Method and apparatus for identifying lithology and weathering degree based on drill pipe vibration signals

Abstract

This invention relates to the field of mine roadway and traffic tunnel technology. To accurately identify the lithology and weathering degree of rocks within boreholes in real time, it provides a method and device for identifying lithology and weathering degree based on drill rod vibration signals. Based on a six-component vibration sensing architecture with an octahedral structure, it achieves six-component full-dimensional sensing of drill bit rock-breaking vibration acceleration signals. A triple filtering method is used to remove interference, and the vibration frequency is corrected by combining the drill rod length to eliminate the interference of drill rod length changes on the vibration frequency during drilling. The corrected signal is processed by frame segmentation, extracting time-frequency features and time-domain features respectively. Finally, a borehole lithology and weathering degree identification model composed of CNN branches and MLP branches is constructed. The CNN branch automatically learns the local correlation patterns of time-frequency features, and the MLP branch automatically combines the nonlinear relationships in the time-domain features, thereby achieving efficient and accurate identification of borehole lithology and weathering degree.

Description

Technical Field

[0001] This invention relates to the field of mine roadway and traffic tunnel technology, specifically a method and device for identifying lithology and weathering degree based on drill rod vibration signals. Background Technology

[0002] The lithology and weathering degree of rock are two fundamental conditions that determine the physical and mechanical parameters of the surrounding rock, and are also important bases for the excavation and support parameters of mine roadways and traffic tunnels. Because mine roadways and traffic tunnels involve large spatial areas and require traversing various strata with constantly changing lithology and weathering conditions, it is crucial to dynamically, in real-time, and accurately grasp the specific lithology and weathering conditions of the working face. This is of great significance for guiding design and construction and ensuring project safety.

[0003] Currently, the most direct and accurate surveying method is to determine geological parameters through core drilling tests. However, core drilling is costly and the number of boreholes is limited, making it difficult to accurately determine the geological parameters between adjacent boreholes. Although there are comprehensive advanced geological prediction methods such as geophysical exploration, electrical resistivity tomography, and ground-penetrating radar, these technologies have drawbacks such as limited applicability, multiple interpretations of detection results, and difficulty in predicting rock lithology and weathering degree. Experienced geological engineers are often required to make judgments, but manual judgment is costly, untimely, and highly subjective. Therefore, there is an urgent need to develop a technology that can accurately and automatically sense the lithology and weathering degree of strata within a certain range ahead during tunnel excavation. Summary of the Invention

[0004] To identify the lithology and weathering degree of rocks inside boreholes in real time and accurately, this invention provides a method and apparatus for identifying lithology and weathering degree based on drill pipe vibration signals.

[0005] The technical solution adopted by the present invention to solve the above problems is:

[0006] Methods for identifying lithology and weathering degree based on drill pipe vibration signals include:

[0007] Step 1: Collect measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during the drilling process. The sensing carrier is a regular octahedron, and a single-axis accelerometer is installed at the midpoint of each edge of the regular octahedron.

[0008] Step 2, based on measurement vector ,according to The vibration state vector X and H are calculated as the structure matrix determined by the position vector of the three-dimensional coordinate point of the single-axis accelerometer and the unit vector of the sensitive axis direction; based on the vibration state vector Extract the six-degree-of-freedom vibration acceleration vector;

[0009] Step 3: Filter the extracted six-degree-of-freedom vibration acceleration vector to obtain a clean six-degree-of-freedom vibration signal;

[0010] Step 4: Perform a fast Fourier transform on the pure six-degree-of-freedom vibration signal to obtain the vibration acceleration signal spectrum, and correct the frequency based on the drill pipe length;

[0011] Step 5: Perform frame segmentation on the corrected signal, extract the time-frequency features and time-domain features of the framed signals respectively, and normalize the extracted features.

[0012] Step 6: Using the normalized features as input, determine the borehole lithology and weathering degree based on the trained borehole lithology and weathering degree identification model.

[0013] Furthermore, step 2 specifically involves:

[0014] Step 21: Set the cost function , ;

[0015] Step 22: With the goal of minimizing the cost function, determine the optimal solution of the vibration state vector using the least squares method. , ,in: As the carrier center Acceleration in the direction of; For the center of the carrier Angular velocity of the axis of rotation; For the center of the carrier Angular acceleration of the axis of rotation;

[0016] Step 23: Based on the vibration state vector Extracting the six-degree-of-freedom vibration acceleration vector Where d is the direction number of the degree of freedom. ; The sampling time.

[0017] Furthermore, step 3 employs triple filtering, including: sequentially applying bandpass filtering, wavelet limiting filtering, and wavelet threshold filtering.

