Foundation bearing capacity dynamic prediction method and system based on machine learning

Abstract

The invention discloses a foundation bearing capacity dynamic prediction method and system based on machine learning, and relates to the technical field of foundation bearing capacity prediction.The foundation bearing capacity multi-modal associated data is collected and comprises geological radar data, underground water level historical sequence data and soil sample CT scanning image data; fusing the preprocessed geological radar data sequence, the preprocessed underground water level sequence and the preprocessed soil sample CT scanning feature vector to form a foundation bearing capacity associated multi-dimensional feature vector; and inputting the fused multi-dimensional feature vector, and inputting the trained dynamic foundation bearing capacity prediction model based on the LSTM. According to the method, the LSTM network is used, and the long-term dependency relationship in the time sequence data can be effectively captured, which is crucial for understanding the dynamic change of the foundation bearing capacity. And the output of the LSTM layer is connected to the full connection layer, so that the prediction result can be further refined, and the precision of foundation bearing capacity prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of foundation bearing capacity prediction, and in particular to a method and system for dynamic prediction of foundation bearing capacity based on machine learning. Background Art

[0002] The bearing capacity of the foundation refers to the ability of the foundation to withstand the load of the building without excessive settlement or damage. Accurately predicting the bearing capacity of the foundation is crucial to ensure the safety and economy of the building. Underestimating the bearing capacity may lead to foundation damage, while overestimating it will cause unnecessary economic waste. Therefore, accurately predicting the bearing capacity of the foundation is a key link in civil engineering.

[0003] The existing foundation bearing capacity prediction technologies mainly include empirical formula method, field test method and numerical simulation method; Among them, the empirical formula method: such as the empirical formula proposed by Terzaghi, Meyerhof, etc., calculates the bearing capacity based on soil parameters (such as internal friction angle, cohesion, etc.).

[0004] Field test method: such as standard penetration test (SPT), cone penetration test (CPT), etc., to obtain soil parameters through field testing.

[0005] Numerical simulation methods: such as finite element analysis (FEA), predict the bearing behavior of the foundation through numerical simulation.

[0006] The empirical formula method is too dependent on empirical parameters and is difficult to adapt to complex geological conditions. The field test method is costly, time-consuming, and limited by field conditions. The numerical simulation method is complex in calculation, sensitive to input parameters, and difficult to update in real time. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a method and system for dynamic prediction of foundation bearing capacity based on machine learning. The following technical solutions are adopted: The dynamic prediction method of foundation bearing capacity based on machine learning includes the following steps: Step 1: Collect multi-modal associated data of foundation bearing capacity, which includes geological radar data, groundwater level historical sequence data and soil sample CT scanning image data; Step 2, preprocessing the foundation bearing capacity multimodal correlation data to obtain geological radar data sequence, groundwater level sequence and soil sample CT scanning feature vector; Step 3, fusing the preprocessed geological radar data sequence, groundwater level sequence and soil sample CT scanning feature vector to form a multi-dimensional feature vector associated with foundation bearing capacity; Step 4, input the fused multi-dimensional feature vector and the trained LSTM-based dynamic prediction model of foundation bearing capacity; Step 5, the output of the LSTM layer of the foundation bearing capacity dynamic prediction model is connected to the fully connected layer for the final foundation bearing capacity prediction; Step 6: The output layer of the foundation bearing capacity dynamic prediction model outputs the bearing capacity change trend, and outputs the foundation bearing capacity value corresponding to the current foundation bearing capacity multimodal correlation data based on the bearing capacity change trend.

[0008] By adopting the above technical solutions, by collecting geological radar data, groundwater level historical series data and soil sample CT scan image data, and fusing them, preprocessing the multimodal related data, such as standardization, normalization and feature extraction, noise can be removed, key information can be highlighted, and high-quality data input can be provided for subsequent models, which helps to improve the performance of the model. The use of LSTM network can effectively capture long-term dependencies in time series data, which is crucial for understanding the dynamic changes of foundation bearing capacity. By connecting the output of the LSTM layer to the fully connected layer, the prediction results can be further refined and the accuracy of foundation bearing capacity prediction can be improved. It can more comprehensively capture various factors affecting foundation bearing capacity, thereby improving the accuracy and reliability of prediction.

[0009] It can dynamically predict the changing trend of foundation bearing capacity, which is of great significance for real-time monitoring and early warning in engineering design and construction.

