Optuna-CNN-LSTM deformation multi-step prediction method and system based on space-time effect of foundation pit

Abstract

The invention discloses an Optuna-CNN-LSTM deformation multi-step prediction method and system based on the space-time effect of a foundation pit, and the method comprises the steps: carrying out the preprocessing of monitoring data, extracting key features, and obtaining a key feature data set; constructing a CNN-LSTM model, extracting spatial features of foundation pit deformation by using a convolutional neural network CNN, and capturing time features of foundation pit deformation by using a long short-term memory network LSTM; hyper-parameters of the CNN-LSTM model are optimized through an Optuna framework, and an optimal parameter combination is obtained; and foundation pit deformation prediction of a plurality of time steps in the future is carried out based on the optimized model. According to the method, the spatial features of the foundation pit deformation are extracted by adopting the CNN, the time features of the foundation pit deformation are captured by adopting the LSTM, the spatial and time features are comprehensively considered, and the foundation pit deformation prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the fields of engineering construction and computer technology, and in particular to an Optuna-CNN-LSTM deformation multi-step prediction method and system based on foundation pit spatiotemporal effects. Background Art

[0002] In modern engineering construction, deep foundation pit excavation is a common and important construction activity. Monitoring and prediction of foundation pit deformation is of great significance to ensure the safety and stability of the project. Existing methods for predicting foundation pit deformation mainly rely on numerical simulation and field monitoring data, but these methods have certain limitations. Numerical simulation often simplifies the distribution and mechanical parameters of the soil layer, making it difficult to accurately describe the stress-strain relationship, and methods based on finite element simulation are difficult to accurately predict daily deformation. Traditional field monitoring data analysis methods are less effective when dealing with highly nonlinear and time-dependent data. Summary of the invention

[0003] Purpose of the invention: The present invention aims to provide a high-precision and high-stability Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pits; another purpose of the present invention is to provide an Optuna-CNN-LSTM deformation multi-step prediction system based on the spatiotemporal effect of foundation pits.

[0004] Technical solution: The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit described in the present invention comprises the following steps:

[0005] Step S1, collecting monitoring data during foundation pit excavation, preprocessing the monitoring data, extracting key features, and obtaining a key feature data set;

[0006] Step S2, constructing a CNN-LSTM model, using a convolutional neural network (CNN) to extract the spatial characteristics of foundation pit deformation, and using a long short-term memory (LSTM) network to capture the temporal characteristics of foundation pit deformation;

[0007] Step S3: Optimize the hyperparameters of the CNN-LSTM model through the Optuna framework to obtain the optimal parameter combination;

[0008] Step S4: predicting foundation pit deformation for multiple future time steps based on the optimized model.

[0009] Furthermore, in step S2, the convolutional neural network CNN includes a convolution layer, an activation function layer, a batch normalization layer and a fully connected layer; wherein, the convolution layer does not add a pooling layer, keeping the time dimension unchanged; nonlinearity is introduced to the spatial features after convolution by introducing nonlinearity through the activation function, and whether to use the pooling layer for downsampling is determined according to the amount of data.

[0010] Furthermore, step S1 is specifically as follows:

[0011] Data collection and collation: Collect monitoring data during foundation pit excavation, including displacement data, settlement data and soil pressure data of underground continuous walls;

[0012] Data denoising: Use denoising algorithm to denoise the data, preferably wavelet transform;

[0013] Missing data filling: interpolation method is used to fill in the missing data;

[0014] Feature extraction and selection: Extract key features based on the characteristics of foundation pit deformation, including maximum lateral displacement, deformation rate and soil pressure change;

[0015] Data normalization: Normalize the key feature data to obtain the key feature data set.

[0016] Furthermore, step S2 is specifically as follows:

[0017] Construct a CNN-LSTM model; the convolutional neural network CNN includes a convolutional layer, an activation function layer, a batch normalization layer, and a fully connected layer; the long short-term memory network LSTM includes an input layer, an LSTM layer, and an output layer;

[0018] Convolution operation: The convolution layer of the convolutional neural network (CNN) performs convolution operations on the input data to extract the spatial relationship and features between the foundation pit monitoring points, generate a spatial feature map, and capture the deformation features of different locations of the foundation pit; no pooling layer is added to keep the time dimension unchanged;

[0019] Activation and pooling: After the convolution operation, use the activation function to activate the spatial features and add nonlinearity; determine whether to use the pooling layer for downsampling based on the amount of data to retain key feature information;

[0020] Time series processing: The input layer of the long short-term memory network LSTM receives the spatial features extracted by CNN and inputs the data into the LSTM network in time series. The LSTM unit captures the dynamic change characteristics of foundation pit deformation through the memory and forgetting mechanism;

[0021] Hidden state update: The LSTM network continuously updates the hidden state and cell state, memorizes the historical information of the pit deformation, and outputs the prediction result at the current moment.

