A dynamic determination method for stability of a supporting structure for deep foundation pit construction

Abstract

This invention discloses a dynamic method for determining the stability of support structures in deep foundation pit construction, relating to the field of engineering structure monitoring and diagnosis technology. The method includes: collecting multi-source heterogeneous data of the deep foundation pit support structure for preprocessing to generate a complete reconstruction sequence; constructing stability labels and combining them with the complete reconstruction sequence to build training samples, constructing a master decision model, using the training samples as input to the master decision model, outputting a predicted probability distribution, and selecting a final decision label; obtaining auxiliary labels based on the final decision label, and performing label consistency determination to obtain consistent-inconsistent results; then, based on the inconsistent results, extracting predicted probabilities for penalty correction, and performing label credibility-untrustworthiness determination to obtain credible-untrustworthiness results. This invention improves the accuracy, robustness, and engineering feasibility of determining the stability of support structures.

Description

Technical Field

[0001] This invention relates to the field of engineering structure monitoring and diagnosis technology, and in particular to a method for dynamically determining the stability of support structures used in deep foundation pit construction. Background Technology

[0002] With the high-intensity development and utilization of urban underground space, deep foundation pit engineering has been widely used in rail transit, super high-rise buildings, municipal infrastructure, and other fields. As construction depths increase and surrounding environmental conditions become increasingly complex, the safety and stability of the support structure have become key factors restricting project safety and construction efficiency. Diaphragm walls, as a commonly used retaining and support structure in deep foundation pit engineering, possess advantages such as good integrity, high rigidity, and excellent water-stopping performance, and are widely used in deep foundation pit construction in soft soil areas and complex geological conditions. To ensure construction safety, various monitoring sensors are typically installed on the diaphragm wall in engineering practice to monitor wall deformation, stress state, and groundwater changes in real time, and to determine and provide early warnings based on the monitoring data. Existing technologies mainly include empirical judgment methods based on standard thresholds, analytical or numerical calculation methods based on mechanical models, and data-driven intelligent judgment methods that have gradually emerged in recent years. Among them, the empirical threshold method relies on standard limits and engineering experience to make judgments, the mechanical model method analyzes the structural response through finite element or limit equilibrium theory, while the data-driven method uses machine learning or deep learning models to extract features and identify states from monitoring data, which improves the automation and real-time performance of stability determination to a certain extent.

[0003] Existing technologies lack a mechanism for evaluating and interpreting the reliability of judgment results. When conflicts arise between different judgment models or judgment results, it is difficult to effectively correct the prediction results. Furthermore, there is a lack of an adaptive adjustment mechanism for parameters, which limits their engineering applicability and safety reliability in high-risk deep foundation pit engineering. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a dynamic determination method for the stability of support structures used in deep foundation pit construction, which solves the problems of existing technologies that are difficult to effectively correct prediction results and lack adaptive adjustment mechanisms for parameters, thus limiting their engineering applicability and safety reliability in high-risk deep foundation pit projects.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for dynamically determining the stability of a support structure used in deep foundation pit construction, comprising,

[0008] Multi-source heterogeneous data of deep foundation pit support structure are collected, preprocessed, and a complete reconstruction sequence is generated.

[0009] A stability label is constructed and combined with the complete reconstructed sequence to build training samples. The main decision model is then constructed, and the training samples are used as the input of the main decision model to output the predicted probability distribution and select the final decision label.

[0010] Auxiliary labels are obtained based on the final judgment label, and label consistency judgment is performed to obtain consistent-inconsistent results. Then, based on the inconsistent results, the predicted probability is extracted and penalized and corrected. Finally, label credibility-untrustworthiness judgment is performed to obtain credibility-untrustworthiness results.

[0011] Based on unreliable results, a probability sequence is constructed and the posterior state entropy and causal consistency coefficient are calculated. Then, a posterior feedback adjustment factor is generated to update, iterate, and provide feedback to the parameters required for penalty correction, and the specified label is obtained and stored.

[0012] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, the method comprises: constructing training samples by combining the constructed stability labels with the complete reconstruction sequence, constructing a master determination model, using the training samples as input to the master determination model, outputting a predicted probability distribution, and selecting the final determination label, including:

[0013] Stability labeling of diaphragm wall structures in deep foundation pits using the Delphi method, including stable states. Critical state and instability The complete reconstructed sequence and stability label are constructed as input pairs and used as training sample pairs;

[0014] A master decision model is constructed using a one-dimensional convolutional neural network and a bidirectional long short-term memory neural network structure, including an input layer, a convolutional extraction layer, a max pooling layer, a bidirectional LSTM layer, a fully connected layer, and a softmax output layer. The cross-entropy loss function is used as the loss function of the master decision model to minimize the gap between the predicted distribution and the true label.

[0015] The Adam optimizer is used to update and optimize the parameters of the master decision model. During the iteration process, when the loss value of the loss function no longer decreases significantly, the iteration stops and the trained master decision model is output.

