Rock mass dynamic mechanical parameter sensing method and system based on while-drilling parameters

Abstract

The invention relates to the field of geological exploration and geotechnical engineering, provides a rock mass dynamic mechanical parameter sensing method and system based on while-drilling parameters in order to dynamically obtain rock mass dynamic mechanical parameter characteristics in situ in real time in the drilling process, and provides a rock mass dynamic mechanical parameter sensing method and system based on while-drilling parameters by using multi-modal signals such as vibration, acoustic emission, torque and bit pressure through uniform characteristic engineering and multi-task learning. The instantaneous elastic modulus, the dynamic Poisson's ratio, the transient uniaxial compressive strength, the cohesive force and the internal friction angle can be synchronously and dynamically predicted in single drilling, and the limitation that only a single static parameter is estimated in a traditional method is broken through; a rock mass dynamic mechanical parameter acquisition model constructed on the basis of a multi-modal adaptive graph attention sequential network can effectively couple the physical proximity and the statistical association relationship of a sensor, and integrates expansion sequential convolution and a time perception Transform module, so that the characterization capability of emergencies (such as jamming / friction resistance) and long-term trends is considered on the premise of low time delay; therefore, the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geological exploration and geotechnical engineering, and particularly relates to a rock mass dynamic mechanical parameter sensing method and system based on drilling parameters. BACKGROUND

[0002] In deep resource mining, tunneling, geothermal development, and oil and gas drilling, the in-situ rock mass mechanical parameters are mainly obtained through laboratory core testing, geophysical logging and exploration, simplified drilling speed model strength estimation, and drilling formation evaluation. Laboratory core testing requires stopping drilling to take cores, which is time-consuming and costly, destroys the in-situ state, and cannot reflect the dynamic response under drilling disturbance; geophysical logging and exploration have limited resolution and insufficient real-time performance, the inversion results are mainly macroscopic average characteristics, and it is difficult to capture local stress changes and dynamic responses; simplified drilling speed model strength estimation can only estimate a single static parameter (such as UCS), and the precision is limited, and multiple parameter dynamic mechanical characteristics cannot be obtained; part of the drilling formation evaluation mainly focuses on lithology identification, porosity and pore pressure, and the real-time, multi-parameter, in-situ sensing ability of dynamic mechanical parameters is insufficient. At present, there is no method to obtain the dynamic mechanical parameter characteristics of rock mass in-situ, in real time, and dynamically during drilling, which restricts the safe and efficient construction of deep engineering. SUMMARY

[0003] In order to obtain the dynamic mechanical parameter characteristics of rock mass in-situ, in real time, and dynamically during drilling, a rock mass dynamic mechanical parameter sensing method and system based on drilling parameters are provided.

[0004] The technical scheme adopted by the present application to solve the above problems is: The rock mass dynamic mechanical parameter sensing system based on drilling parameters comprises: A drill bit integrated sensor module, including a probe body, a drill bit measurement unit for obtaining rock drilling vibration spectrum, a drill string monitoring unit for monitoring mechanical load, and an acoustic wave acquisition unit for acquiring acoustic waves, the drill bit measurement unit, the drill string monitoring unit, and the acoustic wave acquisition unit are all arranged on the probe body; A data preprocessing module for preprocessing the original data collected by the drill bit integrated sensor module, including synchronization alignment, hybrid noise reduction, and feature extraction, and standardizing and feature dimensionality reduction of the extracted features; A rock mass dynamic mechanical parameter acquisition model module for constructing a rock mass dynamic mechanical parameter acquisition model based on a multi-modal adaptive graph attention time series network and training; A dynamic mechanical parameter acquisition module for acquiring mechanical parameters based on the preprocessed data and the trained rock mass dynamic mechanical parameter acquisition model.

[0005] Further, the drill bit measurement unit comprises a three-axis MEMS acceleration sensor and a piezoelectric impact stress wave sensor; the drill string monitoring unit comprises a torque sensor and an axial drilling pressure sensor; the acoustic wave acquisition unit comprises a piezoelectric acoustic wave transducer, a differential amplifier and a programmable band-pass filter; the drill bit measurement unit and the acoustic wave acquisition unit are arranged behind the drill bit, and the drill string monitoring unit is arranged at both ends of the probe rod.

[0006] Further, it also comprises a pre-warning module which gives an alarm when the dynamic mechanical parameters obtained by the dynamic mechanical parameter acquisition module exceed the corresponding pre-warning threshold, and the pre-warning threshold is determined based on historical statistics and multi-task weight.

