Mine underground safety prevention and control method, device and equipment based on vehicle-mounted equipment and storage medium

Abstract

The application provides a mine underground safety prevention and control method and device based on a vehicle-mounted device, equipment and a storage medium, and relates to the technical field of mine safety management and control. Through local preprocessing of the vehicle-mounted end to eliminate environmental noise interference, real-time fusion and global optimization of data are realized in combination with an edge-cloud collaborative architecture, and a network interruption autonomous mechanism and a dynamic risk assessment control algorithm are innovatively introduced, effectively solving the problems of sensing failure and network interruption safety loss of control caused by extreme working conditions underground, solving the reliability problem of environmental sensing data in an extreme environment underground, realizing local autonomous safety control when the network is interrupted, and improving the real-time performance and reliability of mine underground safety prevention and control.

Description

Methods, devices, equipment, and storage media for underground mine safety control based on vehicle-mounted equipment. Technical Field

[0001] This application relates to the field of mine safety management and control technology, and in particular to a method, device, equipment and storage medium for underground mine safety control based on vehicle-mounted equipment. Background Technology

[0002] As a core support for national strategic resource security and industrial development, mines face extreme operating conditions in their underground mining environments, including high temperatures, high humidity, strong dust, and sudden changes in light intensity, posing severe challenges to safety control devices. In deep mining operations, environmental sensing equipment often suffers from limited visibility due to pervasive dust, and low-light conditions render visual systems ineffective. Single sensors, such as pure vision or UWB positioning devices, are difficult to operate stably and are susceptible to interference from metal structure reflections or partial obstruction, leading to data loss and misjudgments. Simultaneously, mines suffer from weak digital infrastructure, with safety monitoring and risk warnings heavily reliant on human experience and lacking real-time quantification capabilities for dynamic human-vehicle interaction scenarios. In narrow tunnels, personnel frequently interact with large mobile equipment such as loaders, resulting in significant blind spots. Traditional fixed-threshold safety strategies cannot adapt to dynamic changes in risk levels, leading to frequent false alarms and missed alarms. More importantly, existing system architectures rely heavily on cloud computing. When the underground network fails due to interference or interruption, the on-board terminal immediately loses its intelligent decision-making capabilities, creating a "window of opportunity" for safety supervision, which can easily trigger mechanical injury accidents in emergencies. Furthermore, the control strategy unilaterally emphasizes safety objectives while neglecting production efficiency and operational comfort, making it difficult to meet actual production needs. The system as a whole has significant deficiencies in environmental adaptability, reliability, and real-time performance.

[0003] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0004] This application provides a method, device, equipment, and storage medium for underground mine safety control based on vehicle-mounted equipment. It has advantages such as strong environmental adaptability, high robustness to network interruptions, and precise real-time control.

[0005] Firstly, the mine underground safety control method based on vehicle-mounted equipment provided in this application adopts the following technical solution:

[0006] A method for underground mine safety control based on vehicle-mounted equipment includes:

[0007] The vehicle-mounted terminal collects downhole environmental data through multi-source sensors and performs local preprocessing on the data, including noise reduction and format conversion.

[0008] A lightweight target detection model is deployed on the vehicle to process the pre-processed data in real time, identify targets, and generate preliminary perception results.

[0009] The edge server receives preliminary perception results uploaded by multiple vehicle terminals within the same area for secondary reasoning and multimodal data fusion, complex model reasoning, and local map management;

[0010] The cloud platform aggregates data from the entire mine for global monitoring, algorithm model training and iteration, and distributes the optimized model parameters to edge servers and vehicle terminals;

[0011] The edge server and the vehicle-mounted terminal run a control algorithm based on dynamic risk assessment in parallel and establish a heartbeat detection mechanism. When the vehicle-mounted terminal detects that the network interruption exceeds the threshold, it immediately switches to the network outage autonomous mode and directly executes braking control through the vehicle controller based on the latest historical data cached locally and real-time sensor data.

[0012] Optionally, the multi-source sensors include at least an industrial camera, a lidar, and an inertial measurement unit; the local preprocessing includes passing-through filtering to denoise the laser point cloud, and performing grayscale correction and JPG compression format conversion on the image.

[0013] Optionally, the lightweight target detection model is a pruned and optimized YOLO11-N model, specifically designed to identify personnel, mining vehicles, and fixed equipment in underground mining scenarios; the inference frame rate for real-time processing is no less than 10fps.

[0014] Optionally, the multimodal data fusion performed by the edge server specifically includes:

[0015] For targets with a confidence level below 0.7 uploaded from the vehicle terminal, a second inference is performed using the more computationally powerful YOLO11-L model to improve the completeness of target detection.

[0016] Based on the real-time lighting stability and dust concentration in the well, the fusion weight of visual data and laser point cloud data is dynamically adjusted: when the lighting is stable and the dust concentration is low, the visual data is assigned a weight of 0.6 to 0.7.

[0017] When vision is obstructed or there is significant dust interference, assign a weight of 0.5 to 0.6 to the laser point cloud data.

[0018] Optionally, the local map management employs the following steps to achieve high-precision map construction and updating in large-scale scenes:

[0019] The LIO-SAM framework, based on tight coupling of multiple sensors, generates local point cloud sub-maps. This framework provides high-frequency initial pose values ​​through visual odometry and performs feature point matching and residual optimization through laser odometry to provide accurate pose correction.

[0020] An incremental map stitching strategy is adopted. First, the FPFH fast point feature histogram of the sub-map is extracted for coarse registration. Then, an improved ICP algorithm with a distance weight factor is used for fine registration to minimize the Euclidean distance error between point clouds.

[0021] The quality of the registered sub-maps is evaluated, and the sub-maps are only integrated into the global map when the map matching error MME ≤ 0.05m and the map point cloud variance MPV ≤ 0.03.

[0022] Establish a timed or triggered map update mechanism to incrementally update roadway areas that have changed due to mining operations and to replace outdated areas without loss.

[0023] Optionally, the control algorithm based on dynamic risk assessment uses a risk assessment model that integrates the following multi-source heterogeneous data to calculate a comprehensive risk value:

[0024] The relative distance, relative speed, and trajectory of the target acquired by the sensing system;

[0025] The absolute coordinates and speed of the vehicle are calculated by comparing the point cloud scanned in real time by the vehicle-mounted LiDAR with the preset global map.

[0026] Centimeter-level precise positioning data, provided by UWB smart bracelets worn by personnel, is used to compensate for blind spots in perception.

[0027] Optionally, the control algorithm is a multi-objective optimization adaptive braking strategy, and its specific execution logic includes:

[0028] The calculated comprehensive risk value is dynamically divided into four levels: low risk, medium risk, medium-high risk, and high risk.

[0029] Different braking modes are triggered based on the real-time risk level: emergency braking mode is activated when the risk is high to prioritize safety; forced braking mode is activated when the risk is medium to high to quickly shorten the safe distance; and gradual braking is used when the risk is medium or lower to balance work efficiency and operating comfort.

[0030] By setting a maximum deceleration limit of ≤3m / s² and employing a brake valve opening maintenance mechanism, stable deceleration output is maintained, thus avoiding braking shock.

[0031] Secondly, this application provides a mine underground safety control device based on vehicle-mounted equipment, which performs the method described above, including:

[0032] The data acquisition module is used to collect downhole environmental data from multiple sources via vehicle-mounted sensors and to perform local preprocessing on the data, including noise reduction and format conversion.

[0033] The data recognition module is used to deploy a lightweight target detection model on the vehicle to process the pre-processed data in real time, identify targets, and generate preliminary perception results.

[0034] The data processing module is used by the edge server to receive the preliminary perception results uploaded by multiple vehicle terminals within the same area for secondary reasoning and multimodal data fusion, complex model reasoning, and local map management.

[0035] The data training module is used to aggregate data from the entire mine on the cloud platform for global monitoring, algorithm model training and iteration, and to distribute the optimized model parameters to the edge server and vehicle terminal.

