TBM soft surrounding rock stratum change real-time sensing method and system

Abstract

The application provides a TBM soft surrounding rock stratum change real-time sensing method and system, relates to the tunnel engineering technical field, and the application realizes the probability distribution prediction of the future evolution trend of the TBM tunneling parameter by introducing the Bayesian time series prediction and the MC Dropout technology, simultaneously, the deviation degree between the measured data and the predicted probability distribution is utilized, the "probability residual" is constructed as a quantitative criterion, so that a soft surrounding rock discrimination mechanism with self-adaptive ability without relying on a fixed threshold is established; the application can effectively filter out sporadic noise interference, accurately capture the persistence anomaly caused by the soft surrounding rock stratum change, provide a real-time, reliable and statistically interpretable geological risk probability early warning for the TBM construction, and significantly improve the construction safety and decision efficiency under complex geological conditions.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering technology, and in particular to a method and system for real-time sensing of changes in weak surrounding rock strata in TBMs. Background Technology

[0002] TBM is an abbreviation for Tunnel Boring Machine. During TBM tunnel construction, the equipment's operating status and the surrounding rock conditions jointly determine changes in tunneling parameters such as thrust, cutterhead torque, penetration depth, advance speed, and power current. With the widespread use of sensors and data acquisition systems, TBMs can continuously acquire high-frequency, multi-dimensional tunneling and operational parameter data, providing a data foundation for construction status analysis, tunneling performance prediction, and risk control. How to utilize this real-time data to improve the ability to identify geological changes has become an important research and engineering requirement for intelligent TBM tunneling and safe construction.

[0003] Under complex geological conditions, the surrounding rock exhibits heterogeneity and abrupt changes. Especially in sections with weak surrounding rock, fractured zones, and water-rich interlayers, changes in surrounding rock conditions often lead to abnormal fluctuations or coupled changes in tunneling parameters. If these changes are not identified and adjusted in a timely manner, over-excavation, increased deformation, machine jamming, and other sudden risks can easily occur. Current methods for identifying changes in surrounding rock largely rely on advanced geological forecasts, design documents, and on-site experience, which suffer from deviations from actual excavation, response lags, and difficulties in continuous real-time quantitative assessment. Furthermore, relying solely on single parameter thresholds or post-hoc statistical analysis is insufficient to stably and reliably characterize changes in weak surrounding rock formations. Therefore, a technological means is needed that can combine real-time TBM tunneling and operational data to continuously perceive and quantify changes in weak surrounding rock formations.

[0004] Existing TBM surrounding rock condition sensing technologies largely rely on static data from advanced geological forecasts or on tunneling parameter predictions based on traditional deterministic deep learning models (such as standard LSTM and RNN). These methods share several drawbacks: First, they lack the ability to quantify prediction uncertainty. Existing models typically output only a single "point prediction" value, failing to reflect the model's confidence level under different operating conditions and making it difficult to distinguish between random noise in the data itself and anomalies caused by geological shifts. Second, the criteria for judgment are rigid and lack universality. Anomaly identification is mostly based on the residual between predicted and measured values, relying on manually set fixed empirical thresholds for alarms. These static thresholds cannot adapt to the complex and ever-changing weak surrounding rock environment, leading to false alarms during periods of large fluctuations in normal operation, or missed alarms and delayed responses during periods of gradual geological change. Third, they lack statistically significant risk quantification indicators. Existing results are difficult to directly translate into intuitive probabilities of weak surrounding rock risk, making it difficult to make quick and accurate decisions on-site.

[0005] In summary, there is an urgent need for a real-time sensing method and system for changes in weak surrounding rock formations using TBMs to solve the aforementioned problems. Summary of the Invention

[0006] To address the aforementioned technical problems in related technologies, this invention proposes a real-time sensing method and system for changes in weak surrounding rock formations using a TBM (Toyota Burma Machine).

[0007] In a first aspect, the present invention provides a method for real-time sensing of changes in weak surrounding rock strata in a TBM, comprising the following steps:

[0008] S1. Collect multi-source high-frequency time series data of TBM under various construction conditions, and construct a multi-source time series database by combining the correspondence between surrounding rock grade and construction mileage section.

[0009] S2. Divide the multi-source time-series database into cyclic tunneling segments according to the continuity of construction mileage to obtain a cyclic tunneling segment dataset. After outlier detection and interpolation processing, the cyclic tunneling segment data is standardized to obtain a standardized cyclic tunneling segment dataset.

[0010] S3. Using the Bayesian gated cyclic unit as the basic prediction model, a Bayesian time series prediction model is constructed using MC Dropout technology, and its prediction target is set as the predicted value of the cutterhead torque. Then, the Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset to obtain the trained Bayesian time series prediction model.

[0011] S4. Construct a weak surrounding rock strata discrimination model based on the prediction residuals according to the Bayesian time series prediction model;

[0012] S5. The weak surrounding rock strata discrimination model is used for subsequent TBM construction to provide early warning of weak surrounding rock.

[0013] Specifically, step S1 includes the following steps:

[0014] S11. Collect multi-source high-frequency time-series data of the same model of TBM under various construction conditions through a TBM monitoring system; the multi-source high-frequency time-series data includes timestamps, tunneling parameters, and operating parameters; wherein, one timestamp corresponds to a set of tunneling parameters and operating parameters; the tunneling parameters include thrust, cutterhead torque, penetration depth, propulsion speed, and power; the operating parameters include speed setting values ​​and thrust setting values; the various construction conditions include conventional geological conditions and different surrounding rock conditions;

[0015] S12. Combine advanced geological forecast results, design documents and field survey data to determine the surrounding rock grade information of the corresponding construction mileage section, and establish the correspondence between the surrounding rock grade and the construction mileage section.

[0016] S13. The surrounding rock grade information is associated with the multi-source high-frequency time series data using a mileage alignment method to form a multi-source time series database.

[0017] Specifically, step S2 includes the following steps;

[0018] S21. Select continuous data segments that meet the construction mileage continuity condition from the multi-source time series database to obtain the cyclic tunneling segment dataset;

[0019] S22. Perform outlier detection and interpolation on each cycle tunneling segment data in the cycle tunneling segment dataset to obtain corrected cycle tunneling segment data; this step is the data preprocessing process.

[0020] S23. Standardize the tunneling parameters and operating parameters to normalize parameters with different dimensions to a unified scale, and obtain a standardized cyclic tunneling section dataset.

[0021] Specifically, step S3 includes the following steps:

[0022] S31. A Bayesian gated recurrent unit is used as the basic prediction model, and the weight parameters and bias parameters corresponding to the input gate, forget gate and output gate in the basic prediction model are set to follow a Gaussian probability distribution.

[0023] S32. Based on the basic prediction model, the MC Dropout technique is used to approximate efficient Bayesian inference and construct a Bayesian time series prediction model.

[0024] S33. The Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset, and the prediction target of the Bayesian time series prediction model is set as the predicted value of the cutterhead torque, so as to obtain the trained Bayesian time series prediction model.

[0025] Specifically, step S4 includes the following steps:

[0026] S41. Obtain the first N seconds of data from the tunneling section with potentially weak surrounding rock strata during actual construction as input data.

