Tunnel geological identification online self-learning method based on while-drilling parameters

Abstract

This invention relates to the field of geological information processing and discloses an online self-learning method for tunnel geological identification based on drilling parameters. The method includes: real-time acquisition and preprocessing of multiple types of drilling parameters to generate test sequences; inputting the sequences into a geological identification model, simultaneously monitoring prediction confidence, temporal rate of change, and phase lag characteristics; activating the self-learning engine when four thresholds are simultaneously met: confidence level, low confidence percentage, integral of temporal rate of change, and phase lag change; extracting high-dimensional features using hook functions, and discovering new geological patterns through topological persistence analysis and geometric verification; freezing the backbone network and dynamically expanding classifier nodes, using hot-start initialization, differentiated learning rate, and lightweight fine-tuning of orthogonal regularization terms; locking and deploying the normalizer parameters. This invention achieves online self-learning and continuous evolution of the tunnel geological identification model, improving the adaptive capability and accuracy of advanced geological prediction.

Description

Technical Field

[0001] This invention relates to the field of geological information processing, and in particular to an online self-learning method for tunnel geological identification based on drilling parameters. Background Technology

[0002] The safety of tunnel and underground engineering construction is highly dependent on the accurate detection of geological conditions ahead of the tunnel face. Traditional forecasting methods, such as advanced horizontal drilling, TSP seismic reflection, and ground-penetrating radar, play an important role in engineering practice, but they generally suffer from bottlenecks such as poor timeliness, high detection costs, limited resolution, and strong subjectivity in data interpretation, making it difficult to meet the urgent needs of modern tunnel excavation for efficient, real-time, and intelligent detection.

[0003] Measurement while drilling (MWD) technology provides a new approach to identifying lithological changes by monitoring the mechanical parameters (such as drilling speed, torque, and feed pressure) of the drilling rig's response to the rock mass in real time. However, existing machine learning applications mostly employ traditional algorithms such as Super Vector Machines (SVM) and Long Short-Term Memory Networks (LSTM), which are insufficient for mining the high-frequency, high-dimensional, and highly nonlinear temporal features of MWD parameters. Furthermore, their applications are often limited to single intensity grading or coarse lithological classification, making it difficult to capture subtle changes under complex geological conditions.

[0004] The core challenge facing current technologies in engineering implementation lies in the conflict between static classification models and the dynamic and variable nature of geological environments. Existing models heavily rely on pre-set discrete labeling systems and methodological indoor experimental data, but there is a significant "domain gap" between indoor simulated materials and real rock masses, and laboratory data cannot exhaustively represent the continuously distributed lithological series in nature. When the model encounters novel lithologies or extreme working conditions at the construction site without prior training, the recognition accuracy often drops precipitously. Summary of the Invention

[0005] The purpose of this invention is to propose an online self-learning method for tunnel geological identification based on drilling parameters, which solves the shortcomings of existing static geological identification models, such as weak adaptability under complex working conditions, inability to identify new lithology outside the preset system, and lagging model updates.

[0006] Specifically, this invention provides an online self-learning method, device, and medium for tunnel geological identification based on drilling parameters. The method includes the following steps: S1. Real-time synchronous acquisition of multiple core drilling parameters during tunnel drilling, and preprocessing of the raw drilling parameters to generate standardized test sequence fragments; S2. Input the sequence fragment to be tested into the pre-trained geological identification model, and simultaneously monitor the prediction confidence, the time-series change rate of the drilling parameters, and the dynamic phase lag characteristics between at least two different parameters; when the prediction confidence is lower than the first threshold, the proportion of low-confidence samples exceeds the second threshold, the cumulative integral of the time-series change rate exceeds the third threshold, and the change in the dynamic phase lag characteristics exceeds the fourth threshold, activate the online self-learning engine. S3. In response to the activation of the online self-learning engine, the high-dimensional feature vector of the intermediate feature representation layer of the geological identification model is extracted using the hook function. The high-dimensional feature vector is then subjected to multi-scale clustering verification through topological persistence analysis. Combined with geometric feature verification, it is determined whether there are new feature clusters different from the known categories. If they exist, it is confirmed that a new geological pattern has been discovered. S4. Once a new geological pattern is confirmed, freeze the weights of the backbone network of the geological identification model, dynamically expand the neuron nodes corresponding to the new category in the output layer of its classifier, initialize the weights of the newly added nodes using a hot start method, set differentiated learning rates for the newly added nodes and the original nodes with different decay curves, add an orthogonal regularization term to the loss function, and perform lightweight fine-tuning using field verification data to obtain the evolved geological identification model. S5. Lock and apply the global mean and standard deviation of the normalizer during the initial training phase, perform performance backtesting on the evolved geological identification model, and save and deploy the evolved model after passing the backtest.

