A neural network-based method for monitoring the deformation of engineering foundation pits

Abstract

This invention relates to the field of foundation pit deformation monitoring technology, and particularly to a neural network-based method for monitoring engineering foundation pit deformation. The method includes: acquiring monitoring data from multiple monitoring points; calculating the temporal covariance intensity of the deformation rate between any two monitoring points within a historical time window; calculating the geotechnical mechanics influence factor; calculating the causal perception activity vector for each monitoring point, where the causal perception activity vector contains the activity intensity at the current moment; calculating the adaptive adjacency weights between monitoring points; using the adaptive adjacency weights as the adjacency matrix of a graph neural network model; and inputting the monitoring values ​​and causal perception activity vectors of each monitoring point as node features into the trained graph neural network model to obtain the current deformation pattern of the foundation pit. This invention improves the accuracy of graph neural networks in recognizing deformation patterns.

Description

Technical Field

[0001] This invention relates to the field of foundation pit deformation monitoring technology, and in particular to a method for monitoring engineering foundation pit deformation based on neural networks. Background Technology

[0002] Deformation monitoring of engineering foundation pits is a core component in ensuring construction safety and the stability of the surrounding environment. By deploying various sensors, such as settlement point sensors, inclinometers, and stress gauges, within the foundation pit retaining structure and affected area, monitoring data can be acquired. Analyzing this data identifies potential abnormal deformation patterns and establishes a correlation with specific engineering risks (such as instability and leakage).

[0003] Graph Neural Networks (GNNs) have been introduced into the field of foundation pit deformation monitoring due to their powerful ability to process non-Euclidean space structure data. Typically, each sensor is abstracted as a monitoring point in a graph, and the physical or logical relationships between these monitoring points are abstracted as edges in the graph. The overall deformation pattern is identified by learning the interactions between the monitoring points.

[0004] However, existing graph neural network-based methods generally suffer from a technical problem: the prior knowledge of the graph structure relies on the adaptive learning problem. Specifically, the performance of graph neural network models is highly dependent on a predefined graph topology (i.e., the adjacency matrix). Graph topologies are typically constructed based on simplified prior knowledge, such as defining edges based on the physical spatial distance between sensors or direct structural connections. This fixed, experience-based graph structure has limitations. First, it may fail to accurately reflect the complex, nonlocal, real mechanical relationships transmitted through the soil medium within the excavation pit. For example, dewatering operations at one corner of the pit may significantly affect retaining piles on the other side, a coupling relationship that cannot be represented in distance-based graphs. Second, this static graph topology cannot be adjusted with the construction process (such as excavation and support) and the evolution of deformation states, ignoring the dynamic, time-varying characteristics of the influence relationships between monitoring points. Therefore, an improperly pre-defined or unadaptable graph topology limits the learning upper limit of the model, preventing it from discovering real risk transmission paths and thus reducing the accuracy and reliability of deformation pattern recognition. Summary of the Invention

[0005] To address the issue that fixed adjacency matrices lead to low accuracy in deformation pattern recognition by graph neural networks, this invention provides a neural network-based method for monitoring engineering foundation pit deformation.

[0006] This invention provides a method for monitoring the deformation of engineering foundation pits based on neural networks, employing the following technical solution:

[0007] A neural network-based method for monitoring the deformation of engineering foundation pits includes the following steps: acquiring monitoring data from multiple monitoring points; calculating the temporal covariance intensity of the deformation rate between any two monitoring points within a historical time window; calculating the geotechnical mechanics influence factor, which is positively correlated with the temporal covariance intensity and the soil transfer coefficient of the path connecting the two monitoring points, and negatively correlated with the spatial distance between the two monitoring points; calculating the causal perception activity vector for each monitoring point, which includes the activity intensity at the current moment; the activity intensity is positively correlated with the magnitude of the change in the monitored value within a set window;

[0008] The adaptive adjacency weights between monitoring points are calculated. The adaptive adjacency weights are positively correlated with the absolute value of the geotechnical mechanics influence factor and the cosine similarity of the causal perception active vectors of any two monitoring points. The adaptive adjacency weights are used as the adjacency matrix of the graph neural network model. The monitoring values ​​of each monitoring point and its causal perception active vectors are used as node features and input into the trained graph neural network model to obtain the deformation mode of the current foundation pit.

