Early warning method and system based on foundation pit engineering ensemble learning model

Abstract

The invention provides an early warning method and system based on a foundation pit engineering ensemble learning model. The method comprises the following steps: acquiring standard monitoring data, a trend component and a fluctuation component; determining a first time window, over-threshold and low-threshold time lags, a first stable stage and a new change typical interval, and further determining a training window and a prediction window; obtaining training and prediction standard monitoring data, training the base learner model and performing prediction, and obtaining a prediction result; and obtaining a corrected model prediction value based on prediction standard monitoring data, and comparing the corrected model prediction value with an actual observation value to obtain a historical prediction error: if the historical prediction error exceeds a threshold value, adjusting the two windows until the error is lower than the threshold value, obtaining an integrated learning model, and performing settlement deformation prediction on the foundation pit to obtain a prediction result. According to the early warning method and system based on the foundation pit engineering ensemble learning model provided by the invention, the model can be updated in real time, the prediction precision is improved, the early warning reliability is enhanced, the deformation trend is accurately grasped, and the construction and surrounding safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of underground engineering technology, and in particular to an early warning method and system based on an integrated learning model for foundation pit engineering. Background Art

[0002] With the rapid development of urbanization and the continued expansion of urban areas, underground space development has become a critical component of urban development. As rail transit construction continues to expand across cities, monitoring excavation in densely populated urban areas is crucial for ensuring the safety of surrounding areas. Traditionally, excavation monitoring data is used only for comparison with control values. Exceeding control limits triggers alarms to prevent major engineering accidents. The rapid development of machine learning and deep learning models has prompted the exploration of broader uses for monitoring data, such as using monitoring data to train models for early warning and prediction, spatial relationship reasoning, and comprehensive risk assessment. However, excavation monitoring data is essentially streaming data, regularly updated as construction progresses. As construction progresses, the spatiotemporal patterns of excavation deformation are likely to change, causing models trained on earlier data to lose accuracy over time. The task of integrating newly received data to update machine learning models is defined as "streaming machine learning." Currently, there is a lack of relevant research and technology for streaming machine learning in excavation engineering, making it difficult to ensure the long-term performance and stability of models. Several common streaming machine learning methods are already in use in other fields, including online learning, sliding window learning, and adaptive learning. Online learning is a method for incrementally updating models. Whenever new data arrives, the model is immediately updated. Its advantages include fast updates without the need to retrain the entire model. However, it is sensitive to noise and is generally suitable for simple linear models, but struggles with complex models. The sliding window method uses a fixed-size window and only utilizes the latest data for training. This effectively avoids the influence of old data when the data volume is large, but selecting the window size is difficult and may overlook long-term trends. Adaptive learning dynamically adjusts the model to adapt to changes in the data distribution, making it particularly suitable for cases of concept drift. Although it can effectively address data evolution, its implementation is complex and resource-intensive.

[0003] In the current state of the art, the lack of an update mechanism for machine learning models in foundation pit projects results in them typically being trained only once using early monitoring data. This lacks an effective update mechanism, and as construction progresses, the model's prediction accuracy declines due to the deformation of the foundation pit and the constant changes in its mechanical response characteristics. This makes it unable to accurately reflect the actual working conditions in the later stages of construction, seriously impacting the reliability of project safety warnings. Furthermore, general streaming learning methods are not well-suited to the characteristics of foundation pit engineering: online learning methods, while offering fast update speeds, are primarily applicable to simple linear models and struggle to handle the complex nonlinear relationships found in foundation pit engineering. Sliding window methods, while able to mitigate the influence of old data, may miss important long-term deformation patterns. Adaptive learning methods, while capable of addressing concept drift, have complex update rules that are incompatible with the relatively small amount of monitoring data and can easily lead to unstable model performance. Furthermore, data analysis and model updating are not effectively integrated. Existing technologies fail to fully guide model updates based on the results of foundation pit mechanical response pattern analysis. There is a lack of mechanisms to combine construction phase feature recognition with model update strategies. A model update mechanism based on data analysis results has not yet been established, and there is also a lack of real-time evaluation methods for model performance changes during the update process. Under the influence of these comprehensive factors, foundation pit projects may experience reduced reliability of safety warnings and an inability to accurately grasp the deformation trend of the foundation pit, which in turn seriously affects the construction safety of the foundation pit project and the safety of the surrounding environment. Summary of the Invention

[0004] The present invention aims to provide an early warning method and system based on an integrated learning model for foundation pit engineering to solve the above-mentioned technical problems, avoid the problems of decreased prediction accuracy, mismatch of model update methods, and disconnection of machine learning model updates. By constructing a dynamic update mechanism, real-time model updates can be achieved, prediction accuracy can be improved, and the reliability of early warning can be enhanced, thereby ensuring the safety of construction and surrounding areas.

[0005] In order to solve the above technical problems, the present invention provides an early warning method based on an integrated learning model for foundation pit engineering, comprising the following steps:

[0006] Based on the pre-acquired initial monitoring data, standard monitoring data is obtained, and then the trend component and the fluctuation component are obtained based on the standard monitoring data;

[0007] determining a first time window based on the trend component;

[0008] Based on the fluctuation component and the preset autocorrelation threshold, the above-threshold time lag and the below-threshold time lag are obtained;

[0009] Based on the fluctuation components, standard monitoring data and preset sensitivity thresholds, the first stable stage and the typical interval of new changes are obtained;

[0010] Determining a training window and a prediction window based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable stage, and the typical interval of new changes;

[0011] Acquire training standard monitoring data based on the training window, and acquire prediction standard monitoring data based on the prediction window;

[0012] Train the base learner model based on the training standard monitoring data and make predictions to obtain the prediction results;

[0013] Based on the prediction results and the prediction standard monitoring data, the revised model prediction value is obtained;

[0014] Obtain historical prediction errors based on the corrected model prediction value and the actual observation value: If the historical prediction error exceeds the preset prediction error threshold, adjust the training window and the prediction window based on the preset training window adjustment amount and the preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, obtain the latest training window and the latest prediction window, and then obtain the integrated learning model;

[0015] The settlement and deformation of the foundation pit are predicted based on the integrated learning model, and the settlement and deformation prediction results of the foundation pit in different future construction stages are obtained.