[0007] A storage medium storing instructions and data for implementing an online self-learning method for tunnel geological identification based on drilling parameters.

[0008] An online self-learning device for tunnel geological identification based on drilling parameters includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement an online self-learning method for tunnel geological identification based on drilling parameters.

[0009] The beneficial effects provided by this invention are as follows: First, by triggering online learning through temporal matching of low-confidence continuity and drilling response mutation index, false triggering caused by noise and sporadic uncertainties is reduced; Second, by jointly verifying new geological models through density gradient, density valley, normalized separation degree and temporal continuity, the credibility of discovering unknown geological models is improved; Third, by guiding the initialization of new nodes through inter-class confusion and combining it with the stability constraints of old class boundaries, rapid convergence and anti-forgetting are achieved; Fourth, model updates only involve new nodes and a small number of gating adaptation parameters, making it suitable for deployment on tunnel site control computers and meeting real-time construction needs. Attached Figure Description

[0010] Figure 1 This is a simplified flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the dynamic expansion and weight protection mechanism of the output layer of the model in this invention; Figure 3 This is a schematic diagram of the triggering timing of the present invention under typical operating condition transitions; Figure 4 This is a schematic diagram of the hardware device operation according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0012] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0013] Example 1 Please refer to Figure 1 The present invention provides an online self-learning method for tunnel geological identification based on drilling parameters, comprising: S1. Real-time synchronous acquisition of multiple core drilling parameters during tunnel drilling, and preprocessing of the raw drilling parameters to generate standardized test sequence fragments; It should be noted that the multiple core drilling parameters mentioned in step S1 include impact pressure, impact flow rate, rotational speed, rotational pressure, propulsion pressure, drilling stroke, drilling speed, impact frequency, and propulsion force; the preprocessing includes outlier detection using Mahalanobis distance, noise and geological abrupt change separation using discrete wavelet transform, state estimation using Kalman filtering, and downsampling to a preset frequency.

[0014] As one embodiment, in step S1 of the present invention, the multiple types of core drilling parameters include, but are not limited to, nine types of parameters such as impact pressure, impact flow rate, rotational speed, rotational pressure, propulsion pressure, drilling stroke, drilling speed, impact frequency, and propulsion force. These parameters are collected in real time by high-frequency sensors installed on the hydraulic system and rotating mechanism of the drilling rig, with a sampling frequency of not less than 2 kHz.

[0015] Preprocessing specifically includes: firstly, using Mahalanobis distance to detect outliers in the raw data, marking data points whose Mahalanobis distance exceeds a set threshold (e.g., the 95th quantile of the chi-square distribution) as outliers and removing or correcting them; Then, the discrete wavelet transform (DWT) is used to decompose the signal into low-frequency approximate components and high-frequency detail components, where the high-frequency detail components are treated as noise removal and the low-frequency components retain geological change information; then, Kalman filtering is used to perform state estimation on the denoised signal to smooth random fluctuations. Finally, the processed data is downsampled to 20 Hz using linear interpolation or decimation to form a time-continuous standardized sequence segment of the test data. For example, in a tunnel project, the original impact pressure sequence contained impulse noise. After the above processing, a smooth 20 Hz sequence was obtained, with each segment being 5 seconds long (i.e., 100 sampling points).

[0016] In another embodiment, the detailed acquisition and preprocessing process of the nine types of parameters in step S1 of the present invention is as follows: At a TBM tunnel construction site, pressure sensors, flow meters, and speed encoders are installed on the main oil circuit, pilot oil circuit, and rotary drive motor of the drilling rig hydraulic system, respectively, with the sampling frequency uniformly set to 2 kHz. The nine types of raw data streams are collected, aggregated by the on-site PLC, and then transmitted to the industrial control computer.

[0017] The first step is outlier detection: The Mahalanobis distance is calculated independently for each parameter. Using historical normal drilling data as a reference, a threshold is set at the 0.975th quantile of the chi-square distribution (approximately 5.02 when the degrees of freedom are 1). Points with Mahalanobis distances exceeding this threshold are marked as outliers and replaced with linear interpolation. For example, the impact pressure sensor, due to occasional vibration spikes, has a Mahalanobis distance as high as 12, which is corrected to the average of its preceding and following neighborhood values.

[0018] The second step is discrete wavelet transform: using the Daubechies-4 wavelet basis, a 5-level decomposition is performed. The high-frequency detail coefficients of the first and second levels are set to zero to achieve low-pass filtering, and the detail coefficients of the third to fifth levels are subjected to soft thresholding based on an adaptive threshold to preserve abrupt changes. For example, when entering the fracture zone, the low-frequency approximate component of the propulsion pressure exhibits a step change, while high-frequency noise is suppressed.

