Tunnel robot geological forecasting method and system based on active sensing regulation and control

Abstract

The invention provides a tunnel robot geological forecasting method and system based on active sensing regulation and control, and relates to the technical field of geological forecasting. Carrying out self-adaptive vibration compensation and phase correction on an original radar signal acquired by the tunnel robot by adopting the vibration data to obtain a corrected radar signal; processing the corrected radar signal into an image, inputting the image into a pre-trained geological target recognition model, and outputting the category, position information and recognition confidence of geological anomalies in front of the tunnel face; judging whether to trigger an active rechecking mechanism or not based on the recognition confidence; if the confidence is in a preset fuzzy judgment interval, controlling the robot to perform secondary fine scanning on the target area, and inputting a result into the geological target recognition model again to perform result confirmation; and fusing the confirmed detection result with the space positioning information of the robot, and constructing and generating a three-dimensional geological risk distribution map of the tunnel face. Therefore, the anti-interference capability, the high-risk disaster identification precision and the detection comprehensiveness of tunnel geological forecast are improved.

Description

Technical Field

[0001] This invention relates to the field of geological prediction technology, and in particular to a geological prediction method and system for tunnel robots based on active perception and control. Background Technology

[0002] Advanced geological forecasting for tunnels is a core guarantee for tunnel construction safety. Its core lies in acquiring geological information about the tunnel face and ahead using detection equipment to accurately identify unfavorable geological bodies such as water-rich areas and fault fracture zones, thus preventing fatal accidents such as water inrush, mudslides, and collapses from the outset. As tunnel projects expand into deeper and more complex geological areas, the uncertainty of geological conditions increases significantly, placing higher demands on the timeliness, comprehensiveness, and accuracy of forecasts. Traditional detection methods and existing automated equipment are gradually becoming inadequate for actual construction needs.

[0003] Currently, tunnel geological exploration is mainly divided into two categories: manual inspection and robot-assisted inspection. Traditional manual inspection is labor-intensive, inefficient, highly subjective, and poses multiple safety threats. To address the pain points of manual inspection, tunnel geological prediction robots have emerged, but they still face significant technical bottlenecks in complex real-world scenarios.

[0004] The core shortcomings of existing tunnel geological prediction robots are concentrated in the following aspects: weak anti-interference ability, uneven tunnel surface causing robot vibration, radar antenna is prone to undesirable displacement, resulting in distortion of the original signal and seriously affecting the subsequent identification accuracy; at the same time, the detection is not targeted enough, the monitoring mode of fixed sensors is difficult to flexibly adapt to different geological scenarios, the ability to capture weak features such as micro-fractures and early-stage karst caves is limited, and there is a lack of priority identification mechanism for high-risk geological targets; in addition, the level of intelligence is low, mostly passive scanning detection, unable to dynamically adjust the working posture and detection strategy according to real-time detection results, making it difficult to achieve full-section, high-precision advanced geological detection, which restricts its practical application value in complex tunnel engineering. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a tunnel robot geological prediction method and system based on active perception and control. Through adaptive vibration compensation, enhanced geological identification, and confidence-driven active verification closed-loop control, it effectively improves the anti-interference capability, high-risk disaster identification accuracy, and detection comprehensiveness of tunnel geological prediction, providing precise and efficient geological assurance for tunnel construction safety.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a geological prediction method for tunnel robots based on active perception and control, comprising: Based on real-time acquired vibration data, adaptive vibration compensation and phase correction are performed on the raw radar signals collected by the tunnel robot to obtain the corrected radar signals. The corrected radar signal is processed into a ground-penetrating radar image, which is then input into a pre-trained geological target recognition model for analysis. The model outputs the type, location information, and corresponding recognition confidence of the geological anomaly in front of the tunnel face. Based on the recognition confidence level, it is determined whether to trigger the active verification mechanism; if the confidence level is within the preset fuzzy judgment range, the robot's robotic arm is controlled to adjust the radar detection pose, perform a second fine scan of the target area, and the second scan data is input into the geological target recognition model again for result confirmation. By integrating verified test results with the robot's spatial positioning information, a three-dimensional geological risk distribution map of the tunnel face is constructed and generated.

