Foundation pit displacement prediction and intelligent early warning method and system

Abstract

The invention relates to a foundation pit displacement prediction and intelligent early warning method. Firstly, a multi-parameter sensor is adopted in a data acquisition module to collect key data of a foundation pit; then extracting data closely related to foundation pit displacement prediction and early warning in a data processing module by adopting an improved fine complex multivariate multi-scale symbol dynamic entropy (RCMMSDE) feature extraction method; a quantum derived characteristic modal decomposition (Q-FMD) technology is introduced to decompose the screened data; then a K-fold cross validation method (K-CV) is selected to be combined with a biogeography optimization algorithm (BBO), alpha and phi hyper-parameters of the ETSformer model are optimized, and the model performance is improved; and finally, foundation pit displacement evaluation is established according to expert experience in a grade evaluation and monitoring early warning module, the specific displacement distance is observed through camera videos, and early warning is completed through a central control display platform.

Description

Technical Field

[0001] The present invention relates to the field of construction engineering, in particular to a method and system for foundation pit displacement prediction and intelligent early warning. Background Technique

[0002] With the acceleration of the urbanization process, deep foundation pit engineering plays an important role in various construction projects. However, the construction environment of foundation pit engineering is complex and changeable. Traditional monitoring and early warning methods mainly rely on manual regular measurement and empirical judgment, with limitations such as untimely acquisition of monitoring data, low efficiency of data processing and analysis, insufficient early warning accuracy, and inability to achieve intelligent decision-making. These methods are difficult to accurately predict the displacement of the foundation pit, easily miss the critical early warning time, and increase the engineering risk.

[0003] In recent years, the development of sensor technology, Internet of Things technology, big data analysis, and artificial intelligence technology has provided new technical means for foundation pit displacement prediction and intelligent early warning. By collecting data in real time through high-precision displacement sensors and using big data analysis and machine learning algorithms to establish a prediction model, accurate prediction and risk assessment of foundation pit displacement can be achieved. However, existing systems still face challenges in data fusion and processing, model generalization ability, timeliness and accuracy of early warning, and user interaction and decision support. Summary of the Invention

[0004] Object of the Invention: The present invention relates to a platform for foundation pit displacement prediction and intelligent early warning. In the data acquisition module of this platform, a variety of sensors are used to collect key data of the foundation pit, covering multi-dimensional information such as soil strength, soil layer distribution, foundation pit plane size, wall horizontal displacement, and wall pressure change. Then, the data processing module adopts an improved feature extraction method of refined composite multi-scale multi-symbol dynamic entropy (RCMMSDE) to deeply analyze the collected data and accurately extract the feature data closely related to foundation pit displacement prediction and early warning. Further, the quantum-derived feature mode decomposition (Q-FMD) technology is used to decompose the selected data, effectively reducing the data complexity and improving the data processing efficiency. In the prediction module, the α and φ hyperparameters of the ETSformer model are optimized by combining the K-fold cross-validation method (K-CV) and the biogeography-based optimization algorithm (BBO), significantly improving the prediction performance of the model. Finally, in the grade evaluation and monitoring early warning module, a foundation pit displacement evaluation system is constructed based on expert experience, and the specific displacement distance is observed in real time through a camera video, and accurate early warning is realized through the central control display platform to ensure the safe construction of the foundation pit project.

[0005] Technical Solution: A method for foundation pit displacement prediction and intelligent early warning includes the following steps:

[0006] Step 1: Use multiple multi-parameter sensors to collect key data of the foundation pit in construction engineering, including soil strength, soil layer distribution, foundation pit plane size, wall horizontal displacement, and wall pressure change. Use the Maximum Entropy Partitioning algorithm MEP to divide the time series into ε intervals, and use the improved refined composite multi-scale symbolic dynamic entropy RCMMSDE feature extraction method to extract data closely related to foundation pit displacement prediction;

[0007] Step 2: Perform quantum derivative feature mode decomposition Q-FMD on the data after feature extraction;

[0008] Step 3: Establish a foundation pit displacement prediction model based on the ETSformer model;

[0009] Step 4: Combine the K-fold cross-validation method with the Biogeography-Based Optimization algorithm BBO to form the K-CV-BBO algorithm, and use the combined algorithm to optimize the hyperparameters in the ETSformer model to obtain the optimized K-CV-BBO-ETSformer model. The hyperparameters include: smoothing parameter α, trend damping parameter

[0010] Step 5: Use the optimized K-CV-BBO-ETSformer model to predict the foundation pit displacement and obtain the prediction result of the foundation pit displacement;

[0011] Step 6: Establish a foundation pit displacement level assessment based on expert experience. Finally, display the specific position of the displacement point through a camera video and complete the early warning through the central control display platform.

[0012] Further, the specific method for using the improved refined composite multi-scale symbolic dynamic entropy RCMMSDE to extract features from the collected data in Step 1 is as follows:

[0013] Step 1.1: First, for each time series, use the Maximum Entropy Partitioning algorithm MEP to divide the time series into ε intervals, and map the values of each time series to a symbol;

[0014] Step 1.2: For each symbolized time series, describe the subsequent states and transition probabilities between symbol sequences through a D-Markov machine;

[0015] Step 1.3: Sum the state transition matrices of all channels, including soil strength, soil layer distribution, foundation pit plane size, wall horizontal displacement, and wall pressure change. During the fusion process of the multi-variable time series state transition matrix, introduce the channel weight coefficient μ k , the state transition matrix Π of the multi-variable time series m :

[0016]

[0017] where: Π k,m is the m-th element in the k-th sequence transition matrix, D is the depth of the D-Markov machine, and l is the number of channels;

[0018] Step 1.4: Calculate the RCMMSDE according to the state transition matrix Π m .

[0019] From the above process of feature extraction, it can be obtained that finally, from the time series: P, T, V, R, W, the sequences: P, T, R, W are extracted. (P soil strength, T soil layer distribution, V foundation pit plan dimension, R wall horizontal displacement, W wall pressure change)

[0020] The improved refined composite multi-scale symbolic dynamic entropy (RCMMSDE) can significantly improve the accuracy and reliability of prediction, reflect the influence differences of different monitoring parameters (such as soil strength, soil layer distribution, etc.) on the foundation pit safety through the weight coefficient, overcome the limitation of equal-weight addition of the original method, and through enhancing the anti-noise ability and fine multi-scale analysis, make the features extracted from the foundation pit data more objective and accurate, and effectively identify abnormal foundation pit displacements.