[0009] By comprehensively calculating the temporal covariance intensity, geotechnical mechanics influence factors, and causal perception activity vectors among monitoring points, a dynamically changing adaptive adjacency weight is constructed as the adjacency matrix of the graph neural network. This integrates the temporal synchronization of data, the physical properties of the soil medium, and the causal roles of each monitoring point (such as triggerers or responders), enabling the graph's topology to dynamically and more realistically simulate the mechanical transmission paths and influence relationships within the foundation pit. This improves the accuracy and physical interpretability of the graph neural network model in recognizing complex deformation patterns.

[0010] Preferably, the expression for the temporal covariance intensity is:

[0011] ;

[0012] in, Indicates monitoring point and At the present moment The temporal covariance intensity; Indicates the length of the historical time window; and They represent monitoring points respectively. and Within the time window The deformation rate at any given time; and monitoring points In the time window Mean and standard deviation of internal deformation rate and monitoring points In the time window Mean and standard deviation of internal deformation rate.

[0013] By standardizing and integrating the time series of deformation rates at two monitoring points, their synchronicity (positive or negative correlation) in dynamic behavior can be quantitatively measured. This provides an accurate and reliable quantitative basis for subsequently constructing geotechnical mechanics influence factors that reflect the true physical correlation.

[0014] The preferred expression for the geotechnical mechanics influence factor is:

[0015] ;

[0016] in, Indicates monitoring point and The geotechnical influencing factors between; Indicates monitoring point and The temporal covariance intensity; Indicates monitoring point and Spatial distance between them; Indicates monitoring point and Path between The soil transfer coefficient.

[0017] By combining the temporal covariance intensity at the data level with a physical model that comprehensively considers spatial distance and soil transfer coefficient, it is possible to more accurately identify strong correlations between monitoring points that are far apart but connected by a dominant medium, thus avoiding misjudgments caused solely by distance.

[0018] Preferably, the expression for the soil transfer coefficient is:

[0019]

[0020] in, Indicates monitoring point and Path between Soil transfer coefficient; Indicates monitoring point and Spatial distance between them; Representing a path Upper Normalized Young's modulus of soil types; Indicates the first The length of the soil layer of each type; Representing a path Total number of upper soil layer types.

[0021] By weighted averaging the physical properties and lengths of different soil types along the path connecting two monitoring points, the transmission efficiency of deformation or stress in complex strata is quantified. This accurately reflects the modulating effect of geological conditions on the interaction between monitoring points, enhancing the model's physical realism and prediction accuracy.

[0022] The preferred expression for activity intensity is:

[0023] ;

[0024] in, Indicates monitoring point At any moment The intensity of activity at that time; Indicates monitoring point At any moment The monitoring value at that time; Indicates the time step; This indicates the preset deformation threshold. This represents the hyperbolic tangent function.

[0025] By calculating the changes in deformation at monitoring points over a short period of time, the activity level of these points can be dynamically assessed. This allows the model to automatically identify monitoring points undergoing drastic changes, i.e., high-risk points, thus giving them greater attention in subsequent analyses. This achieves dynamic focusing on risk areas and improves the sensitivity and early warning efficiency of the monitoring system.

[0026] Preferably, the causal perception activity vector also includes the activity type at the current moment, which includes active triggerer, passive responder and independent responder.

[0027] Building upon activity intensity, this paper further introduces activity types, including active triggers, passive responders, and independent responders. This addresses the problem that existing technologies can typically only detect correlations but struggle to determine causal direction. By distinguishing whether a deformation at a monitoring point is a source or a consequence, the physical interpretability of the model results is enhanced. It not only identifies deformations but also helps analyze their root causes and propagation paths.

