A Method and System for Risk Prediction in Deep Foundation Pit Engineering Based on Multi-Source Data Fusion
By introducing a dynamic marginal mechanism and a risk decision boundary smoothing constraint, the problems of high false alarm rate and missed alarm in traditional deep foundation pit risk prediction models are solved, and accurate identification and efficient prediction of deep foundation pit risks are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING JIANYETONG ENG TESTING TECH CO LTD
- Filing Date
- 2026-06-18
- Publication Date
- 2026-07-17
Smart Images
Figure CN122413352A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a method and system for predicting risks in deep foundation pit engineering based on multi-source data fusion. Background Technology
[0002] The risk prediction method for deep foundation pit engineering involves collecting multi-source data, including monitoring deformation, structural internal forces, geological and hydrological conditions, and construction status throughout the entire construction process. Utilizing data preprocessing, feature mining, and intelligent modeling, this method uncovers the temporal deformation evolution patterns of the foundation pit, enabling early prediction and tiered warning of future safety risks such as instability and excessive deformation. However, traditional deep foundation pit risk prediction models cannot adapt to the distribution shifts in monitoring data caused by construction timeline evolution, soil consolidation and drift, and changes in the site environment. This leads to a high false alarm rate, as normal deformation fluctuations and minor dynamic variations in construction are easily misjudged as risk anomalies. Furthermore, traditional deep foundation pit risk prediction models exhibit an imbalanced sensitivity to minute deformation fluctuations and lack the ability to identify and judge gradual implicit deformation, localized hidden hazards, and sudden changes in working conditions, making them highly susceptible to early risk underestimation and missed hazard reporting. Summary of the Invention
[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for predicting risks in deep foundation pit engineering based on multi-source data fusion. Traditional deep foundation pit risk prediction models cannot adapt to the evolution of construction timelines, soil consolidation drift, and changes in the site environment that cause monitoring data distribution shifts. This often leads to misjudging normal deformation fluctuations and minor dynamic variations in construction as risk anomalies, resulting in a high false alarm rate. This solution introduces a dynamic marginal mechanism. By quantifying the differences in the distribution shift of deep foundation pit working conditions, an adaptive curvature control factor is constructed. This factor accurately adapts to the long-term temporal drift of deep foundation pits and the dynamic changes in environmental working conditions, significantly reducing the probability of misjudging normal working condition fluctuations. Based on the curvature control factor, a risk boundary loss is constructed, enabling accurate identification of sudden risk anomalies in deep foundation pits and accurate prediction of minor dynamic variations in normal construction. This approach, by incorporating tolerance, effectively improves the overall accuracy of risk prediction for deep foundation pit projects. Addressing the shortcomings of traditional deep foundation pit risk prediction models—such as their imbalanced sensitivity to minute deformation fluctuations, insufficient ability to identify and judge gradual implicit deformations, localized hidden hazards, and sudden changes in operating conditions—leading to frequent early risk omissions and underreporting of potential hazards, this solution designs a risk decision boundary smoothing constraint term. This term applies a numerical saturation effect to monitoring outlier noise samples and extreme distortion samples that are too far from the risk decision boundary, effectively limiting the excessive stretching and distortion of the global risk boundary by extreme anomaly samples. Furthermore, it introduces a risk margin enhancement loss term to specifically strengthen the penalty and identification capabilities for new construction disturbances, implicit gradual hazards, and localized hidden instability characteristics. This significantly improves the accuracy of identifying high-risk hidden risks in deep foundation pits and effectively reduces the risk of underreporting in deep foundation pit project risk prediction.
[0004] The technical solution adopted by this invention is as follows: The method for predicting the risk of deep foundation pit engineering based on multi-source data fusion provided by this invention includes the following steps:
[0005] Step S1: Data collection for deep foundation pit monitoring;
[0006] Step S2: Feature extraction network design;
[0007] Step S3: Constructing the risk boundary loss;
[0008] Step S4: Smoothing the risk decision boundary;
[0009] Step S5: Design of deep foundation pit risk prediction model;
[0010] Step S6: Engineering risk prediction.
