Deep foundation pit deformation prediction method and system based on neural network and rough set classification

Abstract

The invention relates to the technical field of foundation pit detection, and discloses a deep foundation pit deformation prediction method and system based on neural network and rough set classification, and the method comprises the following steps: data collection and preprocessing; carrying out attribute reduction based on a rough set theory; constructing and training a model based on an improved attention mechanism recurrent neural network (A-RNN); carrying out credibility verification and dynamic feedback adjustment; and outputting a result and an early warning response. By integrating multi-parameter monitoring data of displacement, soil pressure, underground water level, support stress and the like, driving factors of deformation of the deep foundation pit are comprehensively captured, misjudgment caused by data deviation of a single sensor is reduced, the combination of bidirectional LSTM and an attention mechanism effectively captures local features of long-term dependence and key time points in time sequence data, and the accuracy of deep foundation pit deformation detection is improved. And the accuracy of deformation prediction is obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of foundation pit detection technology, and in particular to a deep foundation pit deformation prediction method and system based on neural network and rough set classification. Background Art

[0002] With the acceleration of urbanization, deep foundation pit engineering is widely used in high-rise building and subway construction. However, deformation problems caused by complex geological conditions and dynamic environmental changes during foundation pit construction are frequent, potentially leading to structural instability or even collapse. Current deep foundation pit deformation prediction technologies still have the following limitations.

[0003] 1. Insufficient data dimension and reliability

[0004] Traditional monitoring methods rely heavily on single parameters like displacement and stress, ignoring the coupling effects of multiple factors, including soil pressure, groundwater level, and support stress. This can easily lead to omissions of key drivers. Data deviations from a single sensor can easily lead to misjudgments, impacting the comprehensiveness and accuracy of forecast results.

[0005] 2. Weak timing modeling capabilities

[0006] Existing prediction models, such as traditional RNNs, struggle to capture the long-term temporal dependencies of deep foundation pit deformation and are incapable of identifying local features at key time nodes. The high computational complexity associated with complex network structures makes it difficult to adapt to the real-time demands of on-site edge computing devices.

[0007] 3. Model generalization and dynamic adaptation defects

[0008] Static models are prone to false alarms due to overfitting or abnormal data under long-term complex working conditions, and lack the ability to adjust online to dynamic changes in the construction environment. Fixed risk thresholds are difficult to adapt to different geological conditions, resulting in insufficient warning sensitivity.

[0009] 4. Redundant features and resource consumption issues

[0010] High-dimensional monitoring data contains a large amount of redundant information. Traditional methods fail to effectively reduce the dimensionality, resulting in low efficiency of neural network training and excessive memory usage, which restricts the feasibility of embedded system deployment.

[0011] To address the above issues, there is an urgent need for a deep foundation pit deformation prediction method that integrates multi-source data perception, efficient time series modeling, and dynamic adaptive capabilities to improve prediction accuracy while meeting the real-time requirements of the engineering site. Summary of the Invention

[0012] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method and system for predicting deformation of deep foundation pits based on neural network and rough set classification.

[0013] The present invention provides the following technical solution: a method for predicting deformation of a deep foundation pit based on a neural network and rough set classification, comprising the following steps:

[0014] Step S1: Data acquisition and preprocessing

[0015] The displacement, soil pressure, groundwater level and support structure stress data of deep foundation pits are collected in real time by multi-source sensor arrays to construct a dynamic monitoring data set D raw ;

[0016] Step S2: Attribute reduction based on rough set theory

[0017] The preprocessed data is constructed into a decision table S = (U, A, V, f), where U is the domain, that is, the set of all samples; A = C ∪ D, C is the conditional attribute set, including monitoring data attributes such as displacement and earth pressure, and D is the decision attribute set, such as the deformation state of the deep foundation pit (normal, warning, dangerous, etc.); V is the set of attribute values; f: U × A → V is the information function used to determine the attribute value of each sample;

[0018] Calculate attribute importance: Use the attribute importance calculation method based on information entropy, let H(D) be the information entropy of decision attribute D, H(D|C-{a i}) is to remove attribute a i The conditional information entropy of the decision attribute D is then attribute a i Importance Sig(a i ) is: Sig(a i )=H(D|C-{a i})-H(D|C) where p(d j ) is the decision attribute D with the value d j probability; p(c i ) is the conditional attribute set C with the value c i The probability, p(d j |c i ) is the conditional attribute value c i When the decision attribute value is d j The conditional probability of

[0019] Step S3: Model construction and training based on improved attention mechanism recurrent neural network (A-RNN)

