A dynamic prediction method and system for leakage weak points of an ultra-deep circular shaft foundation pit

Abstract

This invention relates to the field of leakage early warning technology for foundation pit engineering, and in particular to a method and system for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits. The method includes the following steps: acquiring a set of static structural parameters, a dataset of geological environment, and a dynamic dataset of construction procedures to construct a multi-source heterogeneous data fusion matrix; extracting key features and preprocessing and encoding the data to obtain a spatiotemporal fusion feature tensor; obtaining the circumferential topological features of the shaft to obtain the temporal features of construction disturbances; performing feature fusion to obtain a feature fusion vector and generating a probability distribution map of weak points in the shaft leakage; and deploying sensors in high-risk areas to acquire real-time monitoring data and establish a data feedback model and a weight update model to achieve dynamic prediction of weak points in the shaft leakage. This invention solves the problems of difficult multi-source heterogeneous data fusion, significant lag in leakage detection technology, and lack of circumferential stress effect modeling, and achieves proactive and precise control of leakage risks in ultra-deep circular vertical shaft foundation pits.

Description

Technical Field

[0001] This invention relates to the field of early warning technology for foundation pit engineering, and in particular to a dynamic prediction method and system for weak points in ultra-deep circular vertical shaft foundation pits. Background Technology

[0002] With the continuous development of urban underground space, ultra-deep circular vertical shaft foundation pits, as an important component of hydropower stations, foundation pits, and urban water storage wells, have received increasing attention for their design and construction safety. Due to the large burial depth, complex retaining system, high groundwater level, and strong construction disturbance of vertical shaft projects, the seepage prevention performance of the structure directly affects the overall stability and durability of the project. In recent years, as foundation pit engineering has shifted from strength control to the coordinated control of deformation and seepage, the traditional seepage prevention design based primarily on experience or safety factors has gradually revealed its limitations. Simply relying on passive waterproofing or high-strength materials is no longer effective in addressing the risk of multi-source seepage. How to achieve proactive prediction and precise control of weak seepage areas during the construction phase has become an important research direction for the intelligent construction of deep foundation pit projects.

[0003] While existing leakage detection and early warning technologies have been applied in numerous projects, they remain largely at the stage of post-event identification and localized monitoring. Detection methods based on electrical resistivity to locate leakage areas by measuring conductivity anomalies, but they are only effective after the leakage channel has formed. Monitoring methods based on tracer diffusion rely on periodic manual observation of water level changes in wells, limiting their real-time performance and sensitivity. Meanwhile, some studies have attempted to use sensor networks for predictive assessment, but their modeling often neglects the combined effects of circumferential stress concentration in shafts and joint anti-seepage degradation, resulting in insufficient spatial resolution. Due to the heterogeneity of different data sources in time, space, and scale, traditional algorithms have significant shortcomings in multi-dimensional data fusion, temporal correlation mining, and adaptive model updates, often resulting in predictions that lag behind the actual occurrence of leakage.

[0004] Therefore, there is an urgent need to propose a new prediction method that can comprehensively consider the circumferential stress effect, the permeability characteristics of the formation and the evolution law of construction disturbance, so as to realize the dynamic identification of weak points of leakage in ultra-deep circular vertical shaft foundation pits. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits. The method includes the following steps: acquiring a set of static structural parameters, a geological environment dataset, and a dynamic dataset of construction procedures; constructing a multi-source heterogeneous data fusion matrix through spatiotemporal index mapping; extracting key features and preprocessing the multi-source heterogeneous data fusion matrix to obtain a spatiotemporal fusion feature tensor; based on the spatiotemporal fusion feature tensor, obtaining the circumferential topological features of the shaft through a spatial feature extraction network, and obtaining the temporal features of construction disturbances using a temporal feature extraction network; fusing the circumferential topological features of the shaft and the temporal features of construction disturbances to obtain a feature fusion vector, and generating a probability distribution map of weak points in the shaft through a leakage risk prediction classifier; and deploying sensors in high-risk areas based on the probability distribution map of weak points in the shaft, acquiring real-time monitoring data, establishing a data feedback model and a weight update model, and realizing dynamic prediction of weak points in the shaft leakage. This invention accurately locates high-risk leakage areas, providing a scientific basis for construction protection, improving the timeliness and accuracy of prediction, reducing engineering risks caused by leakage hazards, and ensuring the safety and stability of ultra-deep circular vertical shaft foundation pit construction.

[0007] Optionally, the acquisition of the structural static parameter set, the geological environment dataset, and the construction process dynamic dataset, and the construction of a multi-source heterogeneous data fusion matrix through spatiotemporal index mapping, includes: the structural static parameter set including reinforcement information, joint characteristics, and concrete strength grade of the ultra-deep circular vertical shaft foundation pit; the geological environment dataset including geological borehole exploration data, soil permeability parameters, groundwater level changes, and geological structure information; and the construction process dynamic dataset including tunneling progress, grouting pressure, structural strain, and surrounding rock deformation. The spatiotemporal index mapping unifies the timestamps and spatial coordinates of the structural static parameter set, the geological environment dataset, and the construction process dynamic dataset to construct the multi-source heterogeneous data fusion matrix. This invention achieves standardized integration and efficient mapping of multi-source heterogeneous data, reduces data redundancy and errors, improves data utilization efficiency, ensures the reliability of subsequent prediction results, and provides comprehensive and accurate data support for leakage risk assessment.