[0016] In this approach, model prediction is performed by combining the prediction results of the base learner model with standard monitoring data, further improving the model's robustness and obtaining more accurate revised model predictions to adapt to changes in working conditions during foundation pit construction. The revised model predictions are then compared with actual observations, and historical prediction errors are calculated. By analyzing these historical prediction errors, periods of unstable performance of the foundation pit engineering ensemble learning model can be identified, and the updated parameters of the foundation pit engineering ensemble learning model can be corrected, providing a basis for revising the foundation pit engineering ensemble learning model. After training and predicting the revised foundation pit engineering ensemble learning model, its prediction performance is evaluated. When the historical prediction error exceeds the preset prediction error threshold, indicating poor prediction performance, the training and prediction windows need to be adjusted based on feedback from the historical prediction error. This allows the foundation pit engineering ensemble learning model to better adapt to data changes caused by changes in construction conditions, improve model stability and prediction accuracy, and ensure consistently high prediction accuracy during foundation pit construction. Until the historical prediction error is lower than the preset prediction error threshold, the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained; based on the finally updated integrated learning model, the settlement and deformation of the foundation pit are predicted, and the settlement and deformation prediction results of the foundation pit in different future construction stages are obtained. This can avoid the problems of decreased prediction accuracy, mismatch of model update methods and disconnection of machine learning model updates. By building a dynamic update mechanism, real-time model updates, improved prediction accuracy, and enhanced early warning reliability can be achieved, thereby ensuring the safety of construction and surrounding areas.

[0017] Furthermore, the method of obtaining standard monitoring data based on the pre-acquired initial monitoring data, and then obtaining trend components and fluctuation components based on the standard monitoring data, is specifically as follows:

[0018] Standardize the pre-acquired initial monitoring data to obtain standard monitoring data:

[0019]

[0020] Among them, X t represents the standard monitoring data at time point t, X t ′ represents the initial monitoring data at the pre-acquired time point t, μ and σ represent the mean and standard deviation of the standard monitoring data;

[0021] Based on standard monitoring data, the trend component and the volatility component are separated and obtained through filters:

[0022] X t =T t +C t

[0023] Among them, T t represents the trend component, C t Represents the fluctuation component.

[0024] In the above scheme, the pre-acquired initial monitoring data is standardized to eliminate the effects of data scale. During the excavation construction process, standard monitoring data will exhibit certain temporal characteristics, such as seasonal variations, long-term trends, and cyclical fluctuations. These variations can affect the determination of the training and prediction windows. Therefore, by calculating the trend and fluctuation components of the standard monitoring data, it is possible to determine whether the data has a clear long-term trend. If the trend component shows a clear linear or nonlinear trend, the training window should cover the evolution of this trend.

[0025] Furthermore, the determining of the first time window based on the trend component is specifically as follows:

[0026] Estimate the correlation of the trend component in the time series and select the time span that makes the trend component change relatively stably as the first time window:

[0027]

[0028] Among them, h trend represents the first time window, N represents the standard monitoring data sequence X={X1,X2,…,X n}, where the standard monitoring data sequence X is composed of standard monitoring data at n time points.

[0029] In the above scheme, the time span of relatively stable changes in the trend component is selected as the first time window, which is used as the candidate value of the training window.

[0030] Furthermore, the above-threshold time lag and the low-threshold time lag are obtained based on the fluctuation component and the preset autocorrelation threshold, specifically:

[0031] The autocorrelation function is constructed based on the fluctuation components:

[0032]

[0033] Among them, ρ k Represents the constructed autocorrelation function, C={C1,C2,…,C n} represents the standard monitoring data sequence X={X1,X2,…,X n}, represents the mean of the volatility component, k represents the time lag variable;

[0034] Based on the preset autocorrelation threshold, the autocorrelation function is constructed to obtain the above-threshold time lag and the below-threshold time lag:

[0035]

[0036] in, represents the time lag after exceeding the threshold value, represents the low threshold time hysteresis, {τ1|ρ(τ1)>ρ threshold} means all the values ​​satisfying ρ(τ1)>ρ threshold The set of time delay τ1, {τ2|ρ(τ2)<ρ threshold} means all the values ​​satisfying ρ(τ2)<ρ threshold The set of time delays τ2, ρ threshold Indicates the preset autocorrelation threshold.

[0037] In the above scheme, since different stages of foundation pit construction may show autocorrelation, especially when the construction technology or environmental conditions are similar, an autocorrelation function is constructed to analyze the autocorrelation of the fluctuation component of the standard monitoring data. In the constructed autocorrelation function, ρ k Reflects the degree of autocorrelation of the volatility component when the time lag variable is k. ρ k The larger it is, the stronger the correlation between the fluctuation components separated by k time intervals is. The mean of the fluctuation components is used to calculate the deviation of the fluctuation components from the mean to measure the degree of fluctuation. Next, the maximum time lag for which the autocorrelation function value exceeds the preset autocorrelation threshold, i.e., the over-threshold time lag, is found. In foundation pit engineering, this can determine the longest time interval during which the strong correlation of the fluctuation components persists. The minimum time lag for which the autocorrelation function value falls below the preset autocorrelation threshold, i.e., the under-threshold time lag, can be found. This can determine the time point when the correlation of the fluctuation components decreases to a certain level. This is important for analyzing fluctuation patterns such as foundation pit deformation and for selecting subsequent training and prediction windows.

[0038] Furthermore, the first stable phase and the typical interval of new changes are obtained based on the fluctuation component, standard monitoring data and preset sensitivity threshold, specifically:

[0039] Based on the fluctuation components, standard monitoring data and preset sensitivity thresholds, the change time point is detected and the change time point is detected by the CUSUM algorithm:

[0040] S t′ =max(0,S t′-1 +(C t′ -μ-m))

[0041] Among them, t ′ is the time point of change, S t′ represents the statistic at the changing time point t′, C t′ represents the standard monitoring data fluctuation component at the time point of the change, and m represents the preset sensitivity threshold;

[0042] There are multiple t', which are respectively represented as the change time point t' j ;

[0043] At multiple changing time points t′ j Select the typical interval between the first stable phase and new changes:

[0044]

[0045] in, Indicates the first stable stage, Indicates the typical interval of new changes.

[0046] In the above scheme, the change time point will cause a significant change in the mechanical response pattern of the foundation pit, which will affect the selection of the training window and prediction window. In foundation pit engineering, a longer stable phase means that the construction conditions are relatively stable and the mechanical response pattern changes less. By determining the first stable phase, we can understand the relatively stable time period of the foundation pit during the construction process, providing a reference for stable data segments for the training window; selecting the median of the new typical interval of change can avoid the impact of individual abnormally large or small time intervals on the overall situation, and can better represent the typical time interval between change time points, which helps to determine the reasonable time interval for model updates, timely capture changes in the construction process, and improve the accuracy of model predictions.

[0047] Furthermore, the training window and the prediction window are determined based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable stage and the typical interval of new changes, specifically:

[0048]

[0049] Among them, h1 represents the training window and h2 represents the prediction window.

[0050] In the above scheme, the first time window, the super-threshold time lag and the maximum value in the first stable stage are selected as the training window. The training window determined in this way integrates multiple factors such as trend changes, fluctuation correlation and the stable stage of construction. The training window contains sufficiently long stable stage data to ensure the learning effect of the model under stable working conditions. Taking the maximum value of the three can make the training window contain as comprehensive and critical information as possible, provide sufficient data support for model training, and improve the accuracy and stability of the integrated learning model of foundation pit engineering. By selecting the minimum value in the low threshold time lag and the typical interval of new changes as the prediction window, the prediction window determined in this way ensures the reliability of the prediction, avoids the prediction being too sensitive and unstable due to the window being too small, and can find a balance between timely capturing changes and ensuring prediction accuracy, so that the prediction results of the model are more in line with the actual situation of the foundation pit engineering.

