Deep foundation pit support structure damage identification method and system combined with deep learning

Abstract

The application discloses a deep foundation pit supporting structure damage identification method and system combined with deep learning, relates to the technical field of deep foundation pit engineering safety monitoring and structure damage identification, and comprises the following steps: collecting multi-source monitoring data and performing standardized processing, constructing a damage causal diagram model to extract causal characteristics, determining causes and tracing the initial damage position through a causal quantization algorithm; abstracting components into a time-varying topological dynamic graph, predicting damage evolution states by using a space-time attention network, establishing a normal benchmark grading early warning, and realizing closed-loop iteration through data retraining; the application fuses multi-source monitoring data and engineering knowledge to construct a damage causal diagram model, combines a time-varying topological dynamic graph and a space-time attention network, realizes causal tracing, accurate positioning and dynamic evolution prediction of deep foundation pit supporting structure damage, and through closed-loop iteration self-adaptive optimization, breaks through the limitations of traditional static methods, provides multi-scale fine analysis and grading early warning, and improves the monitoring scientificity and engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and structural damage identification technology for deep foundation pit engineering, specifically to a method and system for identifying damage to deep foundation pit support structures using deep learning. Background Technology

[0002] Deep foundation pit engineering is a key component of modern urban underground space development and building foundation construction. As the core structure that resists soil deformation and ensures the safety of foundation pit construction, the service condition of the support structure directly determines the safety of the foundation pit itself, surrounding buildings, and underground pipelines. With the continuous advancement of urban construction, deep foundation pit engineering is gradually showing characteristics of greater depth, wider span, and more complex geological and construction environments. During the entire construction process, the support structure will be subjected to the combined effects of multiple factors such as soil pressure, groundwater fluctuations, construction disturbances, and external loads, resulting in frequent structural damage problems and significantly increasing the difficulty of safety management. At present, the engineering field has built a basic monitoring system with the help of sensing equipment and Internet of Things technology, which can collect relevant data on the support structure and the surrounding environment. Deep learning technology is also gradually being applied in the field of structural condition analysis. The industry urgently needs a technical solution that can integrate multi-source data and take into account causal logic and dynamic evolution to solve the core problem of damage management of deep foundation pit support structures.

[0003] Traditional deep foundation pit support structure damage monitoring and identification technologies have many limitations and are difficult to adapt to the safety management needs of complex projects. Traditional methods mostly rely on single-type monitoring data for analysis, failing to achieve systematic processing of multi-source heterogeneous data. The data preprocessing process is imperfect, which can easily lead to data distortion or loss of effective information. Existing technologies mostly use experience-based judgment or simple data statistics to determine damage, without establishing an intrinsic correlation mechanism between influencing factors and structural damage. This makes it impossible to accurately locate the cause and initial location of damage, and the identification results lack scientific rigor and reliability. At the same time, traditional technologies adopt a static analysis mode, which cannot adjust the analysis logic in real time according to changes in the structural state. It is difficult to achieve multi-dimensional prediction of damage evolution, and the early warning mechanism lacks a unified benchmark, resulting in insufficient early warning accuracy and timeliness. In addition, traditional systems do not have continuous optimization capabilities, and the models and algorithms cannot adapt to different working conditions. The analysis effect continues to decline over long-term use, failing to meet the actual requirements of intelligent safety management. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a method and system for damage identification of deep foundation pit support structures that combines deep learning. Through standardized processing of multi-source heterogeneous monitoring data, damage causal graph modeling, and causal feature extraction, it accurately determines the causes of damage and traces the initial location. It constructs a time-varying topological dynamic graph and, relying on a spatiotemporal attention dynamic graph convolutional network and spatiotemporal evolution algorithm, completes the prediction of damage evolution states at multiple spatiotemporal scales. It establishes a hierarchical early warning mechanism and forms a data-driven closed-loop iterative system to continuously optimize model and algorithm parameters. This invention integrates causal analysis and deep learning technologies, overcoming the limitations of traditional methods such as static analysis, lack of causal correlation, and insufficient predictive ability. It significantly improves the accuracy and timeliness of damage identification and safety early warning, providing reliable technical support for the full-cycle intelligent safety management of deep foundation pit support structures and effectively ensuring the stability and safety of foundation pit construction and the surrounding environment.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a deep learning-based method for identifying damage in deep foundation pit support structures is provided, the specific steps of which are as follows:

[0006] S1. Data Acquisition and Processing: Sensors are deployed on the support piles, steel supports, anchor rods, walers, and surrounding soil and environment of the foundation pit to collect multi-source heterogeneous monitoring data. The data is aggregated and stored through the Internet of Things, and the data is deduplicated, range verified, normalized, missing value filled, and outlier processed to obtain a standardized monitoring dataset.

[0007] S2. Causal Modeling and Feature Extraction: Construct a damage causal graph model based on standardized datasets, engineering knowledge, and historical cases; extract causal features from it using a causal screening algorithm to form a causal feature set;

[0008] S3. Causal Analysis and Damage Source Tracing: Based on the damage causal graph model and causal feature set, spatial hierarchy is divided and causal intervention is carried out. The comprehensive causal contribution coefficient of each influencing factor is calculated using the causal quantification algorithm to determine the damage cause. The initial damage location is traced by combining the mechanical transmission relationship of the component.

[0009] S4. Dynamic graph construction and evolution prediction: The support components are abstracted as graph nodes, and the mechanical connections between components are abstracted as graph edges and assigned initial weights. The edge weights and topology are adjusted in real time according to the damage causes and initial damage locations obtained in step S3 to form a time-varying topological dynamic graph. Spatiotemporal attention dynamic graph convolutional network is used to extract spatiotemporal correlation features, and spatiotemporal evolution algorithm is used to predict the damage evolution state quantities at multiple time and spatial scales.

[0010] S5. Anomaly Warning and Closed-Loop Iteration: Based on historical working condition data, establish a normal damage evolution statistical benchmark using the model and algorithm constructed in steps S2 to S4, compare the damage evolution state quantity obtained in step S4 with the benchmark to classify the warning level; continuously collect field data, and periodically retrain the model and algorithm and update the parameters.

[0011] Furthermore, in S1, during data acquisition and processing, displacement and stress sensors are installed at the top of the support piles and at different depths within the pile body; stress and vibration sensors are installed at the two end joints and mid-span of the steel supports; tension sensors are installed at the anchor head of the anchor bolts; displacement and stress sensors are installed at the mid-span and node connections of the walers; earth pressure and water pressure sensors are installed inside the soil at different depths around the foundation pit; settlement and tilt sensors are installed at environmental points around the foundation pit; and temperature and construction load sensors are installed at the construction site. The sampling frequency of the displacement, stress, and tension sensors is set to once per minute; the sampling frequency of the earth pressure and water pressure sensors is set to once every five minutes; and the sampling frequency of the temperature and construction load sensors is set to once every ten minutes. When the monitored data fluctuates, the sampling frequency of the corresponding sensor and adjacent related sensors is switched to once per second.