[0020] Improved A-RNN network structure: establish input layer, bidirectional LSTM layer, attention mechanism layer and output layer. The input layer receives the dataset D after attribute reduction. red, the bidirectional LSTM layer extracts features from the input data. and are the hidden states of the forward and backward LSTM units at time t, respectively. The output h of the bidirectional LSTM layer is t for: Attention mechanism layer: The attention mechanism is mainly used to weight the output of the bidirectional LSTM layer to highlight important features; training model: the simplified data D red Divide into training set, validation set and test set;

[0021] Step S4: Credibility Verification and Dynamic Feedback Adjustment

[0022] Credibility verification: The credibility of the neural network output is verified through the rough set decision rule base; let the prediction result of the neural network be y pred , according to the rough set decision rule base, it is judged whether it complies with the corresponding rules. If it complies, it is considered to have high credibility. If it does not comply, the deviation Δ between the prediction result and the rule is calculated;

[0023] Dynamic feedback adjustment: If the error exceeds the preset threshold τ, the dynamic feedback mechanism is triggered. First, the cause of the error is analyzed. If the attribute reduction is unreasonable, the attribute reduction parameters are readjusted and the minimum condition attribute set C is updated. min If the problem is with the neural network model parameters, then incremental learning is used to fine-tune the model using new data and update the network weights.

[0024] Step S5: Output results and warning response

[0025] Output deep foundation pit deformation prediction curve and risk level classification results. Risk levels are divided according to the predicted deformation amount and preset thresholds, such as low risk, medium risk, and high risk;

[0026] The on-site early warning devices are linked to implement a graded response. For low-risk conditions, regular monitoring and recording can be carried out; for medium-risk conditions, early warning signals are issued and the monitoring frequency is increased; for high-risk conditions, emergency measures are immediately initiated, such as evacuating personnel and reinforcing support structures.

[0027] Preferably, the data collected in step S1 is preprocessed, including noise removal, missing value filling and other operations. For noise removal, a sliding average filter algorithm is used. Suppose the original data sequence is x1, x2, ... x n ; The window size is w, then the filtered data y i for is the floor function, is an upward rounding function. For missing values, linear interpolation is used to fill them. If x kMissing, and x is known k-1 and x k+1 ,but

[0028] Preferably, the attention mechanism layer: let s t is the context vector at time t, α t , i is the attention weight, then: Among them, v and W h 、W s is the learnable weight matrix, b is the bias vector, and T is the time step.

[0029] Preferably, the training model uses the mean square error loss function Train the model, where N is the number of samples, y i is the true value, To predict the value, the model parameters are updated using optimization algorithms such as stochastic gradient descent (SGD).

[0030] A deep foundation pit deformation prediction system based on neural network and rough set classification consists of the following modules:

[0031] Distributed sensor array module: This module contains various types of sensors, such as displacement sensors, earth pressure sensors, groundwater level sensors, and stress sensors. These sensors are distributed at different locations in the deep foundation pit, collect relevant data in real time, and transmit the data to edge computing nodes.

[0032] Edge computing node: integrated data preprocessing unit to perform preprocessing operations such as removing noise and filling missing values ​​on the data collected by sensors;

[0033] Integrated rough set reduction processor: simplifies the attributes of the data according to the method in step S2;

[0034] Integrated lightweight neural network inference chip: runs the improved A-RNN model to perform deformation prediction on simplified data;

[0035] Cloud-based collaborative training platform: stores large amounts of historical monitoring data and model training results, providing data support for model training and optimization;

[0036] Equipped with a rule base version management module: manages and updates the rough set decision rule base, provides a model iteration optimization interface, and supports optimization of neural network models through incremental learning and other methods;

[0037] 3D visualization terminal: supports spatial mapping of prediction results and BIM models, intuitively displays the deformation prediction results of deep foundation pits on the BIM model, and provides multi-working condition comparative analysis functions. Users can select different time points and different working conditions for comparative analysis, and display the deformation of the stratum in the form of 3D graphics. At the same time, users can view the deep foundation pit deformation prediction results and related data at any time point in the past.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) By integrating multi-parameter monitoring data such as displacement, soil pressure, groundwater level, and support stress, the driving factors of deep foundation pit deformation are fully captured, and the misjudgment caused by the deviation of single sensor data is reduced. The combination of bidirectional LSTM and attention mechanism effectively captures the long-term dependence in time series data and the local characteristics of key time points, significantly improving the accuracy of deformation prediction. The prediction results of the neural network are verified by the rough set rule library to avoid model overfitting or false alarms under abnormal working conditions, thereby ensuring the reliability of early warning decisions.