[0021] Furthermore, in Step 2, quantum-derived feature mode decomposition Q-FMD is performed on the preprocessed data. Taking the soil strength P parameter as an example, the specific steps are as follows:

[0022] Step 2.1: Initialize the parameters, and set the soil strength signal sequence as P(n) = P1, P2,..., P n , where n is the length of the signal. According to the quantum state initialization formula:

[0023]

[0024] where, |t> is the quantum ground state, representing the quantum state at time point t, P(t) is the amplitude of the original signal at time t, ||P|| is the L2 norm of the signal for normalization, and Φ(t) maps the classical signal into a quantum superposition state to achieve parallel processing;

[0025] Set the parameter combination: the number of decomposition modes n = 8, the number of filters K = 10, the iteration parameter i = 1; the filter length L = 32.

[0026] Step 2.2: Quantum feature decomposition formula:

[0027] u QFD = e -iHτ

[0028]

[0029] Obtain the decomposed modal components according to the above quantum feature decomposition formula; where k = 1, 2, ···, K; the fault characteristic frequency γ, the Hamiltonian H, which describes the energy characteristics of the system; The number operator of particles, which characterizes the energy of the k-th frequency component, the evolution time τ, which controls the splitting degree of the quantum state; the iteration parameter i;

[0030] Convert the frequency-domain signal into a time-domain signal through quantum measurement and component reconstruction, and reconstruct the IMF component. The formula is as follows:

[0031]

[0032] f -1 Inverse Fourier transform, IMF K (t) Measure the statistical quantum state multiple times to reconstruct the IMF component. The expected value of the quantum measurement of the k-th mode.

[0033] Step 2.3: Reconstruct the filter coefficients according to the quantum component IMF K (t) to obtain the modal components and estimate the fault period of the signal for updating;

[0034] Step 2.4: Determine whether the current iteration process reaches the maximum number of iterations. If it reaches, return to Step 2.3; otherwise, continue to execute Step 2.5.

[0035] Step 2.5: Calculate the correlation coefficient between two adjacent modal components, establish a K×K correlation matrix CC, and calculate the correlation kurtosis between the correlation matrix CC and the estimated signal fault period. Select the modal component with a larger correlation kurtosis as the decomposed component;

[0036] Step 2.6: Determine whether the current number of modes reaches the set number n. If it reaches, stop the decomposition and output the result; if it does not reach, return to Step 2.2 and continue the above operations;

[0037] Step 2.7: Finally, output multiple modal components. Each modal component represents the characteristics of different frequency ranges in the soil strength signal. Finally, 8 modal components u1, u2, …, u8 are output, and each modal component represents the characteristics of different frequency ranges in the soil strength signal. To better characterize the 8 modal components of the soil strength, P can be expressed as: P' = [u P1 , u P2 , …, u P8 .

[0038] Similarly, for the other three time series T, R, and W, we can get: T' = [u T1 , u T2,…,u T8 ; R' = [u R1 , u R2 ,…, u R8 ; W' = [u W1 , u W2 ,…, u W8 .

[0039] Merge the above four groups of data into the final dataset X = [P', T', R', W'].

[0040] This method accurately captures the key features in the foundation pit displacement signal through an adaptive filter bank, effectively filters out high-frequency noise, extracts the pure foundation pit displacement signal, and lays a clear data foundation for anomaly detection. The multi-scale analysis ability of Q-FMD makes it possible to identify the early signs of foundation pit displacement at different time scales, thus enhancing the foresight of prediction.

[0041] Furthermore, the hyperparameters in the ETSformer model include:

[0042] Smoothing parameter α:

[0043]

[0044] L t represents the trend level at the current moment, Y t represents the observed value, is the seasonal component

[0045] Trend damping parameter

[0046]

[0047] represents the growth component at the current time step t extracted in the nth layer encoder.

[0048] The main advantage of the ETSformer model lies in its innovative exponential smoothing attention (ESA) and frequency attention (FA) mechanisms, which can accurately capture the long-term dependencies and seasonal patterns in time series. For the analysis of foundation pit displacement, ETSformer summarizes the characteristic structure of foundation pit displacement data by performing hierarchical, growth, and cycle decomposition. This decomposition method helps to decompose the complex foundation pit displacement data into easily understandable components, such as horizontal displacement, growth trend, and periodic changes, so as to more accurately identify anomalies.

[0049] Furthermore, step 4 combines the K-fold cross-validation method with the biogeography-based optimization algorithm BBO to form the K-CV-BBO algorithm. The specific process is as follows:

[0050] Introduce the combination of K-CV and BBO optimization algorithms to optimize the key parameters α of the ETSformer model, perform optimization. Use K-CV to construct the loss function of ETSformer, and use the BBO algorithm to find the global optimal solution of the loss function; the loss function optimized by the BBO algorithm under the meaning of K-CV is:

[0051]

[0052] R j = ω1R j,1 + ω2R j,2

[0053] In the formula: R j，1 is the mean square error loss function, R j,2 is the cross-entropy difference loss function, and the loss function optimized by the BBO algorithm under the meaning of K-CV is R j , ω1 and ω2 are weight values, which are determined by optimizing with the BBO algorithm; K is the number of cross-validation times; n is the total number of training samples; y i is the original value of the i-th training sample; is the predicted value of the i-th training sample when the parameters are α i , ;

[0054] Generate the initial solution, randomly generate N survival environments, that is, N solutions, and each solution represents a set of parameters of the ETSformer model, namely α and

[0055] Calculate the fitness. Divide the training samples into K parts, calculate the loss function, and train the ETSformer model on each set of parameters on each training set, and calculate R j，1 , R j,2 , and finally calculate the optimized loss function of each parameter under K-fold cross-validation as R j ;

[0056] Migration operation. According to the fitness value of the solution, perform the migration operation to simulate the immigration and emigration of species between different survival environments;

[0057] Mutation operation. Perform the mutation operation on each solution in the solution set to simulate the impact of unexpected events in the survival environment on the number of species;

[0058] Iterative update until the stopping condition is met. The stopping condition includes reaching the maximum number of iterations or the fitness change is less than the preset threshold;

[0059] Step 4.17: Determine the optimal parameters, select the solution with the highest fitness from the iterative process as the optimal solution, and a set of parameters α of the corresponding ETSformer model and are what we want.