[0008] Optionally, the step of extracting key features and preprocessing the data to encode the spatiotemporal fusion feature tensor from the multi-source heterogeneous data fusion matrix includes: acquiring physical and mechanical indicators; performing indicator screening and feature mapping based on the multi-source heterogeneous data fusion matrix to obtain key parameter factors; and combining the key parameter factors to preprocess and encode the multi-source heterogeneous data fusion matrix to obtain the spatiotemporal fusion feature tensor. This invention effectively reduces model computational complexity, accelerates prediction speed, and simultaneously ensures the effectiveness and adaptability of feature data, providing strong support for accurate spatiotemporal feature extraction and improved prediction accuracy.

[0009] Optionally, the step of obtaining the shaft circumferential topological features through a spatial feature extraction network based on the spatiotemporal fusion feature tensor, and obtaining the construction disturbance temporal features using a temporal feature extraction network, includes: the spatial feature extraction network adopting a graph convolutional network architecture to obtain the shaft circumferential topological features based on the spatiotemporal fusion feature tensor; and the temporal feature extraction network adopting a long short-term memory network architecture to obtain the construction disturbance temporal features in combination with the spatiotemporal fusion feature tensor. This invention overcomes the limitations of single feature extraction, improves the comprehensiveness and depth of feature fusion, enhances the model's adaptability to complex engineering scenarios, and makes leakage risk prediction more closely aligned with actual construction conditions, resulting in more accurate prediction results.

[0010] Optionally, obtaining the circumferential topological features of the shaft based on the spatiotemporal fusion feature tensor includes: dividing the ultra-deep circular shaft pit into several sector-shaped partitions along the circumferential direction, using the several sector-shaped partitions as input nodes of the graph convolutional network architecture; determining adjacent partitions through the several sector-shaped partitions and establishing an adjacency matrix; and based on the adjacency matrix, performing multi-layer convolution and aggregation on the spatiotemporal fusion feature tensor through the graph convolutional network architecture to obtain the circumferential topological features of the shaft. This invention effectively captures the spatial correlation of each partition, improves the targeting and accuracy of spatial feature representation, provides precise spatial dimension support for subsequent leakage risk assessment, and helps identify high-risk areas.

[0011] Optionally, obtaining the construction disturbance temporal features by combining the spatiotemporal fusion feature tensor includes: inputting the spatiotemporal fusion feature tensor into the long short-term memory network architecture, and outputting the construction disturbance temporal features using a recursive state update formula. This invention overcomes the shortcomings of traditional methods in quantifying temporal impacts, improves the response capability to dynamic construction disturbances, ensures that prediction results can match the construction progress in real time, and enhances the dynamic adaptability of predictions.

[0012] Optionally, the step of fusing the circumferential topological features of the shaft and the temporal features of construction disturbance to obtain a feature fusion vector, and generating a probability distribution map of weak points in the shaft leakage through a leakage risk prediction classifier, includes: performing a vector concatenation operation on the circumferential topological features of the shaft and the temporal features of construction disturbance to obtain a concatenated feature vector; introducing an attention mechanism to adaptively weight the concatenated feature vector to obtain the feature fusion vector; and inputting the feature fusion vector into the leakage risk prediction classifier to output the probability distribution map of weak points in the shaft leakage. This invention provides a precise targeting basis for subsequent sensor deployment and protection measures formulation through the probability distribution map of weak points in shaft leakage, improving the pertinence of engineering protection.

[0013] Optionally, the step of targeting and deploying sensors in high-risk areas based on the probability distribution map of weak points in the vertical shaft leakage, acquiring real-time monitoring data, establishing a data feedback model and a weight update model, and realizing dynamic prediction of weak points in the vertical shaft leakage includes: obtaining leakage risk probability values ​​based on the probability distribution map of weak points in the vertical shaft leakage, and identifying high-risk areas based on the leakage risk probability values; targeting and deploying the sensors in the high-risk areas to acquire the real-time monitoring data; constructing the data feedback model through the difference between the real-time monitoring data and the probability distribution map of weak points in the vertical shaft leakage, and obtaining an error feedback signal based on the data feedback model; using the error feedback signal as a driver, establishing the weight update model using an adaptive optimization strategy; constructing an update strategy as a constraint condition, and performing online updates through the data feedback model and the weight update model to realize dynamic prediction of weak points in the vertical shaft leakage. This invention breaks through the limitations of traditional static prediction, realizes real-time early warning and dynamic control of leakage risks, and minimizes the incidence of leakage accidents.

[0014] Optionally, the online update via the data feedback model and the weight update model includes: online correction of the difference using the error feedback signal; and online update of the weight parameters of the spatial feature extraction network and the temporal feature extraction network via the weight update model. This invention effectively improves the stability and prediction accuracy of long-term operation, avoids prediction failure due to iterative operating conditions, ensures the continuity and reliability of dynamic prediction, and provides continuous technical support for the safety protection of the entire engineering process.