[0027] S42. Input the input data into the Bayesian time series prediction model and perform T repeated prediction samplings to obtain T prediction samples for each prediction time from N seconds to M seconds in the future.

[0028] S43. Based on the T predicted samples at each prediction time, use the sample mean at time t. variance at time t Predicted probability distribution of parameters at approximately future time t As shown in the following formula:

[0029] ,

[0030] in, ; ; This is the set of predicted cutterhead torque values ​​at time t; Let be the predicted value of the cutterhead torque based on the i-th sampling at time t; Indicates used for prediction The historical input feature sequence of the cutterhead torque at each moment; D represents the sample dataset used to train the Bayesian time series prediction model;

[0031] S44. Based on the predicted probability distribution obtained in step S43 To obtain the predicted probability distribution sequence for each time point from N seconds to M seconds in the future;

[0032] S45. Based on the predicted probability distribution sequence, calculate the single-point probability residual by quantifying the deviation probability between the actual tunneling parameters and the predicted probability distribution, specifically including:

[0033] The deviation probability between the quantified actual tunneling parameters and the predicted probability distribution Calculated using the following formula:

[0034] ,

[0035] in, Indicates the actual tunneling parameters;

[0036] The single-point probability residual Calculated using the following formula:

[0037] ,

[0038] in, Φ(·) represents the outlier threshold; Φ(·) is the CDF formula for the standard normal distribution.

[0039] S46. Define a sliding time window and calculate the cumulative probability residual within the sliding time window at time t; the size of the sliding time window is m seconds, representing the time period m seconds before time t.

[0040] S47. Based on the cumulative probability residual, construct the weak surrounding rock discrimination rule to obtain the weak surrounding rock strata discrimination model.

[0041] Specifically, the cumulative probability residual Calculated based on the following formula;

[0042] ,

[0043] in, Let represent the single-point probability residual at time j.

[0044] Specifically, step S47, which involves constructing a weak surrounding rock discrimination rule based on the cumulative probability residual, includes: if the cumulative probability residual over K consecutive time periods... If the risk exceeds the preset risk threshold, it is marked as a high-probability weak surrounding rock section, and a weak surrounding rock section warning is triggered.

[0045] Specifically, the construction mileage continuity condition mentioned in step S21 is: a continuous data segment with a tunneling parameter thrust greater than 5kN and a time length greater than 1000 seconds.

[0046] Specifically, the Bayesian time series prediction model described in step S32 includes an input layer, a Bayesian LSTM layer, a Dropout layer, a fully connected layer, and an output layer; wherein:

[0047] The input layer receives the standardized cyclic tunneling segment dataset and organizes it into the time-series input format required by the model. The Bayesian LSTM layer is a time-series feature extraction unit formed by probabilistically modeling the weight and bias parameters in the LSTM network. The Dropout layer performs Monte Carlo Dropout sampling during the model prediction phase, performing multiple forward propagations through multiple randomly deactivated neurons to obtain the distribution information of the prediction results. The fully connected layer maps and fuses the hidden state features output by the Bayesian LSTM layer to achieve feature dimension transformation. The output layer outputs the predicted cutterhead torque value.

[0048] Secondly, the present invention provides a real-time sensing system for changes in weak surrounding rock strata during TBM operations, based on the real-time sensing method for changes in weak surrounding rock strata during TBM operations described in the first aspect above, comprising the following modules:

[0049] The multi-source time series database construction module is used to collect multi-source high-frequency time series data of TBM under conventional strata and different surrounding rock conditions, and to construct a multi-source time series database by combining the correspondence between surrounding rock grade and construction mileage section.

[0050] The cyclic tunneling section dataset construction module is used to divide the multi-source time-series database into cyclic tunneling sections according to the continuity of construction mileage to obtain the cyclic tunneling section dataset. After outlier detection and interpolation processing, the cyclic tunneling section data is standardized to obtain the standardized cyclic tunneling section dataset.

[0051] The Bayesian time series prediction model construction module is used to construct a Bayesian time series prediction model using the Bayesian gated cyclic unit as the basic prediction model and MC Dropout technology. The prediction target is set as the predicted value of the cutterhead torque. Then, the Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset to obtain the trained Bayesian time series prediction model.

[0052] The weak surrounding rock strata discrimination model construction module is used to construct a weak surrounding rock strata discrimination model based on the prediction residuals according to the Bayesian time series prediction model.

[0053] The early warning application module is used to apply the weak surrounding rock strata discrimination model to the subsequent construction of TBM to provide early warning of weak surrounding rock.

[0054] This invention provides a real-time sensing method and system for changes in weak surrounding rock formations during TBM construction. It aims to overcome the limitations of existing deterministic prediction models by introducing Bayesian time series prediction and MC Dropout technology to predict the probability distribution of future evolution trends of TBM tunneling parameters. Simultaneously, it utilizes the deviation between measured data and the predicted probability distribution to construct a "probability residual" as a quantitative criterion, thereby establishing a weak surrounding rock discrimination mechanism that does not rely on manually fixed thresholds and possesses adaptive capabilities. This invention can effectively filter out occasional noise interference and accurately capture persistent anomalies caused by changes in weak surrounding rock formations, providing real-time, reliable, and statistically interpretable geological risk probability early warnings for TBM construction, significantly improving construction safety and decision-making efficiency under complex geological conditions. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of a TBM-based real-time sensing method for changes in weak surrounding rock strata, provided in an embodiment of the present invention.

[0057] Figure 2 A schematic diagram of a TBM-based real-time sensing system for changes in weak surrounding rock strata provided in an embodiment of the present invention;

[0058] Figure 3 This is a comparison chart of cutterhead torque prediction results under hard rock and soft rock conditions provided in an embodiment of the present invention; wherein, Figure 3 Figure (a) shows the predicted cutterhead torque under hard rock conditions. Figure 3(b) in the figure shows the predicted cutterhead torque under soft rock conditions;

[0059] Figure 4 This is a comparison chart of probabilistic residuals under hard rock and soft rock conditions provided in an embodiment of the present invention; wherein, Figure 4 (a) in the figure shows the variation of probability residuals under hard rock conditions. Figure 4 (b) in the figure shows the variation of the probability residual under soft rock conditions;

[0060] Figure 5 This is a graph showing the relationship between the cumulative probability residual and the early warning threshold and alarm threshold provided in an embodiment of the present invention.

[0061] Figure 6 This is a diagram showing the early warning results of changes in weak surrounding rock strata provided in an embodiment of the present invention.

[0062] Figure 7 Histograms of probability residual distribution under hard and soft rock conditions provided in embodiments of the present invention. Detailed Implementation

[0063] The invention will be explained in detail through the following embodiments. The purpose of this invention is to protect all technical improvements within its scope. In the description of this invention, it should be understood that 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. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0064] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0065] Example 1

[0066] refer to Figure 1 This embodiment provides a real-time sensing method for changes in weak surrounding rock strata of a TBM, including the following steps:

[0067] S1. Collect multi-source high-frequency time series data of TBM under various construction conditions, and construct a multi-source time series database by combining the correspondence between surrounding rock grade and construction mileage section.