[0010] It can quickly output the predicted value of foundation bearing capacity based on multi-modal correlation data collected in real time, which has great practical value for actual engineering applications.

[0011] By accurately predicting changes in foundation bearing capacity, engineering technicians can identify potential risks and take timely measures to reduce the risk of engineering accidents.

[0012] It has good adaptability and scalability, can be applied to the prediction of foundation bearing capacity under different geological conditions, and the prediction ability can be further improved by introducing more related data.

[0013] Optionally, the preprocessing method of the geological radar data in step 2 is: extracting the amplitude, frequency and phase characteristics of the radar wave, using signal processing technology to extract frequency domain features, converting the radar data into a time series form, ,in is the radar feature matrix, T is the time step, is the radar characteristic dimension, represents real numbers; The preprocessing method of the historical series data of groundwater levels is to interpolate the missing values, perform normalization, and directly use the water level values ​​as the characteristics of the groundwater level series: ; The preprocessing method of soil sample CT scan image data is: use the extracted image features, input CT image ,in is the image size, C is the number of channels, and a pre-trained convolutional neural network is used to extract high-level features and convert image features into vector form: ,in is the feature dimension extracted by the convolutional neural network. If features in the form of time series are required, the CT image features at multiple time points are concatenated into .

[0014] By adopting the above technical solutions, the amplitude, frequency and phase characteristics of radar waves can be extracted in the preprocessing of geological radar data to more accurately capture subtle changes in geological structures, which are crucial for predicting the bearing capacity of foundations. Using signal processing technology to extract frequency domain features helps to reveal the characteristics of radar signals in the frequency domain, thereby better understanding the composition and properties of geological structures. Converting radar data into time series form can maintain the temporal continuity of the data, which is very useful for capturing dynamic changes and long-term trends.

[0015] The preprocessed radar data is more suitable for input into prediction models based on time series analysis, such as LSTM, thereby improving the prediction performance of the model.

[0016] Interpolation of missing values ​​in the preprocessing of groundwater level historical series data can ensure the integrity of the data set and avoid prediction bias caused by missing data. Normalization processing makes data at different time points have the same scale, which helps the model learn and generalize better. Directly using water level values ​​as features simplifies the data processing process, while retaining key information and improving processing efficiency.

[0017] In the preprocessing of soil sample CT scan image data, a pretrained convolutional neural network (CNN) is used to extract high-level features, which can deeply explore the complex structures and patterns in the image. The image features are converted into vector form so that these features can be integrated with other types of data, enhancing the comprehensive analysis ability of the model. By splicing CT image features at multiple time points, the time dimension information is retained, which is helpful for analyzing the time change trend of foundation bearing capacity.

[0018] Optionally, in step 3, the geological radar data, groundwater level data and CT image features are spliced ​​in the feature dimension to form a unified multidimensional feature vector X represented as: ; The multidimensional feature vector X is the input item of the dynamic prediction model of foundation bearing capacity, and the output item of the dynamic prediction model of foundation bearing capacity is the foundation bearing capacity label Y.

[0019] Optionally, the multidimensional feature vector X is reduced in dimension to reduce feature redundancy. The dimension reduction process is expressed as: ; in is the multidimensional feature vector after dimensionality reduction, Represents the PCA dimensionality reduction function, X represents the multidimensional feature vector before dimensionality reduction, and k represents the feature dimension after dimensionality reduction.

[0020] By adopting the above technical solution, the integration of multi-source information is achieved by splicing geological radar data, groundwater level data and CT image features in the feature dimension, which can more comprehensively reflect the various factors affecting the bearing capacity of the foundation. The spliced ​​multi-dimensional feature vector X provides a unified input format for the model, so that the features of different data sources can be directly used for the training and prediction of the machine learning model. The features of different data sources may reflect different aspects of the bearing capacity of the foundation. The spliced ​​feature vector can make full use of this complementary information to improve the accuracy of the prediction.

[0021] Dimensionality reduction reduces the dimension of the feature vector, thereby reducing the computational complexity of the model and speeding up training and prediction. The dimensionality reduction process helps remove redundancy and noise in the data, allowing the model to focus more on the key features that have the greatest impact on the prediction results. By reducing the number of features, overfitting can be avoided and the generalization ability of the model can be improved, thereby achieving better prediction performance on new data.

[0022] The eigenvectors after dimensionality reduction may have better interpretability, allowing researchers to more easily understand which features have an important impact on the bearing capacity of the foundation.

