Intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pit

Abstract

The present application relates to a kind of adjacent subway foundation pit bidirectional disturbance intelligent monitoring early warning system, specifically, to the field of foundation pit engineering information construction and control, by the synergistic cooperation of data acquisition module, digital twin prediction module, optimization decision module and online fine-tuning module, bidirectional disturbance intelligent monitoring and closed-loop control of adjacent subway foundation pit construction whole process are realized;Digital twin prediction module outputs deformation prediction trajectory based on historical monitoring state sequence and historical construction parameter sequence, provides basis for risk prediction;Optimization decision module automatically solves optimal construction parameter sequence according to deformation prediction trajectory and executes the construction parameter of first control step, upgrades artificial experience decision to data-driven intelligent regulation and control;Online fine-tuning module updates the output layer weight of lightweight prediction model using prediction residual, so that digital twin prediction module adaptively tracks the change of soil mechanics properties, and long-term maintains prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of information-based construction and control of foundation pit engineering, and more specifically, to a two-way intelligent monitoring and early warning system for bidirectional disturbances in adjacent subway foundation pits. Background Technology

[0002] As the density of urban underground space development continues to rise, more and more deep and large foundation pit projects are being built adjacent to existing operational subway tunnels. Under such conditions, the excavation and unloading of the foundation pit will inevitably cause a redistribution of the stress field in the surrounding strata, thereby inducing additional settlement and horizontal displacement of the tunnel structure. At the same time, the repeated cyclic dynamic load of subway trains will also weaken the long-term stability of the foundation pit support system, forming a typical two-way disturbance effect. To control such risks, current engineering practices generally deploy automated monitoring systems with total stations and microelectromechanical systems (MEMS) as the core, which collect multi-dimensional data such as tunnel settlement and support structure displacement in real time, and use two-way disturbance prediction models to classify risk levels. When the monitoring data reaches a preset threshold, the system issues a yellow or orange alarm signal, prompting on-site technicians to intervene. However, the specific implementation of intervention measures still relies heavily on the personal experience of engineers, which usually manifests as temporarily adjusting construction parameters such as excavation speed, layer thickness, number of dewatering wells opened and closed, or pre-applied axial force to the supports based on the deformation rate.

[0003] The aforementioned manual decision-making model has significant limitations. On the one hand, there is a strong coupling relationship between foundation pit excavation, dewatering operations, and support construction. Changes in a single parameter often trigger chain reactions. For example, blindly reducing the excavation speed may lead to delayed manifestation of additional settlement caused by continuous dewatering, which may exacerbate the deformation trend. On the other hand, soil consolidation and pore water pressure dissipation have obvious time lag characteristics. The actual effect of current adjustments often takes several hours or even longer to be fully reflected in monitoring data, and engineers cannot accurately predict this delayed response based on intuition. In addition, construction decisions are also subject to multiple rigid constraints, including that the construction schedule cannot be excessively slowed down, dewatering operations cannot exceed the groundwater level protection line, and the axial force of the supports cannot exceed the structural design bearing capacity. When the deformation rate suddenly increases, on-site personnel often make hasty decisions without quantitative analysis tools, which can easily lead to two extreme situations: excessive adjustment causing new stress concentration or insufficient adjustment failing to curb deformation development. In severe cases, it may lead to excessive differential settlement of the tunnel, triggering subway shutdown for maintenance, causing huge economic losses and social impacts.

[0004] Currently, there have been research attempts in academia to introduce model predictive control theory into the regulation of foundation pit excavation. However, most schemes assume that the system is a linear model and do not consider the continuous disturbance source of subway train dynamic load, which is significantly different from the complex working conditions of nonlinearity, time-varying and multiple disturbances in actual engineering. Therefore, there is an urgent need for an intelligent regulation method that can solve constrained nonlinear optimization problems online. This method uses a two-way disturbance prediction model as an internal predictor to continuously optimize the sequence of construction parameters for several future steps. Under the premise of meeting multiple constraints such as construction period, safety and environmental protection, it can realize the real-time adaptive adjustment of excavation and dewatering parameters, fundamentally making up for the lack of predictability of human experience-based decision-making. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing an intelligent monitoring and early warning system for bidirectional disturbances in adjacent subway foundation pits, thereby resolving the issues raised in the background section.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a data acquisition module, a digital twin prediction module, an optimization decision-making module, and an online fine-tuning module that are sequentially connected in communication, wherein; Data acquisition module: acquires multi-dimensional monitoring data and construction parameters in real time during the construction of the adjacent subway foundation pit. The data acquisition module stores the acquired multi-dimensional monitoring data and construction parameters in chronological order as historical monitoring status sequence and historical construction parameter sequence, respectively. Digital twin prediction module: Receives historical construction parameter sequences and historical monitoring status sequences, constructs a lightweight prediction model, takes the historical construction parameter sequences and historical monitoring status sequences as input, and outputs the deformation prediction trajectory for the next 24 hours. The deformation prediction trajectory includes the time-series prediction value of the tunnel convergence deformation rate. Optimization Decision Module: Constructs a finite-time domain optimization objective function within each control cycle. The finite-time domain optimization objective function includes a deformation exceeding penalty term, a construction cost term, and a safety barrier function term. Using the construction parameter sequence of multiple future control steps as decision variables, and under the conditions of satisfying the upper and lower limits of construction parameters and the rate of change constraints, it solves for the optimal construction parameter sequence that minimizes the value of the finite-time domain optimization objective function. Only the construction parameters of the first control step in the optimal construction parameter sequence are executed to adjust the current construction state. Online fine-tuning module: After each control cycle, multi-dimensional monitoring data is obtained from the data acquisition module, deformation prediction trajectory is obtained from the digital twin prediction module, the time-series prediction values ​​at corresponding times in the multi-dimensional monitoring data and deformation prediction trajectory are compared, the prediction residual is calculated, and the output layer weights of the lightweight prediction model are updated online based on the recursive least squares algorithm, so that the digital twin prediction module can adaptively track the changes in soil mechanical properties. In a preferred embodiment, the data acquisition module acquires tunnel arch settlement through a total station, horizontal displacement of retaining pile top through a microelectromechanical system sensor, pore water pressure through a pore water pressure gauge, and excavation rate, dewatering flow rate and support axial force increment through a construction control system. The total station, microelectromechanical system sensor, pore water pressure gauge and construction control system are all sensors. The data acquisition module records the unprocessed values ​​directly acquired from the sensors as the raw measurement values. The data acquisition module maintains a circular buffer for each sensor, which stores the most recent raw measurement values ​​and their hardware timestamps in a first-in-first-out manner. The data acquisition module emits synchronization pulses at fixed intervals, mapping the hardware timestamp of each raw measurement value to the most recent system clock tick; For sensors that have multiple samples within the same system clock cycle, the arithmetic mean is taken as the representative value. System clock cycles that do not generate samples are marked as missing, thus obtaining the discrete sequence of each sensor on the system time axis. The data acquisition module performs sliding window processing on the discrete sequences of each sensor. The length of the sliding window is adaptively adjusted according to the degree of local fluctuation in the discrete sequence. The local fluctuation metric is defined as the median of the absolute value of the difference between two adjacent original measurements within the sliding window.

[0007] In a preferred embodiment, the data acquisition module calculates the median and absolute deviation of the median within a defined sliding window. The absolute deviation of the median is equal to the median of the absolute differences between each original measurement value and the median within the window. The data acquisition module constructs a robust statistical score, the numerator of which is the difference between the current original measurement value and the median within the window multiplied by 0.6745, and the denominator is the sum of the absolute deviation of the median within the window and 10 to the power of negative 6. If the absolute value of the robust statistical score is greater than 3.5, the current original measurement value is identified as an outlier and replaced with the median within the window; otherwise, the original value is retained, and the replaced sequence is recorded as the cleaned sequence.