[0028] Preferably, the method for determining the activity type is as follows: Granger causality test is performed on the deformation rate sequence of the monitoring point and other monitoring points to obtain the Granger causality strength. The deformation rate, deformation acceleration, distance to nearby construction events, and Granger causality strength of the current monitoring point are input into the classification model to obtain the one-hot encoding vector, and the activity type of the monitoring point is further obtained.

[0029] By utilizing methods such as Granger causality tests, and performing causal analysis on the deformation rate sequences between monitoring points, combined with information such as deformation acceleration and distance from construction events, an objective and effective data analysis method is provided to distinguish between active triggers and passive responders, thereby improving the accuracy and automation level of the judgment.

[0030] The preferred expression for adaptive adjacency weight is:

[0031] ;

[0032] in, Indicates monitoring point and At any moment Adaptive adjacency weights at different times; Indicates monitoring point and The geotechnical influencing factors between; Indicates monitoring point At any moment The causal perception activity vector at time Indicates monitoring point At any moment The causal perception active vector at that time; This represents the cosine similarity between vectors.

[0033] When two monitoring points are physically correlated and have similar causal roles, the connection weight between them will be significantly enhanced, enabling the constructed graph structure to capture the synergistic effects and key transmission paths in the foundation pit deformation process more sensitively and accurately.

[0034] Preferably, the monitoring values ​​of each monitoring point and its causal perception activity vector are input into the trained graph neural network model as node features, including: the monitoring values ​​of each monitoring point at the current time, the first difference of the monitoring values, the second difference of the monitoring values, and the causal perception activity vector together constitute the node features of the graph neural network model.

[0035] By using the instantaneous value of the monitored value, the first-order difference (representing velocity), the second-order difference (representing acceleration), and the aforementioned causal perception active vector as node features, the model is provided with comprehensive information about the current state, trend of change, and role of each monitoring point in the causal network. This enriches the amount of input information for the model, helps the model to learn and understand complex deformation behaviors more deeply, and thus improves the accuracy of the final deformation pattern classification.

[0036] Preferably, the deformation modes include normal settlement, overall tilting, local bulging, and uneven settlement.

[0037] The present invention has the following technical effects:

[0038] By integrating the temporal covariance of monitoring data, the physical transmission law based on soil mechanical properties, and the dynamic causal role of monitoring points, an adaptive adjacency weight matrix is ​​constructed that can reflect the real mechanical correlation and risk transmission path inside the foundation pit in real time. This enables the model to accurately identify complex deformation patterns, improves the physical interpretability of monitoring results, and enhances the accuracy of graph neural networks in identifying deformation patterns. Attached Figure Description

[0039] Figure 1 This is a flowchart of a neural network-based method for monitoring the deformation of engineering foundation pits according to the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] This invention discloses a method for monitoring the deformation of engineering foundation pits based on neural networks, referring to... Figure 1 This includes the following steps:

[0042] S1: Obtain monitoring data about the foundation pit.

[0043] Monitoring data from multiple monitoring points in the foundation pit project were acquired over a continuous time series. These monitoring points included: surface settlement points, horizontal displacement points at the top of retaining piles, points at specific depths within inclinometer tubes, and support axial force. Data from different types of monitoring points were time-synchronized, and sampling frequencies were aligned to form a multimodal time-series dataset. For any given monitoring point... At any given moment The monitored value is recorded as .

[0044] S2: Calculate the temporal covariance intensity between any two monitoring points.

[0045] When two physically related monitoring points are affected by the same or related stress field changes, the fluctuations in their deformation rates often exhibit a high degree of consistency or a regular inverse relationship. Therefore, by analyzing the degree of synchronization of their change rates, we can gain a preliminary understanding of their potential intrinsic connection. Thus, by calculating the time-series covariant intensity... To measure the synchronicity of deformation rate changes at two monitoring points and reflect the mutual influence in the dynamic process, the temporal covariance intensity is calculated by constructing a time window with the current time t as the endpoint along the historical direction, the length of which is... The expression for the temporal covariance intensity is:

[0046] ;

[0047] in, Indicates monitoring point and At the present moment The temporal covariance intensity; Indicates the length of the historical time window; and They represent monitoring points respectively. and Within the time window The deformation rate at any given time; and monitoring points In the time window Mean and standard deviation of internal deformation rate and monitoring points In the time window Mean and standard deviation of internal deformation rate.