[0015] Secondly, this invention provides a dynamic prediction system for weak points in ultra-deep circular vertical shaft foundation pits that facilitate leakage. The system executes the dynamic prediction method for weak points in ultra-deep circular vertical shaft foundation pits provided by this invention. The system includes input devices, output devices, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions, and the processor is configured to invoke the program instructions. This invention, through the collaborative use of high-performance hardware, achieves efficient implementation and stable operation of the method, providing reliable hardware support for the prevention and control of leakage in ultra-deep circular vertical shaft foundation pits. Attached Figure Description

[0016] Figure 1 This is a flowchart of a dynamic prediction method for leakage weak points in an ultra-deep circular vertical shaft foundation pit according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the discretization and topology of the circumferential nodes of an ultra-deep circular vertical shaft foundation pit according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatiotemporal fusion feature tensor construction process according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the process of generating a probability distribution map of weak points in vertical shaft leakage according to an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the high-risk area identification in an embodiment of the present invention; Figure 6 This is a schematic diagram of the online update process according to an embodiment of the present invention; Figure 7 This is a framework diagram of a dynamic prediction system for weak points in the leakage of an ultra-deep circular vertical shaft foundation pit, according to an embodiment of the present invention. Detailed Implementation

[0017] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0018] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0019] Please see Figure 1 One embodiment of the present invention provides a method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits, the method comprising the following steps: S1. Obtain the set of static structural parameters, the dataset of geological environment, and the dynamic dataset of construction procedures, and construct a multi-source heterogeneous data fusion matrix through spatiotemporal index mapping.

[0020] In this embodiment, based on the design and construction organization of the shaft retaining structure, the on-site data collection includes multi-source fusion collection of three types of information: foundation pit structure, surrounding strata, and construction procedures.

[0021] Specifically, based on the design and construction conditions of the shaft retaining structure, a structural information database interface is established to extract vectorized data from the design drawings. This vectorized data is then used to construct a set of static structural parameters, satisfying the following relationships: in, For the set of static parameters of the structure, For the thickness of the diaphragm wall, For the design strength of concrete, The connector type is coded (0-locked pipe, 1-I-beam). For other vectorized data.

[0022] Static parameters such as reinforcement details, joint characteristics, concrete strength grade, dimensional information, and material grade of the shaft retaining structure are obtained through a set of structural static parameters to reflect the overall stress characteristics and impermeability of the shaft circumferential structure.

[0023] Specifically, based on the geological survey report, a three-dimensional geological grid is established. For any depth of the shaft, a vector is constructed as a stratigraphic environment dataset, satisfying the following relationship: in, For geological environment datasets, For any depth of the shaft, The horizontal permeability coefficient, The vertical permeability coefficient, The soil weight, This refers to the moisture content.

[0024] Based on the geological environment dataset, borehole exploration data, soil permeability parameters, groundwater level changes, and geological structure information are obtained to characterize the seepage conditions and hydraulic response characteristics of the soil outside the shaft.

[0025] Specifically, time-series data is collected in real time through interfaces deployed on construction equipment and on-site sensors at a sampling frequency (preferably 1Hz), and a time-series dataset is constructed as a dynamic dataset of construction procedures, satisfying the following relationship: in, For dynamic datasets of construction procedures, For tunneling speed, For real-time grouting pressure, The torque of the cutter head. This represents the real-time soil displacement.

[0026] Based on the dynamic dataset of construction procedures, real-time monitoring of construction disturbance parameters, including time-varying data such as tunneling progress, grouting pressure, structural strain, and surrounding rock deformation during the construction cycle, is used to characterize the dynamic impact of the construction process on the shaft structure and surrounding strata.

[0027] In this embodiment, the set of static structural parameters, the dataset of geological environment, and the dataset of dynamic construction procedures are aligned using timestamps and spatial coordinates to form a multi-dimensional and temporally sequential multi-source heterogeneous data fusion matrix. This matrix serves as the input data basis for subsequent feature tensor construction and model training.

[0028] Specifically, since the structural static parameter set is static, the geological environment dataset is spatially distributed, and the construction procedure dynamic dataset is dynamic, they need to be fused based on time and spatial sector numbering. The multi-source heterogeneous data fusion matrix is ​​constructed using spatiotemporal alignment mapping technology, satisfying the following relationship: in, For multi-source heterogeneous data fusion matrix, Number the spatial sectors, For a moment, For space sector The set of static structural parameters This represents a vector concatenation operation. For space sector Stratigraphic environment dataset, For a moment Dynamic dataset of construction procedures.

[0029] It should be noted that if the sampling frequencies of different source data are inconsistent, linear interpolation is used to map the low-frequency data onto the high-frequency time axis.

[0030] In this embodiment, at the multi-source data acquisition level, the data acquisition process can combine fiber optic sensing technology, video image recognition technology, and formation drilling survey results to construct a multi-channel information acquisition network composed of circumferential stress, formation permeability, and construction disturbance. This network can form a vertical shaft spatial discretization model through a partitioned and layered deployment, providing an initial topological foundation for node partitioning and spatiotemporal feature mapping of graph convolutional networks.

[0031] Please see Figure 2 The diagram shows the discrete and topological schematic of the circumferential nodes of an ultra-deep circular vertical shaft foundation pit; where... For circular partition nodes, The weights of adjacent partitions.

[0032] Specifically, data acquisition can adopt a combination of fiber optic strain, pore pressure monitoring and video image recognition, and the data points can be deployed according to circumferential partitioning and depth layering to form a discrete initial topology for the vertical shaft space.

