Drilling machine three-dimensional forcible entry method fused with PCA graph neural network

Abstract

The invention discloses a drilling machine three-dimensional forcible entry method fused with a PCA graph neural network, and the method comprises the steps: determining a contact point between a drill bit and a rock, building a local coordinate system, carrying out the ultrasonic signal processing and environment parameter correction, generating a rock surface three-dimensional model, predicting an adjustment angle through a PCA fusion graph neural network, and obtaining a rock surface three-dimensional model. And automatic adjustment is carried out through three-axis decoupling, and exception handling is carried out. According to the method, the problems of insufficient terrain adaptability, lack of environment compensation and low intelligence in the traditional technology are solved, multi-source data fusion modeling and dynamic self-adaptive adjustment are realized, the angle adjustment precision and efficiency of the drill bit are remarkably improved, the drill jamming risk is reduced, and a rock stratum breaking and dismantling target can be completed at low cost and high efficiency in the mineral resource exploration process.

Description

Technical Field

[0001] The present invention relates to the field of demolition measurement technology, and in particular to a three-dimensional demolition method for a drilling rig integrated with a PCA graph neural network. Background Art

[0002] During mineral resource exploration, mineral resources such as coal and iron are often buried in hard sedimentary rock, requiring rock breaking and removal to clear obstacles. During this rock breaking process, precise adjustment of the drill bit angle directly impacts drilling efficiency, equipment life, and construction safety. In traditional drilling operations, drill bit angle adjustment relies primarily on operator experience or simple inclination sensor feedback. However, the rock surface lacks obvious regularity, making it difficult to capture the three-dimensional topography of complex rock surfaces in real time. For rock formations with uneven terrain and developed fractures, the drill bit's posture cannot be dynamically adjusted to maintain the optimal drilling angle, which can easily lead to drill bit deflection or sticking. This can damage equipment and increase costs, while also compromising drilling efficiency.

[0003] To address these issues, existing technologies use auxiliary drilling peeps and lithologic detection instruments to assess the current rock formation and ultimately formulate a breaching and adjustment strategy. However, these methods lack real-time performance and, due to a lack of environmental compensation and the accumulation of instrument errors, can reduce the accuracy of breaching drilling angles. Therefore, there is an urgent need for an intelligent breaching prediction method that comprehensively considers environmental parameters and offers greater real-time performance. This method could help achieve rock breaching objectives efficiently and cost-effectively during mineral resource exploration. Summary of the Invention

[0004] Purpose of the invention: To solve the problems mentioned in the background technology, the present invention discloses a three-dimensional demolition method for drilling rigs that integrates PCA graph neural network. Rock formation data is acquired through ultrasonic scanning and signal acquisition, supplemented by dynamic sound velocity correction of temperature and humidity, and combined with octree space segmentation and MarchingCubes to reconstruct voxel space to accelerate modeling while reducing the amount of calculation. Finally, the optimal demolition angle of the drill bit is predicted and adjusted through PCA-graph neural network fusion, thereby achieving the purpose of cost-effective demolition.

[0005] Technical solution:

[0006] The present invention discloses a three-dimensional demolition method for a drilling rig integrated with a PCA graph neural network, the method comprising the following steps:

[0007] S1 determines the contact point between the drill bit and the rock surface and establishes a local coordinate system to obtain the initial pitch and yaw angles of the drill pipe;

[0008] The ultrasonic transmitter of the S2 drill bit transmits ultrasonic signals and collects reflected signals. After noise suppression processing is performed on the reflected signals, the time delay sequence is obtained from the ultrasonic receiver.

[0009] S3 corrects the ultrasonic ranging error according to the environmental parameters and the initial pitch angle, and generates a three-dimensional point cloud of the rock surface based on the corrected sound velocity and the time delay sequence;

[0010] S4 processes the three-dimensional point cloud using an octree space segmentation method, and reconstructs the voxelized space using MarchingCubes to generate a three-dimensional model of the rock surface;

[0011] Based on the three-dimensional model, S5 uses PCA-graph neural network fusion to predict the adjustment angle of the drill bit, automatically adjusts the drill bit according to the adjustment angle, and visually verifies and handles exceptions during the adjustment process.