[0060] K-CV divides the dataset into a training set and a validation set, repeats model training and validation, ensures that the model has good prediction performance for unseen data, and thus avoids overfitting. BBO simulates the migration and mutation processes in biogeography, effectively explores the parameter space, and finds the global optimal solution. When used in combination, K-CV provides a robust performance evaluation for BBO, while BBO selects the optimal parameters for the model under K-CV, jointly improving the prediction accuracy and robustness of the model.

[0061] Furthermore, the said Step 5 includes the following steps:

[0062] Step 5.1: Divide the final dataset into a training set, a validation set, and a test set according to a preset ratio;

[0063] Step 5.2: Use the divided training set to train the improved K-CV-BBO-ETSformer model, enabling the model to learn the patterns and relationships in the data;

[0064] Step 5.3: Use the validation set to validate the model, adjust the model parameters, and complete the training;

[0065] Step 5.4: Use the trained model to predict the test set and obtain the prediction results of the foundation pit displacement.

[0066] Furthermore, the said Step 6 includes:

[0067] Step 6.1: Based on expert experience, establish a foundation pit displacement evaluation model. This model classifies the stability level of the foundation pit based on soil layer parameters, wall displacement, and excavation depth factors;

[0068] Step 6.2: Input the inverted soil layer parameters and the on-site monitored wall horizontal displacement data into the evaluation model to evaluate the stability level of the foundation pit in real time; when the stability level of the foundation pit reaches a relatively dangerous or dangerous level, send a warning signal through the central control display platform;

[0069] Step 6.3: While sending the warning signal, use the cameras installed around the foundation pit to capture video images of the foundation pit in real time, and use image processing technology to automatically identify and measure the displacement distance of the foundation pit wall;

[0070] Step 6.4: Compare and analyze the displacement distance obtained from video observation with the early warning result of the evaluation model. If the two are consistent, confirm that there are potential safety hazards in the foundation pit and promptly notify the construction personnel to take corresponding emergency measures; if the two are inconsistent, further analyze the reasons, adjust the evaluation model or re - perform parameter inversion to ensure the accuracy and reliability of the early warning.

[0071] A foundation pit displacement prediction and intelligent early warning system includes the following modules:

[0072] Data acquisition module: Use a variety of sensors to collect key data of the foundation pit, including soil strength, soil layer distribution, foundation pit plane size, wall horizontal displacement, and wall pressure change of the foundation pit;

[0073] Data processing module: Use the improved refined composite multi - scale symbolic dynamic entropy RCMMSDE feature extraction method to extract data closely related to foundation pit displacement prediction and early warning; and use quantum - derived feature modal decomposition Q - FMD to decompose the selected data to reduce data complexity;

[0074] Model prediction module: Select the combination of K - fold cross - validation method and biogeography - based optimization algorithm BBO to optimize the α hyperparameters of the ETSformer model to improve the model performance and predict the foundation pit displacement;

[0075] Level evaluation, monitoring and early warning module: Establish an evaluation of the severity level of foundation pit displacement according to expert experience, display the specific position of the displacement point through a camera video, and complete early warning through the central control display platform.

[0076] Furthermore, for the quantum - derived feature modal decomposition Q - FMD, taking the soil strength P parameter as an example, the specific steps are as follows:

[0077] Step 2.1: Initialize the parameters, set the soil strength signal sequence as P(n)=P1, P2,..., P n , where n is the length of the signal. According to the quantum state initialization formula:

[0078]

[0079] where, |t> is the quantum ground state, representing the quantum state at time point t, P(t) is the amplitude of the original signal at time t, ||P|| is the L2 norm of the signal for normalization, and Φ(t) maps the classical signal into a quantum superposition state for parallel processing;

[0080] Set the parameter combination: the number of decomposition modes n = 8, the number of filters K = 10, the iteration parameter i = 1; the filter length L = 32.

[0081] Step 2.2: Quantum feature decomposition formula:

[0082] u QFD = e -iHτ

[0083]

[0084] Obtain the decomposed modal components according to the above quantum feature decomposition formula; where k = 1, 2, ···, K; the fault characteristic frequency γ, the Hamiltonian H, which describes the energy characteristics of the system; The number operator of particles, which characterizes the energy of the k-th frequency component, the evolution time τ, which controls the splitting degree of the quantum state; the iteration parameter i;

[0085] Convert the frequency-domain signal into a time-domain signal through quantum measurement and component reconstruction, and reconstruct the IMF components. The formula is as follows:

[0086]

[0087] f -1 Inverse Fourier transform, IMF K (t) Statistically measure the quantum state multiple times to reconstruct the IMF components, The expected value of the quantum measurement for the k-th mode.

[0088] Step 2.3: Reconstruct the filtering coefficient according to the quantum component IMF K (t) to obtain the modal component and estimate the fault period of the signal for updating;

[0089] Step 2.4: Determine whether the current iteration process reaches the maximum number of iterations. If it reaches, return to Step 2.3; otherwise, continue to execute Step 2.5.

[0090] Step 2.5: Calculate the correlation coefficient between two adjacent modal components, establish a K×K correlation matrix CC, and calculate the correlation kurtosis between the correlation matrix CC and the estimated signal fault period. Select the modal component with a larger correlation kurtosis as the decomposed component;

[0091] Step 2.6: Determine whether the current number of modes reaches the set number n. If it reaches, stop the decomposition and output the result; if not, return to Step 2.2 and continue the above operations;

[0092] Step 2.7: Finally, output multiple modal components, and each modal component represents the characteristics of different frequency ranges in the soil strength signal.