[0012] Furthermore, in S1, the data acquisition and processing, the raw electrical signal output by the sensor is wirelessly transmitted using a combination of LoRa and NB-IoT transmission methods. For long-distance monitoring points, 4G or 5G networks are used as supplementary transmission links. All monitoring data is aggregated to a cloud-based distributed storage center via the MQTT communication protocol. The data deduplication stage removes duplicate data records corresponding to the same timestamp, the range verification stage removes data exceeding the sensor's calibrated range, the normalization stage uses minimum-maximum normalization to map the monitoring data to a unified numerical range, monitoring data with a missing duration of less than one hour is completed using linear interpolation, and monitoring data with a missing duration of ≥ one hour is completed using time-series prediction. The outlier handling stage removes abnormal monitoring data and replaces the data using the average of multiple neighboring valid monitoring data to obtain a standardized monitoring dataset.

[0013] Furthermore, in S2, causal modeling and feature extraction, the damage causal graph model sets up two types of nodes. The first type of node is the influencing factor node, which specifically includes soil strength, groundwater level, excavation speed, dewatering rate, construction load, ambient temperature, and surrounding traffic load. The second type of node is the damage node, which specifically includes support pile displacement, support pile cracking, steel support instability, steel support buckling, anchor bolt tension attenuation, anchor bolt detachment, waler deformation, waler fracture, pit heave, and pit leakage. Based on engineering experience, industry standards, and historical engineering cases, an initial causal graph structure is built. Directed edges are used to establish the association between different nodes. Using a standardized monitoring dataset, the number of connecting edges and the direction of directed edges in the initial causal graph are adjusted through a causal discovery algorithm. The probability parameters corresponding to each directed edge are calculated using the maximum likelihood estimation method. For discrete nodes, the conditional probability table corresponding to different combinations of parent nodes is calculated, and for continuous nodes, the mean and variance of the probability distribution are calculated.

[0014] Furthermore, in S3, causal analysis and damage tracing, the spatial hierarchy is divided into three levels: component level, local level, and overall level. The component level corresponds to a single support pile, a single steel support, a single anchor rod, or a single waler section. The local level corresponds to all support components within an independently divided area of ​​the foundation pit. The overall level corresponds to the complete deep foundation pit support system. The causal intervention operation adopts a single variable adjustment mode, keeping the values ​​of all remaining influencing factors fixed, adjusting the value of a single influencing factor individually, and simultaneously collecting the support structure status data corresponding to the three spatial levels: component level, local level, and overall level.

[0015] Furthermore, in S3, causal analysis and damage tracing, when the support structure is monitored to produce an abnormal state, the abnormal state information is input into the damage causal graph model. The posterior probability values ​​corresponding to all influencing factor nodes are obtained through Bayesian inference. The posterior probability values ​​and the comprehensive causal contribution coefficient are combined to complete the sorting process. The damage cause is determined based on the sorting results. The damage propagation path is constructed based on the fixed mechanical transmission relationship between the support components. The abnormal position obtained by monitoring is used as the starting point for deduction. The initial damage position is located by deducing backward layer by layer along the damage propagation path.

[0016] Furthermore, in S4, the dynamic graph construction and evolution prediction, a single support pile, a single steel support, a single anchor rod, and a single waler section correspond to independent graph nodes in the dynamic graph, and the mechanical connection relationship between support components corresponds to graph edges in the dynamic graph. Each graph node is configured with an independent multidimensional feature vector, which consists of three types of data: causal features, component geometric parameters, and component material properties. The initial weight of the graph edges is assigned based on the mechanical connection strength of the components. The weight parameters of the graph edges are adjusted based on the real-time structural state corresponding to the damage cause and the initial damage location, combined with the change in the force transmission coefficient between components. When the mechanical connection of a component fails, all graph edges between the component and its adjacent components are deleted, thus completing the dynamic graph topology update.

[0017] Furthermore, in S4, dynamic graph construction and evolution prediction, the spatiotemporal attention dynamic graph convolutional network is composed of multiple layers of spatiotemporal attention blocks stacked together. Each spatiotemporal attention block is sequentially set with a temporal attention layer, a spatial attention layer, and a dynamic graph convolutional layer. The temporal attention layer completes the temporal data operation and extracts the correlation features of different time nodes. The spatial attention layer completes the graph node data operation and extracts the correlation features of different support components. The dynamic graph convolutional layer combines the real-time updated dynamic graph topology to extract the information transmission features between components. The dynamic graph node feature sequence is input into the spatiotemporal attention dynamic graph convolutional network, and after the feature operation is completed layer by layer, the spatiotemporal correlation features are output.

[0018] Furthermore, in S5, the anomaly warning and closed-loop iteration, monitoring data and structural status data corresponding to historical normal operating conditions are collected, and a multivariate Gaussian statistical distribution is constructed in combination with the damage causal graph model. The multivariate Gaussian statistical distribution serves as the statistical benchmark for normal damage evolution. The deviation between the damage evolution state quantity and the statistical benchmark for normal damage evolution is calculated. Combining the damage induction and comprehensive causal contribution coefficient, a comprehensive anomaly index is calculated by weighted summation. Based on the numerical range of the comprehensive anomaly index, three warning levels are divided: the comprehensive anomaly index below the first threshold corresponds to the first warning level; the comprehensive anomaly index between the first and second thresholds corresponds to the second warning level; and the comprehensive anomaly index above the second threshold corresponds to the third warning level.

[0019] On the other hand, a deep learning-based damage identification system for deep foundation pit support structures includes: a data acquisition and preprocessing module, a causal modeling and feature extraction module, a causal analysis and damage tracing module, a dynamic graph construction and evolution prediction module, and an early warning and closed-loop iteration module.

[0020] Data acquisition and preprocessing module: Deploys multiple types of sensors to collect multi-source monitoring data, transmits it wirelessly and stores it in the cloud, completes data deduplication, verification, normalization, missing value and outlier processing, and outputs a standardized monitoring dataset;

[0021] Causal modeling and feature extraction module: Constructs a damage causal graph model containing influencing factors and damage nodes, optimizes the model structure and parameters, and extracts causal feature sets through causal screening algorithms;

[0022] Causal analysis and damage tracing module: Divide the spatial hierarchy into component level, local level and global level and perform causal intervention, calculate the comprehensive causal contribution coefficient to determine the damage cause, and locate the initial damage location by combining the mechanical transmission relationship;

[0023] The dynamic graph construction and evolution prediction module abstracts support components and mechanical connections into dynamic graph nodes and edges and updates the topology in real time. It extracts features through a spatiotemporal attention dynamic graph convolutional network and predicts damage evolution state quantities through a spatiotemporal evolution algorithm.

[0024] Early warning and closed-loop iteration module: Establish a normal damage evolution benchmark, compare data to classify early warning levels; continuously collect field data to update the knowledge base, and periodically retrain the model and algorithm and update parameters.

