A method for early warning of collapse risk of a hole wall of an impact drilled pile foundation

Abstract

This invention provides a method for early warning of borehole wall collapse risk in impact-drilled pile foundations, belonging to the field of construction risk early warning technology. This invention collects multi-source signals by deploying an ultrasonic transducer array and a mechanical vibration acceleration sensor array at multiple depth nodes within the borehole. It extracts borehole wall collapse characteristic components from strong vibration interference using empirical mode decomposition and independent component analysis. Simultaneously, it inputs the multi-dimensional time-series data of the entire borehole into a minimum description length change point detection algorithm to detect abrupt changes in borehole wall state and outputs a hazard level score. Sensor data from each depth node is input into a physically embedded spatiotemporal graph attention model to output the collapse risk probability and estimated collapse depth location. Furthermore, an adaptive learning rate adjustment function driven by a stratum change intensity index dynamically adjusts the online update learning rate of the artificial intelligence model. Finally, it integrates multi-level early warning signals to output a final early warning conclusion, solving the technical problem of the inability to provide real-time and accurate early warning of borehole wall collapse risk during impact-drilled pile construction.

Description

Technical Field

[0001] This invention belongs to the field of construction risk early warning technology, and specifically relates to a method for early warning of the risk of borehole wall collapse in impact drilling pile foundations. Background Technology

[0002] Impact drilling pile construction is a crucial process in foundation engineering for bridges, high-rise buildings, and other projects, with the core challenge being ensuring borehole wall stability. Traditional methods for monitoring borehole wall stability mainly include manual inspection of mud levels, empirical mud density control, and static borehole wall lateral pressure monitoring. On-site, engineers typically rely on experience to judge the borehole wall condition, and some projects use a single sensor to collect mud parameters and trigger alarms based on fixed thresholds. With the increasing demand for deep-hole pile construction and projects in complex geological formations, the aforementioned methods are widely used in various pile foundation construction projects.

[0003] However, the strong mechanical vibration and broadband noise generated by the impact drill seriously interfere with the sensor signal, causing the weak collapse signal to be masked; the traditional static threshold method cannot detect the gradual instability process of the wall differential pressure; single-dimensional monitoring data is difficult to characterize the multi-dimensional spatiotemporal evolution of the borehole wall state under complex strata conditions; existing models have a catastrophic forgetting problem when crossing new strata and cannot continuously adapt to strata changes.

[0004] In existing technologies, due to the strong mechanical vibration interference of impact drilling and the complex and variable geological conditions, traditional monitoring methods cannot effectively extract weak collapse signals from strong vibration noise, nor can they predict the dynamic evolution trend of borehole wall instability in real time. As a result, the risk of borehole wall collapse cannot be accurately and timely warned. In other words, existing technologies have the technical problem of not being able to provide real-time and accurate early warning of the risk of borehole wall collapse during the construction of impact drilled piles. Summary of the Invention

[0005] In view of this, the present invention provides a method for early warning of the risk of borehole wall collapse in impact drilled pile foundations, which can solve the technical problem in the prior art that the risk of borehole wall collapse during the construction of impact drilled piles cannot be warned in real time and accurately.

[0006] This invention is implemented as follows: This invention provides a method for early warning of borehole wall collapse risk in impact drilling pile foundations, comprising the following steps:

[0007] An array of ultrasonic transducers and an array of mechanical vibration acceleration sensors are deployed at multiple depth nodes inside the borehole to collect multi-frequency ultrasonic echo signals and mechanical vibration signals, and simultaneously collect data on mud level, mud density, pump flow rate and borehole wall side pressure.

[0008] Empirical mode decomposition was performed on the multi-frequency ultrasonic echo signal and the mechanical vibration signal to obtain the intrinsic mode function component matrix of each channel. Then, independent component analysis was performed on the intrinsic mode function component matrix of each channel to separate the characteristic component of the borehole wall collapse.

[0009] Using mud level, mud density, and pump flow rate as inputs, a system of simultaneous differential equations based on mass and momentum conservation is established. The fourth-order Runge-Kutta method is used to solve the time evolution curve of the wall differential pressure online numerically. An exponentially moving weighted sliding window is introduced into the wall differential pressure time evolution curve for trend extrapolation. When the extrapolated predicted value is about to cross the Rankine active earth pressure critical value, a first-level warning is triggered. When the second derivative of the wall differential pressure exceeds the second-level warning threshold, a second-level emergency warning is triggered.

[0010] The mud level, mud density, pump flow rate, borehole wall lateral pressure, and borehole wall collapse characteristic components are used to construct multidimensional time-series data of the entire borehole. The data is input into the minimum description length change point detection algorithm, and dynamic programming segmentation is performed using the Bayesian information criterion as the cost function. The time of abrupt change in borehole wall state is detected, and the Körbek-Leibler divergence of the multidimensional joint distribution before and after the change point is calculated. When the Körbek-Leibler divergence exceeds the divergence judgment threshold, a collapse risk event is determined and a hazard level score is output.

[0011] The borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement value, and mud level time series data of each depth node are input into the physical embedded spatiotemporal map attention model, which outputs the collapse risk probability and estimated collapse depth location. The online update learning rate of the physical embedded spatiotemporal map attention model is dynamically adjusted by the formation mutation intensity index-driven formation mutation adaptive learning rate adjustment function.

[0012] By combining the Level 1 warning status, Level 2 emergency warning status, hazard level score and collapse risk probability, the final warning signal is output according to the highest hazard level among the four, and the current depth node data and stratum type label are written into the historical sample library for continuous learning and updating of the physical embedded spatiotemporal map attention model.

[0013] The ultrasonic transducer array consists of multiple ultrasonic transducers of different frequencies, ranging from 40 to 500 kHz. The mechanical vibration acceleration sensor array is deployed on the drill pipe and at the borehole opening, with a sampling frequency of not less than 2000 Hz. .

[0014] The empirical mode decomposition refers to first adaptively decomposing the multi-frequency ultrasonic echo signals and mechanical vibration signals of each sensor channel, and then stitching together the intrinsic mode function component matrices of all channels before inputting them into independent component analysis. The frequency range of the borehole wall collapse characteristic components is 100–800 Hz. .

[0015] In the calculation of the time evolution curve of the wall differential pressure, the time step of the fourth-order Runge-Kutta method is 0.5 to 2 seconds. The noise of the mud level sensor is estimated using Kalman filtering before being input into the calculation.

[0016] The window duration of the exponentially moving weighted sliding window is 30–120. The prediction time domain for trend extrapolation is 10–60. The critical value of Rankine active earth pressure was determined by numerical simulation using an elastic-visco-plastic constitutive model based on in-situ test data of the strata.

[0017] The secondary early warning threshold is determined by taking 3 to 5 times the upper limit of the normal fluctuation range of the second derivative of the differential pressure of the wall in the normal drilling section.

[0018] The minimum description length variable point detection algorithm employs a pruning dynamic programming algorithm to prune the cost function increment, reducing computational complexity to... Level; the divergence determination threshold is taken as the 97.5th percentile of the Körbek-Leibler divergence distribution in the normal drilling section; the hazard level score is obtained by mapping the ratio of Körbek-Leibler divergence to the divergence determination threshold.

[0019] The physical embedded spatiotemporal graph attention model uses each depth node sensor as the vertex of the graph and the spatial distance and fluid connectivity between nodes as edge weights. The graph convolutional layer uses Chebyshev graph convolution kernels, and the gated temporal convolutional layer uses alternating stacks of one-dimensional causal convolution and gated linear units. The graph convolutional layer and the gated temporal convolutional layer are stacked alternately in 4 to 8 blocks.