[0093] Furthermore, select the K-fold cross-validation method combined with the biogeography-based optimization algorithm BBO to optimize the α of the ETSformer model, The specific process of the hyperparameters is as follows:

[0094] Introduce the combination of K-CV and BBO optimization algorithms to optimize the key parameters α of the ETSformer model, Perform optimization. Use K-CV to construct the loss function of ETSformer, and use the BBO algorithm to find the global optimal solution of the loss function; the BBO algorithm optimization loss function in the sense of K-CV is:

[0095]

[0096] R j = ω1R j,1 + ω2R j,2

[0097] In the formula: R j，1 is the mean square error loss function, R j,2 is the cross-entropy difference loss function, and the BBO algorithm optimization loss function in the sense of K-CV is R j , ω1, ω2 are weight values, determined by the BBO algorithm optimization; K is the number of cross-validation times; n is the total number of training samples; y i is the original value of the i-th training sample; is the predicted value of the i-th training sample when the parameter is α i , ;

[0098] Generate the initial solution, randomly generate N survival environments, that is, N solutions, and each solution represents a set of parameters of the ETSformer model, that is, α and

[0099] Calculate the fitness. Divide the training samples into K parts, calculate the loss function, and for each set of parameters Train the ETSformer model on each training set, and calculate R j，1 , R j,2 , and finally calculate the optimization loss function of each parameter under K-fold cross-validation as R j ;

[0100] Migration operation. According to the fitness value of the solution, perform the migration operation to simulate the immigration and emigration of species between different survival environments;

[0101] Mutation operation. Perform the mutation operation on each solution in the solution set to simulate the impact of unexpected events in the survival environment on the number of species;

[0102] Iterative update until the stop condition is met; determine the optimal parameters, select the solution with the highest fitness from the iterative process as the optimal solution, and a set of parameters α and of the corresponding ETSformer model are what we want

[0103] Beneficial effects:

[0104] The improved refined composite multi - scale symbolic dynamic entropy (RCMMSDE) can significantly improve the accuracy and reliability of prediction. Through enhancing the anti - noise ability and fine multi - scale analysis, the features extracted from foundation pit data are more objective and accurate, effectively identifying anomalies. This method shows good robustness to noise and interference, ensuring the stability of the prediction model in a changing environment. By providing more interpretable features, this method also enhances the transparency of the model, facilitating in - depth analysis and decision - making by experts, thus providing important support for the safety management and maintenance of foundation pits.

[0105] The robustness of Q - FMD to noise and small changes in signals ensures the stability of prediction in the face of natural environmental changes and measurement errors. The automated feature extraction process reduces the dependence on expert knowledge, simplifies the analysis process, and significantly improves the anomaly prediction accuracy and response speed of the combined appropriate prediction model.

[0106] The ETSformer model utilizes the exponential smoothing attention (ESA) and frequency attention (FA) mechanisms, which can accurately capture long - term dependencies and seasonal patterns in time series, being crucial for analyzing trends and periodicities. In addition, the modular decomposition architecture of ETSformer helps to decompose complex data into easily understandable components, such as level, growth, and seasonal components, maintaining robustness to noise and non - stationarity, and thus more accurately identifying anomalies.

[0107] K - CV ensures good prediction performance of the model for unseen data by splitting the dataset into training and validation sets and repeating model training and validation, thus avoiding overfitting. BBO simulates the migration and mutation processes in biogeography, effectively exploring the parameter space to find the global optimal solution. When used in combination, K - CV provides a robust performance evaluation for BBO, while BBO selects the optimal parameters for the model under K - CV, jointly improving the prediction accuracy and robustness of the model. Brief description of the drawings

[0108] Figure 1 It is a schematic diagram of the Q - FMD feature mode decomposition of the present invention;

[0109] Figure 2 It is a schematic diagram of the improved refined composite multi - scale symbolic dynamic entropy (RCMMSDE) feature extraction method of the present invention;

[0110] Figure 3 It is a schematic diagram of the (K - CV - BBO) algorithm of the present invention;

[0111] Figure 4 It is a schematic diagram of the framework of the present invention. Detailed implementation manners

[0112] In order to enable those skilled in the art of the present technology to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0113] The specific method for extracting features from the collected data (soil strength, soil layer distribution, foundation pit plane size, wall horizontal displacement, wall pressure change) using the improved refined composite multi - scale symbolic dynamic entropy (RCMMSDE) is as follows:

[0114] Step 1.1: First, for each time series (P soil strength, T soil layer distribution, V foundation pit plane size, R wall horizontal displacement, W wall pressure change), use the maximum entropy partitioning algorithm (Maximum entropy partitioning, MEP) to divide the time series into ε intervals, and map the values of each time series to a symbol. For example, for the soil strength time series {P(t)}, we can symbolize it as {yp(t)}. And perform symbolic processing on the time series:

[0115]

[0116] In the formula, is the time series; is the symbol sequence; F(·) is the mapping from the time series to the symbol sequence; N is the length of the time series; k = 1, 2, … l, where l is the number of channels of the multi - variable time series.

[0117] Step 1.2: For each symbolized time series, describe the subsequent states and transition probabilities between the symbol sequences through the D - Markov machine. Define different state sequences and convert the time series into a time - related matrix:

[0118]

[0119] In the formula, is the state sequence; i = 1, 2, …, N−(D−1), and D is the depth of the D - Markov machine. There are a total of ε D different time series. To highlight the transition probability situation of different state sequences, the state transition matrix:

[0120]

[0121] In the formula: type(·) is the mapping from the symbol sequence to the state sequence; q a and q b are different state sequences.

[0122] Step 1.3: Sum up the state transition matrices of all channels (soil strength, soil layer distribution, foundation pit plan dimension, wall horizontal displacement, wall pressure change), and introduce the channel weight coefficient μ during the fusion process of the multi-variable time series state transition matrix. k , the state transition matrix Π of the multi-variable time series m :

[0123]

[0124] In the formula: Π k,m is the m-th element in the k-th sequence transition matrix.

[0125] Step 1.4: Calculate RCMMSDE according to the state transition matrix Π m :

[0126]

[0127] In the formula: X is the set of multi-variable time series; ε is the number of intervals during symbolization; D is the depth of the D-Markov machine; τ is the scale factor.

[0128] From the above process of feature extraction, it can be obtained that finally, from the time series: P, T, V, R, W, the sequences: P, T, R, W are extracted.

[0129] The improved refined composite multi-variable multi-scale symbolic dynamic entropy (RCMMSDE) can significantly improve the accuracy and reliability of prediction, reflect the influence differences of different monitoring parameters (such as soil strength, soil layer distribution, etc.) on the foundation pit safety through the weight coefficient, overcome the limitations of equal-weight addition in the original method, and through enhancing the anti-noise ability and fine multi-scale analysis, make the features extracted from the foundation pit data more objective and accurate, and effectively identify abnormal foundation pit displacements.