[0036] Thirdly, this application provides a computer device, the device comprising: a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method described above.

[0037] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method described above.

[0038] In summary, this application eliminates environmental noise interference through local preprocessing on the vehicle-mounted end, achieves real-time data fusion and global optimization through an edge-cloud collaborative architecture, and innovatively introduces a network outage self-governance mechanism and dynamic risk assessment and control algorithm. This effectively solves the problems of perception failure and network interruption safety control caused by extreme underground working conditions, solves the reliability problem of environmental perception data in extreme underground environments, realizes local autonomous safety control during network interruption, and improves the real-time performance and reliability of underground mine safety control. Attached Figure Description

[0039] Figure 1 is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of this application;

[0040] Figure 2 is a flowchart illustrating the first embodiment of the mine underground safety control method based on vehicle-mounted equipment of this application.

[0041] Figure 3 is a structural block diagram of the first embodiment of the mine underground safety control device based on vehicle-mounted equipment of this application. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] Referring to Figure 1, which is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of this application.

[0044] As shown in Figure 1, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0045] Those skilled in the art will understand that the structure shown in Figure 1 does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0046] As shown in Figure 1, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a mine underground safety control program based on vehicle-mounted equipment.

[0047] In the computer device shown in Figure 1, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this application can be set in the computer device. The computer device calls the mine underground safety control program based on vehicle-mounted equipment stored in the memory 1005 through the processor 1001 and executes the mine underground safety control method based on vehicle-mounted equipment provided in the embodiment of this application.

[0048] This application provides a method for underground mine safety control based on vehicle-mounted equipment. Referring to Figure 2, which is a flowchart of the first embodiment of the method for underground mine safety control based on vehicle-mounted equipment, this application provides a method for underground mine safety control based on vehicle-mounted equipment.

[0049] In this embodiment, the mine underground safety control method based on vehicle-mounted equipment includes the following steps:

[0050] Step S10: The vehicle-mounted terminal collects downhole environmental data through multi-source sensors and performs local preprocessing on the data, including noise reduction and format conversion.

[0051] It should be noted that in traditional underground operations in metal mines, the environmental data collected by sensing equipment is easily distorted due to extreme working conditions such as high temperature, high humidity, strong dust, and sudden changes in light. This results in the inability of the accuracy and real-time performance of safety sensing to meet the needs of dynamic mining. At the same time, the instability of network communication causes the system to lose control in the event of an interruption, creating a break in the safety supervision process in the spatiotemporal dimension. As a result, the risks of human-machine collaborative operations cannot be effectively identified and avoided.

[0052] For example, in a dynamic filling mining area of ​​a copper-nickel mine with a mining depth of over 1,300 meters, loaders and workers move crisscrossing in narrow tunnels; strong dust causes the image information captured by the visual sensors to become blurred, sudden changes in lighting cause the camera to malfunction temporarily, and network signal fluctuations cause data transmission interruptions between the vehicle-mounted terminal and the edge server, so the initial perception results cannot be uploaded in time, and the risk warning mechanism is delayed in being triggered.

[0053] The existing technology mainly has four main problems:

[0054] 1. The sensing system has poor environmental adaptability:

[0055] In complex environments such as underground mining with high dust levels, sudden changes in lighting, and frequent human-vehicle interactions, single-type sensors (such as pure vision systems or UWB ranging alarm wristbands) are easily interfered with and fail. Vision systems are susceptible to occlusion, while UWB wristbands suffer from false alarms and missed alarms. Due to the lack of complementary and fusion mechanisms for multimodal sensing, the overall stability and completeness of the system's sensing results are poor.

[0056] 2. Insufficient map building and positioning capabilities:

[0057] Existing SLAM algorithms heavily rely on the richness and uniqueness of environmental features. In environments like mine roadways, where features are sparse and highly repetitive, matching failures and severe pose estimation drift are common problems. Furthermore, the lack of effective incremental map updates and full lifecycle management mechanisms makes them unable to adapt to dynamic changes in roadways caused by mining and backfilling operations, resulting in rapidly outdated maps and poor usability.

[0058] 3. The security management strategy is rigid and has low reliability:

[0059] Traditional safety strategies are mostly based on fixed thresholds (such as fixed safety distances), which cannot accurately quantify and classify the risks of dynamically changing human-vehicle interactions, leading to frequent false alarms and missed alarms. Control systems heavily rely on cloud-based central computing; once the underground network is interrupted, the on-board unit loses its intelligent decision-making capability, creating a "control window" and posing a serious safety hazard. Control strategies often only consider the single objective of safety, neglecting production efficiency and operational comfort, making them difficult to promote and apply in actual production.

[0060] 4. Centralized system architecture:

[0061] Existing systems often adopt a simple architecture of "sensor-cloud" or "sensor-vehicle terminal", which fails to fully utilize the capabilities of edge computing, resulting in bottlenecks in real-time performance, reliability and scalability.

[0062] If the above problems are not resolved, the safety supervision gap will persist for a long time, the risk of mechanical injury in human-machine collaborative operations cannot be monitored and responded to in real time, the safety prevention and control system will be difficult to form a closed loop, and the continuity and controllability of the underground operation process will be fundamentally restricted.

[0063] In practical implementation, the vehicle-mounted terminal, as the first-line execution unit directly connected to the operation scenario, is designed with "real-time response, offline controllability, and lightweight deployment" as its core design principles. It is equipped with an industrial-grade embedded controller (computing power ≥8 TOPS, supporting TensorRT acceleration) and multiple types of sensor interfaces (compatible with Gigabit Ethernet and CAN bus). The key is to achieve efficient perception through the lightweight YOLO11. The specific functions and technical details are as follows:

[0064] Real-time data acquisition and preprocessing: Four types of core data are acquired simultaneously, including visual data, laser point cloud data, inertial navigation data, and vehicle status data (engine speed, braking pressure, etc., read in real time via CAN bus). After acquisition, the local preprocessing module performs data denoising (laser point cloud pass-through filtering, image grayscale correction) and format conversion (point cloud PCD to binary, image JPG compression) to ensure lightweight data transmission and efficient subsequent processing.

[0065] It should be noted that multi-source sensors refer to various sensing devices used to collect downhole environmental data. These can be implemented using acoustic sensors, infrared sensors, or temperature sensors. For example, an acoustic sensor can be used to detect the distance to obstacles. This is primarily to achieve comprehensive environmental data collection to address the problem of single sensor failure. Specifically, local preprocessing refers to the process of denoising and format conversion of the collected raw data. This can be achieved using wavelet transform denoising or PNG format conversion. For example, wavelet transform can be used to eliminate image noise caused by dust. This is mainly to improve data quality and ensure the reliability of subsequent processing. Furthermore, lightweight target detection models refer to compressed and optimized neural network models. These can be implemented using MobileNet or SqueezeNet models. For example, deploying a MobileNet model for target recognition is mainly to achieve real-time processing with low computational resource consumption. The autonomous mode during network outages refers to a control mechanism that operates independently on the vehicle during network interruptions. This can be implemented using predictive control algorithms based on historical data. For example, cached data can be used to predict vehicle trajectories and execute braking. This is mainly to achieve safe control under network outage conditions to avoid loss of control. Therefore, this embodiment, through a layered collaborative architecture including local processing on the vehicle end, fusion processing on the edge server, and global optimization in the cloud, combined with a network outage self-governance mechanism, effectively solves problems such as inaccurate safety perception, poor real-time performance, and loss of control during network interruptions in harsh environments such as high temperature, high humidity, strong dust, and sudden changes in light in underground mines. It achieves efficient and reliable safety control, reduces the risks of human-machine collaborative operations, and eliminates safety supervision gaps. As a preferred implementation, the control algorithm, which runs in parallel on the edge server and the vehicle end, is based on dynamic risk assessment. This can be achieved using a risk calculation model that integrates multi-source data fusion, such as combining distance and speed parameters to generate risk levels, primarily to ensure the real-time accuracy of risk assessment. Thus, this architecture ensures the continuity of perception data and the timeliness of control decisions under extreme operating conditions, avoiding safety hazards caused by environmental interference or communication failures.