[0068] Specifically, the following steps are included:

[0069] S11. Collect multi-source high-frequency time-series data of the same type of TBM under various construction conditions through the TBM monitoring system; the sampling frequency of the data is 1Hz; the multi-source high-frequency time-series data includes at least timestamps, tunneling parameters and operating parameters; the various construction conditions include conventional geological conditions and different surrounding rock conditions.

[0070] One timestamp corresponds to a set of tunneling parameters and operating parameters; the tunneling parameters include one or more of thrust, cutterhead torque, penetration depth, propulsion speed and power; the operating parameters include one or more of rotation speed setpoint and thrust setpoint; the timestamp is to clarify the time index relationship of multi-source high-frequency time series data, and is not to add independent construction parameters.

[0071] S12. Combine advanced geological forecast results, design documents and field survey data to determine the surrounding rock grade information of the corresponding construction mileage section, and establish the correspondence between the surrounding rock grade and the construction mileage section.

[0072] S13. The surrounding rock grade information is associated with the multi-source high-frequency time series data (tunneling parameters and operation parameters) collected in step S11 using the mileage alignment method to form a multi-source time series database containing tunneling parameters, operation parameters, timestamps, mileage sections and surrounding rock grades; the multi-source time series database serves as the basic dataset for subsequent data preprocessing, model training and weak surrounding rock identification.

[0073] S2. The multi-source time-series database is divided into cyclic tunneling segments according to the continuity of construction mileage to obtain a cyclic tunneling segment dataset. After outlier detection and interpolation, the cyclic tunneling segment data is standardized to obtain a standardized cyclic tunneling segment dataset. Specifically, this includes:

[0074] S21. Select continuous data segments that meet the construction mileage continuity condition from the multi-source time series database to obtain the cyclic tunneling segment dataset;

[0075] The continuity condition for the construction mileage is: a continuous data segment in which the thrust of the tunneling parameters is greater than 5kN and the time length is greater than 1000 seconds.

[0076] Based on the continuity of construction mileage, the multi-source time-series database is filtered according to the condition that the thrust of the tunneling parameters is greater than 5kN. Continuous data segments (circular tunneling segment data) that meet the above conditions are extracted and segmented to obtain a continuous data segment set. Then, continuous data segments with a time length of less than 1000s are removed from the continuous data segment set to obtain the circular tunneling segment data segment set.

[0077] Among them, a thrust greater than 5kN is used to exclude data in the state of shutdown, standby, no load or non-tunneling; a time length greater than 1000 seconds is used to ensure that the selected data segments have sufficient continuity and time sequence length, which is convenient for subsequent model training samples.

[0078] The continuous data segment refers to a time-series data segment where adjacent sampling points are sequentially arranged according to timestamps, and each sampling point satisfies the condition of a thrust greater than 5kN. Since the sampling frequency is 1Hz, a time length greater than 1000 seconds can be understood as the continuous data segment containing at least approximately 1000 continuous sampling points. This continuous data segment corresponds to the data of one cyclic tunneling section.

[0079] S22. Using box plot and linear interpolation methods, outlier detection and interpolation are performed on each cyclic tunneling segment in the cyclic tunneling segment dataset to obtain corrected cyclic tunneling segment data; this step is the data preprocessing process.

[0080] S23. The Z-score standardization method is used to standardize the input features (tunneling parameters and operation parameters) to normalize parameters of different dimensions to a unified scale, thereby obtaining a standardized cyclic tunneling segment dataset for model training and inference.

[0081] The input features refer to the numerical feature variables that participate in model training and inference, mainly including tunneling parameters and operational parameters. Timestamps are mainly used to determine the time sequence of each sampled data, construct continuous time series and sliding windows, and are not used as input features that require Z-score standardization; mileage sections are mainly used to locate the construction location corresponding to the data or to segment the data; surrounding rock grade is mainly used to identify or verify the strata / surrounding rock condition.

[0082] Z-score standardization (also known as standardization) is a statistical method that transforms raw data into a distribution with a mean of 0 and a standard deviation of 1. This is a mature existing technology, and how to standardize input features using the Z-score standardization method is also a mature existing technology, which will not be elaborated here.

[0083] S3. Using the Bayesian gated recurrent unit as the basic prediction model, a Bayesian time series prediction model is constructed using MC Dropout technology, and its prediction target is set as the predicted value of the cutterhead torque. Then, the Bayesian time series prediction model is trained using the standardized recurrent tunneling section dataset to obtain the trained Bayesian time series prediction model; specifically including the following steps:

[0084] S31. After completing the data preprocessing and standardization process in step S2 to obtain the standardized cyclic tunneling segment dataset, a Bayesian gated cyclic unit is used as the basic prediction model. The weight parameters and bias parameters corresponding to the input gate, forget gate and output gate in the basic prediction model are set to follow a Gaussian probability distribution to characterize the uncertainty of the model parameters.

[0085] The Gaussian probability distribution, also known as the normal distribution, is one of the most important probability distributions in probability theory. It is a continuous probability distribution proposed by the German mathematician Karl Gauss in 1809, hence its name. The probability density function of the Gaussian probability distribution is a bimodal curve with a peak value, extending outwards on both sides, resembling a bell shape, and is therefore also called a bell curve.

[0086] The structure of the Bayesian gated recurrent unit (Bayesian LSTM) is as follows: Figure 2 As shown, it includes an input gate and a forget gate.

[0087] Output gates and memory units, among which, For a moment The input vector, For a moment The hidden state, For a moment The cell state, For input merge nodes; , , and These are the weight parameters corresponding to the output gate, candidate cell state, input gate, and forget gate, respectively. For output gate, The candidate cell state is... For input gate, For the Gate of Oblivion For a moment The cell state, For a moment The hidden state; It is the Sigmoid activation function. It is the hyperbolic tangent activation function;

[0088] S32. Based on the basic prediction model described in step S31, the MC Dropout technique is used to approximate efficient Bayesian inference and construct a Bayesian time series prediction model.

[0089] Monte Carlo Dropout (MC Dropout) is a Bayesian interpretation of dropout, viewing it as a Bayesian approximation of a Gaussian process. It's a deep learning inference method that combines Monte Carlo methods and Dropout regularization, primarily used to quantify the predictive uncertainty of neural network models. It estimates the model's predictive distribution given an input by simulating multiple forward propagation processes, thus providing an uncertainty assessment. As a current technique, it will not be elaborated upon further here.

[0090] The Bayesian time series prediction model is a time series prediction network built using Bayesian LSTM as the core prediction unit. Its overall structure includes an input layer, a Bayesian LSTM layer, a Dropout layer, a fully connected layer, and an output layer; wherein:

[0091] 1) The input layer is used to receive the standardized cyclic tunneling segment dataset obtained from step S2 and organize it into the time series input form required by the model;

[0092] 2) The Bayesian LSTM layer is a temporal feature extraction unit formed by probabilistically modeling the weight parameters and bias parameters in the LSTM network. Its basic computational structure still uses LSTM units, including input gate, forget gate, output gate and memory unit. The difference is that the weight parameters and bias parameters involved in the above gated computation no longer take fixed values, but are represented by probability distribution and learned through variational Bayesian methods.