[0018] Furthermore, the specific steps for correcting the frequency based on the drill pipe length are as follows:

[0019] Step 41: Measure the actual drilling length of the drill pipe in real time. ;

[0020] Step 42: Obtain the elastic wave velocity of the drill pipe ;

[0021] Step 43, based on Determine the theoretical frequency difference ;

[0022] Step 44, according to Determine the frequency difference correction factor In the formula, They are respectively The frequency corresponding to a given moment;

[0023] Step 45: Based on the frequency difference correction coefficient ,right Frequency of time Make corrections: , This is the corrected frequency.

[0024] Furthermore, during framing, adjacent frames overlap with each other.

[0025] Furthermore, the borehole lithology and weathering degree identification model includes:

[0026] CNN branching models are used to process time-frequency features;

[0027] MLP branching model is used to process temporal features;

[0028] The feature concatenation layer is used to concatenate the features output by the CNN branch model and the MLP branch model.

[0029] A fully connected fusion layer is used to perform fully connected transformations and nonlinear activations on the spliced ​​fusion features;

[0030] The borehole lithology and weathering degree results are output layers. Based on the output results of the fused fully connected layer, the borehole lithology output head is used to output the borehole lithology, and the weathering degree output head is used to output the weathering degree.

[0031] Furthermore, when training the borehole lithology and weathering degree identification model, the loss function is composed of a weighted sum of the borehole lithology output head loss function and the weathering degree output head loss function.

[0032] A lithology and weathering degree identification device based on drill pipe vibration signals includes:

[0033] The data acquisition unit is used to acquire measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during drilling. The sensing carrier is a regular octahedron, and a single-axis accelerometer is installed at the midpoint of each edge of the regular octahedron.

[0034] Six-DOF vibration acceleration vector acquisition unit: based on measurement vector ,according to Solving for the vibration state vector H is the structure matrix determined based on the position vector of the three-dimensional coordinate point of the single-axis accelerometer and the unit vector of the sensitive axis direction; based on the vibration state vector Extract the six-degree-of-freedom vibration acceleration vector;

[0035] Filtering unit: Used to filter the extracted six-degree-of-freedom vibration acceleration vector to obtain a clean six-degree-of-freedom vibration signal;

[0036] Frequency correction unit: Performs fast Fourier transform on the pure six-degree-of-freedom vibration signal to obtain the vibration acceleration signal spectrum, and corrects the frequency based on the drill pipe length;

[0037] Feature extraction unit: performs frame segmentation on the corrected signal, extracts the time-frequency features and time-domain features of the framed signal respectively, and normalizes the extracted features.

[0038] Identification Unit: Using normalized features as input, the borehole lithology and weathering degree are determined based on the trained borehole lithology and weathering degree identification model.

[0039] Furthermore, the sensing carrier is located inside the first section of the drill pipe connected to the drill bit.

[0040] Furthermore, the frequency correction unit obtains the drill pipe length through a wire displacement sensor located at the tail of the last drill pipe section.

[0041] The advantages of this invention compared to the prior art are:

[0042] (1) By installing the first and last section of the drill rod with integrated sensing devices on the traditional drilling rig, and with the help of deep learning algorithms, the lithology and weathering degree of the drilling rig strata can be simultaneously perceived and identified during the normal drilling process. This method does not require manual judgment, does not interfere with normal drilling construction, and has the ability to automatically extract and classify features from end to end. It significantly improves the level of intelligent geological information acquisition and has the advantages of high accuracy, good timeliness and strong practical value.

[0043] (2) An innovative six-component vibration sensing architecture based on a regular octahedral structure was designed. By constructing an analytical model of 12-channel redundant measurement vectors to solve the six-degree-of-freedom vibration vector, the "six-component" full-dimensional sensing of drill bit rock-breaking vibration acceleration signal monitoring was realized. The three-dimensional time-frequency tensor was extracted by wavelet packet decomposition, and the time-domain statistical feature vector was calculated at the same time. A multi-dimensional feature data system containing vibration direction, frequency band, time and statistical characteristics was constructed. This feature data system comprehensively reflects the time-frequency structure and waveform of drill rod vibration, providing rich and complementary vibration information support for subsequent identification of lithology and weathering degree.

[0044] (3) A rod length correction method for the resonant main frequency of "drill rod-rock" was proposed. The frequency difference was corrected by combining elastic wave theory, which eliminated the interference of the drill rod length change on the vibration frequency during the drilling process, making the subsequent model recognition more accurate and having good engineering practicality and promotion value.

[0045] (4) Construct a borehole lithology and weathering degree identification model consisting of a CNN branch and an MLP branch, which integrates heterogeneous features. The CNN branch automatically learns the local correlation patterns in the time-frequency graph, and the MLP branch automatically combines the nonlinear relationships in the statistical features. Without the need for manual design of complex features or reliance on expert experience, the model can adaptively optimize parameters, has strong generalization ability, robustness and transferability, and can be conveniently applied to different types of lithology and weathering degree identification tasks. It also has strong scalability. Attached Figure Description

[0046] Figure 1 The flowchart shows a method for identifying lithology and weathering degree based on drill pipe vibration signals.