[0007] Secondly, the present invention provides a tunnel robot geological prediction system based on active perception and control, comprising: The signal correction module is configured to perform adaptive vibration compensation and phase correction on the original radar signal collected by the tunnel robot based on the real-time vibration data, and obtain the corrected radar signal. The anomaly identification module is configured to process the corrected radar signal into a ground-penetrating radar image, input it into a pre-trained geological target identification model for analysis, and output the type, location information and corresponding identification confidence of the geological anomaly in front of the tunnel face. The active verification module is configured to determine whether to trigger the active verification mechanism based on the recognition confidence level; if the confidence level is within the preset fuzzy judgment range, the robot's robotic arm is controlled to adjust the radar detection pose, perform a secondary fine scan of the target area, and the secondary scan data is input into the geological target recognition model again for result confirmation. The results generation module is configured to integrate the verified test results with the robot's spatial positioning information to construct and generate a three-dimensional geological risk distribution map of the tunnel face.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the tunnel robot geological prediction method based on active perception and control described in the first aspect.

[0009] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the geological prediction method for a tunnel robot based on active perception and control described in the first aspect.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively eliminates radar signal distortion caused by robot vibration through adaptive compensation and phase correction based on real-time vibration data, significantly improving the signal-to-noise ratio and fidelity, and solving the problem of inaccurate detection data under complex tunnel surfaces. Simultaneously, based on an improved geological target recognition model, it strengthens long-distance layer-dependent capture and preservation of weak texture features, combined with a high-risk category-priority recognition logic, significantly reducing the risk of missed detection of fatal disasters such as water and mud inrushes. Furthermore, it proposes an innovative confidence-driven active verification mechanism, achieving secondary fine scanning through dynamic adjustment of the robotic arm's detection posture, balancing detection efficiency and recognition accuracy. Finally, it integrates positioning information to generate a three-dimensional geological risk distribution map, providing intuitive and accurate basis for construction decisions, and significantly improving the safety and intelligence level of tunnel construction.

[0011] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 A main flowchart of a tunnel robot geological prediction method based on active perception and control provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the tunnel robot structure provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a geological prediction method for tunnel robots based on active perception and control, including the following steps: S1: Based on the real-time acquired vibration data, adaptive vibration compensation and phase correction are performed on the original radar signal acquired by the tunnel robot to obtain the corrected radar signal. S2: The corrected radar signal is processed into a ground-penetrating radar image, which is then input into a pre-trained geological target recognition model for analysis. The model outputs the type, location information, and corresponding recognition confidence of the geological anomaly in front of the tunnel face. S3: Based on the recognition confidence level, determine whether to trigger the active verification mechanism; if the confidence level is within the preset fuzzy judgment range, control the robot's robotic arm to adjust the radar detection pose, perform a secondary fine scan of the target area, and input the secondary scan data back into the geological target recognition model for result confirmation; S4: Integrate the verified test results with the robot's spatial positioning information to construct and generate a three-dimensional geological risk distribution map of the tunnel face.

[0016] Next, combined Figure 1 This embodiment provides a detailed description of a tunnel robot geological prediction method based on active perception and control.

[0017] (I) Hardware support device for tunnel geological prediction robot The device includes an off-road chassis module consisting of four sets of three-axis driving modules, which is equipped with a four-axis robotic arm module and a face detection module. The overall module equipment is as follows: Figure 2 As shown.

[0018] The three-axis driving module consists of off-road AT tires, hydraulic drive motors, hydraulic cylinders, and a frame. When the robot is moving, a driving command is sent from the command terminal, and then the robot controller sends a command to the robot to drive the hydraulic drive motors, causing the robot to move in the corresponding direction. By controlling the hydraulic cylinders, the robot can avoid terrain, obstacles, and other external objects.

[0019] The chassis module houses the hydraulic power unit, robot main controller, and other key components. The hydraulic power unit provides power to all the robot's hydraulic components; the main controller controls the movement of each part of the robot. The chassis module also integrates a high-precision inertial measurement unit (IMU) for real-time monitoring of the vehicle's attitude and vibration frequency, providing data support for motion compensation of radar signals.