[0019] The third step is Kalman filtering: A one-dimensional linear state-space model is established for each parameter, and the process noise covariance is calculated. and observation noise covariance The EM algorithm is used for online estimation, and the filtered output is a smooth state estimate.

[0020] The fourth step is downsampling: the 2 kHz data is passed through an anti-aliasing low-pass filter, and then a value is taken every 100 points to obtain a 20 Hz sequence. Finally, 9-dimensional feature vectors are obtained for 100 time points in each 5-second time window, which are used as the sequence segments to be tested.

[0021] S2. Input the sequence fragment to be tested into the pre-trained geological identification model, and simultaneously monitor the prediction confidence, the time-series change rate of the drilling parameters, and the dynamic phase lag characteristics between at least two different parameters; when the prediction confidence is lower than the first threshold, the proportion of low-confidence samples exceeds the second threshold, the cumulative integral of the time-series change rate exceeds the third threshold, and the change in the dynamic phase lag characteristics exceeds the fourth threshold, activate the online self-learning engine. As one embodiment, in step S2 of this invention, the pre-trained geological identification model adopts the MOE-Transformer architecture, with a Softmax classifier as the output layer. The maximum probability value output by the model is the prediction confidence level. The first threshold is set to a relatively high confidence level, such as 0.95, meaning that attention is triggered when the model's confidence in classifying the current sample is below 95%. Simultaneously, the method maintains a buffer with a sliding window length (e.g., 5 seconds, corresponding to 100 samples), and counts the percentage of samples within this window whose confidence level is below the first threshold. When this percentage exceeds a second threshold (e.g., 0.1, or 10%), it indicates that the model has recently encountered a series of uncertain samples.

[0022] Simultaneously, the method calculates the absolute values ​​of the rate of change of drilling speed and propulsion pressure in real time, and performs time integration on both within the same window and then performs a weighted summation (for example, the weights of the important parameters are set to 0.5 and 0.5) to obtain the cumulative integral of the time-series rate of change. When the integral exceeds the third threshold (for example, set to 3 times the average integral under normal tunneling conditions), it indicates that the drilling conditions have fluctuated drastically.

[0023] Furthermore, this invention specifically introduces the dynamic phase lag characteristic between impact pressure and rotational speed: performing Hilbert transform on the two signals respectively to extract the instantaneous phase. and Calculate the instantaneous phase difference Then, the change in the mean phase difference is calculated within a short sliding window (e.g., 0.5 seconds). .when When the rock mass exceeds the fourth threshold (e.g., 5 times the average change in phase difference during normal tunneling), it is considered that the rock mass is transitioning from a brittle fracture mode to a plastic cutting mode.

[0024] The online self-learning engine is activated only when all four conditions are met. For example, in a granite stratum, the model's confidence level for samples of the fault fracture zone ahead drops to 0.85, the proportion of low-confidence samples in the window reaches 15%, and at the same time, the integral of the rate of change of propulsion pressure suddenly increases, and the phase difference between impact pressure and rotational speed reaches 8 times the normal value. At this time, all four criteria are met simultaneously, and the self-learning engine is immediately started.

[0025] As one embodiment, the triggering timing of the four criteria in step S2 of the present invention in actual engineering is as follows: Figure 3 As shown in the figure, the upper curve illustrates the changes in drilling parameters such as ROP and feed pressure over time, depicting the dramatic parameter fluctuations as the model transitions from the granite stage to the fractured zone. The lower curve synchronously displays the real-time response of the model's prediction confidence. When the confidence signal falls below the 0.95 trigger threshold and the proportion of low-confidence samples exceeds 0.1, the method simultaneously detects that the cumulative integral of the time-series rate of change exceeds the limit and the change in dynamic phase lag characteristic exceeds the limit. Since all four criteria are met simultaneously, the method determines that it has encountered an unknown working condition and activates the online self-learning engine.

[0026] It should be noted that in step S2: The first threshold is a preset high confidence threshold, the second threshold is a preset low confidence sample proportion threshold, and the sliding evaluation window length of the low confidence samples is a preset short window length. The cumulative integral value of the time-series rate of change is a weighted sum of the absolute values ​​of the rates of change of drilling speed and propulsion pressure after time integration; The dynamic phase lag characteristic is the instantaneous phase difference between the impact pressure and the rotational speed. It is calculated by performing Hilbert transform on the impact pressure signal and the rotational speed signal respectively, extracting the instantaneous phase, taking the difference and modulus, and then calculating the mean change within a preset short-time sliding window. When the change exceeds the fourth threshold, it reflects the physical precursor of the rock mass changing from brittle fracture to plastic cutting.

[0027] As another embodiment, the specific implementation of the four-fold criterion in step S2 of the present invention is further explained as follows: (1) The first threshold is preferably between 0.90 and 0.99, and the second threshold is preferably between 0.05 and 0.15. In a specific example, the first threshold is 0.95 and the second threshold is 0.10. The sliding evaluation window length is set to 5 seconds, corresponding to 100 consecutive samples at 20 Hz. The method maintains a queue of length 100. Each time the confidence of a new sample is obtained, the queue is updated and the proportion of low-confidence samples is calculated.