[0051] Furthermore, the base learner model is trained based on the training standard monitoring data and prediction is performed to obtain the prediction result, specifically:

[0052] Based on the training standard monitoring data, the base learner model is trained using machine learning methods:

[0053]

[0054] Among them, f i (t) represents the trained base learner model, h1′ represents the time point corresponding to the training standard monitoring data obtained based on the training window h1, represents the training standard monitoring data between the time period t-h1′ to t, represents the actual foundation pit settlement between the time period t-h1′ and t, and there are M base learner models in total;

[0055] Make predictions based on the trained base learner model and obtain prediction results.

[0056] In the above scheme, training the base learner model can improve the accuracy and stability of prediction; It contains the standard monitoring data of the foundation pit during this period. By analyzing these standard monitoring data, the base learner model attempts to find the correlation with the actual foundation pit settlement. is the actual foundation pit settlement between the time period t-h1′ and t, and is the target output of the base learner model. The base learner model continuously adjusts its parameters to ensure that the predicted results are as close to the actual settlement as possible, thereby establishing a mapping relationship between standard monitoring data and actual settlement. The previously trained base learner model is then used to predict foundation pit settlement, generating actual predicted outputs that provide a basis for risk assessment and decision-making in foundation pit projects.

[0057] Furthermore, the modified model prediction value is obtained based on the prediction results and the prediction standard monitoring data, specifically:

[0058] Perform weighted averaging on the prediction results to obtain the integrated prediction results:

[0059]

[0060] in, represents the integrated prediction result, w i represents the weight of the i-th base learner model, Represents the prediction result of the i-th base learner model;

[0061] Based on the integrated prediction results and sliding window technology, the foundation pit engineering integrated learning model is modified to obtain the predicted value of the modified model:

[0062]

[0063] in, represents the predicted value of the modified model, and h2′ represents the corresponding time point of the predicted standard monitoring data obtained based on the prediction window h2.

[0064] In the above scheme, the prediction results of multiple base learner models are weighted averaged, and the weights are adjusted based on the performance of the models on the training set. To avoid overfitting, the weights can be dynamically optimized based on the historical performance of each base learner model. However, for foundation pit projects, since the construction conditions will change significantly as the construction progresses, the performance of newer base models is usually better than that of older base models. Further adaptive correction is performed, and the prediction results of the base learner model are averaged using sliding window technology and prediction standard monitoring data to obtain the corrected model prediction value The prediction results were further optimized and the robustness of the model was improved, enabling it to better adapt to the ever-changing working conditions during foundation pit construction.

[0065] Furthermore, the historical prediction error is obtained based on the corrected model prediction value and the actual observation value: if the historical prediction error exceeds the preset prediction error threshold, the training window and the prediction window are adjusted based on the preset training window adjustment amount and the preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, and the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained, specifically:

[0066] Based on the corrected model prediction value and the actual observation value, the historical prediction error is obtained, and then the mean and standard deviation of the historical prediction error are obtained:

[0067]

[0068] Among them, ∈(t) represents the historical forecast error, and y(t) represents the actual observation value;

[0069] Preset prediction error threshold:

[0070] ∈ threshold =μ ε +γσ ε

[0071] Among them, ∈ threshold represents the prediction error threshold, μ ε and σ ε represents the mean and standard deviation of historical forecast errors, and γ represents the empirical parameter;

[0072] If the historical prediction error exceeds the preset prediction error threshold, the training window and prediction window are adjusted based on the preset training window adjustment amount and the preset prediction window adjustment amount:

[0073] h1(t+1)=h1(t)+Δh1if∈(t)>∈ threshold

[0074] h2(t+1)=h2(t)+Δh2if∈(t)>∈threshold

[0075] Wherein, h1(t+1) represents the latest training window, h2(t+1) represents the latest prediction window, Δh1 represents the preset training window adjustment amount, and Δh2 represents the preset prediction window adjustment amount;

[0076] Δh1 and Δh2 are set to 3-5 days or dynamically adjusted using a data-driven approach:

[0077] Δh 1,2 =α·(∈(t)-∈ threshold )

[0078] Among them, α represents the learning rate;

[0079] Until the historical prediction error is lower than the preset prediction error threshold, the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained.

[0080] In the above scheme, by obtaining historical prediction errors, the degree of deviation between the corrected model predictions and the actual observations can be intuitively measured, reflecting the model's prediction accuracy at each moment. A prediction error threshold is then set to determine whether the model's prediction error is acceptable. This prediction error threshold provides an objective basis for determining whether model performance meets requirements. When the historical prediction error exceeds this threshold, it indicates that the model's prediction performance is poor and requires adjustment. A dynamic adjustment mechanism for the model is then established. When the historical prediction error exceeds a preset threshold, the training window and prediction window are adjusted based on preset training window adjustment amounts and preset prediction window adjustment amounts. This adjustment enables the model to better adapt to changes during foundation pit construction, reoptimizing the model's training and prediction conditions to improve its prediction accuracy. This dynamic adjustment method adaptively adjusts the training window and prediction window based on the size of the model's historical prediction error, enabling the model to more quickly adapt to data changes. Finally, by continuously adjusting the training window and prediction window, the historical prediction error is reduced below the preset prediction error threshold, resulting in the latest training window and prediction window, and ultimately determining the final ensemble learning model. It can achieve real-time model updates, improve prediction accuracy, maintain high prediction accuracy, and enhance the reliability of early warning, thereby ensuring the safety of construction and surrounding areas, and providing reliable support for safety monitoring and early warning of foundation pit projects.

[0081] The present invention provides an early warning system based on an integrated learning model for foundation pit engineering, comprising a data extraction module, a first time window determination module, a time lag calculation module, a time interval acquisition module, a window determination module, a data screening module, a model training module, a correction module, an adjustment module, and a foundation pit early warning module, wherein:

[0082] The data extraction module is used to obtain standard monitoring data based on pre-acquired initial monitoring data, and then obtain trend components and fluctuation components based on the standard monitoring data;

[0083] The first time window determination module is configured to determine the first time window based on the trend component;

[0084] The time lag calculation module is used to obtain an over-threshold time lag and a low-threshold time lag based on the fluctuation component and a preset autocorrelation threshold;

[0085] The time interval acquisition module is used to obtain the first stable stage and the typical interval of the new change based on the fluctuation component, the standard monitoring data and the preset sensitivity threshold;

[0086] The window determination module is configured to determine a training window and a prediction window based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable phase, and the typical interval of new changes;

[0087] The data screening module is used to obtain training standard monitoring data based on the training window and obtain prediction standard monitoring data based on the prediction window;

[0088] The model training module is used to train the base learner model based on the training standard monitoring data and perform prediction to obtain the prediction results;

[0089] The correction module is used to obtain a corrected model prediction value based on the prediction result and the prediction standard monitoring data;

[0090] The adjustment module is configured to obtain a historical prediction error based on the corrected model prediction value and the actual observation value: if the historical prediction error exceeds a preset prediction error threshold, the training window and the prediction window are adjusted based on a preset training window adjustment amount and a preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, and the latest training window and the latest prediction window are obtained, thereby obtaining an integrated learning model;

[0091] The foundation pit early warning module is used to predict the settlement and deformation of the foundation pit based on the integrated learning model, obtain the settlement and deformation prediction results of the foundation pit at different future construction stages, and realize early warning of the foundation pit.