[0020] The four-axis robotic arm module consists of first / second / third / fourth rotary axes, a communication antenna, telescopic components, servo motors, macroscopic adjustment components, and lighting components. The rotary axes adjust the direction of the robotic arm; the servo motors drive the third and fourth rotary axes; the telescopic and macroscopic adjustment components adjust the distance between the robotic arm and the working face; the communication antenna communicates with the control terminal; and the lighting components increase the visibility of the robot's movement.

[0021] The face detection module consists of a ground-penetrating radar, four laser ranging modules, and an end-effector position sensor, forming an active detection end-effector with "sensing-adjustment" capabilities. The ground-penetrating radar detects the face, while the end-effector position sensor and laser ranging modules work together to adjust the distance between the robotic arm and the face.

[0022] In operation, the tunnel robot autonomously travels to the working area at the tunnel face. A four-axis robotic arm carrying a ground-penetrating radar antenna is positioned close to the tunnel face surface. The robotic arm moves along a preset trajectory, driving the ground-penetrating radar to scan the tunnel face point-by-point or continuously, acquiring raw radar echo signals (A-Scan data). During this process, a high-precision inertial measurement unit (IMU) integrated within the robot's chassis continuously and synchronously collects linear acceleration and angular velocity data of the vehicle body along the X, Y, and Z axes.

[0023] The unevenness of the tunnel surface causes mechanical vibrations in the robot during scanning, which are transmitted to the radar antenna via the robotic arm. This causes an undesirable reciprocating displacement of the antenna in the vertical direction (Z-axis). This displacement slightly alters the propagation path length of the radar's transmitted / received electromagnetic waves, introducing periodic or random "sawtooth" waveform distortion and phase shift into the original A-Scan data. This severely reduces the signal-to-noise ratio and the accuracy of subsequent image interpretation and geological identification.

[0024] To address the aforementioned issues, this embodiment proposes an adaptive vibration compensation filter based on IMU data feedback, and performs model-driven reverse displacement correction on the original radar signal.

[0025] Traditional filtering methods typically use fixed-parameter band-stop or high-pass filters to suppress vibration noise at specific frequencies. However, the actual vibration frequency changes in real time with road conditions, rendering fixed filters ineffective. Therefore, this embodiment is based on the concept of "adaptive" filtering, which dynamically estimates the dominant frequency and amplitude of the vibration based on real-time IMU data and adjusts the suppression characteristics of the filter accordingly.

[0026] First, the vertical (Z-axis) acceleration signal acquired by the IMU is processed. Preprocessing is performed, including trend removal and low-pass filtering; then, the main frequency components of the vibration acceleration within the current time window are calculated using real-time Fast Fourier Transform (FFT) or autocorrelation analysis. and its amplitude This forms the basis for adjusting the filter parameters.

[0027] Then, an adjustable band-stop filter bank, such as a second-order infinite impulse response filter, is used and runs on the robot's main controller. Its core stopband frequency... Based on the extracted To perform dynamic settings, i.e. Stopband width The attenuation depth can be related to the amplitude. Correspondingly, the greater the vibration amplitude, the wider the stopband or the stronger the attenuation can be to more thoroughly suppress noise in that frequency band. This process is repeated within each processing time window, ensuring that the filter always tracks and targets the most prevalent mechanical vibration interference frequency.

[0028] Furthermore, considering that filtering primarily targets specific frequency noise caused by vibration, while "reverse displacement correction" directly corrects radar signal delay (or phase) errors caused by antenna physical displacement, a method is proposed to calculate the antenna displacement change during each radar transmit / receive cycle based on the acceleration measured by the IMU, and then perform reverse compensation for this displacement effect in signal processing.

[0029] First, convert acceleration to displacement. For the vertical acceleration measured by the IMU... Performing two integrations is theoretically the way to obtain the displacement. The method, namely .