[0025] Compared with existing technologies, this method and system for damage identification of deep foundation pit support structures, which combines deep learning, has the following advantages:

[0026] I. This invention utilizes standardized acquisition and processing of multi-source heterogeneous monitoring data, combined with engineering knowledge and historical cases to construct a damage causal graph model. It relies on causal screening and quantification algorithms to complete feature extraction, cause determination, and initial damage location tracing. It divides the data into multiple spatial levels for causal intervention analysis, abandoning the limitations of traditional methods that rely solely on experience. This makes damage identification and tracing more aligned with engineering mechanics logic, accurately distinguishing the weight of different influencing factors, and solving the identification bias problem caused by neglecting causal relationships in traditional data-driven methods. It improves the accuracy of damage cause determination and location positioning, enabling multi-scale, refined analysis of deep foundation pit support structure damage. This makes the monitoring results more scientific and engineering-practical, providing a reliable basis for structural safety diagnosis.

[0027] II. This invention abstracts the support components and mechanical connections into a time-varying topological dynamic graph, combines it with a spatiotemporal attention dynamic graph convolutional network to extract the spatiotemporal correlation features of the structure, and relies on a spatiotemporal evolution algorithm to predict the damage evolution state at multiple spatiotemporal scales. It dynamically adapts to real-time changes in the structure's state, overcomes the shortcomings of traditional static models that cannot respond to the dynamic development of structural damage, establishes a normal evolution statistical benchmark to achieve graded anomaly early warning, and continuously improves the system's adaptive capability and prediction accuracy by continuously collecting field data, periodically retraining the model and updating algorithm parameters to form a closed-loop iteration. This enables early warning and dynamic control of damage, effectively avoids the safety risks of deep foundation pit construction, ensures the safety of the support structure throughout its entire life cycle, and enhances the intelligence and long-term effectiveness of engineering monitoring.

[0028] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0030] Figure 1 A flowchart of a deep foundation pit support structure damage identification method incorporating deep learning;

[0031] Figure 2 A framework diagram of a deep foundation pit support structure damage identification system incorporating deep learning;

[0032] Figure 3 This diagram illustrates the data transmission between steps in a deep foundation pit support structure damage identification method incorporating deep learning. Detailed Implementation

[0033] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0034] Reference Figure 1 One embodiment of the present invention proposes a deep learning-based method for identifying damage to deep foundation pit support structures. This method employs a complete technical process including multi-source heterogeneous monitoring data acquisition and processing, damage causal graph modeling and feature selection, multi-spatial hierarchical causal intervention analysis, time-varying topological dynamic graph convolutional feature extraction, multi-scale damage evolution prediction, hierarchical anomaly early warning, and periodic iterative optimization of the model. This method enables accurate identification of support structure damage, quantitative determination of causes, initial location tracing, and evolution trend prediction throughout the entire process of deep foundation pit excavation and support construction. It significantly improves the intelligence level and early warning reliability of deep foundation pit support safety monitoring, and reduces the safety risks of foundation pit construction.

[0035] The method described in this embodiment specifically includes:

[0036] S1. Data Acquisition and Processing: Multiple types of sensors are deployed on key components such as deep foundation pit support piles, steel supports, anchor bolts, and walers, as well as the surrounding soil and environment. These sensors collect multi-source heterogeneous monitoring data, including displacement, stress, vibration, tension, earth pressure, water pressure, settlement, tilt, temperature, and construction load. The data is then uploaded to a cloud-based distributed storage center via multi-mode wireless transmission and supplementary network links. After data deduplication, range verification, minimum-maximum normalization, rule-based missing value filling, and 3σ principle outlier processing, a continuous, regular, and effective standardized monitoring dataset is obtained.

[0037] S2. Causal Modeling and Feature Extraction: Based on engineering experience, industry standards and historical engineering cases, a damage causal graph model containing influencing factor nodes and damage nodes is built. The directed edge structure and orientation of the model are optimized through causal discovery algorithms. Maximum likelihood estimation is used to complete the calibration of the model probability parameters. The standardized monitoring dataset is input into the feature extraction network. The fixed-dimensional causal feature set after removing confounding variables is extracted with the causal screening algorithm as the iterative target.

[0038] S3. Causal Analysis and Damage Source Tracing: The deep foundation pit support system is divided into three spatial levels: component level, local level, and overall level. A single variable adjustment method is used to perform causal intervention. Structural state data of each level are collected. The comprehensive causal contribution coefficient of each influencing factor is calculated by causal quantification algorithm. The posterior probability of the influencing factors is obtained by combining Bayesian inference and ranked to determine the damage inducing factors. The damage propagation path is constructed based on the mechanical transmission relationship of the components, and the initial damage location is located by reverse deduction.

[0039] S4. Dynamic Graph Construction and Evolution Prediction: Single support piles, single steel supports, single anchor rods, and single walers are abstracted as nodes in the dynamic graph, and the mechanical connections between components are abstracted as edges and assigned initial weights. The edge weights are adjusted in real time according to the damage induction and the initial damage location, and the topology is updated to form a time-varying topological dynamic graph. Temporal and spatial correlation features are extracted through a multi-layer stacked spatiotemporal attention dynamic graph convolutional network, and spatiotemporal evolution algorithms are used to calculate the damage evolution state quantities at multiple time scales and multiple spatial scales.

[0040] S5. Anomaly Early Warning and Closed-Loop Iteration: Based on historical normal operating condition data, a multivariate Gaussian statistical distribution is constructed as the statistical benchmark for normal damage evolution. The deviation between the damage evolution state quantity and the benchmark is calculated. A comprehensive anomaly index is obtained by weighting the damage inducing factors and the comprehensive causal contribution coefficient. Three warning levels are defined based on the index value range. On-site damage data, structural state data, and treatment data are continuously collected and entered into the damage causal knowledge base. Using a fixed cycle of natural months, the latest data is used to retrain the damage causal graph model and the spatiotemporal attention dynamic graph convolutional network, synchronously updating the algorithm parameters and network structure. Figure 3 As shown.

[0041] Specific implementation details of the method:

[0042] S1. Acquisition and Standardization Processing of Multi-Source Heterogeneous Monitoring Data:

[0043] Specifically, in S1, the multi-source heterogeneous monitoring data acquisition and standardization processing, the sensor deployment is precisely arranged according to the stress characteristics of the support components and the monitoring requirements:

[0044] Support piles: Laser displacement sensors are installed at the top of the piles, and fiber optic stress sensors are installed every 2m depth in the pile body to collect real-time pile displacement and stress changes.

