Cave rescue three-dimensional semantic perception and dynamic risk self-adaptive avoidance system and method

Abstract

This invention discloses a three-dimensional semantic perception and dynamic risk adaptive avoidance system and method for cave rescue, including a multi-source data access module, a three-dimensional point cloud semantic mapping module, a fracture dynamic risk assessment module, a rescue danger radius dynamic adjustment module, and a robotic arm autonomous positioning and navigation module. Through synchronous acquisition of multi-source data, geometric-texture joint feature extraction, rock mass semantic segmentation and static risk assessment, fracture dynamic evolution prediction, dynamic and static risk fusion cost map construction, danger radius adaptive optimization, and low-risk path planning and dynamic replanning of the robotic arm, it achieves fully closed-loop intelligent control from front-end perception to back-end execution in cave rescue scenarios. This invention solves the problems of single spatial perception dimension, insufficient fusion of dynamic and static information, delayed risk response, and fragmented data chains in existing technologies, significantly improving the safety, timeliness, and success rate of cave rescue, and is applicable to emergency rescue scenarios such as cave collapses and underground space disasters.

Description

Technical Field

[0001] This invention belongs to the fields of underground space emergency rescue, three-dimensional environmental perception and intelligent robot technology, specifically involving a three-dimensional semantic perception and dynamic risk adaptive avoidance system and method for cave rescue. Background Technology

[0002] In emergency rescue scenarios such as cave collapses, mine tunnel accidents, and underground space disasters, the rescue environment is characterized by high concealment, unstable rock structures, lack of GPS signals, limited communication, and high dust and low light levels. This places extremely high demands on the autonomous environmental perception capabilities, intelligent risk decision-making levels, and autonomous operation capabilities of rescue equipment. Currently, modern cave rescue operations widely utilize intelligent equipment such as miniature reconnaissance drones, multi-functional demolition and rescue robotic arms, and tracked search robots, which can collect multi-dimensional spatial and environmental data, including LiDAR 3D point clouds, binocular visual images, and environmental parameters. However, existing emergency rescue technologies still suffer from the following prominent technical deficiencies in practical applications: 1. Limited spatial perception dimension leads to low accuracy in rock mass stability identification. Existing solutions are mostly based on two-dimensional images for texture and fracture analysis, which makes it difficult to capture the three-dimensional geometric abrupt changes in rock mass fractures, bedding planes, and unstable rock masses. They also cannot accurately quantify key geometric parameters that determine rock mass stability, such as fracture width, orientation, dip angle, and connectivity. A few identification solutions based on three-dimensional point clouds only use single geometric features without integrating texture information. In harsh environments such as dusty caves with low illumination, the misjudgment rate of fracture and unstable rock masses remains high, which can easily lead to safety risks.

[0003] 2. Insufficient integration of static and dynamic risk information, and a disconnect between spatiotemporal dimensions. Existing technologies mostly use static 3D maps for environmental modeling, which cannot reflect the evolution trend of fractures over time and the dynamic process of rock mass instability. Meanwhile, dynamic monitoring data such as fracture displacement and microseismic activity are difficult to accurately bind to 3D spatial locations. There is a lack of an effective fusion model between static spatial information and dynamic temporal information. Command personnel cannot comprehensively judge the probability of rock mass instability from the spatiotemporal evolution dimension, and are prone to overlooking the risk correlation across data sources, resulting in delayed risk warnings.

[0004] 3. Rescue of hazardous areas relies on manual, static settings, resulting in a lagging risk response model. In current cave rescue operations, the hazardous operating radius and safe zone are mostly statically set in advance by commanders based on experience, and cannot be adaptively adjusted in real time according to changes in the dynamic risks of the rock mass. When cracks expand rapidly and the risk of rock mass instability increases sharply, there is a significant time delay in manually analyzing multi-source data streams, making it difficult to issue timely evacuation instructions. In some cases, failure to avoid secondary collapse disasters in time can lead to damage to rescue equipment and casualties among rescue personnel.

[0005] 4. The lack of a closed-loop data chain results in low efficiency and poor fault tolerance in the current operation mode. The current rescue system generally adopts an open-loop operation mode of "drone data collection - manual identification and marking of dangerous areas - manual issuance of instructions to the robotic arm - manual planning of the operation path". The 3D point cloud data transmitted back by the drone needs to be manually marked for dangerous areas and risk assessment, and then manually issued to the robotic arm to plan the operation path. The whole process is time-consuming, labor-intensive, and prone to risks due to human error, making it impossible to achieve rapid response and autonomous operation in emergency rescue scenarios.

