An underground tunneling inertial navigation system and method for multi-source information fusion

Abstract

The present invention discloses an underground tunneling inertial navigation system and method for multi-source information fusion, which relates to the technical field of underground tunneling navigation. This system includes multiple functional units such as a shape memory alloy self-calibrating inertial module and a magnetorheological fluid intelligent vibration damping platform. Through technical means such as hardware self-calibration, full-frequency vibration suppression, and dynamic weight allocation, multi-source data fusion processing is achieved. In terms of inertial navigation, the positioning error within 8 hours is less than 5 cm, which is significantly improved compared with traditional systems; the positioning accuracy under complex working conditions reaches ±10 cm / 100 m, which is significantly improved compared with the fixed-weight fusion method. When multiple machines cooperate, the lag time is <50 ms, and the support efficiency is increased by 35%. At the same time, it can respond to geological anomalies in real time, the efficiency of passing through faults is increased by 60%, the mis-cutting rate of coal and rock is reduced by 70%, and the navigation continuity is ensured when the sensor fails. This invention effectively solves the problem of underground tunneling navigation and greatly improves the safety, efficiency, and intelligent level of underground tunneling operations.

Description

Technical Field

[0001] The present invention relates to the technical field of underground tunneling navigation, and specifically to an underground tunneling inertial navigation system and method for multi-source information fusion. Background Art

[0002] In underground tunneling operations, accurate navigation and positioning are crucial for ensuring construction safety, improving tunneling efficiency, and guaranteeing project quality. Traditional underground navigation systems mainly rely on single sensors or simple combinations, which have many limitations in complex underground environments. For example, although the inertial navigation system can work independently, its positioning error will continuously accumulate over time; the measurement accuracy of lidar will be severely affected in harsh environments such as dust and water vapor; the performance of vision sensors will drop significantly under low-light conditions. Moreover, existing systems are difficult to perceive geological structure changes in real time, unable to adjust the tunneling path in a timely manner, easily leading to over-excavation and under-excavation phenomena, resulting in resource waste and safety hazards. In addition, the large communication delay and low synchronization accuracy during multi-machine collaborative operations seriously restrict the intelligent development of underground construction. Therefore, it is urgent to develop an underground tunneling inertial navigation system and method for multi-source information fusion that can adapt to complex underground environments, has high precision, and is intelligent.

[0003] In view of this, the present application is specifically proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide an underground tunneling inertial navigation system and method for multi-source information fusion to solve the problems mentioned in the above background art.

[0005] To solve the above technical problems, an underground tunneling inertial navigation system for multi-source information fusion provided by the present invention includes:

[0006] Shape memory alloy self-calibration inertial module: Integrating a nickel-titanium alloy cantilever beam, triggering mechanical deformation recovery through a heating power of 1.5W, correcting the zero-offset errors of a MEMS accelerometer (accuracy ±0.1g) and a gyroscope (angle random walk 0.01 degrees per square root hour), with a self-calibration time ≤ 80 ms / time, a working temperature range of -40°C to 60°C, and a pressure resistance ≥ 50 MPa; through hardware-level self-calibration technology, reducing the accelerometer zero-drift from 0.2g to 0.03g and the gyroscope zero-drift from 0.5° / h to 0.08° / h, suppressing the accumulation of inertial device errors from the physical level, achieving a positioning error < 5 cm in 8 hours, and the improvement amplitude compared with the traditional system far exceeding the original set value;

[0007] Magnetorheological fluid intelligent vibration reduction platform: It includes 3 Lord RD-1005-3 shock absorbers filled with MRF-132DG magnetorheological fluid, with a response time ≤ 0.5 ms, the viscosity is adjusted by a current ranging from 0 A to 1 A (both 0 A and 1 A are included), the vibration attenuation rate ≥ 92% (20 - 300 Hz), the inertial navigation output noise density is reduced to ≤ 5 micrograms per square root hertz, and the stress is measured in real time to dynamically adjust the vibration reduction parameters; The full-frequency band vibration suppression technology combined with stress measurement reduces the noise density by 90%, improves the stability of inertial data output by 5 times, and optimizes the vibration reduction effect through stress feedback to ensure the navigation accuracy in complex vibration environments;

[0008] Variational Bayesian network processing unit: It fuses 100 Hz inertial navigation data, 10 Hz lidar point cloud, 50 Hz ultrasonic radar ranging, and 30 fps visual images in real time, dynamically allocates sensor weights through posterior probability, and when the dust concentration > 1500 mg / m³, the weight of the ultrasonic radar ≥ 0.7; Based on the dynamic weight allocation algorithm of environmental parameters, it automatically switches the optimal data source under complex working conditions such as dust and low light, and the positioning accuracy is significantly improved compared with the fixed weight fusion method, achieving a dynamic positioning accuracy of ±10 cm / 100 m;

[0009] 5G URLLC communication module: It supports the 2.6 GHz frequency band, the communication delay < 8 ms, the reliability ≥ 99.999%, realizes multi-machine clock synchronization through the time-sensitive network with a deviation < 500 ns, and the single-device bandwidth ≥ 50 Mbps; The ultra-low latency communication and high-precision clock synchronization technology enable the lag time of multi-machine collaborative operation < 50 ms, the bolt positioning accuracy reaches ±10 cm, and the support efficiency is increased by 35%, realizing the precise coordination of "heading - support" parallel operation;