[0130] Then perform quantum-derived feature mode decomposition (Q-FMD):

[0131] Step 2.1: Initialize the parameters, set the soil strength signal sequence as P(n) = P1, P2,..., P n , where N is the length of the signal. According to the quantum state initialization formula:

[0132]

[0133] Among them, |t> is the quantum ground state, representing the quantum state at time point t, P(t) is the amplitude of the original signal at time t, ||P|| is the L2 norm of the signal for normalization, and Φ(t) maps the classical signal into a quantum superposition state to achieve parallel processing. Set the parameter combination: the number of decomposition modes n = 8, the number of filters K = 10, the iteration parameter i = 1; the filter length L = 32.

[0134] Step 2.2: Quantum eigenvalue decomposition formula:

[0135] u QFD = e -iHτ

[0136]

[0137] Obtain the decomposed modal components according to the above quantum eigenvalue decomposition formula, where k = 1, 2, ···, K; the fault characteristic frequency γ, the Hamiltonian H, which describes the energy characteristics of the system; The number operator of particles, which characterizes the energy of the k-th frequency component, the evolution time τ, which controls the splitting degree of the quantum state. Convert the frequency-domain signal into a time-domain signal through quantum measurement and component reconstruction, and reconstruct the IMF component. The formula is as follows:

[0138]

[0139] f -1 Inverse Fourier transform, IMF K (t) Statistically measure the quantum state multiple times and reconstruct the IMF component. The expected value of the quantum measurement for the i-th mode.

[0140] Step 2.3: Update the filtering coefficient according to the quantum component IMF K (t), and reconstruct to obtain the modal component and the estimated signal fault period ( is The time corresponding to the local maximum value reached after the zero crossing of the autocorrelation spectrum ), and let i = i + 1.

[0141] Step 2.4: Determine whether the current iteration process reaches the maximum number of iterations. If it reaches, return to Step 2.3; otherwise, continue to execute Step 2.5.

[0142] Step 2.5: Calculate the correlation coefficient between two adjacent modal components, establish a K×K correlation matrix CC(K×K), and calculate the correlation kurtosis between the correlation matrix and the estimated signal fault period. Select the modal component with a larger correlation kurtosis as the decomposed component, and let K = K - 1.

[0143] Step 2.6: Determine whether the current number of modes reaches the set number n. If it reaches, stop the decomposition and output the result; if it does not reach, return to Step 2.2 and continue the above operations.

[0144] Step 2.7: Finally, eight modal components u1, u2, …, u8 are output. Each modal component represents the characteristics of different frequency ranges in the soil strength signal. To better characterize the eight modal components of the soil strength, P can be expressed after Q-FMD feature modal decomposition as: P' = [u P1 , u P2 , …, u P8 .

[0145] Similarly, for the other three time series T, R, and W, we can obtain: T' = [u T1 , u T2 , …, u T8 ; R' = [u R1 , u R2 , …, u R8 ; W' = [u W1 , u W2 , …, u W8 .

[0146] Merge the above four groups of data into the final dataset X = [P', T', R', W'].

[0147] Step 3.1: Map the input set X t-L:t = {P', T', R', W'} t-L:t to the latent space through the convolutional filter Conv to obtain the input embedding:

[0148]

[0149] where: The latent space representation after the input embedding layer processing, which is the result of the convolutional operation on the original time series data; Embedded input; t-L: t represents from time step t-L to time step t.

[0150] Step 3.2: The encoder consists of multiple layers, and each layer is responsible for extracting the horizontal, growth, and seasonal components of the time series. For each layer n, perform the following operations:

[0151] Extract the seasonal component: Use the frequency attention (FA) module to extract the seasonal component from the input of the encoder

[0152]

[0153] where: is the latent representation output by the encoder in the (n - 1)th layer.

[0154] And remove this part of the information from the original data so that subsequent steps can focus on other features. The formula is as follows:

[0155]

[0156] Wherein: is the final latent representation.

[0157] Perform growth component extraction: Identify the growth component through the Multi-Head Exponential Smoothing Attention (MH-ESA) module

[0158]

[0159] Wherein: represents the growth component extracted from time step t-L to time step t in the nth layer encoder;

[0160] Smoothing parameter α:

[0161]

[0162] L t represents the current moment trend level, Y t represents the observed value.

[0163] Then update the latent representation through layer normalization (LN):

[0164]

[0165] Wherein: represents the latent representation output by the (n-1)th layer encoder.

[0166] Finally, further process the data through a feed-forward network (FF), and the formula is expressed as follows:

[0167]

[0168] Wherein: represents the final output of the nth layer.

[0169] Step 3.3: Generate predictions for the next H steps through the decoder. The prediction steps are as follows:

[0170] First, generate growth and seasonal predictions: The decoder contains N growth and seasonal (G+S) stacks, and a level stack. The G+S stack consists of a growth damping (GD) and an FA module, which respectively utilize and to predict and The formula is as follows:

[0171]

[0172] Wherein: represents the growth component at the current time step t extracted in the n-th layer encoder; TD represents the Trend Damping function; represents the prediction of the growth component from the current time step t to the future time step t + H in the n-th layer decoder; represents the seasonal component from time step t - L to t extracted in the n-th layer encoder; FA represents Frequency Attention; represents the prediction of the seasonal component from the current time step t to the future time step t + H in the n-th layer decoder.

[0173] Trend damping parameter

[0174]

[0175] Secondly, generate the level prediction: The level stack generates the level prediction within the prediction range by repeating the level value of the last time step in the backtracking window:

[0176]

[0177] In the formula: E t:t+H represents the level prediction sequence from the current time step t to the future time step t + H; Repeat H represents the repeat operation, and this function repeats a certain value H times; represents the level component at the current time step t extracted in the N-th layer encoder.

[0178] Finally, it is the final prediction:

[0179]

[0180] In the formula, Y t:t+H is the final prediction result of the foundation pit volume.

[0181] The specific process of combining the K-fold cross-validation method (K-CV) with the biogeography-based optimization algorithm (BBO) to form the K-CV-BBO algorithm is as follows:

[0182] Step 4.1: In the initialization stage, randomly generate a set of solutions (islands) and calculate the fitness of each solution, that is, the quality of the solution. The higher the fitness, the more suitable the island is for species survival and the higher the quality of the solution.