[0047] Figure 2 Schematic diagram of a regular octahedral rigid base;

[0048] Figure 3 This is a schematic diagram of the drill pipe structure;

[0049] Figure 4 A schematic diagram of the borehole lithology and weathering degree identification model structure;

[0050] Figure 5 This is a schematic diagram of a CNN branch model;

[0051] Figure 6 This is a schematic diagram of an MLP branching model;

[0052] Reference numerals in the attached diagram: 1 is an octahedral base, 2 is a single-axis accelerometer, 3 is the first section of the drill rod, 4 is the drill bit, 5 is the last section of the drill rod, 6 is a wire displacement sensor, and 7 is the power head. Detailed Implementation

[0053] 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.

[0054] To achieve real-time and accurate identification of rock lithology and weathering degree within boreholes, this invention designs a six-component vibration sensing architecture based on an octahedral structure. An analytical model for calculating six-degree-of-freedom vibration vectors using 12-channel redundant measurement vectors is constructed, enabling full-dimensional six-component sensing of drill bit rock-breaking vibration acceleration signals. A triple filtering method is then employed to effectively remove interference, and the vibration frequency is corrected based on drill rod length to eliminate interference from drill rod length variations during drilling. Subsequently, the corrected signal is processed in frames, extracting time-frequency and time-domain features from each frame. Finally, a borehole lithology and weathering degree identification model is constructed, consisting of a CNN branch and an MLP branch, which fuse heterogeneous features. The CNN branch automatically learns local correlation patterns in the time-frequency features, while the MLP branch automatically combines nonlinear relationships in the time-domain features, thus achieving efficient and accurate identification of borehole lithology and weathering degree.

[0055] like Figure 1 As shown, the method for identifying lithology and weathering degree based on drill pipe vibration signals includes:

[0056] Step 1: Collect measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during the drilling process. The sensing carrier is a regular octahedron, and a single-axis accelerometer is installed at the midpoint of each edge of the regular octahedron.

[0057] During the rock-breaking process, the drill bit causes axial and torsional vibrations in three orthogonal directions in the first section of the drill rod. This vibration acceleration information is closely related to the lithology and specific degree of weathering of the rock. To collect relevant information, such as... Figure 2 , 3 As shown, this invention integrates a rigid octahedral base 1 inside the first drill rod section 3 connected to the drill bit 4 as an integrated sensing carrier for acceleration sensors. A single-axis acceleration sensor 2 is installed at the midpoint of each of the 12 edges of the octahedral base 1. The sensitive axis of each sensor is parallel to the edge direction, used to collect the axial acceleration generated by the drill bit's rock-breaking vibration along the 12 edges. The sensor leads are routed along the edges for convenient centralized routing.

[0058] Using uniaxial accelerometers on each edge of a rigid octahedral base, the measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during the drilling and rock-breaking process are obtained. .

[0059] Step 2, based on measurement vector ,according to The vibration state vector X and H are calculated as the structure matrix determined by the position vector of the three-dimensional coordinate point of the single-axis accelerometer and the unit vector of the sensitive axis direction; based on the vibration state vector Extract the six-degree-of-freedom vibration acceleration vector.

[0060] according to It can calculate the vibration state vector including the center of the built-in sensing carrier in the drill pipe. , ;in: As the carrier center Acceleration in the direction of; For the center of the carrier Angular velocity of the axis of rotation; For the center of the carrier Angular acceleration of the axis of rotation; The structure is a 12×12 matrix, where the 12 elements are known quantities. The three-dimensional spatial coordinates of the sensor position vector and the Unit vector along the sensing axis of each sensor The determination method is as follows:

[0061] There are 12 sensors in total, so there are 12 rows. Taking the i-th sensor as an example, the method for determining the 12 elements in the i-th row is as follows:

[0062]

[0063] The 12-edge octahedral vibration signal sensing carrier of this invention features a highly redundant design architecture. The process of calculating the measured vibration state vector X using the above steps involves finding the optimal solution for 9 unknowns based on 12 known quantities. To achieve this, this embodiment employs the least squares method, utilizing the following cost function. Determine the vibration state vector optimal solution :

[0064] Cost function: ;

[0065] The optimal solution of the vibration state vector determined by least squares Satisfy the following formula:

[0066] ;

[0067] The optimal solution of the vibration state vector obtained from the above steps Extract the original six-degree-of-freedom vibration acceleration vector at the center of the drill pipe carrier. Where d is the direction number of the degree of freedom. ; The sampling time. It reflects the vibration acceleration characteristic signals generated in six dimensions by the "drill bit-rock" rock breaking process under different lithology and weathering conditions. These signals contain information on the lithology and weathering degree of the rock.