[0012] Furthermore, the contact point between the drill bit and the rock surface is determined by pressing the drill bit tip against the rock surface through the drilling rig control console, and using the drill bit's built-in pressure sensor to detect the contact force in real time. When the pressure value is ≥50N, the system lock command is triggered to complete the physical contact confirmation.

[0013] Furthermore, the local coordinate system is established as follows:

[0014] With the drill bit contact point as the origin, the Z axis is along the mechanical axis of the drill pipe, that is, perpendicular to the rock surface, the X axis is horizontal and consistent with the forward direction of the cockpit, and the Y axis is perpendicular to the XZ plane and determined according to the right-hand rule.

[0015] Furthermore, the specific process of noise suppression is as follows:

[0016] The ultrasonic transmitter transmits pulse signals in a time-sharing manner using a time-determined multiple access (TDMA) method. Eight ultrasonic receivers arranged in a ring array for noise suppression capture reflected signals, and vibration noise is eliminated using an LMS adaptive filter. Furthermore, the ultrasonic ranging error is corrected based on environmental parameters and the initial pitch angle. A multivariate regression model is used to calculate the compensated sound velocity, obtaining the true three-dimensional position of the rock surface. The time delay sequence and initial pitch angle correction are then substituted into the time-of-flight (TOF) formula to calculate the coordinates of each point, forming an initial point cloud of the rock surface.

[0017] Furthermore, the steps for generating the three-dimensional model of the rock surface in S4 are as follows:

[0018] Using the octree space segmentation method, the entire rock space is imagined as a large cube, which is divided into 8 small cubes on average. Each small cube is checked. If the internal structure is simple, it is marked as "no subdivision required"; if it is complex, it is further divided into 8 smaller sub-cubes; the segmentation is repeated until the accuracy requirements are met; the termination condition is determined based on the real-time crushing energy consumption and the compressive strength of the rock; MarchingCubes is used to reconstruct the voxelized space: implicit surfaces are generated through radial basis function interpolation to generate triangular mesh patches.

[0019] Furthermore, the specific process of predicting the adjustment angle of the drill bit by the PCA-graph neural network fusion described in S5 is as follows:

[0020] Multi-region PCA analysis - spatial division: The drill bit is used as the center to divide the area into three levels: core area R = 10 cm, transition area R = 20 cm, and edge area R = 50 cm. PCA dimension reduction is performed through covariance matrix calculation, and feature decomposition is performed independently on each area to obtain the local normal direction N j ;

[0021] Physical constraint graph neural network: The normal vector obtained by PCA is fused with the main frequency of the vibration spectrum and then input into the graph neural network. The final normal direction and predicted angle are obtained by solving the space-time dynamics equation and performing dynamic weight fusion.

[0022] Furthermore, the three-axis decoupling automatic adjustment of the drill bit described in S5 includes a coarse adjustment stage and a fine adjustment stage: in the coarse adjustment stage, PID parameters are dynamically adjusted for the X / Y axes, and in the fine adjustment stage, laser gyroscope closed-loop feedback is used for the Z axis.

[0023] Furthermore, S5 describes an exception handling of the adjustment process, including momentum conservation verification and energy balance check: every 50ms, it is detected whether the rock acceleration exceeds the dynamic threshold, and every 5s, it is verified whether the error between the input mechanical energy and the energy consumed by rock fracture is within the allowable range. If the constraint is violated, the octree node refinement and graph neural network weight online correction are triggered.

[0024] Beneficial effects:

[0025] 1. The present invention is based on environmental compensation three-dimensional coordinate solution, dynamic sound velocity correction temperature and humidity environmental data are substituted into the multivariate regression model to improve the accuracy of predicting the optimal breaking angle in complex environments and ensure the robustness of the method.

