Data-physical mechanism dual-drive foundation pit monitoring data intelligent prediction algorithm

Abstract

The invention discloses a data-physical mechanism dual-driven foundation pit monitoring data intelligent prediction algorithm, which comprises the following steps that: the lateral displacement of a retaining wall at a plurality of past time steps is # imgabs0 #, and the lateral displacement of a future retaining wall to be predicted is # imgabs1 #; establishing a neural network model based on a planar elastic foundation beam method, and introducing a physical intermediate variable layer; setting a loss function based on the retaining wall sidesway and the simulated wall back active load sequence maximum value index, and constraining the simulated wall back active load under the condition that the real value of the wall back active load is unknown; a CNN-LSTM network is adopted to extract space-time information of retaining wall lateral displacement of a plurality of past time steps and simulated active load behind the wall; and fusing the two parts of spatio-temporal information, and outputting future retaining wall sidesway to be predicted through a full-connection neural network. The invention relates to the technical field of constructional engineering, and can solve the problems of poor generalization ability and adaptability of a pure data driven intelligent prediction algorithm in practical application in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of construction engineering, and particularly to an intelligent prediction algorithm for foundation pit monitoring data driven by both data - physical mechanism. Background Technique

[0002] Currently, the research hotspots of intelligent prediction algorithms for foundation pit monitoring data mainly focus on the following aspects: A. Hybrid models based on deep learning CNN - LSTM hybrid model: Combining the feature extraction ability of CNN and the sequence modeling ability of LSTM, first use CNN to extract local features from inclinometer data, and then input the extracted feature sequence into LSTM for long - term dependence modeling and prediction. This hybrid model can give full play to the advantages of both, being able to capture both local detailed features in inclinometer data and consider the long - term change trend of the data, thus improving the accuracy and reliability of prediction. For example, when predicting inclinometer data under complex working conditions of the foundation pit, the CNN - LSTM hybrid model can better handle the non - linearity and non - stationarity in the data.

[0003] LSTM - GRU hybrid model: Combining LSTM and GRU, and flexibly using the advantages of both according to different task requirements and data characteristics. For example, use LSTM in the early stage of the data to better capture the initial long - term dependence relationship, and then use GRU in the later stage to improve the prediction efficiency and real - time performance. Or use LSTM and GRU respectively on different time scales to achieve accurate prediction of inclinometer data within different time ranges.

[0004] The disadvantages of this hybrid model based on deep learning (taking CNN - LSTM as an example) are as follows: High complexity and resource requirements: The model combines the structures of CNN and LSTM, resulting in a significant increase in complexity. It requires a large amount of computing resources, including high - performance GPUs and other hardware support, and ordinary engineering computing devices may be difficult to meet its training requirements.

[0005] Long training time: Due to the complex structure and large amount of calculation, the training process is extremely time - consuming. In actual engineering, if the model needs to be quickly adjusted or real - time monitoring and prediction are required, the long training time will seriously affect work efficiency.

[0006] Difficult parameter tuning: It involves the parameter settings of two types of networks, CNN and LSTM, with a large parameter space. Professional personnel need to carry out careful debugging based on rich experience. Otherwise, it is difficult to achieve ideal prediction performance, and problems such as overfitting or underfitting are likely to occur.

[0007] B. Solutions based on recurrent neural network (RNN) and its variants Simple Recurrent Neural Network (Simple RNN): RNN can process sequential data and is suitable for modeling time series data such as inclinometer data of foundation pits. It can capture the sequential relationships in the data and learn the dependencies between inclinometer data at different times through the state transfer of the hidden layer. For example, when continuous inclinometer data is input into the Simple RNN in chronological order, the network will predict the inclinometer value at the next moment based on the previous inclinometer values, thus achieving short-term prediction of inclinometer data.