[0016] The training sample pairs are used as input to the main decision model. The input layer receives the training samples and passes them to the convolutional extraction layer. The convolutional extraction layer is then used to obtain the feature values ​​of the complete reconstructed sequence in the training samples.

[0017] The temporal features are downsampled using a max pooling layer to generate sampled features;

[0018] The forward and backward output features of the bidirectional LSTM layers are obtained and concatenated to obtain the feature representation of each time step. ;

[0019] Average pooling is performed on all feature representations to generate the overall representation vector;

[0020] A fully connected layer is used to perform a linear transformation on the overall representation vector to obtain a classification scoring vector, which contains the original score value of each stable label;

[0021] Based on the classification scoring vector, the Softmax output layer is used to map the probabilities to obtain the predicted probability distribution, and the stability label corresponding to the highest probability is selected as the final decision label.

[0022] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, wherein: the step of obtaining auxiliary labels based on the final determination label includes:

[0023] Backtracking is performed based on the final judgment label, the corresponding complete reconstruction sequence is extracted and discrete operation is performed to obtain equally segmented discrete intervals, the frequency of occurrence of each discrete interval under different stability labels is counted, and an attribute-state frequency mapping table is constructed.

[0024] For each stability label, calculate the cumulative statistical support of the complete reconstructed sequence for each stability label;

[0025] The cumulative statistical support scores are sorted in descending order, and the stability label corresponding to the highest cumulative statistical support score is selected as the auxiliary label.

[0026] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, the following steps are included: performing label consistency determination to obtain consistent-inconsistent results, then extracting predicted probabilities based on the inconsistent results for penalty correction, and finally performing label credibility-untrustworthiness determination to obtain credibility-untrustworthiness results, including:

[0027] Based on the auxiliary labels and the final judgment labels, determine whether the labels are consistent;

[0028] If the auxiliary label matches the final judgment label, the final judgment label will be used as the designated label; otherwise, the weight of the penalty correction operation will be set, and the penalty correction operation will be performed on the predicted probability.

[0029] A confidence threshold is set. If the predicted probability after penalty correction is greater than or equal to the confidence threshold, it means that the current stability label is reliable. The final judgment label corresponding to the predicted probability after maximum penalty correction is taken as the designated label. Otherwise, it means that the current stability label is unreliable.

[0030] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, wherein: after constructing a probability sequence and calculating the posterior state entropy and causal consistency coefficient based on unreliable results, a posterior feedback adjustment factor is generated to update, iterate, and provide feedback on the parameters required for penalty correction, thereby obtaining a specified label, including:

[0031] Based on unreliable results, the penalized and corrected predicted probabilities are extracted, sorted according to time step order, and a probability sequence is constructed. The posterior state entropy of each penalized and corrected predicted probability in the probability sequence is calculated. Then, based on the posterior state entropy, the posterior entropy change rate is calculated, and the posterior entropy time factor is obtained based on the posterior entropy change rate.

[0032] Based on the probability sequence and the posterior state entropy, the changes in the predicted probability and the posterior state entropy are calculated respectively, and a change vector is generated, including the probability change vector and the posterior entropy change vector.

[0033] Based on the change vector, the posterior causal consistency coefficient is calculated. Combined with the posterior entropy time factor, the posterior feedback regulation formula is constructed, and the posterior feedback regulation factor is calculated.

[0034] Based on the posterior feedback adjustment factor, the weights of the penalty correction operation are updated, and the updated penalty weights are fed back into the penalty correction operation formula. The penalty correction prediction probability is then recalculated to perform a reliable judgment and select a specified label.

[0035] If the label remains unreliable during continuous iterations, a warning message will be sent to regulators, indicating that the currently designated label is unreliable and suggesting a manual re-inspection.

[0036] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, the storage includes:

[0037] Edit the currently specified label and predicted probability into a folder, add a timestamp and corresponding stability label to the folder, store the folder in the database, and sort it according to the timestamp order after storage.

[0038] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, the preprocessing of multi-source heterogeneous data of the deep foundation pit support structure includes:

[0039] Sensors are deployed on the underground continuous wall of the deep foundation pit to collect multi-source heterogeneous data for preprocessing and generate preprocessed data.

[0040] As a preferred embodiment of the dynamic stability determination method for support structures used in deep foundation pit construction according to the present invention, wherein: generating a complete reconstruction sequence includes:

[0041] Based on the preprocessed data, a sliding window technique is used to divide the data into window data, which is then constructed into an input matrix. Uniform random occlusion is then used to occlude each data point in the input matrix to generate a mask matrix. A pre-trained encoding network, including a mask autoencoder and a decoder, is applied. The mask autoencoder performs contextual modeling on the mask matrix to generate a latent space representation. Finally, the pre-trained decoder is used to reconstruct the latent space representation, outputting a complete reconstructed sequence.

[0042] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the dynamic determination method for the stability of support structures for deep foundation pit construction as described in the first aspect of the present invention.

[0043] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the dynamic determination method for the stability of support structures for deep foundation pit construction as described in the first aspect of the present invention.