[0092] The proposed early warning system based on the foundation pit engineering ensemble learning model has a simple structure. In practical applications, it only requires a model training module to combine the prediction results of the base learner model with standard monitoring data for model prediction. This further improves the model's robustness and obtains more accurate revised model predictions to adapt to changing working conditions during foundation pit construction. An adjustment module then compares the revised model predictions with actual observations and calculates historical prediction errors. This analysis of historical prediction errors identifies periods of unstable performance of the foundation pit engineering ensemble learning model and allows for parameter corrections to be made to the updated model, providing a basis for revising the model. After training and predicting the revised foundation pit engineering ensemble learning model, the model's prediction performance is evaluated. When the historical prediction error exceeds a preset prediction error threshold, indicating poor prediction performance, the training and prediction windows need to be adjusted based on feedback from the historical prediction errors. This allows the model to better adapt to data changes caused by changing construction conditions, improving model stability and prediction accuracy, and ensuring consistently high prediction accuracy during foundation pit construction. Until the historical prediction error is lower than the preset prediction error threshold, the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained; the foundation pit early warning module is used to predict the settlement and deformation of the foundation pit based on the finally updated integrated learning model, and the settlement and deformation prediction results of the foundation pit in different future construction stages are obtained. This can avoid the problems of decreased prediction accuracy, mismatch of model update methods and disconnection of machine learning model updates. By building a dynamic update mechanism, real-time model updates, improved prediction accuracy, and enhanced early warning reliability can be achieved, thereby ensuring the safety of construction and surrounding areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A schematic flow chart of an early warning method based on an integrated learning model for foundation pit engineering provided by one embodiment of the present invention;

[0094] Figure 2 A schematic diagram of the architecture of an early warning system based on an integrated learning model for foundation pit engineering provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0095] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0096] This embodiment provides an early warning method based on an integrated learning model for foundation pit engineering. For specific steps, please refer to Figure 1 ,include:

[0097] Based on the pre-acquired initial monitoring data, standard monitoring data is obtained, and then the trend component and the fluctuation component are obtained based on the standard monitoring data;

[0098] determining a first time window based on the trend component;

[0099] Based on the fluctuation component and the preset autocorrelation threshold, the above-threshold time lag and the below-threshold time lag are obtained;

[0100] Based on the fluctuation components, standard monitoring data and preset sensitivity thresholds, the first stable stage and the typical interval of new changes are obtained;

[0101] Determining a training window and a prediction window based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable stage, and the typical interval of new changes;

[0102] Acquire training standard monitoring data based on the training window, and acquire prediction standard monitoring data based on the prediction window;

[0103] Train the base learner model based on the training standard monitoring data and make predictions to obtain the prediction results;

[0104] Based on the prediction results and the prediction standard monitoring data, the revised model prediction value is obtained;

[0105] Obtain historical prediction errors based on the corrected model prediction value and the actual observation value: If the historical prediction error exceeds the preset prediction error threshold, adjust the training window and the prediction window based on the preset training window adjustment amount and the preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, obtain the latest training window and the latest prediction window, and then obtain the integrated learning model;

[0106] The settlement and deformation of the foundation pit are predicted based on the integrated learning model, and the settlement and deformation prediction results of the foundation pit in different future construction stages are obtained.

[0107] In this embodiment, during the foundation pit construction process, mechanical response characteristics such as foundation deformation and foundation pit settlement will change as the construction progresses, and the mechanical response pattern of the foundation pit is usually complex in time and space. In order to enable the foundation pit engineering integrated learning model to adapt to these changes, it is necessary to use data analysis methods to identify mechanical response patterns and recommend appropriate training windows and prediction windows based on these patterns. Model prediction is performed by combining the prediction results of the base learner model and the prediction standard monitoring data to further improve the robustness of the model and obtain more accurate corrected model prediction values ​​to adapt to changes in working conditions during foundation pit construction. The corrected model prediction value is then compared with the actual observation value, and the historical prediction error is calculated. By analyzing the historical prediction error, the time period when the performance of the foundation pit engineering integrated learning model is unstable can be identified, and the update parameters of the foundation pit engineering integrated learning model can be corrected, providing a basis for correcting the foundation pit engineering integrated learning model. After training and predicting the revised ensemble learning model for foundation pit engineering, the model's prediction performance is evaluated. When the historical prediction error exceeds the preset prediction error threshold, it indicates poor prediction performance. The training and prediction windows need to be adjusted based on feedback from the historical prediction error. This allows the ensemble learning model to better adapt to data changes caused by changes in construction conditions, improve the model's stability and prediction accuracy, and ensure consistently high prediction accuracy in foundation pit engineering. Until the historical prediction error falls below the preset prediction error threshold, the latest training and prediction windows are obtained, and then the ensemble learning model is obtained. Based on the final updated ensemble learning model, settlement and deformation predictions are performed on the foundation pit, resulting in predicted settlement and deformation results for different future construction phases. This avoids issues such as decreased prediction accuracy, mismatched model update methods, and disconnected machine learning model updates. By establishing a dynamic update mechanism, the model can be updated in real time, prediction accuracy can be improved, and the reliability of early warnings can be enhanced, thereby ensuring construction and surrounding safety.

[0108] Furthermore, the method of obtaining standard monitoring data based on the pre-acquired initial monitoring data, and then obtaining trend components and fluctuation components based on the standard monitoring data, is specifically as follows:

[0109] Standardize the pre-acquired initial monitoring data to obtain standard monitoring data:

[0110]

[0111] Among them, X t represents the standard monitoring data at time point t, X t ′ represents the initial monitoring data at the pre-acquired time point t, μ and σ represent the mean and standard deviation of the standard monitoring data;

[0112] Based on standard monitoring data, the trend component and the volatility component are separated and obtained through filters:

[0113] X t =T t +C t

[0114] Among them, T t represents the trend component, C t Represents the fluctuation component.

