A deep foundation pit deformation prediction method based on a neural network model

Abstract

This invention belongs to the field of deep foundation pit deformation prediction technology, specifically involving a method for predicting deep foundation pit deformation based on a neural network model. First, a neural network prediction model is established; second, a dataset is constructed; third, the model is initialized; then, the model is trained; and finally, the prediction results are output. It is based on a BiLSTM-Attention combined neural network algorithm. The BiLSTM module performs bidirectional processing on the input time-series data and controls the flow of information through input gates, forget gates, and output gates. The Attention module assigns an attention weight to the features at each time step. The output features of the BiLSTM module are weighted and summed using these attention weights to obtain a key feature representation. This allows the prediction model to focus on the time-series features that have the most impact on the prediction results. Through continuous training, the prediction model will ultimately predict more accurate deep foundation pit excavation deformation results based on the input data.

Description

Technical fields:

[0001] This invention belongs to the field of deep foundation pit deformation prediction technology, specifically involving a deep foundation pit deformation prediction method based on a neural network model, which provides basic data support for engineering construction quality acceptance and safety early warning. Background technology:

[0002] Because deep foundation pit excavation can significantly impact surrounding facilities, predicting the deformation of deep foundation pits is a common prediction problem in engineering. The goal is to accurately predict the deformation of deep foundation pits by learning from and training on historical deformation data, while considering key influencing factors, and then take corresponding control measures to ensure the safety of the construction process. Predicting the deformation of deep foundation pits typically uses monitoring data from the past N days as input, combined with algorithms such as LSTM (Long Short-Term Memory), which are commonly used for processing time series problems. For example, Chinese Patent 201710780941.5 discloses a neural network-based method for monitoring and early warning of foundation pit excavation, which includes the following steps:

[0003] (1) Layout of foundation pit displacement monitoring points

[0004] According to the "Technical Specification for Monitoring of Foundation Pit Engineering" GB50497-2009 and the "Code for Monitoring Construction of Foundation Pit Engineering" (DG / TJ08-2001-2006), M displacement monitoring benchmarks are arranged in a stable and deformation-free area three times the excavation depth (3H) outside the excavation surface of the foundation pit. L displacement monitoring points are arranged along the perimeter of the foundation pit at the top of the slope or the top of the retaining wall. The displacement monitoring benchmarks and displacement monitoring points together form a displacement deformation monitoring control network of the excavation surface. As the foundation pit is excavated, the horizontal and vertical displacements of the top of the slope at different excavation depths are monitored.

[0005] (2) Determination of horizontal and vertical displacements at different excavation depths at the top of the foundation pit slope

[0006] Taking a section around the top of the slope as an example, the number of monitoring points on this section is N;

[0007] (3) Determination of the average horizontal and vertical displacements at different excavation depths at the top of the foundation pit slope

[0008] The average horizontal and vertical displacement of N monitoring points is used as the deformation of this section of the foundation pit during the excavation process;

[0009] (4) Establishment of a BP neural network model based on foundation pit displacement

[0010] ①Establishment of the BP neural network model

[0011] Based on the modeling principle of the BP neural network prediction method, the BP network model is determined to consist of three parts: an input layer, a hidden layer, and an output layer. The input layer has one unit, represented by a set of excavation depth vectors; the output layer has one unit, represented by a set of foundation pit displacement vectors; the weights and thresholds from the input layer to the hidden layer are wjk and θjk, respectively, and the weights and thresholds from the hidden layer to the output layer are vkj and γkt, respectively.