[0008] In a preferred embodiment, the data acquisition module aligns the cleaning sequences of all sensors according to the system clock cycle and categorizes them according to predefined physical quantity categories. At each system clock cycle, it acquires the tunnel arch settlement corresponding to the system clock cycle from the cleaning sequence of the total station, the horizontal displacement of the retaining pile top corresponding to the system clock cycle from the cleaning sequence of the microelectromechanical system (MEMS) sensor, and the pore water pressure corresponding to the system clock cycle from the cleaning sequence of the pore water pressure gauge. The tunnel arch settlement, the horizontal displacement of the retaining pile top, and the pore water pressure are assembled into multi-dimensional monitoring data. Simultaneously, the data acquisition module directly acquires the excavation rate, dewatering flow rate, and support axial force increment corresponding to the system clock cycle from the construction control system, and assembles the excavation rate, dewatering flow rate, and support axial force increment into construction parameters. The data acquisition module assembles the multi-dimensional monitoring data from the most recent system clock cycles into a historical monitoring status sequence in chronological order, and assembles the construction parameters from the most recent system clock cycles into a historical construction parameter sequence in chronological order.

[0009] In a preferred embodiment, the specific process of constructing a lightweight prediction model in the digital twin prediction module is as follows: The historical monitoring state sequence and the historical construction parameter sequence are concatenated along the feature dimension to form a joint feature tensor. The joint feature tensor is then input into a bidirectional long short-term memory recurrent neural network (LSTM). The bidirectional LSTM recurrent neural network is used as an encoder, which consists of a forward LSTM recurrent neural network and a backward LSTM recurrent neural network connected in parallel. For each time step in the joint feature tensor, the forward LSTM recurrent neural network generates a forward hidden state, and the backward LSTM recurrent neural network generates a backward hidden state. The forward hidden state and the backward hidden state are concatenated, and the concatenated result is the encoder hidden state. The digital twin prediction module also includes a decoder, which is a single-layer long short-term memory recurrent neural network. The initial hidden state and initial cell state of the decoder are taken from the forward hidden state and forward cell state of the encoder at the last time step. The digital twin prediction module constructs a construction perception attention mechanism, which calculates the attention weight by concatenating the decoder hidden state, encoder hidden state and construction parameters and then summing them by an attention scoring function to obtain the context vector. For each future prediction time step, the temporal predicted value of the tunnel convergence deformation rate of the previous time step is concatenated with the context vector as the decoder input. During the training process, the digital twin prediction module adds a physical information regularization term, which is defined as the squared mean of the deviation between the second-order difference of the temporal predicted value of the tunnel convergence deformation rate and the rate of change of the construction parameters. The physical information regularization term and the negative log-likelihood loss are weighted and summed to form the total loss function. After training, the model parameters are frozen to obtain a lightweight prediction model.

[0010] In a preferred embodiment, the digital twin prediction module uses a hybrid density network as the output layer of the decoder. The hybrid density network generates the mixing coefficients, mean, and log-variance of the Gaussian mixture model by passing the decoder's hidden state through three independent fully connected branches. The digital twin prediction module calculates the probability density function of the time-series predicted value of the tunnel convergence deformation rate at the current time step based on the mixing coefficient, mean, and log-variance. It uses the weighted expected value of the mean of each component as the time-series predicted value of the tunnel convergence deformation rate at the current time step, and saves the log-variance sequence for each time step. The digital twin prediction module outputs the deformation prediction trajectory and the corresponding log-variance sequence for the next 24 hours. The deformation prediction trajectory includes the time-series predicted value of the tunnel convergence deformation rate.

[0011] In a preferred embodiment, the process of constructing a finite-time domain optimization objective function in the optimization decision module within each control cycle specifically includes: The control time domain length is set to eight control steps, and the decision variable is the sequence of construction parameters for the next eight control steps. The constructed finite-time domain optimization objective function is a weighted sum of a deformation exceeding penalty term, a construction cost term, and a safety barrier function term, which is accumulated over each future control step. The deformation exceeding penalty term adopts a piecewise exponential function form, taking zero when the time-series predicted value of the tunnel convergence deformation rate is below the safety threshold, growing exponentially when it is between the safety threshold and the warning threshold, and increasing sharply according to a quadratic function when it exceeds the warning threshold. The construction cost term is defined as a weighted sum of the square of the deviation of the excavation rate from the reference value, the square of the ratio of precipitation flow to the maximum allowable precipitation flow, and the square of the ratio of the increase in support axial force to the maximum allowable increase in support axial force. The safety barrier function term generates a steep cubic penalty gradient when the time-series predicted value of the tunnel convergence deformation rate approaches the critical threshold. The upper and lower limits of construction parameters and the rate of change constraints are added as inequality constraints to the optimization problem.

[0012] In a preferred embodiment, the process of finding the optimal construction parameter sequence that minimizes the value of the finite-time optimization objective function specifically involves: First, the historical monitoring status sequence, historical construction parameter sequence, and candidate construction parameter sequence at the current moment are passed to the digital twin prediction module. The digital twin prediction module performs forward propagation and returns the time-series predicted values ​​of the tunnel convergence deformation rate and its logarithmic variance sequence for each time step covered by multiple future control steps. Then, the optimization decision module introduces the upper and lower limit constraints and the rate of change constraints of the construction parameters into the finite-time domain optimization objective function through the Lagrange multiplier method, and constructs the augmented Lagrange function. Next, the algorithm in the quasi-Newton method is used for iterative solution. In each iteration, the gradient of the augmented Lagrangian function with respect to each component in the candidate construction parameter sequence is automatically calculated by differentiation, and the candidate solution is updated according to the gradient direction. The maximum number of iterations is set to two hundred, and the gradient convergence threshold is set to ten to the power of negative four. The optimal construction parameter sequence is obtained after iterative convergence. The optimization decision module only sends the excavation rate, dewatering flow rate, and support axial force increment of the construction parameters in the first control step of the optimal construction parameter sequence to the excavator control system, the variable frequency dewatering pump control system, and the hydraulic support system, respectively. The construction parameters of the remaining control steps are discarded and wait for the next control cycle to be solved again.

[0013] In a preferred embodiment, after each control cycle, the online fine-tuning module acquires multi-dimensional monitoring data of all system clock beats within the current control cycle from the data acquisition module, obtains the deformation prediction trajectory at the corresponding time from the digital twin prediction module, extracts the time-series predicted value of the tunnel convergence deformation rate from the deformation prediction trajectory, subtracts the measured value of tunnel arch settlement from the multi-dimensional monitoring data from the corresponding time value of the time-series predicted value of tunnel convergence deformation rate, and calculates the prediction residual for each system clock beat; the online fine-tuning module calculates the absolute median of the prediction residual sequence, executes an update step when the absolute median is less than a preset residual threshold, and executes an update step when the absolute median is not less than the preset residual threshold. Skip this update; the online fine-tuning module extracts the decoder hidden state corresponding to each system clock beat from the lightweight prediction model of the digital twin prediction module, stacks them in chronological order to form a feature matrix, and organizes the measured values ​​of tunnel arch settlement in multidimensional monitoring data in chronological order to form a target value vector; the online fine-tuning module uses a recursive least squares algorithm with a forgetting factor to sequentially update the output layer weights and output layer biases of the lightweight prediction model. For each system clock beat, the gain vector is calculated, the covariance matrix is ​​updated, and the output layer weights and output layer biases are updated using the prediction residuals. After traversing all system clock beats in the current control cycle, the updated output layer weights and updated output layer biases are obtained.