[0010] Digital twin geological modeling unit: Based on lidar with an accuracy of ±3 cm and ground penetrating radar with a detection depth of 15 m, it constructs a three-dimensional point cloud model, and combines reinforcement learning to output cutting path adjustment instructions with a step size of 0.5° and a response time < 150 ms; The real-time geological modeling and path optimization technology enables the response time of the cutting head angle adjustment < 180 ms, the efficiency of passing through faults is increased by 60%, the mis-cutting rate of coal and rock is reduced by 70%, and the ash content of coal quality is reduced by 4% compared with the traditional scheme;

[0011] Conditional generative adversarial network cGAN data filling module: It inputs timestamps and known sensor data to generate inertial navigation data including acceleration and angular velocity, and the training data volume ≥ 1.2×10 6 The mean square error (MSE) of the generated data ≤ 0.01 g 2; A data-driven missing data filling technique that generates inertial navigation data with an accuracy of ±0.01 g when the sensor fails, maintaining the system positioning error within ±10 cm / 100 m, significantly improving compared to traditional interpolation methods and ensuring navigation continuity under complex working conditions;

[0012] Tactile perception unit: Integrates 128 piezoelectric film tactile sensors with a resolution of 0.1 N / mm 2 , measures force in real time to sense the distribution of cutting resistance, and evaluates the load of the cutting head by measuring torque; The high-precision tactile perception technology monitors the cutting resistance and torque in real time, providing direct load feedback for optimizing the cutting path, reducing the number of mis-cutting of the roof and floor in thin coal seam cutting from 5 times per hour to 0.5 times per hour, and improving the purity of coal quality.

[0013] Furthermore, in the shape memory alloy self-calibrating inertial module, the surface of the cantilever beam is coated with an alumina thermal insulation layer with a thickness of 50 μm and a thermal conductivity of 0.15 W / (m·K), and is connected to the edge computing node through a cable with a temperature resistance of 80 °C, and the computing power ≥ 200 TOPS; The design of the alumina thermal insulation layer and the high-temperature resistant cable enables the inertial module to work stably in a high-temperature environment of 60 °C, broadens the system's environmental adaptability range, reduces the influence of high temperature on the accuracy of inertial devices, and ensures reliable operation in high-humidity and high-temperature environments such as deep mines.

[0014] Furthermore, in the magnetorheological fluid intelligent vibration damping platform, it is externally wrapped with an aerogel thermal insulation layer with a thermal conductivity of 0.013 W / (m·K), collects vibration data at 1000 Hz in real time, dynamically adjusts the current of the shock absorber through the PWM module to achieve full-frequency vibration suppression, and measures the fluid pressure at the same time to optimize the working parameters of the magnetorheological fluid; The combination of the aerogel thermal insulation layer and the fluid pressure measurement ensures that the shock absorber can achieve a vibration attenuation rate of more than 92% within the temperature range of -40 °C to 60 °C, improves the robustness of the system under extreme temperatures, and extends the service life of the equipment.

[0015] Furthermore, in the variational Bayesian network processing unit, a built-in fault detection module automatically switches to the fusion positioning of the ultrasonic radar and the magnetic gradient tensor instrument when the number of valid point clouds of the lidar is <5000 for 5 consecutive frames, and the positioning accuracy is ±10 cm / 100 m; The intelligent fault detection and redundant positioning technology automatically switches to the fusion mode of the ultrasonic radar + magnetic gradient tensor instrument when the lidar fails due to dust interference, and the positioning accuracy still remains ±10 cm / 100 m, improving the fault tolerance ability by 90% compared to the problem of the error explosion and growth after the failure of traditional single sensors.

[0016] Furthermore, in the digital twin geological modeling unit, a geological radar with a resolution of 0.1 m is integrated. The accuracy of identifying the dielectric constant difference at the coal-rock interface is ±0.5. A convolutional neural network is combined to detect faults with an identification accuracy of ≥95%. The cutting energy consumption is evaluated by measuring the mechanical power. The high-precision geological identification technology combined with the mechanical power measurement can give an early warning of fault activities 2 hours in advance, and at the same time optimize the cutting power distribution, reduce the equipment energy consumption by 30%, and improve the tunneling efficiency and safety.

[0017] Furthermore, in the conditional generative adversarial network cGAN data filling module, a 5-layer fully connected generator and a 4-layer discriminator network are adopted. The loss function is equal to the Wasserstein distance + gradient penalty, and the peak signal-to-noise ratio (PSNR) of the generated data is ≥32 dB. The advanced loss function and deep network architecture improve the generalization ability of the cGAN model by 40%, and high-precision data filling can be achieved under 20 complex working conditions, ensuring the reliability of the system in extreme environments.

[0018] Furthermore, the tactile perception unit is linked with the cutting head manipulator of the tunneling equipment. The cutting head speed and feed speed are dynamically adjusted by measuring the force and torque, and the mechanical efficiency is measured synchronously to optimize the cutting parameters. The load perception and parameter adaptive technology can increase the load uniformity of the cutting head by 40%, improve the mechanical efficiency by 25%, reduce tool wear, and extend the equipment maintenance cycle.

[0019] A usage method of an underground tunneling inertial navigation system with multi-source information fusion includes the following steps:

[0020] Data acquisition: Data is collected through a shape memory alloy self-calibrating inertial module (100 Hz), lidar (10 Hz), geological radar (once every 5 seconds), and tactile sensors (5 ms response), including position, attitude, geological structure, and cutting resistance information. The multi-source high-frequency data acquisition technology ensures the real-time acquisition of environmental information and equipment status, providing comprehensive data support for intelligent decision-making.