[0020] In the physical constraint regularization module of the physical embedded spatiotemporal graph attention model, the residual terms of the Terzaghi earth pressure balance equation and the Darcy seepage equation are added to the loss function. The weight of the collapse risk probability task of the multi-task output head is 0.6, and the weight of the estimated collapse depth and location task is 0.4.

[0021] The attention mechanism layer of the physical embedded spatiotemporal graph attention model reduces the weights of nodes with a signal-to-noise ratio (SNR) lower than a preset SNR threshold and replaces them with interpolation features from adjacent nodes. The elastic weight solidification module calculates the diagonal elements of the Fisher information matrix of each weight parameter and applies a secondary penalty term to important weights.

[0022] The formation abrupt change intensity index is represented by the weighted average of the Körbek-Leibler divergence, the second derivative of the retaining wall differential pressure, and the probability of collapse risk. The initial suggested values ​​for the weight coefficients of the three factors are 0.3, 0.3, and 0.4, respectively.

[0023] The adaptive learning rate adjustment function for formation mutation is divided according to the range of formation mutation intensity index: when the formation mutation intensity index is less than the low threshold, the online learning rate remains at the basic learning rate; when the formation mutation intensity index is in the low to medium range, the online learning rate is adjusted to twice the basic learning rate and the elastic weight solidification module is activated; when the formation mutation intensity index is in the high to medium range, the online learning rate is adjusted to four times the basic learning rate and the penalty coefficient of the elastic weight solidification module is increased to three times the normal value; when the formation mutation intensity index is not lower than the high threshold, online weight updates are stopped and the system switches to a progressive expansion mode.

[0024] The progressive expansion mode refers to dynamically expanding independent convolutional columns for the current new stratigraphic type, freezing the weights of the original columns, and then laterally connecting and fusing the newly expanded convolutional columns with the original convolutional columns after the new stratigraphic data has accumulated to a sample size threshold.

[0025] In the training dataset of the physical embedded spatiotemporal graph attention model, the data is divided into training set, validation set and test set according to engineering projects, with a ratio of 7:1.5:1.5. The annotation content includes the collapse start time, estimated collapse depth and location and hazard level.

[0026] The low threshold is 0.3, the low-to-medium range is 0.3–0.6, the high-to-medium range is 0.6–0.85, and the high threshold is 0.85; the base learning rate is... ~ The sample size threshold is 300-500 samples; the preset threshold for signal-to-noise ratio is 1.5 times the lower limit of the normal distribution range of signal-to-noise ratio at each depth node; the normalized reference values ​​for the Kulbeck-Leibler divergence and the normalized reference values ​​for the second derivative of the wall differential pressure are respectively taken as the 90th percentile of the corresponding parameters in the historical data of the normal drilling section.

[0027] This invention combines empirical mode decomposition with independent component analysis to extract borehole wall collapse feature components from strong vibration interference signals. It then integrates dynamic early warning of wall differential pressure, minimum description length change point detection, and physical embedded spatiotemporal graph attention model to construct a multi-level, multi-dimensional borehole wall collapse risk early warning system. This solves the technical problem of not being able to provide real-time and accurate early warning of borehole wall collapse risk during impact drilling pile construction.

[0028] Empirical Mode Decomposition (EMD) adaptively decomposes nonlinear, non-stationary signals into physically meaningful intrinsic mode function components. Independent Component Analysis (ICA) further achieves statistically independent separation of impact vibration and weak collapse signals in the component space, thus overcoming the failure of traditional filtering methods due to the high overlap of their frequency bands. The physically embedded spatiotemporal graph attention model utilizes the spatial topology of the sensor network and physical constraint regularization to provide predictions consistent with physical laws even with sparse data or sensor failure. Furthermore, it prevents catastrophic forgetting when traversing new strata through an adaptive learning rate adjustment function based on strata mutation.

[0029] In summary, the present invention solves the technical problem mentioned in the background art of the inability to provide real-time and accurate early warning of the risk of borehole wall collapse during the construction of impact drilled piles. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a probability distribution map of collapse risk at each depth node.

[0032] Figure 3 The curve and trend extrapolation diagram of the time evolution of the differential pressure on the retaining wall at a depth of 28m are shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shown is a flowchart of a method for early warning of borehole wall collapse in impact drilling pile foundations provided by the present invention. This method includes the following steps:

[0035] S01. An array of ultrasonic transducers and an array of mechanical vibration acceleration sensors are deployed at multiple depth nodes inside the borehole to collect multi-frequency ultrasonic echo signals and mechanical vibration signals, and simultaneously collect data on mud level, mud density, pump flow rate and borehole wall side pressure.

[0036] S02. Perform empirical mode decomposition on the multi-frequency ultrasonic echo signal and mechanical vibration signal to obtain the intrinsic mode function component matrix of each channel. Then, perform independent component analysis on the intrinsic mode function component matrix of each channel to separate the characteristic component of the borehole wall collapse.

[0037] S03. Using mud level, mud density, and pump flow rate as inputs, a system of differential equations based on mass conservation and momentum conservation is established. The fourth-order Runge-Kutta method is used to solve the wall differential pressure time evolution curve online numerically. An exponentially moving weighted sliding window is introduced into the wall differential pressure time evolution curve for trend extrapolation. When the extrapolated predicted value is about to cross the Rankine active earth pressure critical value, a first-level warning is triggered. When the second derivative of the wall differential pressure exceeds the second-level warning threshold, a second-level emergency warning is triggered.

[0038] S04. The mud level, mud density, pump flow rate, borehole wall side pressure, and borehole wall collapse characteristic components are used to form multidimensional time series data of the whole borehole. The data is input into the minimum description length change point detection algorithm, and dynamic programming segmentation is performed using the Bayesian information criterion as the cost function. The time of abrupt change in borehole wall state is detected, and the Körbek-Leibler divergence of the multidimensional joint distribution before and after the change point is calculated. When the Körbek-Leibler divergence exceeds the divergence judgment threshold, a collapse risk event is judged and a hazard level score is output.

[0039] S05. Input the borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement value, and mud level time series data of each depth node into the physical embedded spatiotemporal map attention model, output the collapse risk probability and estimated collapse depth location, and dynamically adjust the online update learning rate of the physical embedded spatiotemporal map attention model by the formation mutation intensity index-driven formation mutation adaptive learning rate adjustment function.

[0040] S06. Based on the comprehensive assessment of Level 1 warning status, Level 2 emergency warning status, hazard level score, and collapse risk probability, the final warning signal is output according to the highest hazard level among the four. The current depth node data and stratum type label are written into the historical sample library for continuous learning and updating by the physical embedded spatiotemporal map attention model.

[0041] The ultrasonic transducer array consists of multiple ultrasonic transducers of different frequencies, ranging from 40 to 500 kHz. Different frequencies of ultrasonic echo signals have varying detection sensitivities to spalling and cracks at different depths of the borehole wall. Multi-frequency signals are combined to cover abnormal responses from shallow to deep layers of the borehole wall. The mechanical vibration acceleration sensor array is deployed on the drill pipe and borehole opening, with a sampling frequency of no less than 2000 Hz. It is used to capture broadband mechanical vibration signals generated during the operation of impact drills.