[0011] Further, in step S1, the deep foundation pit monitoring data acquisition involves obtaining historical deep foundation pit engineering monitoring data, including time-series monitoring data, spatial deformation data, and environmental condition data. The acquired data is preprocessed, and sequence alignment and prediction label definition are performed. The historical time-series window is used as the model input, and the future time period's working condition risk status is used as the labeling basis. Normal steady-state samples and risk abnormal samples are labeled to construct an engineering risk prediction dataset.
[0012] Furthermore, in step S2, the feature extraction network design is based on the engineering risk prediction dataset, constructing a multi-flow deep neural network to process different types of deep foundation pit engineering monitoring data and fuse them into feature vectors.
[0013] Furthermore, in step S3, the risk boundary loss construction introduces a dynamic marginal adaptive mechanism to construct a risk decision boundary that can adaptively adjust in real time according to the distribution of deep foundation pit time-series evolution working condition data. The comprehensive feature vector is input into a linear risk mapping layer to obtain a risk prediction score. The historical trend center and the recent trend center are taken, and the Euclidean distance between the comprehensive feature vector and the historical trend center and the recent trend center are calculated respectively. The distribution offset difference is calculated to quantify the degree of abnormal change in the trend of the time-series sample of the deep foundation pit under test relative to the normal working condition evolution distribution, and a curvature control factor is constructed. Then, a risk boundary loss function oriented towards trend extrapolation is constructed.
[0014] Furthermore, in step S4, the risk decision boundary smoothing introduces an engineering risk smoothing constraint term, which applies a numerical saturation effect to time-series distorted samples and instantaneous mutation noise samples that are too far from the risk decision boundary, suppressing the influence of unbalanced gradients of single samples, so that the final learned deep foundation pit risk decision boundary has stronger smoothness.
[0015] Furthermore, in step S5, the deep foundation pit risk prediction model design is based on a feature extraction network and a linear risk mapping layer to construct a deep foundation pit risk prediction model. It relies on a dynamic marginal adaptive mechanism and boundary smoothing constraints to construct a risk decision boundary, and outputs a binary risk discrimination result. The risk boundary loss and smoothing constraint term are integrated, and a risk marginal enhancement loss term is introduced to construct a total loss function. By minimizing the total loss function, the optimal training of the deep foundation pit risk prediction model is completed.
[0016] Furthermore, in step S6, the engineering risk prediction is based on the trained deep foundation pit risk prediction model. Real-time monitoring data of the deep foundation pit project is acquired, preprocessed, and then input into the deep foundation pit risk prediction model. The engineering risk prediction is realized based on the risk discrimination results output by the model.
[0017] The deep foundation pit engineering risk prediction system based on multi-source data fusion provided by the present invention includes a deep foundation pit monitoring data acquisition module, a feature extraction network design module, a risk boundary loss construction module, a risk decision boundary smoothing module, a deep foundation pit risk prediction model design module, and an engineering risk prediction module.
[0018] The deep foundation pit monitoring data acquisition module acquires historical deep foundation pit engineering monitoring data, performs cleaning and preprocessing, and constructs a deep foundation pit engineering risk prediction dataset.
[0019] The feature extraction network design module builds a multi-stream deep neural network based on the deep foundation pit engineering risk prediction dataset, mines different types of data features respectively, and fuses them to obtain a comprehensive feature vector.
[0020] The risk boundary loss construction module calculates the risk decision score of deep foundation pit based on the comprehensive feature vector, constructs the distribution adaptive curvature control factor based on the average value of historical and recent steady-state working conditions of deep foundation pit, and then constructs a risk boundary loss function that adapts to working condition drift.
[0021] The risk decision boundary smoothing module introduces a boundary smoothing constraint term based on risk decision scores to apply a numerical saturation effect to extreme outlier samples, thereby stabilizing the model's risk decision boundary.
[0022] The deep foundation pit risk prediction model design module integrates feature extraction network, risk boundary loss and boundary smoothing constraint term, and combines risk margin enhancement loss term to construct deep foundation pit risk prediction model;
[0023] The engineering risk prediction module uses a trained risk prediction model to predict engineering risks based on real-time deep foundation pit engineering monitoring data.