[0115] In this embodiment, the frequency of collecting the initial monitoring data of the foundation pit is usually once a day, so the frequency of model updating is also set to once a day. First, the necessary preprocessing is performed on the pre-acquired initial monitoring data, including denoising, missing value filling, etc. The initial monitoring data of the foundation pit usually includes time series data, for example, settlement values ​​and displacements measured every day. These data need to be standardized or normalized to eliminate the influence of the data scale, so as to obtain standard monitoring data. During the project process, the monitoring data will show certain time series characteristics, with seasonal changes, long-term trends and periodic fluctuations. These changes may affect the selection of training windows and prediction windows. The Hodrick-Prescott filter is used to separate the trend component and the fluctuation component of the standard monitoring data. By calculating the trend component of the standard monitoring data (such as linear trend, quadratic trend, etc.), it can be determined whether the standard monitoring data has an obvious long-term trend. If the mechanical response of the foundation pit shows an obvious linear or nonlinear trend, the training window should cover the evolution process of this trend.

[0116] Furthermore, the determining of the first time window based on the trend component is specifically as follows:

[0117] Estimate the correlation of the trend component in the time series and select the time span that makes the trend component change relatively stably as the first time window:

[0118]

[0119] Among them, h trend represents the first time window, N represents the standard monitoring data sequence X={X1,X2,…,X n}, where the standard monitoring data sequence X is composed of standard monitoring data at n time points.

[0120] In this embodiment, the correlation of the trend component in the time series calculated using the Hodrick-Prescott filter can be used to estimate the presence of significant periodicity in the foundation pit mechanical response. The time span that minimizes the change in the trend component is selected as the first time window and used as a candidate value for the training window.

[0121] Furthermore, the above-threshold time lag and the low-threshold time lag are obtained based on the fluctuation component and the preset autocorrelation threshold, specifically:

[0122] The autocorrelation function is constructed based on the fluctuation components:

[0123]

[0124] Among them, ρ k Represents the constructed autocorrelation function, C={C1,C2,…,C n} represents the standard monitoring data sequence X={X1,X2,…,X n}, represents the mean of the volatility component, k represents the time lag variable;

[0125] Based on the preset autocorrelation threshold, the autocorrelation function is constructed to obtain the above-threshold time lag and the below-threshold time lag:

[0126]

[0127] in, represents the time lag after exceeding the threshold value, represents the low threshold time hysteresis, {τ1|ρ(τ1)>ρ threshold} means all the values ​​satisfying ρ(τ1)>ρ threshold The set of time delay τ1, {τ2|ρ(τ2)<ρ threshold} means all the values ​​satisfying ρ(τ2)<ρ threshold The set of time delays τ2, ρ threshold Indicates the preset autocorrelation threshold.

[0128] In this embodiment, since different stages of foundation pit construction may show autocorrelation, especially when the construction technology or environmental conditions are similar, an autocorrelation function is constructed to analyze the autocorrelation of the fluctuation component of the standard monitoring data. In the constructed autocorrelation function, ρ k It reflects the degree of autocorrelation of the fluctuation component when the time lag variable is k. For the standard monitoring data sequence X={X1,X2,…,X n}, focusing on analyzing its fluctuation components C={C1,C2,…,C n}, and ρ k The larger it is, the stronger the correlation between the fluctuation components separated by k time intervals is. The mean of the fluctuation component is used to calculate the deviation of the fluctuation component from the mean to measure the degree of fluctuation. By autocorrelation analysis, the maximum time lag that makes the autocorrelation function value greater than the preset autocorrelation threshold is found, that is, the super-threshold time lag. In foundation pit engineering, the longest time interval of strong correlation of the fluctuation component can be determined. The preset autocorrelation threshold ρ thresholdTypically, it is set to 0.2 or 0.3. Finding the minimum time lag that makes the autocorrelation function value less than the preset autocorrelation threshold, that is, the low threshold time lag, can determine the time point when the correlation of the fluctuation components decreases to a certain level. This is of great significance for analyzing the fluctuation patterns of foundation pit deformation and the selection of subsequent training windows and prediction windows.

[0129] Furthermore, the first stable phase and the typical interval of new changes are obtained based on the fluctuation component, standard monitoring data and preset sensitivity threshold, specifically:

[0130] Based on the fluctuation components, standard monitoring data and preset sensitivity thresholds, the change time point is detected and the change time point is detected by the CUSUM algorithm:

[0131] S t′ =max(0,S t′-1 +(C t′ -μ-m))

[0132] Among them, t′ is the time point of change, S t′ represents the statistic at the changing time point t′, C t′ represents the standard monitoring data fluctuation component at the time point of the change, and m represents the preset sensitivity threshold;

[0133] There are multiple t', which are respectively represented as the change time point t' j ;

[0134] At multiple changing time points t′ j Select the typical interval between the first stable phase and new changes:

[0135]

[0136] in, Indicates the first stable stage, Indicates the typical interval of new changes.

[0137] In this embodiment, the change time point detection is used to identify phase changes in the project process, such as process changes, geological condition changes, etc. The change time point will cause a significant change in the foundation pit mechanical response pattern, which will affect the selection of training windows and prediction windows. The CUSUM method can detect the time points with large changes in the sequence. The preset sensitivity threshold is used to control the sensitivity of the detection. When S t′ When the data exceeds a predetermined threshold, it indicates that a significant change has occurred, which is the change time point. The position of the change time point t' can be used to divide the different stages of foundation pit construction. jThe longest stable phase is selected as the first stable phase and the typical interval of new changes during construction is selected as the typical interval of new changes. In foundation pit engineering, a longer stable phase means that the construction conditions are relatively stable and the mechanical response pattern changes little. By determining the first stable phase, we can understand the relatively stable time period of the foundation pit during the construction process, providing a reference for stable data segments for the training window; median() represents the median between the time points of change. Selecting the median of the typical interval of new changes can avoid the impact of individual abnormally large or small time intervals on the overall situation, and can better represent the typical time interval between the time points of change. It helps to determine the reasonable time interval for model updates, timely capture changes in the construction process, and improve the accuracy of model predictions.

[0138] Furthermore, the training window and the prediction window are determined based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable stage and the typical interval of new changes, specifically:

[0139]

[0140] Among them, h1 represents the training window and h2 represents the prediction window.

[0141] In this embodiment, the maximum value in the first time window, the super-threshold time lag and the first stable stage is selected as the training window. The training window means using the training standard monitoring data when training the machine learning model, and the prediction window means using the model of the prediction standard monitoring data for prediction during ensemble learning. The training window determined in this way integrates multiple factors such as trend changes, fluctuation correlation and the stable stage of construction. The training window contains data from the stable stage for a sufficiently long time to ensure the learning effect of the model under stable working conditions. Taking the maximum value of the three can make the training window contain as comprehensive and critical information as possible, provide sufficient data support for model training, and improve the accuracy and stability of the ensemble learning model of the foundation pit project. By selecting the minimum value in the low threshold time lag and the typical interval of new changes as the prediction window, the prediction window determined in this way ensures the reliability of the prediction, avoids the prediction being too sensitive and unstable due to the window being too small, and can find a balance between timely capturing changes and ensuring prediction accuracy, so that the prediction results of the model are more in line with the actual situation of the foundation pit project.