[0033] In an optional embodiment, only geometric features such as structural joint misalignment and moisture seepage marks can be extracted from the image data for subsequent image recognition, avoiding the direct use of the original video to reduce the computational burden on the model.

[0034] Without altering the process of step S1, the following preferred configuration can be used for on-site data acquisition to facilitate project implementation: Structural static parameter set: diaphragm wall thickness preferably ≥600mm, joint type can be interlocking pipe or I-beam joint, concrete grade can be C35-C40, etc.; Geological environment dataset: integrating permeability coefficient order obtained from borehole exploration (e.g., ), soil moisture content (e.g., 0.50), etc.; dynamic dataset of construction procedures: real-time collection of tunneling speed (e.g., ≤0.5m / h), grouting pressure (e.g., 0.8MPa-1.2MPa), etc.

[0035] Without changing the process of step S1, the monitoring network and discretization can adopt the following preferred configuration: sensors are deployed in circumferential zones and depth layers on the inner wall of the shaft; the circumferential zone can be divided into 8 sectors at 45°, and a monitoring layer is set every 5m in the depth direction; image recognition is used for joint misalignment / wetness trace extraction, the recognition accuracy can be preferably ±2mm, and the sampling frequency can be preferably ≥10Hz.

[0036] It should be noted that the above values ​​are preferred and example configurations for the embodiments and can be adjusted according to engineering conditions.

[0037] S2. Extract key features and preprocess and encode data from the multi-source heterogeneous data fusion matrix to obtain a spatiotemporal fusion feature tensor.

[0038] Based on the completion of multi-source data acquisition, the construction of spatiotemporal fusion feature tensors aims to achieve structured processing and quantitative expression of different types of data, so as to form a unified feature system that can be used for deep learning model recognition and training.

[0039] In this embodiment, a key feature extraction module is used to screen and map the multi-source input information in the multi-source heterogeneous data fusion matrix to extract key parameter factors that reflect the evolution of leakage risk in ultra-deep vertical shaft foundation pits. This feature extraction process comprehensively considers the stress characteristics of the vertical shaft circumferential structure, the impermeability of the joints, the seepage conditions of the strata, and the dynamic impact of construction disturbances.

[0040] The core feature set of the key feature extraction module includes one or more physical and mechanical indicators from the following: circumferential stress non-uniformity coefficient, joint impermeability degradation degree, soil permeability gradient abrupt change factor, and construction disturbance intensity factor.

[0041] It should be noted that the circumferential stress non-uniformity coefficient (characterizing the structural stress eccentricity), joint impermeability deterioration degree (calculated based on the joint width identified by image and historical records), soil permeability gradient abrupt change factor (determined by the permeability coefficient of each layer), and construction disturbance intensity factor (characterized by a combination of tunneling parameters) are all included.

[0042] In this embodiment, the key feature extraction module is configured to construct index screening and feature mapping according to the following logic, including the following steps: First, extract fiber stress data from the multi-source heterogeneous data fusion matrix and define nodes. Circumferential stress non-uniformity coefficient This is used to filter and highlight stress concentration areas, satisfying the following relationship: in, This is the coefficient for circumferential stress non-uniformity. For the node index, This represents the peak value of the principal stress at the monitoring point in this sector. The average circumferential stress in the same layer.

[0043] Subsequently, the joint opening amount was extracted from the image recognition data to construct the joint's resistance to seepage degradation. , used to characterize the physical integrity of the waterstop structure, satisfies the following relationship: in, To improve the joint's resistance to seepage degradation, The degradation factor is adjusted based on historical leakage records. For real-time monitoring of joint width, To design the maximum permissible seam width threshold.

[0044] Simultaneously, based on stratigraphic exploration data, a mutation factor for soil permeability gradient is defined. This is used to locate hydraulically active layers, satisfying the following relationship: in, The soil permeability gradient abrupt change factor. strata Vertical permeability coefficient, strata The vertical permeability coefficient.

[0045] It should be noted that the larger the soil permeability gradient abrupt change factor, the more obvious the hydraulic abrupt change between layers, and the higher the risk of leakage.

[0046] Subsequently, the construction progress and grouting parameters are integrated to construct a construction disturbance intensity factor that satisfies the following relationship: in, The construction disturbance intensity factor. These are the weighting coefficients. For tunneling speed, For real-time grouting pressure, This represents the formation fracturing pressure threshold.

[0047] Finally, the key feature extraction module performs variance filtering on the calculated feature indicators, eliminating quasi-static features with variance below the threshold and retaining high-fluctuation features that are sensitive to leakage.

[0048] Please see Figure 3 The figure shows a schematic diagram of the spatiotemporal fusion feature tensor construction process; based on the structural static parameter set, the geological environment dataset, and the construction procedure dynamic dataset, the spatiotemporal fusion feature tensor is obtained through the data preprocessing and encoding module.

[0049] In this embodiment, the data preprocessing and encoding module performs timestamp alignment and dimension normalization on continuous variables, and interpolates or filters missing values ​​to improve the robustness of data fusion, thereby obtaining a spatiotemporal fusion feature tensor with spatially partitioned topology and temporally unified time periods.