[0012] ② Algorithm for establishing the BP neural network model

[0013] 1) According to the BP neural network model diagram, the net input and output of each neuron in the intermediate hidden layer are represented by equations (1) and (2), respectively;

[0014] 2) The net input and actual output of each neuron in the output layer are represented by equations (3) and (4), respectively;

[0015] 3) Based on the given expected output Y(hj), calculate the correction error dt of each neuron in the output layer, expressed by equation (5);

[0016] 4) The correction error ek of each neuron in the hidden layer is expressed by equation (6);

[0017] 5) Adjust the connection weights V from the hidden layer to the output layer and the threshold γ of the output layer neurons;

[0018] 6) Correct the connection weights W from the input layer to the hidden layer and the thresholds θ of the hidden layer neurons;

[0019] The next learning mode pair is randomly selected and provided to the network until all learning mode pairs have been trained. The network global error is then judged to see if it meets the accuracy requirements. If it does, the process ends; otherwise, it continues.

[0020] (5) Development of neural network programs for deep foundation pit deformation prediction

[0021] Based on the BP neural network model determined in step (4), the predicted value of the foundation pit displacement can be obtained by using the MATLAB neural network toolbox. The excavation depth vector group is used as the input sample, and the displacement vector group is used as the target vector. The network output can be obtained by loading it into the toolbox, which is the predicted value of the displacement at the next excavation depth hn+1.

[0022] (6) Prediction of the mean horizontal and vertical displacements of monitoring points during phased excavation of the foundation pit

[0023] Step 5: The BP neural network program is compiled to obtain the displacement prediction value at the excavation depth hn+1 based on the measured displacement values ​​at the excavation depths h1, h2, ..., hn. Then, based on the measured displacement values ​​at the excavation depths h1, h2, ..., hn, hn+1, the predicted value at the excavation depth hn+2 can be obtained. Modeling and prediction are carried out simultaneously and continue until the deformation law of the foundation pit can be reflected.

[0024] (7) Analysis and evaluation of foundation pit stability

[0025] In step 6, the predicted displacement values ​​of the monitoring points are adjusted to meet the error requirements and then compared with different levels of warning values. If the predicted displacement value is less than the blue warning level alarm value, the foundation pit is stable, and excavation proceeds according to the original plan. If the predicted displacement value exceeds the blue alarm value in the table below, it indicates that the foundation pit has entered the initial unstable stage. At this time, a blue warning should be issued to alert construction personnel and strengthen reinforcement measures in advance. If the predicted displacement value exceeds the orange alarm value in the table below, it indicates that the foundation pit has entered the intermediate unstable stage. At this time, an orange warning should be issued, and the excavation plan should be adjusted, with construction personnel further optimizing reinforcement measures. If the predicted displacement value exceeds the red alarm value in the table below, it indicates that the foundation pit is about to become unstable. Excavation will produce excessive displacement after the next step of excavation, leading to instability of the foundation pit sidewalls. At this time, excavation should be stopped, reinforcement measures should be further strengthened, and preparations should be made for timely evacuation of personnel. However, this method has the problems of low prediction accuracy and poor adaptability, but the prediction results have strong guiding significance for the improvement of construction.

[0026] Bidirectional Long Short-Term Memory (BiLSTM) neural networks are a special type of recurrent neural network that captures both forward and backward contextual information by inputting data into two independent LSTM networks and merging their outputs. This results in higher performance than traditional LSTM networks when processing sequential data.

[0027] BiLSTM effectively captures long-term dependencies by introducing three gate structures (input gate, forget gate, and output gate) to control the flow of information. These three gate structures are responsible for controlling the input information, retaining useful information, and outputting information, respectively.

[0028] (1) The forget gate determines which information needs to be forgotten or retained. Through a series of calculations, the forget gate outputs a value between 0 and 1 to each neuron. If a value in the forget gate is close to 1, the corresponding information is retained; if it is close to 0, the corresponding information is forgotten.

[0029] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0030] Where σ represents the sigmoid function, W f and b f These represent the weight and bias of the forget gate, respectively.

[0031] [h t-1 ,x t ] represents the concatenation of the previous hidden state and the current input, while f t It is the output of the forget gate.