[0026] 2. This invention combines octree space segmentation with MarchingCubes to reconstruct voxelized space to model the rock surface, intelligently manage spatial data and selective calculations, and significantly reduces the amount of calculation and lowers computing power costs while increasing modeling speed and improving the feedback efficiency of demolition strategies.

[0027] 3. The present invention uses PCA to fuse vibration spectrum data into graph neural network modeling to form a PCA-graph neural network to predict the optimal demolition angle of the drill bit and adjust it. The drill bit adjustment angle is solved by the space-time dynamics equation, which can respond to changes in rock mechanical properties in real time to provide optimal demolition angle prediction, further improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Specific flow chart of the method of the present invention;

[0029] Figure 2 This is a schematic diagram of the PCA-graph neural network fusion of the present invention;

[0030] Figure 3 This is the final prediction result diagram of the embodiment of the present invention. DETAILED DESCRIPTION

[0031] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0032] like Figure 1 As shown, this embodiment discloses a three-dimensional demolition method for a drilling rig integrated with a PCA graph neural network, comprising the following steps:

[0033] S1 determines the contact point between the drill bit and the rock surface and establishes a local coordinate system to obtain the initial pitch and yaw angles of the drill pipe;

[0034] Manual positioning: The operator presses the drill tip against the rock surface from the drill console. A built-in pressure sensor (range 0-500N, accuracy ±2%) measures contact force in real time. When the pressure reaches 50N or higher, the system locks, confirming physical contact.

[0035] Coordinate system definition: Establish a local coordinate system with the drill contact point O as the origin:

[0036] Z axis: along the mechanical axis of the drill pipe (perpendicular to the rock surface)

[0037] X-axis: horizontal direction, consistent with the direction of the cockpit

[0038] Y-axis: Determined by the right-hand rule (perpendicular to the XZ plane), the electronic horizontal inclinometer (Q / M point) synchronously measures the drill pipe pitch angle θ and yaw angle φ with an accuracy of ±0.1°.

[0039] The ultrasonic transmitter of the S2 drill bit transmits ultrasonic signals and collects reflected signals. After noise suppression processing is performed on the reflected signals, the time delay sequence is obtained from the ultrasonic receiver.

[0040] Dual-transmitter scanning: The controller activates the A / B ultrasonic transmitters (center frequency 30kHz, bandwidth ±5kHz), and transmits pulse signals in a TDMA manner (pulse width 10μs, repetition frequency 200Hz).

[0041] Array Reception and Noise Suppression Eight ultrasonic receivers arranged in a ring (15 cm apart) capture the reflected signal and eliminate vibration noise through an LMS adaptive filter:

[0042] w(n+1)=w(n)+0.02·e(n)·x(n)

[0043] The input signal x(n) is the noise reference collected by the vibration sensor, the step size μ is 0.02, and the filter order is 32.

[0044] The time difference measurement records the delay sequence of each receiving module {t a , t b ,…,t h}, the time resolution is 0.1μS.

[0045] S3 corrects the ultrasonic ranging error according to the environmental parameters and the initial pitch angle, and generates a three-dimensional point cloud of the rock surface based on the corrected sound velocity and the time delay sequence;

[0046] By real-time detection of environmental changes (temperature, humidity) and drill bit inclination angle, the error of ultrasonic ranging is automatically corrected, and the true three-dimensional position of the rock surface is finally obtained.

[0047] Dynamic sound velocity correction temperature and humidity sensor data is substituted into the multivariate regression model:

[0048] v e =331.4+0.6(T-25)+0.04(H-60) (unit: m / s)

[0049] Where T is temperature (°C), H is humidity (%RH), and the sound velocity error after compensation is less than 0.05%.

[0050] The three-dimensional point cloud is generated by calculating the coordinates of each point based on the TOF formula:

[0051]

[0052] A single scan acquires 800-1200 spatial points to form the initial point cloud of the rock surface (density 40 points / cm 2 ).