[0028] (2) Cumulative integral of the time-series rate of change: Let the drilling speed be... The pressure to advance is First, calculate the absolute value of the rate of change. and Then in the window Inner integral, T Given the window length, perform a weighted sum:

[0029] Among them, the weighting coefficient and The threshold can be set based on experience, such as 0.5 for all values, or dynamically adjusted according to the importance of the parameters. The third threshold can be set to 2 to 4 times the average integral value during normal tunneling. For example, during normal tunneling... The average value is 0.5 when entering the fault fracture zone. The value suddenly rose to 2.3, exceeding the threshold of 2.0, and the condition was met.

[0030] (3) Dynamic phase lag characteristics: for impact pressure signals and rotation speed signal Performing Hilbert transforms on each signal yields an analytic signal with the following instantaneous phase:

[0031] j The imaginary unit, For Hilbert transform; instantaneous phase difference ,in It involves modulo operation; then, the change in mean is calculated within a short window (e.g., 0.5 seconds, corresponding to 10 points):

[0032] The fourth threshold can be set during normal tunneling. Three to six times the average. For example, in homogeneous granite, the phase difference between impact pressure and rotational speed is relatively stable. The mean value is approximately 0.1 rad; upon entering the fracture zone, the rock mass changes from brittle fracture to plastic shearing, and the phase relationship between the two becomes disordered. If the value suddenly increases to 0.6 rad, exceeding the threshold of 0.5, the criterion is activated.

[0033] S3. In response to the activation of the online self-learning engine, the high-dimensional feature vector of the intermediate feature representation layer of the geological identification model is extracted using the hook function. The high-dimensional feature vector is then subjected to multi-scale clustering verification through topological persistence analysis. Combined with geometric feature verification, it is determined whether there are new feature clusters different from the known categories. If they exist, it is confirmed that a new geological pattern has been discovered. As one embodiment, in step S3 of the present invention, a hook function is first registered at the last layer output of the model encoder using the register_forward_hook mechanism in the PyTorch or TensorFlow framework.

[0034] Once the engine is activated, the test sequence segment within the current sliding window is propagated forward again, and the hook function captures the 128-dimensional high-dimensional feature vector output by the encoder. Subsequently, the method performs topological persistence analysis on these feature vectors: at different neighborhood distance thresholds... The set of values ​​(e.g.) ,in Given the standard deviation of the feature space, construct a minimum spanning tree based on the Euclidean distance between feature points, and record the first occurrence of each cluster of feature points. Value (birth side length) ) and when the cluster is merged into other clusters Value (length of the dead side) ), calculate the lifetime of each cluster After normalization, the topological durability score is obtained. .

[0035] Select scores that exceed a preset threshold (e.g., persistence score). Candidate clusters are selected as persistent feature clusters. Then, a triple geometric validation is performed on each persistent feature cluster: the Euclidean distance between its cluster center and the centers of all known categories (e.g., sandstone, granite, mudstone) is calculated; if the minimum distance exceeds a geometric validation threshold (e.g., 30% of the feature space diameter); the average distance from all samples within the cluster to the cluster center is calculated; if it is less than a first compactness threshold (e.g., 0.5 times the feature space standard deviation); and the feature density change rate is calculated radially; if the difference from the density gradient distribution of all known categories exceeds a preset gradient threshold (e.g., Kullback-Leibler divergence greater than 0.5), then the persistent feature cluster is confirmed to correspond to a novel geological model. For example, when a tunnel encounters an alteration zone that has not been pre-trained, the extracted feature vectors form independent clusters, and their persistence scores... The distance from the known granite center is 1.2 times the characteristic spatial diameter, the average distance within the cluster is 0.3 times the standard deviation, and the density gradient distribution is significantly different from the known categories. The method successfully confirmed the discovery of a new geological model, the "alteration zone".

[0036] It should be noted that step S3, which involves performing multi-scale clustering verification on the high-dimensional feature vector through topological persistence analysis and combining it with geometric feature verification, specifically includes: S31. Extract the 128-dimensional feature vector output from the encoder layer in the geological identification model using the hook function; For example, suppose there are 100 samples to be tested in the current sliding window, and after being captured by the hook function, a result is obtained. Feature matrix .