[0014] In a preferred embodiment, the online fine-tuning module replaces the weight matrix of the output layer in the lightweight prediction model with the updated output layer weights and the bias of the output layer with the updated output layer bias, while keeping the parameters of the remaining layers in the lightweight prediction model unchanged. The online fine-tuning module calculates the magnitude of the difference between the updated and unupdated output layer weights. When the magnitude exceeds a preset maximum single weight change, the updated output layer weights are pruned so that the magnitude of the difference between the updated and unupdated output layer weights equals the preset maximum single weight change, with the pruning direction consistent with the original update direction. After the online fine-tuning module completes the parameter replacement, the digital twin prediction module uses the updated lightweight prediction model for prediction in the next control cycle, enabling the digital twin prediction module to adaptively track changes in soil mechanical properties.

[0015] The beneficial effects of this invention are as follows: Through the coordinated operation of the data acquisition module, digital twin prediction module, optimization decision-making module, and online fine-tuning module, bidirectional intelligent monitoring and closed-loop control of disturbances throughout the entire construction process of the subway foundation pit are realized; the digital twin prediction module outputs deformation prediction trajectory based on historical monitoring state sequences and historical construction parameter sequences, providing a basis for risk prediction; the optimization decision-making module automatically solves the optimal construction parameter sequence based on the deformation prediction trajectory and executes the construction parameters for the first control step, upgrading manual experience-based decision-making to data-driven intelligent control; the online fine-tuning module uses prediction residuals to update the output layer weights of the lightweight prediction model online, enabling the digital twin prediction module to adaptively track changes in soil mechanical properties and maintain prediction accuracy over the long term. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1: This example provides the following... Figure 1-2 The system illustrates an intelligent monitoring and early warning system for bidirectional disturbances in adjacent subway foundation pits, specifically comprising: a data acquisition module, a digital twin prediction module, an optimization decision-making module, and an online fine-tuning module, which are sequentially connected via communication. Data acquisition module: acquires multi-dimensional monitoring data and construction parameters in real time during the construction of the adjacent subway foundation pit. The data acquisition module stores the acquired multi-dimensional monitoring data and construction parameters in chronological order as historical monitoring status sequence and historical construction parameter sequence, respectively. Digital twin prediction module: Receives historical construction parameter sequences and historical monitoring status sequences, constructs a lightweight prediction model, takes the historical construction parameter sequences and historical monitoring status sequences as input, and outputs the deformation prediction trajectory for the next 24 hours. The deformation prediction trajectory includes the time-series prediction value of the tunnel convergence deformation rate. Optimization Decision Module: Constructs a finite-time domain optimization objective function within each control cycle. The finite-time domain optimization objective function includes a deformation exceeding penalty term, a construction cost term, and a safety barrier function term. Using the construction parameter sequence of multiple future control steps as decision variables, and under the conditions of satisfying the upper and lower limits of construction parameters and the rate of change constraints, it solves for the optimal construction parameter sequence that minimizes the value of the finite-time domain optimization objective function. Only the construction parameters of the first control step in the optimal construction parameter sequence are executed to adjust the current construction state. Online fine-tuning module: After each control cycle, multi-dimensional monitoring data is obtained from the data acquisition module, deformation prediction trajectory is obtained from the digital twin prediction module, the time-series prediction values ​​at corresponding times in the multi-dimensional monitoring data and deformation prediction trajectory are compared, the prediction residual is calculated, and the output layer weights of the lightweight prediction model are updated online based on the recursive least squares algorithm, so that the digital twin prediction module can adaptively track the changes in soil mechanical properties.

[0021] In this embodiment, it should be specifically noted that the data acquisition module obtains the tunnel arch settlement through a total station, the horizontal displacement of the retaining pile top through a microelectromechanical system sensor, and the pore water pressure through a pore water pressure gauge. It also obtains the excavation rate, dewatering flow rate, and support axial force increment through the construction control system. The total station, microelectromechanical system sensor, pore water pressure gauge, and construction control system are all sensors. The data acquisition module records the unprocessed values ​​directly obtained from the sensors as the raw measurement values. The total station sampling frequency is set to 0.1 Hz, corresponding to a data acquisition of tunnel arch settlement every 10 seconds. This frequency is sufficient to capture the slow settlement changes caused by foundation pit excavation while avoiding data redundancy. The microelectromechanical system (MEMS) sensor sampling frequency is set to 100 Hz to capture high-frequency vibrations and instantaneous displacements when subway trains pass by. This frequency can resolve 100 vibration waveforms per second, meeting the requirements of structural dynamics analysis. The pore water pressure gauge sampling frequency is set to 1 Hz, i.e., once per second, matching the time-varying characteristics of the soil consolidation process. The construction control system's data update frequency is... The pulse frequency is set every 30 seconds, consistent with the system synchronization pulse period, facilitating subsequent time alignment. The capacity of the circular buffer is set to 1000 original measurements and their corresponding hardware timestamps. Calculated at the lowest sampling frequency of the total station of 0.1 Hz, 1000 storage units can hold approximately 10,000 seconds (approximately 2.78 hours) of data, ensuring that no data is lost due to buffer overflow within the system synchronization pulse period. The fixed period of the system synchronization pulse is set to 30 seconds, which is consistent with the data update frequency of the construction control system, ensuring that all sensor data can be aligned on the same time reference. The data acquisition module maintains a circular buffer for each sensor. The circular buffer stores the most recent raw measurement values ​​and their hardware timestamps in a first-in-first-out manner. The circular buffer is implemented using a circular queue data structure. When the write pointer reaches the capacity limit, it automatically wraps back to the beginning position and overwrites the oldest data, thus always retaining the most recent 1000 raw measurement values ​​and their hardware timestamps, which is suitable for continuous caching of real-time data streams. The data acquisition module emits synchronization pulses at fixed intervals, mapping the hardware timestamp of each raw measurement value to the nearest system clock beat. The index number of the system clock beat is calculated using