Anti-impact drilling robot navigation system and method based on multi-mode 3D target detection

Abstract

The invention discloses an anti-impact drilling robot navigation system and method based on multi-mode 3D target detection. The anti-impact drilling robot navigation system comprises a multi-sensor sensing module, a 3D target detection module, a path planning module and a control module. The multi-sensor sensing module is used for acquiring surrounding environment data and the real-time position and posture of the anti-impact drilling robot; the 3D target detection module adopts an improved MVXNet model to identify environmental data; the path planning module is of a double-layer architecture, wherein one layer adopts an improved dung beetle optimization algorithm to carry out global path planning; on the other layer, an improved dynamic window method is adopted for dynamic obstacle avoidance and local planning adjustment of a global path, and finally a corresponding control instruction is generated through a control module to enable the anti-impact drilling robot to advance according to a plan. Therefore, high-robustness, high-adaptability and high-precision navigation driving of the anti-impact drilling robot in a complex environment is realized, and the driving efficiency of the anti-impact drilling robot is effectively improved on the premise of ensuring the advancing safety of the anti-impact drilling robot.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent coal mine equipment, in particular to a navigation system and method for a bump-preventing drilling robot based on multi-modal 3D target detection. BACKGROUND

[0002] As an important part of the national economy, the coal industry has become an era trend with the breakthrough development of related disciplines and key technologies such as robots and artificial intelligence. The navigation system of the robot is a key basic technology for the realization of the intelligence of the mobile robot in the coal mine, and the completeness of the navigation function directly affects the stability of the mobile robot in the coal mine. It is also one of the important obstacles to the comprehensive popularization of the intelligence of the mobile robot in the coal mine. At present, the coal industry mainly adopts a semi-automatic mode, and the mining operation is completed by the miners operating mechanical equipment. Drilling and pressure relief is an important method for preventing and controlling rock burst, and personnel need to carry drilling equipment into the dangerous area to perform drilling and pressure relief operation. In the process of pressure relief, rock burst disaster is easily induced, which seriously threatens the life safety of the construction personnel.

[0003] The bump-preventing drilling robot is a fully automatic device, which has been applied in the coal mine. It can automatically perform drilling and pressure relief operation without manual operation. However, due to the limitations of the current underground positioning methods, such as radio frequency positioning, inertial navigation positioning, and video monitoring, there are problems such as limited positioning range, error accumulation, low precision, poor stability, discontinuous positioning results, and unsuitability for long-term positioning, which cannot meet the long-term application of the bump-preventing drilling robot in the complex working conditions of the coal mine. Using traditional static mapping means such as high-precision laser scanners and ground mapping radars to construct the underground map is inefficient and costly, which is not conducive to the accurate modeling and rapid map updating of the roadway expansion and dynamic obstacles in the coal mine. Therefore, the current bump-preventing drilling robot is mainly operated remotely by manual operation to move the robot to the required drilling and pressure relief area. This method not only increases the labor time of personnel, but also has communication delay between the personnel and the robot due to manual remote operation. This leads to the fact that the personnel cannot operate the robot to avoid obstacles in time during the robot's travel process, and ultimately the robot cannot reach the required position smoothly.

[0004] Therefore, how to provide a new navigation system and method, which can automatically collect surrounding environment data in the coal mine environment, form a planned path after analysis and processing, and make the bump-preventing drilling robot travel automatically according to the planned path, and can automatically avoid obstacles during the travel process, so as to ensure that the bump-preventing drilling robot travels to the required position with high precision and autonomy, is the research direction of the present application. SUMMARY

[0005] In view of the problems in the prior art, the present application provides a drilling robot navigation system and method based on multi-modal 3D target detection, which can effectively solve the above technical problems.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is: a drilling robot navigation system based on multi-modal 3D target detection, comprising a multi-sensor perception module, a 3D target detection module, a path planning module and a control module mounted on the drilling robot; The multi-sensor perception module is used to obtain surrounding environment data and real-time position and attitude of the drilling robot; The 3D target detection module is used to identify static environment structures and dynamic obstacles from the surrounding environment data; The path planning module is used to form a path planning and dynamic obstacle avoidance scheme according to the identified static environment structures and dynamic obstacles, in combination with the real-time position and attitude of the drilling robot; The control module comprises a host computer and a lower computer, the host computer is used to obtain the path planning and dynamic obstacle avoidance scheme fed back by the path planning module and form corresponding control instructions, and the lower computer is used to receive the control instructions sent by the host computer and control the drilling robot to travel according to the path planning and dynamic obstacle avoidance scheme.