[0006] To address the shortcomings of existing technologies, there is currently no effective solution. Therefore, there is an urgent need to develop a complete data loop that can connect the front-end perception of the drone to the back-end execution of the robotic arm, achieve deep integration of static geometric features and dynamic time-series data, construct a spatiotemporal integrated risk cost map, and dynamically adjust the rescue danger radius and operation path based on risk, so as to improve the timeliness, safety and success rate of emergency rescue in underground spaces. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a three-dimensional semantic perception and dynamic risk adaptive avoidance system and method for cave rescue. This invention establishes a closed-loop data link from environmental perception, risk assessment, intelligent decision-making to equipment execution, realizing accurate three-dimensional semantic perception in complex cave environments, deep integration of static and dynamic risks, adaptive optimization of the rescue danger radius, and dynamic risk avoidance of the robotic arm's operating path. It effectively solves the technical problems of single perception dimension, insufficient integration of static and dynamic information, delayed risk response, and fragmented data link in the prior art, and significantly improves the safety and operational efficiency of cave rescue.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is: a three-dimensional semantic perception and dynamic risk adaptive avoidance system for cave rescue, including a multi-source data access module, a three-dimensional point cloud semantic mapping module, a crack dynamic risk assessment module, a rescue danger radius dynamic adjustment module, and a robotic arm autonomous positioning and navigation module.

[0009] The multi-source data access module is used to access the emergency rescue network at the cave rescue site and obtain multi-source data that is synchronized with the time.

[0010] The three-dimensional point cloud semantic mapping module is connected to the multi-source data access module. It is used to extract geometric-texture joint features from the original point cloud to target the sensitivity of rock mass fractures to geometric abrupt changes. It also constructs a semantic octree map with fracture stability probability through a semantic segmentation network as a static risk map.

[0011] The crack dynamic risk assessment module is used to perform non-rigid inter-frame registration of key crack regions obtained by semantic recognition, calculate the time series of crack state quantities, predict the future evolution trend of cracks through a lightweight time series model, and output dynamic risk cost.

[0012] The rescue danger radius dynamic adjustment module is connected to the three-dimensional point cloud semantic mapping module and the crack dynamic risk assessment module, respectively. It is used to integrate the static risk map and the dynamic risk cost to construct a total risk cost map, and dynamically optimize the safe operating radius of the rescue equipment based on the multi-objective evolutionary algorithm of adaptive preference adjustment in the rescue stage.

[0013] The robotic arm autonomous positioning and navigation module is used to achieve precise six-degree-of-freedom positioning of the rescue robotic arm in a rock cave environment without GPS through visual inertial SLAM. Based on the total risk cost map and dynamically optimized safe operation hazard radius, it plans the lowest risk operation path and performs local dynamic replanning when the dynamic risk trigger threshold is reached.

[0014] Furthermore, the multi-source data includes 3D point cloud data collected by UAV-borne LiDAR, RGB image data collected by binocular vision cameras, and cave disaster environment parameter data; the multi-source data access module performs hard synchronization of the multi-source data through the PTP precise time protocol, completes the timestamp alignment of point cloud and image data, and performs motion distortion correction and outlier filtering preprocessing on the original point cloud data; the semantic segmentation network used by the 3D point cloud semantic mapping module is the Res16UNet sparse convolutional network based on the MinkowskiEngine deep learning library; the lightweight temporal model used by the fracture dynamic risk assessment module is the temporal convolutional network TCN or the long short-term memory network LSTM; the multi-objective evolutionary algorithm used by the rescue danger radius dynamic adjustment module is the MOEA / D algorithm; and the path planning algorithm used by the robotic arm autonomous positioning and navigation module is the RRT* algorithm.

[0015] The avoidance method of the aforementioned cave rescue 3D semantic perception and dynamic risk adaptive avoidance system includes the following steps: Step 1: Multi-source data access and preprocessing: Through the emergency rescue network at the cave rescue site, acquire lidar point cloud data, binocular visual image data, and disaster environment parameter data collected by drones. Perform time synchronization and preprocessing on the multi-source data to obtain a regular point cloud data stream with color information.

[0016] Step 2: Geometric-texture joint feature extraction for crack geometrical abrupt change sensitivity: On the preprocessed 3D point cloud data, extract local geometrical abrupt change features to characterize the crack edge and the boundary of the intact rock mass, as well as texture features to distinguish the lithological interface and the crack filling material, to form a multi-channel input feature vector.

[0017] Step 3, 3D point cloud semantic model training and inference: Construct a semantic segmentation network based on sparse convolution, use real cave labeled data and synthetic simulation data to complete model training, input the feature vector obtained in step 2 into the trained model, and infer the point-by-point semantic labels of the rock mass, so that the model has robust crack recognition ability in the low-light and dusty cave environment.