[0024] The beneficial effects achieved by the present invention using the above solution are as follows:
[0025] (1) In view of the shortcomings of traditional deep foundation pit risk prediction models, which cannot adapt to the distribution deviation of monitoring data caused by construction time-series evolution, soil consolidation drift, and changes in site environment, and are prone to misjudging normal deformation fluctuations and small dynamic variations in construction as risk anomalies, resulting in a high false alarm rate, this solution introduces a dynamic marginal mechanism. By quantifying the difference in the distribution deviation of deep foundation pit working conditions, an adaptive curvature control factor is constructed to accurately adapt to the long-term time-series drift of deep foundation pits and the dynamic changes in environmental working conditions, which greatly reduces the probability of misjudging the fluctuations of normal working conditions of deep foundation pits. Based on the curvature control factor, a risk boundary loss is constructed to achieve accurate identification of sudden risk anomalies in deep foundation pits and high tolerance for small dynamic variations in normal construction, effectively improving the overall accuracy of deep foundation pit engineering risk prediction.
[0026] (2) In view of the fact that traditional deep foundation pit risk prediction models are not sensitive enough to small deformation fluctuations, and are not good at identifying and judging gradual implicit deformation, local hidden dangers and sudden changes in working conditions, which easily leads to early risk omission and hidden danger underreporting, this scheme designs a risk decision boundary smoothing constraint term to apply a numerical saturation effect to monitoring outlier noise samples and extreme distortion samples that are too far away from the risk decision boundary, effectively limiting the excessive stretching and distortion of the global risk boundary by extreme abnormal samples; and introduces a risk margin enhancement loss term to specifically strengthen the punishment and identification ability of new construction disturbances, implicit gradual dangers and local hidden instability characteristics, significantly improving the identification accuracy of high-risk hidden risks of deep foundation pits and effectively reducing the hidden dangers of underreporting in deep foundation pit engineering risk prediction. Attached Figure Description
[0027] Figure 1 A flowchart illustrating the risk prediction method for deep foundation pit engineering based on multi-source data fusion provided by this invention;
[0028] Figure 2 A schematic diagram of the deep foundation pit engineering risk prediction system based on multi-source data fusion provided by the present invention.
[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0031] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0032] Example 1, see Figure 1 The present invention provides a method for predicting the risks of deep foundation pit engineering based on multi-source data fusion, the method comprising the following steps:
[0033] Step S1: Deep foundation pit monitoring data collection, acquire historical deep foundation pit engineering monitoring data, perform cleaning and preprocessing, and construct a deep foundation pit engineering risk prediction dataset;
[0034] Step S2: Feature extraction network design. A multi-stream deep neural network is built based on the deep foundation pit engineering risk prediction dataset to mine different types of data features and fuse them to obtain a comprehensive feature vector.
[0035] Step S3: Risk boundary loss construction. Calculate the risk decision score for deep foundation pits based on the comprehensive feature vector. Construct a distribution adaptive curvature control factor based on the average characteristics of historical and recent steady-state working conditions of deep foundation pits, and then construct a risk boundary loss function that adapts to working condition drift.
[0036] Step S4: Smoothing the risk decision boundary. A boundary smoothing constraint term based on the risk decision score is introduced to apply a numerical saturation effect to extreme outlier samples and stabilize the risk decision boundary of the model.
[0037] Step S5: Design of deep foundation pit risk prediction model, integrating feature extraction network, risk boundary loss and boundary smoothing constraint term, combined with risk margin enhancement loss term, to construct deep foundation pit risk prediction model;
[0038] Step S6: Engineering risk prediction. Based on the trained risk prediction model, engineering risk prediction is performed on real-time deep foundation pit engineering monitoring data.