[0066] In practical implementation, under the harsh working conditions of high temperature, high humidity, strong dust, and sudden changes in light in underground mines, this safety control method achieves efficient and reliable operation through a layered collaborative architecture. Specifically, the vehicle-mounted terminal uses multi-source sensors to collect underground environmental data. The collected data undergoes local preprocessing operations such as noise reduction and format conversion to eliminate the influence of environmental interference factors on the original signal and ensure the reliability of the input data. The preprocessed data is then processed in real time by a lightweight target detection model deployed on the vehicle-mounted terminal. This model performs target recognition tasks based on the optimized data format, generating preliminary perception results, thereby maintaining recognition stability in dynamic scenes.

[0067] The edge server receives preliminary perception results uploaded from multiple vehicle-mounted terminals within the same mining area. Through secondary inference, multimodal data fusion, complex model inference, and local map management operations, it integrates complementary spatial information from multiple sources to improve perception integrity. The cloud platform aggregates data from the entire mine to perform global monitoring, algorithm model training, and iteration tasks, and distributes the optimized model parameters to the edge server and vehicle-mounted terminals to achieve continuous algorithm optimization. The edge server and vehicle-mounted terminals run a control algorithm based on dynamic risk assessment in parallel and establish a heartbeat detection mechanism. When the vehicle-mounted terminal detects that continuous network interruptions exceed a preset threshold, it immediately switches to a network-disconnected autonomous mode. This mode, based on locally cached recent historical data and real-time sensor data, directly executes braking control through the vehicle controller, avoiding response delays caused by communication failures.

[0068] For example, in a roadway operation scenario in an underground panel, after the vehicle-mounted equipment collects environmental data, the local preprocessing module performs targeted noise filtering and format standardization on the sensor signals. Based on this, the lightweight target detection model identifies people or obstacles ahead in real time and generates preliminary perception results. The edge server receives the results uploaded by multiple vehicle-mounted devices in the same area, performs data correlation analysis and fusion processing. When the network is suddenly interrupted due to dust interference, the vehicle-mounted terminal determines the interruption status based on the heartbeat detection mechanism, and then calls the locally cached historical trajectory data and real-time sensor input. The vehicle controller directly outputs braking commands to avoid collision risks.

[0069] Therefore, the method in this embodiment effectively suppresses the negative impact of harsh environments on data quality through local preprocessing, the real-time processing mechanism of the lightweight model ensures the response speed in dynamic scenarios, the multi-source data fusion operation of the edge server makes up for the blind spots of single-point perception, the cloud model iteration process continuously improves the algorithm's adaptability to complex underground working conditions, and the introduction of the network-disconnected autonomous mode fundamentally eliminates the safety supervision breakpoints caused by network dependence, thereby realizing closed-loop prevention and control of underground safety risks throughout the entire process and significantly reducing the risk of mechanical injury and collision accidents in human-machine collaborative operations.

[0070] In its implementation, this embodiment proposes multi-source sensors and local preprocessing to collect and preprocess underground environmental data. However, during its implementation, due to the extreme harshness of the underground environment (such as high temperature, high humidity, strong dust, and sudden changes in light), the general sensor selection and preprocessing methods lack specificity, resulting in significantly increased noise, incompatible formats, and insufficient stability of the collected data. This, in turn, affects the accuracy and real-time performance of subsequent sensing results and fails to meet the stringent requirements of mine safety control for data reliability.

[0071] In response, this embodiment proposes that the multi-source sensors include at least an industrial camera, a LiDAR, and an inertial measurement unit; local preprocessing includes direct-pass filtering to denoise the laser point cloud, and grayscale correction and JPG compression format conversion of the image.

[0072] Among them, multi-source sensors refer to integrated systems composed of multiple heterogeneous sensing units. These systems can utilize industrial cameras to acquire visual information within the visible spectrum, lidar to collect three-dimensional spatial point cloud data, and inertial measurement units to monitor the motion state parameters of the carrier. The aim is to overcome the failure risk of a single sensor under extreme conditions through multimodal sensing complementarity. Pass-through filtering denoising refers to a noise suppression method that filters point cloud data based on spatial distribution characteristics. This can be achieved by setting a height threshold to remove suspended noise above the ground or filtering far-field interference points based on a distance threshold. The aim is to specifically eliminate false point clouds caused by dust while retaining effective structural information. Grayscale correction is an image enhancement technique that dynamically adjusts the brightness distribution of an image. This can be achieved by using linear compensation or nonlinear mapping functions based on the average brightness of local areas to remap pixel values. The aim is to compensate for brightness distortion caused by sudden changes in illumination and maintain the consistency of visual data. JPG compression format conversion refers to the process of converting image data into a standard compression format. This can be achieved by using lossy compression algorithms to reduce data volume while retaining key edge features. The aim is to adapt to the limited transmission bandwidth and storage resources of the vehicle-mounted terminal.

[0073] Specifically, the solution proposed in this application enhances the robustness of environmental perception through a collaborative acquisition mechanism of multi-source sensors. An industrial camera provides a continuous visual information stream in scenarios with sudden changes in lighting, while a lidar utilizes its low sensitivity to dust interference to acquire precise distance data to compensate for blind spots. An inertial measurement unit simultaneously records vehicle pose changes. Subsequently, the local preprocessing stage sequentially performs pass-through filtering to remove dust noise from the lidar point cloud, grayscale correction to dynamically balance image brightness fluctuations, and JPG compression format conversion to optimize data transmission efficiency. This forms a data purification process customized for the specific working conditions underground, ensuring that the preprocessed environmental data has a high signal-to-noise ratio and format compatibility, providing a reliable input foundation for subsequent real-time target detection.

[0074] For example, in underground mining operations, the vehicle-mounted unit integrates an industrial-grade CMOS camera with an explosion-proof shell as a vision sensor unit, selects a 905nm band solid-state LiDAR for 3D point cloud acquisition, and is equipped with a six-axis microelectromechanical system inertial measurement unit to monitor vehicle acceleration and angular velocity. In the local preprocessing stage, the LiDAR point cloud data is subjected to a pass-through filtering operation based on the tunnel height distribution to filter suspended dust noise, the image data is compensated for local lighting differences in the tunnel using an adaptive grayscale correction algorithm, and the processed image is converted into JPG format to reduce data volume, and finally outputs a standardized environmental perception data stream.

[0075] The above solution effectively suppresses data noise introduced by the harsh underground environment, ensures the uniformity of format and time synchronization of multi-source sensing data, and significantly improves the reliability and real-time processing capability of environmental data, thereby meeting the stringent requirements of mine safety control devices for the quality of raw data.

[0076] Step S20: Deploy a lightweight target detection model on the vehicle to process the preprocessed data in real time, identify targets, and generate preliminary perception results.

[0077] It should be noted that the lightweight intelligent algorithm deployment focuses on "low computational power consumption and high real-time response," deploying customized and optimized lightweight algorithms. Target detection uses a pruned YOLO11-N (Nano version) model, specifically adapted to target features such as miners, loaders, and fixed equipment; point cloud clustering uses the SA-DBSCAN algorithm (computational complexity O(n)) to quickly distinguish obstacles from the background; the local emergency control algorithm has a built-in "risk threshold matrix," pre-setting braking strategies for different operating scenarios such as flat tunnels and uphill / downhill sections, and can dynamically adjust braking intensity according to vehicle load and speed.

[0078] In practical implementation, the lightweight target detection model proposed above is used to process underground environmental data and identify targets in real time. However, in its implementation, due to the extreme conditions of the underground mining environment, such as high temperature, high humidity, strong dust and sudden changes in light, the general lightweight model is difficult to balance recognition accuracy and processing speed under limited on-board computing resources. This makes target detection susceptible to interference, resulting in missed detections or false detections. Moreover, the insufficient frame rate cannot meet the real-time response requirements of dynamic scenes, thus affecting the reliability and timeliness of safety control devices.