[0040] (2) Eliminate redundant features, reduce the input dimension to the minimum necessary set, reduce the neural network training time and memory usage, and adapt to the real-time reasoning requirements of edge computing devices. The improved A-RNN structure simplifies the network layer while ensuring prediction accuracy through the collaborative design of bidirectional LSTM and attention mechanism, reduces computing energy consumption, and is suitable for on-site embedded deployment.

[0041] (3) Through the dynamic feedback adjustment module, the system can automatically update the model parameters according to the changes in the actual engineering environment without the need for complete retraining, adapting to the long-term complex working conditions of deep foundation pits, and dynamically adjusting the risk level threshold based on historical data and real-time monitoring results to avoid misjudgments caused by fixed thresholds and improve warning sensitivity.

[0042] (4) The prediction results are superimposed on the BIM model to generate deformation thermodynamic maps and support structure safety factor distributions, support multi-dimensional comparative analysis, and provide an intuitive basis for engineering decision-making. Through sound and light alarms, SMS push notifications, and equipment linkage, a full closed-loop control from risk identification to emergency response is achieved, reducing manual intervention delays and ensuring construction safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the logic diagram of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure more clear, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings of the embodiments of the present disclosure. In order to keep the following description of the embodiments of the present disclosure clear and concise, the present disclosure omits detailed descriptions of known functions and known components to avoid unnecessary confusion of the concepts of the present disclosure.

[0045] See also Figure 1 A deep foundation pit deformation prediction method based on neural network and rough set classification includes the following steps:

[0046] Step S1: Data acquisition and preprocessing

[0047] The displacement, soil pressure, groundwater level and support structure stress data of deep foundation pits are collected in real time by multi-source sensor arrays to construct a dynamic monitoring data set D raw ;

[0048] The collected data is preprocessed, including noise removal, missing value filling and other operations. For noise removal, high-frequency random interference is eliminated, and linear interpolation is used to repair data breakpoints to ensure data continuity and validity and prevent model training from being affected by outliers.

[0049] Using the sliding average filtering algorithm, let the original data sequence be x1, x2, ...x n ; The window size is w, then the filtered data y i for is the floor function, is an upward rounding function. For missing values, linear interpolation is used to fill them. If x k Missing, and x is known k-1 and x k+1 ,but

[0050] Step S2: Attribute reduction based on rough set theory

[0051] The preprocessed data is constructed into a decision table S = (U, A, V, f), where U is the domain, that is, the set of all samples; A = C ∪ D, C is the conditional attribute set, including monitoring data attributes such as displacement and earth pressure, and D is the decision attribute set, such as the deformation state of the deep foundation pit (normal, warning, dangerous, etc.); V is the set of attribute values; f: U × A → V is the information function used to determine the attribute value of each sample; the preprocessed multidimensional monitoring data is converted into a structured table form to clarify the logical relationship between the conditional attributes (monitoring parameters) and the decision attributes (deformation state).

[0052] Calculate attribute importance: Use the attribute importance calculation method based on information entropy, let H(D) be the information entropy of decision attribute D, H(D|C-{ai}) is to remove attribute a i The conditional information entropy of the decision attribute D is then attribute a i Importance Sig(a i ) is: Sig(a i )=H(D|C-{a i})-H(D|C) where p(d j ) is the decision attribute D with the value d j probability; p(c i ) is the conditional attribute set C with the value c i The probability, p(d j |c i ) is the conditional attribute value c i When the decision attribute value is d j The conditional probability of the deformation state is quantified based on the information entropy theory. The redundant attributes (such as earth pressure parameters with low correlation) are eliminated to generate the minimum conditional attribute set. The simplified decision table is output as the input feature of the neural network, which significantly improves the model training efficiency and generalization ability.

[0053] Step S3: Model construction and training based on improved attention mechanism recurrent neural network (A-RNN)

[0054] Improved A-RNN network structure: establish input layer, bidirectional LSTM layer, attention mechanism layer and output layer. The input layer receives the dataset D after attribute reduction. red , the bidirectional LSTM layer extracts features from the input data. and are the hidden states of the forward and backward LSTM units at time t, respectively. The output h of the bidirectional LSTM layer is t for: By transmitting temporal information in both positive and negative directions, the long-term dependence of the deformation process (such as the impact of continuous rainfall on foundation pit settlement) can be captured.

[0055] Attention mechanism layer: The attention mechanism is mainly used to weight the output of the bidirectional LSTM layer to highlight important features; let s t is the context vector at time t, α t , i is the attention weight, then: Among them, v and W h 、W sis a learnable weight matrix, b is a bias vector, and T is the time step. The time step weight is dynamically allocated to strengthen the focus on monitoring data during key periods (such as construction peaks) and solve the problem that traditional RNNs are not sensitive enough to local features.