[0007] The rock mass dynamic mechanical parameter sensing method based on drilling parameters comprises: Step 1, constructing training sample data: using the drill bit integrated sensor module to collect drilling parameters and pre-processing them, including synchronous alignment, hybrid noise reduction and feature extraction, and standardizing and feature dimensionality reduction of the extracted features; obtaining the mechanical parameters corresponding to the drilling parameters through laboratory core true values and finite element simulation, and constructing training sample data based on the pre-processed data and the corresponding mechanical parameters; Step 2, constructing a rock mass dynamic mechanical parameter acquisition model based on a multi-modal adaptive graph attention time sequence network and training it based on the training sample data; Step 3, using the drill bit integrated sensor module to collect drilling parameters in real time, and using the same pre-processing method as the training sample to pre-process them, and obtaining the mechanical parameters based on the pre-processed data and the trained rock mass dynamic mechanical parameter acquisition model.

[0008] Further, the mechanical parameters are: instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle.

[0009] Further, the pre-processing step is specifically: Aligning the time scales of all sensor channels uniformly, real-time dividing the signal stream with M window length and S step, and processing each window independently; Performing band-pass filtering of different frequencies according to the data characteristics; Performing ensemble empirical mode decomposition on the filtered signal, selecting the first N IMF components for reconstruction; performing singular value decomposition on the reconstructed signal, and retaining the principal components whose singular value energy proportion is more than Y%; Dynamically estimating the noise spectrum of the acoustic wave signal, adaptively reducing the noise energy according to the frequency band, and retaining the rock mass fracture characteristic spectrum line; Parallelly extracting multiple features of the windowed signal, including: Time domain features: calculating the mean, variance, peak value, RMS energy, kurtosis, skewness, instantaneous mechanical specific energy and drilling power in the window; Frequency domain features: power spectral density is extracted by fast Fourier transform, and the dominant frequency, 1-5 kHz band energy ratio and spectral kurtosis are calculated; the resonance peak frequency is identified by cepstrum analysis; Time-frequency domain features: short-time Fourier transform is performed on the band-passed signal to generate time-frequency spectrum, and energy entropy and spectral centroid trajectory are extracted; multi-scale sub-band coefficients are obtained by wavelet packet decomposition, and the energy, information entropy and frequency band variability of each sub-band are calculated; Dynamic event labeling: a mutation point detection algorithm is used to detect torque mutations, label the stuck / drilling resistance event time, and embed [0,1] event flag in the feature vector; Standardize the extracted features: Z-score normalization is used for time domain features, and Min-Max scaling is used for frequency domain features to [0,1]; PCA dimensionality reduction is performed on time domain, frequency domain and time-frequency domain features respectively.

[0010] Further, the rock mass dynamic mechanical parameter acquisition model comprises: A sensor topology module for constructing a spatial correlation graph structure of sensor data; A graph attention layer module for calculating attention coefficients based on the spatial correlation graph structure and aggregating features based on the attention coefficients; A time series convolution graph module for multi-scale extraction of time series dimension based on the sequence features output by the graph attention layer module, and for enhancing feature nonlinear expression and local feature learning accuracy through causal dilated convolution group and lightweight regression head based on local supervision optimization training process; A time-aware Transformer enhancement module: based on the long-term and short-term dependence relationship modeling of time series features, relative position encoding is introduced through multi-head self-attention mechanism, and attention weight is adjusted based on exponential decay factor to enhance the simultaneous modeling capability of sudden events and long-term trends; A multi-task regression head module: based on the final feature sequence, multiple target parameters are predicted in parallel.

[0011] Further, the rock mass dynamic mechanical parameter acquisition model adopts hierarchical deployment, and the sensor topology module, the graph attention layer module and the time series convolution graph module are deployed on the downhole edge computing node, and the time-aware Transformer enhancement module and the multi-task regression head module are deployed on the ground server.

[0012] Further, a composite loss function L is used in the model training process: Where y is the true value; is the predicted value output by the model; is the weight; symbol is the L1 norm, , symbol is the L2 norm, , denotes the predicted value of the i-th physical quantity in the final output vector for the current sample, is the true value of the i-th physical quantity in the final output vector for the current sample, when i = 1, …, 5, corresponding to instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle respectively; wherein denotes the predicted value of the i-th physical quantity in the final output vector for the current sample, is the true value of the i-th physical quantity in the final output vector for the current sample, when i = 1, …, 5, corresponding to instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle respectively; wherein is the true value of the i-th physical quantity in the final output vector for the current sample, when i = 1, …, 5, corresponding to instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle respectively; , is the number of regression heads; is the model parameter.

[0013] Further, it further comprises step 4, when the dynamic mechanical parameter acquisition module obtains the mechanical parameter exceeding the corresponding early warning threshold, an alarm is given, and the early warning threshold is determined based on historical statistics and multi-task weight.

[0014] The present application has the beneficial effects compared with the prior art: (1) Multi-source while-drilling parameter to multi-target mechanical parameter joint prediction: using vibration, acoustic emission, torque, drilling pressure and other multi-modal signals, through unified feature engineering and multi-task learning, instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle can be dynamically predicted simultaneously in a single drilling, which breaks through the limitation of traditional methods that only estimate a single static parameter.