[0093] 3) The Dropout layer is used to perform Monte Carlo Dropout sampling during the model prediction phase. It performs multiple forward propagations through multiple randomly deactivated neurons to obtain the distribution information of the prediction results.

[0094] 4) Fully connected layers are used to map and fuse the hidden state features output by the Bayesian LSTM layers, thereby achieving feature dimension transformation;

[0095] 5) The output layer is used to output the predicted value of the tool turret torque;

[0096] The Bayesian time series prediction model adds Dropout masks to the fully connected layer and the output layer, randomly discarding 20% ​​to 50% of neurons during model training and inference to simulate weight uncertainty.

[0097] The role of neurons: Neurons in each layer perform nonlinear mapping and feature extraction on the input features through nonlinear activation functions, thereby establishing the mapping relationship between the tunneling parameter time series and the cutterhead torque, and obtaining the predicted distribution of the cutterhead torque through multiple sampling predictions, providing a basis for subsequent probability discrimination of weak surrounding rock.

[0098] S33. The Bayesian time series prediction model is trained using a standardized cyclic tunneling section dataset, and the prediction target of the Bayesian time series prediction model is set as the predicted value of the cutterhead torque, thereby obtaining the trained Bayesian time series prediction model, which is used for subsequent tunneling parameter prediction and weak surrounding rock probability discrimination.

[0099] This step employs a Bayesian time series prediction model to predict tunneling parameters and uses MC Dropout technology to quantify model uncertainty. This expands the prediction results from traditional single-point predictions to a probability distribution that includes both the prediction mean and the prediction uncertainty, which is the key protection point of this invention. While existing research has involved using Bayesian probability models for prediction analysis, it has not utilized the probability distribution output by this model for the discrimination analysis of weak surrounding rock formations. Therefore, this invention can provide a new technical foundation for subsequent probability-based identification of weak surrounding rock. Compared to existing technologies that only output deterministic prediction values ​​and use them to perform residual subtraction or anomaly alarms, this invention can simultaneously obtain the prediction mean and prediction uncertainty during the prediction stage, providing a more stable and interpretable probabilistic basis for subsequent weak surrounding rock identification, thereby improving the reliability of identification and engineering applicability.

[0100] S4. Construct a weak surrounding rock strata discrimination model based on the prediction residuals according to the Bayesian time series prediction model;

[0101] Specifically, it includes:

[0102] S41. Based on the training of the Bayesian time series prediction model described in step S3, the first N seconds of data of the tunneling section with potentially weak surrounding rock strata during actual construction are obtained as input data.

[0103] S42. Input the input data into the Bayesian time series prediction model and perform T repeated prediction samplings to obtain T prediction samples for each prediction time from N seconds to M seconds in the future.

[0104] Specifically, using the same input data, the tunneling parameters (propulsion force, torque, etc.), operating parameters, and surrounding rock conditions of the first N seconds of the tunneling section are used as historical input variables, and the operating parameters and surrounding rock conditions of the next N to M seconds are used as future auxiliary variables. The historical input variables are input into the Bayesian time series prediction model trained in step S3, and T forward propagations are performed. Each forward propagation uses a random Dropout mask and outputs the prediction results for each prediction time in the future prediction period, so that each prediction time in the future obtains T corresponding prediction samples.

[0105] Wherein, N takes values ​​in [30, 90], M takes values ​​in [800, 1200], and T takes values ​​in [60, 140]; preferably, in this embodiment, N is 60, M is 1000, and T is 100.

[0106] S43. Based on the T predicted samples obtained in step S42 for each prediction time, use the sample mean at time t. variance at time t Predicted probability distribution of parameters at approximately future time t As shown in the following formula:

[0107] ,

[0108] in, ; ; This is the set of predicted cutterhead torque values ​​at time t; Let be the predicted value of the cutterhead torque based on the i-th sampling at time t; Indicates used for prediction The historical input feature sequence of the cutterhead torque at any given time, where D represents the sample dataset used to train the Bayesian time series prediction model.

[0109] The predicted probability distribution Approximated by a normal distribution;

[0110] S44. Based on the predicted probability distribution obtained in step S43 To obtain the predicted probability distribution sequence for each time point from N seconds to M seconds in the future;

[0111] S45. Based on the predicted probability distribution sequence, calculate the single-point probability residual by quantifying the deviation probability between the actual tunneling parameters and the predicted probability distribution, specifically including:

[0112] The deviation probability between the quantified actual tunneling parameters and the predicted probability distribution Calculated using the following formula:

[0113] ,

[0114] in, Indicates the actual tunneling parameters. Let be the sample mean at time t. Let be the variance at time t;

[0115] The single-point probability residual Calculated using the following formula:

[0116] ,

[0117] in, The abnormal threshold is selected in [1.5, 3] based on the actual engineering situation; Φ(·) is the standard normal distribution CDF formula, which is an extremely mature existing technology and belongs to the basic theory of probability theory and mathematical statistics, and will not be elaborated here;

[0118] S46. Based on the single-point probability residual calculated in step S45, a sliding time window is defined, and the cumulative probability residual within the sliding time window at time t is calculated; the size of the sliding time window is m seconds, representing the time period m seconds before time t;

[0119] The cumulative probability residual Calculated based on the following formula;

[0120] ,

[0121] in, This represents the single-point probability residual at time j;

[0122] m takes a value in [15, 25], and preferably, in this embodiment, m is 20;

[0123] The purpose of step S4 is to introduce a sliding time window, based on the single-point probability residuals obtained in step S45, to accumulate and average the single-point probability residuals of the most recent m sampling times, so as to reduce the impact of single-point fluctuations on the judgment result. Since the sampling frequency is 1Hz and the sliding time window size is m seconds, it corresponds to the most recent m consecutive sampling points.

[0124] S47. Based on the cumulative probability residual, construct the weak surrounding rock discrimination rule to obtain the weak surrounding rock strata discrimination model.

[0125] Specifically, this includes: the cumulative probability residual over K consecutive time periods. If the risk exceeds the preset risk threshold, it is marked as a high-probability weak surrounding rock section and a weak surrounding rock section warning is triggered.

[0126] like If the preset risk threshold is not reached, or if the condition that the current tunneling parameters are greater than the preset risk threshold for K consecutive sampling times is not met, it indicates that the deviation between the current tunneling parameters and the model prediction distribution is still within the normal fluctuation range. It is determined that the current surrounding rock condition has not shown any continuous abnormality, the system maintains normal monitoring status, and does not trigger a weak surrounding rock warning.

[0127] It is worth noting that the cumulative probability residual in this step It does not directly represent the true probability of weak surrounding rock, but rather is a statistical indicator used to characterize the degree of deviation between tunneling parameters and the model's predicted distribution. When the properties of the surrounding rock change (such as entering a weak surrounding rock section), the actual cutterhead torque and other tunneling parameters often continuously deviate from the predicted distribution obtained based on normal surrounding rock conditions, resulting in the probability residual gradually accumulating and increasing within a continuous time window. Therefore, the cumulative probability residual can serve as a characterizing indicator of the degree of risk associated with weak surrounding rock.