[0045] Steel supports: Vibrating wire stress sensors and piezoelectric vibration sensors are installed at the joints connecting the two ends to the walers and at the mid-span of the supports to monitor the stress and vibration amplitude of the supports;

[0046] Anchor bolts: Axial force tension sensors are installed at the connection point between the anchor head and the retaining structure to collect real-time tension values ​​of the anchor bolts;

[0047] Waler: Displacement sensors are installed at the mid-span and stress sensors are installed at the node connection points to monitor the deformation and stress state of the waler;

[0048] Soil surrounding the foundation pit: Earth pressure sensors and seepage water pressure sensors are installed every 3m depth to monitor changes in soil pressure and groundwater level.

[0049] Surrounding environment of the foundation pit: Settlement and tilt sensors are installed at building and road locations to monitor the deformation of surrounding structures;

[0050] At the construction site: Digital temperature sensors and load sensors are deployed to collect data on ambient temperature and the loads of construction machinery and slabs.

[0051] The sensor sampling frequency adopts a hierarchical adaptive setting mode: for devices that reflect the instantaneous state of the structure, such as displacement sensors, stress sensors, and tension sensors, the basic sampling frequency is set to 1 time / minute; for soil parameter devices such as earth pressure sensors and water pressure sensors, the basic sampling frequency is set to 1 time / 5 minutes; for environmental parameter devices such as temperature sensors and construction load sensors, the basic sampling frequency is set to 1 time / 10 minutes; when the fluctuation range of any sensor monitoring data exceeds 10%, the sampling frequency of the sensor and related sensors within a 3m radius around it will automatically switch to 1 time / second to fully capture the details of abnormal dynamic changes in the structure.

[0052] Data transmission adopts a multi-link redundancy guarantee scheme: the original electrical signals of the sensors are preferentially transmitted through a combination of LoRa and NB-IoT. For long-distance monitoring points more than 50m away from the edge of the foundation pit, 4G / 5G network is automatically activated as a supplementary transmission link. All monitoring data is stably uploaded to the cloud distributed storage center in the form of data packets through the MQTT communication protocol. The latency of single packet data transmission is controlled within 200ms, and the packet loss rate is less than 0.1%.

[0053] Data standardization processing includes five refined processing steps:

[0054] Data deduplication: Traverse all monitoring data, remove duplicate data records corresponding to the same timestamp and the same sensor number, and retain unique and valid data;

[0055] Range calibration: Compare with the sensor calibration parameters and filter out invalid data that exceeds the measurement range. For example, for a stress sensor with a range of 0-1000kN, data collected at 1200kN should be directly rejected.

[0056] Normalization: The minimum-maximum normalization method is used to map monitoring data with different dimensions and different numerical ranges to a unified interval of [0, 1] to eliminate the influence of dimension differences on model calculation;

[0057] Missing value imputation: Monitoring data with a missing duration of less than one hour are imputed using linear interpolation, while monitoring data with a missing duration of one hour or more are imputed using time-series prediction.

[0058] Outlier handling: The 3σ principle is used to identify outlier data, and outliers that deviate from the mean by 3 times the standard deviation are removed. The outliers are then replaced with the arithmetic mean of 10 consecutive valid data sets to ensure the continuity and validity of the data sequence.

[0059] For example, when carrying out data acquisition and processing operations at a deep foundation pit construction site, all sensors are deployed and debugged according to a preset plan to continuously collect various monitoring data of the support structure and surrounding environment; deduplication is performed on all collected data, eliminating duplicate data generated by the same sensor at the same time stamp, retaining only a single valid data record; range verification is performed to screen out all invalid data that exceeds the sensor's calibrated measurement range and exclude it from subsequent processing; min-max normalization is used to transform all valid data to unify different types of monitoring data into the same numerical range; for missing segments in the data sequence, linear interpolation or time series prediction is used to complete them according to the missing duration to ensure the integrity and continuity of the data sequence; abnormal values ​​in the data are identified and eliminated according to preset rules, and the average value of multiple adjacent valid data sets is used to replace the abnormal positions, finally obtaining a standardized monitoring dataset that is time-continuous, free of redundant information, and free of abnormal values, ensuring that the data meets all the requirements of subsequent model calculations.

[0060] S2. Damage Causal Modeling and Causal Feature Extraction:

[0061] Specifically, in S2, damage causal modeling and causal feature extraction, the damage causal graph model sets two independent node types: the first type is influencing factor nodes, with a total of 7 types, including soil strength, groundwater level, excavation speed, dewatering rate, construction load, ambient temperature, and surrounding traffic load; the second type is damage nodes, with a total of 10 types, including support pile displacement, support pile cracking, steel support instability, steel support buckling, anchor bolt tension attenuation, anchor bolt detachment, waler deformation, waler fracture, pit heave, and pit leakage.

[0062] The initial causal graph structure is built based on engineering experience, historical deep foundation pit engineering cases, and geotechnical engineering experience. Directed edges are used to represent the causal relationships between nodes. The standardized monitoring dataset is input into the PC causal discovery algorithm to automatically adjust the number of directed edges, connection relationships, and directions of the initial causal graph, thereby optimizing the rationality of the model structure. The maximum likelihood estimation method is used to calculate the probability parameters of each directed edge. Discrete nodes generate conditional probability tables for different combinations of parent nodes, and continuous nodes calculate the mean and variance of the probability distribution to complete the accurate calibration of model parameters.

[0063] Causal feature extraction employs a causal screening algorithm, the mathematical expression of which is:

[0064]

[0065] in, To standardize the monitoring dataset, For causal feature set, For feature reconstruction mapping function, - For feature reconstruction constraints, For the set of confusing variables, For mutually exclusive causal constraints, These are adaptive weighting coefficients that dynamically adjust according to the distribution of data features; adaptive weighting coefficients The mathematical expression for calculation is: ; The basic adjustment coefficient is used to determine... The initial value of α is the baseline, δ is the decay adjustment factor used to control the decay rate of α as the variance of the monitoring data changes, Var(X) is the global variance of the standardized monitoring dataset X, used to characterize the overall dispersion of the monitoring data; feature reconstruction mapping function. The mathematical expression for calculation is: ; The weight matrix is ​​reconstructed for the features, which is used to realize the mapping transformation from causal features to the monitoring data space. The bias vector is reconstructed for the features to correct the offset error in the mapping calculation, and the causal mutual exclusion constraint term is used. The mathematical expression for calculation is: ;cov is the covariance calculation operator. For the first in the causal feature set Each feature component For the first in the set of confounding variables The mutual exclusivity of the two variables is characterized by the sum of the absolute values ​​of their covariances.

[0066] The execution flow of the causal screening algorithm is as follows: the standardized monitoring dataset is input into the fully connected feature extraction network, the above mathematical expression is used as the iterative optimization target, more than 99% of the effective information in the monitoring data is retained through feature reconstruction constraints, and irrelevant and confusing variables such as soil moisture content, construction lighting, and wind speed are eliminated through causal mutual exclusion constraints. After iterative calculation, a fixed-dimensional causal feature set is output to provide accurate data input for subsequent causal analysis.