[0015] (2) Self-adaptive graph attention time sequence network: the network structure of "trainable edge weight + GATv2" is proposed, which effectively couples the physical proximity of sensors and statistical correlation, and integrates dilated time convolution and time perception Transformer module, and under the premise of low time delay, the representation ability of sudden events (such as sticking / drilling resistance) and long-term trend is considered, thereby improving the prediction accuracy of the model.

[0016] (3) Engineering usability: hierarchical deployment (downhole edge + ground server), online adaptive fine-tuning, dynamic threshold early warning and explainability analysis, which significantly improves real-time performance, stability and usability.

[0017] Therefore, the present application can realize in-situ, real-time, dynamic and multi-parameter joint perception under the condition of non-stop drilling, and is suitable for safe and efficient construction and risk warning of deep complex strata. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a rock dynamic mechanical parameter perception system structure diagram based on while-drilling parameters; Figure 2This is a schematic diagram of the probe structure; Figure 3 This is a cross-sectional view of the probe. Figure 4 The flowchart shows the method for sensing dynamic mechanical parameters of rock mass based on drilling parameters. Reference numerals in the attached figures: 1 is the probe body, 2 is the sealing cover plate, 3 is the triaxial MEMS accelerometer, 4 is the piezoelectric impact stress wave sensor, 5 is the annular strain torque sensor, 6 is the piezoelectric ceramic axial drilling pressure sensor, 7 is the piezoelectric acoustic transducer, 8 is the temperature sensor, 9 is the piezoresistive pressure sensor, and 10 is the turbine flow meter. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] This method achieves real-time, dynamic, and joint inversion of multiple mechanical parameters during drilling by using multimodal signal fusion and an adaptive time-series graph neural network architecture, significantly improving the intelligence level of geological parameter acquisition during drilling.

[0021] like Figure 1 As shown, the rock mass dynamic mechanical parameter sensing system based on drilling parameters includes: The drill bit integrated sensor module includes a probe body, a drill bit measurement unit for acquiring the vibration spectrum of drilled rock, a drill string monitoring unit for monitoring mechanical load, and an acoustic wave acquisition unit for acquiring sound waves. The drill bit measurement unit, drill string monitoring unit, and acoustic wave acquisition unit are all mounted on the probe body.

[0022] Specifically, such as Figure 2 , 3 As shown, the probe body 1 is a cylindrical rod made of high-strength alloy steel, with sensor mounting holes and slots and wiring channels inside, and a flow-through hole in the annulus. Threads for connection to the drill bit and drill rod are provided at both ends. Axial slots are opened on the sidewalls of the rod body, and the sealing cover plate 2 is connected by bolts to achieve modular sensor installation and waterproofing. A drill bit measurement unit, a drill string monitoring unit, and an acoustic wave acquisition unit are installed inside the probe body, installed in corresponding holes and slots, and fixed with bolts or glue. Specifically, the drill bit measurement unit embeds a triaxial MEMS accelerometer 3 and a piezoelectric impact stress wave sensor 4 behind the drill bit to directly capture the rock vibration spectrum. The drill string monitoring unit installs one set of annular strain gauge torque sensors 5 and a piezoelectric ceramic axial drill pressure sensor 6 at each end of the rod body to monitor mechanical load in real time. The acoustic wave acquisition unit arranges four piezoelectric acoustic transducers 7 near the drill bit, connected to a differential amplifier and a programmable bandpass filter.

[0023] In addition, the probe body also houses an environmental monitoring unit: integrating a PT100 temperature sensor 8, a piezoresistive pressure sensor 9, and a turbine flow meter 10, used to acquire temperature T, pressure p, and flow rate Q in real time, and as online calibration values ​​for each measurement channel to suppress the influence of temperature drift and drilling fluid on the measurement results. The specific calibration method is as follows: (1) Temperature drift compensation: A cubic polynomial compensation model with temperature as the independent variable is established for the channels of acceleration, annular strain torque and piezoelectric drilling pressure. ,in, : Measurement value after temperature drift compensation (unit consistent with this channel, such as g, N·m, kN, Pa); Original output value (units as above); coefficient (k=0,1,2,3): Temperature compensation polynomial coefficients (matched to the channel dimensions; obtained by segmented calibration in the constant temperature chamber and written into the sensor's EEPROM (Electrically Erasable Programmable Read-Only Memory)).

