] Obtain a public data sample set, and annotate the public data sample set to obtain a training set and a test set;

[0030] The Transformer model is trained by using the training set and introducing a temperature parameter as well as a supervised learning loss function and a distillation loss function to obtain a training model;

[0031] The training model is fine-tuned on the task of data scheduling balance for surface subsidence prediction, and tested through the test set to obtain the surface subsidence prediction model.

[0032] Preferably, after obtaining the training model, the method further includes: freezing the bottom-level parameters of the training model and optimizing the upper-level parameters of the training model.

[0033] Preferably, the process of inputting the target area settlement data set into the surface settlement prediction model for calculation includes: inputting the target area settlement data set into the encoder of the surface settlement prediction model and dividing it into a time vector and a space vector, dividing the time vector and the space vector through a time processing module, and then calculating through a decoder to generate the surface settlement prediction result.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] The Transformer model of the present invention is suitable for processing sequence data, and therefore can effectively capture time series data caused by foundation pit precipitation, such as changes in precipitation over time, and its impact on surrounding surface settlement. At the same time, the Transformer model can capture the dependencies between different positions in the input sequence through the self-attention mechanism, so as to better understand the global relationship between precipitation and surface settlement, rather than just being limited to a local spatial or temporal range. The Transformer model can perform end-to-end learning without the need for manual feature extraction, but directly learns features and patterns related to surface settlement from the original precipitation data, thereby simplifying the modeling process. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 The present invention is a flowchart of a method for predicting surrounding surface settlement caused by foundation pit dewatering based on a neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0040] Embodiment 1

[0041] like Figure 1 As shown, this embodiment provides a method for predicting surrounding surface settlement caused by foundation pit dewatering based on a neural network, including:

[0042] Acquire settlement data of a target area, calculate the contribution of the settlement data of the target area, generate a contribution value, and construct a settlement data set of the target area based on the contribution value;

[0043] The target area settlement data set includes the buried depth of the water-stop curtain m, the water level drop S w , water collection well diameter r w , Distance from water collection well r o , permeability coefficient K, aquifer thickness H0.

[0044] Preprocessing the target area settlement data set to obtain a preprocessed data set;

[0045] Linear transformation of the original data so that the resulting value is mapped to [0-1]. The conversion function is as follows:

[0046]

[0047] Among them, max is the maximum value of the sample data, and min is the minimum value of the sample data. This normalization method is more suitable for the case of a data set of settlement data. However, if max and min are unstable, it is easy to make the normalization result unstable, and the subsequent use effect is also unstable. In actual use, max and min can be replaced by empirical constant values. Moreover, when new data is added, max and min may change and need to be redefined.

[0048] Constructing a Transformer model, and training the Transformer model based on the sample set to obtain a surface subsidence prediction model;

[0049] The target area settlement data set is input into the surface settlement prediction model for calculation to generate a surface settlement prediction result.

[0050] Further optimizing the scheme, the process of constructing the target area settlement data set based on the contribution value includes:

[0051] Calculating the characteristic values ​​of the settlement data of the target area by PCA to generate data characteristic values;

[0052] Acquire a sedimentation contribution based on the sum of the data characteristic values ​​and the data characteristic values, and generate a contribution curve based on the sedimentation contribution;

[0053] Setting a contribution curve threshold, screening the contribution curve based on the contribution curve threshold, and obtaining a screening characteristic value;

[0054] The target area sedimentation data set is constructed based on the screening feature values.

[0055] For a given data set, the variance or eigenvalue of each principal component is first calculated through a dimensionality reduction algorithm such as PCA. These variances or eigenvalues ​​represent the amount of information contained in each principal component.

[0056] Calculate the sum of the variances or eigenvalues ​​of all principal components. This sum represents the total variance or total information of the original data set.

[0057] For each principal component, its variance or eigenvalue is divided by the total variance or total eigenvalue to obtain its sedimentation contribution. This step is to measure the proportion of the amount of information contained in each principal component relative to the amount of information in the entire data set.

[0058] The cumulative sedimentation contribution can be calculated by adding up the sedimentation contributions of each principal component to obtain a cumulative contribution curve. This curve can help determine how many principal components to retain to retain most of the information in the data set.

[0059] The scheme is further optimized, and the settlement data set of the target area is preprocessed. The process of obtaining the preprocessed data set includes:

[0060] Obtaining time generation information of data in the target area settlement data set;

[0061] Cleaning abnormal data of the target area settlement data set, and then generating labels for the target area settlement data set based on the time generation information to obtain a label data set;

[0062] The label data set is converted into vector information to generate the preprocessed data set.