[0048] in, Indicates monitoring point and At the present moment The temporal covariance intensity, Indicates the length of the time window. and They represent monitoring points respectively. and Within the time window The deformation rate at any given time can be obtained from the original monitoring value sequence. We obtain it by performing difference or differentiation. and monitoring points In the time window The mean and standard deviation of the internal deformation rate are used to normalize the data and eliminate the influence of dimensions. and monitoring points In the time window The mean and standard deviation of the internal deformation rate are used to normalize the data and eliminate the influence of dimensions.

[0049] If the deformation rates at two monitoring points fluctuate in the same direction (i.e., increase or decrease simultaneously), the product is positive, and the integral result tends towards a large positive number, indicating a strong positive correlation. If the fluctuations are in opposite directions, the integral result tends towards a large negative number, indicating a strong negative correlation. If the fluctuations are irregular, the integral result approaches zero. Therefore, The larger the value, the stronger the synchronicity of the dynamic behavior of the two monitoring points, and the stronger the potential correlation.

[0050] S3: Constructing the influencing factors of rock and soil mechanics.

[0051] In one embodiment, within an engineering foundation pit, stress and deformation are transmitted through the soil medium, and this transmission process is affected by distance and the properties of the medium. Even if two monitoring points show strong synchronicity, if they are separated by a hard rock layer or a long distance, the actual direct impact may be weak; conversely, if they are located in the same soft soil layer and are close to each other, their data synchronicity is likely to stem from actual mechanical transmission. Therefore, in time-varying intensity... Based on this, we introduce geotechnical mechanical properties and construct a geotechnical mechanical influence factor that can reflect the real physical effects. The expression is:

[0052]

[0053] in, Indicates monitoring point and Geotechnical influencing factors between Indicates monitoring point and Spatial distance between them Indicates monitoring point and Path between The soil transfer coefficient characterizes the connection between two points. and path The comprehensive mechanical transmission performance of the soil is determined based on the geological survey report. For example, different soil layers along the path, such as silt, loess, and sandstone, are assigned different transmission coefficient values. The higher the value, the better the transmission performance, such as sandstone > loess > silt. The weighted average is then obtained by using the ratio of the length of different soil layers to the total length of the path as the weight. The term uses a logarithmic form to penalize the distance, in order to mitigate the rapid decay of long-range effects and allow the model to capture some long-range effects.

[0054] Distance term in the denominator This has a damping effect; the greater the distance, the smaller the influence factor in geotechnical mechanics. If the soil mechanical properties between two points are good ( If the value is large, the transmission of influence between them is more effective. This also increases accordingly, raising the assessment of the correlation between monitoring points from the level of pure data to the level of integrating physical models.

[0055] In another embodiment, the soil transfer coefficient is calculated as follows: The soil layer type between two monitoring points, as well as the length and Young's modulus of each soil type, are obtained; the Young's modulus is then normalized using a linear normalization algorithm.

[0056]

[0057] in, Indicates monitoring point and Path between Soil transfer coefficient, This represents the pattern modulus of the k-th type of soil layer after normalization. This represents the length of the k-th type of soil layer. Indicates monitoring point and The spatial distance between them, where N represents the monitoring point. and The number of soil layer types between them, k represents the number of monitoring points. and Index of soil layer types.

[0058] For example, the straight-line distance between monitoring point A and monitoring point B The path is 20 meters long and traverses two soil layers; the first 15 meters pass through silt. =0.20, the last 5 meters pass through sandstone, its =1.00, then the soil transfer coefficient of the path between monitoring point A and monitoring point B is: The soil transfer coefficient for the path between monitoring points A and B is 0.40. This value is higher than 0.20 for pure silt but lower than 1.00 for sandstone because most of its path is in silt with poor transfer properties. The soil transfer coefficient comprehensively includes information on path length, soil layer type, and mechanical properties of each soil layer.