[0021] Preprocessing: The inertial data is subjected to magnetorheological fluid damping and adaptive notch filtering, and the noise density is reduced to less than or equal to 5 micrograms per square root hertz. The tactile data is subjected to force-electric signal conversion to measure the force with a resolution of 0.1 N / mm². The multi-stage filtering and signal conversion technology improves the data quality, reduces the inertial data noise density by 90%, and shortens the tactile signal response time to 5 ms, laying a foundation for the subsequent fusion algorithm.

[0022] Dynamic error suppression: Trigger the SMA heating circuit, 1.5W, 10ms per time, for hardware self-calibration. Combine the variational Bayesian network to online evaluate the random walk parameters of inertial devices, and calibrate the cumulative error of the inertial navigation by ±10 cm every 10 seconds; Combine hardware self-calibration with algorithm compensation to correct the cumulative error of the inertial navigation in real time, avoid the infinite growth of errors over time, and achieve a positioning error of <5 cm in 8 hours, meeting the millimeter-level accuracy requirements of unmanned tunneling.

[0023] Multi-source fusion: The variational Bayesian network dynamically allocates sensor weights according to environmental parameters (dust concentration, illuminance). In a dusty environment, ultrasonic radar is preferred, with a weight ≥0.7. In a normal environment, lidar is preferred, with a weight ≥0.6; The environment-adaptive fusion algorithm automatically selects the optimal data source under complex working conditions, significantly improving the positioning accuracy, reducing manual intervention, and enhancing the system robustness.

[0024] Multi-robot collaboration: Transmit the position information of multiple robots (accuracy ±5 cm) through the 5G URLLC network (delay <8 ms), and optimize the actions of the support robots using the federated learning-blockchain architecture. The time-sensitive network achieves action synchronization (deviation <1 μs); Low-latency communication and distributed optimization technology enable multi-robot sub-meter-level collaborative operation, improving the support efficiency by 30% and promoting the unmanned and intelligent underground construction.

[0025] Geological adaption: The digital twin model updates the three-dimensional geological structure in real time. When encountering a fault with a dielectric constant mutation >5, the reinforcement learning generates an obstacle avoidance path, and the cutting head angle adjustment step is 0.5°, with a response time <180 ms; Combine digital twin and reinforcement learning to achieve automatic identification and rapid response to geological anomalies, improving the fault-crossing efficiency by 60% and reducing the material waste caused by overexcavation and under-excavation by 30%.

[0026] Data filling: When the sensor fails, the conditional generative adversarial network cGAN data filling module inputs the timestamp and known data to generate 60 seconds of continuous inertial navigation data. After filling, the positioning error ≤±10 cm / 100 m; The data filling technology based on deep learning maintains the continuous operation of the system when the sensor fails, significantly improving compared with the traditional interpolation method, and ensuring the continuity of operations.

[0027] Energy efficiency optimization: Dynamically adjust the device operation parameters by measuring the mechanical power and fluid pressure, reducing the system energy consumption by more than 30%; The energy efficiency optimization technology combines real-time power and pressure monitoring to achieve precise matching of device load and energy consumption, reducing energy waste and lowering the underground operation cost.

[0028] Furthermore, in the data acquisition, the tactile sensors cover a distance of 10 meters from the lidar, the lidar point cloud density ≥ 0.1 points / m², the ground penetrating radar detection depth ≥ 15m, and the tactile sensors cover ≥ 80% of the surface area of the cutting head; high-density data acquisition and comprehensive tactile coverage ensure the accurate perception of the geological structure and cutting load, providing a reliable basis for optimizing the cutting path.

[0029] Furthermore, in the preprocessing, the adaptive notch filter algorithm suppresses 50Hz power frequency noise, and the magnetorheological fluid damping platform attenuates the vibration energy of 20 - 300Hz by ≥ 92%; the targeted noise suppression technology improves the purity of inertial data and ensures the stability of the navigation system in a strong vibration and high electromagnetic interference environment.

[0030] Furthermore, in the dynamic error suppression, the update frequency of the variational Bayesian network is synchronized with the inertial navigation at 100Hz, the gradient compression rate of the federated learning model ≥ 80%, and the accuracy of the blockchain timestamp is 1μs; high-frequency algorithm updates and efficient data transmission ensure the real-time optimization of the error model, the non-tampering of data storage, and improve the system error suppression efficiency and data credibility.

[0031] Furthermore, in the multi-source fusion step, in a low-light environment with illuminance < 10 lux, the weight of the inertial navigation is automatically increased to ≥ 0.8, combined with the magnetic gradient tensor instrument for assisted positioning, with an accuracy of ±10 cm / 100m; the multi-sensor redundant fusion technology maintains high-precision positioning in a low-light environment, broadens the applicable scenarios of the system, and improves the adaptability to complex environments.

[0032] Furthermore, in the multi-machine collaboration, the uplink throughput of the 5G URLLC network ≥ 1.2 Gbps, supports ≥ 20 devices to be online simultaneously, the positioning accuracy of the support robot's anchor bolts is ±10 cm, and the lag time < 50 ms; high-bandwidth communication and multi-device collaboration technology enable real-time linkage of large-scale devices, improve the efficiency and safety of underground operations, and promote the upgrade of intelligent construction.

[0033] Furthermore, in the geological adaptation, the digital twin model updates the roadway point cloud every 2 seconds, and the reinforcement learning algorithm takes the coal-rock mis-cutting rate (deduct 10 points for every 1 cm of mis-cutting) and path smoothness (curvature > 10m -1 deduct 5 points) as the optimization objectives; high-frequency model updates and refined optimization objectives ensure the real-time and accuracy of the cutting path, reduce the coal-rock mis-cutting rate by 70%, and improve the quality of coal mining.