[0044] The beneficial effects of this invention are as follows: This invention constructs a deep learning master judgment model based on multi-source monitoring data, introduces statistical support to form auxiliary judgment labels, and constructs a feedback adjustment mechanism based on posterior state entropy and causal consistency when there is a judgment conflict or insufficient confidence, thereby realizing dynamic correction and credibility assessment of the judgment results, thereby improving the accuracy, robustness and engineering feasibility of the stability judgment of the support structure. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the dynamic determination method for the stability of the support structure used in deep foundation pit construction in Example 1.

[0047] Figure 2 This is a flowchart of the label consistency determination and credibility determination in Example 1.

[0048] Figure 3 This is a flowchart of updating the penalty weight in Example 1. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides a method for dynamically determining the stability of a support structure for deep foundation pit construction, including the following steps:

[0053] S1. Collect multi-source heterogeneous data of deep foundation pit support structure, preprocess it, and generate a complete reconstruction sequence;

[0054] Specifically, multi-source heterogeneous data on deep foundation pit support structures are collected and preprocessed, including:

[0055] Sensors are deployed on the underground continuous wall of the deep foundation pit to collect multi-source heterogeneous data for preprocessing and generate preprocessed data;

[0056] The sensors include soil pressure gauges, pore pressure gauges, stress gauges, inclinometers, and displacement gauges. The multi-source heterogeneous data includes normal stress on the wall, pore water pressure in groundwater, tensile stress in steel bars, overall tilt angle of the underground continuous wall, and horizontal displacement data of the top of the wall or the wall surface.

[0057] Preprocessing includes denoising and normalization.

[0058] By deploying multi-source heterogeneous sensors on the underground diaphragm wall and preprocessing the collected data, the system comprehensively perceives the critical stability status of the support structure under complex working conditions, achieving high-dimensional coverage and real-time updates of status information. The preprocessing step effectively unifies the scale and quality of various sensor data, providing standardized input for subsequent model building, while enhancing sensitivity to abnormal changes.

[0059] Furthermore, a complete reconstruction sequence is generated, including:

[0060] Based on the preprocessed data, a sliding window technique is used to divide the data into window data, which is then constructed into an input matrix. Uniform random masking is then used to mask each data point in the input matrix to generate a mask matrix. A pre-trained encoding network, including a masked autoencoder (MAE) and a decoder, is applied. The mask matrix is ​​then used to perform contextual modeling on the mask matrix to generate a latent space representation. Finally, the pre-trained decoder is used to reconstruct the latent space representation, outputting a complete reconstructed sequence.

[0061] It should be noted that uniform random masking is used to mask each data point in the input matrix, specifically as follows:

[0062] For each position of the data in the input matrix, a random number is generated by a random number generator. The random number is used as the sampling mask probability value. When the sampling mask probability value is less than the mask ratio (the mask ratio can be set from 0.1 to 0.3, but according to the stability requirements and generalization ability improvement requirements of the mask autoencoder during training, this invention can take 0.2 as the initial value of the mask ratio), the data is set to 0 (masked); otherwise, the data is set to 1 (retained).

[0063] The pre-trained mask autoencoder structure consists of a Transformer encoder module, which includes a multi-head self-attention mechanism, residual connections, LayerNorm, and a feedforward network (FFN). The pre-trained decoder has a symmetrical structure to the pre-trained mask autoencoder. The mean square error of the mask position is used as the loss function for minimization. The Adam optimizer is used for parameter update optimization. When the loss value of the loss function no longer decreases significantly, the iteration stops, and the trained mask autoencoder is output.

[0064] By uniformly segmenting preprocessed multi-source heterogeneous data using a sliding window, the problem of inconsistent time scales among data from different sensors is effectively solved, improving the standardization of the data structure. Introducing a uniform random occlusion strategy, which actively introduces missing information during the training phase, helps the masked autoencoder learn global contextual features, improving its adaptability to local disturbances and sensor faults. Supported by a Transformer-structured masked autoencoder, the model achieves high-precision context modeling and sequence reconstruction, resulting in reconstructed sequences with good semantic consistency and structural integrity. This method not only improves the input data quality of subsequent stability assessment models but also enhances the system's ability to handle uncertain data, thus providing a solid data foundation for determining the dynamic stability of support structures.

[0065] S2. Construct stable labels and complete reconstructed sequences to build training samples, construct the main decision model, use the training samples as input to the main decision model, output the predicted probability distribution, and select the final decision label.

[0066] Specifically, a stability label is constructed and combined with the complete reconstructed sequence to build training samples, construct the master decision model, use the training samples as input to the master decision model, output the predicted probability distribution, and select the final decision label, including:

[0067] Stability labeling of diaphragm wall structures in deep foundation pits using the Delphi method, including stable states. Critical state and instability The complete reconstructed sequence and stability label are constructed as input pairs and used as training sample pairs;

[0068] steady state The support structure was not deformed, the pore pressure was stable, and no displacement occurred.

[0069] Critical state : The support structure experiences continuous increase in displacement and fluctuation in pore pressure (e.g., displacement increases by more than 0.5 mm / h three times in a row, and pore pressure fluctuation exceeds ±10%).