[0024] (2) Pressure / flow calibration and temperature-viscosity correction: The pressure channel is calibrated using a combination of "two-point zero / full-scale + multi-point linear fitting"; the turbine flow meter is calibrated at 6 points using the volumetric method to obtain a nonlinear compensation table, which is then combined with temperature-viscosity / density correction. , : Compensated volumetric flow rate ( (or L / min); : Raw reading of turbine flow meter (same unit as above); , Temperature-viscosity / density correction coefficient (obtained through regression, dimensionless); Dynamic viscosity of drilling fluid at temperature T (Experimental / field measurement or model estimation); Reference temperature viscosity at the bottom ( ); Drilling fluid density at temperature T and pressure p ( ); Reference state ( Density at ) p: Annular / channel pressure (Pa or MPa).

[0025] (3) Correction for drilling fluid interference (liquid damping / sound absorption): Estimate drilling fluid viscosity based on T and p. With density Calculate bandwidth attenuation using an empirical model Applying equivalent inverse filter to shock / acoustic channel in frequency domain , and viscosity normalization correction to MSE, power, etc. MSE (mechanical specific energy): MSE = (WOB x RPM) / ROP; where, : equivalent inverse filter transfer function (dimensionless, to compensate for medium attenuation); f: frequency (Hz, from frequency variable after FFT to signal); : band-dependent amplitude attenuation coefficient (Np / m; calculated from empirical / semi-empirical model, varies with viscosity μ, density ρ); L: equivalent propagation path length of acoustic / shock signal in medium (m); WOB: weight on bit (kN), from axial WOB sensor; RPM: rotational speed (r / min or r / s); ROP: rate of penetration (m / s or m / h).

[0026] Power: ; : torque (N-m, from ring strain type torque sensor); : angular velocity (rad / s, converted from rotational speed).

[0027] (4) Channel cross-sensitivity elimination: construct cross-sensitivity matrix , solve by online least squares: , : real channel vector after decoupling (each component unit varies with channel); : measured multi-channel vector (original / pre-processed data). I: identity matrix, dimension N x N (N is the number of channels). K: cross-sensitivity matrix (dimensionless); element k_ij represents the cross-talk weight of channel j to channel i (usually k_ii = 0).

[0028] (5) Downhole in-situ zero point / drift suppression: automatically reset zero bias and update compensation coefficients for current temperature section in the tripping static window.

[0029] After the above correction, the temperature drift suppression is better than 0.5% FS / 10°C, the amplitude error caused by drilling fluid is less than 2%, and the long-term zero point drift is less than 0.2% FS.

[0030] Synchronization and calibration: use dual-redundant clock sources (downhole OCXO constant temperature crystal oscillator + GPS tamed clock), time synchronization error < 1 μs; before drilling, automatically calibrate the frequency response characteristics of each sensor through a standard exciter. Centralized device: built-in lithium sub-battery pack (endurance 72 h) and RS-485 / CAN bus interface, transmit the matched impedance of multiple analog signals to the ground system.

[0031] The multi-modal signals such as vibration, acoustic emission, torque, and drilling pressure provide complementary observation perspectives, and the prediction accuracy, reliability, and stability are improved by cross-validation and fusion of the information captured at different physical levels (mechanical, acoustic, and hydraulic).

[0032] The data preprocessing module is configured to preprocess the original data collected by the drill bit integrated sensor module, including synchronization alignment, hybrid noise reduction, and feature extraction, and to standardize and reduce the dimension of the extracted features.

[0033] The rock mass dynamic mechanical parameter acquisition model module is configured to construct a rock mass dynamic mechanical parameter acquisition model based on a multi-modal adaptive graph attention time sequence network and to train the model.

[0034] The dynamic mechanical parameter acquisition module is configured to acquire the mechanical parameters based on the preprocessed data and the trained rock mass dynamic mechanical parameter acquisition model.

[0035] Correspondingly, as shown in Figure 4 The rock mass dynamic mechanical parameter sensing method based on the drilling parameters comprises the following steps: Step 1: Constructing training sample data: Collecting drilling parameters by using a drill bit integrated sensor module and preprocessing the drilling parameters, including synchronization alignment, hybrid noise reduction, and feature extraction, and standardizing and reducing the dimension of the extracted features; acquiring the mechanical parameters corresponding to the drilling parameters through laboratory core true values and finite element simulation, and constructing training sample data based on the preprocessed data and the corresponding mechanical parameters.

[0036] The preprocessing process specifically comprises the following steps: 1. Data synchronization alignment Align all sensor channel time stamps based on the GPS / PTP protocol to ensure time synchronization accuracy. Divide the signal stream in real time with a window length M and a step S, and process each window independently. The selection of the window length M and the step S is based on the frequency characteristics and the sampling rate of the signal, for example, 1.0 s and 0.1 s, respectively.

[0037] 2. Hybrid noise reduction Perform band-pass filtering of different frequencies according to the data characteristics. For example, apply 0-1 kHz band-pass filtering to mechanical parameters such as drilling pressure and rotational speed; apply 0.5-20 kHz band-pass filtering to vibration and acoustic data; and solve low-frequency drift and high-frequency interference filtering.