[0128] In this embodiment, the preset risk threshold is 0.8. This preset risk threshold is a risk discrimination threshold determined based on historical tunneling data statistical analysis and engineering experience, used to distinguish between normal fluctuations and continuous abnormal deviations. In practical applications, this threshold can be adjusted according to different engineering conditions, shield equipment types, or historical data characteristics. For example, it can be determined through historical sample statistical analysis, cross-validation, or empirical calibration. Therefore, this threshold is not fixed and unchangeable. Furthermore, this preset risk threshold is used to identify continuous abnormal deviations in tunneling parameters, thereby determining the possible occurrence of weak surrounding rock sections. It is not applicable to a specific type of surrounding rock, but rather used to identify surrounding rock sections that show significant anomalies compared to normal strata conditions, and can be varied according to different surrounding rock types.

[0129] In step S47, the discrimination model for weak surrounding rock strata is a regular discrimination model constructed based on cumulative probability residuals. Specifically, this model uses time... Cumulative probability residual The input for discrimination is that the cumulative probability residuals at K consecutive time points are all greater than the preset risk threshold. The output for discrimination is whether to mark it as a high-probability weak surrounding rock section and trigger an early warning.

[0130] When the aforementioned conditions are met, it indicates that the actual tunneling parameters deviate abnormally from the predicted probability distribution trained based on normal surrounding rock conditions. Therefore, it is marked as a high-probability weak surrounding rock section and an early warning is triggered. When the aforementioned conditions are not met, it indicates that the degree of deviation is still within the normal fluctuation range, and the system maintains normal monitoring status. Thus, the cumulative probability residual, the preset risk threshold, and the discrimination rules for K consecutive time periods together constitute the weak surrounding rock strata discrimination model.

[0131] This step involves performing T repeated prediction samplings on the same input data to obtain multiple prediction samples for each time point from 60s to 1000s in the future. The sample mean and variance are used to approximate the probability distribution of future tunneling parameters. Then, by quantifying the "probability of deviation between actual tunneling parameters and predicted probability distribution", a probability residual is constructed to characterize the risk of changes in weak surrounding rock strata. This realizes the transformation from the traditional residual identification of "whether it is abnormal" to the discrimination method of "risk probability quantification". Making the weak surrounding rock identification results more interpretable and valuable for decision-making is an important protection point of this invention.

[0132] Furthermore, this invention proposes a method for defining the probability distribution interval of the abnormal threshold. The abnormal threshold is derived from the statistical interval of the predicted probability distribution rather than an empirically fixed threshold, which has cross-project universality and does not require manual adjustment for specific projects.

[0133] Furthermore, this invention proposes cumulative probability residuals as a new discrimination index. By accumulating single-point probability residuals through a sliding time window and combining this with continuity discrimination conditions, "high-probability weak surrounding rock sections" are marked and an early warning is triggered. Compared to existing technologies that rely on instantaneous residuals or single-point thresholds, this invention can distinguish between sporadic noise fluctuations and persistent anomalies caused by changes in surrounding rock conditions, thereby reducing the risk of false alarms and missed alarms, improving the stability and timeliness of weak surrounding rock identification, and enhancing the safe tunneling capability of TBMs under complex geological conditions.

[0134] S5. The weak surrounding rock strata discrimination model is used for subsequent TBM construction to provide early warning of weak surrounding rock.

[0135] This embodiment provides a real-time sensing method for changes in weak surrounding rock formations during TBM construction. It aims to overcome the limitations of existing deterministic prediction models by introducing Bayesian time series prediction and MC Dropout technology to predict the probability distribution of future evolution trends of TBM tunneling parameters. This invention utilizes the deviation between measured data and the predicted probability distribution to construct a "probability residual" as a quantitative criterion, thereby establishing a weak surrounding rock discrimination mechanism that does not rely on manually fixed thresholds and possesses adaptive capabilities. This method can effectively filter out occasional noise interference, accurately capture persistent anomalies caused by changes in weak surrounding rock formations, and provide real-time, reliable, and statistically interpretable geological risk probability early warnings for TBM construction, significantly improving construction safety and decision-making efficiency under complex geological conditions.

[0136] Example 2

[0137] refer to Figure 2 This embodiment provides a real-time sensing system for changes in weak surrounding rock formations during TBM operations. Based on the real-time sensing method for changes in weak surrounding rock formations during TBM operations described in this embodiment, it includes the following modules:

[0138] The multi-source time series database construction module is used to collect multi-source high-frequency time series data of TBM under conventional strata and different surrounding rock conditions, and to construct a multi-source time series database by combining the correspondence between surrounding rock grade and construction mileage section.

[0139] The cyclic tunneling section dataset construction module is used to divide the multi-source time-series database into cyclic tunneling sections according to the continuity of construction mileage to obtain the cyclic tunneling section dataset. After outlier detection and interpolation processing, the cyclic tunneling section data is standardized to obtain the standardized cyclic tunneling section dataset.

[0140] The Bayesian time series prediction model construction module is used to construct a Bayesian time series prediction model using the Bayesian gated cyclic unit as the basic prediction model and MC Dropout technology. The prediction target is set as the predicted value of the cutterhead torque. Then, the Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset to obtain the trained Bayesian time series prediction model.

[0141] The weak surrounding rock strata discrimination model construction module is used to construct a weak surrounding rock strata discrimination model based on the prediction residuals according to the Bayesian time series prediction model.

[0142] The early warning application module is used to apply the weak surrounding rock strata discrimination model to the subsequent construction of TBM to provide early warning of weak surrounding rock.

[0143] Example 3

[0144] This embodiment uses actual TBM tunneling project data as an example to experimentally verify the real-time sensing method for changes in weak surrounding rock strata described in Embodiment 1. The experiment uses normal TBM tunneling data under hard rock strata conditions as model learning and normal state characterization data, and subsequent TBM tunneling data under soft rock strata conditions as verification data for changes in weak surrounding rock strata. By comparing the changes in cutterhead torque prediction results, probability residuals, and cumulative probability residuals under different surrounding rock conditions, the effectiveness of the real-time sensing method for changes in weak surrounding rock strata proposed in this invention is verified.

[0145] This embodiment focuses on verifying the following: First, it verifies the ability of the Bayesian time series prediction model to learn and predict the change law of cutterhead torque under normal hard rock strata conditions; second, it verifies whether there is an identifiable and persistent deviation between the actual cutterhead torque and the predicted probability distribution when the strata change from hard rock conditions to soft rock conditions; and third, it verifies whether the weak surrounding rock discrimination method based on probability residuals and cumulative probability residuals can achieve timely early warning of changes in weak surrounding rock strata.

[0146] In this embodiment, the sampling interval for the raw TBM tunneling data is 10 seconds, and each raw data file includes 495 columns of TBM operation and status parameters. The data acquisition time range for hard rock formations is from October 10, 2025 to March 28, 2026, including a total of 168 raw data files; the data acquisition time range for soft rock formations is from March 29, 2026 to May 21, 2026, including a total of 54 raw data files.