[0053] S4 processes the three-dimensional point cloud using an octree space segmentation method, and reconstructs the voxelized space using MarchingCubes to generate a three-dimensional model of the rock surface;

[0054] Octree compression:

[0055] Using a "grid" approach to intelligently manage spatial data, focusing only on important areas and ignoring irrelevant details, significantly reducing computational effort. Octree compression and grid storage imagine the entire rock space as a large cube, evenly divided into eight smaller cubes. Each cube is examined. If the internal structure is simple (e.g., entirely solid rock), it is marked as "no need for subdivision." If it is complex (e.g., with cracks or holes), it is further divided into eight smaller sub-cubes. This division is repeated until the required accuracy is met.

[0056] Modeling acceleration: Selective calculation, only processing "grids that require details" (such as fine small grids near cracks and holes) during modeling; ignoring "simple large grids" (such as solid rock areas) and directly filling them with homogeneous blocks.

[0057] In this embodiment, the initial cube edge length δ=5cm is defined by the node in the octree space partitioning, and the recursive partitioning is performed to the minimum δ=0.5cm. The termination condition is as follows:

[0058]

[0059] Among them, E break is the real-time crushing energy consumption, calculated by the torque sensor, σ c is the compressive strength of rock.

[0060] Curvature screening: Calculate the eigenvalues ​​of the node point cloud covariance matrix λ1≥λ2≥λ3, preserving the curvature The data compression rate is greater than 65%.

[0061] MarchingCubes reconstructs the voxel space (voxel side length ε = 1 cm) and generates an implicit surface through radial basis function interpolation:

[0062]

[0063] Where f(x, y, z) is the implicit surface function, Ra is the surface roughness (calibrated by laser profilometer), σ is the Gaussian kernel width, and p is the surface roughness. i After compressing the point cloud data points, a triangular mesh (about 30,000 faces) is generated, which is a three-dimensional model of the rock surface. The modeling time is less than 3 seconds.

[0064] S5 is based on the three-dimensional model, such as Figure 2 As shown, PCA-graph neural network fusion is used to predict the adjustment angle of the drill bit, and the drill bit is automatically adjusted by three-axis decoupling according to the adjustment angle, and the adjustment process is visually verified and exceptions are handled.

[0065] Multi-region PCA analysis-spatial division: The drill bit was used as the center to divide the area into three levels (core area R = 10 cm, transition area R = 20 cm, and edge area R = 50 cm).

[0066] Covariance matrix calculation (PCA dimensionality reduction):

[0067]

[0068] Among them, Cov j Regional covariance matrix: describes the point cloud in the local area R j Spatial distribution characteristics of n j is the number of regional points: neighborhood R j The number of valid points in the interior; p i is the point coordinate; c j For regional centers: R j It is a local neighborhood: a spherical area with a radius of r = 15 mm and the target point as the center.

[0069] Perform feature decomposition on each region independently to obtain the local normal direction N j .

[0070] N j : PCA normal direction: Cov j The eigenvector corresponding to the minimum eigenvalue

[0071] Physical Constraint Graph Neural Network

[0072] Node feature initialization: Fusion of PCA normal vector and vibration spectrum main frequency f vib

[0073]

[0074] Among them, h i (0) is the initial state of the node: the initial embedding vector of the graph neural network node; MLP is a multi-layer perceptron: a 3-layer fully connected network; f vib : The main frequency of the vibration spectrum (scalar, unit: Hz), measured by the sensor.

[0075] Solve by space-time dynamics equation:

[0076]

[0077] Among them, h i is the embedding vector of node i; φ θ is a neural network defined by parameters θ; W msg is the weight matrix of message passing; is the fracture energy pair node embedding h i The gradient of H B =10 {(f-80) / 35} is the rock hardness, and the differential equation is solved by CUDA acceleration (Δ t=0.1s).

[0078] Dynamic weight fusion obtains the final normal direction:

[0079]

[0080] Among them, ω j is the local normal N j Fusion weight of d j is the local area R j Center C j The Euclidean distance to the target point; n is the total number of points in the point cloud, used for distance normalization; k j is the local area R j The number of points n j : Reflects regional reliability. N final : The final normal direction after weighted fusion.