[0007] Further, the multi-sensor perception module comprises a depth camera, a laser radar, a UWB locator, an IMU sensing unit and a wheel speed meter, the depth camera is used to obtain depth images of the surrounding environment; the laser radar is used to obtain three-dimensional point clouds of the surrounding environment; the UWB locator is used to obtain position information of the drilling robot; the IMU sensing unit is used to obtain the attitude of the drilling robot; and the wheel speed meter is used to obtain the moving speed of the drilling robot.

[0008] Further, the laser radars are multiple, and the multiple laser radars are distributed at the front and rear parts of the drilling robot. In order to make up for the visual angle blind area of a single laser radar and enable the robot to obtain comprehensive environment information, the multiple laser radars are calibrated in a unified coordinate, taking the coordinate system of the front laser radar as a reference, the point cloud coordinates of the rear laser radars are converted, and the cubic spline interpolation method is adopted in the point cloud fusion process to ensure the alignment of the point cloud data; at the same time, in order to solve the distortion of the laser radar in the motion process, combined with the accurate position and attitude information of the drilling robot, the sensor position and attitude of each point at the collection time are calculated through time synchronization and interpolation, all points are converted to the same time reference, and the correction of the laser three-dimensional point cloud is realized.

[0009] The working method of the above-mentioned drilling robot navigation system based on multi-modal 3D target detection comprises the following steps: Step one, data collection: obtain the surrounding environment data and the real-time position and attitude of the anti-collision drilling robot, and determine the position information of the destination; wherein the surrounding environment data includes three-dimensional point cloud and depth image; Step two, target recognition: identify static environmental structures and dynamic obstacles from the surrounding environment data using an improved MVXNet model; Step three, path planning: through the improved DungBeetle Optimizer (DBO), combined with the position information of the static environment structure and the destination, and using the heuristic search method to realize the optimal path solving in the static scene, the corresponding path planning scheme is obtained; at the same time, through the improved Dynamic Window Approach (DWA), according to the movement of the identified dynamic obstacles, a dynamic obstacle avoidance scheme and local path planning trajectory smoothing are established; Step four, autonomous driving of the anti-collision drilling robot: the anti-collision drilling robot starts driving according to the path planning scheme of step three, adopts the dynamic obstacle avoidance scheme to avoid obstacles during driving, and keeps the driving trajectory smooth; at the same time, the surrounding environment data and the real-time position and attitude of the anti-collision drilling robot are collected in real time during driving, and steps two and three are repeated to update the path planning scheme and the dynamic obstacle avoidance scheme in real time, so that the anti-collision drilling robot finally reaches the destination.

[0010] Further, in the step one, the position information and attitude information of the anti-collision drilling robot are tightly coupled through the extended Kalman filter (EKF) to obtain the accurate position and attitude information of the anti-collision drilling robot.

[0011] Further, in the step two, the improved MVXNet model is improved and optimized with the MVXNet algorithm model as the core, which includes four parts of input end, backbone network, feature fusion and detection head. In the feature fusion part, a semantic difference module (SDM) is introduced to optimize the semantic alignment of multi-modal features. At the same time, since the traditional edge detection operator cannot finely extract the differential graph of the feature, a learnable boundary operator is introduced in the semantic difference module, so that each position of the kernel has different values. Then, multi-scale feature extraction is introduced in the backbone network, and MSPANet module is used instead of the residual block of ResNet network in MVXNet. The MSPANet module contains three core components: HPC module for extracting multi-scale spatial information, SPR module for learning channel attention weight, and Softmax operation for calibrating channel attention weight. Parallel channel attention mechanism and spatial attention mechanism operations are introduced in each Layer layer of the MSPANet module to complete the construction of the improved MVXNet model.