[0034] Furthermore, in the data filling, the conditional generative adversarial network cGAN data filling module's training data covers ≥ 20 working conditions (dust, faults, vibrations, etc.), the peak signal-to-noise ratio between the generated data and the real data ≥ 32 dB, and the mean square error ≤ 0.008g 2Large-scale training data and strict performance metrics enable the cGAN model to achieve high-precision data filling even under extreme working conditions, ensuring the reliability and robustness of the navigation system.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. High-precision positioning and error suppression: By means of a shape memory alloy self-calibrating inertial module, the mechanical deformation recovery of a nickel-titanium alloy cantilever beam is triggered by a heating power of 1.5W, which can quickly correct the zero-offset errors of the MEMS accelerometer and gyroscope. The self-calibration time is short, and the working temperature and pressure resistance range are wide, effectively suppressing the error accumulation of inertial devices at the physical level. The positioning error within 8 hours is less than 5 cm, and the improvement amplitude compared with the traditional system far exceeds the original set value, providing a high-precision and long-time stable positioning basis for underground tunneling operations.

[0037] 2. Enhanced adaptability to vibration environment: The magnetorheological fluid intelligent vibration damping platform adopts 3 Lord RD-1005-3 type shock absorbers and fills MRF-132DG magnetorheological fluid, with rapid response. It can adjust the viscosity through current, and the vibration attenuation rate in the frequency band of 20 - 300 Hz exceeds 92%. By combining stress measurement to dynamically adjust the vibration damping parameters, the inertial navigation output noise density is reduced by 90%, and the stability of inertial data output is increased by 5 times, greatly ensuring the navigation accuracy in complex vibration environments and ensuring the reliable operation of the system under harsh working conditions.

[0038] 3. Precise positioning in complex working conditions: The variational Bayesian network processing unit can fuse multi-source data in real time and dynamically allocate sensor weights according to environmental parameters (such as dust concentration, illuminance). When the dust concentration is greater than 1500 mg / m 3 ³, the weight of the ultrasonic radar ≥ 0.7; in a normal environment, the weight of the lidar ≥ 0.6. This technology automatically switches the optimal data source under complex working conditions, and the positioning accuracy is improved more than that of the fixed-weight fusion method, achieving a dynamic positioning accuracy of ±10 cm / 100 m, effectively solving the positioning problem in complex underground environments.

[0039] 4. Efficient multi-machine collaborative operation: The 5G URLLC communication module supports the 2.6 GHz frequency band, with a communication delay less than 8 ms and a reliability as high as 99.999%. It realizes multi-machine clock synchronization (the deviation is less than 500 ns) through the time-sensitive network. The single-device bandwidth is sufficient, making the lag time of multi-machine collaborative operation < 50 ms, the bolt positioning accuracy reaches ±10 cm, and the support efficiency is increased by 35%, effectively promoting the precise coordination of "tunneling - support" parallel operations and improving the underground construction efficiency and safety.

[0040] 5. Geological Adaptation and Efficient Tunneling: The digital twin geological modeling unit constructs a three-dimensional point cloud model based on lidar and ground-penetrating radar, and outputs cutting path adjustment instructions in combination with reinforcement learning. The step size is accurate to 0.5°, and the response time is less than 150 ms, enabling rapid response to geological anomalies. In practical applications, the fault-crossing efficiency is increased by 60%, the coal-rock mis-cutting rate is reduced by 70%, and the ash content of coal quality is reduced by 4% compared with the traditional scheme, significantly improving the tunneling efficiency, reducing resource waste, and enhancing the quality of coal mining.

[0041] 6. Ensuring Navigation Continuity: The conditional generative adversarial network cGAN data filling module generates inertial navigation data by inputting timestamps and known sensor data when the sensor fails. The training data volume is large, and the generated data accuracy is high, with the mean square error ≤ 0.01g². It maintains the system positioning error within ±10 cm / 100 m, significantly improving compared with the traditional interpolation method, ensuring navigation continuity under complex working conditions, and avoiding operation interruption caused by sensor failures.

[0042] 7. Optimizing Cutting Parameters and Improving Coal Quality: The tactile perception unit integrates 128 high-resolution piezoelectric film tactile sensors to measure the cutting resistance distribution in real time and evaluate the load of the cutting head. It is linked with the cutting head manipulator of the tunneling equipment to dynamically adjust the cutting head speed and feed speed, optimizing the cutting parameters. The number of times of mis-cutting the roof and floor in thin coal seam cutting is reduced from 5 times per hour to 0.5 times per hour, improving the purity of coal quality, reducing equipment wear, and extending the equipment maintenance cycle. Description of the Drawings

[0043] Figure 1 is the principle block diagram of an underground tunneling inertial navigation system with multi-source information fusion;

[0044] Figure 2 is the flow chart of the usage method of an underground tunneling inertial navigation system with multi-source information fusion. Detailed Embodiments

[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] Please refer to Figure 1-2 , the present invention provides a technical solution: an underground tunneling inertial navigation system and method with multi-source information fusion, including:

[0047] I. System Hardware Setup (taking a certain roadheader used in underground tunneling as an example):

[0048] 1. Configuration of the inertial navigation module:

[0049] Shape Memory Alloy (SMA) Self-Calibrating Inertial Measurement Unit (IMU): The IMU is installed on the top of the roadheader cab and includes a nickel-titanium alloy cantilever beam (phase transition temperature 40°C, geometric dimensions 50 mm × 2 mm × 0.5 mm, resistance value 1.2 Ω), with an alumina thermal insulation layer plated on its surface, thickness 50 μm, thermal conductivity 0.15 W / (m·K), to isolate environmental temperature interference. MEMS inertial devices are integrated at both ends of the cantilever beam: The accelerometer selects ADXL345, range ±16g, noise density 3.5 mg per square root hertz, zero drift coefficient <0.05g / °C, meeting the measurement accuracy of ±0.1g; The gyroscope selects MPU-6050, range ±2000° / s, angular random walk 0.01° per square root hour, zero drift <0.1° / h. It is connected to the edge computing node (ECU, model NVIDIA Jetson AGX Orin, computing power 200 TOPS) through a shielded cable (characteristic impedance 50 Ω) that can withstand a temperature of 80°C. The SMA heating circuit is powered by the 5V power supply of the ECU, heating power 1.5W (voltage 12V, current 1.25A), single self-calibration period 80ms (including cooling time), adapting to a temperature range of -40°C to 60°C and a pressure-resistant environment of 50 MPa.

[0050] Example: When the IMU detects that the temperature rises to 60°C, the SMA heating circuit automatically triggers calibration, heats up to the phase transition temperature within 10ms, and the zero drift correction is completed after 80ms. The zero drift of the accelerometer drops from 0.2g to 0.03g, and the zero drift of the gyroscope drops from 0.5° / h to 0.08° / h.

[0051] Magnetorheological Fluid (MRF) Intelligent Vibration Damping Platform:

[0052] Integrate 3 Lord RD-1005-3 type shock absorbers between the IMU and the roadheader frame, filled with MRF-132DG magnetorheological fluid (density 2.8 g / cm³, yield strength 100 kPa at 1A current), externally wrapped with an aerogel thermal insulation layer, thickness 10 mm, thermal conductivity 0.013 W / (m·K). The response time of the shock absorber ≤0.5ms, adjusts the viscosity through a current ranging from 0A to 1A, including both 0A and 1A, achieving a vibration attenuation rate of 92% in the 20 - 300Hz frequency band, suppressing the inertial navigation output noise density from 50 μg per square root hertz to 5 μg per square root hertz (100Hz bandwidth). The ECU real-time collects 1000Hz vibration data through the PWM module (frequency 10kHz) and dynamically adjusts the shock absorption parameters.

[0053] Example: When the vibration amplitude of the cutting arm reaches 5g, the ECU adjusts the shock absorber current to 1A, the MRF viscosity rises to 10 Pa·s, the 200 Hz vibration component decays by 95%, and the inertial navigation output noise density drops suddenly from 50 micrograms per square root hertz to 5 micrograms per square root hertz, ensuring the stability of inertial data.

[0054] 2. Multi-sensor integration scheme:

[0055] Environmental perception sensor group:

[0056] Vision and ranging unit: A 16-line lidar (Velodyne VLP-16) is installed at the front end, with a scanning frequency of 10 Hz, a ranging accuracy of ±3 cm, and a point cloud density of 0.1 point / m² at a distance of 10 m; a panoramic camera (Basler acA2500-14gm), with a resolution of 2592×1944, equipped with an 850 nm near-infrared filter, and a frame rate of 30 fps; an ultrasonic phased array radar (operating frequency 40 kHz), with a scanning range of 360°×120°, a ranging accuracy of ±3 cm, and suitable for a dust environment of 2000 mg / m³.

[0057] Geological exploration unit: A MALA ProEx geological radar is integrated into the cutting arm, with a center frequency of 200 MHz, a detection depth of 15 m, a layer thickness resolution of 0.1 m, and a dielectric constant measurement accuracy of ±0.5; 128 channels of Tekscan piezoelectric film tactile sensors are deployed on the surface of the cutting head, with a resolution of 0.1 N / mm² and a response time of 5 ms, to real-time sense the distribution of cutting resistance.

[0058] Example: When the cutting head contacts sandy mudstone, the tactile sensor array detects an average resistance of 15 N / mm 2 , and the ECU adjusts the cutting head speed to 1500 rpm accordingly, with an efficiency improvement of 40% compared to the traditional empirical adjustment.

[0059] II. System software architecture and algorithm implementation:

[0060] 1. Variational Bayesian network (VBN) dynamic fusion algorithm: The VBN processing unit receives IMU data at a frequency of 100 Hz, lidar point clouds at 10 Hz, ultrasonic radar ranging at 50 Hz, and visual images at 30 fps in real time, and realizes dynamic allocation of sensor weights through the following steps:

[0061] 1.1 Noise modeling: Establish a random walk model for the zero-bias drift of inertial devices and a Gaussian mixture model for the ranging error of lidar.

[0062] 1.2 Weight calculation: Estimate the dust concentration based on the lidar echo intensity. When it is > 1500 mg / m³, the weight of the ultrasonic radar is assigned 0.7, the inertial navigation weight is 0.2, and the vision weight is 0.1; in a normal environment (illuminance > 10 lux, dust < 500 mg / m³), the lidar weight is 0.6, the vision weight is 0.3, and the inertial navigation weight is 0.1;

[0063] 1.3 Fault detection: When the number of valid lidar point clouds in 5 consecutive frames is < 5000 (spot scattering rate > 30%), automatically switch to the fusion positioning of the ultrasonic radar and the magnetic gradient tensor instrument, with a positioning accuracy of ±10 cm / 100 m.