[0008] Long Short-Term Memory Network (LSTM): LSTM is a variant of RNN that can better handle long-term dependency problems in long sequences of data. In foundation pit monitoring, the inclinometer data may have complex variation patterns and correlations over a long time span. The memory cells of LSTM can selectively remember and forget information, thus more accurately capturing these long-term dependency relationships. For example, when predicting the long-term deformation trend of a foundation pit, LSTM can utilize its memory function to comprehensively consider various factors and variation trends in the early stage and make more reliable predictions of future inclinometer data.

[0009] Gated Recurrent Unit (GRU): GRU is also an improved RNN structure that controls the flow and update of information through a simplified gating mechanism. When processing inclinometer data of foundation pits, GRU can reduce the computational complexity and training time of the model while maintaining high prediction accuracy. For example, in a real-time monitoring system where it is necessary to quickly predict and analyze inclinometer data, the high efficiency of GRU makes it more suitable for scenarios with high requirements for timeliness.

[0010] The disadvantages of the solution based on the Recurrent Neural Network (RNN) and its variants are as follows: Simple RNN: Dilemma in processing long sequences: When processing long sequences of foundation pit monitoring data, problems such as vanishing gradients or exploding gradients are likely to occur. This will cause the model to be unable to effectively learn the long-term dependency relationships in the data, resulting in large deviations in long-term prediction results and unable to meet the requirements of long-term stability monitoring and prediction of engineering projects.

[0011] LSTM: High computational cost: Its complex structure and numerous gating units lead to a huge amount of computation. Whether in the training or prediction stage, it requires a large amount of computing resources, increasing the cost of engineering applications and the dependence on hardware facilities.

[0012] Low training efficiency: The training process is slow, especially in the foundation pit monitoring scenario with a large amount of data and high data dimensions, where long-term iterative training is required to converge, which is not conducive to adjusting the model in a timely manner according to the progress of the project.

[0013] GRU: Modeling accuracy limitations: Although the computational efficiency is relatively high, its accuracy is not as good as LSTM when dealing with complex long-term dependencies. For some foundation pit projects that require extremely high prediction accuracy, such as foundation pits adjacent to important buildings or underground pipelines, it may not be able to provide sufficiently accurate prediction results, thereby increasing project risks.

[0014] C. Deep learning scheme based on attention mechanism Combining attention mechanism with RNN: Adding attention mechanism to RNN can make the model pay more attention to important information in the input sequence when predicting. For foundation pit inclination data, the attention mechanism can automatically focus on those moments or data features that have a greater impact on the prediction results, thereby improving the accuracy of the prediction. For example, in the presence of abnormal data or complex data changes, the attention mechanism can help the model ignore noise and irrelevant information and more accurately capture the key inclination data change trends.

[0015] Combining attention mechanism with CNN: Integrating the attention mechanism into CNN can make CNN pay more attention to important areas in the image or data when extracting features. When processing two-dimensional foundation pit monitoring data (such as images composed of inclination data at different locations), the attention mechanism can highlight areas closely related to inclination changes, thereby improving the effectiveness of feature extraction and the accuracy of prediction.

[0016] The disadvantages of this deep learning solution based on the attention mechanism are: Model complexity and computational burden: The attention mechanism itself introduces additional computational modules and parameters, increasing the overall complexity of the model. When processing large-scale foundation pit monitoring data, the computational burden will be further increased, which may lead to prediction delays and fail to meet the needs of real-time monitoring and early warning.

[0017] Data and training dependency: Its effectiveness depends largely on the characteristic distribution and quality of the data. If there are problems such as noise, outliers, or insufficient data in the foundation pit monitoring data, the attention mechanism may not be able to accurately focus on key information, or even be misled, thereby reducing the accuracy of the prediction. At the same time, the training process is very sensitive to the setting of training parameters. Inappropriate parameters will make the attention mechanism ineffective and affect the prediction performance of the entire model.