[0067] For example, the standardized monitoring dataset obtained from the construction site is input into the feature extraction network to initiate the iterative computation process of the causal screening algorithm. During the computation, the feature reconstruction constraint term continuously retains all valid structural information in the monitoring data to avoid the loss of key data, and the causal mutual exclusion constraint term automatically identifies and removes confounding variables that are not related to the causal relationship of damage to reduce the interference of irrelevant information on feature extraction. After a complete iterative computation, a causal feature set with fixed dimensions is output. This feature set covers all core influencing factors such as the stress state of the support structure, soil physical parameters, and on-site environmental loads, achieving reasonable simplification of feature dimensions and effectively improving the efficiency and accuracy of subsequent model computation.

[0068] S3. Multi-spatial hierarchical causal analysis and damage tracing:

[0069] Specifically, in S3, multi-spatial-level causal analysis and damage tracing, the spatial levels are divided into three categories according to the scope of the foundation pit support and the attribution of components:

[0070] Component level: corresponding to a single support pile, a single steel support, a single anchor bolt, or a single waler section, which is the smallest monitoring and analysis unit;

[0071] Local level: All support components within an independent area of ​​the foundation pit divided into construction sections;

[0072] Integral level: a complete support system corresponding to the entire deep foundation pit.

[0073] The causal intervention operation adopts a single variable control mode: keeping the values ​​of other influencing factors fixed, adjusting the value of a single influencing factor, and simultaneously collecting state data such as displacement, stress, and tension of the support structure at three levels: component level, local level, and overall level, to establish a quantitative correlation between the single factor and structural damage.

[0074] When the monitoring system identifies an abnormal state of the support structure, it inputs information such as the abnormal location, abnormal type, and abnormal value into the damage causal graph model. Through Bayesian inference, it obtains the posterior probability values ​​of all influencing factor nodes. The posterior probabilities are combined with the comprehensive causal contribution coefficient and ranked. The influencing factor ranked first is the dominant damage cause.

[0075] The comprehensive causal contribution coefficient is calculated using a causal quantification algorithm, and its mathematical expression is as follows:

[0076]

[0077] in, For the first The comprehensive causal contribution coefficient of the class of influencing factors For the first A subset of causal characteristics of influencing factors. For the first Similar to support structure damage events, The feature-damage causal correlation function, The total number of influencing factors. This represents the total number of damage events. For the first Spatial hierarchy of the effects of various influencing factors This is a spatial scale correction term. This is a dynamic weighting coefficient, with a value range of [value range missing]. Dynamic weighting coefficient The mathematical expression for calculation is: ; This is the proportional adjustment coefficient, used for control. The magnitude of the numerical change, For the first The main spatial hierarchy of influencing factors The hierarchical dispersion is used to reflect the distributional differences in spatial hierarchy. This is an offset correction factor used to ensure... The value of is always in Within the effective interval; Feature-damage causal correlation function The mathematical expression for calculation is: ; The L2 norm operator is used to measure the numerical distance between a feature and a damage event. The correlation smoothing coefficient is used to adjust the numerical smoothness of the correlation function, and is also a spatial scale correction term. The mathematical expression for calculation is: ; This is an indicator function that takes the value when the condition is met. When the condition is not met, the value is taken as follows: , , , These are the hierarchical weight coefficients for the component level, local level, and global level, respectively, and the sum of the three is always equal to... , used to distinguish the correction weights of different spatial levels.

[0078] Causal quantification algorithm execution flow: Import causal feature set and causal intervention data, classify feature subsets according to influencing factors, calculate and normalize the correlation degree of each factor to all damage events through causal correlation degree function, complete hierarchical calculation by combining spatial scale correction term, and output the comprehensive causal contribution coefficient of all influencing factors.

[0079] Damage tracing process: Based on the fixed mechanical transmission relationship of support piles-walers-steel supports-anchors, a damage propagation path is constructed from top to bottom; taking the abnormal location monitored on site as the starting point for deduction, the stress and deformation data of the components are checked layer by layer along the propagation path in reverse, gradually narrowing the scope, and finally locating the specific component and location of the initial damage.

[0080] For example, the construction site monitoring system captures the state of continuous abnormal stress changes in the steel support and inputs the complete information of this structural anomaly into the damage causal graph model. Through Bayesian inference, the posterior probability values ​​corresponding to all influencing factor nodes are obtained. After sorting the posterior probabilities, the primary and secondary influencing factors are determined. The causal feature set and the data collected by causal intervention are imported, and the calculation is completed through the causal quantification algorithm to obtain the comprehensive causal contribution coefficient of various influencing factors. Combining the coefficient values ​​with the posterior probability ranking results, the dominant inducing factor causing structural damage is determined. Taking the location of the abnormal component monitored as the starting point of the deduction, the damage propagation path is constructed according to the established mechanical transmission relationship between the support components. The stress and deformation state of adjacent components are checked layer by layer along the path, gradually narrowing the scope of investigation. Finally, the specific component and connection location where the initial damage occurred are accurately determined, and the direct cause of the damage is clarified.

[0081] S4. Construction of Time-Varying Dynamic Graphs and Prediction of Damage Evolution:

[0082] Specifically, in S4, the construction of time-varying dynamic diagrams and the prediction of damage evolution, the basic construction rules for dynamic diagrams are as follows: each support pile, steel support, anchor rod, and waler on site is treated as an independent diagram node, and the mechanical connection relationship between components is treated as a diagram edge.

[0083] Each graph node is configured with a three-dimensional feature vector: the first dimension is the causal feature, the second dimension is the component's geometric parameters, and the third dimension is the component's material properties, comprehensively characterizing the physical state and stress characteristics of the node's components. The initial weights of the graph edges are assigned based on the mechanical connection strength of the components; the higher the connection strength, the greater the weight. The graph edge weights are adjusted in real time based on the damage induction cause and the initial damage location, combined with changes in the force transmission coefficient. If a component connection fails, all graph edges of that component and its adjacent components are directly deleted, completing the real-time update of the dynamic graph topology.

[0084] Spatiotemporal attention dynamic graph convolutional networks consist of multiple stacked spatiotemporal attention blocks. Each spatiotemporal attention block contains, in sequence, a temporal attention layer, a spatial attention layer, and a dynamic graph convolutional layer.

[0085] Temporal attention layer: performs weighted operations on the temporal features of nodes to extract structural change correlation features at different time nodes;

[0086] Spatial attention layer: performs weighted operations on the spatial features of nodes to extract the force correlation features between different components;

[0087] Dynamic graph convolutional layer: Combines real-time updated dynamic graph topology to extract force transmission and damage propagation features between components. The feature sequence of dynamic graph nodes is input into the network, and spatiotemporal correlation features are output after layer-by-layer computation.