[0068] Step 3: Filter the extracted six-degree-of-freedom vibration acceleration vector to obtain a clean six-degree-of-freedom vibration signal.

[0069] Original vibration acceleration vector of drill rod collected during drilling operation This includes various external environmental interference vibration signals such as machine shaking, motor noise, compressed air circulation within the borehole, friction between the drill bit and the rock wall or cuttings, and the flow of water and slurry within the borehole. To eliminate these interference vibration signals, at each sampling time point... Vibration signal in each dimension The following three-stage filtering process is performed sequentially.

[0070] First-stage filtering: A bandpass filter is used to remove irrelevant signals outside the effective frequency band of the drill pipe, such as low-frequency mechanical vibrations generated during drilling operation, and ultra-high-frequency electronic noise waves generated by power switches or sensor electronic components, while retaining the lower limit frequency of the main frequency vibration signal generated during the drill bit's rock-breaking process. Upper limit frequency Determined by Fast Fourier Transform (FFT) analysis and The value of is taken in this embodiment. .

[0071] This embodiment uses a Butterworth bandpass filter as an example. The square of its low-pass prototype amplitude-frequency response can be expressed as: In the formula, Indicates frequency as The proportion of energy retained in the signal after passing through the filter; The angular frequency of the signal. j is the imaginary unit; ω is the threshold angular frequency (rad / s); f is the frequency of the acquired signal (Hz). This represents the filter order.

[0072] A practical bandpass filter can be obtained from the above lowpass prototype through frequency transformation, and its design parameters (order n, threshold frequency) are... and The appropriate filter order and specific implementation method should be determined based on the required filtering performance. Those skilled in the art can select the appropriate filter order and specific implementation method based on the above principles and actual needs; no restrictions are imposed here.

[0073] Second-stage filtering: Based on the first-stage filtering, wave-limiting filtering is used to filter out the harmonic components generated by electromagnetic fields or electromagnetic radiation from electrical equipment such as AC power, motors, and frequency converters at the drilling rig's working site. This prevents these components from superimposing on the spectrum and forming false peaks after entering the sensor, thereby eliminating deterministic periodic power frequency interference waves.

[0074] This embodiment uses the power frequency interference angular frequency commonly found in drilling in roadway surrounding rock. The transfer function of the two-stage limiter, designed with its harmonics of 100Hz and 150Hz, is as follows: In the formula, To limit the center angular frequency of the wave, The power frequency interference to be filtered out, The sampling frequency of the signal; The width factor is a limiting factor, and its value ranges from 0 to 1. In this embodiment, it is set to 0.95. , These represent shifting the signal forward by one and two sampling periods, respectively.

[0075] The third layer of filtering employs wavelet thresholding to remove random background noise vibrations caused by factors such as friction between the drill rod or drill bit and the rock wall or cuttings, the flow of water and slurry within the borehole, and the circulation of compressed air, while preserving the transient impact signal characteristics generated during the drilling and rock breaking process. This embodiment performs J-layer wavelet decomposition on the signal and applies the following thresholding processing: In the formula, the threshold ,in, Where N is the noise standard deviation and N is the signal length; Wavelet decomposition coefficients; sign function Return coefficient The symbol (+1 or -1).

[0076] After the above three signal filtering processes, a pure six-degree-of-freedom vibration signal can be obtained, which is generated solely by the rock-breaking vibration between the drill bit and the rock.

[0077] Step 4: Perform a fast Fourier transform on the pure six-degree-of-freedom vibration signal to obtain the vibration acceleration signal spectrum, and correct the frequency based on the drill pipe length.

[0078] The drill rod length increases with drilling depth. Since the fixed frequency of the drill rod decreases with increasing length, the resonant frequency between the drill rod and the rock also dynamically changes with the drill rod length. To eliminate the interference of drill rod length variations on the resonant frequency signal characteristics during drilling, this invention corrects the resonant frequency based on the real-time drill rod length during drilling, including:

[0079] (1) such as Figure 3As shown, a wire displacement sensor 6 is integrated at the power head 7, which connects the drill rig body to the tail of the last drill rod 5. The wire end is connected to the tail of the last drill rod 5, enabling real-time measurement of the actual drilling length of the drill rod during drilling. Simultaneously, it collects the vibration acceleration of the internal carrier of the first drill rod as the drill rod advances, ensuring real-time measurement at each sampling point. The vibration data all correspond to a precise measured value of the actual drill pipe feed length. .

[0080] (2) The elastic wave velocity of the drill pipe is obtained in advance based on the calibration test of the propagation velocity of elastic waves inside the drill pipe. The theoretical frequency difference is determined as follows: : .

[0081] (3) Perform a Fast Fourier Transform (FFT) on the filtered vibration acceleration signal to obtain the vibration acceleration signal spectrum. Based on the data from two adjacent moments... corresponding frequency The frequency difference correction factor is determined as follows: .