[0050] Specifically, the data preprocessing and encoding module is used to perform standardization and encoding operations on different types of data. The static structural parameter set, the geological environment dataset, and the dynamic construction process dataset are spatiotemporally fused: for continuous variables, a normalization method is used to unify the amplitude scale; for discrete or categorical variables, one-hot encoding is used to map them into orthogonal binary vectors to achieve an informational representation of structural differences.

[0051] The above normalization for continuous variables, and the Z-Score standardization used for continuous variables such as stress and seepage pressure to eliminate dimensional differences, satisfy the following relationship: in, These are the normalized eigenvalues. These are the original eigenvalues. This is the historical sliding window mean of this feature. The standard deviation is denoted as .

[0052] Furthermore, through timestamp matching and spatial indexing mechanisms, structural static features, formation permeability features, and construction disturbance variables are spatiotemporally aligned to construct a spatiotemporally fused feature tensor. This spatiotemporally fused feature tensor maintains the circumferential partitioning topology in the spatial dimension and represents the evolutionary patterns of the construction phases in the temporal dimension. The dimensions of the spatiotemporally fused feature tensor are constructed as follows: ), spatial dimension ( Maintain the vertical circumferential partition topology (e.g., corresponding to 8 sector nodes); time dimension ( ): Characterizes the evolutionary pattern of the construction phase (corresponding to the historical time step sequence); Feature dimensions ( ): This includes the physical indicators and coding features mapped above.

[0053] In an optional embodiment, a feature importance and weight balancing mechanism is introduced to perform dimensionality reduction or weighting when feature redundancy exists, thereby improving the stability of model training.

[0054] In an optional embodiment, spatiotemporal alignment can also employ sliding time window resampling or event-driven alignment (with reference to key nodes of the construction process) as needed for engineering purposes. For asynchronous sensor data, interpolation / filtering (such as linear or spline interpolation, Kalman smoothing) can be combined. Registration is completed spatially through circumferential partitioning and depth-layered mesh mapping. Mapping methods include, but are not limited to, nearest neighbor, inverse distance weighting, or triangulation-based interpolation.

[0055] It should be noted that the structure of the above core feature set and spatiotemporal fusion feature tensor is only a preferred configuration in this embodiment, and can be replaced or modified in different strata (such as sandy soil or cohesive soil) and construction organization.

[0056] S3. Based on the spatiotemporal fusion feature tensor, the circumferential topological features of the shaft are obtained through a spatial feature extraction network, and the temporal features of construction disturbance are obtained through a temporal feature extraction network.

[0057] In this embodiment, a spatial feature extraction network and a temporal feature extraction network are constructed and run to perform parallel deep feature computation on the spatiotemporal fusion feature tensor.

[0058] In this embodiment, the spatial feature extraction network adopts a graph convolutional network architecture. Its input nodes correspond to several sector partitions divided along the circumferential direction of the circular shaft. The edge weights of the adjacency matrix are determined by normalization of one or more factors, such as the geometric adjacency and mechanical coupling relationship between adjacent partitions.

[0059] Specifically, the structural unit of the circular shaft is discretized into several nodes along the circumferential direction. (like Figure 2 As shown in the figure, an adjacency matrix is ​​constructed based on the mechanical and permeability relationships between adjacent nodes. A spatial feature extraction network performs multi-layer convolution and feature aggregation on the spatiotemporal fusion feature tensor to achieve spatial correlation extraction of features such as stress, joint impermeability, and formation hydraulic gradient between vertical shaft circumferential nodes, obtaining high-order spatial features as the vertical shaft circumferential topological features. The spatial feature extraction network is used to process the topological relationships of the ring structure, and its core spatial feature aggregation formula satisfies the following relationship: in, This represents the circumferential topological features of a vertical shaft. It is a non-linear activation function. It is an adjacency matrix. for Slices of feature tensors input at each time step. This is the weight matrix.

[0060] It should be noted that the feature tensor slices are the raw data such as stress and leakage indicators of each node; the adjacency matrix numerically represents the physical connection relationship between vertical shaft sectors (if two sectors are adjacent, the corresponding position in the matrix is ​​1, otherwise it is 0); the weight matrix is ​​a parameter automatically learned through training, representing the attenuation or enhancement coefficient of different features during spatial propagation; the nonlinear activation function simulates the nonlinear characteristics of the formation medium's response to seepage; the high-order spatial features not only include the monitoring data of this node, but also aggregate the risk information of surrounding neighboring nodes, thereby realizing a comprehensive assessment of the overall structural risk in the circumferential direction.

[0061] In this embodiment, the temporal feature extraction network adopts a long short-term memory network architecture. Its input sequence includes one or more time variables from the construction stage disturbance intensity, tunneling progress, grouting parameters, and seepage response curves. By learning the time dependence patterns in the construction process through memory units and gating mechanisms, dynamic modeling of temporal signals such as construction progress, grouting pressure, disturbance intensity, and hydraulic gradient changes can be achieved.

[0062] Specifically, construction process disturbances (such as grouting pulses) and time-varying feature sequences of the seepage field are input into a long short-term memory network architecture. Recursive units are used to capture the time dependencies during construction, extracting the temporal features of construction disturbances. The temporal feature extraction network captures the temporal dependencies of construction progress, and its core recursive state update formula satisfies the following relationship: in, The timing characteristics of construction disturbance. It is a non-linear activation function. For the input weight matrix, For a moment The input vector, This is a cyclic weight matrix. For the previous moment The hidden layer state, This is the bias vector.