[0042] The empirical mode decomposition (EMD) is an adaptive nonlinear nonstationary signal decomposition method that decomposes a mixed signal into several intrinsic mode function (IMF) components, each representing vibration components at different time-frequency scales. Independent component analysis (ICA) is a blind source separation method that recovers independent source signals from a multi-channel mixed signal based on statistical independence. First, EMD is performed on the multi-frequency ultrasonic echo signals and mechanical vibration signals of each sensor channel to obtain the IMF component matrices for each channel. Then, the IMF component matrices of all channels are concatenated and input into ICA to separate the borehole wall collapse characteristic components, which are statistically independent of the impact vibration. The frequency range of the borehole wall collapse characteristic components is determined by collecting labeled data from test boreholes with known collapse conditions and performing spectral statistical analysis; a suggested range is 100–800 Hz. The combined use of Empirical Mode Decomposition (EMD) and Independent Component Analysis (ICA) effectively separates impact vibration noise from weak collapse signals in a statistical sense, solving the problem of weak collapse signals being masked under strong vibration interference and allowing the characteristic components of borehole wall collapse to be effectively utilized in subsequent steps. The combined processing of EMD and ICA extracts weak collapse signals under strong interference. Its technical advantages are: EMD adaptively decomposes nonlinear and non-stationary mixed signals into physically meaningful intrinsic mode function components, avoiding the distortion of non-stationary signals by traditional Fourier transform; ICA further recovers statistically independent source signals from the component space, fundamentally separating the mechanical vibration of the impact drill from the weak borehole wall collapse signal in the component domain, fundamentally overcoming the problem of traditional filters failing due to the high overlap of their frequency bands.

[0043] The borehole wall differential pressure, defined as the difference between the hydrostatic pressure of the drilling mud column and the lateral active earth pressure of the formation, is a core state variable reflecting the borehole wall's stability. The simultaneous differential equations based on mass and momentum conservation describe the dynamic relationship between drilling mud flow rate, borehole wall seepage rate, and drilling mud density over time. The time step for the fourth-order Runge-Kutta method's online numerical integration is set to 0.5–2. The time step is determined by the rate of change of the on-site pump flow; the window duration of the exponentially weighted sliding window is set to 30–120. The weight decay coefficient was determined by least-squares fitting of historical drilling data; the prediction time domain for trend extrapolation was set to 10–60. The Rankine active earth pressure critical value was determined by numerical simulation of the in-situ formation test data using an elastic-visco-plastic constitutive model. The secondary warning threshold was determined by taking 3 to 5 times the upper limit of the normal fluctuation range after statistically analyzing the normal fluctuation range of the second derivative of the wall differential pressure in multiple test boreholes. The wall differential pressure time evolution curve was calculated by inputting the state estimation of the mud level sensor noise using Kalman filtering. The process noise covariance matrix and the observation noise covariance matrix of the Kalman filter were determined by statistical analysis of the static calibration data of the mud level sensor. The dynamic wall differential pressure adaptive warning algorithm, based on the principle of physical conservation, realized the real-time prediction of the dynamic trend of the wall differential pressure before borehole instability, making up for the shortcomings of the traditional static threshold in not being able to detect the gradual instability process. At the same time, Kalman filtering effectively suppressed the influence of random noise from the level sensor on the accuracy of the wall differential pressure calculation.

[0044] The minimum description length change point detection algorithm is based on the principle of minimum description length. It treats the full-hole multi-dimensional time-series data as an information stream to be compressed, and aims to minimize the total encoding length required to describe the normal drilling section model and the abnormal section model to find the change point location. The formula for the Bayesian information criterion cost function is as follows: ;in The sum of squared residuals of each segment after partitioning and The ratio multiplied by The dimensionless quantity after that, This is the noise variance estimate. To change the number of points, For each segment of model parameters, Total number of data points The reference data points are used; the minimum description length change point detection algorithm uses a pruning dynamic programming algorithm to prune the cost function increment, reducing the computational complexity of change point detection to... The system is designed to meet the requirements of online real-time processing. The divergence judgment threshold is determined by statistical analysis of data from normal drilling sections and marked collapse event sections, taking the 97.5th percentile of the Körbek-Leibler divergence distribution in the normal drilling section. The hazard level score is obtained by hierarchical mapping of the ratio of Körbek-Leibler divergence to the divergence judgment threshold. The minimum description length change point detection algorithm defines anomalies from an information theory perspective, requiring no preset signal model. It has universal detection capability for abrupt changes in multidimensional time series data across the entire borehole under nonlinear and non-stationary strata conditions, solving the problem of slow response of traditional threshold methods to formation abrupt changes. The technical effect of the minimum description length change point detection algorithm is that it uses the shortest coding length in information theory as a unified measure of model complexity and data fit, enabling change point detection to automatically adapt to the statistical characteristics of different strata types without any prior distribution assumptions. This fundamentally overcomes the problems of large prediction deviations and difficulty in setting alarm thresholds in traditional models under nonlinear and non-stationary strata conditions.

[0045] The specific structure of the physical embedded spatiotemporal graph attention model is as follows: Each depth node sensor is constructed as a vertex of the graph, and the spatial distance and fluid connectivity between nodes are used as edge weights, forming a multi-depth node sensor graph. The graph input layer receives time-series data of borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement values, and mud level from each node. The graph convolutional layer uses Chebyshev graph convolution kernels to extract spatial domain structural features by aggregating features from adjacent nodes; the convolution kernel order is set to 2-4. The gated temporal convolutional layer uses alternating stacks of one-dimensional causal convolutions and gated linear units to extract the temporal domain features of each node's time series; the number of stacked layers is set to 3-6. Four to eight blocks of alternating stacked graph convolutional layers and gated temporal convolutional layers form the backbone for joint spatiotemporal feature extraction. The attention mechanism layer dynamically weights the spatiotemporal features of nodes at different depths; the weights are determined by the current signal-to-noise ratio (SNR) and historical importance of each node. Nodes with an SNR lower than a preset threshold have their weights reduced and replaced by interpolated features from adjacent nodes, achieving adaptive jump tolerance. The reasoning and physical constraint regularization module incorporates the residual terms of the Terzaghi earth pressure balance equation and the Darcy seepage equation into the loss function. These two residuals impose constraints on the output of the graph convolutional layer, ensuring that the model follows physical laws even with sparse training data. The multi-task output head branches from the shared encoder into two independent fully connected branches. The first branch outputs the collapse risk probability of each depth node, and the second branch outputs the estimated collapse depth location. The two task losses are weighted and summed. The weight for the collapse risk probability task is set to 0.6, and the weight for the estimated collapse depth location task is set to 0.4. The weights are determined by performing a grid search on the validation set. The elastic weight solidification module calculates the diagonal elements of the Fisher information matrix of each weight parameter during online updates of the physically embedded spatiotemporal graph attention model. It applies a secondary penalty term to important weights to prevent new stratigraphic data from overwriting old stratigraphic knowledge. The preset signal-to-noise ratio threshold is calculated by taking 1% of the lower limit of the normal distribution range of the signal-to-noise ratio of each depth node after statistically analyzing the normal distribution range of the signal-to-noise ratio in multiple sets of test well data.The 5x accuracy is confirmed. The technical advantages of the physically embedded spatiotemporal graph attention model are: graph convolutional layers aggregate information from adjacent nodes using the spatial topology of sensor networks, gated temporal convolutional layers capture the nonlinear dynamic features of time series, and the alternating stacking of these two layers achieves deep fusion of spatiotemporal features; physical constraint regularization enables the model to maintain physical inference capabilities even under conditions of sparse data or sensor failure; a multi-task mutual supervision mechanism allows the two prediction tasks—collapse risk probability and estimated collapse depth / location—to mutually verify each other, improving overall prediction robustness; the elastic weight solidification module fundamentally solves the catastrophic forgetting problem of online models when traversing new strata; the physically embedded... The steps for establishing the training dataset for the spatiotemporal graph attention model specifically include: collecting historical time-series data on borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement values, and mud level at various depth nodes during the construction of impact drilled piles in multiple engineering projects; manually annotating the periods in which borehole wall collapses occurred by geological engineers, with annotations including the collapse start time, estimated collapse depth and location, and hazard level; dividing the data into training, validation, and test sets according to engineering projects in a ratio of 7:1.5:1.5; and normalizing each feature dimension; the steps for training the physically embedded spatiotemporal graph attention model specifically include: using the Adam optimizer with an initial learning rate set to [value missing]. The batch size was set to 32–64, and the training rounds were set to 100–300 rounds. Early stopping was performed when the validation set loss did not decrease for 20 consecutive rounds. The loss function was a weighted sum of the binary cross-entropy loss of the collapse risk probability and the mean square error loss of the estimated collapse depth and location, plus the mean square error of the residual terms of the Terzaghi earth pressure balance equation and the Darcy seepage equation. The weights of each physical residual term were determined by Bayesian hyperparameter optimization on the validation set.