[0082] (4) Based on the frequency difference correction coefficient On the spectrum Frequency of time The equivalent frequency, normalized to the standard length, is obtained by performing correction as follows. : .

[0083] After the above correction process, the influence of the change in drill rod length during drilling on the resonant main frequency characteristics can be eliminated.

[0084] Step 5: Perform frame segmentation on the corrected signal, extract the time-frequency features and time-domain features of the framed signals respectively, and normalize the extracted features.

[0085] The lithology and weathering degree change continuously with the drilling depth. To capture the dynamic changes in drill pipe vibration signal characteristics with lithology and weathering degree, the pure six-degree-of-freedom vibration signal, after filtering and rod length frequency correction, is analyzed. Perform frame segmentation processing.

[0086] In this embodiment, the signal is divided into consecutive frames with N=1024 sampling points, denoted as... ,in, For frame number index, Index the sampling points. The degrees of freedom in the six vibration directions are indexed. Adjacent frames overlap by 50% to ensure temporal continuity.

[0087] Wavelet packet time-frequency feature extraction of framed signals: J-level wavelet packet decomposition is performed on the six-degree-of-freedom vibration signal of each frame. In this embodiment, J=5 is used. The decomposition yields... Wavelet packet coefficients of equal bandwidth And calculate the energy characteristic value for each frequency band: .

[0088] Arrange the energy values ​​of all frames in temporal order T to obtain the three-dimensional feature tensor. , This reflects the vibrational energy in direction (d=6) and frequency band ( =32), time (t=T) and other three dimensions of information.

[0089] Temporal feature extraction of framed signals: Four feature values ​​are calculated for each frame: variance, root mean square value, crest factor, and kurtosis of the energy values ​​in the six degrees of freedom directions. These values ​​reflect the waveform morphology characteristics of the six-degree-of-freedom vibration energy.

[0090] (1) Variance: ,in: The standard deviation of the signal is represented. It is used to reflect the fluctuation amplitude of vibration signals (acceleration);

[0091] (2) Root mean square value: It is used to reflect the average energy of the vibration signal (acceleration);

[0092] (3) Crest factor: It is used to reflect the impact degree of vibration signal (acceleration);

[0093] (4) Kurtosis: It is used to reflect the sharpness of the vibration signal (acceleration) distribution. for The mean.

[0094] Each frame has four waveform morphology feature values ​​in each of the six degrees of freedom directions, for a total of n = 4 × 6 = 24 feature values. These, together with the temporal sequence T of each frame, form the temporal feature value matrix of all frames. .

[0095] Eigenvalue normalization: In order to eliminate the dimensional influence between eigenvalues ​​of different dimensions, this invention normalizes the above-mentioned three-dimensional time-frequency feature tensor. and time-domain eigenvalue matrix The element data in the file is normalized, including,

[0096] For three-dimensional feature tensors Min-Max normalization based on the degree of freedom d yields the following result This maps the numerical range of each degree of freedom to [0,1]. These represent the maximum and minimum values ​​of the vibration characteristic value in the d-th direction across all frequency bands and all time frames, respectively.

[0097] For time-domain eigenvalue matrix The results are obtained by Min-Max normalization based on the feature dimension n (variance, root mean square value, crest factor, and kurtosis in each vibration direction). .in, These represent the maximum and minimum values ​​of the feature dimensions in each direction across all time frames, respectively.

[0098] Step 6: Using the normalized features as input, determine the borehole lithology and weathering degree based on the trained borehole lithology and weathering degree identification model.

[0099] This approach targets two types of heterogeneous feature input data simultaneously present in vibration signals—three-dimensional time-frequency tensor data with spatial structure. and spatially unstructured time-domain statistical data A dual-branch heterogeneous feature fusion neural network deep learning model was designed, including: CNN branch, MLP branch, feature concatenation, fusion fully connected layer, and output layer for borehole lithology and weathering degree results, such as... Figure 4 As shown. The CNN branch is used to process 3D feature tensor input data that exhibits clear local correlations in the orientation, frequency band, and time dimensions. The MLP branch is used to process input data of time-domain eigenvalue matrices that are independent of each dimension. By fully leveraging the complementarity of heterogeneous information from two branches, feature fusion and multi-task output are used to improve the model's accuracy and generalization ability in identifying borehole lithology and weathering degree. The specific model structure includes:

[0100] (1) Constructing and processing three-dimensional time-frequency feature tensor data CNN branch models, such as Figure 5 As shown.

[0101] Normalized three-dimensional time-frequency tensor It possesses clear spatial structural characteristics (direction, frequency band, time) and strong numerical correlation characteristics in local regions. To address these data structure characteristics, this invention constructs a convolutional neural network (CNN) branch model to automatically learn and extract these local features, specifically including:

[0102] Input layer: The input data for the CNN branch comes from the three-dimensional time-frequency tensor after preprocessing the original six-degree-of-freedom vibration signal, correcting the rod length, and normalizing it. The input data retains information on the vibration signals generated by the interaction between the drill rod and the rock in three dimensions: direction, frequency, and time, providing rich data integration for the CNN branch model to learn features related to lithology and weathering degree.