[0059] S4: Calculate the adaptive adjacency weights between monitoring points.

[0060] Deformation risks in foundation pits are often concentrated in areas of severe localized deformation, and understanding the source and propagation path of deformation is crucial for risk early warning. A monitoring point undergoing significant deformation, whether it acts as an active trigger or a passive responder, should have higher importance in the graph structure at the current moment. By introducing causal analysis, it is possible to dynamically focus on high-risk areas, distinguish the causal role of monitoring points, i.e., differentiate whether the monitoring point is an active trigger, a passive responder, or an independent responder, thereby improving the physical interpretability of the graph structure.

[0061] S41: Calculate the activity intensity of each monitoring point.

[0062] The activity intensity is calculated to measure the degree of deformation at monitoring point i in the most recent time step, expressed as:

[0063]

[0064] in, Indicates monitoring point The activity intensity at time t Indicates monitoring point At any moment The monitoring value at that time Indicates the time step. This represents the preset deformation threshold, and tanh represents the hyperbolic tangent function used for normalization.

[0065] When the deformation is much smaller than the threshold A value close to 1 indicates that monitoring point i is in a resting state; when the deformation increases significantly... A value close to 2 indicates that monitoring point i is in an active state.

[0066] S42: Determine the activity type of the monitoring point.

[0067] Determine monitoring points The activity type is categorized into three types: active triggerer, passive responder, and independent responder. Within the sliding time window, the monitoring points are... Deformation rate sequence and Perform Granger causality testing:

[0068] If monitoring point j has a significant Granger causal relationship with monitoring point i, but monitoring point i has no significant relationship with monitoring point j, then monitoring point i tends to be a passive responder; if monitoring point i has a significant Granger causal relationship with monitoring point j, but monitoring point j has no significant relationship with monitoring point i, then monitoring point i tends to be an active trigger; if both relationships are significant or neither is significant, then monitoring point i and monitoring point j may both be co-responders or independent variants. The Granger causality test is existing technology, and its specific steps will not be detailed here. If monitoring point... If the deformation accelerates immediately after a specific construction activity (such as excavation or dewatering) and there is no obvious external monitoring point that has a significant causal effect on it, then it is more likely to be an active trigger.

[0069] The deformation rate, deformation acceleration, distance to nearby construction events, and Granger causality strength (p-value) between monitoring point i and other monitoring points at the current moment are input into the classification model to obtain the one-hot encoding vector, which includes: and ,in, This indicates that the monitoring point is an active trigger. This indicates that the monitoring point is a passive responder. The monitoring point is indicated as either a resting or co-responder, and the classification model is a support vector machine.

[0070] The causal sensing activity vector is constructed using activity intensity and one-hot encoded vector, and its expression is: ,in, Indicates monitoring point The causal perception activity vector at time t Indicates monitoring point The activity intensity at time t Indicates monitoring point The one-hot encoded vector at time t.

[0071] S43: Calculate the adaptive adjacency weights between monitoring points.

[0072]

[0073] in, Indicates monitoring point and The adaptive adjacency weight at time t Indicates monitoring point and Geotechnical influencing factors between Indicates monitoring point The causal perception activity vector at time t This represents the causal sensing activity vector of monitoring point j at time t. Represents the causal perception active vector With causal perception active vector The cosine similarity.

[0074] if A value close to 1 indicates that monitoring points i and j are highly similar in terms of activity type and intensity. For example, both monitoring points may be active triggers or passive responders undergoing drastic changes. This high similarity may mean that they are subject to similar external influences or that there is some kind of synergy between them. In this case, the synergy between the monitoring points should be reflected by increasing the adjacency weight between the two monitoring points.

[0075] S5: Monitoring of foundation pit deformation based on adaptive adjacency weight.

[0076] The calculated adaptive adjacency weights As an adjacency matrix in the graph neural network model, all monitoring points are treated as nodes in the graph, and the monitoring value of each monitoring point at the current time is represented. The first-order difference, second-order difference, and causal sensing active vector are used as monitoring point features and input into a pre-trained graph neural network model to obtain the classification results of the current overall deformation mode of the foundation pit. The classification results include normal settlement, overall tilt, local heave, and uneven settlement.