[0038] Empirical mode decomposition (EEMD)-singular value decomposition (SVD) reconstruction: Perform ensemble empirical mode decomposition on the filtered signal, select the first N (set to 5 in this embodiment) IMF components for reconstruction; perform singular value decomposition (SVD) on the reconstructed signal, and retain the principal components whose singular value energy accounts for more than 95%.

[0039] Adaptive spectral subtraction: dynamically estimate the noise spectrum of the acoustic signal, adaptively reduce the noise energy in the frequency band, and retain the rock mass breaking characteristic spectrum.

[0040] 3. Feature extraction Time domain features: calculate the mean, variance, peak value, RMS energy, kurtosis, skewness in the window; instantaneous mechanical specific energy MSE=(WOB×RPM) / ROP; drilling power .

[0041] Frequency domain features: use fast Fourier transform (FFT) to extract power spectral density (PSD), calculate dominant frequency, 1-5 kHz band energy ratio, spectral kurtosis; identify the resonance peak frequency through cepstrum analysis.

[0042] Time-frequency domain features: perform short-time Fourier transform (STFT) on the band-passed signal to generate a time-frequency spectrum, extract energy entropy and spectral centroid trajectory; apply wavelet packet decomposition (WPD) to obtain multi-scale sub-band coefficients, calculate the energy, information entropy and frequency band variability of each sub-band.

[0043] Dynamic event labeling: use mutation point detection algorithms such as CUSUM algorithm to detect torque mutations, label the stuck / abrasion event time, and embed [0,1] event flag bits in the feature vector.

[0044] 4. Feature standardization and dimensionality reduction The extracted multi-modal features are normalized in time domain (mean=0, variance=1) and scaled to [0,1] in frequency domain.

[0045] Feature dimensionality reduction: PCA dimensionality reduction (cumulative variance>95%) is performed on time domain, frequency domain and time-frequency domain features, respectively, and in this embodiment, it is spliced into a 120-dimensional feature vector.

[0046] In addition, in order to increase the amount of data, this embodiment also carries out sample enhancement, through the addition of Gaussian noise, random scaling, time shift, etc., the data set is expanded to 3 times the original. The expanded data set is divided into 70% training set, 15% validation set and 15% test set.

[0047] Step 2, construct a rock mass dynamic mechanical parameter acquisition model based on a multi-modal adaptive graph attention time sequence network and train based on training sample data.

[0048] The rock mass dynamic mechanical parameter acquisition model is composed of five core functional modules connected in turn, and each module cooperates to complete the whole process from the input of the preprocessed sensor signal to the output of the target parameter, which specifically includes: Sensor topology module: This module is used to construct the spatial correlation graph structure of sensor data, providing the basic topology for subsequent graph attention calculation. The pre-processed signals collected by each sensor are taken as the nodes of the graph; the initial value of the edge weight between nodes is generated by weighted calculation, and the edge weight is set as a trainable parameter, which is dynamically optimized in the network backpropagation process to accurately capture the real correlation strength between sensors. The edge weight calculation formula is as follows: wherein is the Euclidean distance between sensors, is mutual information; is the distance scale; is a numerical stability constant; , are the feature vectors of the first , the second sensor in the current time window obtained by preprocessing; , is the feature vector of any two sensors in the same time window, which is used to calculate the maximum value of mutual information of all channel pairs to realize normalization.

[0049] Graph attention layer module: Based on the topology graph structure, the spatial correlation features of the nodes are extracted, and the influence of the key nodes is highlighted through the graph attention layer module. The graph attention layer module adopts 2-layer GATv2, and 4 attention heads are set in each layer to realize multi-scale feature aggregation; node feature aggregation is completed through attention coefficient weighting, wherein the attention coefficient is combined with the trainable edge weight calculation, so that the network can adaptively focus on the sensor nodes that are more important to the target prediction; the node feature aggregation formula is: wherein the attention coefficient wherein: a is the attention weight vector (learnable parameter); represents the feature vector of node before entering this layer of GAT; : updated node feature (output vector); : nonlinear activation function (such as ReLU / ELU / LeakyReLU), which increases expressiveness and stability; : sum over neighbor set (usually including the self-loop itself, so as to retain its own information and stabilize training). ∈[0,1]: attention weight (scalar) from neighbor to . For fixed , these weights are softmax normalized over neighbors, satisfying , learned by the scoring network, reflecting the "neighbor to Importance of the : Linear mapping matrix (learnable) that projects neighbor features from input dimension to output dimension ; : Input features of neighbor nodes . A "node" is a feature vector (e.g. 120-dimension after pre-processing) of a sensor / channel at the current time window; Trainable edge weight.