[0079] To address this, this embodiment further proposes that the lightweight target detection model is a pruned and optimized YOLO11-N model, specifically designed for identifying personnel, mining vehicles, and fixed equipment in underground mining scenarios; the inference frame rate for real-time processing is no less than 10fps.

[0080] Among them, the pruned and optimized YOLO11-N model refers to the technique of reducing the model size by removing redundant connections in the neural network. It can be implemented by structured pruning or unstructured pruning, such as removing low-weight neuron connections or simplifying the number of channels in convolutional layers. Its purpose is to reduce the consumption of computing resources on the vehicle end while maintaining the ability to identify specific targets in the mine. An inference frame rate of not less than 10fps can be understood as the requirement that the model processes at least 10 frames of images per second. This can be achieved through lightweight model design, hardware accelerator configuration, or input resolution adjustment. Its purpose is to ensure continuous and timely perception results in dynamic underground scenes.

[0081] Specifically, the solution in this embodiment deploys a pruned and optimized YOLO11-N model on the vehicle-mounted terminal to perform real-time inference on the preprocessed underground environment data. This model is customized for mining scenarios and focuses on identifying personnel, mining vehicles, and fixed equipment. At the same time, by setting a hard requirement that the inference frame rate is no less than 10fps, the model can maintain sufficient processing speed under limited computing resources, thereby stably outputting continuous frame perception results under dust interference and lighting changes, providing timely and reliable input for subsequent safety decisions.

[0082] For example, a lightweight object detection model can be deployed on an embedded AI accelerator in an vehicle that supports INT8 quantization inference; the model input is locally preprocessed image data, and a stable frame rate output is achieved through an optimized network structure.

[0083] The above-mentioned solution effectively improves the robustness of target detection in the underground environment, reduces identification errors caused by environmental interference, and ensures that the perception system can respond to dynamic scene changes in a timely manner, providing a reliable basis for safety control.

[0084] Step S30: The edge server receives the preliminary perception results uploaded by multiple vehicle terminals within the same area for secondary reasoning and multimodal data fusion, complex model reasoning, and local map management.

[0085] In practical implementation, edge servers (edges) are deployed in the underground data center, using industrial-grade rack servers (CPU ≥ 16 cores, GPU ≥ 24GB video memory, RAM ≥ 64GB), supporting temperatures from -20℃ to 60℃ and a dustproof and waterproof rating of IP65. As a core node bridging the gap between upstream and downstream systems, they undertake computationally intensive tasks and regional collaborative management, primarily achieving precise perception fusion through the high-performance version of YOLO11. Specific functions are as follows:

[0086] Multimodal fusion perception optimization based on YOLO11: Leveraging the computing power of edge GPUs, YOLO11-L (medium-high computing power version) is deployed to process scattered perception data uploaded by multiple vehicle-mounted terminals (equipped with YOLO11-N) within the same area through "secondary inference - dynamic weighting - accurate output," overcoming the perception limitations of single sensors and lightweight vehicle-mounted models. For low-confidence targets (such as distant personnel or partially occluded equipment) that are not identified by the vehicle-mounted YOLO11-N due to computing power limitations, the more complex feature extraction network of YOLO11-L is used for re-inference, improving the completeness of target detection. The weights of visual and laser data fusion are dynamically adjusted according to the real-time underground environment—visual confidence is higher when lighting is stable and dust concentration is low, with a weight allocation of 0.6~0.7; when vision is obstructed (such as equipment shadows or dust clumps), laser ranging data (accuracy ±10cm) is more reliable, with a weight adjustment of 0.5~0.6. The final output is "target category (such as personnel, fully loaded loader) + three-dimensional coordinates (x / y / z, based on laser point cloud calibration) + straight-line distance between the vehicle terminal and the target + fusion confidence (≥0.7 is considered valid)", which improves the perception accuracy by 35% compared to the single vehicle terminal YOLO11 and reduces the false detection rate (such as misjudging the texture of the tunnel wall as equipment) by 25%.

[0087] Complex Model Inference and Decision Support: Leveraging YOLO11-L's segmentation capabilities, this feature enables segmentation inference for specific targets in mines (such as "personnel wearing UWB wristbands," "empty / fully loaded loaders," and "vehicles on flat / uphill slopes"), providing more accurate target attributes for risk assessment. For example, personnel wearing UWB wristbands can be linked to positioning system data to supplement blind spots; fully loaded loaders and vehicles on uphill slopes require corresponding adjustments to braking distances (the braking distance of a fully loaded loader increases by 20% compared to an empty one, and the braking threshold for downhill vehicles is 1.5 times earlier). Using the target location sequence continuously detected by YOLO11-L (sampling frequency 10Hz) as input, an LSTM time-series prediction model is trained to output the target's movement trajectory within the next 5 seconds. This allows for the early identification of potential risks such as personnel crossing roadways and equipment turning and merging, providing sufficient braking response time for the vehicle-mounted system.

[0088] Local map management and iterative updates: Responsible for the full lifecycle management of a single area's map. Map construction adopts an "incremental stitching" mode. After receiving the sub-map data from the vehicle, it first performs coarse registration using FPFH features (matching is achieved when the overlapping area is ≥30%), and then performs fine registration using an improved ICP algorithm (introducing a distance weight factor) to ensure that the sub-map stitching error is ≤0.1m. Map updates support a dual mode of "triggered + timed". When the vehicle detects changes in the roadway structure (such as the addition of equipment or movement of the filling body boundary), it actively triggers an update request, or automatically starts an update every 2 hours, and performs lossless replacement of outdated areas that have not been updated for more than 7 days. A built-in map quality assessment module selects effective sub-maps based on three indicators: "point cloud density ≥50 points / ㎡, matching error MME ≤0.05m, and flatness MPV ≤0.03", ensuring map reliability.

[0089] Local data caching and scheduling: Construct a hierarchical storage system of "hot data + historical data". Hot data (maps of the past 24 hours, YOLO11 perception results, vehicle trajectories) is stored on SSD hard drives, supporting fast retrieval by the vehicle terminal; historical data (more than 24 hours) is stored on mechanical hard drives (capacity ≥ 2TB), and archived by "disk area - date - vehicle ID - YOLO11 version" to facilitate subsequent model iteration (such as optimizing YOLO11 anchor boxes based on historical false detection samples).

[0090] In practical implementation, the above-mentioned multimodal data fusion using edge servers is used to integrate multi-source sensing data. However, in its implementation process, there is a lack of targeted processing mechanism for targets with low confidence, and the fixed fusion weight cannot adapt to the dynamic environment of sudden changes in underground lighting and fluctuations in dust concentration, resulting in insufficient target detection integrity and decreased fusion accuracy, making it difficult to meet the stringent requirements of sensing reliability for underground mine safety control.

[0091] To address this, this embodiment further proposes that the multimodal data fusion performed by the edge server specifically includes: the edge server performs confidence screening on the preliminary perception results uploaded from the vehicle terminal; for targets with a confidence level lower than a preset threshold (e.g., 0.7), it calls the locally deployed YOLO11-L model for secondary inference. This model has a deeper convolutional structure and attention mechanism, enabling it to perform feature re-extraction and contextual analysis on low-confidence targets, significantly improving the detection rate and recognition accuracy in cases of occlusion, blurriness, or small targets.

[0092] Meanwhile, the edge server has a built-in adaptive multimodal fusion engine that receives real-time light intensity data and dust concentration monitoring values ​​from environmental sensors, and dynamically calculates the fusion weights of visual data and laser point cloud data by combining the output of the visual occlusion detection module. This is achieved through the following steps:

[0093] Confidence screening and secondary inference mechanism: If the target confidence score is lower than 0.7 for the perception results uploaded by the vehicle terminal, the secondary inference process of the YOLO11-L model in the edge server is triggered to improve the completeness of target recognition by enhancing feature extraction and context modeling.