[0046] The adaptive learning rate adjustment function for the formation mutation is designed as follows: the weighted average of the Körbek-Leibler divergence, the second derivative of the retaining wall differential pressure, and the collapse risk probability is used as the formation mutation intensity index. The formula for the formation mutation intensity index is as follows: ;in For the Kulbeck-Leibler divergence, This is the normalized reference value for the Kulbeck-Leibler divergence. The second derivative of the differential pressure on the wall is... This is the normalized reference value for the second derivative of the wall differential pressure. The probability of collapse. , , The three weighting coefficients were initially suggested to be 0.3, 0.3, and 0.4, respectively, and were determined after a grid search on the validation set; the local stratum mutation intensity index At that time, the online learning rate of the physically embedded spatiotemporal graph attention model remained at the base learning rate. ;when At that time, the online learning rate was adjusted to Simultaneously, the elastic weight solidification module is activated to protect important weights; when At that time, the online learning rate was adjusted to The penalty coefficient for the elastic weight solidification module is increased to three times the normal value; when At this point, the physical embedded spatiotemporal graph attention model stops online weight updates and switches to a progressive expansion mode, dynamically expanding the independent graph convolutional columns for the new stratigraphic type, while freezing the weights of the original columns; the base learning rate... The learning rate was determined by conducting a learning rate warm-up experiment on the training set and selecting the learning rate corresponding to the most stable stage of loss function descent. The range is [insert range here]. ~ The progressive expansion model, after accumulating 300-500 samples of new stratigraphic data, performs lateral connection fusion between the convolutional columns of the newly expanded map and the original map. The fusion weights are determined through joint verification of the new and old stratigraphic data. The Kolb-Leibler divergence normalization reference value... Normalized reference value of the second derivative of the wall differential pressure The 90th percentile of the corresponding parameters in the historical data of the normal drilling section is taken respectively; the formation mutation adaptive learning rate adjustment function integrates the Kolb-Leibler divergence, the second derivative of the wall differential pressure and the collapse risk probability into a formation mutation intensity index, and drives the automatic switching of the online learning strategy of the physical embedded spatiotemporal map attention model in the interval division method. This fundamentally solves the problem of catastrophic forgetting of the online model when crossing new formations, while maintaining a low learning rate during smooth drilling to avoid model overfitting.

[0047] The specific implementation of step S01 is as follows: During the construction of the impact drilled pile, ultrasonic transducer arrays and mechanical vibration acceleration sensor arrays are respectively deployed at multiple depth nodes along the hole depth direction. The ultrasonic transducer array consists of multiple transducers with different operating frequencies, covering a frequency range of 40–500 kHz. Different frequencies of ultrasonic waves have varying penetration depths and scattering sensitivities within the borehole wall medium. Low-frequency bands are more sensitive to deep cracks, while high-frequency bands are more sensitive to shallow spalling. Combining multiple frequencies can cover anomalous responses from the shallow to deep layers of the borehole wall. Mechanical vibration acceleration sensor arrays are deployed at the drill pipe and borehole opening, with a sampling frequency of no less than 2000 Hz. It can capture broadband mechanical vibration signals generated during the operation of impact drills. At the same time, the system synchronously collects data on borehole mud level, mud density, pump flow rate, and borehole wall side pressure, providing a complete input dataset for subsequent differential pressure calculation and multidimensional time series analysis.

[0048] The specific implementation of step S02 is as follows: First, empirical mode decomposition (EMD) is performed on the multi-frequency ultrasonic echo signals and mechanical vibration signals of each sensor channel. EMD is an adaptive nonlinear nonstationary signal decomposition method. Through a sieving process, the mixed signal is decomposed step by step into several intrinsic mode function (IMF) components. Each IMF component represents a vibration component at a specific time-frequency scale. It does not require preset basis functions and is suitable for processing nonstationary strong interference signals. After the decomposition of each channel is completed, the corresponding IMF component matrix is ​​obtained. The IMF component matrices of all channels are then concatenated and input into the independent component analysis (ICA) algorithm. ICA is a blind source separation method that uses statistical independence as a criterion to recover the independent source signals from the multi-channel mixed input. Since the mechanical vibration signal of the impact drill and the weak signal of borehole wall collapse are statistically independent, ICA can essentially separate the two in the component domain and extract the borehole wall collapse characteristic component that is statistically independent of the impact vibration. The frequency range of this characteristic component is recommended to be 100–800 kHz. The specific range was determined by collecting labeled data from test boreholes with known collapse conditions and performing spectral statistical analysis.

[0049] The specific implementation of step S03 is as follows: Using mud level, mud density, and pump flow rate as dynamic inputs, a system of simultaneous differential equations based on mass and momentum conservation is established to describe the dynamic relationship between mud supply flow rate, borehole wall seepage rate, and mud density over time. To reduce the impact of random noise from the mud level sensor on calculation accuracy, Kalman filtering is used to estimate the state of the level signal before input. The process noise covariance matrix and observation noise covariance matrix of the Kalman filter are determined through statistical analysis of the sensor's static calibration data. The fourth-order Runge-Kutta method is used for online numerical integration of the simultaneous differential equations, with a time step set to 0.5–2. The real-time output of the wall differential pressure time evolution curve is determined by the rate of change of the on-site pump flow. An exponentially weighted sliding window is introduced into this curve for trend extrapolation, with the window duration set to 30–120 seconds. The weighted decay coefficient was determined by least-squares fitting of historical drilling data, and the prediction time domain was set to 10–60. A Level 1 warning is triggered when the extrapolated predicted value crosses the Rankine active earth pressure critical value, which is determined by numerical simulation of the in-situ test data of the strata using an elastic-visco-plastic constitutive model. A Level 2 emergency warning is triggered when the second derivative of the wall differential pressure exceeds the Level 2 warning threshold, which is 3 to 5 times the upper limit of the normal fluctuation range.