[0063] It should be noted that the construction disturbance time series features integrate the current construction actions and historical cumulative effects, and will serve as an important input to the prediction module; the input vector includes monitoring data such as tunneling speed, grouting pressure, and real-time seepage pressure; the hidden layer state at the previous moment represents the cumulative response of the formation to previous construction disturbances; the input weight matrix and the cyclic weight matrix are used to adjust the proportion of the influence of the current disturbance and historical memory on the final risk; the bias vector is used to adjust the activation threshold of the model.

[0064] S4. The circumferential topological features of the shaft and the temporal features of the construction disturbance are fused to obtain a feature fusion vector, and a probability distribution map of the weak points of shaft leakage is generated by a leakage risk prediction classifier.

[0065] In this embodiment, a spatiotemporal feature attention fusion layer is constructed, which concatenates and weights the vertical shaft circumferential topology features output by the spatial feature extraction network and the construction disturbance temporal features output by the temporal feature extraction network. By introducing an attention mechanism into the fusion layer, the importance of multi-source features is adaptively weighted, highlighting the dominant factors that contribute significantly to the identification of weak points in leakage.

[0066] Specifically, in the spatiotemporal feature attention fusion layer, the vertical shaft circumferential topological features and construction disturbance temporal features are concatenated into a concatenated feature vector. An attention mechanism is then introduced to adaptively weight the concatenated feature vector to obtain the feature fusion vector.

[0067] Please see Figure 4 The diagram illustrates the process of generating a probability distribution map of weak points in vertical shaft leakage. The spatiotemporal fusion feature tensor is used as input and fed into a leakage risk prediction classifier. This classifier uses two parallel branches, a "spatial feature extraction network" and a "temporal feature extraction network," and then performs fusion processing by a "spatiotemporal feature attention fusion layer" to finally output the probability distribution map of weak points in vertical shaft leakage.

[0068] Specifically, the feature fusion vector is input into the leakage risk prediction classifier, and the Softmax function is used to generate a probability distribution map of weak points in the vertical shaft leakage.

[0069] It should be noted that the leakage risk prediction classifier adopts a fully connected layer architecture and receives the feature fusion vector output by the spatiotemporal feature attention fusion layer.

[0070] In this embodiment, the probability distribution map of weak points in vertical shaft leakage is visualized in the circumferential space in the form of "thermal ring / filling segment".

[0071] S5. Based on the probability distribution map of weak points in the vertical shaft leakage, sensors are deployed in high-risk areas to obtain real-time monitoring data, establish a data feedback model and a weight update model, and realize dynamic prediction of weak points in the vertical shaft leakage.

[0072] In this embodiment, the leakage risk probability value corresponding to each node is output through the probability distribution map of leakage weak points in the shaft; when the leakage risk probability value of any node exceeds the optimal threshold determined based on the Receiver Operating Characteristic Curve (ROC), the area corresponding to that node is marked as a high-risk leakage weak point, thereby realizing the spatial identification and visual labeling of high-risk areas in the shaft.

[0073] The formula for calculating the above leakage risk probability value satisfies the following relationship: in, This represents the probability value of leakage risk. For activation function, This is the weight matrix of the fully connected layer. The timing characteristics of construction disturbance. This represents the circumferential topological features of a vertical shaft. This is a bias term.

[0074] It should be noted that the leakage risk probability value represents the first... The predicted probability value of leakage occurring in each sector node (between 0 and 1); The feature concatenation operator merges the evolutionary patterns of the temporal dimension with the topological features of the spatial dimension into a high-dimensional joint feature vector; the weight matrix of the fully connected layer (i.e., the attention weighting parameters) automatically identifies the importance of features through model training, such as automatically amplifying them when there is severe construction disturbance. The weight is automatically amplified during the resting period. The weights; the bias term is used to adjust the intercept of the classification plane.

[0075] Please see Figure 5 The diagram illustrates the identification of high-risk areas; in this embodiment, the leakage risk probability value is compared with a preset threshold. Threshold determination: When the leakage risk probability value of a certain sector... At that time, mark the location as a high-risk sector (e.g. Figure 5 High-risk sectors 1, 2, and 3 are designated as high-risk areas.

[0076] It should be noted that the preset threshold Instead of using fixed empirical values, dynamic optimization is performed based on the ROC curve of the validation set. Candidate thresholds ranging from 0 to 1 are traversed, and the corresponding precision and recall are calculated. The probability corresponding to the maximum value of the ROC curve is selected as the optimal threshold, which is the preset threshold. (like This is to balance the false alarm rate and the missed alarm rate.

[0077] In an optional embodiment, when applied in the soft soil area of ​​Hangzhou, the weight matrix of the fully connected layer obtained by model training usually shows a high degree of attention to the disturbance intensity during the construction stage. This phenomenon is consistent with the physical law of strong thixotropy of soft soil in this area. The weight distribution is only an empirical observation under specific geological conditions, and the model has the generalization ability to automatically adjust the weights according to different strata conditions.

[0078] In this embodiment, dense sensors are deployed in high-risk areas. The prediction method iterative optimization mechanism uses real-time monitoring data to adaptively and dynamically update the data feedback model and weight update model, thereby continuously improving and dynamically correcting the prediction accuracy of weak points in vertical shaft leakage, and realizing iterative evolution of prediction capabilities.