[0147] Continuous tunneling segments were extracted from the raw data. In this embodiment, a total thrust greater than 0 and a cutterhead rotation speed greater than 0 were used as screening criteria for the TBM to be in an effective tunneling state. Data areas continuously meeting these criteria were divided into continuous tunneling segments, and short-term data segments with less than 20 sampling points were discarded. Since the data sampling time interval in this embodiment is 10 seconds, 20 sampling points correspond to 200 seconds of continuous tunneling time. After processing, a total of 366 hard rock continuous tunneling segments and 233 soft rock continuous tunneling segments were obtained. The hard rock continuous tunneling segments were used for training the Bayesian time series prediction model and establishing the probability distribution and risk threshold of normal tunneling states, while the soft rock continuous tunneling segments were used for testing and verifying changes in weak surrounding rock strata.

[0148] This embodiment uses cutterhead torque as the target predictor variable. The historical exogenous variables used in the model include cutterhead rotational speed, total thrust, feed rate, cutterhead penetration, torque penetration index (TPI), field penetration index (FPI), unit energy consumption (SE), average motor operating current, average motor temperature, average motor output torque, and average motor output frequency; future exogenous variables use the given motor frequency.

[0149] Among them, the Torque Penetration Index (TPI), Field Penetration Index (FPI), and Energy Consumption per Unit (SE) are calculated according to the following formulas:

[0150] TPI = Tool head torque / (Number of tools × Penetration);

[0151] FPI = Total Thrust / (Number of Cutters × Penetration);

[0152] SE = Total thrust / cutter head area + (2π / cutter head area) × cutter head rotation speed × cutter head torque / (feed speed × 1000 / 60).

[0153] In this embodiment, 34 TBM cutters are used, with a cutterhead area of ​​21.2372 m². The average operating current, average temperature, average output torque, and average output frequency of the motors are obtained by averaging the corresponding parameters of 16 motors (M1 to M16). Through the above feature construction, the model can comprehensively utilize the force state of the TBM cutterhead, the propulsion state, the interaction state between the rock machine and the TBM, and the operating state of the drive motor to model the torque variation law of the cutterhead.

[0154] The model's historical input window (input_size) is set to 24 sampling points, and the prediction window (h) is also set to 24 sampling points. With a 10-second sampling interval, each input window contains historical tunneling information for 24 consecutive time points. The model continuously outputs the predicted cutterhead torque for the next 24 sampling time points.

[0155] This embodiment uses a Bayesian LSTM as the core temporal encoder. The model encoder is set to a 2-layer Bayesian LSTM with a hidden layer dimension of 512; the decoder uses a 2-layer Bayesian MLP with a hidden layer dimension of 512. The model training steps are set to 60,000, the learning rate is set to 0.001, the batch size is set to 32, the gradient clipping value is set to 1.0, the learning rate decay times are set to 15, and the KL divergence weight complexity_cost_weight is set to 1×10^-6.

[0156] To reduce the impact of different TBM parameter dimensions and extreme values ​​on the model training results, this embodiment uses a robust approach for data scaling, where both scaler_type and local_scaler_type are set to robust. The upper limit for trimming the predicted cutter head torque value is set to 18000 kN·m.

[0157] The experimental calculations in this embodiment used an NVIDIA GeForce RTX 5090 D graphics processor and a PyTorch 2.11.0+cu128 deep learning framework.

[0158] Two Monte Carlo sampling operations are performed during model training; during model prediction, the Dropout is kept in a random deactivated state, and the same input data is repeatedly forward-propagated 30 times, i.e., T=30. Since each forward propagation corresponds to a different random Dropout mask, 30 toolhead torque prediction samples can be obtained for each future prediction time.

[0159] Based on 30 predicted samples, the sample mean and sample variance of the cutterhead torque at each predicted time are calculated, and the predicted probability distribution of the cutterhead torque at the corresponding time is approximated accordingly. Therefore, the model output is no longer a single deterministic predicted cutterhead torque value, but can simultaneously obtain the predicted center location and prediction uncertainty information, providing a foundation for subsequent identification of weak surrounding rock based on the predicted probability distribution.

[0160] After model training, the performance of the model in predicting cutterhead torque under normal formation conditions is first evaluated using hard rock formation data. In this embodiment, mean absolute error (MAE), coefficient of determination (R²), and mean absolute percentage error (MAPE) are used as evaluation metrics for deterministic prediction results.

[0161] Under hard rock formation conditions, the model predicted a MAE of 175.58 kN·m, an R² of 0.8889, and a MAPE of 7.62%. These results indicate that the model can effectively learn the temporal mapping relationship between TBM operating parameters and cutterhead torque under normal hard rock tunneling conditions, and the predicted results show high consistency with the actual cutterhead torque.

[0162] Subsequently, the trained Bayesian time series prediction model was applied to soft rock formation data. Under soft rock formation conditions, the MAE of the model prediction increased to 1035.33 kN·m, and the R² was 0.8193.

[0163] refer to Figure 3 The input and prediction time is 240s. Under hard rock conditions, the model prediction results can follow the actual cutterhead torque variation trend well, and most of the actual values ​​are within the normal variation range represented by the model prediction results. When the stratum conditions change to soft rock, there is a more obvious deviation between the actual cutterhead torque and the prediction results obtained based on the normal tunneling law of hard rock. Moreover, this deviation is not only manifested as an instantaneous change at a few moments, but also exists continuously within a certain continuous time range.

[0164] The results, showing a MAE of 175.58 kN·m under hard rock conditions and an increase to 1035.33 kN·m under soft rock conditions, indicate that the cutterhead torque time-series mapping relationship established under the original normal formation conditions changes significantly when the TBM transitions from hard rock to soft rock formations. This demonstrates that changes in formation properties can be characterized by the deviation between the TBM's real-time tunneling parameters and the model's normal predicted state, thus providing a basis for further utilizing predicted probability distributions to identify changes in weak surrounding rock in this invention.

[0165] Using only the ordinary absolute residual between the actual value and the predicted mean is easily affected by TBM operation adjustments, local shocks, and measurement noise. Therefore, this embodiment further calculates the probability residual of the actual cutterhead torque relative to the corresponding predicted probability distribution based on the predicted mean and prediction uncertainty output by the Bayesian time series prediction model, following the method described in Embodiment 1.

[0166] refer to Figure 4 During the normal tunneling stage in hard rock, the probability residuals remain within a relatively stable range of variation. However, after entering soft rock formations, the fluctuation characteristics of the probability residuals change accordingly as the actual cutterhead torque gradually deviates from the predicted probability distribution established based on the normal state of hard rock.

[0167] Statistical analysis showed that the mean probability residual for hard rock data was -0.0020, while the mean probability residual for soft rock data was -0.0259. These results indicate that the statistical deviation characteristics of the actual cutterhead torque from the model's predicted probability distribution differ under different surrounding rock conditions. Therefore, compared to using only the deterministic difference between the actual value and the predicted mean, the probability residual can further incorporate the model's own prediction uncertainty, enabling a unified quantification of the degree of prediction deviation under different working conditions.

[0168] Considering the potential for localized impacts, instantaneous parameter adjustments, and short-term sensor fluctuations during TBM construction, the probability residual at a single moment cannot fully demonstrate that the surrounding rock condition has changed. To reduce the impact of transient anomalies on the judgment results, this embodiment further accumulates the probability residual within a sliding time window.