[0039] Example 2, see Figure 1This embodiment is based on the above embodiment. In step S1, the deep foundation pit monitoring data acquisition involves obtaining historical deep foundation pit engineering monitoring data, including time-series monitoring data, spatial deformation data, and environmental condition data, and performing preprocessing. The time-series monitoring data includes time-series data of the internal forces of the support structure (recording the complete continuous sequence of the deep foundation pit from the initial stage of construction to the current stage, including the rate of stress change and time-series interval fluctuations) and time-series data of displacement and settlement (for the deep foundation pit retaining structure, collecting settlement, horizontal displacement, and deformation rate). The spatial deformation data records the deformation path of the deep foundation pit slope and retaining structure in the spatial domain. The model includes coordinate points, deformation gradient, deformation acceleration, and spatial distribution angles. The environmental conditions data include geological parameters, groundwater depth, and construction time sequence characteristics. The preprocessing uses a Savitzky-Golay filter to smooth the deformation coordinates to eliminate noise from monitoring equipment. All numerical features are Z-score standardized, and sequence alignment and prediction label definitions are performed. Using historical time series windows as model input and future time period risk status as labeling basis, normal steady-state samples (y=1) and risk abnormal samples (y=-1) are labeled to construct an engineering risk prediction dataset.
[0040] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the feature extraction network design is based on the engineering risk prediction dataset. A multi-stream deep neural network is constructed to process different types of deep foundation pit engineering monitoring data and fuse them into a highly discriminative feature vector. The network architecture details specifically include:
[0041] The temporal feature extraction stream uses two layers of bidirectional long short-term memory network to stack internal force and settlement time series data to capture the time dependence and local patterns in deformation sequences and obtain temporal feature vectors.
[0042] The spatial feature extraction stream uses a one-dimensional convolutional neural network to extract local shape patterns and obtain spatial feature vectors for deep foundation pit deformation spatial data.
[0043] The environmental feature embedding stream uses an embedding layer for categorical features and a fully connected layer for continuous features to obtain the environmental feature vector.
[0044] Feature fusion combines the extracted temporal feature vectors, spatial feature vectors, and environmental feature vectors to obtain a comprehensive feature vector z.
[0045] Example 4, see Figure 1This embodiment is based on the above embodiment. In step S3, the risk boundary loss construction is to avoid the vulnerability of prediction performance when facing new engineering disturbances or changes in the distribution of monitoring data. Therefore, a dynamic marginal adaptive mechanism is introduced to construct a risk decision boundary that can be adaptively adjusted in real time with the distribution of deep foundation pit construction condition data, so as to achieve strong suppression of future sudden risk samples of deep foundation pits and high tolerance for small evolutionary variations in normal construction time. The comprehensive feature vector z is input into a linear risk mapping layer to obtain the risk prediction score. The mean of the comprehensive eigenvectors of all historical normal steady-state samples of deep foundation pits is taken as the historical trend center, and the mean of the comprehensive eigenvectors of the normal time-series evolution samples of deep foundation pits in the short term (one month) is taken as the recent trend center. The Euclidean distance between the comprehensive eigenvector z and the historical trend center and the recent trend center is calculated respectively, and expressed as: and Calculate the distribution offset difference The degree of abrupt changes in the trend of the evolution distribution of the deep foundation pit under test relative to normal working conditions was quantified, and a curvature control factor was constructed. , represented as: It adapts to the evolution of construction progress, soil consolidation time drift, and environmental and meteorological time changes in deep foundation pit engineering scenarios, thereby constructing a risk boundary loss function oriented towards trend prediction. , represented as: Where y is the sample label, with 1 for normal steady-state deep foundation pit samples and -1 for risky and abnormal samples; It is the curvature control factor; w and b are the weights and biases of the linear risk mapping layer, respectively; It is the basic curvature coefficient, with a value ranging from 0.5 to 5.0; It is the distribution offset adjustment coefficient, with a value ranging from 0.1 to 1.0; It is a smooth term, a very small positive number.
[0046] By performing the above operations, this solution addresses the shortcomings of traditional deep foundation pit risk prediction models, which cannot adapt to the evolution of construction timelines, soil consolidation drift, and changes in the site environment that cause monitoring data distribution shifts. These models are prone to misjudging normal deformation fluctuations and minor dynamic variations in construction as risk anomalies, leading to a high false alarm rate. This solution introduces a dynamic marginal mechanism, which constructs an adaptive curvature control factor by quantifying the differences in the distribution of deep foundation pit working conditions. This accurately adapts to the long-term temporal drift of deep foundation pits and the dynamic changes in environmental working conditions, significantly reducing the probability of misjudging normal working condition fluctuations. Based on the curvature control factor, a risk boundary loss is constructed, enabling accurate identification of sudden risk anomalies in deep foundation pits and high tolerance for minor dynamic variations in normal construction, effectively improving the overall accuracy of deep foundation pit engineering risk prediction.