[0023] The reduced eigenvectors are easier to visualize and analyze, helping researchers discover patterns and nonlinear relationships in the data.

[0024] The feature vector after dimensionality reduction occupies less storage space, which can significantly improve the storage and processing efficiency of data for large-scale data sets.

[0025] Optional, foundation bearing capacity label Y, , Y represents the foundation bearing capacity value at each time step T, and the foundation bearing capacity label Y is obtained through historical data.

[0026] Optionally, geological radar data of historical data, groundwater level historical sequence data and soil sample CT scanning image data are collected as input data, and historical data are collected to obtain corresponding foundation bearing capacity data, and the foundation bearing capacity data is used as output data. Multiple groups of input data and output data are used as training data to train a foundation bearing capacity dynamic prediction model based on LSTM. The training process is expressed as follows: ; in is the predicted value of the foundation bearing capacity at time step t, is the multidimensional feature vector at time step t.

[0027] By adopting the above technical solution, the dynamic prediction model of foundation bearing capacity is trained based on real historical data, which is beneficial to improving the prediction accuracy of the dynamic prediction model of foundation bearing capacity.

[0028] Optionally, the architecture of the foundation bearing capacity dynamic prediction model includes: an input layer, a feature fusion layer, a dimensionality reduction layer, an LSTM layer, a fully connected layer, and an output layer; The input items of the input layer include , and ; The feature fusion layer is connected to the input layer to , and Fusion into a multi-dimensional feature vector X; The dimension reduction layer is connected to the feature fusion layer to reduce the multidimensional feature vector X into ; The LSTM layer is connected to the dimension reduction layer, and multi-layer LSTM is used to capture the long-term dependency between the multi-dimensional feature vector based on time series and the bearing capacity of the foundation; The fully connected layer is connected to the LSTM layer, the final hidden state of the LSTM is input into the fully connected layer, and the carrying capacity change trend is output; The output layer is connected to the fully connected layer, and outputs the foundation bearing capacity value corresponding to the multimodal correlation data of the current foundation bearing capacity based on the bearing capacity change trend.

[0029] Optionally, the candidate cell state formula for the LSTM layer is: ; in is the weight matrix of candidate cell states, is the bias of the candidate cell state, is the hidden state of the previous time step at time t, represents the multidimensional feature vector after dimensionality reduction at time step t, is the hyperbolic tangent activation function; The formula for cell state update is: ; in is the cell state at the previous time step at time t, is the cell state after forgetting output by the forget gate, is element-wise multiplication, is the output of the input gate, is the candidate cell state; By adopting the above technical solutions, the feature fusion layer can effectively integrate the features from different data sources (geo-radar data, groundwater level data and CT image features) to form a comprehensive multi-dimensional feature vector, which helps the model to understand the changes in foundation bearing capacity more comprehensively. By fusing different types of data, the model can learn more complex feature representations, thereby improving the accuracy of predictions. The dimension reduction layer reduces the computational complexity of the subsequent LSTM layer by reducing the feature dimensions, making model training more efficient. Dimension reduction helps remove irrelevant or redundant features and reduce the impact of noise on model performance. The LSTM layer can capture long-term dependencies in time series data, which is crucial for understanding the dynamic changes in foundation bearing capacity. The LSTM layer can process time series data and model the changing trend of foundation bearing capacity over time. The fully connected layer can perform nonlinear mapping on the output of the LSTM layer and extract the features most relevant to the prediction of foundation bearing capacity. The bearing capacity change trend output by the fully connected layer provides a direct basis for the final prediction of the foundation bearing capacity value.

[0030] The output layer outputs specific foundation bearing capacity values ​​based on the change trend provided by the fully connected layer, providing accurate prediction results for engineering practice. The prediction results of the output layer can be directly used for engineering decision-making, such as determining construction plans and assessing engineering risks.