a rounding function. Specifically, the difference between the current raw timestamp and the system startup time is subtracted, and then divided by the fixed period of the system synchronization pulse. The integer part of the result is the index number of the corresponding system clock beat. The system startup time t0 is a unified time reference recorded when the data acquisition module is first powered on. All sensor hardware timestamps are converted to seconds relative to t0. The rounding function uses a floor function to ensure that each raw measurement value uniquely belongs to one system clock beat, avoiding ambiguity in cross-beat attribution. For sensors that have multiple samples within the same system clock cycle, the arithmetic mean is used as the representative value. System clock cycles that do not generate samples are marked as missing, thus obtaining the discrete sequence of each sensor on the system time axis. For high-frequency sensors such as those in microelectromechanical systems (MEMS), a 30-second system clock cycle may contain 3000 raw measurement values. The arithmetic mean can effectively suppress random noise while retaining the average amplitude information within that cycle. The missing value is marked with a null placeholder. Subsequent sliding window processing will skip the cycle containing the null value, but it will not affect the calculation of the window length. The data acquisition module performs sliding window processing on the discrete sequences of each sensor. The length of the sliding window is adaptively adjusted according to the degree of local fluctuation in the discrete sequence. The local fluctuation metric is defined as the median of the absolute values ​​of the differences between two adjacent original measurements within the sliding window. When the local fluctuation metric is higher than a high threshold, the sliding window length shrinks to its minimum value; when the local fluctuation metric is lower than a low threshold, the sliding window length expands to its maximum value; when the local fluctuation metric is between the high and low thresholds, the sliding window length remains unchanged. The minimum sliding window length L_min is set to 10 system clock cycles, corresponding to 5 minutes; the maximum sliding window length L_max is set to 50 system clock cycles, corresponding to 2... 5 minutes; the high threshold θ_high and low threshold θ_low are set according to the sensor type: for total station, θ_high is 0.5 mm and θ_low is 0.1 mm; for microelectromechanical system (MEMS) sensors, θ_high is 0.02 mm and θ_low is 0.005 mm; for pore water pressure gauge, θ_high is 5 kPa and θ_low is 1 kPa; this adaptive mechanism expands the window when the data is stable (such as at night when there are no trains running) to improve the statistical robustness of anomaly detection, and shrinks it when the data changes drastically (such as at the moment of unloading during foundation pit excavation) to retain the true information of the sudden change and avoid misjudging normal deformation caused by construction as anomaly; The data acquisition module calculates the median and absolute deviation of the median within a defined sliding window. The absolute deviation of the median is equal to the median of the absolute differences between each original measurement and the median within the window. The module constructs a robust statistical score. The numerator of the robust statistical score is the difference between the current original measurement and the median within the window multiplied by 0.6745, and the denominator is the sum of the absolute deviation of the median and 10 to the power of -6. If the absolute value of the robust statistical score is greater than 3.5, the current original measurement is identified as an outlier and replaced with the median within the window; otherwise, the original value is retained. The replaced sequence is denoted as the cleaned sequence. The constant 0.6745 is chosen based on the assumption of normal distribution. Suppose that when the data follows a normal distribution, the absolute deviation of the median is approximately 0.6745 times the standard deviation. Multiplying by this constant makes the robust statistical score consistent with the classic Z-score scale, facilitating the use of the empirical criterion of 3.5 times the standard deviation. The small amount ε added to the denominator is to prevent division by zero errors when the absolute deviation of the median in the window is zero. The threshold of 3.5 corresponds to a 99.95% confidence level, meaning that under a normal distribution, only about 0.05% of normal data will be misclassified as abnormal. This threshold has been verified in engineering practice to effectively balance the false negative rate and false alarm rate in the foundation pit monitoring scenario. Outliers are replaced with the median within the window instead of the mean because the median is not sensitive to extreme values, and the replacement will not introduce new biases. The data acquisition module aligns the cleaning sequences of all sensors according to the system clock cycle and categorizes them according to predefined physical quantity types. At each system clock cycle, it acquires the tunnel arch settlement from the total station's cleaning sequence, the horizontal displacement of the retaining pile tops from the microelectromechanical system (MEMS) sensor's cleaning sequence, and the pore water pressure from the pore water pressure gauge's cleaning sequence. This data is then assembled into multi-dimensional monitoring data. Simultaneously, the data acquisition module directly acquires the opening... The excavation rate, dewatering flow rate, and support axial force increment are assembled into construction parameters. Multidimensional monitoring data are stored in the form of a three-dimensional column vector, with the dimension order fixed as tunnel arch settlement, retaining pile top horizontal displacement, and pore water pressure, which facilitates subsequent digital twin prediction module reading by index. Construction parameters are also stored in a three-dimensional column vector, with the dimension order as excavation rate, dewatering flow rate, and support axial force increment. For the construction control system, since its data update frequency is exactly equal to the system synchronization pulse period of 30 seconds, each system clock cycle has one and only one construction parameter value, and no interpolation or averaging processing is required. The data acquisition module assembles the multidimensional monitoring data from several recent system clock cycles into a historical monitoring state sequence in chronological order, and assembles the construction parameters from several recent system clock cycles into a historical construction parameter sequence in chronological order. The historical monitoring state sequence and the historical construction parameter sequence have the same length and match the input window length required by the digital twin prediction module. The historical monitoring state sequence and the historical construction parameter sequence are arranged in chronological order for direct use by the digital twin prediction module. The number M of several recent system clock cycles is 720, corresponding to 6 hours of data, which is consistent with the input time window length of the long short-term memory recurrent neural network in the digital twin prediction module. This length has been experimentally verified to cover the main time-effect cycle of foundation pit excavation and unloading, while also taking into account computational efficiency. The historical monitoring state sequence and the historical construction parameter sequence are both M×3 matrices. The row indices of the two matrices correspond to the same system clock cycle, while the column indices correspond to different physical quantities. After receiving these two matrices, the digital twin prediction module concatenates them into an M×6 input tensor, which serves as the feature input of the long short-term memory recurrent neural network.