[0030] However, direct integration is prone to drift errors. This embodiment uses an improved frequency domain segmentation process combined with high-pass filtering to obtain the relative displacement: First, the acceleration signal is converted to the frequency domain. The spectrum is divided into a low-frequency portion (mainly containing integral drift and slow changing trends) and a mid-to-high-frequency portion containing the main energy of the vibration.

[0031] Acceleration spectrum in the mid-to-high frequency range conduct The operation corresponds to the equivalent operation of two integrations in the frequency domain; then, by converting back to the time domain, the relative displacement component caused by high-frequency vibration is obtained. .

[0032] By setting the high-pass cutoff frequency, the low-frequency drift is effectively avoided from being amplified by integration, thus obtaining a more stable estimate of the antenna vibration displacement.

[0033] Next, time delay compensation for the radar signal is performed. The ground-penetrating radar signal can be viewed as an electromagnetic wave emitted from the antenna, reflected by the medium, and returning to the receiving antenna. The antenna is displaced in the vertical direction. This is equivalent to changing the two-way propagation path length of the wave. During correction, for time... A single A-Scan data point acquired corresponds to the target's reflected echo in time. Launch, among which For two-way travel, the additional time delay change caused by antenna displacement. Approximately ,in It is the speed of electromagnetic waves in the air (approximate).

[0034] Reverse displacement correction involves applying time-shift compensation in the reverse direction along the time axis to the A-Scan data. In practice, this is typically done in the frequency domain for radar A-Scan signals. Spectrum Multiplied by a phase compensation factor Then, the signal is transformed back to the time domain to obtain the corrected signal. This operation is equivalent to finely realigning the signal in the time domain, aiming to eliminate waveform jagged artifacts and phase distortion caused by antenna jitter.

[0035] In this embodiment, the original radar A-Scan data stream first passes through the aforementioned adaptive vibration compensation filter to suppress specific frequency electrical noise or coupling interference caused by vibration and independent of displacement. Subsequently, the filtered data is subjected to reverse displacement correction based on IMU displacement estimation to directly compensate for the propagation delay error caused by the physical movement of the antenna. Together, these actions effectively improve the signal-to-noise ratio and fidelity of the radar signal.

[0036] (II) Geological Target Recognition Model - Geological Feature Enhancement Perception and Active Control Algorithm Model (Geo-GPR-Net) Continuously corrected radar signals Arranged and imaged according to spatial location, these images form ground-penetrating radar (GPR) images for analysis. Addressing three core challenges of GPR images in underground tunnel environments—weak texture features (weak signal), large geological stratigraphic spans (long-term dependence), and high risk of missed disaster detection (asymmetric risk)—this embodiment constructs a Geo-GPR-Net model that integrates geological texture preservation and proactive perception decision-making. Based on an improved YOLOv5 architecture, this model deeply couples lightweight feature extraction, spatial attention enhancement, and a risk-sensitive loss function, and uses a confidence-based closed-loop control system to manage the robot's physical actions.

[0037] The overall workflow of the model is as follows: First, the ground-penetrating radar image is input into the model's backbone network for initial multi-scale feature extraction. Then, at the end of the backbone, the feature map is fed into the Stratigraphic-CA module for geological strata-sensitive attention enhancement. Next, the enhanced multi-scale features are passed to the Neck (feature fusion layer), where semantic and detail information are deeply fused through the Geo-GSConv module. Finally, the refined features output from the Neck are fed into the Head (detection head), where object classification and bounding box regression are performed in parallel, directly outputting the category, location, and confidence level of the geological anomaly.

[0038] Specifically, a Stratigraphic-CA coordinate attention module is embedded at the end of the backbone to capture long-distance stratigraphic dependencies of fault fracture zones; a Geo-GSConv module is constructed in the Neck feature fusion layer, which forces the fusion of semantic information from standard convolutions and high-frequency texture information from depthwise separable convolutions through a channel shuffling strategy. During training, a Geo-Risk Loss function is loaded, and a geological risk coefficient is set. High weights are assigned to high-risk categories such as "water inrush and mud inrush," forcing the model to prioritize optimizing the recall rate of high-risk samples during gradient descent, generating the final detection weight file.