[0018] In the application of deep learning prediction algorithms for inclinometer data in foundation pit monitoring, RNN and its variants have their own advantages and disadvantages. SimpleRNN is suitable for initial monitoring but performs poorly in processing long sequences. LSTM is strong in handling long dependencies but consumes a large amount of resources. GRU has good real-time performance but slightly lacks in accuracy. In terms of CNN, one-dimensional CNN is good at capturing local features but difficult to handle long-term dependencies. Two-dimensional CNN has good multi-dimensional analysis but has high data requirements. Hybrid models such as CNN-LSTM take into account multiple aspects but have high complexity. LSTM-GRU is flexible but difficult to optimize. The attention mechanism-based solution can focus on key information to improve accuracy, but the complexity and computational volume also increase, and it depends on data features and training quality. Overall, each solution needs to be weighed and selected according to the specific situation of the foundation pit project in practical applications.

[0019] In the practical application of foundation pit engineering, although the deep learning hybrid model (such as CNN-LSTM) can combine the advantages of multiple networks to improve the prediction accuracy, the model has high complexity, consumes a large amount of computational resources and takes a long time during training, the parameter tuning process is complex, and it is necessary to take into account multiple network parameters, which requires high professional technology. Based on the recurrent neural network (RNN) and its variants, like Simple RNN has weak ability to process long sequence data, is vulnerable to gradient vanishing or explosion, and the long-term prediction is inaccurate. Although LSTM can handle long dependencies, its structure is complex and the training is slow. GRU has high computational efficiency but lacks accuracy in modeling complex long-term dependencies. The deep learning solution based on the attention mechanism can focus on key information to improve the prediction quality, but the attention mechanism itself increases the model complexity and computational volume, and its effect highly depends on data features and training quality. If the data preprocessing is improper or the training parameter settings are unreasonable, it may not be able to effectively play its role or even interfere with the prediction results, and it is limited in application scenarios with less data volume or less obvious features.

[0020] Therefore, the pure data-driven intelligent prediction algorithm of the existing technology has poor generalization ability and poor adaptability to different projects in the practical application of foundation pit engineering, and there is a need to provide a data-physical mechanism dual-driven intelligent prediction algorithm for foundation pit monitoring data to solve the above problems. Summary of the Invention

[0021] The purpose of the present invention is to provide a data-physical mechanism dual-driven intelligent prediction algorithm for foundation pit monitoring data, which can solve the problems of poor generalization ability and poor adaptability to different projects of the pure data-driven intelligent prediction algorithm in the practical application of foundation pit engineering.

[0022] The present invention is implemented as follows: A data-physical mechanism dual-driven intelligent prediction algorithm for foundation pit monitoring data includes the following steps: Step 1: Denote the lateral displacement of the retaining wall in the past several time steps as where represents from To the time dimension, represents the depth dimension from 1 to , that is, the spatial dimension; the future lateral displacement of the retaining wall to be predicted is denoted as ; Step 2: Based on the plane elastic foundation beam method, establish a neural network model to predict the future lateral displacement of the retaining wall, and introduce a physical intermediate variable layer to simulate the active load behind the wall effect; Step 3: Set a loss function based on the maximum value index of the sequence of the lateral displacement of the retaining wall in the past several time steps of the model input and the simulated active load behind the wall to constrain the simulated active load behind the wall when the true value is unknown; ; Step 4: After calculating the simulated active load behind the wall , use the CNN-LSTM network to extract the spatio-temporal information of the lateral displacement of the retaining wall in the past several time steps and the simulated active load behind the wall ; Step 5: Integrate the two parts of spatio-temporal information extracted in Step 4, and output the final future lateral displacement of the retaining wall to be predicted , which is used to comprehensively reflect the lateral displacement of the retaining wall in the past several time steps and the role of the simulated active load behind the wall in predicting the future lateral displacement of the retaining wall.

[0023] In the said Step 2, the input of the neural network model is the lateral displacement of the retaining wall in the past several time steps , the output is the predicted future lateral displacement of the retaining wall , and the simulated active load behind the wall is calculated through the LSTM network that maintains monotonicity , and the simulated active load behind the wall is continuously updated during the training process of the neural network model .