[0094] Adaptive weight allocation model: A dynamic weight calculation model based on environmental perception is established. Based on real-time collected data including light intensity variance, dust concentration sensor data, and visual occlusion detection results, the fusion weights of visual data and laser point cloud data are dynamically allocated. The weight calculation uses the following formula:

[0095]

[0096]

[0097] in, For visual data weights, The light stability coefficient (value range 0~1). The normalized value of dust concentration is (0~1). This is the reliability coefficient of the lidar in a dusty environment, typically ranging from 1.2 to 1.5.

[0098] Weighting and data fusion: Under stable lighting and low dust concentration ( ≥0.8, ≤0.2), visual data weights When the accuracy is between 0.6 and 0.7, the system primarily relies on visual detection results for target tracking and localization; however, when vision is obstructed or there is significant dust interference ( ≤0.4 or ≥0.6), weight of laser point cloud data The value is increased to between 0.5 and 0.6, and enhanced filtering and feature completion processing of point cloud data are triggered to ensure that reliable environmental perception capabilities can still be maintained under visual degradation.

[0099] Among them, confidence level refers to the reliability quantification index of the target detection model output results, which can be expressed in the form of probability values. The specific threshold can be dynamically set within the range of 0.65 to 0.75 according to the downhole dust concentration and lighting conditions. Its purpose is to accurately screen low reliability detection results that require secondary verification. The YOLO11-L model is a deep learning target detection model with stronger computing power. It can be implemented using a convolutional neural network-based architecture. Its purpose is to perform deep feature extraction and re-identification of targets that are susceptible to environmental interference. Dynamic adjustment of fusion weights refers to the real-time calculation of the contribution ratio of multimodal data based on environmental parameters. It can be implemented based on a preset rule base or a lightweight environmental perception algorithm. Its purpose is to ensure that the fusion process closely follows the real-time changes in downhole working conditions.

[0100] Specifically, the solution in this embodiment first performs confidence screening on the preliminary perception results uploaded by the vehicle-mounted terminal. When a target with a confidence level below a threshold is detected, a secondary inference process is immediately triggered, utilizing the more computationally powerful YOLO11-L model to refine the specific target. Simultaneously, the system continuously collects downhole environmental parameters and dynamically calculates and applies corresponding fusion weights based on real-time monitoring data of illumination stability and dust concentration, ensuring that visual data and laser point cloud data can achieve optimal performance under different operating conditions. This mechanism achieves organic synergy between confidence screening, secondary inference, and weight adjustment, ensuring that the perception system can maintain high-precision target detection and data fusion even during sudden environmental changes.

[0101] For example, when the dust concentration in a certain working area underground suddenly increases due to blasting operations, the system automatically adjusts the weight of the laser point cloud data to 0.55 and starts the secondary inference process of the YOLO11-L model for targets with a confidence level of less than 0.7. By enhancing the contribution ratio of the point cloud data and the depth recognition of key targets, the system effectively compensates for the performance degradation of the visual data, thereby accurately identifying personnel and mining vehicles in the roadway in a high-dust environment.

[0102] The above solution effectively solves the problems of insufficient target detection completeness and decreased fusion accuracy, significantly improves the robustness and adaptability of the downhole sensing system in dynamic environments, and provides a stable and reliable environmental cognition foundation for safety control.

[0103] In specific implementation, this embodiment proposes local map management to support environmental modeling and real-time monitoring of underground safety control devices. However, in this process, due to the drastic dynamic changes in roadway structure in large-scale underground mining scenarios, high dust concentration leading to sensor data distortion, and the inadequacy of traditional map construction methods in balancing computational efficiency and accuracy, map matching errors accumulate and updates lag, making it difficult to meet the stringent requirements of high-precision safety control for real-time and reliable environmental perception.

[0104] To address this, this embodiment further proposes the following steps for local map management to achieve high-precision map construction and updating in large-scale scenarios: Based on the above method, a local point cloud sub-map is generated using a multi-sensor tightly coupled LIO-SAM framework. This framework provides initial high-frequency pose values ​​through visual odometry and performs feature point matching and residual optimization through laser odometry to provide accurate pose correction. An incremental map stitching strategy is adopted, first extracting the FPFH fast point feature histogram of the sub-map for coarse registration, and then using an improved ICP algorithm with a distance weight factor for fine registration to minimize the Euclidean distance error between point clouds. The registered sub-map is quality-assessed, and only when the map matching error MME ≤ 0.05m and the map point cloud variance MPV ≤ 0.03 is the sub-map integrated into the global map. A timed or triggered map update mechanism is established to incrementally update roadway areas changed due to mining operations and to replace outdated areas without loss.

[0105] Among them, the LIO-SAM framework refers to a synchronous localization and map construction method that integrates LiDAR and visual sensors. It can be implemented using a multi-sensor data tightly coupled fusion algorithm, aiming to improve the robustness of pose estimation to cope with complex downhole working conditions. The Fast Point Feature Histogram (FPFH) can be understood as a local geometric feature descriptor for point cloud registration. Specifically, it can be implemented using standardized modules in the point cloud feature extraction library, aiming to efficiently acquire point cloud structural features to support fast coarse registration. The improved ICP algorithm specifically introduces an iterative nearest-point optimization algorithm based on distance weight factors. For example, it can dynamically adjust the matching weights based on the spatial distribution characteristics of the point cloud, aiming to enhance the adaptability of the fine registration process to noise and occlusion. The quality assessment indicators MME and MPV can be understood as key parameters that quantify the map matching accuracy and point cloud consistency, used to objectively determine the reliability of the sub-map.

[0106] Specifically, the solution in this embodiment generates local point cloud sub-maps using the LIO-SAM framework. Visual odometry quickly provides high-frequency initial pose values ​​when downhole lighting conditions are relatively stable, avoiding positioning interruptions caused by single sensor failure. Simultaneously, laser odometry achieves precise pose correction through feature point matching and residual optimization, combining the dust-resistant capabilities of laser point clouds with the detail-enhancing properties of visual data to effectively overcome pose drift problems under harsh conditions. An incremental map stitching strategy is adopted, first extracting geometric features of the sub-maps using FPFH (Fast Point Feature Histogram) for coarse registration, significantly reducing the computational complexity of large-scale point cloud data. Then, an improved method incorporating distance weighting factors is used... The ICP algorithm performs precise registration, dynamically adjusting the point cloud matching weights based on the roadway geometry. This allows the algorithm to focus on reliable data points in areas with dust obscuring or structural abrupt changes, thereby minimizing the Euclidean distance error between point clouds. A dual-threshold quality assessment is applied to the registered sub-maps, integrating them into the global map only when both MME and MPV meet strict thresholds. This effectively filters out low-quality data caused by sensor noise and environmental interference. Furthermore, a timed or triggered map update mechanism is established. For roadway changes caused by mining operations, incremental updates quickly integrate new regional data, while a lossless replacement strategy is used to process older areas, avoiding the resource consumption of full reconstruction and ensuring the map model continuously reflects the latest underground environmental conditions.

[0107] In practical implementation, when the vehicle-mounted equipment moves in narrow tunnels, the tightly coupled multi-sensor LIO-SAM framework generates local point cloud sub-maps in real time. Visual odometry provides initial pose estimation, and laser odometry performs feature point matching and residual optimization to complete pose correction. The system then extracts the FPFH fast point feature histogram of the sub-map for coarse registration, and then applies an improved ICP algorithm with a distance weight factor for fine registration. After registration, MME and MPV indices are calculated. If the threshold requirements are met, the sub-map is integrated into the global map. At the same time, the system monitors the dynamic changes in mining operations, triggers incremental updates to integrate data from newly exposed areas, and performs non-destructive replacement of old areas that have become ineffective due to mining.

[0108] Through the above technical solution, this embodiment effectively solves the problems of map matching error accumulation and update lag in large-scale underground scenes, ensuring the real-time and reliability of environmental perception, and providing continuous and accurate environmental model support for safety control devices.