[0037] S32, at different neighborhood distance thresholds Given the set of possible values, a minimum spanning tree is constructed based on the high-dimensional feature vector, recording the first occurrence of each feature point cluster. Value as the birth side length and when the cluster is merged into other clusters Value as the length of the dead edge Calculate the lifetime of each cluster The normalized lifetime is used as the topological durability score for the cluster. ; Specifically, first, L2 normalization is performed on the feature vectors, and then a set of neighborhood distance thresholds is selected. ,in Let be the standard deviation of the distances between all points in the feature space. For each Build a graph Where nodes are sample points, and the distance between edges is less than 1. The point pairs. Then calculate the birth edge length of each connected component (cluster). The minimum value when the cluster first appears. Death side length When this cluster merges with other clusters .life Normalized persistence score ,in , This represents the maximum and minimum lifetimes for all clusters. For example, newly discovered clusters in... Occurring at times, in When it merges with a neighboring cluster, then If the whole , ,but .

[0038] S33, Persistence Score Candidate clusters that exceed a preset persistence threshold are marked as persistent feature clusters; In this invention, a preset persistence threshold is used. Then Clusters are considered persistent feature clusters. In the example above... If not satisfied, it will not be considered. If another cluster... Then proceed to the next verification step.

[0039] S34. For each persistent feature cluster, verify in sequence: whether the Euclidean distance between its cluster center and the center of any known category exceeds the geometric verification threshold, whether the average distance of all samples in the cluster to the cluster center is less than the first compactness threshold, and whether the feature density change rate along the radial direction from the cluster center to the cluster boundary differs from the corresponding density gradient distribution of all known categories by more than a preset gradient threshold. Specifically, for candidate clusters with a persistence score of 0.85 Let its center be Known category set The centers are respectively Calculate the minimum Euclidean distance. The geometric verification threshold is set to 30% of the feature space diameter (for example, if the maximum distance between point pairs in the feature space is 2.5, then the threshold is 0.75). In this example... The first condition is met. Then, all samples within the cluster are calculated to... average distance The first density threshold is set as the standard deviation of the feature space. 0.5 times. If Threshold 0.5, in this example The second condition is satisfied. Finally, along the radial direction from the center to the boundary, the space is divided into several concentric rings. The density of feature points in each ring is statistically analyzed, the rate of change of density with radius (i.e., gradient) is calculated, and the KL divergence is compared with the average density gradient of known categories. If the KL divergence is greater than a preset gradient threshold (e.g., 0.5), the third condition is satisfied. In this example, the KL divergence between the density gradient curve and the gradient distribution of granite is 0.8, satisfying the third condition.

[0040] S35. When all of the above triple geometric verification conditions are met, the discovery of a new geological model is confirmed.

[0041] Overall, all three conditions mentioned above are met, and the method confirms the candidate cluster. A new geological model is identified and temporarily labeled as "Unknown Category-1" for further incremental learning.

[0042] S4. Once a new geological pattern is confirmed, freeze the weights of the backbone network of the geological identification model, dynamically expand the neuron nodes corresponding to the new category in the output layer of its classifier, initialize the weights of the newly added nodes using a hot start method, set differentiated learning rates for the newly added nodes and the original nodes with different decay curves, add an orthogonal regularization term to the loss function, and perform lightweight fine-tuning using field verification data to obtain the evolved geological identification model. In step S4, during the lightweight fine-tuning process, known class prototype vectors, old model output distributions, or a small number of historical samples are used as stabilizing constraints to ensure that the change in the old class output boundary during the learning process of new nodes is less than the preset boundary drift threshold.

[0043] As one embodiment, in step S4 of the present invention, after confirming the new geological model, all trainable parameters of the MOE-Transformer backbone network (including the embedding layer, encoder layer, and hybrid expert layer) are first set to requires_grad=False to achieve weight freezing.

[0044] Then, based on the original K nodes (corresponding to K known categories) in the output layer of the classifier, a new neuron node is added, and its output dimension becomes K+1.

[0045] like Figure 2 As shown on the left, the weight protection mechanism ensures that the weight parameters of the initial classification layer and known categories are locked and protected, preventing catastrophic forgetting of existing knowledge. Figure 2 As shown on the right, the initial classification layer nodes contain known geological labels such as sandstone, granite, and mudstone. When a new geological pattern (such as an alteration zone) is detected, the method uses dynamic node expansion technology to add a new neuron node (labeled as "alteration zone") while keeping the original node weights unchanged, and assigns it independent new node weight parameters. Figure 2 The solid boxes in the diagram represent frozen weights, and the dashed boxes represent newly added trainable nodes.

[0046] The weights of newly added nodes are initialized using a warm-start approach: the average feature vector of multiple low-confidence samples (e.g., all low-confidence samples within the last minute) that triggered the self-learning engine is calculated in the high-dimensional feature space. Obtain the original mapping matrix of the classifier. (in For feature dimension, (The original number of categories), calculate its pseudo-inverse. Then through Obtain the initial weights of the newly added nodes.