[0031] The foundation bearing capacity dynamic prediction system based on machine learning is used to realize the foundation bearing capacity dynamic prediction method based on machine learning. The foundation bearing capacity dynamic prediction system includes a geological radar data acquisition module, a groundwater level historical sequence data acquisition module, a soil sample CT scanning image data acquisition module and a computer. The geological radar data acquisition module communicates and exchanges historical geological radar data and current geological radar data with the data output end of the geological radar instrument. The groundwater level historical sequence data acquisition module communicates and exchanges historical groundwater level historical sequence data and current groundwater level historical sequence data with the data output end of the groundwater detector. The soil sample CT scanning image data acquisition module communicates and exchanges historical groundwater level historical sequence data and current groundwater level historical sequence data. The image data acquisition module communicates and exchanges historical soil sample CT scanning image data and current soil sample CT scanning image data with the data output end of the soil CT multifunctional analyzer. The computer is respectively communicated and connected with the foundation bearing capacity dynamic prediction system including the geological radar data acquisition module, the groundwater level historical sequence data acquisition module and the soil sample CT scanning image data acquisition module. The computer presets a foundation bearing capacity dynamic prediction model based on LSTM, inputs the current geological radar data, the current groundwater level historical sequence data and the current soil sample CT scanning image data into the foundation bearing capacity dynamic prediction model, and the foundation bearing capacity dynamic prediction model outputs the foundation bearing capacity value.

[0032] In summary, the present invention includes at least one of the following beneficial technical effects: The present invention can provide a dynamic prediction method and system for foundation bearing capacity based on machine learning. By collecting geological radar data, groundwater level historical sequence data and soil sample CT scanning image data, and fusing them, preprocessing the multimodal associated data, such as standardization, normalization and feature extraction, noise can be removed, key information can be highlighted, and high-quality data input can be provided for subsequent models, which helps to improve the performance of the model. Using the LSTM network, long-term dependencies in time series data can be effectively captured, which is crucial for understanding the dynamic changes of foundation bearing capacity. By connecting the output of the LSTM layer to the fully connected layer, the prediction results can be further refined and the accuracy of foundation bearing capacity prediction can be improved. It can more comprehensively capture various factors affecting foundation bearing capacity, thereby improving the accuracy and reliability of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the method for dynamic prediction of foundation bearing capacity based on machine learning of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0035] The embodiments of the present invention disclose a method and system for dynamically predicting foundation bearing capacity based on machine learning.

[0036] Reference Figure 1 Embodiment 1, a method for dynamic prediction of foundation bearing capacity based on machine learning, comprising the following steps: Step 1: Collect multi-modal associated data of foundation bearing capacity, which includes geological radar data, groundwater level historical sequence data and soil sample CT scanning image data; Step 2, preprocessing the foundation bearing capacity multimodal correlation data to obtain geological radar data sequence, groundwater level sequence and soil sample CT scanning feature vector; Step 3, fusing the preprocessed geological radar data sequence, groundwater level sequence and soil sample CT scanning feature vector to form a multi-dimensional feature vector associated with foundation bearing capacity; Step 4, input the fused multi-dimensional feature vector and the trained LSTM-based dynamic prediction model of foundation bearing capacity; Step 5, the output of the LSTM layer of the foundation bearing capacity dynamic prediction model is connected to the fully connected layer for the final foundation bearing capacity prediction; Step 6: The output layer of the foundation bearing capacity dynamic prediction model outputs the bearing capacity change trend, and outputs the foundation bearing capacity value corresponding to the current foundation bearing capacity multimodal correlation data based on the bearing capacity change trend.

[0037] By collecting geological radar data, groundwater level historical series data and soil sample CT scan image data, and fusing them, preprocessing the multimodal associated data, such as standardization, normalization and feature extraction, noise can be removed, key information can be highlighted, and high-quality data input can be provided for subsequent models, which helps to improve the performance of the model. The use of LSTM networks can effectively capture long-term dependencies in time series data, which is crucial for understanding the dynamic changes in foundation bearing capacity. By connecting the output of the LSTM layer to the fully connected layer, the prediction results can be further refined and the accuracy of foundation bearing capacity prediction can be improved. It can more comprehensively capture the various factors affecting foundation bearing capacity, thereby improving the accuracy and reliability of predictions.

[0038] It can dynamically predict the changing trend of foundation bearing capacity, which is of great significance for real-time monitoring and early warning in engineering design and construction.

[0039] It can quickly output the predicted value of foundation bearing capacity based on multi-modal correlation data collected in real time, which has great practical value for actual engineering applications.

[0040] By accurately predicting changes in foundation bearing capacity, engineering technicians can identify potential risks and take timely measures to reduce the risk of engineering accidents.

[0041] It has good adaptability and scalability, can be applied to the prediction of foundation bearing capacity under different geological conditions, and the prediction ability can be further improved by introducing more related data.