[0039] 1. A Geologically Sensitive Coordinate Attention Enhancement Mechanism (Stratigraphic-CA) Tunnel geological hazards (such as fault fracture zones and underground rivers) typically have spatial distribution characteristics of "large vertical penetration depth and wide lateral distribution range," making it difficult for ordinary attention mechanisms to capture this long-distance stratigraphic dependence.

[0040] Based on this, this embodiment embeds a coordinate attention (CA) module at the backbone end of the traditional YOLOv5 network. The input feature tensor of this module... This refers to the feature map initially extracted by Backbone, which is then aggregated along the tunnel excavation direction (X-axis) and the exploration depth direction (Y-axis). For the input feature tensor... Its height (Corresponding to detection depth) and width The pooling formula on (corresponding scan mileage) is defined as:

[0041] in, and These represent the global perceptual features of the c-th channel in the depth and horizontal directions, respectively. The height and width of the input feature map; For feature map The value of a specific pixel at a specific coordinate on the c-th channel.

[0042] By aggregating features in orthogonal directions, the algorithm can effectively suppress isolated noise caused by electromagnetic interference and accurately locate the depth of continuous geological faults, providing spatial coordinate guidance for subsequent fixed-point scanning by the robotic arm.

[0043] 2. Mechanism for preserving weak geological textures based on channel washing (Geo-GSConv Neck) While traditional depthwise separable convolution (DSC) can reduce the edge computing burden on robots, its channel separation characteristics can easily lead to the loss of weak high-frequency signals that characterize early caverns or micro-cracks in GPR images during downsampling.

[0044] To address this, this embodiment constructs a Geo-GSConv module in the Neck feature fusion layer. This module introduces a channel shuffle strategy, forcing cross-channel interaction between the strong semantic information of the rock mass contour extracted by standard convolution and the micro-fracture texture information extracted by DSC. Its mathematical model is expressed as follows:

[0045]

[0046] in, The feature map is derived from the Backbone and enhanced with attention, representing the output from the previous layer. This is a standard convolution operation used to preserve the main outline of the rock strata; Depthwise separable convolutions are used to extract high-frequency details; For channel concatenation, the two types of features are concatenated along the channel dimension to obtain intermediate mixed features. ; The output feature map refers to the final features that have been fully fused and contain both semantics and details. It is input into the next layer and will be sent to the next layer or Head for final prediction. This is the channel shuffling operator.

[0047] Channel shuffling employs a "Reshape-Transpose-Flatten" tensor operation. This involves first splitting the channel dimension into "number of groups - number of channels per group," then permuting and recombining them. This forces the standard convolution, responsible for extracting strong semantic information such as rock strata contours, to interact across channels with DSC, which captures high-frequency texture information such as micro-fractures. This allows each output feature to fuse semantic and texture information, building an information bridge between "strong semantic features" and "weak texture features." Through this mechanism, the model significantly enhances its ability to preserve features of minor underground geological defects while maintaining a lightweight design.

[0048] 3. Asymmetric weighting of geological risks and proactive review triggering strategy In response to the asymmetric risk characteristics in tunnel construction, where "missed detection of sudden water or mud inrushes can have fatal consequences, while false detections in ordinary rock strata have a relatively small impact," this embodiment improves the VariFocal Loss and proposes a geological risk-weighted loss function (Geo-RiskLoss), which is used as the core trigger for the robot's active actions.

[0049] First, for the loss function, a geological risk coefficient is introduced. For high-risk geological categories (such as water-rich areas and mudslides). Setting a larger value forces the model to prioritize optimizing the recall rate of high-risk samples:

[0050] in, The total loss value is the target of model training and optimization. To predict the probability (0~1) that the confidence model believes there is a geological disaster in the current area; The target IoU score is the degree of overlap between the predicted bounding box and the actual geological defect bounding box, used to measure the accuracy of localization. Dynamic weights are set based on the level of geological hazards (e.g., water-rich areas). ordinary rocks ); This is an adjustment factor, which can be a constant, used to address the imbalance between positive and negative samples.