[0070] Instability The support structure has been damaged and significantly deformed.

[0071] A master decision model is constructed using a 1D Convolutional Neural Network (1DCNN) and a Bidirectional Long Short-Term Memory (BiLSTM) neural network. This model includes an input layer, convolutional extraction layer, max pooling layer, bidirectional LSTM layer, fully connected layer, and softmax output layer. The cross-entropy loss function is used as the loss function of the master decision model to minimize the difference between the predicted distribution and the true label. The formula is as follows:

[0072]

[0073] In the formula, This represents the loss function value of the primary decision model. Indicates the total number of stability tags. Indicates the indicator function (if the actual label) equal to the If a stability label is provided, the function value is 1; otherwise, it is 0. This represents a real label (which can be generated through manual annotation). Indicates the index of the stability label. Represents the natural logarithm function. Indicates that, given a complete reconstruction sequence Under the condition that it belongs to the first The predicted probability of a stable label. Indicates the complete reconstruction sequence;

[0074] The Adam optimizer is used to update and optimize the parameters of the master decision model. During the iteration process, when the loss value of the loss function no longer decreases significantly, the iteration stops and the trained master decision model is output.

[0075] The training sample pairs are used as input to the main decision model. The input layer receives the training samples and passes them to the convolutional extraction layer. The convolutional extraction layer is then used to obtain the feature values ​​of the fully reconstructed sequence from the training samples. The formula is as follows:

[0076]

[0077] In the formula, Indicates the first The time step The feature values ​​output by each convolutional channel express Activation function This indicates the size of the convolution kernel's time window. Indicates the total number of sensors. Indicates the first The first time step Sensor values ​​for each input feature channel. Indicates from the first The input feature channel to the th At the nth convolutional channel The weights at the convolution kernel positions at each time step Indicates the first Bias terms for each convolutional channel;

[0078] The temporal features are downsampled using a max pooling layer to generate sampled features, as shown in the formula:

[0079]

[0080] In the formula, Represents the sampled feature value. This indicates a max pooling operation. and This represents two adjacent feature values ​​in the time dimension;

[0081] The forward and backward output features of the bidirectional LSTM layers are obtained and concatenated to obtain the feature representation of each time step. The formula is:

[0082]

[0083] In the formula, Indicates time step Time-related characteristics Indicates at time step The sampling feature value at that time, Indicates the feedforward LSTM layer. Indicates a reverse LSTM layer;

[0084] Average pooling is performed on all feature representations to generate the overall representation vector, as shown in the formula:

[0085]

[0086] In the formula, This represents the overall vector. Indicates the total number of time steps;

[0087] A fully connected layer is used to perform a linear transformation on the overall representation vector to obtain a classification scoring vector, which contains the original score value of each stable label. The formula is as follows:

[0088]

[0089] In the formula, Represents the classification scoring vector. This represents the weight matrix of the fully connected layer. This represents the bias term of the fully connected layer;

[0090] Based on the classification scoring vector, a Softmax output layer is used to map the probabilities, obtaining the predicted probability distribution. The stability label corresponding to the highest probability is then selected as the final decision label, as shown in the formula:

[0091]

[0092] In the formula, The base of the natural logarithm. and They respectively represent the corresponding number and the The classification score vector value of each stability label.

[0093] By introducing the Delphi method to construct a stability labeling system for stable, critical, and unstable states, this invention integrates dispersed engineering experience, regulatory requirements, and monitoring data characteristics. This eliminates the reliance on single thresholds or individual experience in defining stability states, significantly improving the consistency and reproducibility of stability determination labels under different engineering conditions and providing a reliable supervised foundation for subsequent model training. Secondly, the main decision model constructed in this invention employs a combination of a one-dimensional convolutional neural network and a bidirectional long short-term memory network. The one-dimensional convolutional neural network extracts local response patterns from multi-source monitoring data over short timescales, while the bidirectional long short-term memory network characterizes the bidirectional dependency of support structure stability over time. This allows the model to simultaneously capture local anomaly features and overall evolutionary trends, avoiding misjudgments caused by relying solely on short-term fluctuations or long-term trends. Furthermore, the bidirectional temporal modeling mechanism enables the model to comprehensively consider historical response processes and evolutionary constraints within the overall time window when determining stability states, effectively improving the accuracy and robustness of critical state identification. Furthermore, this invention provides a quantitative basis for subsequent reliability assessment, probability correction, and feedback adjustment by outputting the determination results in the form of a probability distribution using Softmax. This ensures that the stability determination results not only have clear state labels but also confidence information that can be used for dynamic adjustment.

[0094] S3. Obtain auxiliary labels based on the final judgment label, and perform label consistency judgment to obtain consistent-inconsistent results. Then, based on the inconsistent results, extract the predicted probability, perform penalty correction, and perform label trustworthiness-untrustworthiness judgment to obtain trustworthiness-untrustworthiness results.