[0063] Further optimizing the scheme, the process of converting the label data set into vector information includes:

[0064] Converting the label data set into data vector information;

[0065] Constructing an attention mechanism module, inputting the data vector information into the attention mechanism module for processing, and obtaining a feature map of the target settlement area in time and space;

[0066] Processing the spatial and temporal feature map of the target subsidence area based on the SaveReLU activation function to obtain a time series weight vector;

[0067] The timing weight vector is updated to generate the vector information.

[0068] To further optimize the scheme, the process of obtaining the surface subsidence prediction model includes:

[0069] Constructing a time processing module and an encoder-decoder architecture, introducing the time processing module into the encoder-decoder architecture, and generating the Transformer model;

[0070] The Transformer model is trained using a public data sample set to obtain the surface subsidence model.

[0071] As with the original Transformer architecture, the relevance of each value to a given query is determined by the dot product between the query and the key corresponding to that value, such as:

[0072]

[0073] Among them, Q i ,K i ,V i Based on the characteristic map of foundation pit dewatering, different linear changes are obtained. Softmax is a normalized exponential function, T represents the transposition operation, and d k is the smoothing factor;

[0074] The softmax function is used to convert a vector with an arbitrary real number range into a probability distribution, and its expression is as follows:

[0075] Given a vector containing (n) real values ​​(\mathbf{z}=(z_1,z_2,...,z_n)), the softmax function converts it into a probability distribution vector with the same dimension (\mathbf{\sigma}=(\sigma_1,\sigma_2,...,\sigma_n)), where each element (\sigma_i) is defined as:

[0076] [\sigma_i=\frac{e^{z_i}}{\sum_{j=1}^{n}e^{z_j}}];

[0077] Where (e) is the base of natural logarithms and (\sum_{j=1}^{n}e^{z_j}) is the sum of the exponential functions of all elements in the vector (\mathbf{z}).

[0078] In this way, the softmax function ensures that each element in the output vector is in the range of ([0,1]) and the sum of all elements is equal to 1, so it can be interpreted as a probability distribution. The softmax function is often used in multi-class classification problems, where the output of the model needs to be expressed as the probability of each category.

[0079] Encoder loss function L GAN for:

[0080]

[0081] Where G(SP,Eg) represents the image data synthesized by the feature point matrix SP, and Ig is the real data. D(·) is the output of the encoder, Indicates an operation that seeks expectations on data.

[0082] Perceptual loss L P Used to improve the perceived quality of data, the formula is as follows:

[0083]

[0084] where F i (·) represents the feature representation of the input at the i-th layer in pre-training, and N is the number of extracted feature layers.

[0085] In order to increase the nonlinearity of the model, this embodiment applies a feedforward network to the output of the multi-head attention, that is, two sets of linear transformations and ReLU activation functions, such as:

[0086] H=FNN(M)=ReLU(MW1+b1)W2+b2

[0087] Where M represents the multi-head attention output, W1, W2, b1 and b2 are learnable parameters during training.

[0088] In the self-attention mechanism, the Transformer model can simultaneously focus on different positions in the input sequence. Multi-scale self-attention enables the model to acquire information at different scales by using attention weights of different scales in different layers or heads. This approach helps the model consider both long-distance dependencies and local relationships at the same time.

[0089] To further optimize the solution, the Transformer model can introduce multi-scale information through cross-layer connections. By combining or fusing the outputs of different layers, the model can obtain multi-scale representations at different levels. This approach helps the model capture features of different scales at different levels.

[0090] A further optimization scheme is provided, in which the time processing module is used to segment the data according to the time axis of the surrounding surface settlement caused by foundation pit dewatering. When the time interval is greater than the time threshold, the time processing module segments the data of the target area settlement data set and uses the segmented data as new data for calculation. At the attention mechanism level, the weight of the data of the target area settlement data set whose interval time exceeds the threshold is assigned to 0.

[0091] To further optimize the scheme, the process of training the Transformer model using a public data sample set includes:

[0092] Obtain a public data sample set, and annotate the public data sample set to obtain a training set and a test set;

[0093] The Transformer model is trained by using the training set and introducing a temperature parameter as well as a supervised learning loss function and a distillation loss function to obtain a training model;

[0094] The training model is fine-tuned on the task of data scheduling balance for surface subsidence prediction, and tested through the test set to obtain the surface subsidence prediction model.