[0047] Example 5, see Figure 1This embodiment is based on the above embodiment. In step S4, risk decision boundary smoothing is introduced to further limit the excessive stretching of the overall risk decision boundary by extreme temporal distortion samples and noisy outlier samples in deep foundation pits. An engineering risk smoothing constraint term is applied to temporal distortion samples and instantaneous noise samples that are too far from the risk decision boundary, applying a numerical saturation effect to suppress the influence of unbalanced gradients of single samples. This makes the finally learned deep foundation pit risk decision boundary have stronger smoothness and generalization stability. Represented as: Where p is the distance metric index, with a value of 1 to 2; It is the distance threshold, with a value ranging from 1.0 to 3.0.
[0048] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the deep foundation pit risk prediction model is designed based on a feature extraction network and a linear risk mapping layer. The deep foundation pit risk prediction model architecture consists of three parts connected in series: Feature extraction network: includes temporal feature extraction flow, spatial feature extraction flow, and environmental feature embedding flow, and obtains a comprehensive feature vector through feature concatenation; Linear risk mapping layer: maps the comprehensive feature vector into a risk prediction score; Risk decision layer: constructs the risk decision boundary based on dynamic marginal adaptive mechanism and boundary smoothing constraint, and outputs binary risk classification. The results show that after model convergence, the predicted score s=0 corresponding to the risk decision boundary. If s>0 for the engineering risk prediction data, the prediction is considered normal steady state; if s<0, the prediction is considered risk anomaly; if s=0, it is determined to be a fuzzy critical sample of the working condition, and anomalies are reported separately. The risk boundary loss and smoothing constraint terms are integrated, and a risk margin enhancement loss term is introduced to construct the total loss function. This improves the model's ability to identify new construction disturbances, implicit temporal gradual deformations, and minor hidden long-term hazards in deep foundation pits. By minimizing the total loss function, the optimal training of the deep foundation pit risk prediction model is completed. Represented as: ; Where i is the deep foundation pit monitoring sample index, It is the comprehensive feature vector of the corresponding sample; It is a loss item with increased risk margin; is the hazard penalty coefficient, with a value of 0.1 to 1.0; k is the marginal threshold for abnormal risk samples, with a value of -1.0 to 0, which strengthens the gradient penalty for abnormal samples of deep foundation pits; It is an abnormal sample indicator function, with 1 for risky abnormal samples and 0 for normal steady-state samples; It is the balance coefficient, with a value ranging from 0.01 to 0.1;
[0049] Regarding model training: The engineering risk prediction dataset is divided into training, validation, and test sets (70%, 15%, and 15% respectively), based on the deep foundation pit project ID. This avoids the same deep foundation pit project sample being distributed in both the training and test sets simultaneously, preventing data leakage from projects with the same source and preventing virtual model aggregation. Stratified sampling is used to ensure balanced data distribution. The parameters of the deep foundation pit risk prediction model are initialized. For each batch, forward propagation is performed to output a comprehensive feature vector, calculate the risk prediction score, calculate the total loss, backpropagate the gradient, and update the network parameters through the optimizer. After each round of training, performance evaluation is conducted on the validation set, and the deep foundation pit risk prediction model with the best overall performance is saved. The model's performance in identifying long-term hidden risks and sudden emergencies in foundation pits is evaluated by combining accuracy and F1 score.