[0095] Specifically, auxiliary tags are obtained based on the final determined tag, including:

[0096] Backtracking is performed based on the final judgment label, the corresponding complete reconstruction sequence is extracted and discrete operation is performed to obtain equally segmented discrete intervals, the frequency of occurrence of each discrete interval under different stability labels is counted, and an attribute-state frequency mapping table is constructed.

[0097] For each stability label, calculate the cumulative statistical support of the complete reconstructed sequence for each stability label, using the following formula:

[0098]

[0099] In the formula, Indicates the first The cumulative statistical support of each stability label, Indicates the complete reconstruction sequence at the 1st... The frequency of occurrence of each stability label (obtained by querying the attribute-state frequency mapping table). Indicates the first The discrete range of multi-source heterogeneous data values ​​from a single sensor. Indicates the first Within the first time step Multi-source heterogeneous data values ​​from individual sensors (pre-processed values);

[0100] The cumulative statistical support scores are sorted in descending order, and the stability label corresponding to the highest cumulative statistical support score is selected as the auxiliary label.

[0101] By backtracking the final judgment label and extracting the corresponding complete reconstruction sequence, this invention achieves a traceable association between the judgment result and the original monitoring information, effectively avoiding the "black box" problem common in deep learning judgment processes, and ensuring that subsequent auxiliary judgments are no longer divorced from specific physical quantities and sensor data. This process provides interpretable data evidence for the stability judgment results, significantly enhancing the credibility of the judgment process in engineering applications. Secondly, the complete reconstruction sequence is processed using equidistant segmentation, transforming continuous monitoring data into a structured set of discrete attributes, thereby reducing the impact of differences in the dimensions of different sensors and inconsistencies in numerical scales on statistical analysis. This processing method improves the stability and robustness of statistical frequency calculation while maintaining the overall distribution characteristics of the data, providing a unified expression space for multi-source data fusion. Thirdly, by constructing an attribute-state frequency mapping table and calculating the cumulative statistical support corresponding to each stability label, this invention introduces a statistical judgment mechanism independent of the main judgment model. This mechanism does not rely on the parameters or structure of the deep model, but rather makes judgments based on the co-occurrence relationship of "attribute-state" in historical data, effectively compensating for the instability that a single model may produce under sample distribution drift or extreme conditions. Finally, by using the stability label with the highest cumulative statistical support as an auxiliary label, this invention achieves cross-validation and supplementary judgment of the main decision result.

[0102] Further, a label consistency determination is performed to obtain consistent-inconsistent results. Based on the inconsistent results, the predicted probabilities are extracted and penalized. Then, a label credibility-untrustworthiness determination is performed to obtain credibility-untrustworthiness results, including:

[0103] Based on the auxiliary labels and the final judgment labels, determine whether the labels are consistent;

[0104] If the auxiliary label matches the final decision label, the final decision label is used as the designated label; otherwise, the weight of the penalty correction operation is set (the weight of the penalty correction operation). The value range is from 0.2 to 0.4. To ensure that attention is always drawn when a conflict occurs, a balance point of 0.3 can be taken as the default value. A penalty correction operation is performed on the predicted probability, and the formula is:

[0105]

[0106] In the formula, This represents the predicted probability value after penalty correction. This indicates the weight of the penalty correction operation. This represents the predicted probability value;

[0107] A confidence threshold is set (based on the fault tolerance range of engineering safety, the confidence threshold can be set to a range of 0.6 to 0.8, and 0.7 is taken as the default value of the confidence threshold, so as to ensure that the present invention achieves a balance between judgment sensitivity and engineering reliability). When the predicted probability after penalty correction is greater than or equal to the confidence threshold, it means that the current stability label is reliable, and the final judgment label corresponding to the predicted probability after maximum penalty correction is taken as the designated label; otherwise, it means that the current stability label is unreliable.

[0108] By checking the consistency between auxiliary labels and the final decision label, this invention introduces a dual-channel decision mechanism into the state determination process. This step significantly improves the confidence level of the current stability state and reduces unnecessary subsequent calculations when the two decision results are consistent; conversely, when they are inconsistent, it can promptly identify potential uncertainties or conflict risks, providing clear triggering conditions for subsequent probability correction and reliability assessment. This consistency judgment mechanism is equivalent to "decision consistency verification" in control system diagnosis, effectively reducing the impact of single-model misjudgment on system decision-making. Secondly, a penalty correction operation is introduced in the case of inconsistent decisions, through parameters... By proportionally attenuating the predicted probability, the amplification effect of uncertainty caused by decision conflict can be explicitly reflected at the numerical level. This step allows the present invention to map the logical event of "decision conflict" into a quantifiable change in the probability space, thereby providing continuous and computable input for subsequent credibility determination. Furthermore, by setting a confidence threshold and performing credibility judgment based on the penalty-corrected predicted probability, the present invention achieves a secondary evaluation of whether the stability determination result is "sufficient to be used as the final output." When the corrected predicted probability is not lower than the confidence threshold, it indicates that after considering the impact of conflict, the determination result still has sufficient concentration and reliability, and the stability label corresponding to the highest probability can be directly selected as the specified label output; conversely, it indicates that the uncertainty of the current determination result is still high and it is not suitable for direct use in engineering decision-making. Therefore, the present invention elevates the traditional "direct output of classification results" determination logic to a controlled determination process of "first evaluating the credibility of the determination, then deciding whether to output," significantly enhancing the system's security and robustness under complex operating conditions.