[0047] The learning rate is set using a differentiated strategy: a higher initial learning rate (e.g., 0.01) is set for new nodes, while a lower initial learning rate (e.g., 0.0001) is set for existing nodes; furthermore, the learning rate varies with each fine-tuning epoch according to different decay curves, with the learning rate for new nodes decaying exponentially. ,in Smaller (e.g., 0.95) learning rates decay faster; the learning rate of existing nodes is adjusted according to... ,in A larger value (e.g., 0.99) results in slower decay. Simultaneously, an orthogonal regularization term is added to the loss function. ,in For the newly added node weight vector, This is the original node weight matrix. is the regularization coefficient (e.g., 0.01). This regularization term encourages the feature representations of newly added nodes to be orthogonal to the feature spaces of the original categories, thereby reducing feature confusion.

[0048] Finally, lightweight fine-tuning was performed over 5 epochs using validation data collected on-site (e.g., 10 typical new category samples), with the Adam optimizer and a batch size of 4. After fine-tuning, the model gained the ability to identify the new category "alteration band" while maintaining the recognition accuracy of the original categories.

[0049] It should be noted that the hot-start method described in step S4 for initializing the weights of newly added nodes specifically involves: calculating the average feature vector of the low-confidence samples that triggered the self-learning engine in the high-dimensional feature space. The average feature vector is then processed through the original mapping matrix of the classifier. pseudo-inverse transform As the initial weights for newly added neuron nodes.

[0050] As one embodiment, the hot-start initialization process of the newly added node in step S4 of the present invention is as follows: Assume that when the self-learning engine is triggered, a total of [number] nodes have been collected. Low-confidence samples (e.g.) Each sample corresponds to a 128-dimensional feature vector. Then the average eigenvector Suppose the original classifier contains Given categories, the weight matrix of its fully connected layer is: The bias term is (Bias is temporarily ignored for simplification). Pseudo-inverse Through formula Calculate the initial weight vector of the newly added node. The physical meaning of this initialization is to project the feature centers of the new class onto the weight space of the original classifier, minimizing the confusion between its initial output and the original class, thereby accelerating the convergence of subsequent fine-tuning. For example, in a specific instance, The average features of the alteration zone samples are used to obtain the initial weight values ​​of the newly added nodes after pseudo-inverse transformation. This results in the original output score of the node for the alteration zone samples being approximately 0.6, while the output score for other categories is close to 0. Compared to random initialization (where the output score fluctuates around 0), hot-start initialization enables the model to correctly distinguish the new category in the first epoch, significantly improving learning efficiency.

[0051] It should be noted that step S4 involves setting differentiated learning rates that decay according to different curves and adding an orthogonal regularization term to the loss function, specifically as follows: A first initial learning rate is set for new nodes, and a second initial learning rate is set for existing nodes, with the first initial learning rate being greater than the second initial learning rate. The learning rate of new nodes decays with each fine-tuning round according to a first decay function, and the learning rate of existing nodes decays with each fine-tuning round according to a second decay function, wherein the decay rate of the first decay function is faster than that of the second decay function. Add an orthogonal regularization term to the total loss function ,in For the newly added node weight matrix, This is the original node weight matrix. The preset regularization coefficients, denoted by Frobenius norm, this regularization term is used to force the feature representation of newly added nodes to be orthogonal to the feature space of the original categories.

[0052] As one embodiment, the specific implementation of the differential learning rate and orthogonal regularization term in step S4 of the present invention is as follows: The learning rate is set as follows: Let the total number of fine-tuning rounds be... The initial learning rate of the newly added nodes The initial learning rate of the original nodes The first decay function uses exponential decay: ,in The second decay function is ,in After 5 epochs, , The learning rate of newly added nodes is consistently about two orders of magnitude higher than that of existing nodes, and decays faster, allowing new categories to adapt quickly while the weights of old categories remain almost unchanged.

[0053] The orthogonal regularization terms are as follows: Let the original node weight matrix be... Add node weight vector (as a matrix) ).calculate , get one A row vector, where each element is... With the The dot product of the old class weight vectors. The square of the Frobenius norm is the sum of the squares of these dot products: Orthogonal regularization term ,in For example, the regularization coefficient. Total loss function ,in This is the cross-entropy loss. This regularization term encourages the weights of new nodes to be orthogonal to all old class weights, so that in the feature space, the representation of the new class is as perpendicular as possible to the representation direction of the old classes, effectively reducing feature confusion. For example, in a fine-tuning process, the orthogonal regularization term reduced the product of the new class weight and the granite class weight from 0.3 to 0.05, significantly reducing the risk of misidentifying alteration zones as granite.

[0054] S5. Lock and apply the global mean and standard deviation of the normalizer during the initial training phase, perform performance backtesting on the evolved geological identification model, and save and deploy the evolved model after passing the backtest.