[0109] Step S40: The cloud platform gathers all the mine's data for global monitoring, algorithm model training and iteration, and distributes the optimized model parameters to the edge server and vehicle terminal.

[0110] The edge server and the vehicle-mounted terminal run a control algorithm based on dynamic risk assessment in parallel and establish a heartbeat detection mechanism. When the vehicle-mounted terminal detects that the network interruption exceeds the threshold, it immediately switches to the network outage autonomous mode and directly executes braking control through the vehicle controller based on the latest historical data cached locally and real-time sensor data.

[0111] In practical implementation, the cloud platform is deployed in the mining area's data center to achieve mine-wide data aggregation, model iteration, and global control. Specific functions include: Full-domain data aggregation and standardized processing: receiving perception results, positioning information, equipment status, map data, etc., uploaded from all edge nodes. After data access, standardization is achieved through a "cleaning-labeling-association" process: the cleaning module removes abnormal data (such as invalid point clouds caused by sensor malfunctions or incomplete images caused by network packet loss); the labeling module automatically adds scene tags (operation type, environmental conditions) to the data; and the association module establishes a four-dimensional data index of "vehicle-target-location-time," supporting data traceability and correlation analysis across edge nodes and time periods. All data is stored in a distributed database.

[0112] Algorithm model training and iterative deployment: Utilizing comprehensive big data for model optimization, a closed loop of "offline training - online evaluation - incremental update" is constructed. During offline training, core algorithms such as multimodal fusion models and risk assessment models are trained based on historical mine operation data, with cross-validation ensuring model generalization ability. In the online evaluation phase, the new model is deployed in a cloud-based testing environment, using real-time data to verify performance (such as target detection rate and risk assessment accuracy). Once the target is met, incremental updates are triggered, and the model is deployed in batches to edge devices and vehicle-mounted devices. The edge devices receive the data and synchronously update the inference model, while the vehicle-mounted devices replace the core parameters of the local lightweight algorithm. The entire iterative process does not affect the normal operation of the system.

[0113] It should be noted that this embodiment breaks away from the traditional mine intelligent management and control system's single control logic of "centralized (relying on cloud-based single computing)" or "standalone (relying on vehicle-mounted independent computing)," and constructs a three-level distributed intelligent collaborative system of "end-edge-cloud." The core differences are reflected in the following three aspects, and are deeply bound to the architecture design of each layer:

[0114] Differences in control logic: From "single node dependence" to "layered collaborative redundancy", the traditional centralized architecture relies on the cloud to uniformly process perception and decision-making tasks, and the vehicle terminal only serves as a "data acquisition terminal" without independent emergency control capabilities; the traditional stand-alone architecture only relies on local computing on the vehicle terminal, without edge / cloud collaborative optimization support.

[0115] This embodiment adopts a layered design of "vehicle terminal (local emergency control) + edge terminal (local optimization decision-making) + cloud terminal (global coordination)" to form a three-level control logic of "local autonomy + edge collaboration + cloud optimization". It also clearly defines the core responsibilities of each node (such as the vehicle terminal focusing on real-time emergency response, the edge terminal focusing on intensive computing, and the cloud terminal focusing on global iteration), which solves the pain point of traditional architectures that "either rely on communication or have insufficient capabilities".

[0116] Decentralization of key capabilities: From “cloud / single machine monopoly on computing” to “edge + vehicle terminal undertaking core tasks” In traditional architecture, computationally intensive tasks (such as multimodal fusion and complex model inference) either rely entirely on the cloud (leading to high latency) or are entirely pushed to the vehicle terminal (leading to insufficient computing power); and there is no dedicated “edge node” to undertake intermediate layer functions such as “local map management and high-frequency data caching”.

[0117] This embodiment pushes computationally intensive tasks such as "multimodal fusion perception, local map stitching, and complex model inference" to the edge, and pushes "real-time emergency control and lightweight perception" to the vehicle, while the cloud focuses only on "global data aggregation, model training, and operation and maintenance management"—this "capability decentralization + layered division of labor" design is a core innovation not addressed in traditional architectures.

[0118] Network outage adaptability: From "out of control due to network outage" to "autonomy of the vehicle terminal when the network is out of control", the traditional centralized architecture causes a management gap when the vehicle terminal lacks independent decision-making ability during network interruption; although the traditional stand-alone architecture can make decisions locally, it lacks model iteration and data support from the edge / cloud, resulting in limited decision accuracy.

[0119] This embodiment specifically designs the vehicle-mounted "autonomous operation without network" capability (with a clearly defined "braking decision based on the last known valid data + preset rules"), and at the same time, it forms data / model redundancy through "local caching at the edge + cloud model distribution", which solves the core pain point of "network dependence" in traditional architecture.

[0120] The corresponding beneficial effects are: high reliability: no loss of control even when the network is interrupted, and no gaps in management and control. Relying on the "autonomous governance when the network is interrupted" capability of the vehicle terminal, combined with the dual-end (vehicle + cloud) collaborative redundancy mechanism of "emergency control of the vehicle terminal + local decision-making of the edge terminal + global monitoring of the cloud", even if the network is interrupted, the vehicle terminal can immediately take over control, completely eliminating the "loss of control when the network is interrupted" problem of the traditional architecture. In actual testing, seamless switching of management and control was achieved, meeting the safety requirements of mining operations.

[0121] Low latency: Real-time response, adapted to high-speed mobile scenarios. As computationally intensive tasks (multimodal fusion, complex reasoning) are moved to the edge, lightweight perception and emergency control tasks are completed locally on the vehicle, which greatly reduces the communication latency of "data upload-cloud processing-command issuance" in the traditional centralized architecture. The response speed meets the real-time requirements of high-speed mobile devices in underground mines.

[0122] Strong scalability: Adapting to the dynamic development needs of mines, this embodiment adopts a modular layered architecture. Edge nodes can independently undertake the calculation and map management tasks of specific panels / segments. When adding a new work area, only the new edge node needs to be connected, without the need to reconstruct the entire system. At the same time, the cloud model can be iteratively optimized and then distributed to the edge / vehicle terminal to achieve flexible functional expansion. This solves the problem of "adding new areas requires expanding cloud computing power and is difficult to modify" in traditional centralized architectures, and adapts to the continuous development needs of mines.

[0123] Specifically, in some of the above-described embodiments of this example, a control algorithm based on dynamic risk assessment is proposed to realize underground safety braking decisions. However, in its implementation process, due to the extreme environment of underground mines such as high temperature, high humidity, strong dust and sudden changes in light, the reliability of single sensor data is low, and there are blind spots in visual and lidar perception. This makes it impossible for the risk assessment model to accurately capture target dynamics and personnel positions by relying only on limited data sources, resulting in distortion of the comprehensive risk value calculation, which in turn leads to misjudgment or response lag in braking strategies, making it difficult to effectively prevent human-machine collision accidents.

[0124] To address this, this embodiment further proposes the control algorithm based on dynamic risk assessment. Its risk assessment model integrates the following multi-source heterogeneous data to calculate the comprehensive risk value: the target's relative distance, relative speed, and trajectory obtained by the perception system; the vehicle's absolute coordinates and speed calculated by comparing the point cloud scanned in real time by the vehicle-mounted LiDAR with a preset global map; and centimeter-level precise positioning data provided by a UWB smart bracelet worn by the personnel to compensate for perception blind spots.