[0169] In this embodiment, the scrolling window size is set to 30 sampling points. Since the sampling time interval is 10 seconds, one scrolling window corresponds to a continuous tunneling process of 300 seconds. The cumulative probability residual (CPR) is calculated using the probability residual within the window to characterize the degree of continuous deviation between the actual tunneling state and the model's normal predicted state over a continuous time period.

[0170] First, the early warning threshold and alarm threshold for weak surrounding rock are established using the statistical characteristics of the cumulative probability residuals from hard rock training data. Let the statistical mean of the cumulative probability residuals under hard rock conditions be μ_CPR, and the standard deviation be σ_CPR. Then, in this embodiment, the following will be used:

[0171] The warning threshold is set to μ_CPR+2σ_CPR;

[0172] The alarm threshold is set to μ_CPR+3σ_CPR;

[0173] Among them, the early warning threshold is used to identify formation changes that have deviated significantly from the normal hard rock tunneling state but have a relatively low degree of abnormality; the alarm threshold is used to identify high-risk abnormal states where the degree of deviation has further increased.

[0174] Based on the statistical results of the hard rock training data in this embodiment, the warning threshold is 0.3247 and the alarm threshold is 0.4504. Simultaneously, a calibration verification of the hard rock model's prediction probability interval was performed. The coverage rate of hard rock data within the model's 95% prediction confidence interval reached 99.9%, indicating that the constructed Bayesian time series prediction model can stably characterize the uncertainty range of cutterhead torque under normal hard rock conditions.

[0175] refer to Figure 5 Under normal hard rock strata conditions, most of the cumulative probability residuals remain within the normal statistical range established based on historical hard rock data; however, when the strata enter the soft rock region, the cumulative probability residuals show a significant increase and continuously exceed the warning threshold and alarm threshold.

[0176] In this embodiment, the maximum cumulative probability residual for hard rock data is 3.4899, and the maximum cumulative probability residual for soft rock data is 3.7895. More importantly, compared to single-point probability residuals, the cumulative probability residuals obtained using a sliding window can comprehensively reflect the persistence of abnormal deviations within a certain time range, thereby reducing the impact of isolated anomalies on the identification results of weak surrounding rock.

[0177] Based on the aforementioned cumulative probability residuals, a weak surrounding rock early warning judgment is performed. When the cumulative probability residuals exceed the early warning threshold, the system enters a weak surrounding rock early warning state; when the cumulative probability residuals further exceed the alarm threshold, the system enters a weak surrounding rock alarm state. For continuous judgment, the alarm result is further confirmed according to the judgment rule for K consecutive sampling times described in Example 1, thereby avoiding false triggering caused by anomalies in a single sampling time.

[0178] refer to Figure 6 After monitoring of soft rock test data began, the cumulative probability residual gradually changed as the deviation between the actual TBM tunneling condition and the predicted distribution of normal hard rock increased. The system first triggered a weak surrounding rock warning at the 156th sampling point, corresponding to approximately 26.0 minutes of monitoring time; and further triggered a weak surrounding rock alarm at the 162nd sampling point, corresponding to approximately 27.0 minutes of monitoring time.

[0179] As can be seen from the results of the first warning and the first alarm, the present invention can automatically generate risk warnings after the tunneling state continues to be abnormal due to weak surrounding rock by using the operating parameters obtained in real time during the TBM tunneling process, without relying solely on manual observation of whether the torque and propulsion parameters of a single cutterhead exceed the empirical limits.

[0180] Furthermore, in order to compare the differences in probability residuals under different surrounding rock conditions from a statistical distribution perspective, the probability residuals corresponding to hard rock data and soft rock data were statistically analyzed, and histograms of probability residual distribution were plotted.

[0181] refer to Figure 7 The probability residuals of hard rock data and soft rock data differ in their overall distribution characteristics. Under hard rock conditions, the actual tunneling state of the TBM is highly consistent with the predicted probability distribution learned from hard rock data; therefore, the probability residuals are mainly concentrated within the statistical range corresponding to the normal state. Under soft rock conditions, changes in the surrounding rock properties lead to changes in the cutterhead stress state, propulsion state, and rock-machine interaction, resulting in a more significant statistical deviation of the actual cutterhead torque from the original normal predicted probability distribution, thus altering the distribution pattern of the probability residuals.

[0182] Figure 7 The overall statistical results shown further illustrate that Figures 3-6 The anomalies in the weak surrounding rock observed in the data do not originate from a small number of isolated sampling points, but rather exhibit probability deviations that differ from those of normal hard rock in the overall statistical characteristics of the soft rock data.

[0183] comprehensive Figures 3-7As can be seen from the above quantitative results, in this embodiment, the Bayesian time series prediction model can accurately predict the change of cutterhead torque under normal hard rock formation conditions, and obtain the uncertainty information of the prediction results through Monte Carlo Dropout; when the formation changes from hard rock to soft rock, the actual cutterhead torque of the TBM gradually deviates from the predicted probability distribution formed under normal hard rock conditions, causing the probability residual and cumulative probability residual to change accordingly.

[0184] By utilizing the statistical characteristics of the cumulative probability residuals of normal hard rock data to establish early warning and alarm thresholds, the engineering adaptability problems caused by relying entirely on manual experience to set fixed cutterhead torque or propulsion force thresholds can be reduced. Furthermore, by using a sliding time window to accumulate the probability residuals, the impact of instantaneous noise and local working condition adjustments on the judgment results can be reduced, making persistent anomalies more prominently characterized.

[0185] In this embodiment, the system successfully triggered a weak surrounding rock early warning at the 156th sampling point and triggered a further alarm at the 162nd sampling point, thus realizing the effective identification of changes in the weak surrounding rock strata. This verifies the feasibility and effectiveness of the present invention for real-time perception of changes in weak surrounding rock strata using TBM based on Bayesian time series prediction, probability residuals, and cumulative probability residuals.

[0186] It should be noted that the sampling time interval, input window length, prediction window length, Monte Carlo sampling count, Bayesian LSTM hidden layer dimension, sliding window length, warning threshold, and alarm threshold in this embodiment are parameter settings determined based on specific engineering data in this embodiment, and do not constitute a limitation on the scope of protection of this invention. In other embodiments, the above parameters can be adjusted accordingly based on the TBM equipment type, data sampling frequency, geological conditions, and statistical characteristics of historical tunneling data.

[0187] It should be noted that this specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It should 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 processor, 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 and / or block diagrams. Figure 1 A process, multiple processes, and / or boxes Figure 1 Devices that specify the functions in one or more boxes.

[0188] 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 an instruction device, which is implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0189] 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.

[0190] The parts of this invention not described in detail are prior art. It will be apparent to those skilled in the art that this invention is not limited to the details of the above exemplary embodiments, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and are intended to encompass all changes falling within the meaning and scope of equivalents within this invention.