[0051] Through the loss function described above, the model is forced to focus on high-risk geological targets during training.

[0052] During training, the LabelImg annotation software was used to annotate adverse geological features (such as water-rich areas, karst caves, and faults) in the acquired GPR images with rectangular bounding boxes, generating label files containing categories and coordinates. To improve the model's robustness to complex environments, in addition to conventional geometric transformations, physical field simulation enhancement was introduced. By superimposing simulated electromagnetic wave attenuation noise and multiple wave interference, highly realistic "hard samples" were generated to expand the training dataset.

[0053] It should be understood that superimposing simulated electromagnetic wave attenuation noise and multiple wave interference is something that those skilled in the art can achieve. First, data is collected in the field to statistically analyze the intensity, frequency, propagation patterns, and other characteristics of the two types of interference in the real environment. Then, based on statistical patterns, attenuation noise that conforms to the distribution is generated using a random function. Multiple wave interference is generated by superimposing the simulated interference onto the original GPR image according to the proportion of the real interference to generate "hard samples". Finally, invalid samples are removed and labeled to expand the training dataset, thereby improving the robustness of the model to complex interference environments.

[0054] Geo-Risk Loss assigns highly dynamic weights to high-risk categories such as "sudden water and mud" during training, forcing the model to prioritize optimizing the recall rate of high-risk samples. The resulting model is more accurate in identifying high-risk areas, and the output confidence level quantifies the certainty of risk assessment. In real-time applications, the confidence level becomes the basis for action triggering: if it exceeds the certainty threshold, the result is directly output; if it falls within the ambiguity range, a secondary fine scan by the robotic arm is triggered, forming a complete logic of loss function-enabled model, confidence level-quantified risk, and proactive action closed loop.

[0055] Specifically, real-time acquired and vibration-compensated radar data is input into the trained Geo-GPR-Net model to execute active verification closed-loop control. The model outputs the category and predicted confidence level p of suspected geological defects in the current field of view. If the confidence level p is higher than the preset confidence threshold, it is determined as a confirmed target and the result is directly output; if the confidence level p is within the preset fuzzy range, the active verification mechanism is triggered. At this time, the algorithm calculates the spatial coordinates of the suspected target through inverse kinematics, generates robotic arm control commands, drives the four-axis robotic arm to perform a "focusing action," and uses end-effector laser ranging feedback to shorten the coupling distance between the radar antenna and the tunnel face to the optimal detection value, or rotates the antenna polarization direction to perform a secondary fine scan of the suspected area.

[0056] Specifically, based on the confidence level of the model output. Establish the robot's active control decision function Action(p):

[0057] in, This is the robot's decision-making instruction function; To ensure accuracy, the threshold and noise filtering threshold can be set manually; when the detection result falls into the fuzzy region... The algorithm uses inverse kinematics to calculate and outputs control commands to drive a four-axis robotic arm to adjust the radar antenna attitude (rotate the polarization direction or shorten the coupling distance) and perform a secondary fine scan of the suspected area. Based on active verification closed-loop control, it achieves a technological leap from "passive guessing" to "active verification".

[0058] The high-resolution data from the second scan is then input into the model to confirm the results. If the confidence level of the second detection is significantly improved, the geological hazard is confirmed to exist; otherwise, it is marked as an interference signal. Finally, by combining the robot's odometry positioning information with the final detection results, a three-dimensional geological risk cloud map of the tunnel face is constructed, which intuitively displays the distribution location, type, and hazard level of adverse geological bodies, guiding construction personnel to formulate targeted blasting or support plans.

[0059] This specific embodiment effectively eliminates radar signal distortion caused by robot vibration through adaptive compensation and phase correction technology based on real-time vibration data, significantly improving anti-interference capability and signal fidelity, and overcoming the pain point of inaccurate original data. It utilizes an improved geological target recognition model, integrating spatial attention enhancement and semantic-texture feature interaction mechanisms to strengthen the capture of weak features such as micro-fractures and early-stage karst caves. Combined with an asymmetric weighting strategy for geological risk, it constructs a high-risk target priority identification logic, compensating for the shortcomings of insufficient detection targeting. An innovative confidence-driven active perception and control mechanism achieves secondary fine scanning through dynamic adjustment of the robotic arm's detection posture, elevating passive detection to an active closed loop of "real-time analysis - dynamic adjustment - accurate verification," significantly improving the level of intelligence and the accuracy of full-section detection. This provides more reliable geological forecasting support for complex tunnel engineering, breaking through the limitations of existing technologies.