[0109] Furthermore, this invention constructs a two-level logical structure of "label consistency judgment → judgment credibility constraint." This structure allows the main judgment to be responsible for fine-grained differentiation of stable states driven by high-dimensional features, while the auxiliary judgment is only used to answer "whether this judgment conflicts with historical statistical patterns." Only when conflicts or low-confidence situations occur does this invention introduce penalty correction and subsequent feedback mechanisms. This design ensures that the auxiliary judgment does not interfere with the decision freedom of the main judgment in high-confidence scenarios, while also constraining the "overconfidence" of the deep model in risky scenarios.

[0110] S4. Based on unreliable results, construct a probability sequence and calculate the posterior state entropy and causal consistency coefficient. Then, generate a posterior feedback adjustment factor to update, iterate, and provide feedback to the parameters required for penalty correction, and obtain a specified label for storage.

[0111] Specifically, based on unreliable results, a probability sequence is constructed and the posterior state entropy and causal consistency coefficient are calculated. Then, a posterior feedback adjustment factor is generated to update, iterate, and provide feedback to the parameters required for penalty correction, resulting in a specified label, including:

[0112] Based on unreliable results, the penalized and corrected predicted probabilities are extracted, sorted according to time step order, and a probability sequence is constructed. The posterior state entropy of each penalized and corrected predicted probability in the probability sequence is calculated. Then, based on the posterior state entropy, the posterior entropy change rate is calculated, and the posterior entropy time factor is obtained based on the posterior entropy change rate.

[0113] The posterior state entropy of the predicted probability after penalty correction in the probability sequence is calculated using the following formula:

[0114]

[0115] In the formula, Indicates the first The posterior state entropy at each time step. Indicates the first The time step The predicted probability value after penalty correction for each stability label;

[0116] The formula for calculating the rate of change of the posterior entropy is:

[0117]

[0118] In the formula, This represents the rate of change of posterior entropy. Represents absolute value. Indicates the first The posterior state entropy at each time step;

[0119] The formula for obtaining the posterior entropy time factor is:

[0120]

[0121] In the formula, Represents the posterior entropy time factor. This represents the rate of change of the maximum posterior entropy (which can be obtained through statistical analysis of historical monitoring data).

[0122] Based on the probability sequence and the posterior state entropy, the changes in the predicted probability and the posterior state entropy are calculated respectively, and a change vector is generated, including the probability change vector and the posterior entropy change vector.

[0123] The time difference for calculating the predicted probability and the posterior state entropy is given by the following formula:

[0124]

[0125]

[0126] In the formula, Indicates the first With the The time step The change in predicted probability after stability label penalty correction. Indicates the first Time steps down The predicted probability value after penalty correction for each stability label Indicates the first With the The change in posterior state entropy over each time step;

[0127] Based on the change vector, the posterior causal consistency coefficient is calculated. Combined with the posterior entropy time factor, the posterior feedback regulation formula is constructed, and the posterior feedback regulation factor is calculated.

[0128] The formula for calculating the posterior causal consistency coefficient is as follows:

[0129]

[0130] In the formula, Represents the posterior causal consistency coefficient. Represents norm operations;

[0131] The formula for calculating the posterior feedback adjustment factor is:

[0132]

[0133] In the formula, This represents the posterior feedback moderating factor;

[0134] Based on the posterior feedback adjustment factor, the weights of the penalty correction operation are updated, and the updated penalty weights are fed back into the penalty correction operation formula. The penalty correction prediction probability is then recalculated to perform a reliable judgment and select a specified label.

[0135] The formula for updating the weights of the penalty correction operation is as follows:

[0136]

[0137] In the formula, This indicates an update to the penalty weights;

[0138] If the label remains unreliable during continuous iterations, a warning message will be sent to regulators, indicating that the currently designated label is unreliable and suggesting a manual re-inspection.

[0139] By transforming the originally discrete and static stability determination problem into a probabilistic evolution analysis problem with temporal continuity, this invention overcomes the limitation of traditional methods that rely solely on single-moment determination results. By constructing a probability sequence of the penalized and corrected predicted probabilities in chronological order and calculating the posterior state entropy, the changing pattern of determination uncertainty over time can be revealed, providing a reliable basis for subsequent dynamic adjustment. By introducing the posterior entropy change rate and its normalized posterior entropy time factor, this invention effectively avoids the risk of misjudgment caused by relying solely on probability magnitude for reliable determination. When the predicted probability exhibits drastic fluctuations or persistent instability over time, even if a single predicted probability is high, the posterior entropy time factor can still reflect potential risks, thereby improving the safety redundancy of the overall determination process. Furthermore, by constructing a predicted probability change vector and a posterior entropy change vector and calculating the posterior causal consistency coefficient, this invention introduces the higher-order criterion of "change consistency" into the feedback adjustment mechanism. This allows the adjustment of penalty weights to no longer rely on empirical settings but rather on the inherent logic of whether probability changes are accompanied by reasonable changes in uncertainty for adaptive updates. This mechanism significantly enhances the interpretability of the adjustment process. Based on this, the penalty correction weight is iteratively updated through a posterior feedback adjustment factor, enabling the invention to gradually reduce the risk of information loss due to excessive penalty when a judgment is deemed unreliable, while avoiding the spread of uncertainty caused by insufficient penalty. If a reliable judgment result cannot be obtained within a limited number of iterations, a manual review prompt is triggered, thereby achieving a reasonable division of labor between automatic judgment and manual decision-making.