[0032] (2) The input gate is responsible for updating the cell state. First, the sigmoid function determines which values ​​need to be updated; second, the tanh function creates a new vector of candidate values ​​that can be added to the state; finally, the output of the sigmoid function is multiplied by the output of the tanh function to determine the actual state value to be updated.

[0033] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0034]

[0035] Among them, i t It is the output of the input gate, W i and b i These represent the weights and biases of the input gate, respectively. It represents the candidate cell state, tanh denotes the hyperbolic tangent function, and W... C and b C f represents the weights and biases of the candidate cell states, respectively. t It is the output of the forget gate, C t-1 It is the previous cell state, while C t This is the updated cell state.

[0036] (3) The output gate determines the next hidden state. First, the sigmoid function determines which parts of the cell state will be output; then, the cell state is processed by the tanh function, and the result is multiplied by the output of the sigmoid function to obtain the final output.

[0037] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0038] h t =o t ·tanh(C t )

[0039] Among them, o t It is the output of the output gate. W oand b o C represents the weights and biases of the output gate, respectively. t It represents the current cell state, while h t It is currently in a hidden state.

[0040] Therefore, based on a neural network model, a method for predicting the deformation of deep foundation pits was developed to solve the above problems, so as to obtain accurate prediction results and provide better reference for actual construction in specific problems. Summary of the Invention:

[0041] The purpose of this invention is to overcome the shortcomings of existing technologies and develop a method for predicting the deformation of deep foundation pits based on a neural network model. By introducing an attention mechanism, the generalization ability and prediction accuracy of the model are improved, and the sparrow search algorithm is integrated to improve the efficiency of model parameter optimization, thus providing correct guidance for improving construction.

[0042] To achieve the above objectives, the specific process of a deep foundation pit deformation prediction method based on a neural network model, as disclosed in this invention, is as follows:

[0043] S1, Establish a neural network prediction model

[0044] A prediction model based on a neural network model is established, comprising an input layer, a combined network layer, a parameter optimization layer, and an output layer.

[0045] The input layer is used to input feature vectors;

[0046] The combined network layer consists of a BiLSTM (Bidirectional Long Short-Term Memory) module and an Attention module. The BiLSTM module is used to process time series information, and the Attention module is used to enhance the prediction model's ability to capture key information.

[0047] The parameter optimization layer is the Sparrow Search Algorithm (SSA) layer, which is used to globally optimize the parameters of the entire neural network;

[0048] The output layer is used to output the prediction results;

[0049] S2, Building the dataset

[0050] First, the data required by the prediction model is normalized, and the training set and test set are divided in a 9:1 ratio.

[0051] Then, the input feature data is divided into three categories: the first category is environmental factors, the second category is construction factors, and the third category is monitoring data.

[0052] Finally, the monitoring data is normalized and split into input vectors and output vectors. The input vectors include all three types of feature data, and the output vectors are the third type of monitoring data.

[0053] S3, Model Initialization

[0054] S31, Set the parameters of the BiLSTM module according to the features of the input vector;

[0055] S32, set the number of heads for the Attention module and initialize the weight matrix;

[0056] S33, set the population size and number of iterations for the SSA algorithm, and use the mean squared error (MSE) as the fitness function;

[0057] S4, Model Training

[0058] S41, forward propagation, inputs the training set data into the BiLSTM module step by step through the input layer, and calculates the hidden state at each time step;

[0059] S42, calculate the attention score and weights, and input the output of the BiLSTM module into the Attention module to calculate the attention weights of the hidden state at each time step;

[0060] S43, calculate the fitness of the current prediction model, and the SSA algorithm updates the prediction model parameters based on the fitness;

[0061] S44, backpropagation is performed based on the updated thick prediction model parameters;

[0062] S45, Train according to the operations of S41-S44 until the termination condition is met;

[0063] S5, the output layer outputs the prediction results.