[0022] In this embodiment, the specific process of constructing a lightweight prediction model in the digital twin prediction module is as follows: The historical monitoring status sequence and the historical construction parameter sequence are spliced ​​along the feature dimension to form a joint feature tensor. The row index of the joint feature tensor corresponds to the system clock beat, and the column indexes are, in order, the tunnel arch settlement, the horizontal displacement of the retaining pile top, the pore water pressure, the excavation rate, the dewatering flow rate, and the support axial force increment. The joint feature tensor is input into a bidirectional long short-term memory recurrent neural network (LSTM). The bidirectional LSTM recurrent neural network is used as the encoder. The encoder consists of a forward LSTM recurrent neural network and a backward LSTM recurrent neural network connected in parallel. The number of hidden layer neurons in both the forward and backward LSTM recurrent neural networks is 32. For each time step in the joint feature tensor, the forward LSTM recurrent neural network generates a forward hidden state, and the backward LSTM recurrent neural network generates a backward hidden state. The forward and backward hidden states are concatenated, and the concatenated result is the encoder hidden state. The dimension of the encoder hidden state is 64. The digital twin prediction module also includes a decoder, which is a single-layer long short-term memory recurrent neural network. The initial hidden state and initial cell state of the decoder are taken from the forward hidden state and forward cell state of the encoder at the last time step. The digital twin prediction module constructs a construction perception attention mechanism. The calculation process of the attention score function is as follows: First, the decoder hidden state at the previous time step, the encoder hidden state at the i-th time step, and the construction parameters at the i-th time step are concatenated to form a combined feature vector. Second, the combined feature vector is multiplied by a learnable attention vector, a bias term is added, and then the hyperbolic tangent activation function is used to calculate the attention score. The raw score is calculated; finally, a normalized exponential function is applied to the raw score to convert it into attention weights. All encoder hidden states are weighted and summed according to the attention weights to obtain the context vector. The context vector condenses the spatiotemporal features most relevant to the current decoding moment in the historical sequence and carries the modulation information of construction parameters. In the construction-aware attention mechanism, the learnable attention vector dimension is set to sixteen. The construction-aware attention mechanism enables the decoder to automatically focus on key time steps in the historical sequence where construction parameters change drastically or monitoring data shows inflection points when generating predictions for the current time step, thereby improving prediction accuracy. The attention weights can be visualized to help engineers understand the basis of model decisions. For each future prediction time step, the digital twin prediction module first calculates the context vector based on the decoder's hidden state in the previous time step using a construction-aware attention mechanism. Then, it concatenates the temporal prediction value of the tunnel convergence deformation rate in the previous time step with the context vector, using it as the input to the decoder in the current time step. The decoder outputs the decoder's hidden state in the current time step. Specifically, for the first future prediction time step, the decoder's hidden state in the previous time step is the decoder's initial hidden state, and the temporal prediction value of the tunnel convergence deformation rate in the previous time step is zero. For the second and subsequent future prediction time steps, the decoder's hidden state in the previous time step has already been generated and stored in the previous decoding process. The digital twin prediction module incorporates a physical information regularization term during training. This term is defined as the squared mean of the deviations between the second-order difference of the time-series predicted tunnel convergence deformation rate and the rate of change of construction parameters. The calculation process is as follows: First, extract the time-series predicted values ​​of the tunnel convergence deformation rate at the current time step, the previous time step, and the two time steps prior from the sequence of predicted tunnel convergence deformation rates. Then, calculate the time-series predicted value of the tunnel convergence deformation rate at the current time step. The second difference of the tunnel convergence deformation rate is obtained by subtracting twice the tunnel convergence deformation rate predicted at the previous time step from the predicted tunnel convergence deformation rate, and then adding the difference between the predicted tunnel convergence deformation rates predicted two time steps prior. Next, the rate of change of construction parameters at the current time step is extracted, which is the difference between the current excavation rate and the previous excavation rate. Then, the second difference of the tunnel convergence deformation rate is subtracted from the rate of change of construction parameters, and the square of the difference is calculated. Finally, the time series prediction of the squared value within a predetermined future time period is calculated. The sum of the total number of time steps is calculated, and the quotient of this sum to the total number of time steps within a predetermined future time period is obtained to yield the mean squared deviation. The physical information regularization term and the negative log-likelihood loss are weighted and summed to form the total loss function, with the weighting coefficient set to 0.1. After training, the model parameters are frozen to obtain a lightweight prediction model. The design of the physical information regularization term is based on the physical law in soil consolidation theory that "deformation acceleration is proportional to the construction disturbance rate." The second-order difference of the time-series predicted value of the tunnel convergence deformation rate physically represents the deformation rate. Acceleration and the rate of change of construction parameters represent the rate of change of excavation disturbance; the smaller the mean of the square of the difference between the two, the more the deformation trend predicted by the model conforms to the common sense of mechanics; the tradeoff coefficient of 0.1 was determined through cross-validation, which effectively suppressed oscillating predictions that violate physical laws while ensuring prediction accuracy; the Adam optimizer was used for training, with the initial learning rate set to 0.001, the batch size set to 32, and the early stopping patience value set to 10 cycles; after training, the model parameters were frozen, and the digital twin prediction module only performed forward propagation during each inference, without the need for retraining; The digital twin prediction module uses a hybrid density network as the output layer of the decoder. The hybrid density network generates the mixing coefficients, mean, and log-variance of the Gaussian mixture model by passing the decoder's hidden state through three independent fully connected branches. There are three mixing coefficients. The generation process of the mixing coefficients is to input the original output value of the fully connected branch into a normalized exponential function for processing to ensure that the mixing coefficients are non-negative and their sum is one. The generation process of the mean is to directly output the linear activation result of the fully connected branch. The generation process of the log-variance is to output the result of the fully connected branch through linear activation to ensure that the variance is always positive. The digital twin prediction module calculates the probability density function of the time-series predicted value of the tunnel convergence deformation rate at the current time step based on the mixing coefficients, mean, and log-variance. The calculation process is as follows: the contribution of each mixing component is weighted and summed, where the contribution of each mixing component equals the mixing coefficient of that component multiplied by the value of a Gaussian distribution function. The Gaussian distribution function is calculated at the time-series predicted value of the tunnel convergence deformation rate at the current time step, using the mean and variance of the mixing component as parameters. The weighted expected value of the means of each component is then used as the time-series predicted value of the tunnel convergence deformation rate at the current time step, while simultaneously saving the log-variance sequence for each time step. The mixing density network outputs... The log-variance sequence is used to quantify prediction uncertainty. The larger the log-variance, the less certain the model is about the prediction at that moment. In the subsequent optimization decision module, the log-variance can be used as a risk weight in the calculation of the objective function, enabling the system to adopt a more conservative control strategy in the high uncertainty region. The setting of three mixture components can fit a multimodal distribution. For example, the deformation rate shows a peak distribution during train passage and a flat distribution during no-train periods. A single Gaussian distribution cannot simultaneously characterize these two modes. The weighted expectation of the mean of each component is calculated by multiplying the mean of each mixture component by the corresponding mixing coefficient and then adding the three products. The result is the point prediction value at the current time step. The digital twin prediction module outputs the deformation prediction trajectory and corresponding log-variance sequence for the next 24 hours. The deformation prediction trajectory includes the time-series prediction value of the tunnel convergence deformation rate. The next 24 hours correspond to 48 time steps, each time step being 30 minutes long. The deformation prediction trajectory is stored in array form with a length of 48, where each element is a floating-point number representing the predicted value of the tunnel convergence deformation rate at that time step. The log-variance sequence is also an array of length 48, where each element is the log-variance of the predicted value at that time step. Both sets of data are fed into the optimization decision module to construct the deformation over-limit penalty term and the safety barrier function term in the finite-time domain optimization objective function.