[0140] Secondly, existing technologies, by introducing entropy, confidence level, or probability thresholds, are typically only used for static evaluation of single judgment results, failing to answer the crucial question of whether the root cause of unreliable judgments is short-term fluctuations or systemic instability. This invention, however, constructs a posterior state entropy based on the predicted probability after penalty correction, and further introduces the posterior entropy change rate and posterior entropy time factor, enabling the judgment system to distinguish between uncertainties caused by random noise and uncertainties caused by structural evolution. Moreover, the posterior causal consistency coefficient is not a traditional correlation analysis or similarity calculation, but a dedicated criterion constructed specifically for the core issue of judgment credibility: whether changes in predicted probability are accompanied by reasonable changes in uncertainty. This coefficient, by combining the probability change vector and the posterior entropy change vector, establishes an inherent consistency constraint between the judgment result adjustment behavior and the uncertainty response, giving the feedback adjustment process a clear causal explanation path. Thirdly, the posterior entropy time factor and the causal consistency coefficient form a strong coupling relationship and act on the penalty correction weight, a "judgment generation process parameter." This technical approach of "adjusting the judgment process rather than directly correcting the judgment result" enables the present invention to gradually approach a credible judgment state while maintaining the consistency of the judgment logic, thus avoiding the judgment oscillation or logic distortion caused by multiple direct corrections of the judgment result in traditional methods.

[0141] Furthermore, storage includes:

[0142] Edit the currently specified label and predicted probability into a folder, add a timestamp and corresponding stability label to the folder, store the folder in the database, and sort it according to the timestamp order after storage.

[0143] By structurally encapsulating the current specified label and predicted probability, and combining this with timestamp information for database storage and sorting, unified management and traceable analysis of stability assessment results are achieved. This method effectively enhances the organization and auditability of assessment data, supports evolutionary modeling of state trends and anomaly feedback control, and improves the system's ability to dynamically identify the stability state of support structures.

[0144] This embodiment also provides a computer device applicable to the dynamic determination method for the stability of support structures used in deep foundation pit construction, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the dynamic determination method for the stability of support structures used in deep foundation pit construction as proposed in the above embodiment.

[0145] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0146] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the dynamic determination method for the stability of support structures used in deep foundation pit construction, as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for dynamically determining the stability of a support structure for deep foundation pit construction, characterized in that: include, Multi-source heterogeneous data of deep foundation pit support structure are collected, preprocessed, and a complete reconstruction sequence is generated. A stability label is constructed and combined with the complete reconstructed sequence to build training samples. The main decision model is then constructed, and the training samples are used as the input of the main decision model to output the predicted probability distribution and select the final decision label. Auxiliary labels are obtained based on the final judgment label, and label consistency judgment is performed to obtain consistent-inconsistent results. Then, based on the inconsistent results, the predicted probability is extracted and penalized and corrected. Finally, label credibility-untrustworthiness judgment is performed to obtain credibility-untrustworthiness results. Based on unreliable results, a probability sequence is constructed and the posterior state entropy and causal consistency coefficient are calculated. Then, a posterior feedback adjustment factor is generated to update, iterate, and provide feedback to the parameters required for penalty correction, and the specified label is obtained and stored.

2. The method for dynamically determining the stability of the support structure for deep foundation pit construction according to claim 1, characterized in that: The constructed stability label is combined with the complete reconstructed sequence to build training samples, construct the main decision model, use the training samples as input to the main decision model, output the predicted probability distribution, and select the final decision label, including: Stability labeling of diaphragm wall structures in deep foundation pits using the Delphi method, including stable states. Critical state and instability The complete reconstructed sequence and stability label are constructed as input pairs and used as training sample pairs; A master decision model is constructed using a one-dimensional convolutional neural network and a bidirectional long short-term memory neural network structure, including an input layer, a convolutional extraction layer, a max pooling layer, a bidirectional LSTM layer, a fully connected layer, and a softmax output layer. The cross-entropy loss function is used as the loss function of the master decision model to minimize the gap between the predicted distribution and the true label. The Adam optimizer is used to update and optimize the parameters of the master decision model. During the iteration process, when the loss value of the loss function no longer decreases significantly, the iteration stops and the trained master decision model is output. The training sample pairs are used as input to the main decision model. The input layer receives the training samples and passes them to the convolutional extraction layer. The convolutional extraction layer is then used to obtain the feature values ​​of the complete reconstructed sequence in the training samples. The temporal features are downsampled using a max pooling layer to generate sampled features; The forward and backward output features of the bidirectional LSTM layers are obtained and concatenated to obtain the feature representation of each time step. ; Average pooling is performed on all feature representations to generate the overall representation vector; A fully connected layer is used to perform a linear transformation on the overall representation vector to obtain a classification scoring vector, which contains the original score value of each stable label; Based on the classification scoring vector, the Softmax output layer is used to map the probabilities to obtain the predicted probability distribution, and the stability label corresponding to the highest probability is selected as the final decision label.