[0050] Temporal Convolutional Graph Module (TCGM): Multi-scale extraction of temporal dimension on sequence features output by graph attention, while introducing local supervision optimization training process. The core is the causal dilated convolution group: input the graph attention output sequence of each node into 4 parallel convolution channels, each channel with dilated rate of 1, 2, 4, 8, and expand the temporal receptive field through dilated convolution; After the parallel channel outputs are spliced, they are processed by LayerNorm standardization and ReLU activation function to enhance the nonlinear expression of features; A lightweight regression head (2-layer fully connected network) is connected after each TCGM unit to calculate the local MAE loss through deep supervision, which helps the network converge early and improves the accuracy of local feature learning.

[0051] Time-aware Transformer Enhancement Module: Further modeling of long-term and short-term dependencies of temporal features to enhance the ability to capture sudden events and long-term trends simultaneously. Receive the feature sequence output by TCGM and input into a lightweight TransformerEncoder; Introduce relative position encoding in multi-head self-attention to explicitly represent the temporal position relationship; Multiply an exponential decay factor in the attention weight calculation to enhance the ability to model sudden events and long-term trends simultaneously, where and represent the timestamps (or sampling time) corresponding to the th time step and the th time step in the feature sequence, respectively, in seconds (s) or milliseconds (ms), obtained by accumulating the sampling period .

[0052] Multi-task Regression Head Module: Realize parallel prediction of multiple target parameters based on the final feature sequence, and improve the overall prediction accuracy through joint loss optimization. Contains 5 independent fully connected regression branches, respectively corresponding to instantaneous elastic modulus E, dynamic Poisson's ratio , transient uniaxial compressive strength , cohesion C and internal friction angle .

[0053] An adaptive graph attention temporal network is constructed to dynamically capture the physical association and statistical dependence between sensor nodes by using trainable edge weights and GATv2, and to realize multi-scale feature representation of drilling time series signals by combining the local feature extraction capability of dilated convolution and the long-term context modeling of time-aware Transformer, thereby improving the prediction accuracy of the model.

[0054] When training the model, the loss function adopts a composite loss function L, wherein: is the true value (given by core experiment / simulation); is the output mechanical parameter result, , , , , are the predicted values of the instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle of the rock mass respectively; is the weight; is the model parameter; is the mean absolute error of the th intermediate regression head (or local output), defined as: wherein represents the prediction value of the th regression head for the th sample; n is the number of samples in the batch; is the true value of the th sample on the th regression head, is the set number of intermediate supervision / regression heads (for example, the local outputs of each dilated branch or level of TCGM). The composite loss function L is defined as:The prediction uncertainty is estimated by Monte Carlo dropout (p=0.2) and plays a regularization effect; the optimization strategy adopts the AdamW optimizer, combined with linear warm-up and cosine annealing learning rate scheduling; the label smoothing, stochastic depth and spectral jitter techniques are applied at the network level. The AdamW optimizer improves the generalization ability of the model by more reasonably processing weight decay; the linear warm-up stage starts from a small learning rate and gradually increases to the preset maximum value to help the model converge stably. Subsequently, the cosine annealing strategy gradually reduces the learning rate in the form of a cosine function, which helps the model jump out of the local optimal solution and explore a wider parameter space; label smoothing: by smoothing the label distribution, the model's excessive confidence in the training data is reduced, thereby improving the generalization ability; stochastic depth: by randomly discarding certain layers in the network during training, the diversity of the model is increased to prevent overfitting; spectral jitter: by adding random noise in the frequency domain to enhance the data, the robustness of the model to different frequency changes is improved.

[0055] Step 3, real-time acquisition of drilling parameters using drill bit integrated sensor module, and the same preprocessing method as the training sample is used for preprocessing, and the mechanical parameters are obtained based on the preprocessed data and the trained rock mass dynamic mechanical parameters.

[0056] The preprocessed data is sequentially input into: sensor topology graph module→graph attention layer module (GATv2)→time series convolution graph module (TCGM)→time-aware Transformer enhancement module→multi-task regression head, output parameters: instantaneous elastic modulus E, dynamic Poisson's ratio , transient uniaxial compressive strength , cohesion C and internal friction angle In this embodiment, the 1s sliding window and 0.1s step rolling input data sequence are used to update the parameter prediction results in real time.

[0057] To improve processing efficiency, this embodiment uses downhole edge + ground server hierarchical deployment: sensor topology graph module, graph attention layer module (GATv2) and TCGM are deployed on downhole edge computing nodes (GPU / FPGA) to ensure that the single-step processing delay is ≤20 ms; the time-aware Transformer enhancement module and the multi-task regression head are deployed on the ground high-performance server to undertake reading buffer, centralized inference and visualization tasks.

[0058] To improve prediction accuracy, an online adaptive update was implemented: the prediction residuals were continuously monitored, and the error trend was tracked using an exponentially weighted moving average (EWMA); when the residuals exceeded a preset threshold, online fine-tuning was triggered only for the Transformer enhancement layer or the regression head; and field test data was periodically imported to perform incremental retraining, thereby improving the model's adaptability and stability to new geological formations.