[0077] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A neural network-based engineering foundation pit deformation monitoring method, characterized in that, The method comprises the steps of: acquiring monitoring data of a plurality of monitoring points; calculating time series correlation strength of deformation rates of any two monitoring points within a historical time window; calculating a geotechnical mechanics influence factor, which is positively correlated with the time series correlation strength and a soil body transmission coefficient of a path connecting the two monitoring points, and negatively correlated with a spatial distance between the two monitoring points; and calculating a causal perception active vector of each monitoring point, which comprises an active strength at a current time; the active strength is positively correlated with a change amount of a monitoring value within a set window. An adaptive adjacency weight between the monitoring points is calculated, which is positively correlated with an absolute value of the geotechnical mechanics influence factor and a cosine similarity of the causal perception active vectors of any two monitoring points; the adaptive adjacency weight is used as an adjacency matrix of a graph neural network model; monitoring values of the monitoring points and the causal perception active vectors thereof are input into the trained graph neural network model as node features, so as to obtain a deformation mode of the current foundation pit.

2. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The calculation method of the time correlation intensity is: taking the current time t as an end point, constructing a time window in the historical direction, and the expression of the time correlation intensity is: ; wherein, denotes a monitoring point with at the current time a time-variant intensity; denotes the length of a historical time window; and denote a monitoring point and a deformation rate at a time within the time window; and are the mean and standard deviation of the deformation rate of a monitoring point within the time window , and are the mean and standard deviation of the deformation rate of a monitoring point within the time window .

3. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The expression of the geotechnical mechanics influence factor is: ； wherein, representing monitoring points and a geotechnical influence factor between the monitoring points; representing monitoring points and a time series correlation strength between the monitoring points; representing monitoring points and a spatial distance between the monitoring points; representing monitoring points with between paths of the soil mass transfer coefficient.

4. The neural network-based engineering foundation pit deformation monitoring method according to claim 3, characterized in that, The expression of the soil body transmission coefficient is: ； in, Indicates monitoring point and Path between Soil transfer coefficient; Indicates monitoring point and Spatial distance between them; Representing a path Upper Normalized Young's modulus of soil types; Indicates the first The length of the soil layer of each type; Representing a path Total number of upper soil layer types.

5. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The expression of the active strength is: ； wherein, represents a monitoring point at time ; and represents a monitoring point at time ; and represents a time step represents a preset deformation threshold value represents a hyperbolic tangent function.

6. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The causal perception active vector further comprises an active type at the current time, and the active type comprises an active trigger, a passive responder and an independent responder.

7. The neural network-based engineering foundation pit deformation monitoring method according to claim 6, characterized in that, The judgment method of the active type is: performing Granger causality test on deformation rate sequences of the monitoring point and other monitoring points to obtain Granger causality strength; inputting the deformation rate, deformation acceleration and distance from a nearby construction event of the monitoring point at the current time and the Granger causality strength obtained by performing Granger causality test on the deformation rate sequences of the monitoring point and other monitoring points into a classification model to obtain a one-hot encoding vector, and further obtaining the active type of the monitoring point.

8. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The expression of the adaptive adjacency weight is: ； wherein, represents a monitoring point and an adaptive adjacency weight at time ; represents a monitoring point and a geotechnical influence factor between ; represents a monitoring point a causal-aware active vector at time ; represents a monitoring point a causal-aware active vector at time ; represents a cosine similarity between vectors.

9. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The monitoring values of the monitoring points and the causal perception active vectors thereof are input into the trained graph neural network model as node features, comprising: the monitoring value of each monitoring point at the current time, the first-order difference of the monitoring value, the second-order difference of the monitoring value and the causal perception active vector jointly constitute node features of the graph neural network model.

10. The neural network-based engineering foundation pit deformation monitoring method according to claim 1, characterized in that, The deformation mode comprises normal settlement, overall tilting, local uplift and uneven settlement.

Images