[0050] The specific implementation of step S04 is as follows: The mud level, mud density, pump flow rate, borehole wall side pressure, and the borehole wall collapse feature components extracted in step S02 are spliced ​​together into full-bore multidimensional time-series data, which is then input into the minimum description length change point detection algorithm. The minimum description length change point detection algorithm is based on the minimum description length principle, treating the multidimensional time-series data as an information stream to be compressed. It uses the Bayesian information criterion as the cost function, which comprehensively considers the residual sum of squares, noise variance estimation, number of change points, model parameter dimensions, and number of data points for each segment. A pruning dynamic programming algorithm is used to prune the increment of the cost function, reducing the computational complexity to... The system is designed to meet the requirements for online real-time processing. After detecting a sudden change in the borehole wall condition, the Kulbeck-Leibler divergence of the multidimensional joint distribution before and after the change point is calculated. The divergence threshold is taken as the 97.5th percentile of the Kulbeck-Leibler divergence distribution in the normal drilling section. When the divergence exceeds this threshold, a collapse risk event is identified. The hazard level score is output after being mapped from the divergence to the threshold.

[0051] The specific implementation of step S05 is as follows: The borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement values, and mud level time-series data of each depth node are input into a physically embedded spatiotemporal graph attention model. This model constructs each depth node sensor as a graph vertex, using the spatial distance between nodes and fluid connectivity as edge weights. The graph convolutional layer uses Chebyshev graph convolution kernels to aggregate spatial neighborhood features, with kernel order 2-4. The gated temporal convolutional layer extracts temporal dynamic features by alternately stacking one-dimensional causal convolutions and gated linear units, with 3-6 stacked layers. The graph convolutional layer and the gated temporal convolutional layer are alternately stacked in 4-8 blocks to achieve deep fusion of spatiotemporal features. The attention mechanism layer reduces the weights of nodes with a signal-to-noise ratio below a preset threshold, replacing them with interpolated features from adjacent nodes, achieving adaptive fault-tolerant inference. The physical constraint regularization module superimposes the Terzaghi earth pressure balance equation residuals and the Darcy seepage equation residuals into the loss function, ensuring that the model still follows physical laws even with sparse training data. The multi-task output head branches into two independent fully connected branches, outputting the collapse risk probability and estimated collapse depth / location respectively. The loss weights for the two tasks are 0.6 and 0.4, respectively. (Seismic abrupt change intensity index) The adaptive learning rate adjustment function driving formation mutation, expressed as the weighted average of the Körbek-Leibler divergence, the second derivative of the retaining wall differential pressure, and the collapse risk probability, is divided into intervals: when Online learning rate maintains basic learning rate ;when Time adjustment to And activate the elastic weight solidification module; when Time adjustment to Furthermore, the penalty coefficient for the elastic weight solidification module is increased to three times the normal value; when When the online weight update stops, it switches to a progressive expansion mode, which dynamically expands the independent map convolution columns for new strata, freezes the weights of the original columns, and performs lateral connection fusion after the new strata data has accumulated to 300-500 samples.

[0052] The specific implementation of step S06 is as follows: Combining the Level 1 and Level 2 emergency warning states output in step S03, the hazard level score output in step S04, and the collapse risk probability output in step S05, the final warning signal is output according to the highest hazard level among the four, ensuring the conservatism and safety of the warning. Simultaneously, the current data of each depth node and the corresponding stratigraphic type are written into the historical sample database in real time, allowing the physically embedded spatiotemporal map attention model to continuously learn and update during subsequent construction, forming a closed-loop adaptive warning mechanism that combines data-driven and physical constraints.

[0053] It should be noted that the key technologies of this invention include: a cascaded signal processing mechanism of empirical mode decomposition and independent component analysis, an early warning mechanism for dynamic trend extrapolation of wall differential pressure, and a physical embedded spatiotemporal map attention model and formation abrupt change adaptive learning rate adjustment mechanism. Empirical mode decomposition adaptively decomposes nonlinear, non-stationary mixed signals into intrinsic mode function components, while independent component analysis reconstructs the source signal in the component space using statistical independence criteria. The cascaded nature of these two mechanisms essentially separates weak collapse signals from strong vibration noise, overcoming the defect of traditional filters failing due to frequency band overlap. The dynamic trend extrapolation mechanism for wall differential pressure predicts instability trends in real time based on physical conservation equations, compensating for the inability of static thresholds to detect gradual changes. The physical embedded spatiotemporal graph attention model integrates spatial topology, temporal dynamics, and physical constraints to achieve spatiotemporal collaborative prediction of multiple deep nodes. The formation mutation adaptive learning rate adjustment mechanism automatically switches the learning strategy according to the formation mutation intensity index, and together with the elastic weight solidification module, it prevents catastrophic forgetting when crossing new formations. The three key technologies work together to enable the early warning system to maintain real-time and accurate early warning capabilities under adverse conditions such as strong interference, complex formations, and sparse data.

[0054] It should be noted that this invention also solves the following technical problem: During the construction of impact drilled piles, when the physically embedded spatiotemporal graph attention model traverses new strata, the statistical characteristics of the new strata differ significantly from those of the old strata. During online learning, the new strata data overwrites the weights already learned for the old strata, leading to a sharp degradation in the model's predictive ability for the old strata—a catastrophic forgetting problem. This invention calculates the diagonal elements of the Fisher information matrix for each weight parameter using an elastic weight solidification module, applying a secondary penalty term to important weights to constrain the update magnitude of key weights by new data. Simultaneously, when the strata mutation intensity index exceeds a high-value threshold, it switches to a progressive expansion mode, dynamically expanding the independent graph convolution columns for the new strata and freezing the original column weights. Lateral connection fusion is performed only after sufficient new strata data has accumulated. This effectively learns the characteristics of the new strata while retaining knowledge of the old strata, thus solving the technical problem of catastrophic forgetting in online models when traversing new strata.

[0055] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in the following points. First, empirical mode decomposition does not rely on any preset basis functions and can adaptively decompose mixed signals into intrinsic mode function components of different time-frequency scales. Independent component analysis reconstructs the source signal from the component space based on statistical independence. The cascaded use of the two methods essentially separates the impact vibration noise from the weak collapse signal, providing reliable feature input for subsequent early warning steps. Second, the dynamic model of the retaining wall differential pressure established based on the simultaneous differential equations of mass conservation and momentum conservation, combined with online numerical integration and exponential moving weighted trend extrapolation using the fourth-order Runge-Kutta method, can predict the moment when the retaining wall differential pressure crosses the critical value before instability occurs, achieving early perception of the gradual instability process. Furthermore, the minimum description length change point detection algorithm is based on information theory, requires no prior distribution assumptions, and can universally detect abrupt changes in multidimensional time series data. Finally, the physical constraint regularization and elastic weight solidification module together ensure the model's generalization ability and knowledge retention ability under new geological conditions, logically forming a complete closed-loop early warning system from signal extraction, physical modeling, information theory detection to deep spatiotemporal learning.

[0056] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0057] The specific implementation of step S01 is as follows: An ultrasonic transducer array and a mechanical vibration acceleration sensor array are deployed at multiple depth nodes within the hole. The ultrasonic transducer array consists of multiple transducers of different frequencies, ranging from 40 to 500 kHz. Multi-frequency combined coverage of anomalous responses from shallow to deep layers of the borehole wall. An array of mechanical vibration acceleration sensors is deployed on the drill pipe and borehole opening, with a sampling frequency of no less than 2000 Hz. Simultaneously collect data on mud level, mud density, pump flow rate, and borehole wall side pressure as input for subsequent steps.