[0012] Furthermore, step three specifically involves: constructing a global navigation path by fusing static environment structure and destination location information using an improved dung beetle optimization algorithm to obtain a path planning scheme; simultaneously, employing an improved dynamic window method to dynamically correct the global navigation path of the path planning scheme. This method optimizes the path by fusing real-time spatial distribution information of dynamic objects, obstacle spatiotemporal distribution data predicted by the extended Kalman filter algorithm, and local trajectory evaluation function, ultimately generating a real-time control command sequence that conforms to robot kinematic constraints, and establishing trajectory smoothing between the dynamic obstacle avoidance scheme and the local path planning.

[0013] Furthermore, since the traditional dung beetle optimization algorithm may not perform ideally when dealing with high-dimensional complex problems, exhibiting slow convergence speed and susceptibility to getting trapped in local optima due to parameter settings, the traditional dung beetle optimization algorithm is optimized in four aspects: chaotic mapping strategy, global exploration strategy, foraging strategy, and t-mutation perturbation. Specifically, the improved dung beetle optimization algorithm is as follows: First, addressing the shortcomings of existing chaotic mapping in fragile chaos and weak dynamic behavior, a chaotic mapping CTBCS structure is introduced, and... Replace with Piecewise chaotic mapping. The algorithm is first replaced with a Logistic chaotic mapping; secondly, the Harris Eagle optimization algorithm is selected, and its exploration mechanism is integrated into the dung beetle optimization algorithm. The reason for this is that the Harris Eagle optimization algorithm exhibits excellent global exploration capabilities in complex search spaces, especially showing significant local extremum escape characteristics in multimodal function optimization scenarios. Then, a hybrid optimization model is constructed, which enhances the optimization accuracy and convergence speed of the dung beetle optimization algorithm in high-dimensional nonlinear optimization problems. Thirdly, the foraging strategy of the parrot optimization algorithm, which has good global search capabilities and environmental adaptability, is introduced to improve the performance of the dung beetle optimization algorithm in multimodal and complex optimization problems, increasing search efficiency and optimization accuracy. Finally, a mutation strategy with a higher tail probability density t-distribution is introduced to enhance the search performance of the dung beetle optimization algorithm, avoid premature convergence, and form an improved dung beetle optimization algorithm.

[0014] Furthermore, since the traditional DWA algorithm struggles to accurately predict the trajectory of obstacles in complex dynamic environments such as rapidly moving obstacles, leading to the possibility of collisions and significant deviations in path planning results, resulting in inaccurate target point arrival, optimizations are made to the traditional DWA algorithm in obstacle trajectory prediction and trajectory planning. Specifically, the improved dynamic window method involves: in obstacle trajectory prediction, introducing an extended Kalman filter algorithm into the dynamic window method, integrating the prediction results into the cost function; and in robot local trajectory planning, considering the relationship between the robot and the target point in terms of azimuth angle, travel distance changes, and speed fluctuations, firstly, improving the azimuth angle evaluation sub-function by introducing the distance relationship with the target point and dynamically adjusting the azimuth angle weight; secondly, improving the distance evaluation sub-function with the target point by introducing new evaluation metrics. As the target point approaches, the weight of the trajectory closest to the target point is gradually increased; finally, the velocity smoothness evaluation sub-function is improved, and as a secondary optimization, a new evaluation index is introduced. Furthermore, the evaluation function is comprehensively improved to predict the future position of dynamic obstacles, plan avoidance paths in advance, and achieve smooth optimization of local path trajectories.

[0015] Compared with existing technologies, the present invention combines a multi-sensor perception module, a 3D target detection module, a path planning module, and a control module, which has the following advantages: 1. This invention combines a multimodal 3D target detection algorithm, a global dynamic path planning algorithm, and a dynamic obstacle avoidance algorithm. The multimodal 3D target detection algorithm uses an improved MVXNet model to identify environmental data; the global dynamic path planning algorithm uses an improved dung beetle optimization algorithm for global path planning; and the dynamic obstacle avoidance algorithm uses an improved dynamic window method for dynamic obstacle avoidance and local planning and adjustment of the global path. The control module generates corresponding control commands to make the anti-impact drilling robot move according to the plan, thereby achieving high robustness, high adaptability, and high precision navigation and driving of the anti-impact drilling robot in complex environments, effectively ensuring the safety of the anti-impact drilling robot's driving.