[0022] Furthermore, step S3 is specifically as follows:

[0023] In the Optuna framework, the range of hyperparameters is determined according to the collected foundation pit monitoring data; any set of hyperparameters is determined from the determined range of hyperparameters as the current hyperparameters; the CNN-LSTM model is trained on the training set using the current hyperparameters, and the trained CNN-LSTM model is evaluated on the validation set to obtain the evaluation results, and the current hyperparameters are adjusted according to the evaluation results; this step is repeated until the preset number of iterations or time is reached; on the validation set, a set of hyperparameters with the best evaluation results is selected as the optimal hyperparameters; wherein the training set and the validation set are obtained by dividing the key feature data set in step S1 in proportion.

[0024] Furthermore, the Optuna framework uses intelligent pruning technology to terminate poorly performing hyperparameter configurations early.

[0025] Furthermore, the evaluation index is mean square error or cross entropy.

[0026] Furthermore, a regularization term is introduced into the mean square error as the loss function J(θ):

[0027]

[0028] in, is the mean square error loss function of the CNN-LSTM model at the i-th sample, m is the number of samples in the training set, is the square of the jth parameter, and λ is the regularization coefficient.

[0029] Furthermore, the CNN-LSTM model adopts a multi-output strategy with the input time step being x1 and the output time step being y1.

[0030] The Optuna-CNN-LSTM deformation multi-step prediction system based on the spatiotemporal effect of foundation pit described in the present invention comprises:

[0031] The data acquisition and processing module is used to collect and pre-process the monitoring data during the foundation pit excavation process, extract key features, and obtain a key feature data set;

[0032] The model building module is used to build a CNN-LSTM model, which uses the convolutional neural network (CNN) to extract the spatial characteristics of foundation pit deformation and the long short-term memory network (LSTM) to capture the temporal characteristics of foundation pit deformation;

[0033] The hyperparameter optimization module is used to optimize the hyperparameters of the CNN-LSTM model through the Optuna framework to obtain the optimal parameter combination;

[0034] The solution prediction module is used to predict the foundation pit deformation in multiple future time steps based on the optimized model.

[0035] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The present invention adopts convolutional neural network (CNN) to extract spatial characteristics of foundation pit deformation, and long short-term memory network (LSTM) to capture temporal characteristics of foundation pit deformation, comprehensively considering spatial and temporal characteristics to improve the accuracy of foundation pit deformation prediction; 2. The present invention optimizes hyperparameters through the Optuna framework to improve the stability and robustness of the model and ensure the reliability of the prediction results; 3. The present invention adopts CNN convolutional layer without adding pooling layer to keep the time dimension unchanged, and determines whether to use pooling layer for downsampling according to the amount of data, further retains key feature information, and improves the accuracy and reliability of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings.

[0038] The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit described in the present invention comprises the following steps:

[0039] Step S1: Collect monitoring data during foundation pit excavation, perform noise reduction and missing data filling on the monitoring data to ensure data quality, extract key features, and obtain a key feature data set.

[0040] Data collection and collation: Collect relevant monitoring data from the foundation pit excavation site, including displacement data of underground continuous walls, settlement data, soil pressure data, etc. Ensure that the data covers different stages of foundation pit excavation and important monitoring points.

[0041] Data noise reduction: Since the data collected on-site may contain noise, wavelet transform is used to reduce the noise of the data to ensure the accuracy and reliability of the data.

[0042] Missing data filling: For missing data in the monitoring process, linear interpolation, Lagrange interpolation or other appropriate methods are used to fill in the missing data to ensure the continuity and integrity of the data.

[0043] Feature extraction and selection: According to the characteristics of foundation pit deformation, key parameters are extracted, such as maximum lateral displacement, deformation rate, soil pressure change, etc. These features will be used as input data for subsequent model training and prediction.