[0055] As one embodiment, in step S5 of the present invention, the normalizer persistence module stores the global mean vector of each feature dimension calculated during the initial pre-training phase. and standard deviation vector .

[0056] During model evolution and subsequent inference, all input data are standardized using this set of parameters: This ensures that new and old data are always mapped at the same feature scale, avoiding catastrophic forgetting caused by data distribution drift.

[0057] After evolution, the method performs performance backtesting on the retained validation set: calculating the recognition accuracy for each of the original categories before and after the model update, as well as the recognition accuracy for the new categories. If the decrease in accuracy for the original categories does not exceed a preset tolerance threshold (e.g., 5%), and the accuracy for the new categories exceeds a preset usable threshold (e.g., 80%), the backtest is considered successful. Upon success, the evolved model weights, structure configuration, and normalizer parameters are saved as a new model version and deployed to the on-site industrial control computer for subsequent tunneling use. For example, after incremental learning of alteration zones, the model's recognition accuracy for granite slightly decreased from 98% to 96%, while the recognition accuracy for alteration zones reached 93%, indicating a successful backtest. The method automatically replaces the old model, achieving continuous evolution.

[0058] It should be noted that the geological identification model adopts a hybrid expert-Transformer architecture, namely the MOE-Transformer architecture, and uses a Top-2 gating mechanism to achieve sparse activation of experts in this architecture.

[0059] As one embodiment, the geological identification model used in this invention is the MOE-Transformer, whose specific structure is as follows: The input layer maps the 9-dimensional drilling parameters to a 128-dimensional embedding vector through a linear embedding layer, and adds position encoding. The Transformer encoder contains 6 layers, each with 4 self-attention heads. After each encoder layer, a hybrid expert layer (MoE) is connected. The MoE layer contains 4 parallel expert networks, each of which is a two-layer fully connected network with a hidden layer dimension of 1024 and an activation function of GELU. The gating network receives the encoder output, outputs a 4-dimensional weight vector, and then uses a Top-2 gating mechanism: only the top 2 experts with the largest weights are retained, the weights of the remaining experts are reset to zero, the retained weights are then normalized, and finally the outputs of the selected 2 experts are weighted and summed. The mathematical expression is:

[0060] in This represents the original weight vector output by the gating network. This mechanism implements sparse activation of experts, with only two experts participating in the computation during each forward propagation, significantly reducing computational cost while maintaining the model's expressive power. For example, when identifying hard granite, experts 1 and 3, responsible for the high-impedance mode, might be activated; while when identifying weak fractured zones, experts 2 and 4, responsible for the low-impedance mode, are activated. This sparse activation mechanism allows the model to dynamically allocate computational resources according to different geological conditions, improving the efficiency of online inference.

[0061] Example 2: Please see Figure 4 , Figure 4 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: an online self-learning device 401 for tunnel geological identification based on drilling parameters, a processor 402, and a storage medium 403.

[0062] A tunnel geological identification online self-learning device 401 based on drilling parameters: The tunnel geological identification online self-learning device 401 based on drilling parameters implements the tunnel geological identification online self-learning method based on drilling parameters.

[0063] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the online self-learning method for tunnel geological identification based on drilling parameters.