[0060] Example 2 This embodiment provides a tunnel robot geological prediction system based on active perception and control, including: The signal correction module is configured to perform adaptive vibration compensation and phase correction on the original radar signal collected by the tunnel robot based on the real-time vibration data, and obtain the corrected radar signal. The anomaly identification module is configured to process the corrected radar signal into a ground-penetrating radar image, input it into a pre-trained geological target identification model for analysis, and output the type, location information and corresponding identification confidence of the geological anomaly in front of the tunnel face. The active verification module is configured to determine whether to trigger the active verification mechanism based on the recognition confidence level; if the confidence level is within the preset fuzzy judgment range, the robot's robotic arm is controlled to adjust the radar detection pose, perform a secondary fine scan of the target area, and the secondary scan data is input into the geological target recognition model again for result confirmation. The results generation module is configured to integrate the verified test results with the robot's spatial positioning information to construct and generate a three-dimensional geological risk distribution map of the tunnel face.

[0061] As one implementation method, such as Figure 2 As shown, the tunnel robot includes: The three-axis driving module is used to drive the robot to autonomously travel along a preset path to the working face area; The chassis module integrates an inertial measurement unit and a main controller. The inertial measurement unit is used to collect the vehicle's attitude and vibration data in real time. A four-axis robotic arm module is mounted on the chassis module; The face detection module, installed at the end of the four-axis robotic arm module, includes a ground-penetrating radar antenna and a laser ranging sensor for ranging and pose feedback. The inertial measurement unit provides data support for the adaptive vibration compensation and phase correction; the main controller runs the geological target identification model and generates control commands based on the output confidence level; the four-axis robotic arm module responds to the control commands and performs an active verification action to adjust the radar detection pose.

[0062] The tunnel robot, by integrating a high-precision IMU and adaptive filtering algorithms, effectively suppresses the interference of mechanical vibration on radar signals, significantly improving data quality. Its Geo-GPR-Net model intelligently identifies geological anomalies and, combined with a confidence-driven active verification mechanism, achieves secondary, precise detection of suspicious areas. Ultimately, the system automatically generates a three-dimensional geological risk cloud map, thereby significantly improving the accuracy, real-time performance, and automation level of geological hazard prediction ahead of the tunnel, providing reliable technical support for construction safety.

[0063] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the tunnel robot geological prediction method based on active perception and control as described in Embodiment 1 above.

[0064] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the geological prediction method for a tunnel robot based on active perception and control as described in Embodiment 1 above.

[0065] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tunnel robot geological prediction based on active sensing regulation, characterized in that, The method comprises the following steps: Based on real-time collected vibration data, the original radar signals collected by the tunnel robot are adaptively compensated for vibration and phase corrected to obtain corrected radar signals; The corrected radar signals are processed into geological radar images and input into a pre-trained geological target recognition model for analysis, and the category, location information and corresponding recognition confidence of the geological anomaly in front of the tunnel face are output; Based on the recognition confidence, it is judged whether to trigger the active review mechanism; if the confidence is in the preset fuzzy judgment interval, the robot's mechanical arm is controlled to adjust the radar detection pose, the target area is scanned again, and the secondary scanning data is input into the geological target recognition model again for result confirmation; The confirmed detection results and the spatial positioning information of the robot are fused to construct and generate a three-dimensional geological risk distribution map of the tunnel face.