[0024] In the said Step 3, in the loss function, the analytical solution of the ground settlement caused by the foundation pit excavation is incorporated into the loss function, and several correction parameters are introduced into the loss function, and the analytical solution of the ground settlement is dynamically corrected during the training process.

[0025] Compared with the prior art, the present invention has the following beneficial effects: 1. Since the present invention introduces a physical intermediate variable layer into a pure data-driven intelligent prediction algorithm, through a dual-drive mode of data-physical mechanism, it can predict relevant data more accurately. In the prediction of retaining wall lateral displacement, the introduction of the physical intermediate variable layer simulates the active load action behind the wall, fully considering the physical mechanism and reducing the prediction error.

[0026] 2. Since the present invention introduces a physical intermediate variable layer into the retaining wall lateral displacement prediction algorithm based on the plane elastic foundation beam method, the algorithm can exhibit good stability in the complex and changeable foundation pit engineering environment when facing different geological conditions, construction techniques, and external interference factors, maintain relatively stable prediction performance, improve the generalization ability and robustness of the algorithm, and will not produce large deviations due to small fluctuations or anomalies in the data, providing a reliable basis for the long-term stable monitoring of the project.

[0027] 3. Since the present invention introduces a physical intermediate variable layer and integrates the physical mechanism, it has stronger interpretability compared with traditional pure data-driven intelligent prediction algorithms. For example, in the prediction of retaining wall lateral displacement, based on the physical process simulated by the physical intermediate variable layer, the correlation between the algorithm prediction result and the actual physical phenomenon can be intuitively explained, making it easier for engineering and technical personnel to understand and accept the output of the algorithm, facilitating analysis, optimization, and decision-making in engineering practice, and being conducive to the popularization and application of the technology.

[0028] 4. Since the present invention sets a loss function, it effectively reduces the dependence on the amount of training data. By introducing an analytical solution and correction parameters, good prediction results can still be achieved when the amount of data is relatively small. This not only saves the time and cost of data collection and collation, but also improves the applicability of the algorithm in scenarios with limited data resources, speeds up the project progress, and reduces the upfront investment in the project.

[0029] 5. The neural network model guided by physical mechanism designed in the present invention embeds the calculation theory of retaining wall lateral displacement in the plane elastic foundation beam method. By introducing a physical intermediate variable layer to simulate the action of the active load behind the wall, and through the LSTM network that maintains monotonicity and the loss function based on the sequence maximum index, a certain degree of constraint is imposed on the simulated active load behind the wall when the true value of the active load behind the wall is unknown, making it more conform to physical laws and engineering experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the structural diagram of the intelligent prediction algorithm for foundation pit monitoring data with a dual-drive mode of data-physical mechanism of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The present invention will be further described below in conjunction with the drawings and specific embodiments.

[0032] Please refer to the attached Figure 1, An intelligent prediction algorithm for foundation pit monitoring data driven by both data and physical mechanisms, comprising the following steps: Step 1: Denote the lateral displacement of the retaining wall in the past several time steps as , where represents the time dimension from to , and represents the depth dimension from 1 to , that is, the spatial dimension; Denote the future lateral displacement of the retaining wall to be predicted as.

[0033] Step 2: Based on the plane elastic foundation beam method, establish a neural network model to predict the future lateral displacement of the retaining wall, and introduce a physical intermediate variable layer to simulate the action of the active load behind the wall .

[0034] The input of the neural network model is the lateral displacement of the retaining wall in the past several time steps , and the output is the predicted future lateral displacement of the retaining wall , and calculate the simulated active load behind the wall through the LSTM network that maintains monotonicity, and continuously update the simulated active load behind the wall during the training process of the neural network model.