[0142] Furthermore, the base learner model is trained based on the training standard monitoring data and prediction is performed to obtain the prediction result, specifically:

[0143] Based on the training standard monitoring data, the base learner model is trained using machine learning methods:

[0144]

[0145] Among them, fi (t) represents the trained base learner model, h1′ represents the time point corresponding to the training standard monitoring data obtained based on the training window h1, represents the training standard monitoring data between the time period t-h1′ to t, represents the actual foundation pit settlement between the time period t-h1′ and t, and there are M base learner models in total;

[0146] Make predictions based on the trained base learner model and obtain prediction results.

[0147] In this embodiment, in order to avoid the shortcomings of traditional streaming learning methods (such as the sensitivity of online learning to noise and the possibility that the sliding window method ignores long-term trends), an ensemble learning method is used to implement the update of the foundation pit engineering ensemble learning model. The ensemble learning method improves the stability and accuracy of the prediction by combining multiple machine learning base models. After obtaining new monitoring data every day, a new base learner model is retrained using the training standard monitoring data obtained by the latest training window h1. Assume that there are M base learner models, each of which is a model learned from the training data through machine learning methods (such as support vector machines, decision trees, and neural networks). Training the base learner model can improve the accuracy and stability of the prediction; It contains the standard monitoring data of the foundation pit during this period. By analyzing these standard monitoring data, the base learner model attempts to find the correlation with the actual foundation pit settlement. is the actual foundation pit settlement between the time period t-h1′ and t, and is the target output of the base learner model. The base learner model continuously adjusts its parameters to ensure that the predicted results are as close to the actual settlement as possible, thereby establishing a mapping relationship between standard monitoring data and actual settlement. The previously trained base learner model is then used to predict foundation pit settlement, generating actual predicted outputs that provide a basis for risk assessment and decision-making in foundation pit projects.

[0148] Furthermore, the modified model prediction value is obtained based on the prediction results and the prediction standard monitoring data, specifically:

[0149] Perform weighted averaging on the prediction results to obtain the integrated prediction results:

[0150]

[0151] in, represents the integrated prediction result, w i represents the weight of the i-th base learner model, Represents the prediction result of the i-th base learner model;

[0152] Based on the integrated prediction results and sliding window technology, the foundation pit engineering integrated learning model is modified to obtain the predicted value of the modified model:

[0153]

[0154] in, represents the predicted value of the modified model, and h2′ represents the corresponding time point of the predicted standard monitoring data obtained based on the prediction window h2.

[0155] In this embodiment, the prediction results of the M base learner models are Then, in the model prediction phase, the prediction results of multiple base learner models are weighted averaged and fused, and the weights are adjusted based on the performance of the model on the training set. In order to avoid overfitting, the weights can be dynamically optimized based on the historical performance of each base learner model to obtain an integrated prediction result. However, for foundation pit projects, since the construction conditions will change significantly as the construction progresses, the performance of newer base models is usually better than that of older base models. Further adaptive correction is performed, and the prediction results of the base learner model are averaged using sliding window technology and prediction standard monitoring data. During each inference, the model prediction is performed based on the prediction standard monitoring data obtained from the prediction window h2, and the prediction results of multiple models are weighted averaged to obtain the corrected model prediction value. The prediction results were further optimized and the robustness of the model was improved, enabling it to better adapt to the ever-changing working conditions during foundation pit construction.

[0156] Furthermore, the historical prediction error is obtained based on the corrected model prediction value and the actual observation value: if the historical prediction error exceeds the preset prediction error threshold, the training window and the prediction window are adjusted based on the preset training window adjustment amount and the preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, and the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained, specifically:

[0157] Based on the corrected model prediction value and the actual observation value, the historical prediction error is obtained, and then the mean and standard deviation of the historical prediction error are obtained:

[0158]

[0159] Among them, ∈(t) represents the historical forecast error, and y(t) represents the actual observation value;

[0160] Preset prediction error threshold:

[0161] ∈ threshold =μ ε +γσ ε

[0162] Among them, ∈ threshold represents the prediction error threshold, μ ε and σ ε represents the mean and standard deviation of historical forecast errors, and γ represents the empirical parameter;

[0163] If the historical prediction error exceeds the preset prediction error threshold, the training window and prediction window are adjusted based on the preset training window adjustment amount and the preset prediction window adjustment amount:

[0164] h1(t+1)=h1(t)+Δh1if∈(t)>∈ threshold

[0165] h2(t+1)=h2(t)+Δh2if∈(t)>∈ threshold

[0166] Wherein, h1(t+1) represents the latest training window, h2(t+1) represents the latest prediction window, Δh1 represents the preset training window adjustment amount, and Δh2 represents the preset prediction window adjustment amount;

[0167] Δh1 and Δh2 are set to 3-5 days or dynamically adjusted using a data-driven approach:

[0168] Δh 1,2 =α·(∈(t)-∈ threshold )

[0169] Among them, α represents the learning rate;

[0170] Until the historical prediction error is lower than the preset prediction error threshold, the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained.

[0171] In this embodiment, when performing real-time updates, relying solely on current monitoring data for model training may not ensure optimal prediction results. Therefore, a dynamic update mechanism based on historical prediction performance feedback is proposed, which uses historical prediction errors to correct model update parameters. After each model prediction, the historical prediction error is calculated. Analysis of the historical prediction error can identify periods of unstable model performance and intuitively measure the degree of deviation between the corrected model predictions and the actual observed values, reflecting the model's prediction accuracy at each moment. A prediction error threshold is then set to determine whether the model prediction error is acceptable. This prediction error threshold provides an objective basis for determining whether model performance meets requirements. When the historical prediction error exceeds this threshold, it indicates that the model prediction performance is poor and requires adjustment. In practical applications, the prediction error threshold is defined based on statistical analysis. The empirical parameter γ is typically set to 1.96 or 2.58, corresponding to a 95% or 99% confidence interval. Furthermore, a dynamic model adjustment mechanism is established to dynamically adjust the parameters of the model update. When the historical prediction error exceeds a preset threshold, the training window and prediction window in the model parameters are adjusted based on preset training window adjustment amounts and preset prediction window adjustment amounts. This adjustment allows the model to better adapt to changes during foundation pit construction, reoptimizing model training and prediction conditions to improve model stability and prediction accuracy. The preset training window adjustment amounts, Δh1 and Δh2, can be determined empirically. Depending on the characteristics of different construction phases, a fixed step size of 3-5 days can be manually set, or a data-driven approach can be used to dynamically adjust the step size based on error trends. α represents the learning rate, typically ranging from 0.1 to 0.5, with larger errors indicating larger adjustments. This dynamic adjustment method adaptively adjusts the training and prediction windows based on the model's historical prediction errors, enabling the model to more quickly adapt to data changes. Finally, by continuously adjusting the training and prediction windows, the historical prediction error is reduced below the preset prediction error threshold, resulting in the latest training and prediction windows. After each new foundation pit engineering ensemble learning model training and prediction is completed, the prediction performance of the ensemble learning model is evaluated in real time, and parameters are adjusted based on prediction error and historical performance feedback. This dynamic optimization mechanism ensures that the model maintains high accuracy throughout the project. The final ensemble learning model is then determined. It can achieve real-time model updates, improve prediction accuracy, maintain high prediction accuracy, and enhance the reliability of early warning, thereby ensuring the safety of construction and surrounding areas, and providing reliable support for safety monitoring and early warning of foundation pit projects.