[0088] The damage evolution state variables are calculated using a spatiotemporal evolution algorithm, and their mathematical expression is as follows:

[0089]

[0090] in, for Time and space Internal damage evolution state quantity, The region of the previous time step The state of damage, Let be the time evolution function. For causal feature set, For the region The set of adjacent regions, This represents the number of adjacent regions. Adjacent area The damage status at the current moment, For spatial topology and causal discontinuity functions The comprehensive causal contribution coefficient corresponding to the identified injury causes. The adaptive weighting coefficients are dynamically adjusted based on the current stage of damage evolution, and their values ​​range from [value range missing]. Adaptive weighting coefficients The mathematical expression for calculation is: ; This is the amplitude adjustment coefficient, used to limit... Adjustment range The damage state at the previous moment The evolution rate is used to reflect how quickly structural damage develops. The benchmark compensation coefficient is used to ensure The value of conforms to Usage requirements; time evolution function The mathematical expression for calculation is: ;in , These are the temporal evolution weighting coefficients, used to adjust the influence weights of the damage state and causal discontinuity at the previous time step on the temporal evolution result, respectively, along with the spatial topology and causal discontinuity function. The mathematical expression for calculation is: ; This is a spatial constraint coefficient used to regulate the coupling strength between the damage state of adjacent regions and the comprehensive causal contribution coefficient. The damage status is in the adjacent area. The comprehensive causal contribution coefficient corresponding to the injury inducing factors.

[0091] The execution flow of the spatiotemporal evolution algorithm is as follows: import the dynamic graph node features, historical damage states, causal feature sets, and comprehensive causal contribution coefficients; determine the spatial adjacency relationship according to the dynamic graph topology; calculate the temporal evolution term and the spatial neighborhood aggregation term respectively; fuse the two results through adaptive weight coefficients; and output the damage evolution state quantity at multiple time scales and multiple spatial scales.

[0092] For example, the dynamic graph node features corresponding to all components of the deep foundation pit support system are input into a spatiotemporal attention dynamic graph convolutional network. The operation process of temporal attention, spatial attention, and dynamic graph convolution is completed sequentially according to the network hierarchy. The temporal correlation features and spatial correlation features in the data are extracted layer by layer to obtain complete spatiotemporal correlation features. The extracted spatiotemporal correlation features, historical damage state data, causal feature set, and comprehensive causal contribution coefficient are imported into a spatiotemporal evolution algorithm. According to the algorithm process, the temporal evolution term operation and spatial neighborhood aggregation operation are completed respectively. The fusion calculation of the two operation results is completed through adaptive weight coefficients. Finally, the damage evolution state quantities corresponding to different time dimensions and different spatial dimensions are output, which clearly reflects the future development trend and propagation law of the support structure damage and clarifies the difference in damage development rate of different components.

[0093] S5. Damage Anomaly Early Warning and Model Closed-Loop Iteration:

[0094] Specifically, in S5, damage anomaly early warning and model closed-loop iteration, the construction of the statistical benchmark for normal damage evolution is as follows: the monitoring data and structural status data of historical normal construction conditions are collected, and a multivariate Gaussian statistical distribution is constructed in combination with the damage causal graph model. This distribution is used as the benchmark to measure the degree of structural damage anomaly, and the benchmark covers the range of structural status fluctuations under all normal working conditions.

[0095] Calculation of the comprehensive anomaly index: First, the deviation between the damage evolution state quantity and the multivariate Gaussian statistical benchmark is calculated. Then, the comprehensive causal contribution coefficients of the damage causes are weighted and summed to obtain the comprehensive anomaly index. The first and second thresholds are calculated through the statistical distribution of the comprehensive anomaly index under historical normal operating conditions. The mathematical expression for the calculation is as follows: , ;in The first threshold, The second threshold, This is the average of the comprehensive anomaly index under historical normal operating conditions, used to characterize the central trend of the comprehensive anomaly index under normal conditions. The standard deviation of the comprehensive anomaly index under historical normal working conditions is used to characterize the dispersion of the comprehensive anomaly index under normal conditions. The warning level is set according to the calculated threshold: Level 1 warning: The comprehensive anomaly index is lower than the first threshold, the structural condition is slightly abnormal, and there is no safety risk; Level 2 warning: The comprehensive anomaly index is in the range between the first and second thresholds, the structural condition is obviously abnormal, and monitoring needs to be strengthened; Level 3 warning: The comprehensive anomaly index is higher than the second threshold, the structural condition is seriously abnormal, and the construction emergency plan should be activated immediately.

[0096] The system's closed-loop iterative process involves: real-time collection of damage location data, structural status data, and on-site treatment data from the construction site; cleaning, labeling, and classifying the collected data; and simultaneously entering new damage cases, causal relationships of components, damage causes, and treatment results into the damage causality knowledge base, thus dynamically expanding the knowledge base; using a fixed iteration cycle of natural months, retrieving the latest standardized monitoring dataset from the cloud-based distributed storage center, and simultaneously retrieving all new case data from the damage causality knowledge base; fusing the two types of data as training samples to sequentially complete the damage causality graph model and the spatiotemporal attention dynamic graph convolutional network. The model undergoes retraining, during which the stochastic gradient descent algorithm is used to optimize the model loss function. Based on the model output results and accuracy verification metrics after retraining, the weight parameters of the causal screening algorithm, causal quantization algorithm, and spatiotemporal evolution algorithm are updated. Simultaneously, based on the training convergence speed and prediction accuracy, the number of network layers, number of neurons, learning rate, and regularization coefficient of the spatiotemporal attention dynamic graph convolutional network are adjusted. The updated model parameters, algorithm parameters, and network structure are then deployed to the on-site monitoring terminal to complete a full closed-loop iteration, continuously improving the accuracy of model damage identification, source tracing, prediction, and early warning. For example, after constructing a statistical benchmark for normal damage evolution based on historical normal working condition data, the comprehensive anomaly index sequence under historical normal working conditions is extracted, and the mean and standard deviation of the sequence are calculated. The first threshold and the second threshold are obtained through the threshold calculation formula. The deviation between the current damage evolution state and the statistical benchmark is calculated, and the comprehensive causal contribution coefficient corresponding to the damage cause is combined to complete the weighted summation to obtain the comprehensive anomaly index. The warning level is determined according to the numerical range of the comprehensive anomaly index. When the index reaches the highest level warning standard, the system immediately pushes the warning information to the on-site construction management personnel to guide the on-site structural inspection and construction adjustment work. The system continuously collects damage-related data generated at the construction site. After data cleaning and labeling, the newly added damage cases and causal association information are synchronously entered into the damage causal knowledge base. According to a fixed cycle of natural months, the system calls the full data of the damage causal knowledge base and the latest standardized monitoring dataset to retrain the damage causal graph model and the spatiotemporal attention dynamic graph convolutional network. Based on the retraining results, the weight parameters of various algorithms are updated, the structural parameters of the deep learning network are adjusted, and the optimized model and algorithm are deployed to the on-site terminal to complete the closed-loop iteration and continuously improve the damage identification and warning performance of the system.