[0035] Since it is difficult to accurately obtain the true value of the active load behind the wall in actual foundation pit engineering, in the neural network model, based on the input lateral displacement of the retaining wall in the past several time steps , calculate the simulated active load behind the wall through network calculation, and continuously update it during the training process of the neural network model. It should be noted that this physical intermediate variable layer is used to simulate the action of the active load behind the wall in the calculation process of the lateral displacement of the retaining wall, and its output value does not actually represent the true value.

[0036] The plane elastic foundation beam method is a method commonly used in load structure analysis, especially suitable for buildings and civil engineering considering foundation effects. Its basic principle is to assume that there is full bonding between the structure and the foundation, and the interaction between the structure, the foundation and the load is described by an elastic beam model. This method is applicable to static load analysis and design, and can consider complex load conditions, such as concentrated loads, uniform loads, variable loads, etc. The plane elastic foundation beam method is a conventional analysis method in this field, and its principle and analysis process will not be elaborated here.

[0037] According to the calculation theory of the plane elastic foundation beam method, since the water and soil pressure both increase continuously with depth, the active load It has the property of non-decreasing monotonically along the depth direction. To incorporate this physical law into the neural network model, by improving the traditional LSTM network, an LSTM network that maintains monotonicity is adopted and used to calculate the output of the physical intermediate variable layer, so that the simulated active load behind the wall can satisfy the property of non-decreasing monotonically along the depth direction, so that the physical intermediate variable layer can more realistically simulate the active load behind the wall effect.

[0038] LSTM can better capture the initial long-term dependence relationship. Using the LSTM network that maintains monotonicity can better adapt to the property of the active load behind the wall of non-decreasing monotonically along the depth direction. The LSTM network is a commonly used mathematical model in this field, and its specific structure, working principle and calculation process will not be elaborated here.

[0039] Through the physical process simulated by the physical intermediate variable layer, the physical mechanism is incorporated into the traditional pure data-driven neural network model. Compared with the traditional pure data-driven intelligent prediction algorithm, the present invention has stronger interpretability. The physical intermediate variable layer can intuitively explain the relationship between the algorithm prediction result, that is, the future displacement of the retaining wall to be predicted and the actual physical phenomenon, making it easier for engineering and technical personnel to understand and accept the output of the algorithm, facilitating analysis, optimization and decision-making in engineering practice, and being conducive to the popularization and application of technology. At the same time, it also enables the algorithm to maintain relatively stable prediction performance in the face of different geological conditions, construction techniques and external interference factors, and will not produce large deviations due to small fluctuations or anomalies in the data, providing a reliable basis for the long-term stable monitoring of the project.

[0040] Step 3: Set a loss function based on the maximum value index of the retaining wall displacement at several past time steps of the model input and the simulated active load behind the wall sequence, which is used to constrain the simulated active load behind the wall when the true value is unknown.

[0041] In the loss function described above, the analytical solution of the ground settlement caused by the foundation pit excavation is incorporated into the loss function of the traditional neural network algorithm, and several trainable correction parameters are introduced into the loss function to dynamically correct the analytical solution during the training process.

[0042] The number and specific parameters of the correction parameters can be adaptively adjusted according to actual application requirements. Preferably, they are three, enabling the entire neural network model to still achieve good prediction effects even when the amount of data is relatively small, effectively reducing the dependence on the amount of training data. Since it is often difficult and costly to obtain a large amount of high-quality training data in actual foundation pit engineering, setting the loss function not only saves the time and cost of data collection and collation, but also improves the applicability of the algorithm in scenarios with limited data resources, speeds up the project progress, and reduces the upfront investment in the project.

[0043] According to existing foundation pit engineering experience, the positions of the maximum retaining wall lateral displacement and the maximum active load behind the wall generally are both near the depth of the excavation surface. By setting the loss function based on the index of the maximum value of the retaining wall lateral displacement and the simulated active load behind the wall in the model input sequence, this engineering experience is incorporated into the neural network model.