Claims

1. A method for real-time sensing of changes in weak surrounding rock strata during TBM (Tunnel Boring Machine) formations, characterized in that, Includes the following steps: S1. Collect multi-source high-frequency time series data of TBM under various construction conditions, and construct a multi-source time series database by combining the correspondence between surrounding rock grade and construction mileage section. S2. Divide the multi-source time-series database into cyclic tunneling segments according to the continuity of construction mileage to obtain a cyclic tunneling segment dataset. After outlier detection and interpolation processing, the cyclic tunneling segment data is standardized to obtain a standardized cyclic tunneling segment dataset. S3. Using the Bayesian gated cyclic unit as the basic prediction model, a Bayesian time series prediction model is constructed using MC Dropout technology, and its prediction target is set as the predicted value of the cutterhead torque. Then, the Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset to obtain the trained Bayesian time series prediction model. S4. Construct a weak surrounding rock strata discrimination model based on the prediction residuals according to the Bayesian time series prediction model; S5. The weak surrounding rock strata discrimination model is used for subsequent TBM construction to provide early warning of weak surrounding rock.

2. The method according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Collect multi-source high-frequency time-series data of the same model of TBM under various construction conditions through a TBM monitoring system; the multi-source high-frequency time-series data includes timestamps, tunneling parameters, and operating parameters; wherein, one timestamp corresponds to a set of tunneling parameters and operating parameters; the tunneling parameters include thrust, cutterhead torque, penetration depth, propulsion speed, and power; the operating parameters include speed setting values ​​and thrust setting values; the various construction conditions include conventional geological conditions and different surrounding rock conditions; S12. Combine advanced geological forecast results, design documents and field survey data to determine the surrounding rock grade information of the corresponding construction mileage section, and establish the correspondence between the surrounding rock grade and the construction mileage section. S13. The surrounding rock grade information is associated with the multi-source high-frequency time series data using a mileage alignment method to form a multi-source time series database.

3. The method according to claim 2, characterized in that, Step S2 specifically includes the following steps; S21. Select continuous data segments that meet the construction mileage continuity condition from the multi-source time series database to obtain the cyclic tunneling segment dataset; S22. Perform outlier detection and interpolation on each cycle tunneling segment data in the cycle tunneling segment dataset to obtain corrected cycle tunneling segment data; this step is the data preprocessing process. S23. Standardize the tunneling parameters and operating parameters to normalize parameters with different dimensions to a unified scale, and obtain a standardized cyclic tunneling section dataset.

4. The method according to claim 3, characterized in that, Step S3 specifically includes the following steps: S31. A Bayesian gated recurrent unit is used as the basic prediction model, and the weight parameters and bias parameters corresponding to the input gate, forget gate and output gate in the basic prediction model are set to follow a Gaussian probability distribution. S32. Based on the basic prediction model, the MC Dropout technique is used to approximate efficient Bayesian inference and construct a Bayesian time series prediction model. S33. The Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset, and the prediction target of the Bayesian time series prediction model is set as the predicted value of the cutterhead torque, so as to obtain the trained Bayesian time series prediction model.

5. The method according to claim 4, characterized in that, Step S4 specifically includes the following steps: S41. Obtain the first N seconds of data from the tunneling section with potentially weak surrounding rock strata during actual construction as input data. S42. Input the input data into the Bayesian time series prediction model and perform T repeated prediction samplings to obtain T prediction samples for each prediction time from N seconds to M seconds in the future. S43. Based on the T predicted samples at each prediction time, use the sample mean at time t. variance at time t Predicted probability distribution of parameters at approximately future time t As shown in the following formula: ， in, ; ; This is the set of predicted cutterhead torque values ​​at time t; Let be the predicted value of the cutterhead torque based on the i-th sampling at time t; Indicates used for prediction The historical input feature sequence of the cutterhead torque at each moment, where D represents the sample dataset used to train the Bayesian time series prediction model; S44. Based on the predicted probability distribution obtained in step S43 To obtain the predicted probability distribution sequence for each time point from N seconds to M seconds in the future; S45. Based on the predicted probability distribution sequence, calculate the single-point probability residual by quantifying the deviation probability between the actual tunneling parameters and the predicted probability distribution, specifically including: The deviation probability between the quantified actual tunneling parameters and the predicted probability distribution Calculated using the following formula: ， in, Indicates the actual tunneling parameters; The single-point probability residual Calculated using the following formula: ， in, Φ(·) represents the outlier threshold; Φ(·) is the CDF formula for the standard normal distribution. S46. Define a sliding time window and calculate the cumulative probability residual within the sliding time window at time t; the size of the sliding time window is m seconds, representing the time period m seconds before time t. S47. Based on the cumulative probability residual, construct the weak surrounding rock discrimination rule to obtain the weak surrounding rock strata discrimination model.

6. The method according to claim 5, characterized in that, The cumulative probability residual Calculated based on the following formula; ， in, Let represent the single-point probability residual at time j.

7. The method according to claim 6, characterized in that, Step S47, which constructs a weak surrounding rock discrimination rule based on the cumulative probability residual, specifically includes: if the cumulative probability residual for K consecutive time periods... If the risk exceeds the preset risk threshold, it is marked as a high-probability weak surrounding rock section, and a weak surrounding rock section warning is triggered.

8. The method according to claim 3, characterized in that, The continuity condition for construction mileage mentioned in step S21 is: a continuous data segment with a thrust of tunneling parameters greater than 5kN and a time length greater than 1000 seconds.

9. The method according to claim 4, characterized in that, The Bayesian time series prediction model described in step S32 includes an input layer, a Bayesian LSTM layer, a Dropout layer, a fully connected layer, and an output layer; wherein: The input layer receives the standardized cyclic tunneling segment dataset and organizes it into the time-series input format required by the model. The Bayesian LSTM layer is a time-series feature extraction unit formed by probabilistically modeling the weight and bias parameters in the LSTM network. The Dropout layer performs Monte Carlo Dropout sampling during the model prediction phase, performing multiple forward propagations through multiple randomly deactivated neurons to obtain the distribution information of the prediction results. The fully connected layer maps and fuses the hidden state features output by the Bayesian LSTM layer to achieve feature dimension transformation. The output layer outputs the predicted cutterhead torque value.

10. A real-time sensing system for changes in weak surrounding rock strata during TBM (Tunnel Boring Machine) operations, characterized in that: The real-time sensing method for changes in weak surrounding rock formations using TBM according to any one of claims 1-9 includes the following modules: The multi-source time series database construction module is used to collect multi-source high-frequency time series data of TBM under conventional strata and different surrounding rock conditions, and to construct a multi-source time series database by combining the correspondence between surrounding rock grade and construction mileage section. The cyclic tunneling section dataset construction module is used to divide the multi-source time-series database into cyclic tunneling sections according to the continuity of construction mileage to obtain the cyclic tunneling section dataset. After outlier detection and interpolation processing, the cyclic tunneling section data is standardized to obtain the standardized cyclic tunneling section dataset. The Bayesian time series prediction model construction module is used to construct a Bayesian time series prediction model using the Bayesian gated cyclic unit as the basic prediction model and MC Dropout technology. The prediction target is set as the predicted value of the cutterhead torque. Then, the Bayesian time series prediction model is trained using the standardized cyclic tunneling section dataset to obtain the trained Bayesian time series prediction model. The weak surrounding rock strata discrimination model construction module is used to construct a weak surrounding rock strata discrimination model based on the prediction residuals according to the Bayesian time series prediction model. The early warning application module is used to apply the weak surrounding rock strata discrimination model to the subsequent construction of TBM to provide early warning of weak surrounding rock.

Images