[0095] The model of this embodiment uses a masking mechanism in the encoder and decoder, that is, when obtaining the dot product matrix Q i K i T Then replace the upper triangular part with -∞. The attention weights of subsequent positions will all be reset to zero due to the softmax operation. The union of h attention heads is multiplied by W o, to summarize the outputs of different attention heads. This union tensor is the final output of multiple attention heads, as shown below.

[0096]

[0097] MultiHead(Q i ,K i ,V i )=Concat(head1,…,head h )W o ;

[0098] A further optimization scheme, after obtaining the training model, also includes: freezing the bottom-level parameters of the training model and optimizing the upper-level parameters of the training model.

[0099] A further optimization scheme includes the following steps: inputting the target area settlement data set into the surface settlement prediction model for calculation, the process comprising: inputting the target area settlement data set into the encoder of the surface settlement prediction model and dividing it into a time vector and a space vector; dividing the time vector and the space vector through a time processing module; and then calculating through a decoder to generate the surface settlement prediction result.

[0100] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering, characterized in that: The following steps are involved: Acquire settlement data of a target area, calculate the contribution of the settlement data of the target area, generate a contribution value, and construct a settlement data set of the target area based on the contribution value; Preprocessing the target area settlement data set to obtain a preprocessed data set; Constructing a Transformer model, and training the Transformer model based on the sample set to obtain a surface subsidence prediction model; Inputting the target area settlement data set into the surface settlement prediction model for calculation to generate a surface settlement prediction result; The process of constructing the target area settlement data set based on the contribution value includes: Calculating the characteristic values ​​of the settlement data of the target area by PCA to generate data characteristic values; Acquire a sedimentation contribution based on the sum of the data characteristic values ​​and the data characteristic values, and generate a contribution curve based on the sedimentation contribution; Setting a contribution curve threshold, screening the contribution curve based on the contribution curve threshold, and obtaining a screening characteristic value; The target area sedimentation data set is constructed based on the screening feature values.

2. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 1 is characterized in that: The target area settlement data set is preprocessed to obtain the preprocessed data set, including: Obtaining time generation information of data in the target area settlement data set; Cleaning abnormal data of the target area settlement data set, and then generating labels for the target area settlement data set based on the time generation information to obtain a label data set; The label data set is converted into vector information to generate the preprocessed data set.

3. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 2 is characterized in that: The process of converting the label data set into vector information includes: Converting the label data set into data vector information; Constructing an attention mechanism module, inputting the data vector information into the attention mechanism module for processing, and obtaining a feature map of the target settlement area in time and space; Processing the spatial and temporal feature map of the target subsidence area based on the SaveReLU activation function to obtain a time series weight vector; The timing weight vector is updated to generate the vector information.

4. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 1 is characterized in that: The process of obtaining the surface subsidence prediction model includes: Constructing a time processing module and an encoder-decoder architecture, introducing the time processing module into the encoder-decoder architecture, and generating the Transformer model; The Transformer model is trained using a public data sample set to obtain the surface subsidence prediction model.

5. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 4 is characterized in that: The time processing module is used to segment the data according to the time axis of the surrounding surface settlement caused by foundation pit dewatering. When the time interval is greater than the time threshold, the time processing module segments the data of the target area settlement data set and uses the segmented data as new data for calculation. At the attention mechanism level, the weight of the data of the target area settlement data set whose interval time exceeds the threshold is assigned to 0.

6. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 4 is characterized in that: The process of training the Transformer model using a public data sample set includes: Obtain a public data sample set, and annotate the public data sample set to obtain a training set and a test set; The Transformer model is trained by using the training set and introducing a temperature parameter as well as a supervised learning loss function and a distillation loss function to obtain a training model; The training model is fine-tuned on the task of data scheduling balance for surface subsidence prediction, and tested through the test set to obtain the surface subsidence prediction model.

7. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 6 is characterized in that: After obtaining the training model, the method further includes: freezing the bottom-layer parameters of the training model and optimizing the upper-layer parameters of the training model.

8. The neural network-based method for predicting surrounding surface settlement caused by foundation pit dewatering according to claim 1 is characterized in that: The process of inputting the target area settlement data set into the surface settlement prediction model for calculation includes: inputting the target area settlement data set into the encoder of the surface settlement prediction model and dividing it into a time vector and a space vector, dividing the time vector and the space vector through a time processing module, and then calculating through a decoder to generate the surface settlement prediction result.

Images