[0042] In Example 2, the preprocessing method of the geological radar data in step 2 is: extracting the amplitude, frequency and phase characteristics of the radar wave, using signal processing technology to extract frequency domain characteristics, converting the radar data into a time series form, ,in is the radar feature matrix, T is the time step, is the radar characteristic dimension, represents real numbers; The preprocessing method of the historical series data of groundwater levels is to interpolate the missing values, perform normalization, and directly use the water level values ​​as the characteristics of the groundwater level series: ; The preprocessing method of soil sample CT scan image data is: use the extracted image features, input CT image ,in is the image size, C is the number of channels, and a pre-trained convolutional neural network is used to extract high-level features and convert image features into vector form: ,in is the feature dimension extracted by the convolutional neural network. If features in the form of time series are required, the CT image features at multiple time points are concatenated into .

[0043] By extracting the amplitude, frequency and phase characteristics of radar waves in the preprocessing of geological radar data, subtle changes in geological structures can be captured more accurately, which are crucial for predicting the bearing capacity of foundations. Using signal processing technology to extract frequency domain features helps to reveal the characteristics of radar signals in the frequency domain, thereby better understanding the composition and properties of geological structures. Converting radar data into time series form can maintain the temporal continuity of the data, which is very useful for capturing dynamic changes and long-term trends.

[0044] The preprocessed radar data is more suitable for input into prediction models based on time series analysis, such as LSTM, thereby improving the prediction performance of the model.

[0045] Interpolation of missing values ​​in the preprocessing of groundwater level historical series data can ensure the integrity of the data set and avoid prediction bias caused by missing data. Normalization processing makes data at different time points have the same scale, which helps the model learn and generalize better. Directly using water level values ​​as features simplifies the data processing process, while retaining key information and improving processing efficiency.

[0046] In the preprocessing of soil sample CT scan image data, a pretrained convolutional neural network (CNN) is used to extract high-level features, which can deeply explore the complex structures and patterns in the image. The image features are converted into vector form so that these features can be integrated with other types of data, enhancing the comprehensive analysis ability of the model. By splicing CT image features at multiple time points, the time dimension information is retained, which is helpful for analyzing the time change trend of foundation bearing capacity.

[0047] Embodiment 3: In step 3, the geological radar data, groundwater level data and CT image features are spliced ​​in the feature dimension to form a unified multidimensional feature vector X represented as: ; The multidimensional feature vector X is the input item of the dynamic prediction model of foundation bearing capacity, and the output item of the dynamic prediction model of foundation bearing capacity is the foundation bearing capacity label Y.

[0048] In Example 4, the multidimensional feature vector X reduces feature redundancy by dimensionality reduction. The dimensionality reduction process is expressed as: ; in is the multidimensional feature vector after dimensionality reduction, Represents the PCA dimensionality reduction function, X represents the multidimensional feature vector before dimensionality reduction, and k represents the feature dimension after dimensionality reduction.

[0049] By splicing geological radar data, groundwater level data and CT image features in the feature dimension, the integration of multi-source information is achieved, which can more comprehensively reflect the various factors affecting the bearing capacity of the foundation. The spliced ​​multi-dimensional feature vector X provides a unified input format for the model, so that the features of different data sources can be directly used for the training and prediction of the machine learning model. The features of different data sources may reflect different aspects of the bearing capacity of the foundation. The spliced ​​feature vector can make full use of this complementary information to improve the accuracy of the prediction.

[0050] Dimensionality reduction reduces the dimension of the feature vector, thereby reducing the computational complexity of the model and speeding up training and prediction. The dimensionality reduction process helps remove redundancy and noise in the data, allowing the model to focus more on the key features that have the greatest impact on the prediction results. By reducing the number of features, overfitting can be avoided and the generalization ability of the model can be improved, thereby achieving better prediction performance on new data.

[0051] The eigenvectors after dimensionality reduction may have better interpretability, allowing researchers to more easily understand which features have an important impact on the bearing capacity of the foundation.

[0052] The reduced eigenvectors are easier to visualize and analyze, helping researchers discover patterns and nonlinear relationships in the data.

[0053] The feature vector after dimensionality reduction occupies less storage space, which can significantly improve the storage and processing efficiency of data for large-scale data sets.

[0054] Example 5, foundation bearing capacity label Y, , Y represents the foundation bearing capacity value at each time step T, and the foundation bearing capacity label Y is obtained through historical data.