[0056] Training model: The simplified data D red Divide into training set, validation set and test set; use mean square error loss function Train the model, where N is the number of samples, y i is the true value, To predict the value, the model parameters are updated using optimization algorithms such as stochastic gradient descent (SGD). A staged optimization (training set → validation set → test set) is adopted, combined with an adaptive learning rate algorithm (Adam) to ensure that the model converges to the global optimal solution.

[0057] Step S4: Credibility Verification and Dynamic Feedback Adjustment

[0058] Credibility verification: The credibility of the neural network output is verified through the rough set decision rule base; let the prediction result of the neural network be y pred , according to the rough set decision rule base, it is judged whether it complies with the corresponding rules. If it complies, it is considered to have high credibility. If it does not comply, the deviation Δ between the prediction result and the rule is calculated;

[0059] Dynamic feedback adjustment: If the error exceeds the preset threshold τ, the dynamic feedback mechanism is triggered. First, the cause of the error is analyzed. If the attribute reduction is unreasonable, the attribute reduction parameters are readjusted and the minimum condition attribute set C is updated. min If the problem is with the neural network model parameters, then incremental learning is used to fine-tune the model using new data and update the network weights.

[0060] Step S5: Output results and warning response

[0061] Output deep foundation pit deformation prediction curve and risk level classification results. Risk levels are divided according to the predicted deformation amount and preset thresholds, such as low risk, medium risk, and high risk;

[0062] The on-site early warning devices are linked to implement a graded response. For low-risk conditions, regular monitoring and recording can be carried out; for medium-risk conditions, early warning signals are issued and the monitoring frequency is increased; for high-risk conditions, emergency measures are immediately initiated, such as evacuating personnel and reinforcing support structures.

[0063] A deep foundation pit deformation prediction system based on neural network and rough set classification consists of the following modules:

[0064] Distributed sensor array module: contains various types of sensors, such as displacement sensors, soil pressure sensors, groundwater level sensors, stress sensors, etc., which are distributed at different locations of the deep foundation pit, collect relevant data of the deep foundation pit in real time, and transmit the data to the edge computing node.

[0065] Edge computing node: An integrated data preprocessing unit performs preprocessing operations such as removing noise and filling missing values ​​on the data collected by sensors.

[0066] Integrated rough set reduction processor: Reduce the attributes of the data according to the method in step S2.

[0067] Integrated lightweight neural network inference chip: runs the improved A-RNN model to perform deformation prediction on the simplified data.

[0068] Cloud-based collaborative training platform: stores large amounts of historical monitoring data and model training results, providing data support for model training and optimization.

[0069] Equipped with a rule base version management module: manages and updates the rough set decision rule base, provides a model iteration optimization interface, and supports optimization of neural network models through incremental learning and other methods.

[0070] 3D visualization terminal: supports spatial mapping between prediction results and BIM models, intuitively displays the deformation prediction results of deep foundation pits on the BIM model, and provides multi-working condition comparative analysis functions. Users can select different time points and different working conditions for comparative analysis to better understand the deformation of deep foundation pits.

[0071] The stratum deformation body drawing based on the improved MarchingCubes algorithm is realized, and the deformation of the stratum is displayed in the form of three-dimensional graphics.

[0072] The spatial overlay information of the risk heat map and the safety factor of the support structure is displayed, allowing users to intuitively understand the risk distribution of deep foundation pits and the safety status of the support structure.

[0073] With the historical backtracking function, users can view the deep foundation pit deformation prediction results and related data at any time point in the past.

[0074] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and scope of protection of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present invention.