2. The tunnel robot geological prediction method based on active sensing regulation according to claim 1, wherein, The method comprises the following steps: The vertical direction acceleration signal collected by the inertial measurement unit is preprocessed and frequency analyzed to extract the dominant frequency and amplitude of the current mechanical vibration; According to the dominant frequency and amplitude, the center stop band frequency and suppression strength of the adaptive band-stop filter are dynamically adjusted, and the original radar signal is filtered to suppress the specific frequency noise caused by vibration; Based on the acceleration signal, the relative displacement amount of the radar antenna in the vertical direction is calculated by frequency domain integration combined with high-pass filtering; According to the relative displacement amount, the phase of the filtered radar signal is compensated to correct the signal time delay error caused by the physical displacement of the antenna.

3. The tunnel robot geological prediction method based on active sensing regulation of claim 1, wherein, In the geological target recognition model, a coordinate attention module is embedded at the end of the Backbone layer to enhance the model's perception and positioning ability for geological anomalies with long-distance layer dependence characteristics by aggregating features along the tunnel excavation direction and the detection depth direction respectively.

4. The tunnel robot geological prediction method based on active sensing regulation of claim 1, wherein, In the geological target recognition model, a Geo-GSConv module is constructed at the Neck feature fusion layer to forcibly fuse the semantic information extracted by standard convolution and the texture information extracted by depth separable convolution through channel mixing strategy, so as to enhance the representation and retention ability of weak geological features while keeping the model lightweight.

5. The tunnel robot geological prediction method based on active sensing regulation of claim 1, wherein, The pre-trained geological target recognition model is trained by a geological risk weighted loss function; the geological risk weighted loss function is specifically: ； wherein, is a total loss value; is a prediction confidence; is a target IoU score, the degree of overlap between the prediction box and the real geological defect box, used to measure whether the positioning is accurate; is a dynamic weight set according to the geological disaster grade; is an adjustment factor.

6. The tunnel robot geological prediction method based on active sensing regulation of claim 1, wherein, Based on the recognition confidence, it is judged whether to trigger the active review mechanism; if the confidence is higher than the confidence threshold, the detection result is directly output; if the confidence is lower than the noise filtering threshold, the corresponding signal is filtered out; if the confidence is in the fuzzy judgment interval formed by the confidence threshold and the noise filtering threshold, the active review mechanism is triggered, and the mechanical arm is controlled to perform fine scanning.

7. A tunnel robot geological prediction system based on active sensing regulation, characterized in that, The method comprises the following steps: The signal correction module is configured to, based on real-time collected vibration data, adaptively compensate for vibration and phase correct the original radar signals collected by the tunnel robot to obtain corrected radar signals; The anomaly identification module is configured to process the corrected radar signal into a ground penetrating radar image, input the ground penetrating radar image into a pre-trained geological target identification model for analysis, and output a category, location information, and corresponding identification confidence of a geological anomaly in front of the tunnel face. The active review module is configured to determine whether to trigger an active review mechanism based on the identification confidence. If the confidence is in a preset fuzzy determination interval, the robot arm is controlled to adjust the radar detection pose, the target area is finely scanned again, and the secondary scanning data is input into the geological target identification model again for result verification. The result generation module is configured to fuse the verified detection result and the spatial positioning information of the robot, construct, and generate a three-dimensional geological risk distribution map of the tunnel face.

8. The tunnel robot geological prediction system based on active sensing regulation of claim 7, wherein, The tunnel robot comprises: A three-axis driving module for driving the robot to autonomously drive along a preset path to a tunnel face operation area; A chassis module having an inertial measurement unit and a main controller integrated therein, the inertial measurement unit being configured to collect attitude and vibration data of the vehicle body in real time; A four-axis mechanical arm module installed on the chassis module; A tunnel face detection module installed at the end of the four-axis mechanical arm module, comprising a ground penetrating radar antenna and a laser ranging sensor for ranging and pose feedback; The inertial measurement unit is configured to provide data support for the adaptive vibration compensation and phase correction; the main controller is configured to run the geological target identification model and generate a control instruction according to the output confidence; and the four-axis mechanical arm module is configured to perform an active review action of adjusting the radar detection pose in response to the control instruction.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the tunnel robot geological prediction method based on active sensing regulation according to any one of claims 1-6.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the tunnel robot geological prediction method based on active sensing regulation according to any one of claims 1-6.

Images