[0044] The loss function is a commonly used evaluation index in the neural network model and can be adaptively set according to actual usage requirements, which will not be elaborated here. Different geological conditions and engineering boundary conditions will affect the model accuracy, such as in soft soil and rock areas, circular and rectangular foundation pits, etc. The settings of the loss function, correction parameters, etc. can be adjusted accordingly according to the actual working conditions.

[0045] Step 4: After calculating the simulated active load behind the wall , use the CNN-LSTM network to extract the spatio-temporal information of the retaining wall lateral displacement and the simulated active load behind the wall at several past time steps respectively.

[0046] Considering the spatio-temporal characteristics of the retaining wall lateral displacement and the simulated active load behind the wall at several past time steps, use the CNN-LSTM network to extract their spatio-temporal information. The CNN-LSTM network has good capabilities for capturing and processing spatio-temporal information. The CNN-LSTM network is a conventional processing method in this field, and its principle and calculation process will not be elaborated here.

[0047] Step 5: Fuse the two parts of spatio-temporal information extracted in Step 4, and output the final future retaining wall lateral displacement to be predicted through a fully connected neural network, which is used to comprehensively reflect the influence of the retaining wall lateral displacement and the simulated active load behind the wall at several past time steps on the prediction of the future retaining wall lateral displacement.

[0048] After the present invention is applied in a certain foundation pit project, the experimental results of its monitoring data show that the intelligent prediction algorithm for foundation pit monitoring data driven by data-physical mechanism can accurately predict the lateral displacement of the retaining wall in the next 3 days when the lateral displacement of the retaining wall in the past 7 days is input, and the maximum relative norm error is 0.104. The value exceeds 0.9, and the lateral displacement curve of the retaining wall predicted by the algorithm is basically consistent with the measured lateral displacement curve of the retaining wall. Therefore, the intelligent prediction algorithm for foundation pit monitoring data driven by the data-physical mechanism of the present invention has good prediction accuracy and adaptability.

[0049] The above is only a preferred embodiment of the present invention and is not used to limit the protection scope of the invention. Therefore, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A data-physical mechanism dual-driven intelligent prediction algorithm for foundation pit monitoring data, characterized by: The following steps are involved: Step 1: Record the lateral displacement of the retaining wall in the past several time steps as ,in Representative from arrive The time dimension, Represents from 1 to The depth dimension, that is, the spatial dimension; the future lateral displacement of the retaining wall to be predicted is recorded as ; Step 2: Based on the plane elastic foundation beam method, a neural network model is established to predict the future lateral displacement of the retaining wall, and a physical intermediate variable layer is introduced to simulate the active load behind the wall. The role of; Step 3: Set a retaining wall displacement based on the model input for a number of past time steps Active load behind the wall Loss function of the maximum index of the sequence for active loads behind the wall Simulated active load behind the wall when the true value is unknown To impose restraints; Step 4: Calculate the simulated active load behind the wall Then, the CNN-LSTM network is used to extract the lateral displacement of the retaining wall at several time steps in the past. and simulated active load behind the wall spatiotemporal information; Step 5: Fuse the two parts of spatiotemporal information extracted in step 4, and output the final predicted future retaining wall lateral displacement through a fully connected neural network. , which is used to comprehensively reflect the lateral displacement of the retaining wall in the past several time steps and simulated active load behind the wall The role of predicting future retaining wall lateral displacement.

2. The data-physical mechanism dual-driven foundation pit monitoring data intelligent prediction algorithm according to claim 1 is characterized by: In step 2, the input of the neural network model is the lateral displacement of the retaining wall in the past several time steps. The output is the predicted future wall displacement The simulated active load behind the wall is calculated by the LSTM network that maintains monotonicity , the simulated active load behind the wall is continuously updated during the training of the neural network model .

3. The data-physical mechanism dual-driven foundation pit monitoring data intelligent prediction algorithm according to claim 1 is characterized by: In the step 3, the analytical solution of the surface settlement caused by foundation pit excavation is integrated into the loss function, and several correction parameters are introduced into the loss function, and the analytical solution of the surface settlement is dynamically corrected during the training process.

Images