Claims

1. A deep foundation pit deformation prediction method based on neural network and rough set classification, characterized in that: The following steps are involved: Step S1: Data acquisition and preprocessing The displacement, soil pressure, groundwater level and support structure stress data of deep foundation pits are collected in real time by multi-source sensor arrays to construct a dynamic monitoring data set D raw ; Step S2: Attribute reduction based on rough set theory The preprocessed data is constructed into a decision table S = (U, A, V, f), where U is the domain, that is, the set of all samples; A = C ∪ D, C is the conditional attribute set, including monitoring data attributes such as displacement and earth pressure, and D is the decision attribute set, such as the deformation state of the deep foundation pit (normal, warning, dangerous, etc.); V is the set of attribute values; f: U × A → V is the information function used to determine the attribute value of each sample; Calculate attribute importance: Use the attribute importance calculation method based on information entropy, let H(D) be the information entropy of decision attribute D, H(D|C-{a i }) is to remove attribute a i The conditional information entropy of the decision attribute D is then attribute a i Importance Sig(a i ) is: Sig(a i )=H(D|C-{a i })-H(D|C) where p(d j ) is the decision attribute D with the value d j probability; p(c i ) is the conditional attribute set C with the value c i The probability, p(d j |c i ) is the conditional attribute value c i When the decision attribute value is d j The conditional probability of Step S3: Model construction and training based on improved attention mechanism recurrent neural network (A-RNN) Improved A-RNN network structure: establish input layer, bidirectional LSTM layer, attention mechanism layer and output layer. The input layer receives the dataset D after attribute reduction. red , the bidirectional LSTM layer extracts features from the input data. and are the hidden states of the forward and backward LSTM units at time t, respectively. The output h of the bidirectional LSTM layer is t for: Attention mechanism layer: The attention mechanism is mainly used to weight the output of the bidirectional LSTM layer to highlight important features; Training model: The simplified data D red Divide into training set, validation set and test set; Step S4: Credibility Verification and Dynamic Feedback Adjustment Credibility verification: The credibility of the neural network output is verified through the rough set decision rule base; let the prediction result of the neural network be y pred , according to the rough set decision rule base, it is judged whether it complies with the corresponding rules. If it complies, it is considered to have high credibility. If it does not comply, the deviation Δ between the prediction result and the rule is calculated; Dynamic feedback adjustment: If the error exceeds the preset threshold τ, the dynamic feedback mechanism is triggered. First, the cause of the error is analyzed. If the attribute reduction is unreasonable, the attribute reduction parameters are readjusted and the minimum condition attribute set C is updated. min If the problem is with the neural network model parameters, then incremental learning is used to fine-tune the model using new data and update the network weights. Step S5: Output results and warning response Output deep foundation pit deformation prediction curve and risk level classification results. Risk levels are divided according to the predicted deformation amount and preset thresholds, such as low risk, medium risk, and high risk; Linking on-site early warning devices to implement graded responses, and regular monitoring and recording of low-risk conditions; For medium-risk conditions, early warning signals are issued and monitoring frequency is increased; for high-risk conditions, emergency measures are immediately initiated, such as evacuating personnel and reinforcing support structures.

2. The method for predicting deep foundation pit deformation based on neural network and rough set classification according to claim 1, characterized in that: The data collected in step S1 are preprocessed, including noise removal, missing value filling and other operations. For noise removal, a sliding average filter algorithm is used. Suppose the original data sequence is x1, x2, ... x n ; The window size is w, then the filtered data y i for is the floor function, is an upward rounding function. For missing values, linear interpolation is used to fill them. If x k Missing, and x is known k-1 and x k+1 ,but 3. The method for predicting deep foundation pit deformation based on neural network and rough set classification according to claim 1, characterized in that: Attention mechanism layer: Let s t is the context vector at time t, α t , i is the attention weight, then: Among them, v and W h 、W s is the learnable weight matrix, b is the bias vector, and T is the time step.

4. The method for predicting deep foundation pit deformation based on neural network and rough set classification according to claim 1, characterized in that: Training model: using mean square error loss function Train the model, where N is the number of samples, y i is the true value, To predict the value, the model parameters are updated using optimization algorithms such as stochastic gradient descent (SGD).

5. A deep foundation pit deformation prediction system based on neural network and rough set classification, characterized by: A deep foundation pit deformation prediction method based on neural network and rough set classification according to any one of claims 1 to 4 is adopted, which is composed of the following modules: Distributed sensor array module: This module contains various types of sensors, such as displacement sensors, earth pressure sensors, groundwater level sensors, and stress sensors. These sensors are distributed at different locations in the deep foundation pit, collect relevant data in real time, and transmit the data to edge computing nodes. Edge computing node: integrated data preprocessing unit to perform preprocessing operations such as removing noise and filling missing values ​​on the data collected by sensors; Integrated rough set reduction processor: simplifies the attributes of the data according to the method in step S2; Integrated lightweight neural network inference chip: runs the improved A-RNN model to perform deformation prediction on simplified data; Cloud-based collaborative training platform: stores large amounts of historical monitoring data and model training results, providing data support for model training and optimization; Equipped with a rule base version management module: manages and updates the rough set decision rule base, provides a model iteration optimization interface, and supports optimization of neural network models through incremental learning and other methods; 3D visualization terminal: supports spatial mapping of prediction results and BIM models, intuitively displays the deformation prediction results of deep foundation pits on the BIM model, and provides multi-working condition comparative analysis functions. Users can select different time points and different working conditions for comparative analysis, and display the deformation of the stratum in the form of 3D graphics. At the same time, users can view the deep foundation pit deformation prediction results and related data at any time point in the past.

Images