[0183] Step 4.2: Migration operation, simulating the process of species migrating from one island to another. The migration rate depends on the fitness of the island. Islands with high fitness have a high migration out rate and a low migration in rate; conversely, islands with low fitness have a high migration in rate and a low migration out rate. The migration operation is implemented through the following formula:

[0184]

[0185] In the formula, S i is the number of species; S max is the maximum carrying capacity of the species; μ i represents the emigration rate of the i-th living environment; λ i represents the immigration rate of the i-th living environment.

[0186] Step 4.3: Mutation operation, simulating natural disasters or mutation events on the island, resulting in changes in the number of species. The mutation operation is achieved through the following formula:

[0187]

[0188] In the formula, P mut is the probability of mutation occurring on island i; q is the mutation index.

[0189] Step 4.4: Recalculate the fitness of each island and evaluate the quality of the solution.

[0190] Step 4.5: Repeat steps 4.2 - 4 until the preset number of iterations is reached or the fitness no longer changes significantly. Step 4.6: Divide the dataset, randomly divide the entire dataset into K subsets (folds) of equal size.

[0191] Step 4.7: Training and testing loop. For each fold, select one of the subsets as the test set and the remaining K - 1 subsets as the training set. Train the model using the training set and then evaluate the model performance on the test set.

[0192] Step 4.8: Repeat the process, repeat the above steps K times, and each time select a different subset as the test set.

[0193] Step 4.9: Performance evaluation, calculate the average of the K evaluation results as the final evaluation of the model performance.

[0194] Introduce the combination of K - CV and BBO optimization algorithms to optimize the key parameters α of the ETSformer model, perform optimization. Use K - CV to construct the loss function of ETSformer, and use the BBO algorithm to find the global optimal solution of the loss function. The optimization of the loss function by the BBO algorithm under the meaning of K - CV is:

[0195]

[0196] R j = ω1R j,1 + ω2R j,2

[0197] In the formula: R j，1is the mean squared error loss function, R j,2 is the cross - entropy difference loss function, and the optimization loss function of the BBO algorithm in the sense of K - CV is R j , where ω1 and ω2 are weight values determined by the optimization of the BBO algorithm. K is the number of cross - validations; n is the total number of training samples; y i is the original value of the i - th training sample; is the predicted value of the i - th training sample when the parameter is α i 、 .

[0198] Step 4.10: Data normalization processing. Use mapminmax to normalize the input samples to eliminate the influence of different dimensions.

[0199] Step 4.11: Initialize the BBO algorithm. Set the core parameters of the BBO algorithm, including the maximum number of species (S max ), immigration rate (I), emigration rate (E), maximum mutation probability (q max ), etc.

[0200] Step 4.12: Generate the initial solution. Randomly generate N survival environments, that is, N solutions, and each solution represents a set of parameters (α and ) of the ETSformer model.

[0201] Step 4.13: Calculate the fitness. Divide the training samples into K parts. Calculate the loss function. For each set of parameters (α i 、 ), train the ETSformer model on each training set and calculate R j，1 on the corresponding validation set. Finally, calculate the optimization loss function of each parameter under K - fold cross - validation as R j,2 . j .

[0202] Step 4.14: Migration operation. According to the fitness values of the solutions, implement the migration operation to simulate the immigration and emigration of species between different survival environments.

[0203] Step 4.15: Mutation operation. Implement the mutation operation on each solution in the solution set to simulate the impact of unexpected events in the survival environment on the number of species.

[0204] Step 4.16: Iterative update. Repeat steps 4.13 - 15 until the stop condition is met (such as reaching the maximum number of iterations or the change in fitness is less than the preset threshold).

[0205] Step 4.17: Determine the optimal parameters. Select the solution with the highest fitness from the iterative process as the optimal solution, and the corresponding set of parameters (α and )That's what we want.

[0206] Step 5.1: Divide the final dataset into a training set, a validation set, and a test set according to a preset ratio.

[0207] Step 5.2: Use the divided training set to train the improved K-CV-BBO-ETSformer model, enabling the model to learn the patterns and relationships in the data.

[0208] Step 5.3: Use the validation set to validate the model and adjust the model parameters to complete the training.

[0209] Step 5.4: Use the trained model to predict the test set to obtain the prediction results of the foundation pit displacement.

[0210] Step 6.1: Establish an evaluation model for the foundation pit displacement based on expert experience. This model can classify the stability level of the foundation pit according to factors such as soil layer parameters, wall displacement, and excavation depth. For example, the stability of the foundation pit is divided into five levels: safe, relatively safe, general, relatively dangerous, and dangerous.

[0211] Step 6.2: Input the inverted soil layer parameters and the on-site monitored wall horizontal displacement data into the evaluation model to evaluate the stability level of the foundation pit in real time. When the stability level of the foundation pit reaches the relatively dangerous or dangerous level, a warning signal is sent through the central control display platform.

[0212] Step 6.3: While sending the warning signal, use the cameras installed around the foundation pit to capture video images of the foundation pit in real time. Utilize image processing technology to automatically identify and measure the displacement distance of the foundation pit wall to further verify the accuracy of the warning.

[0213] Step 6.4: Compare and analyze the displacement distance obtained from video observation with the warning result of the evaluation model. If the two are consistent, confirm that there are potential safety hazards in the foundation pit and promptly notify the construction personnel to take corresponding emergency measures, such as stopping excavation and strengthening the support structure. If the two are inconsistent, further analyze the reasons, adjust the evaluation model or re-invert the parameters to ensure the accuracy and reliability of the warning.