[0103] Convolutional layer 1: It includes 32 3×3 convolutional kernels (filters) that slide sequentially on the "frequency band J-time T" data plane with a stride of 1 to extract local correlation feature values ​​on 6 vibration directions d, and output a feature size map with a feature size of 32×32×T.

[0104] Pooling layer 1: Construct 2×2 max pooling to reduce the dimension of the feature size map to 2×32×T / 2, which is used as the input data for convolutional layer 2;

[0105] Convolutional layer 2: consists of 64 3×3 convolutional kernels (filters), with the activation function being... Extract further abstracted combined feature values ​​from the dimension-reduced compressed feature size map data, and output an abstracted feature size map with a size of 64×16×T / 2;

[0106] Pooling layer 2: Construct 2×2 max pooling to reduce the dimension of the abstracted feature size map to 64×8×T / 4;

[0107] The flattening layer flattens the 64×8×T / 4 into a one-dimensional feature vector. In this embodiment, taking T=100 as an example, the flattened dimension is 64×8×25=12800, that is... ;

[0108] Fully connected layer: In this embodiment, the fully connected layer constructs 128 neurons, using the weight matrix of the model parameters. and bias vector Using activation functions The 12800-dimensional feature vector of the flattened layer is compressed into a 128-dimensional feature vector. .

[0109] The above method extracts a 128-dimensional feature vector related to borehole lithology and weathering degree from a CNN branch model. It contains a wealth of local information.

[0110] (2) Constructing and processing time-domain statistical data MLP branching models, such as Figure 6 As shown.

[0111] Normalized time-domain eigenvalue matrix This includes variance, root mean square, crest factor, and kurtosis in each direction for d=6. These time-domain features are unordered scalars with no spatial relationship between their dimensions, making them suitable for nonlinear combination using a fully connected multilayer microprocessor (MLP) network. This invention constructs a multilayer fully connected MLP network model to automatically learn the high-order interactions between these statistical features. Specifically, it includes:

[0112] Input layer : From the time-domain eigenvalue matrix Constructing a 24-dimensional input layer vector The input data reflects the time-domain statistical features in six vibration directions, which have clear statistical meanings. The 24-dimensional vector can avoid the curse of dimensionality while retaining key statistical information, making it suitable as input data for a fully connected network MLP. Furthermore, the time-domain statistical data of MLP is suitable for describing the global features of the signal, while the time-frequency data of CNN is suitable for describing local features, which can complement each other.

[0113] Fully Connected Layer 1 (Hidden Layer 1): In this embodiment, the fully connected layer 1 (hidden layer 1) constructs 64 neurons, which are connected through the weight matrix of the model parameters. and bias vector Using the ReLU activation function, the 24-dimensional input parameters are transformed. Mapping to a 64-dimensional fully connected layer 1 (hidden layer 1) yields a 64-dimensional feature vector. ;

[0114] Fully Connected Layer 2 (Hidden Layer 2): In this embodiment, the fully connected layer 2 (hidden layer 2) constructs 32 neurons, which are connected through the weight matrix of the model parameters. and bias vector Using the ReLU activation function, the 64-dimensional feature vector is transformed... Compressed into a 32-dimensional output feature vector .

[0115] The above method extracts a 32-dimensional feature vector related to borehole lithology and weathering degree from the MLP branch model. It contains the global statistical regularity of drill pipe vibration signals, which, along with the rich local feature information extracted by the aforementioned CNN branch model, is... They complement each other.

[0116] (3) Feature concatenation layer: The feature vectors output by the above CNN branch model are concatenated. The output feature vector of the MLP branch model These two types of heterogeneous features are concatenated to obtain a 160-dimensional fused feature vector. , as input feature data for the fusion fully connected layer.

[0117] (4) Fusing fully connected layers: By fusing the weight matrices of the fully connected layers and bias vector The ReLU activation function is used to output a 64-dimensional fused feature vector. This enables cross-modal feature interaction and fusion across two branches.

[0118] The fully connected layer constructed in this invention is trained to simultaneously serve as the borehole lithology output head and the borehole weathering degree output head of the output layer, thereby learning a more robust and general feature representation.

[0119] (5) Output layer of borehole lithology and weathering degree results: Based on the fact that lithology and weathering degree are two relatively independent and weakly correlated attributes, this invention uses two parallel output heads such as borehole lithology output head and weathering degree output head as output layers. At the same time, the two output heads share the 64-dimensional fusion feature vector of the fusion fully connected layer.