[0050] By performing the above operations, this solution addresses the problems of traditional deep foundation pit risk prediction models being unbalanced in their sensitivity to small deformation fluctuations, and lacking the ability to identify and judge gradual implicit deformations, local hidden hazards, and sudden changes in working conditions, which easily lead to missed early risk assessments and unreported hidden dangers. This solution designs a risk decision boundary smoothing constraint term, applying a numerical saturation effect to monitoring outlier noise samples and extreme distortion samples that are too far from the risk decision boundary, effectively limiting the excessive stretching and distortion of the global risk boundary by extreme abnormal samples. Furthermore, a risk margin enhancement loss term is introduced to specifically strengthen the penalty and identification capabilities for new construction disturbances, implicit gradual hazards, and local hidden instability characteristics, significantly improving the accuracy of identifying high-risk hidden risks in deep foundation pits and effectively reducing the risk of missed reporting in deep foundation pit engineering risk prediction.
[0051] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, the engineering risk prediction is based on the trained deep foundation pit risk prediction model. The monitoring data of the deep foundation pit project is acquired in real time, preprocessed and then input into the deep foundation pit risk prediction model. The engineering risk prediction is realized based on the risk discrimination result output by the model.
[0052] Example 8, see Figure 2 Based on the above embodiments, the deep foundation pit engineering risk prediction system based on multi-source data fusion provided by the present invention includes a deep foundation pit monitoring data acquisition module, a feature extraction network design module, a risk boundary loss construction module, a risk decision boundary smoothing module, a deep foundation pit risk prediction model design module, and an engineering risk prediction module.
[0053] The deep foundation pit monitoring data acquisition module acquires historical deep foundation pit engineering monitoring data, performs cleaning and preprocessing, and constructs a deep foundation pit engineering risk prediction dataset.
[0054] The feature extraction network design module builds a multi-stream deep neural network based on the deep foundation pit engineering risk prediction dataset, mines different types of data features respectively, and fuses them to obtain a comprehensive feature vector.
[0055] The risk boundary loss construction module calculates the risk decision score of deep foundation pit based on the comprehensive feature vector, constructs the distribution adaptive curvature control factor based on the average value of historical and recent steady-state working conditions of deep foundation pit, and then constructs a risk boundary loss function that adapts to working condition drift.
[0056] The risk decision boundary smoothing module introduces a boundary smoothing constraint term based on risk decision scores to apply a numerical saturation effect to extreme outlier samples, thereby stabilizing the model's risk decision boundary.
[0057] The deep foundation pit risk prediction model design module integrates feature extraction network, risk boundary loss and boundary smoothing constraint term, and combines risk margin enhancement loss term to construct deep foundation pit risk prediction model;
[0058] The engineering risk prediction module uses a trained risk prediction model to predict engineering risks based on real-time deep foundation pit engineering monitoring data.
[0059] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0061] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A method for predicting the risks of deep foundation pit engineering based on multi-source data fusion, characterized in that: The method includes the following steps: Step S1: Deep foundation pit monitoring data collection, acquire historical deep foundation pit engineering monitoring data, perform cleaning and preprocessing, and construct a deep foundation pit engineering risk prediction dataset; Step S2: Feature extraction network design. A multi-stream deep neural network is built based on the deep foundation pit engineering risk prediction dataset to mine different types of data features and fuse them to obtain a comprehensive feature vector. Step S3: Risk boundary loss construction. Calculate the risk decision score for deep foundation pits based on the comprehensive feature vector. Construct a distribution adaptive curvature control factor based on the average characteristics of historical and recent steady-state working conditions of deep foundation pits, and then construct a risk boundary loss function that adapts to working condition drift. Step S4: Smoothing the risk decision boundary. A boundary smoothing constraint term based on the risk decision score is introduced to apply a numerical saturation effect to extreme outlier samples and stabilize the risk decision boundary of the model. Step S5: Design of deep foundation pit risk prediction model, integrating feature extraction network, risk boundary loss and boundary smoothing constraint term, combined with risk margin enhancement loss term, to construct deep foundation pit risk prediction model; Step S6: Engineering risk prediction. Based on the trained risk prediction model, engineering risk prediction is performed on real-time deep foundation pit engineering monitoring data.