[0097] System Implementation Example:

[0098] Based on the same inventive concept, such as Figure 2As shown, the present invention also provides a deep learning-based damage identification system for deep foundation pit support structures. The system is used to execute the deep learning-based damage identification method for deep foundation pit support structures described above. The system includes: a data acquisition and preprocessing module, a causal modeling and feature extraction module, a causal analysis and damage tracing module, a dynamic graph construction and evolution prediction module, and an early warning and closed-loop iteration module.

[0099] Data Acquisition and Preprocessing Module: This module is used to deploy displacement, stress, tension, earth pressure, water pressure, settlement, tilt, temperature, and construction load sensors on deep foundation pit support components, surrounding soil, and environmental areas. It collects multi-source heterogeneous monitoring data at graded adaptive frequencies. Through a combination of LoRa, NB-IoT, and 4G / 5G transmission links, and relying on the MQTT protocol, the data is uploaded to a cloud-based distributed storage center. The module sequentially performs data deduplication, range verification, minimum-maximum normalization, rule-based missing value filling, and 3σ outlier processing to output a highly effective, continuous, and standardized monitoring dataset, providing data support for subsequent analysis.

[0100] Causal Modeling and Feature Extraction Module: This module is used to build an initial damage causal graph containing influencing factor nodes and damage nodes based on engineering experience and historical cases. It optimizes the directed edge structure of the model through PC causal discovery algorithm and completes the calibration of model probability parameters using maximum likelihood estimation. The standardized monitoring dataset is input into a fully connected feature extraction network. With the causal screening algorithm as the iterative target, after removing confounding variables, a fixed-dimensional causal feature set is output, completing the initial construction of the causal model and features.

[0101] Causal Analysis and Damage Source Tracing Module: This module is used to divide the support system into three spatial levels: component level, local level, and overall level. It implements causal intervention using a single variable adjustment mode and collects structural state data at each level. It calculates the comprehensive causal contribution coefficient of each influencing factor through a causal quantification algorithm and determines the damage inducing factors by combining Bayesian inference ranking. Based on the mechanical transmission relationship of components, it constructs the damage propagation path and reverse-engineers from the abnormal location to accurately locate the specific component and location of the initial damage.

[0102] Dynamic graph construction and evolution prediction module: This module is used to abstract support components as graph nodes and mechanical connections as graph edges. It adjusts the graph edge weights and topology based on the real-time damage state to build a time-varying dynamic topology graph. The three-dimensional feature sequence of the nodes is input into a multi-layer spatiotemporal attention dynamic graph convolutional network to extract temporal and spatial correlation features. The spatiotemporal evolution algorithm is then imported to calculate the damage evolution state quantities at multiple time scales and multiple spatial scales, and the damage evolution prediction results are output.

[0103] Early warning and closed-loop iteration module: This module is used to construct a multivariate Gaussian statistical distribution based on historical normal working condition data as a statistical benchmark for normal damage evolution; calculate the deviation between the damage evolution state quantity and the benchmark, weight it to obtain a comprehensive anomaly index, divide it into three levels of early warning according to the numerical range and push early warning information; continuously collect on-site damage data, structural state data, and treatment data to update the damage causal knowledge base; and use the latest data to retrain the model and network on a natural month cycle, update the algorithm parameters and network structure, and realize continuous iterative optimization of the system.

[0104] It should be noted that the data transmission and logical connections between the various modules described above do not necessarily represent direct data interaction. Indirect transmission, interface calls, cloud data sharing, and other methods can be applied to the embodiments of this invention as long as they achieve the technical objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for damage identification of deep foundation pit support structures combining deep learning, characterized in that, The specific steps of this method are as follows: S1. Data Acquisition and Processing: Sensors are deployed on the support piles, steel supports, anchor rods, walers, and surrounding soil and environment of the foundation pit to collect multi-source heterogeneous monitoring data. The data is aggregated and stored through the Internet of Things, and the data is deduplicated, range verified, normalized, missing value filled, and outlier processed to obtain a standardized monitoring dataset. S2. Causal Modeling and Feature Extraction: Construct a damage causal graph model based on standardized datasets, engineering knowledge, and historical cases; extract causal features from it using a causal screening algorithm to form a causal feature set; The causal feature extraction employs a causal screening algorithm, the mathematical expression of which is: in, To standardize the monitoring dataset, For causal feature set, For feature reconstruction mapping function, - For feature reconstruction constraints, For the set of confusing variables, For mutually exclusive causal constraints, The weighting coefficients are adaptive and dynamically adjusted according to the distribution of data features. S3. Causal Analysis and Damage Source Tracing: Based on the damage causal graph model and causal feature set, spatial hierarchy is divided and causal intervention is carried out. The comprehensive causal contribution coefficient of each influencing factor is calculated using the causal quantification algorithm to determine the damage cause. The initial damage location is traced by combining the mechanical transmission relationship of the component. The comprehensive causal contribution coefficient is calculated using a causal quantification algorithm, and its mathematical expression is as follows: in, For the first The comprehensive causal contribution coefficient of the class of influencing factors For the first A subset of causal characteristics of influencing factors. For the first Similar to support structure damage events, The feature-damage causal correlation function, The total number of influencing factors. This represents the total number of damage events. For the first Spatial hierarchy of the effects of various influencing factors This is a spatial scale correction term. This is a dynamic weighting coefficient, with a value range of [value range missing]. ; S4. Dynamic graph construction and evolution prediction: The support components are abstracted as graph nodes, and the mechanical connections between components are abstracted as graph edges and assigned initial weights. The edge weights and topology are adjusted in real time according to the damage causes and initial damage locations obtained in step S3 to form a time-varying topological dynamic graph. Spatiotemporal attention dynamic graph convolutional network is used to extract spatiotemporal correlation features, and spatiotemporal evolution algorithm is used to predict the damage evolution state quantities at multiple time and spatial scales. S5. Anomaly Warning and Closed-Loop Iteration: Based on historical working condition data, establish a normal damage evolution statistical benchmark using the model and algorithm constructed in steps S2 to S4, compare the damage evolution state quantity obtained in step S4 with the benchmark to classify the warning level; continuously collect field data, and periodically retrain the model and algorithm and update the parameters.

2. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S1 data acquisition and processing, displacement and stress sensors are installed at the top of the support piles and at different depths of the pile body; stress and vibration sensors are installed at the two end joints and mid-span of the steel supports; tension sensors are installed at the anchor head of the anchor rods; displacement and stress sensors are installed at the mid-span and node connection of the walers; earth pressure and water pressure sensors are installed inside the soil at different depths around the foundation pit; settlement and tilt sensors are installed at environmental points around the foundation pit; and temperature and construction load sensors are installed at the construction site. The sampling frequency of the displacement, stress, and tension sensors is set to once per minute; the sampling frequency of the earth pressure and water pressure sensors is set to once every five minutes; and the sampling frequency of the temperature and construction load sensors is set to once every ten minutes. When the monitoring data fluctuates, the sampling frequency of the corresponding sensor and adjacent related sensors is switched to once per second.

3. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S1 data acquisition and processing stage, the raw electrical signals output by the sensor are wirelessly transmitted using a combination of LoRa and NB-IoT transmission methods. For long-distance monitoring points, 4G or 5G networks are used as supplementary transmission links. All monitoring data are aggregated to a cloud-based distributed storage center via the MQTT communication protocol. The data deduplication stage removes duplicate data records corresponding to the same timestamp. The range verification stage removes data that exceeds the sensor's calibrated range. The normalization stage uses a minimum-maximum normalization method to map the monitoring data to a unified numerical range. Monitoring data with a missing duration of less than one hour is completed using linear interpolation, while monitoring data with a missing duration of one hour or more is completed using time-series prediction. The outlier handling stage removes abnormal monitoring data and replaces the data with the average of multiple neighboring valid monitoring data to obtain a standardized monitoring dataset.

4. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S2 causal modeling and feature extraction, the damage causal graph model sets two types of nodes. The first type of node is the influencing factor node, which specifically includes soil strength, groundwater level, excavation speed, dewatering rate, construction load, ambient temperature, and surrounding traffic load. The second type of node is the damage node, which specifically includes support pile displacement, support pile cracking, steel support instability, steel support buckling, anchor bolt tension attenuation, anchor bolt detachment, waler deformation, waler fracture, pit heave, and pit leakage. Based on engineering experience, industry standards, and historical engineering cases, an initial causal graph structure is built. Directed edges are used to establish the association between different nodes. Using a standardized monitoring dataset, the number of connecting edges and the direction of directed edges in the initial causal graph are adjusted through a causal discovery algorithm. The probability parameters corresponding to each directed edge are calculated using the maximum likelihood estimation method. For discrete nodes, the conditional probability table corresponding to different combinations of parent nodes is calculated, and for continuous nodes, the mean and variance of the probability distribution are calculated.

5. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S3 causal analysis and damage tracing, the spatial hierarchy is divided into three levels: component level, local level, and overall level. The component level corresponds to a single support pile, a single steel support, a single anchor rod, or a single waler section. The local level corresponds to all support components within an independently divided area of ​​the foundation pit. The overall level corresponds to the complete deep foundation pit support system. The causal intervention operation adopts a single variable adjustment mode, keeping the values ​​of all remaining influencing factors fixed, adjusting the value of a single influencing factor individually, and simultaneously collecting the support structure status data corresponding to the three spatial levels of component level, local level, and overall level.

6. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In S3, during causal analysis and damage tracing, when the support structure is monitored to produce an abnormal state, the abnormal state information is input into the damage causal graph model. The posterior probability values ​​corresponding to all influencing factor nodes are obtained through Bayesian inference. The posterior probability values ​​and the comprehensive causal contribution coefficient are combined to complete the sorting process. The damage cause is determined based on the sorting results. The damage propagation path is constructed based on the fixed mechanical transmission relationship between the support components. The abnormal position obtained by monitoring is used as the starting point for deduction. The initial damage position is located by deducing backward layer by layer along the damage propagation path.

7. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S4 dynamic graph construction and evolution prediction, a single support pile, a single steel support, a single anchor rod, and a single waler section correspond to independent graph nodes in the dynamic graph, and the mechanical connection relationship between support components corresponds to graph edges in the dynamic graph. Each graph node is configured with an independent multi-dimensional feature vector, which consists of three types of data: causal features, component geometric parameters, and component material properties. The initial weight of the graph edges is assigned based on the mechanical connection strength of the components. The weight parameters of the graph edges are adjusted based on the real-time structural state corresponding to the damage cause and the initial damage location, combined with the change in the force transmission coefficient between components. When the mechanical connection of a component fails, all graph edges between the component and its adjacent components are deleted, thus completing the dynamic graph topology update.

8. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S4 dynamic graph construction and evolution prediction, the spatiotemporal attention dynamic graph convolutional network is composed of multiple layers of spatiotemporal attention blocks stacked together. Each spatiotemporal attention block is sequentially set with a temporal attention layer, a spatial attention layer, and a dynamic graph convolutional layer. The temporal attention layer completes the temporal data operation and extracts the correlation features of different time nodes. The spatial attention layer completes the graph node data operation and extracts the correlation features of different support components. The dynamic graph convolutional layer combines the real-time updated dynamic graph topology to extract the information transmission features between components. The dynamic graph node feature sequence is input into the spatiotemporal attention dynamic graph convolutional network, and after the feature operation is completed layer by layer, the spatiotemporal correlation features are output.

9. The method for damage identification of deep foundation pit support structures combined with deep learning according to claim 1, characterized in that, In the S5, abnormal early warning and closed-loop iteration, the monitoring data and structural status data corresponding to the historical normal working conditions are collected, and a multivariate Gaussian statistical distribution is constructed in combination with the damage causal graph model. The multivariate Gaussian statistical distribution is used as the statistical benchmark for normal damage evolution. The deviation between the damage evolution state quantity and the statistical benchmark of normal damage evolution is calculated. Combined with the damage inducing factors and the comprehensive causal contribution coefficient, a comprehensive anomaly index is calculated by weighted summation. Based on the numerical range of the comprehensive anomaly index, three warning levels are divided: the comprehensive anomaly index below the first threshold corresponds to the first warning level; the comprehensive anomaly index in the range between the first and second thresholds corresponds to the second warning level; and the comprehensive anomaly index above the second threshold corresponds to the third warning level.

10. A deep learning-based damage identification system for deep foundation pit support structures, the system being applicable to the deep learning-based damage identification method for deep foundation pit support structures as described in any one of claims 1-9, characterized in that, The system includes: a data acquisition and preprocessing module, a causal modeling and feature extraction module, a causal analysis and damage tracing module, a dynamic graph construction and evolution prediction module, and an early warning and closed-loop iteration module. Data acquisition and preprocessing module: Deploys multiple types of sensors to collect multi-source monitoring data, transmits it wirelessly and stores it in the cloud, completes data deduplication, verification, normalization, missing value and outlier processing, and outputs a standardized monitoring dataset; Causal modeling and feature extraction module: Constructs a damage causal graph model containing influencing factors and damage nodes, optimizes the model structure and parameters, and extracts causal feature sets through causal screening algorithms; Causal analysis and damage tracing module: Divide the spatial hierarchy into component level, local level and global level and perform causal intervention, calculate the comprehensive causal contribution coefficient to determine the damage cause, and locate the initial damage location by combining the mechanical transmission relationship; The dynamic graph construction and evolution prediction module abstracts support components and mechanical connections into dynamic graph nodes and edges and updates the topology in real time. It extracts features through a spatiotemporal attention dynamic graph convolutional network and predicts damage evolution state quantities through a spatiotemporal evolution algorithm. Early warning and closed-loop iteration module: Establish a normal damage evolution benchmark, compare data to classify early warning levels; continuously collect field data to update the knowledge base, and periodically retrain the model and algorithm and update parameters.

Images