[0120] Based on the classification of potential lithological types in the tunnel crossing project area revealed by engineering geological survey data, this embodiment uses a 5-dimensional vector for the borehole lithology output head, representing [limestone, sandstone, shale, mudstone, phyllite]; and uses a lithology output head weight matrix... and lithological output head bias The following Softmax function is used to convert the lithology category probability results into an output vector. ,in .

[0121] ,in The function is defined as This ensures that the sum of the outputs of the probability function is 1.

[0122] The weathering degree of the borehole rock is classified into five categories: [completely weathered, strongly weathered, moderately weathered, slightly weathered, and unweathered]; this is determined by the lithology output head weight matrix. and lithological output head bias The following Softmax function is used to convert the rock weathering degree probability result into an output vector. , .

[0123] (6) Model training and damage function:

[0124] In a specific embodiment of the present invention, cross-entropy is used as the loss function for the error between the predicted output layer and the actual target output layer. The loss function is composed of a weighted sum of the lithology output head loss function and the weathering degree output head loss function.

[0125] ,in, Let lithology output head loss function be used. , This represents the true lithological label of the i-th sample (in one-hot encoded form, i.e., the correct category corresponds to a value of 1, and the rest are 0). Let be the probability distribution of the i-th sample belonging to each lithology category as predicted by the model; Output head loss function for weathering degree. , The true label of the weathering degree of the i-th sample (also one-hot encoded). The probability distribution of the i-th sample belonging to each weathering degree category as predicted by the model; This is a weighted balance factor used to balance the importance of lithology identification and weathering degree identification. When This indicates that both are equally important; This indicates that lithological identification is more important; Indicating the degree of weathering is more important for identification.

[0126] Using known input and output vector data obtained from actual drilling activities in tunnel engineering as training samples, the loss function is calculated according to the backpropagation method. The minimum value is used to determine the neural network structure parameters such as weights and biases for each layer in the model.

[0127] In this embodiment, vibration signals of the drill rod are collected during actual tunnel construction drilling, and the three-dimensional time-frequency feature tensor of the vibration frame signal, after filtering, rod length correction, and normalization, is used. and time domain eigenvalues The data is input into a trained deep learning network model, and after forward propagation, the borehole lithology probability distribution is obtained from the two output heads. Degree of weathering of borehole rocks The maximum a posteriori probability criterion is adopted, and the lithology or weathering degree represented by the index corresponding to the maximum probability value is taken as the final identification result of borehole rock lithology and weathering degree: Borehole rock lithology = Degree of weathering of drilled rock = ,in, Let be the probability vector of the lithology of the borehole rocks. To output the lithology represented by the position with the highest probability, for example: output position 2 represents sandstone. Let be the probability vector of the weathering degree of the drilled rock. To output the degree of weathering represented by the position with the highest probability, for example: output position 2 represents strong weathering.

[0128] Correspondingly, the present invention also provides a lithology and weathering degree identification device based on drill pipe vibration signals, comprising:

[0129] The data acquisition unit is used to acquire measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during drilling. The sensing carrier is a regular octahedron, and a single-axis accelerometer is installed at the midpoint of each edge of the regular octahedron.

[0130] Six-DOF vibration acceleration vector acquisition unit: based on measurement vector ,according to Solving for the vibration state vector H is the structure matrix determined based on the position vector of the three-dimensional coordinate point of the single-axis accelerometer and the unit vector of the sensitive axis direction; based on the vibration state vector Extract the six-degree-of-freedom vibration acceleration vector;

[0131] Filtering unit: Used to filter the extracted six-degree-of-freedom vibration acceleration vector to obtain a clean six-degree-of-freedom vibration signal;

[0132] Frequency correction unit: Performs fast Fourier transform on the pure six-degree-of-freedom vibration signal to obtain the vibration acceleration signal spectrum, and corrects the frequency based on the drill pipe length;

[0133] Feature extraction unit: performs frame segmentation on the corrected signal, extracts the time-frequency features and time-domain features of the framed signal respectively, and normalizes the extracted features.

[0134] Identification Unit: Using normalized features as input, the borehole lithology and weathering degree are determined based on the trained borehole lithology and weathering degree identification model.

[0135] Specifically, the sensing carrier is installed inside the first section of the drill pipe connected to the drill bit. The frequency correction unit obtains the drill pipe length through a wire displacement sensor located at the tail of the last section of the drill pipe.