[0172] See Figure 2This embodiment provides an early warning system based on an integrated learning model for foundation pit engineering, comprising a data extraction module, a first time window determination module, a time lag calculation module, a time interval acquisition module, a window determination module, a data screening module, a model training module, a correction module, an adjustment module, and a foundation pit early warning module, wherein:

[0173] The data extraction module is used to obtain standard monitoring data based on pre-acquired initial monitoring data, and then obtain trend components and fluctuation components based on the standard monitoring data;

[0174] The first time window determination module is configured to determine the first time window based on the trend component;

[0175] The time lag calculation module is used to obtain an over-threshold time lag and a low-threshold time lag based on the fluctuation component and a preset autocorrelation threshold;

[0176] The time interval acquisition module is used to obtain the first stable stage and the typical interval of the new change based on the fluctuation component, the standard monitoring data and the preset sensitivity threshold;

[0177] The window determination module is configured to determine a training window and a prediction window based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable phase, and the typical interval of new changes;

[0178] The data screening module is used to obtain training standard monitoring data based on the training window and obtain prediction standard monitoring data based on the prediction window;

[0179] The model training module is used to train the base learner model based on the training standard monitoring data and perform prediction to obtain the prediction results;

[0180] The correction module is used to obtain a corrected model prediction value based on the prediction result and the prediction standard monitoring data;

[0181] The adjustment module is configured to obtain a historical prediction error based on the corrected model prediction value and the actual observation value: if the historical prediction error exceeds a preset prediction error threshold, the training window and the prediction window are adjusted based on a preset training window adjustment amount and a preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, and the latest training window and the latest prediction window are obtained, thereby obtaining an integrated learning model;

[0182] The foundation pit early warning module is used to predict the settlement and deformation of the foundation pit based on the integrated learning model, obtain the settlement and deformation prediction results of the foundation pit at different future construction stages, and realize early warning of the foundation pit.

[0183] The early warning system based on the foundation pit engineering integrated learning model provided by this embodiment has a simple structure. In practical application, it only requires a model training module to combine the prediction results of the base learner model with the prediction standard monitoring data for model prediction, further improving the robustness of the model and obtaining more accurate revised model prediction values ​​to adapt to changes in working conditions during foundation pit construction. An adjustment module then compares the revised model prediction values ​​with actual observations and calculates historical prediction errors. By analyzing the historical prediction errors, it is possible to identify time periods when the performance of the foundation pit engineering integrated learning model is unstable, and to correct the update parameters of the foundation pit engineering integrated learning model, providing a basis for correcting the foundation pit engineering integrated learning model. After training and predicting the revised foundation pit engineering integrated learning model, the prediction performance of the foundation pit engineering integrated learning model is evaluated. When the historical prediction error exceeds a preset prediction error threshold, it indicates that the prediction performance is poor. It is necessary to adjust the training window and prediction window based on the feedback of the historical prediction error. This allows the foundation pit engineering integrated learning model to better adapt to data changes caused by changes in construction conditions, improve the model's stability and prediction accuracy, and ensure that high prediction accuracy is consistently maintained during foundation pit construction. Until the historical prediction error is lower than the preset prediction error threshold, the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained; the foundation pit early warning module is used to predict the settlement and deformation of the foundation pit based on the finally updated integrated learning model, and the settlement and deformation prediction results of the foundation pit in different future construction stages are obtained. This can avoid the problems of decreased prediction accuracy, mismatch of model update methods and disconnection of machine learning model updates. By building a dynamic update mechanism, real-time model updates, improved prediction accuracy, and enhanced early warning reliability can be achieved, thereby ensuring the safety of construction and surrounding areas.

[0184] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An early warning method based on an integrated learning model for foundation pit engineering, characterized in that: The following steps are involved: Based on the pre-acquired initial monitoring data, standard monitoring data is obtained, and then the trend component and the fluctuation component are obtained based on the standard monitoring data; determining a first time window based on the trend component; Based on the fluctuation component and the preset autocorrelation threshold, the above-threshold time lag and the below-threshold time lag are obtained; Based on the fluctuation components, standard monitoring data and preset sensitivity thresholds, the first stable stage and the typical interval of new changes are obtained; Determining a training window and a prediction window based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable stage, and the typical interval of new changes; Acquire training standard monitoring data based on the training window, and acquire prediction standard monitoring data based on the prediction window; Train the base learner model based on the training standard monitoring data and make predictions to obtain the prediction results; Based on the prediction results and the prediction standard monitoring data, the revised model prediction value is obtained; Obtain historical prediction errors based on the corrected model prediction value and the actual observation value: If the historical prediction error exceeds the preset prediction error threshold, adjust the training window and the prediction window based on the preset training window adjustment amount and the preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, obtain the latest training window and the latest prediction window, and then obtain the integrated learning model; Based on the integrated learning model, the settlement and deformation of the foundation pit are predicted, and the settlement and deformation prediction results of the foundation pit in different future construction stages are obtained to achieve early warning of the foundation pit.

2. The early warning method based on the foundation pit engineering integrated learning model according to claim 1 is characterized in that: The method of obtaining standard monitoring data based on the pre-acquired initial monitoring data and then obtaining trend components and fluctuation components based on the standard monitoring data is as follows: Standardize the pre-acquired initial monitoring data to obtain standard monitoring data: Among them, X t represents the standard monitoring data at time point t, X t ′ represents the initial monitoring data at the pre-acquired time point t, μ and σ represent the mean and standard deviation of the standard monitoring data; Based on standard monitoring data, the trend component and the volatility component are separated and obtained through filters: X t =T t +C t Among them, T t represents the trend component, C t Represents the fluctuation component.

3. The early warning method based on the foundation pit engineering integrated learning model according to claim 2 is characterized in that: The determining of the first time window based on the trend component is specifically as follows: Estimate the correlation of the trend component in the time series and select the time span that makes the trend component change relatively stably as the first time window: Among them, h trend represents the first time window, N represents the standard monitoring data sequence X={X1,X2,…,X n }, where the standard monitoring data sequence X is composed of standard monitoring data at n time points.