[0064] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the online self-learning method for tunnel geological identification based on drilling parameters.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tunnel geological identification online self-learning method based on drilling parameters, characterized in that: Includes the following steps: S1. Real-time synchronous acquisition of multiple core drilling parameters during tunnel drilling, and preprocessing of the raw drilling parameters to generate standardized test sequence fragments; S2. Input the sequence fragment to be tested into the pre-trained geological identification model, and simultaneously monitor the prediction confidence, the time-series change rate of the drilling parameters, and the dynamic phase lag characteristics between at least two different parameters; when the prediction confidence is lower than the first threshold, the proportion of low-confidence samples exceeds the second threshold, the cumulative integral of the time-series change rate exceeds the third threshold, and the change in the dynamic phase lag characteristics exceeds the fourth threshold, activate the online self-learning engine. S3. In response to the activation of the online self-learning engine, the high-dimensional feature vector of the intermediate feature representation layer of the geological identification model is extracted using the hook function. The high-dimensional feature vector is then subjected to multi-scale clustering verification through topological persistence analysis. Combined with geometric feature verification, it is determined whether there are new feature clusters different from the known categories. If they exist, it is confirmed that a new geological pattern has been discovered. S4. Once a new geological pattern is confirmed, freeze the weights of the backbone network of the geological identification model, dynamically expand the neuron nodes corresponding to the new category in the output layer of its classifier, initialize the weights of the newly added nodes using a hot start method, set differentiated learning rates for the newly added nodes and the original nodes with different decay curves, add an orthogonal regularization term to the loss function, and perform lightweight fine-tuning using field verification data to obtain the evolved geological identification model. S5. Lock and apply the global mean and standard deviation of the normalizer in the initial training phase, perform performance backtesting on the evolved geological identification model, and save and deploy the evolved model after passing the backtest. Step S3, which involves performing multi-scale clustering verification on the high-dimensional feature vector through topological persistence analysis and combining it with geometric feature verification, specifically includes: S31. Extract the 128-dimensional feature vector output from the encoder layer in the geological identification model using the hook function; S32, at different neighborhood distance thresholds Given the set of possible values, a minimum spanning tree is constructed based on the high-dimensional feature vector, recording the first occurrence of each feature point cluster. Value as the birth side length and when the cluster is merged into other clusters Value as the length of the dead edge Calculate the lifetime of each cluster The normalized lifetime is used as the topological durability score for the cluster. ; S33, Persistence Score Candidate clusters that exceed a preset persistence threshold are marked as persistent feature clusters; S34. For each persistent feature cluster, verify in sequence: whether the Euclidean distance between its cluster center and the center of any known category exceeds the geometric verification threshold, whether the average distance of all samples in the cluster to the cluster center is less than the first compactness threshold, and whether the feature density change rate along the radial direction from the cluster center to the cluster boundary differs from the corresponding density gradient distribution of all known categories by more than a preset gradient threshold. S35. When all of the above triple geometric verification conditions are met, the discovery of a new geological model is confirmed. Step S4 involves setting differentiated learning rates that decay according to different curves and adding an orthogonal regularization term to the loss function, specifically as follows: A first initial learning rate is set for new nodes, and a second initial learning rate is set for existing nodes, with the first initial learning rate being greater than the second initial learning rate. The learning rate of new nodes decays with each fine-tuning round according to a first decay function, and the learning rate of existing nodes decays with each fine-tuning round according to a second decay function, wherein the decay rate of the first decay function is faster than that of the second decay function. Add an orthogonal regularization term to the total loss function ,in For the newly added node weight matrix, This is the original node weight matrix. The preset regularization coefficients, denoted by Frobenius norm, this regularization term is used to force the feature representation of newly added nodes to be orthogonal to the feature space of the original categories.

2. The online self-learning method for tunnel geological identification based on drilling parameters as described in claim 1, characterized in that, The multiple core drilling parameters mentioned in step S1 include impact pressure, impact flow rate, rotational speed, rotational pressure, propulsion pressure, drilling stroke, drilling speed, impact frequency, and propulsion force; the preprocessing includes outlier detection using Mahalanobis distance, noise and geological abrupt change separation using discrete wavelet transform, state estimation using Kalman filtering, and downsampling to a preset frequency.

3. The online self-learning method for tunnel geological identification based on drilling parameters as described in claim 1, characterized in that: In step S2: The first threshold is a preset high confidence threshold, the second threshold is a preset low confidence sample proportion threshold, and the sliding evaluation window length of the low confidence samples is a preset short window length. The cumulative integral value of the time-series rate of change is a weighted sum of the absolute values ​​of the rates of change of drilling speed and propulsion pressure after time integration; The dynamic phase lag characteristic is the instantaneous phase difference between the impact pressure and the rotational speed. It is calculated by performing Hilbert transform on the impact pressure signal and the rotational speed signal respectively, extracting the instantaneous phase, taking the difference and modulus, and then calculating the mean change within a preset short-time sliding window. When the change exceeds the fourth threshold, it reflects the physical precursor of the rock mass changing from brittle fracture to plastic cutting.

4. The online self-learning method for tunnel geological identification based on drilling parameters as described in claim 1, characterized in that, The hot-start method described in step S4 initializes the weights of newly added nodes, specifically by calculating the average feature vector of the low-confidence samples that triggered the self-learning engine in the high-dimensional feature space. The average feature vector is then processed through the original mapping matrix of the classifier. pseudo-inverse transform As the initial weights for newly added neuron nodes; where .

5. The online self-learning method for tunnel geological identification based on drilling parameters as described in claim 1, characterized in that, The geological identification model adopts a hybrid expert-Transformer architecture, namely the MOE-Transformer architecture, and uses a Top-2 gating mechanism to achieve sparse activation of experts in this architecture.

6. The online self-learning method for tunnel geological identification based on drilling parameters as described in claim 1, characterized in that, In step S4, during the lightweight fine-tuning process, known class prototype vectors, old model output distributions, or a small number of historical samples are used as stabilizing constraints to ensure that the change in the old class output boundary during the learning process of new nodes is less than the preset boundary drift threshold.

7. A storage medium, characterized in that: The storage medium stores instructions and data to implement the online self-learning method for tunnel geological identification based on drilling parameters as described in any one of claims 1 to 6.

8. An online self-learning device for tunnel geological identification based on drilling parameters, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the online self-learning method for tunnel geological identification based on drilling parameters as described in any one of claims 1 to 6.

Images