[0044] Data normalization: In order to eliminate the differences between different feature dimensions, the data is normalized and the feature values ​​are scaled to the range of 0 to 1 to improve the efficiency and effect of model training.

[0045] Step S2: Construct a CNN-LSTM model, use the convolutional neural network (CNN) to extract the spatial characteristics of foundation pit deformation, and use the long short-term memory (LSTM) network to capture the temporal characteristics of foundation pit deformation.

[0046] The CNN part extracts the spatial features of foundation pit deformation through convolutional layers, and does not add pooling layers to keep the time dimension unchanged.

[0047] Construct CNN model: Design the convolutional neural network structure, including convolutional layer, activation function, batch normalization layer and fully connected layer. The convolutional layer is used to extract the spatial characteristics of foundation pit deformation, the activation function introduces nonlinearity, the batch normalization layer accelerates the training process, and the fully connected layer is used to map the high-dimensional feature vector to the low-dimensional output.

[0048] Convolution operation: The convolution layer performs convolution operation on the input data to extract the spatial relationship and features between the monitoring points of the foundation pit. The convolution kernel slides on the input data to generate a feature map to capture the deformation characteristics of different locations of the foundation pit.

[0049]

[0050] Among them, I is the input, K is the filter, θ represents the convolution operation, ij is the position of the output feature map, and (m, n represent the element position in the filter.

[0051] Feature map generation: The feature map generated by the convolution operation reflects the spatial distribution and changes of foundation pit deformation. Through the superposition of multiple convolutional layers, high-level spatial features are gradually extracted.

[0052] Activation and pooling: After the convolution operation, activation functions such as ReLU are applied to activate the feature map and add nonlinearity. Depending on the characteristics and needs of the data, decide whether to use the pooling layer for downsampling to retain key feature information. In the case of low data volume, the pooling layer can be omitted to maintain the fine features of the data.

[0053] Max pooling formula:

[0054]

[0055] Average pooling formula:

[0056]

[0057] The LSTM part captures the temporal characteristics of foundation pit deformation through a gating mechanism and outputs deformation prediction values ​​for multiple future time steps.

[0058] Build LSTM model: Design LSTM network structure, including input layer, LSTM layer and output layer. LSTM layer captures the time-dependent characteristics of foundation pit deformation through gating mechanism (input gate, forget gate and output gate), the input layer receives the spatial features extracted by CNN, and the output layer generates prediction results.

[0059] Time series processing: The preprocessed data is input into the LSTM network in time series. The LSTM unit captures the dynamic change characteristics of foundation pit deformation through memory and forgetting mechanisms.

[0060] Hidden state update: The LSTM network continuously updates the hidden state and cell state, memorizes the historical information of the pit deformation, and outputs the prediction result at the current moment.

[0061] LSTM unit calculation formula:

[0062] [I t =σ(W i x t +U i H t-1 +b i )]

[0063] [F t =σ(W f x t +U f H t-1 +b f )]

[0064] [O t =σ(W o x t +U o H t-1 +b o )]

[0065] [C t =F t ⊙C t-1 +I t ⊙tanh(C t )]

[0066] [H t =O t ⊙tanh(C t )]

[0067] Where: (x t ) represents the current input data, (H t-1 ) represents the output of the hidden layer at the previous moment, (I t )、(F t ) and (O t) represent the input gate, forget gate and output gate respectively, (W) represents the weight coefficient matrix of different gating units,

[0068] (U) represents the weight of the hidden layer of different gating units at the previous moment, (b) represents the bias term of different gating units,

[0069] (σ) is the sigmoid activation function, (C t-1 ) represents the state memory of the previous moment,

[0070] (C t ) represents the memory at the current moment, and the hyperbolic tangent activation function is represented by (tanh).

[0071] ⊙) is the product of the vector elements, (H t ) represents the output of the hidden layer at the current moment,

[0072] (C t ) represents the candidate cell state.

[0073] Multi-step prediction: The LSTM network generates deformation prediction values ​​for multiple future time steps based on historical data and current status, thus realizing multi-step prediction of foundation pit deformation.

[0074] Step S3: Optimize the hyperparameters of the CNN-LSTM model through the Optuna framework to obtain the optimal parameter combination.