[0079] Please see Figure 6 The diagram illustrates the online update process. Based on the output probability distribution map of weak points in the vertical shaft leakage, error feedback signals for online correction are obtained through real-time monitoring data. Combined with the update strategy, the weight coefficients are updated to complete the online update of the probability distribution map of weak points in the vertical shaft leakage.

[0080] Specifically, dense sensor arrays are deployed in high-risk areas to collect and upload stress changes, seepage pressure fluctuations, and construction image information in real time; a data feedback model is constructed based on the difference between the probability distribution map of weak points in vertical shaft leakage output by the model and the real-time monitoring data; error feedback signals are obtained for online correction after error calculation and consistency verification, and the difference is corrected using the error feedback signals.

[0081] Furthermore, a weight update model is constructed, driven by the error feedback signal, and an adaptive optimization strategy is adopted to update the weight parameters of the spatial feature extraction network and the temporal feature extraction network online.

[0082] Meanwhile, an update strategy is constructed to dynamically regulate the model training cycle and update frequency. The update frequency and iteration step size are dynamically adjusted based on the rate of change of monitored data and the prediction error threshold, so as to ensure the real-time performance of predictions while avoiding model overfitting.

[0083] In an optional embodiment, the weight update model may employ different optimization schemes, including but not limited to: first-order adaptive optimization (adaptive moment estimation with decoupled weight decay, root mean square propagation, etc.); learning rate scheduling (exponential decay, cosine annealing, etc.); and robust training methods (gradient clipping, exponential moving average, etc.).

[0084] In an optional embodiment, under conditions of adjacent shafts or layered formations, online transfer learning can be introduced to transfer some weights of the feature extraction layer and fine-tune them on the target scenario to improve the generalization ability across working conditions.

[0085] It should be noted that online updates can be performed either in a streaming, small-batch manner or in batch processing during nighttime construction shutdown periods.

[0086] Please see Figure 7In one optional embodiment, the present invention provides a dynamic prediction system for weak points in ultra-deep circular vertical shaft foundation pits, the system comprising an input device, an output device, a processor, and a memory, the hardware facilities being interconnected. The memory stores a computer program, the computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute specific steps as described in the relevant embodiments of the dynamic prediction method for weak points in ultra-deep circular vertical shaft foundation pits provided by the present invention. The dynamic prediction system for weak points in ultra-deep circular vertical shaft foundation pits provided by the present invention has a complete structure, is objective and stable, and improves the overall applicability and practical application capability of the present invention.

[0087] The beneficial effects of this invention compared to existing technologies include: by constructing a physical-data dual-driven feature engineering system, combining the spatial topology analysis capability of graph convolutional networks with the temporal evolution learning capability of long short-term memory networks, it can achieve synergistic characterization of the circumferential stress distribution, joint seepage resistance, and soil permeability characteristics of shaft structures. The spatiotemporal fusion feature tensor and spatiotemporal feature attention fusion layer proposed in this invention can not only accurately reflect the coupling relationship between the structural stress state and the seepage response at the model level, but also automatically identify and amplify key risk factors in construction disturbances. Simultaneously, the proposed dynamic feedback and weight update mechanism enables the prediction system to possess self-learning and evolutionary capabilities, allowing for continuous correction and optimization during the construction process, avoiding the accuracy decay problem of traditional static models. Through this complete prediction-feedback-optimization closed-loop system, this invention provides a new technical path for intelligent seepage prevention monitoring of ultra-deep shafts and can also provide a general methodological framework for leakage risk assessment of other complex underground engineering projects.

[0088] In summary, this invention provides a dynamic prediction method and system for weak points in ultra-deep circular vertical shaft foundation pits. It constructs a set of static structural parameters, a dataset of geological environment, and a dynamic dataset of construction procedures, and builds a multi-source heterogeneous data fusion matrix through spatiotemporal index mapping. A key feature extraction module filters physical and mechanical indicators, which are then processed by data preprocessing and encoding modules to construct a spatiotemporal fusion feature tensor targeting weak point location. Spatial feature extraction networks and temporal feature extraction networks run in parallel to analyze the circumferential topological features of the shaft and the evolution law of construction disturbances, respectively. Feature splicing and adaptive weighting are performed in the spatiotemporal feature attention fusion layer, and the result is input into a leakage risk prediction classifier to generate a probability distribution map. Based on the probability distribution, sensors are deployed in targeted areas of high-risk zones, and the predictive capability is iteratively evolved using a data feedback model and a weight update model. This invention is easy to understand, computationally simple, requires minimal workload, and is convenient for engineering applications, providing a theoretical foundation and technical support for the further development of the field of foundation pit engineering leakage early warning technology.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits, characterized in that, Includes the following steps: Acquire the set of static structural parameters, the dataset of geological environment, and the dynamic dataset of construction procedures, and construct a multi-source heterogeneous data fusion matrix through spatiotemporal index mapping; The spatiotemporal fusion feature tensor is obtained by extracting key features and preprocessing and encoding the multi-source heterogeneous data fusion matrix. Based on the spatiotemporal fusion feature tensor, the circumferential topological features of the shaft are obtained through a spatial feature extraction network, and the temporal features of construction disturbance are obtained through a temporal feature extraction network. The circumferential topological features of the shaft and the temporal features of construction disturbance are fused to obtain a feature fusion vector, and a probability distribution map of weak points in shaft leakage is generated by a leakage risk prediction classifier. Based on the probability distribution map of weak points in the vertical shaft leakage, sensors are deployed in high-risk areas to obtain real-time monitoring data, establish a data feedback model and a weight update model, and realize dynamic prediction of weak points in vertical shaft leakage.

2. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pit leakage according to claim 1, characterized in that, The acquisition of the structural static parameter set, the geological environment dataset, and the construction procedure dynamic dataset, and the construction of a multi-source heterogeneous data fusion matrix through spatiotemporal index mapping, includes: The set of static structural parameters includes reinforcement information, joint characteristics, and concrete strength grade of the ultra-deep circular vertical shaft foundation pit; The geological environment dataset includes geological borehole exploration data, soil permeability parameters, groundwater level changes, and geological structure information; The dynamic dataset of the construction process includes tunneling progress, grouting pressure, structural strain, and surrounding rock deformation; The spatiotemporal index mapping unifies the timestamps and spatial coordinates of the structural static parameter set, the geological environment dataset, and the construction process dynamic dataset to construct the multi-source heterogeneous data fusion matrix.

3. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pit leakage according to claim 1, characterized in that, The step of extracting key features and preprocessing and encoding data from the multi-source heterogeneous data fusion matrix to obtain a spatiotemporal fusion feature tensor includes: Obtain physical and mechanical indicators, and perform indicator screening and feature mapping based on the multi-source heterogeneous data fusion matrix to obtain key parameter factors; By combining the key parameter factors, the spatiotemporal fusion feature tensor is obtained by preprocessing and encoding the multi-source heterogeneous data fusion matrix.

4. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits according to claim 1, characterized in that, The process of obtaining the vertical shaft circumferential topological features through a spatial feature extraction network based on the spatiotemporal fusion feature tensor, and obtaining the construction disturbance temporal features through a temporal feature extraction network, includes: The spatial feature extraction network adopts a graph convolutional network architecture and obtains the circumferential topological features of the vertical shaft based on the spatiotemporal fusion feature tensor. The temporal feature extraction network adopts a long short-term memory network architecture and combines the spatiotemporal fusion feature tensor to obtain the temporal features of the construction disturbance.

5. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits according to claim 4, characterized in that, The process of obtaining the circumferential topological features of the shaft based on the spatiotemporal fusion feature tensor includes: The ultra-deep circular vertical shaft foundation pit is divided into several sector-shaped partitions along the circumferential direction, and the several sector-shaped partitions are used as input nodes of the graph convolutional network architecture. Adjacent partitions are determined through the aforementioned sector partitions, and an adjacency matrix is ​​established; Based on the adjacency matrix, the spatiotemporal fusion feature tensor is subjected to multi-layer convolution and aggregation through the graph convolutional network architecture to obtain the vertical shaft circumferential topological features.

6. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits according to claim 4, characterized in that, The process of obtaining the construction disturbance temporal features by combining the spatiotemporal fusion feature tensor includes: The spatiotemporal fusion feature tensor is input into the long short-term memory network architecture, and the construction disturbance temporal features are output using a recursive state update formula.

7. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits according to claim 1, characterized in that, The step of fusing the circumferential topological features of the shaft and the temporal features of construction disturbance to obtain a feature fusion vector, and generating a probability distribution map of weak points in the shaft leakage using a leakage risk prediction classifier, includes: The circumferential topological features of the vertical shaft and the temporal features of the construction disturbance are vector-concatenated to obtain a concatenated feature vector. An attention mechanism is introduced, and the concatenated feature vector is adaptively weighted to obtain the feature fusion vector. The feature fusion vector is input into the leakage risk prediction classifier to output the probability distribution map of the weak points of leakage in the vertical shaft.

8. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits according to claim 1, characterized in that, The step of deploying sensors in high-risk areas based on the probability distribution map of weak points in the vertical shaft leakage, acquiring real-time monitoring data, establishing a data feedback model and a weight update model, and realizing dynamic prediction of weak points in vertical shaft leakage includes: Based on the probability distribution map of weak points in the vertical shaft, the probability value of leakage risk is obtained, and the high-risk area is identified according to the probability value of leakage risk. The sensors are deployed in a targeted manner in the high-risk area to obtain the real-time monitoring data; The data feedback model is constructed by the difference between the real-time monitoring data and the probability distribution map of the weak points of the vertical shaft leakage, and the error feedback signal is obtained based on the data feedback model. Using the error feedback signal as a driving force, an adaptive optimization strategy is employed to establish the weight update model; An update strategy is constructed as a constraint, and online updates are performed through the data feedback model and the weight update model to achieve dynamic prediction of weak points in vertical shaft leakage.

9. The method for dynamic prediction of weak points in ultra-deep circular vertical shaft foundation pits according to claim 8, characterized in that, The online update via the data feedback model and the weight update model includes: The difference is corrected online using the error feedback signal; The weight parameters of the spatial feature extraction network and the temporal feature extraction network are updated online using the weight update model.

10. A dynamic prediction system for weak points in ultra-deep circular vertical shaft foundation pits, characterized in that, The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the dynamic prediction method for weak points in leakage of ultra-deep circular vertical shaft foundation pits as described in any one of claims 1-9.

Images