[0016] 2. This invention enables the anti-impact drilling robot to acquire a real-time static map of a three-dimensional point cloud with rich structural contours of the surrounding environment through a multi-sensor perception module, and effectively eliminates interfering dynamic point clouds that affect the robot's navigation behavior through a 3D target detection module, thereby improving the driving efficiency of the anti-impact drilling robot.

[0017] 3. The path planning module of this invention is designed with a two-level architecture. One level is used for global path planning, and the other level is used for dynamic obstacle avoidance and local path adjustment. Through the two-level architecture, the anti-impact drilling robot can maintain a good balance between the local development capability and global exploration capability of global dynamic path planning, and can efficiently predict the real-time movement trajectory of dynamic obstacles. Thus, while ensuring the safety of the anti-impact drilling robot, its driving efficiency can be effectively improved. Attached Figure Description

[0018] Figure 1 This is a block diagram of the system composition in this invention.

[0019] Figure 2 This is a communication schematic diagram in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below.

[0021] like Figure 1 As shown, a navigation system for an anti-impact drilling robot based on multimodal 3D target detection includes a multi-sensor perception module, a 3D target detection module, a path planning module, and a control module mounted on the anti-impact drilling robot. The multi-sensor perception module acquires surrounding environmental data and the real-time position and attitude of the anti-impact drilling robot. The 3D target detection module identifies static environmental structures and dynamic obstacles from the surrounding environmental data. The path planning module formulates a path planning and dynamic obstacle avoidance scheme based on the identified static environmental structures and dynamic obstacles, combined with the real-time position and attitude of the anti-impact drilling robot. The control module includes a host computer and a slave computer. The host computer is an industrial control computer, and the slave computer is a PLC controller. The host computer acquires the path planning and dynamic obstacle avoidance scheme fed back by the path planning module and generates corresponding control commands. The slave computer receives the control commands sent by the host computer and controls the anti-impact drilling robot to move according to the path planning and dynamic obstacle avoidance scheme. The multi-sensor perception module includes a depth camera, a lidar, a UWB locator, an IMU sensing unit, and a wheel speed meter. The depth camera is used to acquire depth images of the surrounding environment; the lidar is used to acquire 3D point clouds of the surrounding environment; the UWB locator is used to acquire the position information of the anti-collision robot; the IMU sensing unit is used to acquire the attitude of the anti-collision drilling robot; and the wheel speed meter is used to acquire the movement speed of the anti-collision robot. Figure 2As shown, the LiDAR and depth camera generate massive amounts of image and environmental perception data during operation. These data are relayed via an industrial switch and connected to the industrial control computer via Ethernet. The data collected by the UWB positioner and IMU sensor unit is relatively compact and uses a USB connection. The collected sensor data is processed and communicated via ROS. The path planning and PLC controller communicate via the Modbus TCP protocol. To avoid safety risks such as line damage and signal interference in complex field environments, a wireless mesh base station is used to connect the control console and the industrial control computer on the robot. During movement, the industrial control computer on the anti-blowout drilling robot acts as the control module, converting the pre-planned path information into speed commands and transmitting these commands to the PLC controller (programmable logic controller). After receiving the speed commands, the PLC controller controls the hydraulic motor's operating state by adjusting the hydraulic oil pressure and flow, thus achieving drive and steering control of the anti-blowout drilling robot. By monitoring and adjusting the hydraulic oil input and output, the PLC controller ensures that the hydraulic motor provides stable and controllable power output under different load conditions. The entire process seamlessly integrates path planning, speed control, and the hydraulic system, ensuring the robot can reach its target location efficiently and safely. Ground monitoring connects the underground control console with the mine's surface control center via wired cables, establishing a stable and reliable communication link for data transmission and information exchange.

[0022] As an improvement of this invention, multiple lidars are used, distributed at the front and rear of the anti-impact drilling robot. To compensate for the blind spots of a single lidar and enable the robot to obtain comprehensive environmental information, the multiple lidars are calibrated with unified coordinates. Using the coordinate system of the front lidar as a reference, the point cloud coordinates of the rear lidar are transformed. During the point cloud fusion process, cubic spline interpolation is used to ensure the alignment of the point cloud data. Simultaneously, to address the distortion of the lidars during movement, combined with the precise position and attitude information of the anti-impact drilling robot, the sensor position and attitude of each point at the acquisition time are calculated through time synchronization and interpolation. All points are transformed to the same time reference, achieving correction of the laser three-dimensional point cloud.