[0125] Among them, the target's relative distance, relative speed, and trajectory refer to the dynamic behavior parameters of surrounding objects relative to the vehicle. These can be achieved using point cloud clustering analysis based on vehicle-mounted LiDAR or visual target tracking algorithms. The purpose is to provide real-time motion state characteristics of the target, avoid misjudgments caused by environmental noise, and ensure the continuity and reliability of the basic data for risk assessment. The vehicle's absolute coordinates and speed refer to the vehicle's precise pose information in the underground tunnel. This can be achieved using a pose calculation method that matches real-time LiDAR scanning data with features from a high-precision map. The purpose is to use the stable reference frame of the global map to filter environmental interference factors and provide baseline data for risk assessment that is not affected by sudden environmental changes. The centimeter-level precise positioning data provided by the UWB smart bracelet refers to the high-precision coordinate information of the personnel's location. This can be achieved using ultra-wideband wireless communication technology combined with a multi-base station triangulation system. The purpose is to compensate for the perception loss of vehicle-mounted sensors in blind spots, enabling the risk assessment model to fully cover human-computer interaction scenarios.

[0126] Specifically, the risk assessment model receives and integrates target dynamic information output by the perception system, vehicle pose information calculated by comparing LiDAR and the global map, and personnel positioning data from UWB smart bracelets, forming a multi-dimensional input source. During data processing, the system dynamically adjusts the weight allocation of each data source based on real-time downhole lighting stability and dust concentration. For example, when dust concentration is high, the weight of visual data is reduced while the weight of LiDAR and UWB data is increased, thereby cross-validating the target's motion state and filtering environmental interference. Finally, the comprehensive risk value is calculated based on the weighted multi-source heterogeneous data, providing accurate risk assessment results for the braking control algorithm and ensuring the timeliness and effectiveness of safety control in extreme downhole conditions.

[0127] For example, during underground loader operations, when the vehicle enters a dusty tunnel area, the onboard industrial camera's target detection confidence level may fall below the threshold due to dust interference. In this situation, the risk assessment model primarily relies on vehicle coordinates calculated by LiDAR and personnel location data transmitted from a UWB smart bracelet. The UWB smart bracelet can be a wearable device integrating a UWB chip, and the LiDAR can be a 16-line mechanical LiDAR. The global map is constructed based on the LIO-SAM framework. By fusing this data, the system accurately identifies the relative positions of personnel and vehicles and triggers corresponding braking modes based on the calculated comprehensive risk value, effectively avoiding collision risks in areas where the vision system fails.

[0128] Through the above solution, this embodiment can significantly improve the robustness and accuracy of dynamic risk assessment in extreme environments, effectively reduce braking misjudgments or response delays caused by sensor data distortion, thereby reliably preventing man-machine collision accidents in mines and ensuring operational safety.

[0129] Specifically, in some of the above-described embodiments of this example, a control algorithm based on dynamic risk assessment is proposed. However, in its implementation, the risk assessment results fail to be effectively transformed into an adaptive braking behavior control mechanism. As a result, in the complex environment of dynamic changes in underground mines, the system cannot adjust the braking intensity and response speed in real time according to the degree of risk. This may lead to safety accidents due to delayed or insufficient braking response, or reduce work efficiency due to excessive braking. At the same time, the braking process lacks stability control and is prone to impact vibration, which affects vehicle handling safety and personnel comfort.

[0130] To address this, this embodiment further proposes a control algorithm that is a multi-objective optimized adaptive braking strategy, the specific execution logic of which includes:

[0131] The calculated comprehensive risk value is dynamically divided into four levels: low risk, medium risk, medium-high risk, and high risk.

[0132] Different braking modes are triggered based on the real-time risk level: emergency braking mode is activated when the risk is high to prioritize safety; forced braking mode is activated when the risk is medium to high to quickly shorten the safe distance; and gradual braking is used when the risk is medium or lower to balance work efficiency and operating comfort.

[0133] By setting a maximum deceleration limit of ≤3m / s² and employing a brake valve opening maintenance mechanism, stable deceleration output is maintained, thus avoiding braking shock.

[0134] Among them, dynamic classification of comprehensive risk value refers to automatically determining the risk level based on the real-time calculated risk value range. This can be achieved using an adaptive mapping algorithm based on continuous threshold boundaries. For example, by dynamically adjusting the threshold range to adapt to sudden changes in the downhole environment, the aim is to avoid the rigidity of fixed threshold processing and to accurately map the risk assessment results to different working conditions, providing fine-grained basis for braking decisions. Triggering different braking modes based on real-time risk levels can be understood as a differentiated braking behavior strategy preset for risk levels. This can include multiple parallel implementation methods such as emergency braking mode, forced braking mode, and gradual braking, with the aim of realizing the correlation between risk level and braking. Dynamic matching of braking intensity ensures safety as a priority in high-risk scenarios, while balancing operational efficiency and comfort in medium- and low-risk scenarios. The maximum deceleration limit of ≤3m / s² specifically refers to the maximum allowable deceleration during braking, which can be set to 3m / s² or lower. Its purpose is to constrain the upper limit of braking intensity and prevent the risk of vehicle loss of control due to sudden braking. The brake valve opening maintenance mechanism refers to the technical means of maintaining the stability of the brake valve state through control algorithms. It can be implemented using state feedback control or adaptive adjustment strategies. Its purpose is to eliminate acceleration abrupt changes caused by valve fluctuations and maintain the continuity of deceleration output.

[0135] Specifically, the solution in this embodiment inputs the real-time calculated comprehensive risk value into the dynamic classification module, which automatically maps it to four risk levels based on preset continuous threshold boundaries. This classification directly triggers the corresponding braking mode execution unit: when the risk level is determined to be high, the system immediately activates the emergency braking mode to achieve safe braking with the shortest response time; when determined to be medium-high risk, the system initiates the forced braking mode, quickly shortening the safe distance by optimizing braking intensity; when determined to be medium or low risk, the system adopts progressive braking, smoothly adjusting the braking output to maintain operational continuity. During this process, the maximum deceleration limit module monitors the braking intensity in real time, ensuring that the deceleration does not exceed the safety threshold of 3 m / s². Simultaneously, the braking valve opening maintenance mechanism stabilizes the valve state through closed-loop control, keeping the deceleration output continuous and smooth. These components form a closed-loop control chain. The risk level classification serves as the decision input driving the braking mode selection, while the deceleration limit and valve mechanism act as execution constraints ensuring the stability of the braking process. This achieves precise conversion of risk assessment results into braking behavior in a dynamically changing downhole environment.

[0136] For example, when the vehicle-mounted unit detects that the relative distance to an obstacle ahead is less than a safety threshold and the relative speed exceeds a critical value, the comprehensive risk value calculation module outputs a high-risk level signal. The braking control unit immediately triggers an emergency braking mode. At this time, the brake valve opening maintenance mechanism is activated, maintaining a stable opening through closed-loop control of the electro-hydraulic proportional valve. Simultaneously, the deceleration monitoring unit ensures that the output deceleration does not exceed 3 m / s². If the risk level is determined to be medium to high risk, the system adopts a forced braking mode, achieving rapid deceleration by adjusting the brake valve opening, but the deceleration is still constrained by an upper limit. In medium-risk scenarios, the system uses progressive braking, with the brake valve opening changing smoothly with the risk value to avoid sudden deceleration changes. In this embodiment, the braking control unit can specifically use an embedded microcontroller to implement the risk level determination logic, and the braking actuator can specifically use an electro-hydraulic proportional valve in conjunction with a pressure sensor to form a closed-loop control system. The valve opening maintenance mechanism achieves stable output through PWM signal adjustment by the microcontroller.

[0137] Through the above-mentioned solution, this embodiment achieves the precise conversion of risk assessment results into braking behavior in the complex and dynamically changing environment of underground mines. It ensures timely and reliable braking response in high-risk scenarios to effectively avoid safety accidents, while also guaranteeing operational efficiency and comfort in medium- and low-risk scenarios. At the same time, by constraining deceleration limits and controlling valve opening stability, it completely eliminates impact vibrations during braking, significantly improving vehicle handling safety and personnel comfort.