Claims

1. A method for predicting foundation pit displacement and intelligent early warning, characterized in that It includes the following steps: Step 1: Use multiple multi-parameter sensors to collect key data of the foundation pit in construction engineering. Apply the Maximum Entropy Partitioning algorithm MEP to divide the time series into ε intervals, and use the improved Refined Composite Multivariate Multi-scale Symbolic Dynamic Entropy RCMMSDE feature extraction method to extract data closely related to foundation pit displacement prediction; Step 2: Perform Quantum Derivative Feature Mode Decomposition Q-FMD on the data after feature extraction; Step 3: Establish a foundation pit displacement prediction model based on the ETSformer model; Step 4: Combine the K-fold cross-validation method with the biogeography-based optimization algorithm BBO to form the K-CV-BBO algorithm, and use the combined algorithm to optimize the hyperparameters in the ETSformer model to obtain the optimized K-CV-BBO-ETSformer model. The hyperparameters include: the smoothing parameter α, the trend damping parameter Step 5: Use the optimized K-CV-BBO-ETSformer model to predict the foundation pit displacement and obtain the prediction result of the foundation pit displacement; Step 6: Establish a foundation pit displacement grade evaluation according to expert experience. Finally, display the specific position of the displacement point through a camera video, and complete the early warning through the central control display platform.

2. The method for predicting foundation pit displacement and intelligent early warning according to claim 1, wherein Step 1 includes the following steps: Step 1.1: First, for each time series, use the Maximum Entropy Partitioning algorithm MEP to divide the time series into ε intervals, and map the values of each time series to a symbol; Step 1.2: For each symbolized time series, describe the subsequent states and transition probabilities between symbol sequences through a D-Markov machine; Step 1.3: Sum the state transition matrices of all channels, including soil strength, soil layer distribution, foundation pit plane size, wall horizontal displacement, and wall pressure change, and introduce the channel weight coefficient μ in the fusion process of the multivariate time series state transition matrix k , the state transition matrix Π of the multivariate time series m : where: Π k,m is the m-th element in the k-th sequence transition matrix, D is the depth of the D-Markov machine, and l is the number of channels; Step 1.4: Calculate RCMMSDE according to the state transition matrix Π.

3. The method for predicting foundation pit displacement and intelligent early warning according to claim 1, characterized in that, Step 2 performs Quantum Derivative Feature Mode Decomposition Q-FMD on the preprocessed data. Taking the soil strength P parameter as an example, the specific steps are as follows: Step 2.1: Initialize the parameters and set the soil strength signal sequence as P(n) = P1, P2, ..., P n , where n is the length of the signal. According to the quantum state initialization formula: Among them, |t> is the quantum ground state, representing the quantum state at time point t, P(t) is the amplitude of the original signal at time t, ||P|| is the L2 norm of the signal for normalization, and Φ(t) maps the classical signal to a quantum superposition state to achieve parallel processing; Step 2.2: Quantum feature decomposition formula: Obtain the decomposed modal components according to the above quantum feature decomposition formula; where k = 1, 2, ···, K, the number of filters is K; the fault characteristic frequency is γ, H is the Hamiltonian, which describes the energy characteristics of the system; The number operator of particles, which characterizes the energy of the k-th frequency component, τ is the evolution time, which controls the splitting degree of the quantum state; i is the iteration parameter; Convert the frequency-domain signal to the time-domain signal through quantum measurement and component reconstruction, and reconstruct the IMF component. The formula is as follows: f -1 Inverse Fourier transform, IMF K (t) Multiple measurements of the statistical quantum state, reconstruct the IMF components, Expected value of the quantum measurement for the k-th mode; Step 2.3: The filtering coefficients are reconstructed according to the quantum component IMF K (t) to obtain the modal component and the estimated signal fault period are updated; Step 2.4: Determine whether the current iteration process reaches the maximum number of iterations. If it reaches, return to Step 2.3; otherwise, continue to execute Step 2.5; Step 2.5: Calculate the correlation coefficient between two adjacent modal components, establish a K×K correlation matrix CC, and calculate the correlation kurtosis between the CC and the estimated signal fault period. Select the modal component with a larger correlation kurtosis as the decomposed component; Step 2.6: Determine whether the current number of modes reaches the set number n. If it reaches, the decomposition stops and the result is output; if it does not reach, return to Step 2.2 and continue the above operations; Step 2.7: Finally, output multiple modal components, and each modal component represents the characteristics of different frequency ranges in the soil strength signal.

4. A method for predicting foundation pit displacement and intelligent early warning according to claim 1, characterized in that The hyperparameters in the ETSformer model include: Smoothing parameter α: L t represents the trend level at the current moment, Y t represents the observed quantity, is the seasonal component; Trend damping parameter Represents the growth component at the current time step t extracted in the n-th layer encoder.

5. A method for predicting foundation pit displacement and intelligent early warning according to claim 1, characterized in that, Step 4 includes: Introduce the combination of K-CV and BBO optimization algorithms to optimize the key parameters α of the ETSformer model, Optimize the loss function of ETSformer using K-CV and find the global optimal solution of the loss function using the BBO algorithm; the loss function optimized by the BBO algorithm under the meaning of K-CV is: R j = ω1R j,1 + ω2R j,2 Where: R j，1 is the mean square error loss function, R j,2 is the cross-entropy difference loss function, and the optimization loss function of the BBO algorithm under the meaning of K-CV is R j , ω1 and ω2 are weight values, which are determined by optimizing the BBO algorithm; K is the number of cross-validation times; n is the total number of training samples; y i is the original value of the i-th training sample; is the predicted value of the i-th training sample when the parameters are α i and ; Generate an initial solution, randomly generate N survival environments, that is, N solutions, and each solution represents a set of parameters of the ETSformer model, namely α and Calculate the fitness, divide the training samples into K parts, calculate the loss function for each set of parameters Train the ETSformer model on each training set and calculate R on the corresponding validation set j，1 , R j,2 , and finally calculate the optimized loss function of each parameter under K-fold cross-validation as R j ; Migration operation. According to the fitness value of the solution, implement the migration operation to simulate the immigration and emigration of species between different living environments; Mutation operation. Perform mutation operations on each solution in the solution set to simulate the impact of unexpected events in the living environment on the number of species; Iteratively update until the stopping condition is met; determine the optimal parameters, and select the solution with the highest fitness from the iterative process as the optimal solution, and a set of parameters α of the corresponding ETSformer model and are what we want.

6. The method for predicting foundation pit displacement and intelligent early warning according to claim 1, characterized in that, Step 5 includes the following steps: Step 5.1: Divide the final dataset into a training set, a validation set, and a test set according to a preset ratio; Step 5.2: Use the divided training set to train the improved K-CV-BBO-ETSformer model, enabling the model to learn the patterns and relationships in the data; Step 5.3: Use the validation set to validate the model, adjust the model parameters, and complete the training; Step 5.4: Use the trained model to predict the test set to obtain the prediction results of the foundation pit displacement.