[0023] In this embodiment, it is specifically necessary to explain that the process of constructing the finite-time domain optimization objective function in each control cycle in the optimization decision module is as follows: At the beginning of each control cycle, the control time domain length is set to eight control steps, corresponding to the next sixteen hours. The decision variables are the construction parameter sequence for the next eight control steps. This construction parameter sequence includes the excavation rate, precipitation flow, and support axial force increment from the current control step to the next seventh control step. The constructed finite-time domain optimization objective function is a weighted sum of a deformation exceeding penalty term, a construction cost term, and a safety barrier function term. The calculation process involves iterating through each future control step from the first to the eighth control step, multiplying the current control step's deformation exceeding penalty term by a preset first weight coefficient, adding the preset second weight coefficient multiplied by the current control step's construction cost term, and adding the preset third weight coefficient multiplied by the current control step's safety barrier function term. The eight summations are then accumulated. The preset first, second, and third weight coefficients are set to 0.5, 0.3, and 0.2, respectively. These weights are determined using the analytic hierarchy process (AHP) combined with expert scoring, prioritizing deformation control. The safety of the project is ensured while also considering the economic efficiency of construction. The preset safety threshold, early warning threshold, and critical threshold are determined based on the deformation control standards for subway tunnels and the design level of the foundation pit support: the safety threshold is 0.1 mm per day, the early warning threshold is 0.2 mm per day, and the critical threshold is 0.3 mm per day; the preset excavation rate reference value is 0.5 meters per day, the preset maximum allowable precipitation flow rate is 50 cubic meters per hour, and the preset maximum allowable increase in support axial force is 500 kN; the preset maximum variation is used to limit the adjustment range of construction parameters between adjacent control steps: the variation range of excavation rate shall not exceed 0.1 meters per day, the variation range of precipitation flow rate shall not exceed 10 cubic meters per hour, and the variation range of support axial force increase shall not exceed 100 kN. The deformation over-limit penalty term adopts a piecewise exponential function form. The specific calculation process is as follows: when the time-series predicted value of the tunnel convergence deformation rate output by the digital twin prediction module is below the preset safety threshold, the deformation over-limit penalty term is zero; when the time-series predicted value of the tunnel convergence deformation rate is between the preset safety threshold and the preset warning threshold, the deformation over-limit penalty term increases exponentially; when the time-series predicted value of the tunnel convergence deformation rate exceeds the preset warning threshold, the deformation over-limit penalty term increases sharply according to a quadratic function. The base of the exponential function is set to the natural constant e, and the scaling factor of the exponential part is set to five, so that the deformation over-limit penalty term smoothly transitions near the safety threshold and rises rapidly when approaching the warning threshold; the coefficient of the quadratic term of the quadratic function is set to ten, ensuring that the penalty value increases sharply after exceeding the warning threshold, forcing the optimizer to prioritize avoiding deformation over-limit. The construction cost item is defined as a weighted sum of three components: the first component is the square of the deviation of the current control step excavation rate from the preset excavation rate reference value; the second component is the square of the ratio of the current control step precipitation flow to the preset maximum allowable precipitation flow; and the third component is the square of the ratio of the current control step support axial force increment to the preset maximum allowable support axial force increment. The weight coefficients of the three components are set to 0.6, 0.2, and 0.2, respectively, with the excavation rate deviation having the highest weight because the excavation rate directly affects the construction progress and project cost. All squaring operations ensure that the cost item is always non-negative, and the greater the deviation, the higher the cost. The safety barrier function term is defined as follows: when the time-series predicted value of the tunnel convergence deformation rate at the current control step size approaches a preset critical threshold, the safety barrier function term generates a steep cubic penalty gradient. Specifically, the steep cubic penalty gradient is expressed as follows: the value of the safety barrier function term is equal to the cube of the quotient of the difference between the time-series predicted value of the tunnel convergence deformation rate at the current control step size and the critical threshold, divided by the difference between the critical threshold and the warning threshold, multiplied by a preset barrier strength coefficient, which is set to one hundred. This design ensures that as the predicted value approaches the critical threshold from the warning threshold, the penalty value increases cubically, forming an insurmountable safety barrier. The upper and lower limits of construction parameters and the rate of change constraints are added as inequality constraints to the optimization problem. The upper and lower limits of construction parameters limit the excavation rate to between the minimum and maximum values, the precipitation flow rate to between zero and the maximum value, and the support axial force increment to between zero and the maximum value. The rate of change constraint limits the variation of each construction parameter between adjacent control steps to not exceed the preset maximum variation. The minimum excavation rate is set to 0.1 meters per day, and the maximum value is set to 1.0 meters per day. The maximum precipitation flow rate is set to 50 cubic meters per hour. The maximum support axial force increment is set to 500 kN. The preset maximum variation of the rate of change constraint is as above, which is used to prevent drastic fluctuations in construction parameters from causing sudden changes in soil stress. The specific process of finding the optimal construction parameter sequence that minimizes the value of the finite-time optimization objective function is as follows: First, the historical monitoring status sequence, historical construction parameter sequence, and candidate construction parameter sequence at the current moment are passed to the digital twin prediction module. The digital twin prediction module performs forward propagation and returns the time-series predicted values ​​of the tunnel convergence deformation rate and its log-variance sequence for each time step covered by multiple future control steps. The candidate construction parameter sequence is initialized to the construction parameter value at the current moment in the first iteration and is automatically updated by the optimization algorithm in subsequent iterations. The forward propagation computation of the digital twin prediction module is small, with each call taking about 0.01 seconds, which meets the real-time control requirements. Then, the optimization decision module introduces the upper and lower limits of construction parameters and the rate of change constraints into the finite-time domain optimization objective function through the Lagrange multiplier method, and constructs an augmented Lagrange function. In the Lagrange multiplier method, each inequality constraint corresponds to a Lagrange multiplier and a penalty coefficient. The penalty coefficient is initially set to one thousand, and is automatically adjusted according to the degree of constraint violation after each iteration. The greater the degree of violation, the faster the penalty coefficient increases, forcing the optimization solution to gradually satisfy all constraints. Next, the algorithm in the quasi-Newton method is used for iterative solution. In each iteration, the gradient of the augmented Lagrangian function with respect to each component of the candidate construction parameter sequence is automatically calculated by differentiation, and the candidate solution is updated according to the gradient direction. The maximum number of iterations is set to 200, and the gradient convergence threshold is set to 10 to the power of negative 4. The quasi-Newton method specifically adopts the L-BFGS algorithm, which uses the gradient history information of the first 20 iterations to approximate the inverse of the Hessian matrix. It does not require explicit calculation of the second derivative and is suitable for handling medium-sized non-convex optimization problems. Automatic differentiation is implemented through the automatic differentiation mechanism of the PyTorch framework. Each component in the candidate construction parameter sequence is set as a trainable tensor, and the gradient is automatically calculated after forward propagation. The gradient convergence threshold is 10 to the power of negative 4, that is, convergence is determined when the L2 norm of the gradient is less than this value. If convergence is not achieved after reaching the maximum number of iterations, the current optimal solution is taken as the approximate optimal solution. The optimal construction parameter sequence is obtained after iterative convergence. The optimization decision module only sends the excavation rate, dewatering flow rate, and support axial force increment from the first control step of the optimal construction parameter sequence to the excavator control system, the variable frequency dewatering pump control system, and the hydraulic support system, respectively. The construction parameters for the remaining control steps are discarded and await re-solution in the next control cycle. After receiving the excavation rate command, the excavator control system automatically adjusts the excavator's travel speed and bucket cutting depth. The variable frequency dewatering pump control system adjusts the pump motor frequency according to the dewatering flow rate command to achieve precise water control. The hydraulic support system adjusts the hydraulic cylinder pressure according to the support axial force increment command to change the support preload axial force. The response time of the three actuators is within ten seconds, far less than the control cycle of thirty minutes, ensuring that the commands can take effect in a timely manner. Discarding the construction parameters for the remaining control steps is a standard practice in rolling time-domain control to avoid deviations from the actual optimum in the long term due to model mismatch.