[0023] The working method of the above-mentioned anti-blowout drilling robot navigation system based on multimodal 3D target detection includes the following steps: Step 1: Data Acquisition: Acquire surrounding environmental data and the real-time position and attitude of the anti-impact drilling robot, and determine the location information of the destination; the surrounding environmental data includes 3D point cloud and depth image; the UWB position information and IMU attitude information of the anti-impact drilling robot are tightly coupled by extended Kalman filter (EKF), specifically: the UWB / IMU tightly coupled state vector is as follows: In the formula, For the robot's position, For the robot's speed, For the robot's posture, For the robot's linear acceleration, Let be the robot's angular velocity.

[0024] The aforementioned tightly coupled UWB / IMU state vectors are updated according to a kinematic model; the state equations are constructed as follows: in, Here is the state transition matrix. For process noise, This is the IMU sampling time interval.

[0025] After receiving IMU attitude information each time, a state prediction is performed: in, Let be the process noise covariance matrix.

[0026] During the tight coupling of UWB / IMU, the observations use the current coordinates measured by the UWB locator, and the state is updated after receiving UWB data: in, For Kalman gain, For the observation matrix, It measures the noise covariance matrix. The coordinate information is obtained from UWB measurements. After the above calculations and processing, the precise real-time position and attitude information of the anti-impact drilling robot is obtained.

[0027] Step 2, Target Recognition: An improved MVXNet model is used to identify static environmental structures and dynamic obstacles from the surrounding environmental data. The improved MVXNet model is based on the MVXNet algorithm model, which consists of four parts: input, backbone network, feature fusion, and detection head. In the feature fusion part, a semantic difference module (SDM) is introduced to optimize the semantic alignment of multimodal features. Since traditional edge detection operators cannot extract the differential map of features in a fine manner, a learnable boundary operator is introduced in the semantic difference module, so that each position of the kernel has a different value. Then, multi-scale feature extraction is introduced into the backbone network, and the MSPANet module is used to replace the residual block of the ResNet network in MVXNet. The MSPANet module contains three core components: HPC module for extracting multi-scale spatial information, SPR module for learning channel attention weights, and Softmax operation for calibrating channel attention weights. Parallel channel attention mechanism and spatial attention mechanism operations are introduced into each layer of the MSPANet module to complete the construction of the improved MVXNet model.

[0028] In this embodiment, the designed 3D target detection module achieves accurate identification of dynamic targets in point cloud data. First, the module's input receives LiDAR point cloud and depth image data, preprocesses the data to form regularized voxel feature tensors and two-dimensional feature maps, and outputs them to the image backbone network. Then, the reconstructed backbone network unit processes the point cloud and image features, extracting high-level feature representations. In multimodal target detection in a roadway environment, the introduction of an attention mechanism highlights the features of important targets. The replaced MSPANet module can obtain richer multi-scale feature representations and can adaptively adjust channel attention weights. The optimized feature fusion unit receives the extracted features and performs multimodal alignment, information complementarity, and feature interaction. The introduced semantic difference module improves the recognizability of feature boundaries. Finally, the detection head unit generates the final detection result from the fused feature map, including 3D bounding boxes, target categories, and confidence scores.

[0029] Step 3, Path Planning: A global navigation path is constructed by integrating the static environment structure and destination location information using an improved dung beetle optimization algorithm to obtain a path planning scheme. At the same time, an improved dynamic window method is used to dynamically correct the global navigation path of the path planning scheme. This is achieved by real-time integration of the spatial distribution information of dynamic objects, obstacle spatiotemporal distribution data predicted by the extended Kalman filter algorithm, and local trajectory evaluation function optimization, ultimately generating a real-time control command sequence that conforms to the robot's kinematic constraints, and establishing trajectory smoothing between the dynamic obstacle avoidance scheme and the local path planning.