[0058] The specific implementation of step S02 is as follows: Empirical mode decomposition (EMD) is performed on the multi-frequency ultrasonic echo signals and mechanical vibration signals of each sensor channel. Empirical mode decomposition is an adaptive nonlinear nonstationary signal decomposition method that decomposes the mixed signal step by step into several intrinsic mode function components. Let the first... The original signal of each channel is ,in The channel number has a value range of 100. , For time variables, This represents the total number of sensor channels. It is obtained after empirical mode decomposition. One intrinsic mode function component For the first The total number of intrinsic mode function components obtained from the channel empirical mode decomposition is adaptively determined by the empirical mode decomposition. The intrinsic mode function component matrix for each channel is also shown. The formula is expressed as follows:

[0059] ;

[0060] In the formula, For the first Channel 1 The component of the intrinsic mode function at the th intrinsic mode function in the th... The sampled value at each moment. The total number of sampling points. For the first Each sampling time point. The intrinsic mode function component matrices of all channels are concatenated row-wise to obtain the total component matrix. ,in The sum of the intrinsic mode function components of all channels is expressed by the following formula:

[0061] ;

[0062] right Independent component analysis (ICA) is a blind source separation method that uses statistical independence as a criterion to recover independent source signals from a multi-channel mixed signal. The separation results are... The formula is expressed as follows:

[0063] ;

[0064] In the formula, The unmixing matrix obtained from independent component analysis. Each row represents an independent source signal. The frequency range of the borehole wall collapse characteristic component, which is statistically independent of the impact vibration, is 100–800 kHz. The determination was made by collecting labeled data from test boreholes with known collapse conditions and performing spectral statistical analysis.

[0065] The specific implementation of step S03 is as follows: using mud level, mud density, and pump flow rate as inputs, establish a system of simultaneous differential equations based on mass conservation and momentum conservation to describe the mud replenishment flow rate. , leakage rate of the borehole wall Mud density The dynamic relationship over time, among which This is the measured value of the pump flow rate. Estimated by the permeability coefficient of the pore wall and the pressure difference between the inside and outside of the pore, all units are equal. Mud level The mass conservation equation is expressed as follows:

[0066] ;

[0067] In the formula, The cross-sectional area of ​​the hole is given in units of 1. , The overflow flow rate is measured by an orifice flow meter, and the unit is _____. , The process noise covariance matrix of the Kalman filter is obtained after state estimation of the liquid level sensor noise. Covariance matrix of observation noise The differential pressure of the wall was determined through statistical analysis of the static calibration data of the mud level sensor. The difference between the hydrostatic pressure of the drilling mud column and the lateral active earth pressure of the formation is expressed by the following formula:

[0068] ;

[0069] In the formula, This is the acceleration due to gravity, in units of 1. , The lateral active earth pressure of the stratum is expressed in units of 1. The critical value of Rankine active earth pressure was determined by numerical simulation of in-situ test data of the formation using an elastic-visco-plastic constitutive model. The unit is The fourth-order Runge-Kutta method is used to perform online numerical integration of the above differential equations, with a time step of [missing information]. Set to 0.5~2 This is determined by the rate of change of the flow rate of the on-site pump. An exponentially moving weighted sliding window is introduced for trend extrapolation, with the window duration set to 30–120 seconds. The weight decay coefficient was determined by least-squares fitting of historical drilling data. The exponentially moving weighted sliding window... Time weight The formula is expressed as follows:

[0070] ;

[0071] In the formula, This is the weighted attenuation coefficient, ranging from 0 to 1, determined by least-squares fitting of historical drilling data. The total number of data points within the sliding window, determined by the ratio of window duration to sampling interval. Trend extrapolation forecast. The formula is expressed as follows:

[0072] ;

[0073] In the formula, This is the weighted first derivative estimate of the wall differential pressure within the exponentially moving weighted sliding window, in units of... , For the prediction time domain, the range is set to 10–60. , This is a time-normalized reference value, with an empirical value of 1. This is used to ensure that the dimensions of each term in the formula are... When extrapolated predicted values Will cross A level one early warning is triggered. The second derivative of the wall differential pressure... Exceeding the Level II warning threshold The level 2 emergency warning was triggered at that time. Take 3 to 5 times the upper limit of the normal fluctuation range, in units of .

[0074] The specific implementation of step S04 is as follows: construct a full-hole multi-dimensional time-series data matrix by combining mud level, mud density, pump flow rate, borehole wall side pressure, and borehole wall collapse characteristic components. ,in Total number of data points is the feature dimension. The minimum description length variable point detection algorithm is input, and dynamic programming segmentation is performed using the Bayesian information criterion as the cost function. The cost function formula is as follows:

[0075] ;

[0076] In the formula, The sum of squares of the residuals of each segment after partitioning and The ratio multiplied by The dimensionless quantity after that, This is the noise variance estimate. To change the number of points, For reference data points, The cost function value is dimensionless. The algorithm uses pruning dynamic programming to prune the increment of the cost function, with a computational complexity of O(n log n). Level. After detecting the abrupt change in the pore wall state, calculate the Kourbek-Leibler divergence of the multidimensional joint distribution before and after the change point. The formula is expressed as follows:

[0077] ;

[0078] In the formula, After the change point The probability distribution values ​​of the dimensional feature. Before the change point The probability distribution values ​​of the dimensional features, both of which are dimensionless probability values. This is the dimensionless divergence value. Divergence threshold. Take the normal drilling section The 97.5th percentile of the distribution was determined by statistical analysis. The risk of collapse is determined by the hazard level score. It is obtained after hierarchical mapping.

[0079] The specific implementation of step S05 is as follows: each depth node sensor is constructed as a vertex of a graph, and the spatial distance and fluid connectivity between nodes are used as edge weights of the graph, thus forming a multi-depth node sensor graph. ,in For a set of nodes, Let be the set of edges. This is the adjacency weight matrix. The graph input layer receives borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement values, and mud level time-series data from each node, forming the node feature matrix. ,in The total number of nodes in the graph is equal to the number of depth nodes. The time window length, Input the feature dimension for each node; an empirical value is 4. Normalized graph Laplacian matrix. From the adjacency weight matrix The calculation yields the following formula:

[0080] ;

[0081] In the formula, for An identity matrix of order 1. For a degree matrix, its diagonal elements , For nodes With nodes The edge weights between them. The graph convolutional layer uses Chebyshev graph convolution kernels, with orders set to 2 to 4. Layered graph convolution output The formula is expressed as follows:

[0082] ;

[0083] In the formula, Let the order be the Chebyshev polynomial. For the first Layer Order-learnable parameters, For the normalized graph Laplace matrix The Chebyshev polynomials For activation function, , The network layer is numbered. Gated temporal convolutional layers consist of alternating stacks of one-dimensional causal convolutions and gated linear units, with the number of stacked layers set to 3–6. Layer-gated temporal convolution output The formula is expressed as follows:

[0084] ;

[0085] In the formula, For causal convolution operations, For element-wise multiplication, The hyperbolic tangent activation function is used. , For the first Layer-learnable convolutional kernels, , For the first Layer bias term, The input features are for the gated temporal convolutional layer. The attention mechanism layer dynamically weights the spatiotemporal features of nodes at different depths. Attention weights From signal-to-noise ratio With historical importance The decision is made jointly, and the formula is expressed as follows:

[0086] ;

[0087] In the formula, For nodes The signal-to-noise ratio of the current data, A preset threshold for the signal-to-noise ratio is set, which is 1.5 times the lower limit of the normal distribution range. For nodes The historical importance of a node is determined by the ratio of the number of times that node triggers an alert to the total number of times all nodes trigger alerts within a historical period. The initial value is set to [default value]. , The attention weights are dimensionless. The physical constraint regularization module adds the Terzaghi earth pressure balance equation residuals to the loss function. With Darcy's seepage equation residual term Total loss function The formula is expressed as follows:

[0088] ;

[0089] In the formula, The binary cross-entropy loss is the probability of collapse. To estimate the mean square error loss of the collapse depth location, , The weights for the physical residual terms are determined through Bayesian hyperparameter optimization on the validation set. , , , , All values ​​are dimensionless loss values. The elastic weight fixation module calculates each weight parameter during online model updates. diagonal elements of the Fisher information matrix The formula is expressed as follows:

[0090] ;

[0091] In the formula, This represents the total number of training samples for the old task. For the model in the sample The output probability on For the first One input sample, For the first A tag, This is the vector of all weight parameters of the model. This is the dimensionless Fisher information content. A secondary penalty term is applied to the important weights, and the penalty term... The formula is expressed as follows:

[0092] ;

[0093] In the formula, The first time the old task training ended The value of each weight. For the first The normalized reference value for each weight is taken as follows: The absolute value, For flexible weights, a fixed penalty coefficient is applied. This represents the dimensionless penalty loss value. (Formation abrupt change intensity index) The adaptive learning rate adjustment function for driving formation mutations dynamically adjusts the online update learning rate. The formula is expressed as follows:

[0094] ;

[0095] In the formula, This is the normalized reference value for the Kulbeck-Leibler divergence. The reference value for the normalized second derivative of the differential pressure in the wall protection section is taken as the reference value. Both are taken as the 90th percentile of the corresponding parameters in the historical data of the normal drilling section. The probability of collapse. , , These are weighting coefficients, with empirical values ​​of 0.3, 0.3, and 0.4, determined through a grid search on the validation set. This is a dimensionless index of stratigraphic abrupt change intensity. When... At that time, the online learning rate remained at the baseline learning rate. ;when At that time, the online learning rate was adjusted to Simultaneously activate the elastic weight solidification module; when At that time, the online learning rate was adjusted to The penalty coefficient for the elastic weight solidification module is increased to three times the normal value; when When the model stops updating weights online, it switches to a progressive expansion mode, which dynamically expands the independent graph convolution columns for the new stratigraphic type, freezes the weights of the original columns, and performs lateral connection fusion after the new stratigraphic data has accumulated to 300-500 samples. The learning rate was determined through warm-up experiments on the training set, with a range of [range missing]. ~ .

[0096] The specific implementation of step S06 is as follows: integrating the first-level warning state, the second-level emergency warning state, the hazard level score and the collapse risk probability, outputting the final warning signal according to the highest hazard level among the four, and writing the current depth node data and stratum type label into the historical sample library for continuous learning and updating by the physical embedded spatiotemporal map attention model.

[0097] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To illustrate the technical solution of this invention, technicians built a complete borehole collapse risk early warning system at a large bridge pile foundation construction site, and applied the method of this invention throughout the construction process of an impact drilled pile with a designed borehole depth of 50m for real-time monitoring and early warning verification.

[0098] The geological conditions at the construction site are complex, consisting of fill, gravel, silty clay, and strongly weathered rock layers from top to bottom. The thickness and mechanical parameters of each layer vary significantly, presenting a typical working condition for verifying the cross-stratum adaptive capability of this invention. Ultrasonic transducer arrays and mechanical vibration acceleration sensor arrays were deployed at five depth nodes along the borehole depth: 10m, 20m, 30m, 40m, and 50m. The ultrasonic transducer arrays operated at frequencies of 40 Hz and 40 Hz. 100 250 500 There are a total of 4 frequencies; the sampling frequency of the mechanical vibration acceleration sensor array is set to 4000. The data acquisition is synchronized with sensors for mud level, mud density, pump flow rate, and borehole wall side pressure, with a data acquisition cycle of 0.5 seconds. .

[0099] During the signal acquisition phase of step S01, the system continuously acquires multi-frequency ultrasonic echo signals and mechanical vibration signals at various depth nodes. When the borehole depth reaches approximately 28m, the mechanical vibration acceleration sensor records a significant increase in broadband vibration energy, while at 250... The abnormal fluctuations in the echo amplitude of the ultrasonic transducer indicate that there may be signs of early spalling in the borehole wall in this area.

[0100] In the signal decomposition and feature extraction stage of step S02, the system performs empirical mode decomposition on the signals of each channel. Each channel is decomposed to obtain 6 to 8 intrinsic mode function components. The intrinsic mode function component matrices of all channels are then concatenated and input into independent component analysis. The separated borehole wall collapse feature components are concentrated in the range of 180 to 650. The frequency band is statistically independent of the main frequency band of the impact drill mechanical vibration, and the amplitude of the characteristic component shows a significant increase at the 28m depth node.

[0101] In the dynamic early warning stage of differential pressure on the retaining wall in step S03, based on the simultaneous differential equations of mass conservation and momentum conservation, the fourth-order Runge-Kutta method is used to calculate the differential pressure using a 1-dimensional differential pressure. For online integration over a time step, the mud level signal is input for calculation after noise reduction via Kalman filtering. The time evolution curve of the wall differential pressure is shown below. Figure 3 As shown, it can be clearly seen that the differential pressure of the protective wall shows a continuous decreasing trend when the borehole depth reaches approximately 28m. (At 60...) The exponentially shifted trend extrapolation results for the window length show that the predicted wall differential pressure will be around 35. Upon crossing the Rankine active earth pressure critical value, the system immediately triggered a Level 1 warning. Simultaneously, the calculated value of the second derivative of the retaining wall differential pressure exceeded the Level 2 warning threshold, prompting the system to further trigger a Level 2 emergency warning.

[0102] In step S04, the change point detection and hazard level scoring stage, the system constructs multi-dimensional time-series data from all monitoring data at the five depth nodes throughout the borehole, and inputs this data into the minimum description length change point detection algorithm. The algorithm is applied approximately 15 hours after the borehole depth reaches 28m. When a sudden change in the borehole wall state was detected, the Korbek-Leibler divergence before and after the change point was calculated. The result exceeded the divergence judgment threshold by 1.8 times. After classification mapping, the output hazard level score was 4 (out of 5), and it was judged as a high-risk collapse event. The changes in the main monitoring parameters are shown in Table 1.

[0103] Table 1. Comparison of key monitoring parameters before and after the abrupt change in borehole wall condition at a depth of 28m.

[0104]

[0105] As can be seen from Table 1, the significant decrease in the wall differential pressure and the significant increase in the characteristic component of borehole wall collapse are highly synchronized in time, verifying the consistency of the multi-dimensional fusion early warning mechanism of the present invention.

[0106] In the inference stage of the physically embedded spatiotemporal map attention model in step S05, the system inputs the borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement values, and mud level time series data of the five depth nodes into the trained physically embedded spatiotemporal map attention model. The model outputs a collapse risk probability of 0.91 for the 28m depth node, estimating the collapse depth to be in the range of 26–30m. At this time, the formation abrupt change intensity index... The calculated result is 0.78, which falls within the range. During the specified interval, the system will adjust the online learning rate to [a certain value]. The penalty coefficient of the elastic weight solidification module is increased to three times the normal value to protect the learned upper stratigraphic knowledge from being overwritten by new stratigraphic data. The distribution of collapse risk probability at each depth node is as follows: Figure 2 As shown, it can be clearly seen that the risk probability of the 28m node is significantly higher than that of other nodes, and the spatial positioning results are highly consistent with the actual collapse location.