[0138] The core innovation of this embodiment lies in its hierarchical and collaborative approach, which organically combines on-board local preprocessing with lightweight target detection, multi-source data fusion and local map management on edge servers, and global model training in the cloud. This effectively addresses the challenges of extremely harsh underground sensing environments, weak digital infrastructure, significant human-machine collaboration risks, and a lack of intelligent control capabilities. Specifically, this architecture addresses signal interference caused by high temperature, high humidity, dust, and sudden changes in lighting by utilizing redundant input from multiple sensors combined with local preprocessing to ensure data reliability. Real-time processing via a lightweight target detection model on the onboard unit significantly reduces the risk of collisions caused by blind spots. Secondary inference and multimodal fusion of multi-onboard data from edge servers compensate for single-point sensing blind spots and achieve high-precision environmental modeling. Simultaneously, the cloud platform continuously improves the algorithm's adaptability to dynamic mining environments through global data aggregation and iterative model optimization. Thus, this embodiment forms a closed-loop prevention and control system from data acquisition to model optimization, effectively eliminating "window periods" and breakpoints in safety supervision across time and space, achieving closed-loop prevention and control of safety risks throughout the entire underground operation process, and significantly reducing the risk of mechanical injuries and collisions in human-machine collaborative operations.

[0139] Furthermore, embodiments of this application also propose a computer-readable storage medium storing a program for mine underground safety control based on vehicle-mounted equipment. When the program for mine underground safety control based on vehicle-mounted equipment is executed by a processor, it implements the steps of the method for mine underground safety control based on vehicle-mounted equipment as described above.

[0140] Referring to Figure 3, which is a structural block diagram of the first embodiment of the mine underground safety control device based on vehicle-mounted equipment of this application.

[0141] As shown in Figure 3, the underground mine safety control device based on vehicle-mounted equipment proposed in this application includes:

[0142] The data acquisition module 10 is used to collect downhole environmental data through multi-source sensors on the vehicle-mounted terminal, and to perform local preprocessing on the data, including noise reduction and format conversion.

[0143] The data recognition module 20 is used to deploy a lightweight target detection model on the vehicle to process the pre-processed data in real time, identify targets, and generate preliminary perception results.

[0144] Data processing module 30 is used for the edge server to receive the preliminary perception results uploaded by multiple vehicle terminals in the same area for secondary reasoning and multimodal data fusion, complex model reasoning and local map management;

[0145] The data training module 40 is used to aggregate all mine data on the cloud platform for global monitoring, algorithm model training and iteration, and to send the optimized model parameters to the edge server and vehicle terminal.

[0146] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0147] This embodiment eliminates environmental noise interference through local preprocessing on the vehicle-mounted end, achieves real-time data fusion and global optimization by combining an edge-cloud collaborative architecture, and innovatively introduces a network outage self-governance mechanism and dynamic risk assessment and control algorithm. It effectively solves the problems of perception failure and network interruption safety control caused by extreme underground working conditions, solves the reliability problem of environmental perception data in extreme underground environments, realizes local autonomous safety control during network interruption, and improves the real-time performance and reliability of underground mine safety control.

[0148] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0149] In addition, for technical details not described in detail in this embodiment, please refer to the method for mine underground safety control based on vehicle-mounted equipment provided in any embodiment of this application, which will not be repeated here.

[0150] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0151] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0152] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for underground mine safety control based on vehicle-mounted equipment, characterized in that, include: The vehicle-mounted terminal collects downhole environmental data through multi-source sensors and performs local preprocessing on the data, including noise reduction and format conversion. A lightweight target detection model is deployed on the vehicle to process the pre-processed data in real time, identify targets, and generate preliminary perception results. The edge server receives preliminary perception results uploaded by multiple vehicle terminals within the same area for secondary reasoning and multimodal data fusion, complex model reasoning, and local map management; The cloud platform aggregates data from the entire mine for global monitoring, algorithm model training and iteration, and distributes the optimized model parameters to edge servers and vehicle-mounted terminals. The edge servers and vehicle-mounted terminals run a control algorithm based on dynamic risk assessment in parallel and establish a heartbeat detection mechanism. When the vehicle-mounted terminal detects that continuous network interruptions exceed a threshold, it immediately switches to a network-disconnected autonomous mode, directly executing braking control through the vehicle controller based on locally cached recent historical data and real-time sensor data. Specifically, the multimodal data fusion performed by the edge server includes: for targets with a confidence level below 0.7 uploaded by the vehicle-mounted terminal, using the more computationally powerful YOLO11-L model for secondary inference to improve the completeness of target detection; dynamically adjusting the fusion weights of visual data and laser point cloud data according to the real-time lighting stability and dust concentration underground: assigning a weight of 0.6 to 0.7 to visual data when lighting is stable and dust concentration is low; assigning a weight of 0.5 to 0.6 to laser point cloud data when vision is obstructed or dust interference is high.

2. The method according to claim 1, characterized in that, The multi-source sensors include at least an industrial camera, a lidar, and an inertial measurement unit; the local preprocessing includes direct-pass filtering to denoise the laser point cloud, and grayscale correction and JPG compression format conversion of the image.

3. The method according to claim 1, characterized in that, The lightweight target detection model is a pruned and optimized YOLO11-N model, specifically designed to identify personnel, mining vehicles, and fixed equipment in underground mining scenarios; the inference frame rate for real-time processing is no less than 10fps.

4. The method according to claim 1, characterized in that, The local map management adopts the following steps to achieve high-precision map construction and updating in large-scale scenes: Local point cloud sub-maps are generated based on a multi-sensor tightly coupled LIO-SAM framework. This framework provides high-frequency initial pose values ​​through visual odometry and performs feature point matching and residual optimization through laser odometry to provide accurate pose correction. An incremental map stitching strategy is adopted, first extracting the FPFH fast point feature histogram of the sub-map for coarse registration, and then using an improved ICP algorithm with a distance weight factor for fine registration to minimize the Euclidean distance error between point clouds. The registered sub-maps are quality-evaluated, and only when the map matching error MME ≤ 0.05m and the map point cloud variance MPV ≤ 0.03 are the sub-maps integrated into the global map. A timed or triggered map update mechanism is established to incrementally update roadway areas changed due to mining operations and to replace outdated areas without loss.

5. The method according to claim 1, characterized in that, The control algorithm based on dynamic risk assessment integrates the following multi-source heterogeneous data to calculate the comprehensive risk value: the target's relative distance, relative speed, and trajectory obtained by the perception system; the vehicle's absolute coordinates and speed calculated by comparing the point cloud scanned in real time by the vehicle-mounted LiDAR with a preset global map; and centimeter-level precise positioning data provided by the UWB smart bracelet worn by the personnel to compensate for perception blind spots.

6. The method according to claim 1, characterized in that, The control algorithm is a multi-objective optimized adaptive braking strategy. Its specific execution logic includes: dynamically classifying the calculated comprehensive risk value into four levels: low risk, medium risk, medium-high risk, and high risk; triggering different braking modes based on the real-time risk level: initiating emergency braking mode at high risk with safety as the primary objective; initiating forced braking mode at medium-high risk to quickly shorten the safe distance; and employing gradual braking at medium risk and below to balance operational efficiency and comfort. By setting a maximum deceleration limit ≤3m / s² and using a brake valve opening maintenance mechanism, stable deceleration output is maintained to avoid braking shock.

7. A mine underground safety control device based on vehicle-mounted equipment, characterized in that, The method described in claim 1 comprises: a data acquisition module for collecting downhole environmental data from multiple sources via vehicle-mounted sensors and performing local preprocessing on the data, including denoising and format conversion; a data recognition module for deploying a lightweight target detection model on the vehicle-mounted terminal to process the preprocessed data in real time, identify targets, and generate preliminary perception results; a data processing module for receiving preliminary perception results uploaded by multiple vehicle-mounted terminals within the same mining area via an edge server for secondary inference, multimodal data fusion, complex model inference, and local map management; and a data training module for a cloud platform to aggregate mine-wide data for global monitoring, algorithm model training and iteration, and to distribute the optimized model parameters to the edge server and vehicle-mounted terminals.

8. A computer device, characterized in that, The device includes a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, Includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

Images