2. The method for predicting the risk of deep foundation pit engineering based on multi-source data fusion according to claim 1, characterized in that: In step S3, the risk boundary loss construction introduces a dynamic marginal adaptive mechanism to construct a risk decision boundary that can adaptively adjust in real time according to the distribution of deep foundation pit time-series evolution working condition data. The comprehensive feature vector is input into a linear risk mapping layer to obtain a risk prediction score. The historical trend center and the recent trend center are taken, and the Euclidean distance between the comprehensive feature vector and the historical trend center and the recent trend center are calculated respectively. The distribution offset difference is calculated to quantify the degree of abnormal change in the trend of the time-series sample of the deep foundation pit under test relative to the normal working condition evolution distribution, and a curvature control factor is constructed. Then, a risk boundary loss function oriented towards trend extrapolation is constructed.
3. The method for predicting the risk of deep foundation pit engineering based on multi-source data fusion according to claim 1, characterized in that: In step S4, the risk decision boundary smoothing introduces an engineering risk smoothing constraint term, which applies a numerical saturation effect to time-series distorted samples and instantaneous mutation noise samples that are too far from the risk decision boundary, suppressing the influence of unbalanced gradients of single samples, so that the final learned deep foundation pit risk decision boundary has stronger smoothness.
4. The method for predicting the risk of deep foundation pit engineering based on multi-source data fusion according to claim 1, characterized in that: In step S5, the deep foundation pit risk prediction model design is based on a feature extraction network and a linear risk mapping layer to construct a deep foundation pit risk prediction model. It relies on a dynamic marginal adaptive mechanism and boundary smoothing constraints to construct a risk decision boundary, and outputs a binary risk discrimination result. The risk boundary loss and smoothing constraint term are integrated, and a risk marginal enhancement loss term is introduced to construct a total loss function. By minimizing the total loss function, the optimal training of the deep foundation pit risk prediction model is completed.
5. The method for predicting the risk of deep foundation pit engineering based on multi-source data fusion according to claim 1, characterized in that: In step S1, the deep foundation pit monitoring data acquisition involves obtaining historical deep foundation pit engineering monitoring data, including time-series monitoring data, spatial deformation data, and environmental condition data. The acquired data is preprocessed, and sequence alignment and prediction label definition are performed. The historical time-series window is used as the model input, and the future time period's working condition risk status is used as the labeling basis. Normal steady-state samples and risk abnormal samples are labeled to construct an engineering risk prediction dataset.
6. The method for predicting the risk of deep foundation pit engineering based on multi-source data fusion according to claim 1, characterized in that: In step S6, the engineering risk prediction is based on the trained deep foundation pit risk prediction model. Real-time monitoring data of deep foundation pit engineering is acquired, preprocessed, and then input into the deep foundation pit risk prediction model. Engineering risk prediction is achieved based on the risk discrimination results output by the model.
7. A deep foundation pit engineering risk prediction system based on multi-source data fusion, used to implement the deep foundation pit engineering risk prediction method based on multi-source data fusion as described in any one of claims 1-6, characterized in that: It includes a deep foundation pit monitoring data acquisition module, a feature extraction network design module, a risk boundary loss construction module, a risk decision boundary smoothing module, a deep foundation pit risk prediction model design module, and an engineering risk prediction module; The deep foundation pit monitoring data acquisition module acquires historical deep foundation pit engineering monitoring data, performs cleaning and preprocessing, and constructs a deep foundation pit engineering risk prediction dataset. The feature extraction network design module builds a multi-stream deep neural network based on the deep foundation pit engineering risk prediction dataset, mines different types of data features respectively, and fuses them to obtain a comprehensive feature vector. The risk boundary loss construction module calculates the risk decision score of deep foundation pit based on the comprehensive feature vector, constructs the distribution adaptive curvature control factor based on the average value of historical and recent steady-state working conditions of deep foundation pit, and then constructs a risk boundary loss function that adapts to working condition drift. The risk decision boundary smoothing module introduces a boundary smoothing constraint term based on risk decision scores to apply a numerical saturation effect to extreme outlier samples, thereby stabilizing the model's risk decision boundary. The deep foundation pit risk prediction model design module integrates feature extraction network, risk boundary loss and boundary smoothing constraint term, and combines risk margin enhancement loss term to construct deep foundation pit risk prediction model; The engineering risk prediction module uses a trained risk prediction model to predict engineering risks based on real-time deep foundation pit engineering monitoring data.