[0030] Because traditional dung beetle optimization algorithms may not perform ideally when dealing with high-dimensional complex problems, exhibiting slow convergence speed and susceptibility to local optima due to parameter settings, this paper optimizes the traditional dung beetle optimization algorithm in four aspects: chaotic mapping strategy, global exploration strategy, foraging strategy, and t-mutation perturbation. Specifically, the improved dung beetle optimization algorithm addresses the shortcomings of existing chaotic mapping in fragile chaos and weak dynamic behaviors by introducing a chaotic mapping CTBCS structure, and... Replace with Piecewise chaotic mapping. The algorithm is first replaced with a Logistic chaotic mapping; secondly, the Harris Eagle optimization algorithm is selected, and its exploration mechanism is integrated into the dung beetle optimization algorithm. The reason for this is that the Harris Eagle optimization algorithm exhibits excellent global exploration capabilities in complex search spaces, especially showing significant local extremum escape characteristics in multimodal function optimization scenarios. Then, a hybrid optimization model is constructed, which enhances the optimization accuracy and convergence speed of the dung beetle optimization algorithm in high-dimensional nonlinear optimization problems. Thirdly, the foraging strategy of the parrot optimization algorithm, which has good global search capabilities and environmental adaptability, is introduced to improve the performance of the dung beetle optimization algorithm in multimodal and complex optimization problems, increasing search efficiency and optimization accuracy. Finally, a mutation strategy with a higher tail probability density t-distribution is introduced to enhance the search performance of the dung beetle optimization algorithm, avoid premature convergence, and form an improved dung beetle optimization algorithm.

[0031] Because traditional DWA algorithms struggle to accurately predict obstacle trajectories in complex dynamic environments such as rapidly moving obstacles, leading to potential collisions and significant deviations in path planning, resulting in inaccurate target point arrival, this paper optimizes the traditional DWA algorithm for obstacle trajectory prediction and trajectory planning. Specifically, the improved dynamic window method involves: In obstacle trajectory prediction, an extended Kalman filter algorithm is introduced into the dynamic window method, incorporating the prediction results into its cost function; in robot local trajectory planning, the relationship between the robot and the target point—including changes in azimuth angle, travel distance, and speed fluctuations—is considered. First, the azimuth angle evaluation sub-function is improved by incorporating the distance relationship to the target point and dynamically adjusting the azimuth angle weight; second, the distance evaluation sub-function is improved by introducing new evaluation metrics. As the target point approaches, the weight of the trajectory closest to the target point is gradually increased; finally, the velocity smoothness evaluation sub-function is improved, and as a secondary optimization, a new evaluation index is introduced. Furthermore, the evaluation function is comprehensively improved to predict the future position of dynamic obstacles, plan avoidance paths in advance, and achieve smooth trajectory optimization of local paths; the specific evaluation function formula is as follows: In the formula, This is the azimuth evaluation sub-function, which evaluates the relationship between the robot's heading angle and the angle between the line connecting the robot's center and the target point. For the improved azimuth evaluation sub-function; The distance evaluation sub-function is the closest distance between the simulated trajectory and the obstacle at the current speed. The velocity smoothness evaluation sub-function represents the velocity magnitude of the current trajectory; These are the coefficients of each evaluation subfunction; This corresponds to the distance between the end of the simulated trajectory and the target point; The distance between the starting point and the target point in the local path planning; This is the distance factor between the simulated trajectory and the target point; A new evaluation index for distance from the target point; A new index for evaluating speed smoothness; The coefficients of the sub-function evaluating the distance to the target point; The current velocity and angular velocity of the mobile robot; The spatial velocity is sampled for the mobile robot's velocity.

[0032] In this embodiment, considering the complex and variable geological conditions of underground coal mine roadways, and the potential for sudden situations such as roadway deformation, a two-tier architecture design is adopted: First, an environmental map is constructed based on simultaneous localization and mapping (SMR) technology, using a grid map as the environmental representation model. The robot's starting point and target point coordinates are input through a human-machine interface to initialize the navigation task. Second, a globally optimal path is generated using a multi-strategy improved dung beetle optimization algorithm. The robot's multi-sensor perception module collects environmental data in real time, using a target detection algorithm to perceive the types and locations of obstacles around the robot. An extended Kalman filter algorithm is used to predict the future positions of moving obstacles, and a local trajectory evaluation function is optimized to generate spatiotemporal trajectory data containing velocity vectors and acceleration parameters. Then, when a dynamic obstacle is detected entering the safety warning area, an improved dynamic window method is triggered to generate a dynamic obstacle avoidance scheme in real time. Finally, by adjusting the speed control parameters online, a locally optimal path is planned to avoid moving obstacles while satisfying the robot's kinematic constraints, and the global path is simultaneously corrected online.