[0064] The BiLSTM module involved in this invention performs bidirectional processing on the input time series data (such as daily monitoring data in foundation pit engineering, including earth pressure, groundwater level changes, and retaining structure displacement), and controls the flow of information through input gates, forget gates, and output gates;

[0065] The Attention module assigns an attention weight to the features at each time step. The output features of the BiLSTM module are weighted and summed using the attention weights to obtain a key feature representation, enabling the prediction model to focus on the time series features that have the most impact on the prediction results.

[0066] The parameter optimization layer optimizes the initial weights and biases of neurons in the BiLSTM and Attention modules, simulating the population strategy of sparrows foraging for food to perform global optimization, preventing them from getting trapped in local optima and achieving better training results for the combined network.

[0067] The steps for training and optimizing the combined network layers are as follows:

[0068] S11, randomly initialize the positions of individuals in the population;

[0069] S12, calculate the fitness of each individual, and update the role of each individual based on the fitness;

[0070] S13, Update individual positions, that is, update the parameters of the combined network layer;

[0071] S14, repeat the operations of S12-S13 until the termination condition is met.

[0072] Compared with existing technologies, this invention is based on the BiLSTM-Attention ensemble neural network algorithm. The BiLSTM module performs bidirectional processing on the input time series data and controls the flow of information through input gates, forget gates, and output gates. The Attention module assigns an attention weight to the features at each time step. The output features of the BiLSTM module are weighted and summed using the attention weights to obtain a feature representation of the focus. This allows the prediction model to focus on the time series features that have the most impact on the prediction result. Through continuous training, the prediction model will eventually predict more accurate deep foundation pit excavation deformation results based on the input data. It uses the BiLSTM-Attention ensemble neural network algorithm to train deep foundation pit excavation deformation data and predicts the amount of deep foundation pit excavation deformation under different influencing factors. The parameter optimization layer of the SSA algorithm simulates the population strategy of sparrows foraging to globally optimize the weight parameters of the ensemble neural network, preventing it from getting trapped in local optima and achieving a better training effect for the ensemble network, thereby improving the accuracy of the prediction model. Attached image description:

[0073] Figure 1 This is a schematic diagram of the predictive model involved in the present invention.

[0074] Figure 2 This is a schematic diagram of the prediction model involved in the present invention. Detailed implementation method:

[0075] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0076] Example 1:

[0077] The specific process of the deep foundation pit deformation prediction method based on a neural network model involved in this embodiment is as follows:

[0078] S1, Establish a neural network prediction model

[0079] A prediction model based on a neural network model is established, comprising an input layer, a combined network layer, a parameter optimization layer, and an output layer.

[0080] The input layer is used to input feature vectors;

[0081] The combined network layer consists of a BiLSTM (Bidirectional Long Short-Term Memory) module and an Attention module. The BiLSTM module is used to process time series information, and the Attention module is used to enhance the prediction model's ability to capture key information.

[0082] The parameter optimization layer is the Sparrow Search Algorithm (SSA) layer, which is used to globally optimize the parameters of the entire neural network;

[0083] The output layer is used to output the prediction results;

[0084] S2, Building the dataset

[0085] First, the data required by the prediction model is normalized, and the training set and test set are divided in a 9:1 ratio.

[0086] Then, the input feature data is divided into three categories: the first category is environmental factors, including average temperature, humidity and rainfall; the second category is construction factors, including excavation volume, excavation depth and number of cross braces; and the third category is monitoring data, including average groundwater level, axial force, diaphragm wall horizontal displacement, surrounding surface settlement and pile top horizontal displacement.

[0087] Finally, the monitoring data is normalized and split into input vectors and output vectors. The input vectors include all feature data, and the output vectors are the third type of monitoring data.