[0075] In one possible implementation, the Optuna framework uses Bayesian optimization techniques to optimize hyperparameters such as learning rate, number of training epochs, batch size, and weight decay.

[0076] Define the hyperparameter space: According to the structure of CNN and LSTM models, define the hyperparameter space that needs to be optimized, including learning rate, number of training cycles, batch size, weight decay, etc.

[0077] Bayesian optimization: The Optuna framework uses Bayesian optimization technology to intelligently explore the hyperparameter space and continuously adjust the hyperparameters through trials to find the optimal combination.

[0078] Intelligent pruning technology: During the experiment, the Optuna framework uses intelligent pruning technology to terminate poorly performing hyperparameter configurations in advance, concentrate computing resources on more promising areas, and improve optimization efficiency.

[0079] Evaluation indicators: Set evaluation indicators (such as mean square error MSE), evaluate model performance based on the results of each trial, and adjust hyperparameter configurations based on the evaluation results.

[0080] Obtaining optimal hyperparameters: After multiple experiments, the Optuna framework finally determined a set of optimal hyperparameter combinations to improve the prediction accuracy and stability of the model.

[0081] In one possible implementation, the training data of the model is divided into a training set and a test set, the mean square error (MSE) is used as the loss function, and the L2 regularization term is introduced to prevent overfitting. The loss function of the machine learning model after adding the L2 regularization term is usually expressed as:

[0082]

[0083] Among them, J(θ) is the loss function after adding the regularization term. is the loss function of the model on the ith sample (e.g. mean squared error or cross entropy). m is the number of training samples. is the square of the jth parameter of the model. λ is a hyperparameter of the regularization strength, also known as the regularization coefficient, which controls the impact of the regularization term on the total loss.

[0084] In a possible implementation, the foundation pit deformation prediction model is based on a multi-step prediction framework, and the model is optimized in real time as the monitoring data is continuously updated.

[0085] In one possible implementation, a multi-output strategy is used to construct the model, with the input time step being x and the output time step being y.

[0086] Step S4: predicting foundation pit deformation for multiple future time steps based on the optimized model.

[0087] The foundation pit deformation prediction method based on the Optuna-CNN-LSTM model of the present invention can provide the following guidance for the actual construction process:

[0088] Real-time monitoring and early warning: During the excavation process of the foundation pit, the real-time monitoring data is input into the model to predict the deformation trend of multiple time steps in the future, providing early warning for construction personnel to prevent the occurrence of foundation pit instability accidents.

[0089] Optimize construction plan: According to the model prediction results, adjust the construction plan, optimize the excavation sequence and support measures, reduce foundation pit deformation, and ensure project safety.

[0090] Data-driven decision-making: Using the prediction capability of deep learning models and combining them with on-site monitoring data, data-driven construction decisions are made to improve construction management and project quality.

[0091] The Optuna-CNN-LSTM deformation multi-step prediction system based on the spatiotemporal effect of foundation pit described in the present invention comprises:

[0092] The data acquisition and processing module is used to collect and pre-process the monitoring data during the foundation pit excavation process, extract key features, and obtain a key feature data set;

[0093] The model building module is used to build a CNN-LSTM model, which uses the convolutional neural network (CNN) to extract the spatial characteristics of foundation pit deformation and the long short-term memory network (LSTM) to capture the temporal characteristics of foundation pit deformation;

[0094] The hyperparameter optimization module is used to optimize the hyperparameters of the CNN-LSTM model through the Optuna framework to obtain the optimal parameter combination;

[0095] The solution prediction module is used to predict the foundation pit deformation in multiple future time steps based on the optimized model.

Claims

1. An Optuna-CNN-LSTM deformation multi-step prediction method based on foundation pit spatiotemporal effects, characterized in that: The following steps are involved: Step S1, collecting monitoring data during foundation pit excavation, preprocessing the monitoring data, extracting key features, and obtaining a key feature data set; Step S2, constructing a CNN-LSTM model, using a convolutional neural network (CNN) to extract the spatial characteristics of foundation pit deformation, and using a long short-term memory (LSTM) network to capture the temporal characteristics of foundation pit deformation; Step S3: Optimize the hyperparameters of the CNN-LSTM model through the Optuna framework to obtain the optimal parameter combination; Step S4: predicting foundation pit deformation for multiple future time steps based on the optimized model.

2. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 1 is characterized in that: In step S2, the convolutional neural network CNN includes a convolution layer, an activation function layer, a batch normalization layer and a fully connected layer; wherein, the convolution layer does not add a pooling layer to keep the time dimension unchanged; nonlinearity is introduced to the spatial features after convolution by introducing nonlinearity through the activation function, and whether to use the pooling layer for downsampling is determined according to the amount of data.

3. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 1 is characterized in that: Step S1 is specifically as follows: Data collection and collation: Collect monitoring data during foundation pit excavation, including displacement data, settlement data and soil pressure data of underground continuous walls; Data denoising: Use denoising algorithm to denoise the data; Missing data filling: interpolation method is used to fill in the missing data; Feature extraction and selection: Extract key features based on the characteristics of foundation pit deformation, including maximum lateral displacement, deformation rate and soil pressure change; Data normalization: Normalize the key feature data to obtain the key feature data set.

4. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 3 is characterized in that: Step S2 is specifically as follows: Construct a CNN-LSTM model; the convolutional neural network CNN includes a convolutional layer, an activation function layer, a batch normalization layer, and a fully connected layer; the long short-term memory network LSTM includes an input layer, an LSTM layer, and an output layer; Convolution operation: The convolution layer of the convolutional neural network (CNN) performs convolution operations on the input data to extract the spatial relationship and features between the foundation pit monitoring points, generate a spatial feature map, and capture the deformation features of different locations of the foundation pit; no pooling layer is added to keep the time dimension unchanged; Activation and pooling: After the convolution operation, use the activation function to activate the spatial features and add nonlinearity; determine whether to use the pooling layer for downsampling based on the amount of data to retain key feature information; Time series processing: The input layer of the long short-term memory network LSTM receives the spatial features extracted by CNN and inputs the data into the LSTM network in time series. The LSTM unit captures the dynamic change characteristics of foundation pit deformation through the memory and forgetting mechanism; Hidden state update: The LSTM network continuously updates the hidden state and cell state, memorizes the historical information of the pit deformation, and outputs the prediction result at the current moment.

5. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 4 is characterized in that: Step S3 is as follows: In the Optuna framework, the range of hyperparameters is determined based on the collected foundation pit monitoring data; any set of hyperparameters is determined from the determined range of hyperparameters as the current hyperparameters; Use the current hyperparameters to train the CNN-LSTM model on the training set, evaluate the trained CNN-LSTM model on the validation set, obtain the evaluation results, and adjust the current hyperparameters based on the evaluation results; Repeat this step until the preset number of iterations or time is reached; on the validation set, select a set of hyperparameters with the best evaluation results as the optimal hyperparameters; wherein the training set and the validation set are obtained by dividing the key feature data set in step S1 in proportion.

6. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 5 is characterized in that: The Optuna framework uses intelligent pruning technology to terminate poorly performing hyperparameter configurations early.

7. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 5 is characterized in that: The evaluation metric is mean square error or cross entropy.

8. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 7 is characterized in that: A regularization term is introduced into the mean square error as the loss function J(θ): in, is the mean square error loss function of the CNN-LSTM model at the i-th sample, m is the number of samples in the training set, is the square of the jth parameter, and λ is the regularization coefficient.

9. The Optuna-CNN-LSTM deformation multi-step prediction method based on the spatiotemporal effect of foundation pit according to claim 1 is characterized in that: The CNN-LSTM model adopts a multi-output strategy, with the input time step being x1 and the output time step being y1.

10. An Optuna-CNN-LSTM deformation multi-step prediction system based on foundation pit spatiotemporal effects, characterized in that: include: The data acquisition and processing module is used to collect and pre-process the monitoring data during the foundation pit excavation process, extract key features, and obtain a key feature data set; The model building module is used to build a CNN-LSTM model, which uses the convolutional neural network (CNN) to extract the spatial characteristics of foundation pit deformation and the long short-term memory network (LSTM) to capture the temporal characteristics of foundation pit deformation; The hyperparameter optimization module is used to optimize the hyperparameters of the CNN-LSTM model through the Optuna framework to obtain the optimal parameter combination; The solution prediction module is used to predict the foundation pit deformation in multiple future time steps based on the optimized model.

Images