[0024] In this embodiment, it is specifically explained that after each control cycle, the online fine-tuning module obtains multi-dimensional monitoring data of all system clock beats within the current control cycle from the data acquisition module, obtains the deformation prediction trajectory at the corresponding time from the digital twin prediction module, extracts the time-series predicted value of the tunnel convergence deformation rate from the deformation prediction trajectory, subtracts the measured value of tunnel arch settlement from the multi-dimensional monitoring data from the corresponding time value of the time-series predicted value of tunnel convergence deformation rate, and calculates the prediction residual for each system clock beat; the online fine-tuning module calculates the absolute median of the prediction residual sequence, when the absolute median is... An update step is executed when the median is less than a preset residual threshold; the update is skipped when the absolute median is not less than the preset residual threshold. The preset residual threshold is set to 0.05 mm, which is determined based on the total station measurement accuracy (sub-millimeter level) and the allowable prediction deviation in engineering. As a robust statistic, the absolute median is not sensitive to individual outlier residuals and can effectively avoid invalid updates triggered by occasional sensor noise. When the absolute median of the residuals is less than the threshold, it indicates that the current model prediction accuracy is still within an acceptable range and no update is needed; otherwise, it indicates that the model has a systematic bias and parameter adjustments are required. The online fine-tuning module extracts the decoder hidden state corresponding to each system clock cycle from the lightweight prediction model of the digital twin prediction module. For each system clock cycle, the decoder hidden state is treated as a row of data, and all rows are stacked in chronological order to form a feature matrix. The number of rows in the feature matrix equals the total number of system clock cycles in the current control cycle, and the number of columns equals the dimension of the decoder hidden state. The measured values ​​of tunnel arch settlement in the multidimensional monitoring data are arranged in chronological order to form a target value vector. The dimension of the decoder hidden state is sixty-four, consistent with the dimension of the encoder hidden state. Each row of the feature matrix corresponds to a system clock cycle, and the column index corresponds to the sixty-four components of the decoder hidden state. The target value vector is the measured values ​​of tunnel arch settlement for all system clock cycles in the current control cycle, arranged in ascending chronological order. This feature matrix and the target value vector constitute the input-output pair of the recursive least squares algorithm. The online fine-tuning module employs a recursive least squares algorithm with a forgetting factor to sequentially update the output layer weights and biases of the lightweight prediction model. The forgetting factor is set to 0.98, and the covariance matrix is ​​initialized as the identity matrix multiplied by 1000. The selection of the forgetting factor of 0.98 is based on the time scale of soil mechanical property changes. The typical change period of soil elastic modulus and permeability coefficient during foundation pit excavation is tens of hours. The forgetting factor of 0.98 causes the weights of data from about fifty control periods to decay to about 36%, which can both track slow changes and suppress short-term noise interference. The covariance matrix is ​​initialized as the identity matrix multiplied by 1000, indicating that the uncertainty of the output layer weights is relatively large at the beginning, allowing for a larger initial update step size. As the iteration progresses, the covariance matrix gradually decreases, and the update step size adaptively shrinks. For each system clock cycle, the gain vector for the current cycle is calculated based on the covariance matrix obtained from the previous iteration and the row vector of the feature matrix for the current cycle. The covariance matrix for the current cycle is updated based on the gain vector, the row vector of the feature matrix for the current cycle, and the covariance matrix from the previous iteration. The row vector of the feature matrix for the current cycle is multiplied by the current output layer weights and then the output layer bias is added to obtain the time-series predicted value of the tunnel convergence deformation rate for the current cycle. The prediction residual between the value at the corresponding moment in the target value vector and the time-series predicted value of the tunnel convergence deformation rate for the current cycle is calculated. The output layer weights and output layer bias are updated using the prediction residual, the gain vector, and the row vector of the feature matrix for the current cycle. After traversing all system clock cycles within the current control cycle, the updated value is obtained. The output layer weights and updated output layer biases are calculated. The calculation of the gain vector is essentially the equivalent form of Kalman gain in recursive least squares, and its magnitude determines the contribution of the current prediction residual to the weight update. When the product of the feature matrix row vector and the covariance matrix is ​​large, the gain vector is large, and the update magnitude increases accordingly; otherwise, it decreases. The update of the covariance matrix reflects the recursive decay of the uncertainty of parameter estimation. The forgetting factor gradually amplifies the uncertainty of old data, thus leaving more room for adjustment for new data. After traversing all system clock ticks, the output layer weights and output layer biases are updated once. The whole process does not require storing historical data, but only needs to maintain the covariance matrix and the current weights. The computational complexity is O(d²), where d is the feature dimension of 64, which can be completed in milliseconds. The online fine-tuning module replaces the weight matrix of the output layer in the lightweight prediction model with the updated output layer weights and the output layer bias with the updated output layer bias, while keeping the parameters of the remaining layers in the lightweight prediction model unchanged. The shape of the output layer weight matrix is ​​1 x 64, and the output layer bias is a scalar. The reason for replacing only these two parameters is that the deep feature extraction capability of the lightweight prediction model is provided by the bidirectional long short-term memory recurrent neural network and the attention mechanism. These layers have fully learned the general features of soil deformation during offline training and do not require frequent adjustments. The output layer is responsible for mapping deep features to specific deformation values, and its parameters are most sensitive to the time-varying nature of soil mechanical properties. Therefore, updating only the output layer can achieve effective adaptive tracking, while avoiding the overfitting risk and computational overhead caused by large-scale parameter updates. The online fine-tuning module calculates the magnitude of the difference between the updated and unupdated output layer weights. The specific process is as follows: subtract the corresponding element from the unupdated output layer weight matrix from each element in the updated matrix; sum the squares of all the differences; and then take the square root of the sum to obtain the magnitude. When the magnitude exceeds a preset maximum single weight change, the updated output layer weights are pruned. The pruning process is as follows: first, calculate the difference matrix between the updated and unupdated output layer weights; then, divide each element of the difference matrix by the magnitude to obtain a normalized direction vector; next, multiply the preset maximum single weight change by the normalized direction vector; finally, add the unupdated output layer weights to the product, ensuring that the updated output layer weights are equal to the unupdated output layer weights. The magnitude of the difference is equal to the preset maximum change in weights in a single iteration, and the pruning direction is consistent with the original update direction. The preset residual threshold is set to 0.05 mm, and the preset maximum change in weights in a single iteration is set to 0.1. The preset maximum change in weights in a single iteration of 0.1 limits the maximum Euclidean norm change of the output layer weights in each update cycle, preventing drastic drift of model parameters due to abnormal data in a single control cycle (such as instantaneous sensor failure or sudden train vibration). The pruning operation ensures that the update direction remains unchanged, only scaling the update step size, so that the model can remain stable when encountering noise and will not experience a sudden drop in prediction performance. This pruning mechanism, together with the forgetting factor, constitutes a dual protection: the forgetting factor controls the decay rate of historical data, and pruning controls the magnitude of a single update. The two work together to ensure that the model is both adaptive and robust in long-term operation. After the online fine-tuning module completes parameter replacement, the digital twin prediction module uses the updated lightweight prediction model for prediction in the next control cycle, enabling the digital twin prediction module to adaptively track changes in soil mechanical properties. The updated model takes effect immediately without recompiling or restarting. At the start of the next control cycle, the digital twin prediction module uses the new output layer weights and biases to forward propagate the input joint feature tensor, outputting the deformation prediction trajectory. Since only the output layer parameters change, while other parts of the model (encoder, attention mechanism, hybrid density network) remain unchanged, the overall distribution characteristics of the prediction results are preserved, with only the mean output undergoing a smooth shift, achieving progressive tracking of changes in soil mechanical properties. This lightweight online update strategy enables the system to maintain high prediction accuracy throughout the months-long underground engineering construction cycle without frequent offline retraining.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A smart monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits, characterized in that, Specifically, it includes: The system consists of a data acquisition module, a digital twin prediction module, an optimization decision-making module, and an online fine-tuning module, which are connected in sequence via communication. Data acquisition module: acquires multi-dimensional monitoring data and construction parameters in real time during the construction of the adjacent subway foundation pit. The data acquisition module stores the acquired multi-dimensional monitoring data and construction parameters in chronological order as historical monitoring status sequence and historical construction parameter sequence, respectively. The digital twin prediction module receives historical construction parameter sequences and historical monitoring status sequences, constructs a lightweight prediction model, and outputs the deformation prediction trajectory for the next 24 hours, using these sequences as input. The deformation prediction trajectory includes the time-series prediction value of the tunnel convergence deformation rate. The specific process of constructing the lightweight prediction model is as follows: The historical monitoring state sequence and the historical construction parameter sequence are concatenated along the feature dimension to form a joint feature tensor. The joint feature tensor is then input into a bidirectional long short-term memory recurrent neural network (LSTM). The bidirectional LSTM recurrent neural network is used as an encoder, which consists of a forward LSTM recurrent neural network and a backward LSTM recurrent neural network connected in parallel. For each time step in the joint feature tensor, the forward LSTM recurrent neural network generates a forward hidden state, and the backward LSTM recurrent neural network generates a backward hidden state. The forward hidden state and the backward hidden state are concatenated, and the concatenated result is the encoder hidden state. The digital twin prediction module also includes a decoder, which is a single-layer long short-term memory recurrent neural network. The initial hidden state and initial cell state of the decoder are taken from the forward hidden state and forward cell state of the encoder at the last time step. The digital twin prediction module constructs a construction perception attention mechanism, which calculates the attention weight by concatenating the decoder hidden state, encoder hidden state and construction parameters and then summing them by an attention scoring function to obtain the context vector. For each future prediction time step, the temporal prediction value of the tunnel convergence deformation rate of the previous time step is concatenated with the context vector and used as the decoder input; The digital twin prediction module incorporates a physical information regularization term during training. The physical information regularization term is defined as the squared mean of the deviation between the second-order difference of the time-series predicted value of the tunnel convergence deformation rate and the rate of change of construction parameters. The physical information regularization term and the negative log-likelihood loss are weighted and summed to form the total loss function. After training, the model parameters are frozen to obtain a lightweight prediction model. Optimization Decision Module: Constructs a finite-time domain optimization objective function within each control cycle. The finite-time domain optimization objective function includes a deformation exceeding penalty term, a construction cost term, and a safety barrier function term. Using the construction parameter sequence of multiple future control steps as decision variables, and under the conditions of satisfying the upper and lower limits of construction parameters and the rate of change constraints, it solves for the optimal construction parameter sequence that minimizes the value of the finite-time domain optimization objective function. Only the construction parameters of the first control step in the optimal construction parameter sequence are executed to adjust the current construction state. Online fine-tuning module: After each control cycle, multi-dimensional monitoring data is obtained from the data acquisition module, deformation prediction trajectory is obtained from the digital twin prediction module, the time-series prediction values ​​at corresponding times in the multi-dimensional monitoring data and deformation prediction trajectory are compared, the prediction residual is calculated, and the output layer weights of the lightweight prediction model are updated online based on the recursive least squares algorithm, so that the digital twin prediction module can adaptively track the changes in soil mechanical properties.

2. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 1, characterized in that: The data acquisition module acquires tunnel arch settlement through a total station, horizontal displacement of retaining pile top through a microelectromechanical system sensor, and pore water pressure through a pore water pressure gauge. It also acquires excavation rate, dewatering flow rate, and support axial force increment through a construction control system. The total station, microelectromechanical system sensor, pore water pressure gauge, and construction control system are all sensors. The data acquisition module records the unprocessed values ​​directly acquired from the sensors as the raw measurement values. The data acquisition module maintains a circular buffer for each sensor, which stores the most recent raw measurement values ​​and their hardware timestamps in a first-in-first-out manner. The data acquisition module emits synchronization pulses at fixed intervals, mapping the hardware timestamp of each raw measurement value to the most recent system clock tick; For sensors that have multiple samples within the same system clock cycle, the arithmetic mean is taken as the representative value. System clock cycles that do not generate samples are marked as missing, thus obtaining the discrete sequence of each sensor on the system time axis. The data acquisition module performs sliding window processing on the discrete sequences of each sensor. The length of the sliding window is adaptively adjusted according to the degree of local fluctuation in the discrete sequence. The local fluctuation metric is defined as the median of the absolute value of the difference between two adjacent original measurements within the sliding window.

3. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 2, characterized in that: The data acquisition module calculates the median and absolute deviation of the median within a defined sliding window. The absolute deviation of the median is equal to the median of the absolute differences between each original measurement value and the median within the window. The data acquisition module constructs a robust statistical score. The numerator of the robust statistical score is the difference between the original measurement value at the current time and the median within the window multiplied by 0.6745, and the denominator is the sum of the absolute deviation of the median within the window and 10 to the power of negative 6. If the absolute value of the robust statistical score is greater than 3.5, the original measurement value at the current time is identified as an outlier and replaced with the median within the window; otherwise, the original value is retained. The replaced sequence is recorded as the cleaned sequence.

4. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 3, characterized in that: The data acquisition module aligns the cleaning sequences of all sensors according to the system clock cycle and categorizes them according to predefined physical quantity categories. At each system clock cycle, it acquires the tunnel arch settlement from the total station's cleaning sequence, the horizontal displacement of the retaining pile top from the microelectromechanical system sensor's cleaning sequence, and the pore water pressure from the pore water pressure gauge's cleaning sequence. It then assembles the tunnel arch settlement, the horizontal displacement of the retaining pile top, and the pore water pressure into multidimensional monitoring data. Simultaneously, the data acquisition module directly acquires the excavation rate, dewatering flow rate, and support axial force increment from the construction control system at the corresponding system clock cycle, and assembles these parameters into construction parameters. The data acquisition module assembles the multi-dimensional monitoring data from the most recent system clock cycles into a historical monitoring status sequence in chronological order, and assembles the construction parameters from the most recent system clock cycles into a historical construction parameter sequence in chronological order.

5. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 4, characterized in that: The digital twin prediction module uses a hybrid density network as the output layer of the decoder. The hybrid density network generates the mixing coefficients, mean, and log-variance of the Gaussian mixture model by passing the hidden state of the decoder through three independent fully connected branches. The digital twin prediction module calculates the probability density function of the time-series predicted value of the tunnel convergence deformation rate at the current time step based on the mixing coefficient, mean, and log-variance. It uses the weighted expected value of the mean of each component as the time-series predicted value of the tunnel convergence deformation rate at the current time step, and saves the log-variance sequence for each time step. The digital twin prediction module outputs the deformation prediction trajectory and the corresponding log-variance sequence for the next 24 hours. The deformation prediction trajectory includes the time-series predicted value of the tunnel convergence deformation rate.

6. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 5, characterized in that: In the optimization decision module, the process of constructing the finite-time domain optimization objective function in each control cycle is as follows: The control time domain length is set to eight control steps, and the decision variable is the sequence of construction parameters for the next eight control steps. The constructed finite-time domain optimization objective function is a weighted sum of a deformation exceeding penalty term, a construction cost term, and a safety barrier function term, which is accumulated over each future control step. The deformation exceeding penalty term adopts a piecewise exponential function form, taking zero when the time-series predicted value of the tunnel convergence deformation rate is below the safety threshold, growing exponentially when it is between the safety threshold and the warning threshold, and increasing sharply according to a quadratic function when it exceeds the warning threshold. The construction cost term is defined as a weighted sum of the square of the deviation of the excavation rate from the reference value, the square of the ratio of precipitation flow to the maximum allowable precipitation flow, and the square of the ratio of the increase in support axial force to the maximum allowable increase in support axial force. The safety barrier function term generates a steep cubic penalty gradient when the time-series predicted value of the tunnel convergence deformation rate approaches the critical threshold. The upper and lower limits of construction parameters and the rate of change constraints are added as inequality constraints to the optimization problem.

7. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 6, characterized in that: The process of finding the optimal construction parameter sequence that minimizes the value of the finite-time optimization objective function is as follows: First, the historical monitoring status sequence, historical construction parameter sequence, and candidate construction parameter sequence at the current moment are passed to the digital twin prediction module. The digital twin prediction module performs forward propagation and returns the time-series predicted values ​​of the tunnel convergence deformation rate and its logarithmic variance sequence for each time step covered by multiple future control steps. Then, the optimization decision module introduces the upper and lower limit constraints and the rate of change constraints of the construction parameters into the finite-time domain optimization objective function through the Lagrange multiplier method, and constructs the augmented Lagrange function. Next, the algorithm in the quasi-Newton method is used for iterative solution. In each iteration, the gradient of the augmented Lagrangian function with respect to each component in the candidate construction parameter sequence is automatically calculated by differentiation, and the candidate solution is updated according to the gradient direction. The maximum number of iterations is set to two hundred, and the gradient convergence threshold is set to ten to the power of negative four. The optimal construction parameter sequence is obtained after iterative convergence. The optimization decision module only sends the excavation rate, dewatering flow rate, and support axial force increment of the construction parameters in the first control step of the optimal construction parameter sequence to the excavator control system, the variable frequency dewatering pump control system, and the hydraulic support system, respectively. The construction parameters of the remaining control steps are discarded and wait for the next control cycle to be solved again.

8. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 7, characterized in that: After each control cycle ends, the online fine-tuning module obtains multi-dimensional monitoring data of all system clock beats within the current control cycle from the data acquisition module, obtains the deformation prediction trajectory at the corresponding time from the digital twin prediction module, extracts the time-series prediction value of the tunnel convergence deformation rate from the deformation prediction trajectory, and subtracts the measured value of tunnel crown settlement from the time-series prediction value of tunnel convergence deformation rate at the corresponding time one by one to calculate the prediction residual of each system clock beat. The online fine-tuning module calculates the absolute median of the predicted residual sequence. When the absolute median is less than a preset residual threshold, an update step is executed; when the absolute median is not less than the preset residual threshold, the update is skipped. The online fine-tuning module extracts the decoder hidden state corresponding to each system clock cycle from the lightweight prediction model of the digital twin prediction module, stacks them in chronological order to form a feature matrix, and organizes the measured values ​​of tunnel arch settlement in multidimensional monitoring data in chronological order to form a target value vector. The online fine-tuning module uses a recursive least squares algorithm with a forgetting factor to sequentially update the output layer weights and output layer biases of the lightweight prediction model. For each system clock cycle, the gain vector is calculated, the covariance matrix is ​​updated, and the output layer weights and output layer biases are updated using the predicted residuals. After traversing all system clock cycles in the current control cycle, the updated output layer weights and updated output layer biases are obtained.

9. The intelligent monitoring and early warning system for bidirectional disturbance of adjacent subway foundation pits according to claim 8, characterized in that: The online fine-tuning module replaces the weight matrix of the output layer in the lightweight prediction model with the updated output layer weights and the bias of the output layer with the updated output layer bias, while keeping the parameters of the remaining layers in the lightweight prediction model unchanged. The online fine-tuning module calculates the magnitude of the difference between the updated output layer weights and the original output layer weights. When the magnitude exceeds the preset maximum change in weights in a single operation, the updated output layer weights are pruned so that the magnitude of the difference between the updated and original output layer weights is equal to the preset maximum change in weights in a single operation. The pruning direction is consistent with the original update direction. After the online fine-tuning module completes the parameter replacement, the digital twin prediction module uses the updated lightweight prediction model to make predictions in the next control cycle, enabling the digital twin prediction module to adaptively track changes in soil mechanical properties.

Images