[0088] S3, Model Initialization

[0089] S31, The parameters of the BiLSTM module are set according to the features of the input vector, including the input dimension, the number of hidden units and the activation function;

[0090] S32, set the number of heads for the Attention module and initialize the weight matrix;

[0091] S33, set the population size and number of iterations for the SSA algorithm, and use the mean squared error (MSE) as the fitness function;

[0092] S4, Model Training

[0093] S41, forward propagation, inputs the training set data into the BiLSTM module step by step through the input layer, and calculates the hidden state at each time step;

[0094] S42, calculate the attention score and weights, and input the output of the BiLSTM module into the Attention module to calculate the attention weights of the hidden state at each time step;

[0095] S43, calculate the fitness of the current prediction model, and the SSA algorithm updates the prediction model parameters based on the fitness;

[0096] S44, backpropagation is performed based on the updated thick prediction model parameters;

[0097] S45, Train according to the operations of S41-S44 until the termination condition is met;

[0098] S5, the output layer outputs the prediction results.

[0099] Among them, the BiLSTM module, based on bidirectional processing and gating mechanism, can extract rich time series features, which contain the changing patterns and interrelationships of data in the time dimension;

[0100] The parameter optimization layer utilizes the SSA algorithm to quickly locate a potentially optimal parameter region through random exploration by the scout and guidance by the producer. As the iteration progresses, the search within this optimal region is further refined, and the parameters of the prediction model are adjusted so that the prediction model can better fit the time series data of the foundation pit deformation, thereby improving the accuracy of the prediction.

[0101] Example 2:

[0102] This embodiment relates to a method for predicting the deformation of deep foundation pits based on a neural network model, which is used to predict the deformation of a foundation pit adjacent to the Yangtze River.

[0103] The excavation pit is approximately trapezoidal in shape, with an excavation depth of about 25.15m and a floor area of ​​about 17,000 square meters. 2 On the north side of the excavation, 140m from the edge of the excavation, lies the Yangtze River;

[0104] The retaining structure uses a diaphragm wall and incorporates lateral internal bracing to constrain ground deformation;

[0105] The construction method adopted is an excavation-support-dewatering pattern;

[0106] Within the project area, the underlying soil layers of the foundation pit area are, in order, miscellaneous fill, plain fill, silty clay, fine sand, fine and medium sand, strongly weathered rock layer, and moderately weathered rock layer. The strata within the foundation pit excavation area are mainly clay layers.

[0107] The main monitoring targets are the diaphragm walls. The horizontal and vertical displacement monitoring points of the diaphragm walls are arranged at intervals of approximately 20-25m along the perimeter of the foundation pit. The lateral supports are arranged at the corners of the foundation pit, and the lateral displacement of the pile tops is evenly distributed inside the foundation pit.

[0108] Environmental monitoring mainly focuses on surface settlement around the pit. Surface settlement monitoring points outside the pit are arranged at the center of the four sides of the pit, with five points on each side. Groundwater level monitoring points are arranged at intervals of 30-50 meters around the pit and evenly distributed inside the pit.

[0109] After construction begins, we will continuously analyze the trends in monitoring data, dynamically adjust monitoring strategies, and provide timely feedback on monitoring results.

[0110] The specific prediction process is as follows:

[0111] The dataset was compiled based on 227 days of construction logs. The first 200 days of data were used as the training set to train the parameters of the SSA-BiLSTM-Attention neural network and the parameters were dynamically adjusted. The last 27 days of data were used as the validation set, and the neural network parameters were not adjusted.

[0112] Based on the above parameters, the changes in groundwater level, axial force of cross bracing, surface settlement, horizontal deformation of diaphragm wall and horizontal displacement of pile top during the foundation pit excavation process are predicted. The environmental factors, construction factors and monitoring values ​​of the previous day are used as input features, and the predicted values ​​at the monitoring points are used as output features. After normalizing the input and output features, the loss function value is calculated. MSE and R2 are used as the loss function and evaluation index of the prediction model.

[0113] The results show that the numerical prediction trend is good in the five prediction tasks on the test set, indicating that the SSA-BiLSTM-Atention neural network model can handle the deformation in the excavation process of the foundation pit well.