[0055] Example 6, collecting geological radar data of historical data, groundwater level historical sequence data and soil sample CT scanning image data as input data, collecting historical data to obtain corresponding foundation bearing capacity data, using the foundation bearing capacity data as output data, and using multiple groups of input data and output data as training data to train a foundation bearing capacity dynamic prediction model based on LSTM, the training process is shown as follows: ; in is the predicted value of the foundation bearing capacity at time step t, is the multidimensional feature vector at time step t.

[0056] Training the dynamic prediction model of foundation bearing capacity based on real historical data is conducive to improving the prediction accuracy of the dynamic prediction model of foundation bearing capacity.

[0057] Embodiment 7, the architecture of the foundation bearing capacity dynamic prediction model includes: an input layer, a feature fusion layer, a dimensionality reduction layer, an LSTM layer, a fully connected layer and an output layer; The input items of the input layer include , and ; The feature fusion layer is connected to the input layer to , and Fusion into a multi-dimensional feature vector X; The dimension reduction layer is connected to the feature fusion layer to reduce the multidimensional feature vector X into ; The LSTM layer is connected to the dimension reduction layer, and multi-layer LSTM is used to capture the long-term dependency between the multi-dimensional feature vector based on time series and the bearing capacity of the foundation; The fully connected layer is connected to the LSTM layer, the final hidden state of the LSTM is input into the fully connected layer, and the carrying capacity change trend is output; The output layer is connected to the fully connected layer, and outputs the foundation bearing capacity value corresponding to the multimodal correlation data of the current foundation bearing capacity based on the bearing capacity change trend.

[0058] Example 8, the candidate cell state formula of the LSTM layer is: ; in is the weight matrix of the candidate cell states, is the bias of the candidate cell state, is the hidden state of the previous time step at time t, represents the multidimensional feature vector after dimensionality reduction at time step t, is the hyperbolic tangent activation function; The formula for cell state update is: ; in is the cell state at the previous time step at time t, is the cell state after forgetting output by the forget gate, is element-wise multiplication, is the output of the input gate, is the candidate cell state; The feature fusion layer can effectively integrate features from different data sources (geo-radar data, groundwater level data, and CT image features) to form a comprehensive multi-dimensional feature vector, which helps the model to understand the changes in foundation bearing capacity more comprehensively. By fusing different types of data, the model can learn more complex feature representations, thereby improving the accuracy of predictions. The dimension reduction layer reduces the computational complexity of the subsequent LSTM layer by reducing the feature dimensions, making model training more efficient. Dimension reduction helps remove irrelevant or redundant features and reduce the impact of noise on model performance. The LSTM layer can capture long-term dependencies in time series data, which is crucial for understanding the dynamic changes in foundation bearing capacity. The LSTM layer can process time series data and model the changing trend of foundation bearing capacity over time. The fully connected layer can perform nonlinear mapping on the output of the LSTM layer and extract the features most relevant to the prediction of foundation bearing capacity. The bearing capacity change trend output by the fully connected layer provides a direct basis for the final prediction of the foundation bearing capacity value.

[0059] The output layer outputs specific foundation bearing capacity values ​​based on the change trend provided by the fully connected layer, providing accurate prediction results for engineering practice. The prediction results of the output layer can be directly used for engineering decision-making, such as determining construction plans and assessing engineering risks.

[0060] Embodiment 9, a foundation bearing capacity dynamic prediction system based on machine learning, used to implement a foundation bearing capacity dynamic prediction method based on machine learning, the foundation bearing capacity dynamic prediction system includes a geological radar data acquisition module, a groundwater level historical sequence data acquisition module, a soil sample CT scanning image data acquisition module and a computer, the geological radar data acquisition module communicates and exchanges historical geological radar data and current geological radar data with the data output terminal of the geological radar instrument, the groundwater level historical sequence data acquisition module communicates and exchanges historical groundwater level historical sequence data and current groundwater level historical sequence data with the data output terminal of the groundwater detector, the soil sample CT scanning image data acquisition module communicates and exchanges historical groundwater level historical sequence data and current groundwater level historical sequence data, and the soil sample CT scanning image data acquisition module communicates and exchanges historical groundwater level historical sequence data and current groundwater level historical sequence data. The image data acquisition module communicates and exchanges historical soil sample CT scanning image data and current soil sample CT scanning image data with the data output terminal of the soil CT multifunctional analyzer. The computer is respectively communicated and connected with the foundation bearing capacity dynamic prediction system including the geological radar data acquisition module, the groundwater level historical sequence data acquisition module and the soil sample CT scanning image data acquisition module. The computer presets a foundation bearing capacity dynamic prediction model based on LSTM, inputs the current geological radar data, the current groundwater level historical sequence data and the current soil sample CT scanning image data into the foundation bearing capacity dynamic prediction model, and the foundation bearing capacity dynamic prediction model outputs the foundation bearing capacity value.