3. The method for dynamically determining the stability of support structures used in deep foundation pit construction as described in claim 2, characterized in that: The process of obtaining auxiliary tags based on the final determined tag includes: Backtracking is performed based on the final judgment label, the corresponding complete reconstruction sequence is extracted and discrete operation is performed to obtain equally segmented discrete intervals, the frequency of occurrence of each discrete interval under different stability labels is counted, and an attribute-state frequency mapping table is constructed. For each stability label, calculate the cumulative statistical support of the complete reconstructed sequence for each stability label; The cumulative statistical support scores are sorted in descending order, and the stability label corresponding to the highest cumulative statistical support score is selected as the auxiliary label.

4. The method for dynamically determining the stability of support structures used in deep foundation pit construction as described in claim 3, characterized in that: The label consistency determination is performed to obtain a consistent-inconsistent result. Then, based on the inconsistent result, the predicted probability is extracted and penalized. Finally, the label trustworthiness-untrustworthiness determination is performed to obtain a trustworthiness-untrustworthiness result, including: Based on the auxiliary labels and the final judgment labels, determine whether the labels are consistent; If the auxiliary label matches the final judgment label, the final judgment label will be used as the designated label; otherwise, the weight of the penalty correction operation will be set, and the penalty correction operation will be performed on the predicted probability. A confidence threshold is set. If the predicted probability after penalty correction is greater than or equal to the confidence threshold, it means that the current stability label is reliable. The final judgment label corresponding to the predicted probability after maximum penalty correction is taken as the designated label. Otherwise, it means that the current stability label is unreliable.

5. The method for dynamically determining the stability of support structures used in deep foundation pit construction as described in claim 4, characterized in that: Based on unreliable results, a probability sequence is constructed, and posterior state entropy and causal consistency coefficient are calculated. Then, a posterior feedback adjustment factor is generated to update, iterate, and provide feedback on the parameters required for penalty correction, resulting in a specified label, including: Based on unreliable results, the penalized and corrected predicted probabilities are extracted, sorted according to time step order, and a probability sequence is constructed. The posterior state entropy of each penalized and corrected predicted probability in the probability sequence is calculated. Then, based on the posterior state entropy, the posterior entropy change rate is calculated, and the posterior entropy time factor is obtained based on the posterior entropy change rate. Based on the probability sequence and the posterior state entropy, the changes in the predicted probability and the posterior state entropy are calculated respectively, and a change vector is generated, including the probability change vector and the posterior entropy change vector. Based on the change vector, the posterior causal consistency coefficient is calculated. Combined with the posterior entropy time factor, the posterior feedback regulation formula is constructed, and the posterior feedback regulation factor is calculated. Based on the posterior feedback adjustment factor, the weights of the penalty correction operation are updated, and the updated penalty weights are fed back into the penalty correction operation formula. The penalty correction prediction probability is then recalculated to perform a reliable judgment and select a specified label. If the label remains unreliable during continuous iterations, a warning message will be sent to regulators, indicating that the currently designated label is unreliable and suggesting a manual re-inspection.

6. The method for dynamically determining the stability of support structures for deep foundation pit construction as described in claim 5, characterized in that: The storage includes: Edit the currently specified label and predicted probability into a folder, add a timestamp and corresponding stability label to the folder, store the folder in the database, and sort it according to the timestamp order after storage.

7. The method for dynamically determining the stability of support structures used in deep foundation pit construction as described in claim 6, characterized in that: The preprocessing of the multi-source heterogeneous data collected on the deep foundation pit support structure includes: Sensors are deployed on the underground continuous wall of the deep foundation pit to collect multi-source heterogeneous data for preprocessing and generate preprocessed data.

8. The method for dynamically determining the stability of support structures used in deep foundation pit construction as described in claim 7, characterized in that: The generation of the complete reconstructed sequence includes: Based on the preprocessed data, a sliding window technique is used to divide the data into window data, which is then constructed into an input matrix. Uniform random occlusion is then used to occlude each data point in the input matrix to generate a mask matrix. A pre-trained encoding network, including a mask autoencoder and a decoder, is applied. The mask autoencoder performs contextual modeling on the mask matrix to generate a latent space representation. Finally, the pre-trained decoder is used to reconstruct the latent space representation, outputting a complete reconstructed sequence.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the dynamic determination method for the stability of support structures for deep foundation pit construction as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the dynamic determination method for the stability of the support structure for deep foundation pit construction as described in any one of claims 1 to 8.

Images