[0114] From the perspective of the loss function MSE, the loss functions are relatively low for the prediction tasks of groundwater level, axial force and surface settlement, which are 3.55, 3.40 and 3.12 respectively; while for the prediction tasks of diaphragm wall horizontal deformation and pile top horizontal displacement, the amplitude of the stable loss function is relatively large, which are 4.49 and 3.95 respectively.

[0115] In terms of the coefficient of determination R2, the R2 values ​​for the five prediction tasks are 0.95, 0.91, 0.97, 0.87, and 0.89, respectively. The first three have a better fit, while the latter two have a worse fit.

[0116] Overall, the predicted values ​​are quite close to the actual values, indicating a good overall performance.

Claims

1. A method for predicting the deformation of deep foundation pits based on a neural network model, characterized in that, The specific process is as follows: S1, Establish a neural network prediction model A prediction model based on a neural network model is established, comprising an input layer, a combined network layer, a parameter optimization layer, and an output layer. S2, Building the dataset First, the data required by the prediction model is normalized, and the training set and test set are divided in a 9:1 ratio. Then, the input feature data is divided into three categories: the first category is environmental factors, the second category is construction factors, and the third category is monitoring data. Finally, the monitoring data is normalized and split into input vectors and output vectors. The input vectors include all three types of feature data, and the output vectors are the third type of monitoring data. S3, Model Initialization S31, Set the parameters of the BiLSTM module according to the features of the input vector; S32, set the number of heads for the Attention module and initialize the weight matrix; S33, set the population size and number of iterations for the SSA algorithm, and use the mean squared error as the fitness function; S4, Model Training S41, forward propagation, inputs the training set data into the BiLSTM module step by step through the input layer, and calculates the hidden state at each time step; S42, calculate the attention score and weights, and input the output of the BiLSTM module into the Attention module to calculate the attention weights of the hidden state at each time step; S43, calculate the fitness of the current prediction model, and the SSA algorithm updates the prediction model parameters based on the fitness; S44, backpropagation is performed based on the updated thick prediction model parameters; S45, Train according to the operations of S41-S44 until the termination condition is met; S5, the output layer outputs the prediction results; In step S1, the input layer is used to input the feature vector; The combined network layer consists of a BiLSTM module and an Attention module. The BiLSTM module is used to process time series information, and the Attention module is used to enhance the prediction model's ability to capture key information. The parameter optimization layer is a sparrow search algorithm layer, used to globally optimize the parameters of the entire neural network; The output layer is used to output the prediction results; Environmental factors include average temperature, humidity, and rainfall; construction factors include excavation volume, excavation depth, and number of cross braces; the third type of monitoring data includes average groundwater level, axial force, diaphragm wall horizontal displacement, surrounding surface settlement, and pile top horizontal displacement. The BiLSTM module processes the input time series data bidirectionally and controls the flow of information through input gates, forget gates, and output gates. The Attention module assigns an attention weight to the features at each time step. The output features of the BiLSTM module are weighted and summed using the attention weights to obtain a key feature representation, enabling the prediction model to focus on the time series features that have the most impact on the prediction results.

2. The method for predicting deep foundation pit deformation based on a neural network model according to claim 1, characterized in that, The parameter optimization layer performs global optimization on the initial weights and biases of neurons in the BiLSTM and Attention modules, simulating the population strategy of sparrows foraging.

3. A method for predicting the deformation of deep foundation pits based on a neural network model according to claim 1 or 2, characterized in that, The steps for training and optimizing the combined network layers in the parameter optimization layer are as follows: S11, randomly initialize the positions of individuals in the population; S12, calculate the fitness of each individual, and update the role of each individual based on the fitness; S13, Update individual positions, that is, update the parameters of the combined network layer; S14, repeat the operations of S12-S13 until the termination condition is met.

Images