7. A method for predicting foundation pit displacement and intelligent early warning according to claim 1, characterized in that, The said Step 6 includes: Step 6.1: Establish a foundation pit displacement evaluation model based on expert experience. This model classifies the stability level of the foundation pit based on soil layer parameters, wall displacement, and excavation depth factors; Step 6.2: Input the inversed soil layer parameters and the on-site monitored wall horizontal displacement data into the evaluation model to evaluate the stability level of the foundation pit in real time; when the stability level of the foundation pit reaches a relatively dangerous or dangerous level, send a warning signal through the central control display platform; Step 6.3: While sending the warning signal, use the cameras installed around the foundation pit to capture video images of the foundation pit in real time, and use image processing technology to automatically identify and measure the displacement distance of the foundation pit wall; Step 6.4: Compare and analyze the displacement distance obtained from video observation with the warning result of the evaluation model; if the two are consistent, confirm that there are potential safety hazards in the foundation pit and promptly notify the construction personnel to take corresponding emergency measures; if the two are inconsistent, further analyze the reasons, adjust the evaluation model or re-perform parameter inversion to ensure the accuracy and reliability of the warning.

8. A foundation pit displacement prediction and intelligent early warning system, characterized in that, It includes the following modules: Data acquisition module: Collect key data of the foundation pit using various sensors, including the soil strength of the foundation pit, soil layer distribution, foundation pit plane size, wall horizontal displacement, and wall pressure change; Data processing module: Use the improved refined composite multi-scale symbolic dynamic entropy RCMMSDE feature extraction method to extract data closely related to foundation pit displacement prediction and warning; and use quantum-derived feature mode decomposition Q-FMD to decompose the selected data to reduce data complexity; Model prediction module: Select the K-fold cross-validation method combined with the Biogeography-Based Optimization algorithm BBO to optimize the α of the ETSformer model, hyperparameters, improve the model performance, and predict the displacement of the foundation pit; Level evaluation, monitoring, and warning module: Establish an evaluation of the severity level of foundation pit displacement based on expert experience, display the specific position of the displacement point through camera video, and complete the warning through the central control display platform.

9. The foundation pit displacement prediction and intelligent early warning system according to claim 8, characterized in that, For the quantum-derived feature mode decomposition Q-FMD, taking the soil strength P parameter as an example, its specific steps are as follows: Step 2.1: Initialize parameters and set the soil strength signal sequence as P(n) = P1, P2, ..., P n , where n is the length of the signal. According to the quantum state initialization formula: Among them, |t> is the quantum ground state, representing the quantum state at time point t, P(t) is the amplitude of the original signal at time t, ||P|| is the L2 norm of the signal for normalization, and Φ(t) maps the classical signal into a quantum superposition state to achieve parallel processing; Step 2.2: Quantum feature decomposition formula: u QFD = e -iHτ Obtain the decomposed modal components according to the above quantum feature decomposition formula; where k = 1, 2, ···, K; the fault characteristic frequency γ, the Hamiltonian H, which describes the energy characteristics of the system; The number operator of particles, which characterizes the energy of the k-th frequency component, the evolution time τ, which controls the splitting degree of the quantum state; the iteration parameter i; Convert the frequency-domain signal into a time-domain signal through quantum measurement and component reconstruction, and reconstruct the IMF components. The formula is as follows: f -1 Inverse Fourier transform, IMF K (t) Multiple measurements of the statistical quantum state, reconstruct the IMF components, Expected value of the quantum measurement for the k-th mode; Step 2.3: The filtering coefficient is reconstructed according to the quantum component IMF K (t) to obtain the modal component and the estimated signal fault period are updated; Step 2.4: Determine whether the current iteration process has reached the maximum number of iterations. If it has, return to Step 2.3; otherwise, continue to execute Step 2.5; Step 2.5: Calculate the correlation coefficient between two adjacent modal components, establish a K×K correlation matrix CC, and calculate the correlation kurtosis between the CC and the estimated signal fault period. Select the modal component with a larger correlation kurtosis as the decomposed component; Step 2.6: Determine whether the current number of modes reaches the set number n. If it reaches, the decomposition stops and the result is output; if not, return to Step 2.2 and continue the above operations; Step 2.7: Finally, multiple modal components are output, and each modal component represents the characteristics of different frequency ranges in the soil strength signal.

10. A foundation pit displacement prediction and intelligent early warning system according to claim 8, characterized in that, Select the K-fold cross-validation method combined with the Biogeography-Based Optimization algorithm BBO to optimize the α of the ETSformer model, The specific process of the hyperparameters is as follows: Introduce the combination of K-CV and BBO optimization algorithms to optimize the key parameters α of the ETSformer model, Perform optimization. Use K-CV to construct the loss function of ETSformer, and use the BBO algorithm to find the global optimal solution of the loss function; the loss function optimized by the BBO algorithm under the meaning of K-CV is: R j = ω1R j,1 + ω2R j,2 Where: R j，1 is the mean square error loss function, and R j,2 is the cross entropy difference loss function. The optimization loss function of the BBO algorithm under the K-CV sense is R j , where ω1 and ω2 are weight values determined by optimizing the BBO algorithm; K is the number of cross-validation times; n is the total number of training samples; y i is the original value of the i-th training sample; is the predicted value of the i-th training sample when the parameters are α i and . Generate an initial solution, randomly generate N survival environments, that is, N solutions, and each solution represents a set of parameters of the ETSformer model, namely α and Calculate the fitness, divide the training samples into K parts, calculate the loss function, and for each set of parameters Train the ETSformer model on each training set and calculate R on the corresponding validation set j，1 , R j,2 , and finally calculate the optimized loss function of each parameter under K-fold cross-validation as R j ; Migration operation: According to the fitness value of the solution, implement the migration operation to simulate the immigration and emigration of species between different living environments; Mutation operation: Implement the mutation operation on each solution in the solution set to simulate the impact of unexpected events in the living environment on the number of species; Iteratively update until the stopping condition is met; determine the optimal parameters, and select the solution with the highest fitness from the iterative process as the optimal solution. The corresponding set of parameters α of the ETSformer model and are what we want.

Images