[0059] In addition, based on the model prediction results, this embodiment also provides a visualization and early warning module: the ground platform simultaneously supports two-dimensional curves, heat maps and three-dimensional geological models, which can be interactively zoomed and scrolled over time; dynamic threshold early warning: the early warning threshold is adaptively updated based on historical statistics and multi-task weights, and an audible and visual alarm is triggered immediately when the parameter exceeds the limit; interpretability analysis: it integrates Attention Rollout, Grad-CAM and GAT edge weight visualization to provide engineers with a visual explanation of input features and model decisions.

[0060] The specific method for adaptively updating the early warning threshold based on historical statistics and multi-task weights is as follows: a sliding window robust statistics method is used to give the "center + scale" of each parameter, and then the task weight is formed by "priority ÷ uncertainty". The threshold bandwidth is adaptively scaled according to the weight and updated over time.

[0061] "Center" indicates the typical level of this parameter under the current operating conditions, using an exponentially weighted moving average (EWMA): ,in The predicted value at the current moment. The center value at the current moment. The center value calculated at the previous moment. This is the smoothing coefficient.

[0062] “Scale” represents the fluctuation range, expressed as EWMA absolute deviation: , The fluctuation scale at the current moment. The fluctuation scale calculated at the previous moment. This is the smoothing coefficient.

[0063] The "priority" is calculated based on the event score S according to "deviation degree × duration × importance of the condition", and divided into three levels: Calculation of standardized deviation: Introducing uncertainty reduction: The duration of continuous exceedance of the threshold is τ (seconds), where w_p is the business weight set for different monitoring parameters / different operating conditions (different parameters can be different, default is 1), and the score is: Grading rules (can be calibrated on-site): Level 1 (P1, high-level alarm): Event score S≥12 or continuous over-limit time τ≥5 s; Level 2 (P2, medium-level alarm): Event score... ; Level three (P3, low-level alarm): event score . To prevent jitter, a 10% hysteresis and a minimum duration of 1s strategy is adopted.

[0064] “Uncertainty” is the prediction standard deviation from MC-Dropout (or square root of variance), the normalized and truncated uncertainty coefficient is . Where u_ref is the empirical reference uncertainty (e.g., the average of the training set or historical normal operating conditions), u_max is the upper limit of uncertainty truncation (default 2.0), to avoid extreme uncertainty to pull the threshold bandwidth too large.

[0065] The upper and lower thresholds of the adaptive threshold are calculated as follows, and are updated rolling with time: 、 , where k is the alarm sensitivity (k=2 for early warning, k=3 for out-of-limit), κ is the uncertainty amplification coefficient (recommended 0.5).

Claims

1. A rock mass dynamic mechanical parameter perception system based on drilling parameters, characterized in that, Comprise: Drill bit integrated sensor module: including probe rod body, drill bit measurement unit for acquiring rock drilling vibration frequency spectrum, drill string monitoring unit for monitoring mechanical load and acoustic wave acquisition unit for collecting acoustic waves, drill bit measurement unit, drill string monitoring unit and acoustic wave acquisition unit are arranged on the probe rod body; Data preprocessing module: for preprocessing the original data collected by the drill bit integrated sensor module, including synchronization alignment, mixed noise reduction and feature extraction, and standardizing and feature dimensionality reduction of the extracted features; Rock mass dynamic mechanical parameter acquisition model module: based on a multi-modal adaptive graph attention time series network, a rock mass dynamic mechanical parameter acquisition model is constructed and trained; Dynamic mechanical parameter acquisition module: based on the preprocessed data and the trained rock mass dynamic mechanical parameter acquisition model, the mechanical parameters are acquired.

2. The system for sensing the dynamic mechanical parameters of rock mass based on the while-drilling parameters according to claim 1, characterized in that, The drill bit measurement unit comprises a three-axis MEMS acceleration sensor and a piezoelectric impact stress wave sensor; the drill string monitoring unit comprises a torque sensor and an axial drilling pressure sensor; the acoustic wave acquisition unit comprises a piezoelectric acoustic wave transducer, a differential amplifier and a programmable band-pass filter; the drill bit measurement unit and the acoustic wave acquisition unit are arranged behind the drill bit, and the drill string monitoring unit is arranged at both ends of the probe rod.

3. The system for sensing rock mass dynamic mechanical parameters based on drilling parameters according to claim 1, characterized in that, It also includes a warning module, which alarms when the mechanical parameters acquired by the dynamic mechanical parameter acquisition module exceed the corresponding warning threshold, which is determined based on historical statistics and multi-task weights.