[0061] The following specific embodiments are used to illustrate the implementation principle of the present invention: The underground soil structure data were collected using a geological radar with a time step of T = 100, and 5 features (such as amplitude, frequency, phase, etc.) were collected in each time step.

[0062] The groundwater level data were collected using a groundwater detector with a time step of T = 100, and one feature (water level value) was collected in each time step.

[0063] The soil CT multifunctional analyzer was used to collect the internal structure images of the soil samples. The time step was T = 100, one CT image (size 64 × 64) was collected in each time step, and the number of channels was C = 1.

[0064] The amplitude, frequency and phase characteristics of radar waves are extracted from geological radar data.

[0065] Frequency domain features are extracted using Fourier transform.

[0066] Convert the radar data into a time series format: ; Linear interpolation is performed on the missing values ​​of the historical series data of groundwater levels, and normalization is performed. Water level values ​​are directly used as features:

[0067] The soil sample CT scan image data uses the pre-trained ResNet-18 to extract image features.

[0068] Convert image features into vector form: ; The geological radar data, groundwater level data and CT image features are spliced ​​into a unified multi-dimensional feature vector: ; Use PCA to reduce the dimensionality of the multidimensional feature vector to k=64: ; Input the multidimensional feature vector after dimensionality reduction .

[0069] Use two layers of LSTM, with hidden state dimension H=128, and input the final hidden state of LSTM into the fully connected layer. Output the trend of carrying capacity change; Historical geological radar data, groundwater level data and CT image features.

[0070] Output data: historical foundation bearing capacity value; The training process uses the Adam optimizer with a learning rate of η=0.001. The loss function is the mean square error (MSE). The number of training rounds is 100.

[0071] Collect current geological radar data, groundwater level data and CT image data in real time.

[0072] The real-time data is subjected to the same preprocessing and feature fusion as the historical data.

[0073] Input real-time data into the trained LSTM model.

[0074] Output the current foundation bearing capacity value; This implementation case shows how to use geological radar data, groundwater level historical sequence data and soil sample CT scan image data to dynamically predict foundation bearing capacity through the LSTM model. The system realizes the whole process from data collection, preprocessing, feature fusion, model training to real-time prediction, with high prediction accuracy and engineering application value.

[0075] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A dynamic prediction method of foundation bearing capacity based on machine learning, characterized by: The following steps are involved: Step 1: Collect multi-modal associated data of foundation bearing capacity, which includes geological radar data, groundwater level historical sequence data and soil sample CT scanning image data; Step 2, preprocessing the foundation bearing capacity multimodal correlation data to obtain geological radar data sequence, groundwater level sequence and soil sample CT scanning feature vector; Step 3, fusing the preprocessed geological radar data sequence, groundwater level sequence and soil sample CT scanning feature vector to form a multi-dimensional feature vector associated with foundation bearing capacity; Step 4, input the fused multi-dimensional feature vector and the trained LSTM-based dynamic prediction model of foundation bearing capacity; Step 5, the output of the LSTM layer of the foundation bearing capacity dynamic prediction model is connected to the fully connected layer for the final foundation bearing capacity prediction; Step 6: The output layer of the foundation bearing capacity dynamic prediction model outputs the bearing capacity change trend, and outputs the foundation bearing capacity value corresponding to the current foundation bearing capacity multimodal correlation data based on the bearing capacity change trend.

2. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 1, characterized in that: The preprocessing method of geological radar data in step 2 is: extract the amplitude, frequency and phase characteristics of radar waves, use signal processing technology to extract frequency domain features, convert radar data into time series form, ,in is the radar feature matrix, T is the time step, is the radar characteristic dimension, represents real numbers; The preprocessing method of the historical series data of groundwater levels is to interpolate the missing values, perform normalization, and directly use the water level values ​​as the characteristics of the groundwater level series: ; The preprocessing method of soil sample CT scan image data is: use the extracted image features, input CT image ,in is the image size, C is the number of channels, and a pre-trained convolutional neural network is used to extract high-level features and convert image features into vector form: ,in is the feature dimension extracted by the convolutional neural network. If features in the form of time series are required, the CT image features at multiple time points are concatenated into .

3. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 2, characterized in that: In step 3, the geological radar data, groundwater level data and CT image features are spliced ​​in the feature dimension to form a unified multidimensional feature vector X represented as: ； The multidimensional feature vector X is the input item of the dynamic prediction model of foundation bearing capacity, and the output item of the dynamic prediction model of foundation bearing capacity is the foundation bearing capacity label Y.

4. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 3 is characterized in that: The multidimensional feature vector X reduces feature redundancy by dimensionality reduction. The dimensionality reduction process is expressed as: ； in is the multidimensional feature vector after dimensionality reduction, Represents the PCA dimensionality reduction function, X represents the multidimensional feature vector before dimensionality reduction, and k represents the feature dimension after dimensionality reduction.

5. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 4 is characterized in that: Foundation bearing capacity label Y, , Y represents the foundation bearing capacity value at each time step T, and the foundation bearing capacity label Y is obtained through historical data.

6. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 5 is characterized in that: The geological radar data, groundwater level historical sequence data and soil sample CT scan image data of historical data are collected as input data, and the corresponding foundation bearing capacity data are obtained by collecting historical data. The foundation bearing capacity data is used as output data. Multiple sets of input data and output data are used as training data to train the foundation bearing capacity dynamic prediction model based on LSTM. The training process is shown as follows: ； in is the predicted value of the foundation bearing capacity at time step t, is the multidimensional feature vector at time step t.

7. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 6, characterized in that: The architecture of the foundation bearing capacity dynamic prediction model includes: input layer, feature fusion layer, dimension reduction layer, LSTM layer, fully connected layer and output layer; The input items of the input layer include , and ; The feature fusion layer is connected to the input layer to , and Fusion into a multi-dimensional feature vector X; The dimension reduction layer is connected to the feature fusion layer to reduce the multidimensional feature vector X into ; The LSTM layer is connected to the dimension reduction layer, and multi-layer LSTM is used to capture the long-term dependency between the multi-dimensional feature vector based on time series and the bearing capacity of the foundation; The fully connected layer is connected to the LSTM layer, the final hidden state of the LSTM is input into the fully connected layer, and the carrying capacity change trend is output; The output layer is connected to the fully connected layer, and outputs the foundation bearing capacity value corresponding to the multimodal correlation data of the current foundation bearing capacity based on the bearing capacity change trend.

8. The method for dynamic prediction of foundation bearing capacity based on machine learning according to claim 7, characterized in that: The candidate cell state formula for the LSTM layer is: ； in is the weight matrix of the candidate cell states, is the bias of the candidate cell state, is the hidden state of the previous time step at time t, represents the multidimensional feature vector after dimensionality reduction at time step t, is the hyperbolic tangent activation function; The formula for cell state update is: ； in is the cell state at the previous time step at time t, is the cell state after forgetting output by the forget gate, is element-wise multiplication, is the output of the input gate, is a candidate cell state.

9. The foundation bearing capacity dynamic prediction system based on machine learning is characterized by: Used to implement the dynamic prediction method of foundation bearing capacity based on machine learning as described in claim 8, the dynamic prediction system of foundation bearing capacity includes a geological radar data acquisition module, a groundwater level historical sequence data acquisition module, a soil sample CT scanning image data acquisition module and a computer, the geological radar data acquisition module communicates and exchanges historical geological radar data and current geological radar data with the data output end of the geological radar instrument, the groundwater level historical sequence data acquisition module communicates and exchanges historical groundwater level historical sequence data and current groundwater level historical sequence data with the data output end of the groundwater detector, the soil sample CT scanning image data acquisition module communicates and exchanges historical soil sample CT scanning image data and current soil sample CT scanning image data with the data output end of the soil CT multifunctional analyzer, the computer is respectively communicated and connected with the dynamic prediction system of foundation bearing capacity including the geological radar data acquisition module, the groundwater level historical sequence data acquisition module and the soil sample CT scanning image data acquisition module, the computer presets a dynamic prediction model of foundation bearing capacity based on LSTM, inputs the current geological radar data, the current groundwater level historical sequence data and the current soil sample CT scanning image data into the dynamic prediction model of foundation bearing capacity, and the dynamic prediction model of foundation bearing capacity outputs the foundation bearing capacity value.

Images