4. The early warning method based on the foundation pit engineering integrated learning model according to claim 3 is characterized in that: The above-threshold time lag and the low-threshold time lag are obtained based on the fluctuation component and the preset autocorrelation threshold, specifically: The autocorrelation function is constructed based on the fluctuation components: Among them, ρ k Represents the constructed autocorrelation function, C={C1,C2,…,C n } represents the standard monitoring data sequence X={X1,X2,…,X n }, represents the mean of the volatility component, k represents the time lag variable; Based on the preset autocorrelation threshold, the autocorrelation function is constructed to obtain the above-threshold time lag and the below-threshold time lag: in, represents the time lag after exceeding the threshold value, represents the low threshold time hysteresis, {τ1|ρ(τ1)>ρ threshold } means all the values ​​satisfying ρ(τ1)>ρ threshold The set of time delay τ1, {τ2|ρ(τ2)<ρ threshold } means all the values ​​satisfying ρ(τ2)<ρ threshold The set of time delays τ2, ρ threshold Indicates the preset autocorrelation threshold.

5. The early warning method based on the foundation pit engineering integrated learning model according to claim 4 is characterized in that: The first stable phase and the typical interval of new changes are obtained based on the fluctuation component, standard monitoring data and preset sensitivity threshold, specifically: Based on the fluctuation components, standard monitoring data and preset sensitivity thresholds, the change time point is detected and the change time point is detected by the CUSUM algorithm: S t′ =max(0,S t′-1 +(C t′ -μ-m)) Among them, t′ is the time point of change, S t′ represents the statistic at the changing time point t′, C t′ represents the standard monitoring data fluctuation component at the time point of the change, and m represents the preset sensitivity threshold; There are multiple t', which are respectively represented as the change time point t' j ; At multiple changing time points t′ j Select the typical interval between the first stable phase and new changes: in, Indicates the first stable stage, Indicates the typical interval of new changes.

6. The early warning method based on the foundation pit engineering integrated learning model according to claim 5 is characterized in that: The training window and the prediction window are determined based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable stage and the typical interval of new changes, specifically: Among them, h1 represents the training window and h2 represents the prediction window.

7. The early warning method based on the foundation pit engineering integrated learning model according to claim 1 is characterized in that: The base learner model is trained based on the training standard monitoring data and prediction is performed to obtain the prediction result, specifically: Based on the training standard monitoring data, the base learner model is trained using machine learning methods: Among them, f i (t) represents the trained base learner model, h1′ represents the time point corresponding to the training standard monitoring data obtained based on the training window h1, represents the training standard monitoring data between the time period t-h1′ to t, represents the actual foundation pit settlement between the time period t-h1′ and t, and there are M base learner models in total; Make predictions based on the trained base learner model and obtain prediction results.

8. The early warning method based on the foundation pit engineering integrated learning model according to claim 1 is characterized in that: The modified model prediction value is obtained based on the prediction results and the prediction standard monitoring data, specifically: Perform weighted averaging on the prediction results to obtain the integrated prediction results: in, represents the integrated prediction result, w i represents the weight of the i-th base learner model, Represents the prediction result of the i-th base learner model; Based on the integrated prediction results and sliding window technology, the foundation pit engineering integrated learning model is modified to obtain the predicted value of the modified model: in, represents the predicted value of the modified model, and h2′ represents the corresponding time point of the predicted standard monitoring data obtained based on the prediction window h2.

9. The early warning method based on the foundation pit engineering integrated learning model according to claim 8 is characterized in that: The method of obtaining the historical prediction error based on the corrected model prediction value and the actual observation value is as follows: if the historical prediction error exceeds the preset prediction error threshold, the training window and the prediction window are adjusted based on the preset training window adjustment amount and the preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, and the latest training window and the latest prediction window are obtained, thereby obtaining the integrated learning model. Specifically, Based on the corrected model prediction value and the actual observation value, the historical prediction error is obtained, and then the mean and standard deviation of the historical prediction error are obtained: Among them, ∈(t) represents the historical forecast error, and y(t) represents the actual observation value; Preset prediction error threshold: v threshold =μ ε +gs ε Among them, ∈ threshold represents the prediction error threshold, μ ε and σ ε represents the mean and standard deviation of historical forecast errors, and γ represents the empirical parameter; If the historical prediction error exceeds the preset prediction error threshold, the training window and prediction window are adjusted based on the preset training window adjustment amount and the preset prediction window adjustment amount: h1(t+1)=h1(t)+Δh1if∈(t)>∈ threshold h2(t+1)=h2(t)+Δh2if∈(t)>∈ threshold Wherein, h1(t+1) represents the latest training window, h2(t+1) represents the latest prediction window, Δh1 represents the preset training window adjustment amount, and Δh2 represents the preset prediction window adjustment amount; Δh1 and Δh2 are set to 3-5 days or dynamically adjusted using a data-driven approach: Δh 1,2 =α·(∈(t)-∈ threshold ) Among them, α represents the learning rate; Until the historical prediction error is lower than the preset prediction error threshold, the latest training window and the latest prediction window are obtained, and then the integrated learning model is obtained.

10. An early warning system based on an integrated learning model for foundation pit engineering, characterized in that: It includes a data extraction module, a first time window determination module, a time lag calculation module, a time interval acquisition module, a window determination module, a data screening module, a model training module, a correction module, an adjustment module and a foundation pit early warning module, wherein: The data extraction module is used to obtain standard monitoring data based on pre-acquired initial monitoring data, and then obtain trend components and fluctuation components based on the standard monitoring data; The first time window determination module is configured to determine the first time window based on the trend component; The time lag calculation module is used to obtain an over-threshold time lag and a low-threshold time lag based on the fluctuation component and a preset autocorrelation threshold; The time interval acquisition module is used to obtain the first stable stage and the typical interval of the new change based on the fluctuation component, the standard monitoring data and the preset sensitivity threshold; The window determination module is configured to determine a training window and a prediction window based on the first time window, the over-threshold time lag, the under-threshold time lag, the first stable phase, and the typical interval of new changes; The data screening module is used to obtain training standard monitoring data based on the training window and obtain prediction standard monitoring data based on the prediction window; The model training module is used to train the base learner model based on the training standard monitoring data and perform prediction to obtain the prediction results; The correction module is used to obtain a corrected model prediction value based on the prediction result and the prediction standard monitoring data; The adjustment module is configured to obtain a historical prediction error based on the corrected model prediction value and the actual observation value: if the historical prediction error exceeds a preset prediction error threshold, the training window and the prediction window are adjusted based on a preset training window adjustment amount and a preset prediction window adjustment amount until the historical prediction error is lower than the preset prediction error threshold, and the latest training window and the latest prediction window are obtained, thereby obtaining an integrated learning model; The foundation pit early warning module is used to predict the settlement and deformation of the foundation pit based on the integrated learning model, obtain the settlement and deformation prediction results of the foundation pit at different future construction stages, and realize early warning of the foundation pit.

Images