4. A rock mass dynamic mechanical parameter sensing method based on a drilling parameter, characterized in that, Comprise: Step 1, construct training sample data: adopt the drill bit integrated sensor module to collect the while-drilling parameters and preprocess them, including synchronization alignment, mixed noise reduction and feature extraction, and standardize and feature dimensionality reduction of the extracted features; acquire the mechanical parameters corresponding to the while-drilling parameters through laboratory core true value and finite element simulation, and construct the training sample data based on the preprocessed data and the corresponding mechanical parameters; Step 2, based on a multi-modal adaptive graph attention time series network, a rock mass dynamic mechanical parameter acquisition model is constructed and trained based on the training sample data; Step 3, real-time collect the while-drilling parameters by the drill bit integrated sensor module, and preprocess them by the same preprocessing method as the training sample, and acquire the mechanical parameters based on the preprocessed data and the trained rock mass dynamic mechanical parameter acquisition model.

5. The method of claim 4, wherein, The mechanical parameters are: instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion and internal friction angle.

6. The method of claim 4, wherein, The preprocessing steps are specifically: Align the time scales of all sensor channels uniformly, real-time split the signal stream with M window length and S step, and process each window independently; According to the data characteristics, band-pass filter different frequencies; Perform ensemble empirical mode decomposition on the filtered signal, select the first N IMF components for reconstruction, and perform singular value decomposition on the reconstructed signal, retaining the principal components with singular value energy ratio above Y%; Dynamically estimate the noise spectrum of the acoustic signal, adaptively reduce the noise energy according to the frequency band, and retain the rock mass fracture characteristic spectrum; Parallelly extract multiple features of the windowed signal, including: Time domain features: calculate the mean, variance, peak value, RMS energy, kurtosis, skewness, instantaneous mechanical specific energy and drilling power in the window; Frequency domain features: power spectral density is extracted by fast Fourier transform, and dominant frequency, 1-5 kHz band energy ratio and spectral kurtosis are calculated; resonance peak frequency is identified by cepstrum analysis; Time-frequency domain features: short-time Fourier transform is performed on the band-passed signal to generate time-frequency spectrum, and energy entropy and spectral centroid trajectory are extracted; multi-scale sub-band coefficients are obtained by wavelet packet decomposition, and the energy, information entropy and frequency band variability of each sub-band are calculated; Dynamic event labeling: a mutation point detection algorithm is used to detect torque mutations, label the time of sticking / drilling and grinding resistance events, and embed [0,1] event flag in the feature vector; Standardization processing is performed on the extracted features: Z-score normalization is used for time domain features, and Min-Max scaling is used for frequency domain features to [0,1]; PCA dimensionality reduction is performed on time domain, frequency domain and time-frequency domain features respectively.

7. The method of claim 4, wherein, The rock mass dynamic mechanical parameter acquisition model includes: Sensor topology graph module, used to construct the spatial correlation graph structure of sensor data; Graph attention layer module, based on the spatial correlation graph structure to calculate the attention coefficient, and based on the attention coefficient to aggregate features; Time series convolution graph module, based on the sequence features output by the graph attention layer module to perform multi-scale extraction in the time series dimension, and based on local supervision to optimize the training process, through the causal inflation convolution group and the lightweight regression head to enhance the non-linear expression of features and the accuracy of local feature learning; Time-aware Transformer enhancement module: based on the long-term and short-term dependence relationship modeling of time series features, through the multi-head self-attention mechanism to introduce relative position encoding, and based on the exponential decay factor to adjust the attention weight, to enhance the simultaneous modeling ability of sudden events and long-term trends; Multi-task regression head module: based on the final feature sequence to realize parallel prediction of multiple target parameters.

8. The method according to claim 7, wherein, The rock mass dynamic mechanical parameter acquisition model adopts hierarchical deployment, the sensor topology graph module, the graph attention layer module and the time series convolution graph module are deployed on the downhole edge computing node, and the time-aware Transformer enhancement module and the multi-task regression head module are deployed on the ground server.

9. The method of claim 7, wherein, The model training process uses a composite loss function L: , where y is the truth value; The predicted value output by the model; Weights; symbol It is an L1 norm. ,symbol It is the L2 norm. , This refers to the predicted value of the i-th physical quantity in the final output vector for the current sample. The true value of the i-th physical quantity in the final output vector for the current sample is given. When i=1,…,5, these correspond to the instantaneous elastic modulus, dynamic Poisson's ratio, transient uniaxial compressive strength, cohesion, and internal friction angle, respectively. ,in Indicates the first The predicted value of the regression head for the s-th sample, For the s-th sample in the th case The true values ​​on the regression heads, where n is the number of samples. , This represents the number of regression heads; These are the model parameters.

10. The method of claim 4, wherein, It also includes step 4, when the mechanical parameters obtained by the dynamic mechanical parameter acquisition module exceed the corresponding warning threshold, an alarm is given, and the warning threshold is determined based on historical statistics and multi-task weights.

Images