[0033] Step 4: Autonomous Driving of the Anti-impact Drilling Robot: The anti-impact drilling robot begins to drive according to the path planning scheme in Step 3. During the driving process, it adopts a dynamic obstacle avoidance scheme to avoid dynamic obstacles and maintains a smooth driving trajectory. At the same time, it collects real-time environmental data and the real-time position and attitude of the anti-impact drilling robot during the driving process, and repeats Steps 2 and 3 to update the path planning scheme and dynamic obstacle avoidance scheme in real time, so that the anti-impact drilling robot can reach its destination.

[0034] In this embodiment, the control module acts as the command center for the robot's actions. It subscribes to the target detection result topic to clarify the distribution of the work target and obstacles in the environment. Simultaneously, it combines the high-precision map constructed by the 3D target detection module and uses a path planning algorithm to plan a safe and efficient travel path. Subsequently, the control module converts the planned path into corresponding speed commands and publishes them to the ROS topic. To achieve actual control of the anti-impact drilling robot, the control module subscribes to the speed topic and communicates with the PLC module via the Modbus TCP protocol. After receiving these commands, the PLC module converts them into corresponding signals to control the operation of the hydraulic motor, thereby driving the anti-impact drilling robot to travel along the predetermined path. This achieves real-time transmission and rapid response of data from perception, processing to execution, ensuring that the robot can stably and efficiently complete drilling and decompression operations in complex and ever-changing industrial environments.

[0035] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A navigation system for an anti-blowout drilling robot based on multimodal 3D target detection, characterized in that, This includes a multi-sensor perception module, a 3D target detection module, a path planning module, and a control module mounted on the anti-impact drilling robot; The multi-sensor sensing module is used to acquire surrounding environmental data and the real-time position and attitude of the anti-impact drilling robot. The 3D target detection module is used to identify static environmental structures and dynamic obstacles from surrounding environmental data; The path planning module is used to form a path planning and dynamic obstacle avoidance scheme based on the identified static environmental structure and dynamic obstacles, combined with the real-time position and attitude of the anti-impact drilling robot. The control module includes a host computer and a slave computer. The host computer is used to obtain the path planning and dynamic obstacle avoidance scheme fed back by the path planning module and generate corresponding control commands. The slave computer is used to receive the control commands sent by the host computer and control the anti-impact drilling robot to move according to the path planning and dynamic obstacle avoidance scheme.

2. The anti-blowout drilling robot navigation system based on multimodal 3D target detection according to claim 1, characterized in that, The multi-sensor perception module includes a depth camera, a lidar, a UWB locator, an IMU sensing unit, and a wheel speed meter. The depth camera is used to acquire depth images of the surrounding environment; the lidar is used to acquire three-dimensional point clouds of the surrounding environment; the UWB locator is used to acquire the position information of the anti-collision robot; the IMU sensing unit is used to acquire the attitude of the anti-collision drilling robot; and the wheel speed meter is used to acquire the movement speed of the anti-collision robot.

3. The anti-blowout drilling robot navigation system based on multimodal 3D target detection according to claim 2, characterized in that, The laser radar is multiple, and the multiple laser radars are distributed at the front and rear of the anti-impact drilling robot.

4. A method for operating the anti-blowout drilling robot navigation system based on multimodal 3D target detection according to any one of claims 1 to 3, characterized in that, Includes the following steps: Step 1: Data Acquisition: Obtain surrounding environmental data and the real-time position and attitude of the anti-impact drilling robot, and determine the location information of the destination; the surrounding environmental data includes 3D point cloud and depth image; Step 2, Target Recognition: An improved MVXNet model is used to identify static environmental structures and dynamic obstacles from the surrounding environmental data; Step 3, Path Planning: By using the improved dung beetle optimization algorithm, combined with the static environment structure and destination location information, the optimal path is solved in the static scene, and the corresponding path planning scheme is obtained; at the same time, by using the improved dynamic window method, the dynamic obstacle avoidance scheme and the trajectory smoothing of the local path planning are established based on the identified dynamic obstacle movement. Step 4: Autonomous Driving of the Anti-impact Drilling Robot: The anti-impact drilling robot begins to drive according to the path planning scheme in Step 3. During the driving process, it adopts a dynamic obstacle avoidance scheme to avoid dynamic obstacles and maintains a smooth driving trajectory. At the same time, it collects real-time environmental data and the real-time position and attitude of the anti-impact drilling robot during the driving process, and repeats Steps 2 and 3 to update the path planning scheme and dynamic obstacle avoidance scheme in real time, so that the anti-impact drilling robot can reach its destination.