[0064] Example: When entering a roadway with a dust concentration of 2000 mg / m³, VBN increases the weight of the ultrasonic radar to 0.7, and the positioning accuracy after fusion is improved from ±20 cm / 100 m of the traditional scheme to ±10 cm / 100 m, effectively solving the positioning failure problem in a dust environment.

[0065] 2. Digital twin geological modeling and reinforcement learning control:

[0066] Three-dimensional geological modeling: The lidar point cloud constructs a global coordinate system through the ICP algorithm (initial accuracy ±5 cm). The geological radar scans the 10 m area in front every 5 seconds to identify the coal-rock interface (dielectric constant difference > 5) and faults (reflection coefficient > 0.3), and updates the digital twin model in real time.

[0067] Path optimization algorithm: Adopt the PPO reinforcement learning algorithm. The state space includes the cutting head attitude (3-axis angles, accuracy ±0.1°), geological parameters (interface distance ±0.1 m, hardness ±5 MPa), and inertial navigation error (position deviation ±5 cm, angle deviation ±0.2°), a total of 12 dimensions; the action space is the adjustment of the cutting head angle (step size 0.5°). Use the coal-rock mis-cutting rate (deduct 10 points for every 1 cm of mis-cutting) and path smoothness (curvature > 10 m -1 deduct 5 points) as the reward function, and the response time < 150 ms after training convergence.

[0068] Example: When the digital twin model identifies a normal fault (dip angle 60°, throw 1.5 m) 5 m ahead, the reinforcement learning algorithm outputs an adjustment instruction for the cutting head horizontal angle +3° and pitch angle -2°, and completes the angle correction within 180 ms, with an efficiency improvement of 20 times compared to manual intervention.

[0069] III. Specific implementation steps:

[0070] 1. System initialization and self-calibration (0 - 5 minutes):

[0071] 1.1 Hardware Self-Calibration: The ECU sends a 100Hz square wave signal to the SMA heating circuit, triggering 3 times of heating (1.5W, 10ms each time), and completing the zero-offset correction of the IMU: the zero-drift of the accelerometer drops from 0.2g to 0.03g; the zero-drift of the gyroscope drops from 0.5° / h to 0.08° / h.

[0072] 1.2 Initial Environment Modeling: The lidar completes a 360° scan, generating an initial roadway point cloud map containing 100,000 points, with a size of 5.2m × 3.5m; the ground-penetrating radar scans 15m ahead, identifying a coal seam thickness of 1.2m, and the roof lithology is sandy mudstone (density 2.3g / cm 3 , dielectric constant 8).

[0073] Example: During the initialization process, the SMA self-calibration function completes 3 cycles within 5 minutes. Finally, the IMU zero-drift parameters are stabilized at 0.03g for the accelerometer and 0.08° / h for the gyroscope, providing high-precision initial data for subsequent navigation.

[0074] 2. Real-time Navigation and Error Suppression (5 minutes - 7 hours 55 minutes):

[0075] Data Acquisition and Preprocessing: The IMU outputs raw data at 100Hz. After being attenuated by magnetorheological fluid, the peak value of the vibration acceleration drops from ±5g to ±0.5g, and the 50Hz power frequency noise is suppressed through adaptive notch filtering; the vision sensor detects fluorescent markers at 50m intervals through the YOLOv8 algorithm, combines with the lidar point cloud to calculate the absolute position, and calibrates the cumulative error of the inertial navigation every 10 seconds (average ±10cm).

[0076] Dynamic Weight Switching: When entering the area with a dust concentration of 2000mg / m³, the VBN automatically increases the weight of the ultrasonic radar to 0.7, and the positioning accuracy after fusion is ±10cm / 100m; in a low-light environment (illuminance < 10lux), the weight of the inertial navigation is increased to 0.8, and an auxiliary positioning with an accuracy of ±10cm / 100m is achieved in combination with the magnetic gradient tensor instrument.

[0077] Example: In a low-light roadway (illuminance 5lux), the weight of the inertial navigation is increased to 0.8. Combining with the magnetic anomaly of the anchor bolt detected by the magnetic gradient tensor instrument, the system positioning accuracy is stabilized at ±10cm / 100m, which is 50% higher than the traditional single inertial navigation scheme.

[0078] 3. Geological Anomaly Response (taking a normal fault as an example, triggered at the 4th hour):

[0079] Anomaly Detection: The ground-penetrating radar detects that the dielectric constant in front of 5m increases suddenly from 8 to 18. Combining with the sudden 25% increase in the resistance of the cutting head tactile sensor, the CNN algorithm determines it as a normal fault (dip angle 60°, throw 1.5m) with a 95% confidence level.

[0080] Control response: The digital twin model generates an obstacle - avoiding path with a right deviation of 1.2 m and a slope of +5°. The reinforcement learning outputs an instruction to adjust the angle of the cutting head (horizontal +3°, pitch -2°), and the action is executed within 180 ms. An instruction is sent to the support robot via the 5G network to reduce the bolt spacing in the fault area from 1.2 m to 0.8 m and increase the diameter of the cable bolt to 25 mm, and the enhanced support is completed within 5 minutes.

[0081] Example: When cutting in the fault area, after adjusting the angle of the cutting head, the mis - cutting rate of coal and rock drops from 15% before adjustment to 3%, and the ash content of the coal quality drops from 12% to 8%, improving the coal quality per single shift.