[0081] Output predicted angle θ pred =[Δφ, Δθ],

[0082] like Figure 3 As shown, it is converted into hydraulic actuator control instructions by the ROS system.

[0083] Three-axis decoupling automatic adjustment and verification:

[0084] The hierarchical control strategy is divided into coarse adjustment stage and fine adjustment stage. In the coarse adjustment stage (X / Y axis): PID parameters are dynamically adjusted:

[0085] K p =0.8·(1+0.05(H-5)), H∈[1,10]

[0086] Response time is less than 1.5s, and the angular deviation converges to ±2°. During the fine-tuning phase (Z-axis), laser gyroscope closed-loop feedback and hydraulic servo positioning accuracy are ±0.05°.

[0087] The black solid line represents the current direction vector D current =[sinθcosφ, sinθsinφ, cosθ]

[0088] The red dotted line represents the target normal N final

[0089] Color mapping: Deviation <2° (green), 2°-5° (yellow), >5° (red)

[0090] Efficiency prediction:

[0091]

[0092] When η>95%, the sound and light ready signal is triggered.

[0093] The exception handling mechanism includes momentum conservation verification and energy balance check: every 50ms, the rock acceleration is checked to see if it exceeds the dynamic threshold, and every 5s, the error between the input mechanical energy and the energy consumed by the rock fracture is verified to be within the allowable range. If the constraint is violated, the octree node refinement and graph neural network weight online correction are triggered:

[0094] Momentum conservation verification: check every 50ms

[0095] Among them, ||d 2 h / dt 2 || is the acceleration norm of the rock; τ is the cumulative stress time

[0096] Prevent numerical explosion and detect in real time whether the rock acceleration exceeds the physical reasonable range. Dynamic threshold: The harder the rock (H B The longer the force is applied (τ increases) or the longer the force is applied (τ increases), the lower the maximum acceleration allowed (to prevent sudden loss of control after long-term force application). If the acceleration exceeds the limit, it means that the current simulation may produce non-physical and drastic deformation (such as rocks instantly crushing into dust), and immediate intervention is required.

[0097] Energy balance check: Verify every 5s |∫τ·ωdt-E break |≤0.12E break

[0098] Where τ is the input torque; ω is the angular velocity during rock deformation or fracture; E break is the critical energy threshold required for rock fracture.

[0099] Energy conservation verification: Ensure that the mechanical energy input to the system (the integral of torque τ and angular velocity ω) is equal to the energy consumed by rock fracture (E break ). Fault tolerance threshold: 12% error is allowed (actual physical systems also have energy loss), but if it exceeds, it is judged as energy leakage or excessive loss. If the energy is unbalanced, it means that the model may deviate from the physical law for a long time (such as the error in the crack propagation speed), and global or local correction is required. When the constraint is violated, the octree node refinement (δ new =δ old / 2) and online correction of GNN weights. Octree node refinement doubles the mesh resolution in the abnormal area. Combined with online correction of GNN weights, the parameters of the graph neural network are dynamically adjusted. By refining the mesh, tiny cracks or stress concentration points are captured. The GNN is fine-tuned based on new data to prevent incorrect predictions from spreading to other areas.

[0100] The above description of the embodiments enables one skilled in the art to implement or use the present invention. Various modifications to the embodiments will be readily apparent to those skilled in the art. The general principles of the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should encompass the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional demolition method for drilling rigs integrated with PCA graph neural network, characterized in that: The method comprises the following steps: S1 determines the contact point between the drill bit and the rock surface and establishes a local coordinate system to obtain the initial pitch and yaw angles of the drill pipe; The ultrasonic transmitter of the S2 drill bit transmits ultrasonic signals and collects reflected signals. After noise suppression processing is performed on the reflected signals, the time delay sequence is obtained from the ultrasonic receiver. S3 corrects the ultrasonic ranging error according to the environmental parameters and the initial pitch angle, and generates a three-dimensional point cloud of the rock surface based on the corrected sound velocity and the time delay sequence; S4 processes the three-dimensional point cloud using an octree space segmentation method, and reconstructs the voxelized space using MarchingCubes to generate a three-dimensional model of the rock surface; Based on the three-dimensional model, S5 uses PCA-graph neural network fusion to predict the adjustment angle of the drill bit, automatically adjusts the drill bit according to the adjustment angle, and visually verifies and handles exceptions during the adjustment process.

2. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 1 is characterized in that: The contact point between the drill bit and the rock surface is determined by pressing the drill bit tip against the rock surface through the drilling rig control console, and using the drill bit's built-in pressure sensor to detect the contact force in real time. When the pressure value is ≥50N, the system lock command is triggered to complete the physical contact confirmation.

3. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 2 is characterized in that: The local coordinate system is established as follows: With the drill bit contact point as the origin, the Z axis is along the mechanical axis of the drill pipe, that is, perpendicular to the rock surface, the X axis is horizontal and consistent with the forward direction of the cockpit, and the Y axis is perpendicular to the XZ plane and determined according to the right-hand rule.

4. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 3 is characterized in that: The specific process of noise suppression is as follows: The ultrasonic transmitter transmits pulse signals in a time-sharing manner using TDMA. The eight ultrasonic receivers arranged in a ring array for noise suppression capture the reflected signals and eliminate vibration noise through an LMS adaptive filter.

5. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 4 is characterized in that: The ultrasonic ranging error is corrected according to the environmental parameters and the initial pitch angle: the compensated sound velocity is calculated through a multivariate regression model to obtain the true three-dimensional position of the rock surface, the time delay sequence and the initial pitch angle are substituted into the TOF formula to calculate the coordinates of each point and form the initial point cloud of the rock surface.

6. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 5 is characterized in that: The steps for generating the three-dimensional model of the rock surface described in S4 are as follows: Using the octree spatial segmentation method, the entire rock space is imagined as a large cube, which is then divided into eight smaller cubes on average. Each small cube is inspected and marked as "no subdivision required" if the internal structure is simple. If it is complex, it is further divided into eight smaller sub-cubes. The segmentation is repeated until the accuracy requirements are met. The termination condition is determined based on the real-time crushing energy consumption and the compressive strength of the rock. MarchingCubes is used to reconstruct the voxelized space: implicit surfaces are generated through radial basis function interpolation to generate triangular mesh patches.

7. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 6 is characterized in that: The specific process of predicting the adjustment angle of the drill bit by PCA-graph neural network fusion described in S5 is as follows: Multi-region PCA analysis - spatial division: The drill bit is used as the center to divide the area into three levels: core area R = 10 cm, transition area R = 20 cm, and edge area R = 50 cm. PCA dimension reduction is performed through covariance matrix calculation, and feature decomposition is performed independently on each area to obtain the local normal direction N j ; Physical constraint graph neural network: The normal vector obtained by PCA is fused with the main frequency of the vibration spectrum and then input into the graph neural network. The final normal direction and predicted angle are obtained by solving the space-time dynamics equation and performing dynamic weight fusion.

8. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 7 is characterized in that: The three-axis decoupling automatic adjustment of the drill bit as described in S5 includes a coarse adjustment stage and a fine adjustment stage: in the coarse adjustment stage, the X / Y axes are dynamically adjusted using PID parameters, and in the fine adjustment stage, the Z axis is closed-loop feedback using a laser gyroscope.

9. The three-dimensional demolition method of a drilling rig integrated with a PCA graph neural network according to claim 8 is characterized in that: S5 describes the abnormal handling of the adjustment process, including momentum conservation verification and energy balance check: every 50ms, the rock acceleration is checked to see if it exceeds the dynamic threshold, and every 5s, the error between the input mechanical energy and the energy consumed by rock fracture is verified to be within the allowable range. If the constraint is violated, the octree node refinement and graph neural network weight online correction are triggered.

Images