5. The method according to claim 4, characterized in that, In step one, the position and attitude information of the anti-blowout drilling robot are tightly coupled by extended Kalman filtering to obtain the precise position and attitude information of the anti-blowout drilling robot.

6. The method according to claim 4, characterized in that, The improved MVXNet model in step two is based on the MVXNet algorithm model, which consists of four parts: input, backbone network, feature fusion, and detection head. A semantic difference module is introduced in the feature fusion part to optimize the semantic alignment of multimodal features. A learnable boundary operator is introduced in the semantic difference module, giving each position of the kernel a different value. Next, multi-scale feature extraction is introduced into the backbone network, and the MSPANet module replaces the residual blocks of the ResNet network in MVXNet. The MSPANet module contains three core components: an HPC module for extracting multi-scale spatial information, an SPR module for learning channel attention weights, and a Softmax operation for calibrating channel attention weights. Parallel channel attention and spatial attention mechanisms are introduced into each layer of the MSPANet module, completing the construction of the improved MVXNet model.

7. The method according to claim 4, characterized in that, Step three specifically involves: constructing a global navigation path by fusing static environmental structure and destination location information using an improved dung beetle optimization algorithm, and obtaining a path planning scheme; simultaneously, employing an improved dynamic window method to dynamically correct the global navigation path of the path planning scheme. This method optimizes the path by fusing real-time spatial distribution information of dynamic objects, obstacle spatiotemporal distribution data predicted by the extended Kalman filter algorithm, and local trajectory evaluation function, ultimately generating a real-time control command sequence that conforms to robot kinematic constraints, and establishing trajectory smoothing between the dynamic obstacle avoidance scheme and the local path planning.

8. The method according to claim 4, characterized in that, The improved dung beetle optimization algorithm specifically involves: first, introducing a chaotic mapping CTBCS structure, and then... Replace with Piecewise chaotic mapping. First, the algorithm is replaced with a Logistic chaotic mapping. Second, the Harris Eagle optimization algorithm is selected, and its exploration mechanism is integrated into the dung beetle optimization algorithm to construct a hybrid optimization model, thereby enhancing the optimization accuracy and convergence speed of the dung beetle optimization algorithm in high-dimensional nonlinear optimization problems. Third, the foraging strategy of the parrot optimization algorithm is introduced to improve the performance of the dung beetle optimization algorithm in multimodal and complex optimization problems, thereby improving search efficiency and optimization accuracy. Finally, a mutation strategy with a higher tail probability density t-distribution is introduced to enhance the search performance of the dung beetle optimization algorithm, forming an improved dung beetle optimization algorithm.

9. The method according to claim 4, characterized in that, The improved dynamic window method specifically involves the following steps: In obstacle trajectory prediction, an extended Kalman filter algorithm is introduced into the dynamic window method, incorporating the prediction results into the cost function. In robot local trajectory planning, considering the relationship between the robot's azimuth angle, travel distance, and speed fluctuations relative to the target point during operation, the following improvements are made: First, the azimuth angle evaluation sub-function is improved by introducing the distance relationship with the target point and dynamically adjusting the azimuth angle weights. Second, the distance evaluation sub-function is improved by introducing new evaluation metrics. As the target point approaches, the weight of the trajectory closest to the target point is gradually increased; finally, the velocity smoothness evaluation sub-function is improved, and a new evaluation index is introduced. Furthermore, the evaluation function is comprehensively improved to predict the future position of dynamic obstacles, plan avoidance paths in advance, and achieve smooth optimization of local path trajectories.

Images