[0082] 4. Sensor failure tolerance (simulating a 60 - second disconnection of the lidar):

[0083] In the conditional generative adversarial network cGAN data filling module, after detecting the interruption of lidar data, the current timestamp, ultrasonic radar, and visual data are input, and the three - axis acceleration (accuracy ±0.01 g) and angular velocity (accuracy ±0.1° / s) are generated. During the filling period, the heading angle error is maintained at ±0.3°, and the system positioning error is ±10 cm / 100 m, which is significantly improved compared with the traditional interpolation method.

[0084] Example: During the disconnection of the lidar, the MSE of the data generated by cGAN is 0.007 g², which is significantly lower than 0.1 g² of the traditional linear interpolation method, ensuring the stable operation of the navigation system when the sensor fails.

[0085] 5. Job end processing (7 hours 55 minutes - 8 hours):

[0086] After the roadheader is stationary for 10 seconds, zero - speed filtering is triggered to correct the cumulative positioning error to <5 cm / 8 hours;

[0087] The navigation data, geological model, and equipment status are stored on the blockchain with a timestamp accuracy of 1 μs for subsequent fault tracing and model iteration.

[0088] As shown in Table 1 below:

[0089]

[0090] Table 1: Comparison table of performance parameters between the technical solution of the present invention and the traditional technical solution.

[0091] In summary, in the present invention, through the self-calibration of the shape memory alloy self-calibrating inertial module, the zero-bias error of the inertial device is corrected, error accumulation is suppressed, and high-precision long-term positioning is achieved. With the help of the magnetorheological fluid intelligent vibration damping platform for full-frequency vibration suppression and stress measurement, the output noise density of the inertial navigation is reduced, and the stability of inertial data and navigation accuracy are improved. The variational Bayesian network processing unit is used to dynamically allocate sensor weights, automatically select the optimal data source under complex working conditions, and improve the positioning accuracy. Relying on the low-latency communication and high-precision clock synchronization of the 5G URLLC communication module, multi-machine collaborative precise operation is realized, and the support efficiency is improved. The digital twin geological modeling unit is used for real-time geological modeling and path optimization, quickly responding to geological anomalies, reducing the coal and rock mis-cutting rate, and improving the coal quality. The cGAN data filling module is used to generate inertial navigation data when the sensor fails, maintain the system positioning accuracy, and ensure the continuity of navigation.

Claims

1. An underground tunneling inertial navigation system for multi-source information fusion, characterized in that: Including: Shape Memory Alloy Self-Calibrating Inertial Module: Integrating a nickel-titanium alloy cantilever beam, triggering mechanical deformation recovery through a heating power of 1.5W, correcting the zero-offset errors of MEMS accelerometers and gyroscopes, with a self-calibration time ≤ 80 ms / time, an operating temperature range of -40°C to 60°C, and a pressure resistance ≥ 50 MPa; Magnetorheological Fluid Intelligent Vibration Damping Platform: Comprising 3 shock absorbers filled with MRF-132DG magnetorheological fluid, with a response time ≤ 0.5 ms, adjusting the viscosity through a current ranging from 0 A to 1 A (both 0 A and 1 A are included), a vibration attenuation rate ≥ 92%, reducing the inertial navigation output noise density to ≤ 5 micrograms per square root hertz, and measuring stress in real time to dynamically adjust the vibration damping parameters; Variational Bayesian Network Processing Unit: Fusing 100Hz inertial navigation data, 10Hz lidar point cloud, 50Hz ultrasonic radar ranging, and 30fps visual images in real time, dynamically allocating sensor weights through posterior probability. When the dust concentration > 1500 mg / m³, the weight of the ultrasonic radar ≥ 0.7; 5G URLLC Communication Module: Supporting the 2.6 GHz band, with a communication delay < 8 ms, achieving multi-device clock synchronization through the time-sensitive network, with a deviation < 500 ns, and a single-device bandwidth ≥ 50 Mbps; Digital Twin Geological Modeling Unit: Based on lidar with an accuracy of ±3 cm and ground-penetrating radar with a detection depth of 15 m, constructing a three-dimensional point cloud model, and combining reinforcement learning to output cutting path adjustment instructions with a step size of 0.5° and a response time < 150 ms; Conditional Generative Adversarial Network cGAN Data Filling Module: Input timestamp and known sensor data to generate inertial navigation data including acceleration and angular velocity, with the training data volume ≥ 1.2×10 6 , and the mean squared error MSE of the generated data ≤ 0.01g 2 ; Tactile perception unit: Integrated with 128 piezoelectric film tactile sensors, with a resolution of 0.1 N / mm 2 , measure the force in real time to perceive the distribution of cutting resistance, and evaluate the load of the cutting head by measuring the torque.

2. The downhole tunneling inertial navigation system for multi-source information fusion according to claim 1, characterized in that: In the shape memory alloy self-calibrating inertial module, the surface of the cantilever beam is coated with an alumina thermal insulation layer with a thickness of 50 μm and a thermal conductivity of 0.15 W / (m·K), and is connected to the edge computing node through a cable with a temperature resistance of 80°C, and the computing power ≥ 200 TOPS.

3. The downhole tunneling inertial navigation system for multi-source information fusion according to claim 1, characterized in that: In the magnetorheological fluid intelligent vibration damping platform, the exterior is wrapped with an aerogel thermal insulation layer with a thermal conductivity of 0.013 W / (m·K), collecting 1000Hz vibration data in real time, dynamically adjusting the shock absorber current through the PWM module to achieve full-band vibration suppression, and simultaneously measuring the fluid pressure to optimize the working parameters of the magnetorheological fluid.