Claims

1. A method for identifying lithology and weathering degree based on drill pipe vibration signals, characterized in that, include: Step 1: Collect measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during the drilling process. The sensing carrier is a regular octahedron, and a single-axis accelerometer is installed at the midpoint of each edge of the regular octahedron. Step 2, based on measurement vector ,according to The vibration state vector X and H are calculated as the structure matrix determined by the position vector of the three-dimensional coordinate point of the single-axis accelerometer and the unit vector of the sensitive axis direction; based on the vibration state vector Extract the six-degree-of-freedom vibration acceleration vector; Step 3: Filter the extracted six-degree-of-freedom vibration acceleration vector to obtain a clean six-degree-of-freedom vibration signal; Step 4: Perform a fast Fourier transform on the pure six-degree-of-freedom vibration signal to obtain the vibration acceleration signal spectrum, and correct the frequency based on the drill pipe length; Step 5: Perform frame segmentation on the corrected signal, extract the time-frequency features and time-domain features of the framed signals respectively, and normalize the extracted features. Step 6: Using the normalized features as input, determine the borehole lithology and weathering degree based on the trained borehole lithology and weathering degree identification model.

2. The method for identifying lithology and weathering degree based on drill pipe vibration signals according to claim 1, characterized in that, Step 2 is as follows: Step 21: Set the cost function , ; Step 22: With the goal of minimizing the cost function, determine the optimal solution of the vibration state vector using the least squares method. , ,in: As the carrier center Acceleration in the direction of; For the center of the carrier Angular velocity of the axis of rotation; For the center of the carrier Angular acceleration of the axis of rotation; Step 23: Based on the vibration state vector Extracting the six-degree-of-freedom vibration acceleration vector Where d is the direction number of the degree of freedom. ; The sampling time.

3. The method for identifying lithology and weathering degree based on drill pipe vibration signals according to claim 1, characterized in that, Step 3 employs a triple filtering process, including: sequentially applying bandpass filtering, wavelet limiting filtering, and wavelet threshold filtering.

4. The method for identifying lithology and weathering degree based on drill pipe vibration signals according to claim 1, characterized in that, The specific steps for correcting the frequency based on the drill pipe length are as follows: Step 41: Measure the actual drilling length of the drill pipe in real time. ; Step 42: Obtain the elastic wave velocity of the drill pipe ; Step 43, based on Determine the theoretical frequency difference ; Step 44, according to Determine the frequency difference correction factor In the formula, They are respectively The frequency corresponding to a given moment; Step 45: Based on the frequency difference correction coefficient ,right Frequency of time Make corrections: , This is the corrected frequency.

5. The method for identifying lithology and weathering degree based on drill pipe vibration signals according to claim 1, characterized in that, During frame splitting, adjacent frames overlap.

6. The method for identifying lithology and weathering degree based on drill pipe vibration signals according to claim 1, characterized in that, The borehole lithology and weathering degree identification models include: CNN branching models are used to process time-frequency features; MLP branching model is used to process temporal features; The feature concatenation layer is used to concatenate the features output by the CNN branch model and the MLP branch model. A fully connected fusion layer is used to perform fully connected transformations and nonlinear activations on the spliced ​​fusion features; The borehole lithology and weathering degree results are output layers. Based on the output results of the fused fully connected layer, the borehole lithology output head is used to output the borehole lithology, and the weathering degree output head is used to output the weathering degree.

7. The method for identifying lithology and weathering degree based on drill pipe vibration signals according to claim 6, characterized in that, When training the borehole lithology and weathering degree identification model, the loss function is composed of the weighted sum of the borehole lithology output head loss function and the weathering degree output head loss function.

8. A lithology and weathering degree identification device based on drill pipe vibration signals, characterized in that, include: The data acquisition unit is used to acquire measurement vectors of vibration acceleration in 12 directions generated at the center of the sensing carrier during drilling. The sensing carrier is a regular octahedron, and a single-axis accelerometer is installed at the midpoint of each edge of the regular octahedron. Six-DOF vibration acceleration vector acquisition unit: based on measurement vector ,according to Solving for the vibration state vector H is the structure matrix determined based on the position vector of the three-dimensional coordinate point of the single-axis accelerometer and the unit vector of the sensitive axis direction; based on the vibration state vector Extract the six-degree-of-freedom vibration acceleration vector; Filtering unit: Used to filter the extracted six-degree-of-freedom vibration acceleration vector to obtain a clean six-degree-of-freedom vibration signal; Frequency correction unit: Performs fast Fourier transform on the pure six-degree-of-freedom vibration signal to obtain the vibration acceleration signal spectrum, and corrects the frequency based on the drill pipe length; Feature extraction unit: performs frame segmentation on the corrected signal, extracts the time-frequency features and time-domain features of the framed signal respectively, and normalizes the extracted features. Identification Unit: Using normalized features as input, the borehole lithology and weathering degree are determined based on the trained borehole lithology and weathering degree identification model.

9. The lithology and weathering degree identification device based on drill pipe vibration signal according to claim 8, characterized in that, The sensing carrier is located inside the first section of the drill pipe connected to the drill bit.

10. The lithology and weathering degree identification device based on drill pipe vibration signal according to claim 8, characterized in that, The frequency correction unit obtains the drill pipe length through a wire displacement sensor located at the tail of the last section of the drill pipe.

Images