[0107] In the comprehensive early warning output stage of step S06, the system integrates the Level 1 early warning, Level 2 emergency early warning, Level 4 hazard score, and a collapse risk probability of 0.91, and outputs the final early warning signal according to the highest hazard level. The on-site monitoring terminal immediately issues a red alarm, prompting construction personnel to stop the impact and take grouting measures. Approximately 20 minutes after the early warning signal is triggered... On-site observation revealed a sharp increase in slurry return at the borehole opening, confirming the occurrence of a localized borehole wall collapse. The early warning method of this invention provides warnings approximately 55 minutes before the collapse occurs. The Level 1 early warning has been triggered, achieving early warning. Current depth node data and silty clay layer stratigraphic type labels are being written into the historical sample database in real time for the model to continue learning.

[0108] Compared to traditional manual inspections and static threshold alarms using single sensors, this invention represents a technological advancement in the following aspects: Traditional methods rely on engineers' sensory judgment and comparison with fixed thresholds, failing to extract weak collapse features from strong vibration interference signals. This invention, however, adaptively decomposes non-stationary signals into intrinsic mode function components through empirical mode decomposition, and then achieves statistical source signal separation in the component space through independent component analysis, enabling the weak collapse signal to be fundamentally extracted from strong vibration noise. Traditional static threshold methods cannot detect the gradual instability process of the retaining wall differential pressure. This invention, based on physical conservation equations, constructs a dynamic differential equation system and uses online integration and exponential moving weighted trend extrapolation via the fourth-order Runge-Kutta method to predict the moment of crossing the critical value before instability occurs, achieving early warning. Furthermore, traditional methods cannot automatically adjust model parameters when crossing new strata. This invention, through the synergistic effect of an elastic weighting module and a progressive expansion mode, effectively learns the characteristics of new strata while retaining knowledge of old strata, fundamentally solving the catastrophic forgetting problem of online models and enabling the early warning system to have continuous adaptive capabilities.

[0109] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0110] Table 2. Variable Explanation Table (Part 1)

[0111]

[0112] Table 3. Variable Explanation Table (Part Two)

[0113]

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of borehole wall collapse risk in impact drilled pile foundations, characterized in that, Includes the following steps: An array of ultrasonic transducers and an array of mechanical vibration acceleration sensors are deployed at multiple depth nodes inside the borehole to collect multi-frequency ultrasonic echo signals and mechanical vibration signals, and simultaneously collect data on mud level, mud density, pump flow rate and borehole wall side pressure. Empirical mode decomposition was performed on the multi-frequency ultrasonic echo signal and the mechanical vibration signal to obtain the intrinsic mode function component matrix of each channel. Then, independent component analysis was performed on the intrinsic mode function component matrix of each channel to separate the characteristic component of the borehole wall collapse. Using mud level, mud density, and pump flow rate as inputs, a system of simultaneous differential equations based on mass and momentum conservation is established. The fourth-order Runge-Kutta method is used to solve the time evolution curve of the wall differential pressure online numerically. An exponentially moving weighted sliding window is introduced into the wall differential pressure time evolution curve for trend extrapolation. When the extrapolated predicted value is about to cross the Rankine active earth pressure critical value, a first-level warning is triggered. When the second derivative of the wall differential pressure exceeds the second-level warning threshold, a second-level emergency warning is triggered. The mud level, mud density, pump flow rate, borehole wall lateral pressure, and borehole wall collapse characteristic components are used to construct multidimensional time-series data of the entire borehole. The data is input into the minimum description length change point detection algorithm, and dynamic programming segmentation is performed using the Bayesian information criterion as the cost function. The time of abrupt change in borehole wall state is detected, and the Körbek-Leibler divergence of the multidimensional joint distribution before and after the change point is calculated. When the Körbek-Leibler divergence exceeds the divergence judgment threshold, a collapse risk event is determined and a hazard level score is output. The borehole wall lateral pressure, borehole wall displacement, ultrasonic diameter measurement value, and mud level time series data of each depth node are input into the physical embedded spatiotemporal map attention model, which outputs the collapse risk probability and estimated collapse depth location. The online update learning rate of the physical embedded spatiotemporal map attention model is dynamically adjusted by the formation mutation intensity index-driven formation mutation adaptive learning rate adjustment function. By combining the Level 1 warning status, Level 2 emergency warning status, hazard level score and collapse risk probability, the final warning signal is output according to the highest hazard level among the four, and the current depth node data and stratum type label are written into the historical sample library for continuous learning and updating of the physical embedded spatiotemporal map attention model.

2. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 1, characterized in that, The ultrasonic transducer array consists of multiple ultrasonic transducers of different frequencies, ranging from 40 to 500 kHz. The mechanical vibration acceleration sensor array is deployed on the drill pipe and at the borehole opening, with a sampling frequency of not less than 2000 Hz. .

3. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 2, characterized in that, The empirical mode decomposition refers to first adaptively decomposing the multi-frequency ultrasonic echo signals and mechanical vibration signals of each sensor channel, and then stitching together the intrinsic mode function component matrices of all channels before inputting them into independent component analysis. The frequency range of the borehole wall collapse characteristic components is 100–800 Hz. .

4. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 3, characterized in that, In the calculation of the time evolution curve of the wall differential pressure, the time step of the fourth-order Runge-Kutta method is 0.5 to 2 seconds. The noise of the mud level sensor is estimated using Kalman filtering before being input into the calculation.

5. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 4, characterized in that, The window duration of the exponentially moving weighted sliding window is 30–120 seconds. The prediction time domain for trend extrapolation is 10–60. The critical value of Rankine active earth pressure was determined by numerical simulation using an elastic-visco-plastic constitutive model based on in-situ test data of the strata.

6. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 5, characterized in that, The secondary early warning threshold is determined by taking 3 to 5 times the upper limit of the normal fluctuation range of the second derivative of the differential pressure of the wall in the normal drilling section.

7. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 6, characterized in that, The minimum description length variable point detection algorithm employs a pruning dynamic programming algorithm to prune the cost function increment, reducing the computational complexity to... Level; the divergence determination threshold is taken as the 97.5th percentile of the Körbek-Leibler divergence distribution in the normal drilling section; the hazard level score is obtained by mapping the ratio of Körbek-Leibler divergence to the divergence determination threshold.

8. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 7, characterized in that, The physical embedded spatiotemporal graph attention model uses each depth node sensor as the vertex of the graph, and the spatial distance and fluid connectivity between nodes as edge weights. The graph convolutional layer uses Chebyshev graph convolution kernels, and the gated temporal convolutional layer uses alternating stacks of one-dimensional causal convolution and gated linear units. The graph convolutional layer and the gated temporal convolutional layer are stacked alternately in 4 to 8 blocks.

9. The method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 8, characterized in that, The physical constraint regularization module of the physical embedded spatiotemporal graph attention model adds the residual terms of the Terzaghi earth pressure balance equation and the Darcy seepage equation to the loss function. The weight of the collapse risk probability task of the multi-task output head is 0.6, and the weight of the estimated collapse depth and location task is 0.

4.

10. A method for early warning of borehole wall collapse risk in impact drilling pile foundations according to claim 9, characterized in that, The attention mechanism layer of the physical embedded spatiotemporal graph attention model reduces the weights of nodes with a signal-to-noise ratio (SNR) lower than a preset SNR threshold and replaces them with interpolation features from adjacent nodes. The elastic weight solidification module calculates the diagonal elements of the Fisher information matrix of each weight parameter and applies a secondary penalty term to important weights.

Images