4. A downhole tunneling inertial navigation system for multi-source information fusion according to claim 1, characterized in that: In the variational Bayesian network processing unit, a built-in fault detection module automatically switches to the fusion positioning of the ultrasonic radar and the magnetic gradient tensor instrument when the number of valid lidar point clouds in 5 consecutive frames < 5000, with a positioning accuracy of ±10 cm / 100 m.

5. The downhole tunneling inertial navigation system for multi-source information fusion according to claim 1, characterized in that: In the digital twin geological modeling unit, a ground-penetrating radar is integrated with a resolution of 0.1 m, identifying the dielectric constant difference at the coal-rock interface with an accuracy of ±0.5, detecting faults in combination with a convolutional neural network with an identification accuracy ≥ 95%, and evaluating the cutting energy consumption by measuring the mechanical power.

6. The downhole tunneling inertial navigation system for multi-source information fusion according to claim 1, characterized in that: In the conditional generative adversarial network cGAN data filling module, a 5-layer fully connected generator and a 4-layer discriminator network are adopted.

7. The downhole tunneling inertial navigation system for multi-source information fusion according to claim 1, wherein: The tactile perception unit is linked with the cutting head manipulator of the tunneling equipment, dynamically adjusting the cutting head rotation speed and feed speed by measuring force and torque, and synchronously measuring the mechanical efficiency to optimize the cutting parameters.

8. A method for using an underground tunneling inertial navigation system with multi-source information fusion, which is applied to an underground tunneling inertial navigation system with multi-source information fusion as described in claim 1, and is characterized in that: Including the following steps: Data acquisition: Data is collected through a shape memory alloy self-calibrating inertial module, lidar, ground penetrating radar, and tactile sensors, including position, attitude, geological structure, and cutting resistance information; Preprocessing: Magnetorheological fluid damping and adaptive notch filtering are performed on inertial data, and the noise density is reduced to less than or equal to 5 micrograms per square root hertz. Force-electric signal conversion is performed on tactile data to measure force, with a resolution of 0.1 N / mm²; Dynamic error suppression: The SMA heating circuit is triggered for hardware self-calibration. The variational Bayesian network is combined to online evaluate the random walk parameters of inertial devices, and the cumulative error of the inertial navigation is calibrated by ±10 cm every 10 seconds; Multi-source fusion: The variational Bayesian network dynamically allocates sensor weights according to environmental parameters. In a dusty environment, ultrasonic radar is preferred, with a weight ≥0.

7. In a normal environment, lidar is preferred, with a weight ≥0.6; Multi-robot cooperation: The position information of multiple robots is transmitted through the 5G URLLC network. The actions of the support robots are optimized using the federated learning-blockchain architecture, and the time-sensitive network achieves action synchronization with a deviation <1 μs; Geological adaption: The digital twin model updates the three-dimensional geological structure in real time. When encountering a fault with a dielectric constant mutation >5, the reinforcement learning generates an obstacle avoidance path, the cutting head angle adjustment step is 0.5°, and the response time <180 ms; Data filling: When the sensor fails, the conditional generative adversarial network cGAN data filling module inputs the timestamp and known data to generate 60 seconds of continuous inertial navigation data, and the positioning error after filling ≤±10 cm / 100 m; Energy efficiency optimization: The operating parameters of the equipment are dynamically adjusted by measuring mechanical power and fluid pressure, reducing the system energy consumption by more than 30%.

9. The usage method of an underground tunneling inertial navigation system with multi-source information fusion according to claim 8, characterized in that: In the data acquisition, the detection depth of the ground penetrating radar ≥15 m, the tactile sensor covers a distance of 10 meters from the lidar, the lidar point cloud density ≥0.1 points / m², and the surface of the cutting head ≥80% of the area.

10. The usage method of an underground tunneling inertial navigation system with multi-source information fusion as claimed in claim 8, characterized in that: In the preprocessing, the adaptive notch filtering algorithm suppresses 50 Hz power frequency noise, and the magnetorheological fluid damping platform attenuates the vibration energy of 20 - 300 Hz by ≥92%.

11. The usage method of an underground tunneling inertial navigation system with multi-source information fusion according to claim 8, characterized in that: In the dynamic error suppression, the update frequency of the variational Bayesian network is synchronized with the inertial navigation at 100 Hz, and the accuracy of the blockchain timestamp is 1 μs.

12. The method of using a multi-source information fusion underground tunneling inertial navigation system according to claim 8, characterized in that: In the multi-source fusion, in a low-light environment with illuminance <10 lux, the weight of the inertial navigation is automatically increased to ≥0.8, combined with a magnetic gradient tensor instrument for assisted positioning, with an accuracy of ±10 cm / 100 m.

13. The usage method of an underground tunneling inertial navigation system with multi-source information fusion as claimed in claim 8, characterized in that: In the multi-robot cooperation, the uplink throughput of the 5G URLLC network ≥1.2 Gbps, supporting ≥20 devices to be online simultaneously. The positioning accuracy of the support robot's anchor is ±10 cm, and the latency <50 ms.

14. The usage method of an underground tunneling inertial navigation system with multi-source information fusion as claimed in claim 8, characterized in that: In the geological adaption, the digital twin model updates the roadway point cloud every 2 seconds, and the reinforcement learning algorithm takes the coal-rock mis-cutting rate and path smoothness as the optimization objectives.

15. The usage method of an underground tunneling inertial navigation system with multi-source information fusion according to claim 8, characterized in that: In the data filling, the training data of the conditional generative adversarial network (cGAN) data filling module covers ≥20 working conditions, the peak signal-to-noise ratio between the generated data and the real data is ≥32 dB, and the mean square error is ≤0.008g 2 .

Images