[0018] Step 4, Rock Mass Static Risk Assessment: Construct a semantic octree map based on semantic tags, obtain the storage category label and stability probability of each voxel in the map, and obtain the static risk cost; the stability probability is calculated using an adaptive weighted fusion method based on the rock mass structure surface classification rules.

[0019] Step 5: Dynamic Risk Assessment of Fractures: Using the time-series data continuously transmitted by UAVs, non-rigid registration is performed between frames in key fracture areas to calculate the changes in fracture width, the changes in the main direction of displacement fields on both sides of the fracture, and the accumulation rate of local microseismic energy. A time series of multi-physics state variables is constructed and input into a lightweight time-series model to predict future fracture change trends. The dynamic risk cost of the fusion prediction uncertainty penalty is obtained, and when the predicted value exceeds the preset rock mass instability prior threshold, the risk cost is amplified by a nonlinear coefficient.

[0020] Step Six: Construction of Total Risk Cost Map: Integrate the static risk cost obtained from static risk assessment and the dynamic risk cost obtained from dynamic risk assessment to construct a two-layer total risk cost map that changes with time and space.

[0021] Step 7: Dynamic optimization of rescue danger radius: Based on the total risk cost map, calculate the dynamic basic danger radius that is linked to the risk distribution density, set a dynamic adjustment coefficient that is positively correlated with the dynamic risk change rate of the fracture for adaptive adjustment, and use a multi-objective evolutionary algorithm for adaptive preference adjustment in the rescue stage to select the Pareto optimal solution as the final rescue danger radius.

[0022] Step 8: Autonomous localization and path planning of the robotic arm: Visual inertial SLAM is used to achieve accurate localization of the rescue robotic arm in the absence of GPS. Based on the total risk cost map and dynamically optimized safe operation hazard radius, the lowest risk operation path is planned. When the dynamic risk triggers the threshold, local dynamic replanning is performed to complete the adaptive risk avoidance of the rescue operation.

[0023] Furthermore, in step one, the PTP precise time protocol is used to perform hard synchronization of multi-source data to achieve timestamp alignment between point cloud and image data; the preprocessing includes motion distortion correction and statistical filtering to remove outliers from the original point cloud data, resulting in a preprocessed color point cloud data stream with color information.

[0024] Furthermore, in step two, the extraction of joint geometry-texture features for the sensitivity of crack geometry to abrupt changes specifically involves: For any point p, take its neighborhood point set N(p) and calculate the three-dimensional centroid of the neighborhood point set. Construct the covariance matrix: Performing eigenvalue decomposition on the covariance matrix C yields the following results: eigenvalues , , , and the corresponding feature vectors.

[0025] Calculate roughness .

[0026] Calculate normal vector consistency ,in Let p be the normal vector of point p, and let the minimum eigenvalue be the eigenvector. The corresponding eigenvectors are determined. Let be the normal vector of the neighborhood point q.

[0027] Calculate curvature change .

[0028] In the LAB color space, calculate the color covariance matrix of the neighborhood of point p. ,in Let q be the LAB color vector. Given the color mean vector within the neighborhood, take the trace of the color covariance matrix. As a texture feature.

[0029] The above features are concatenated with the original 3D coordinates and RGB color values ​​of point p to construct a 10-dimensional multi-channel input feature vector: Where x, y, z are the three-dimensional spatial coordinates of point p, and R, G, B are the color channel values ​​of point p. Let be the roughness of the local surface where point p is located. For the normal vector of point p to be consistent, Let p be the curvature change of the local surface where point p is located. Let be the trace of the color covariance matrix of the neighborhood of point p in the LAB region.

[0030] Furthermore, in step three, the semantic segmentation network is constructed using the MinkowskiEngine deep learning library combined with the Res16UNet sparse convolutional network. During model training, self-supervised contrastive learning pre-training is first performed on large-scale unlabeled cave point cloud data, and then fine-tuning is performed using a semi-supervised method combined with real labeled samples. The real labeled data are the labeled data of fissures, intact rock masses, and potentially dangerous rock masses collected from real cave environments, and the synthetic simulation data are cave scene data with fissure annotations generated by a 3D simulation engine.

[0031] Furthermore, in step four, for each voxel in the semantic octree map... The weighting coefficients are dynamically determined based on the rock mass structure plane classification rules, and the stability probability is calculated. : in, voxels The normalized value corresponding to the angle between the fracture orientation and the direction of gravity. voxels The normalized value corresponding to the crack width, voxels Normalized value corresponding to the connectivity of the fracture; Voxels determined based on rock mass structural plane classification rules The structural class of the area is comprehensively evaluated based on the crack density, roughness, filling material characteristics, and spatial relationship with the main structural surface; An adaptive weighting function related to the structural plane level, satisfying Furthermore, a nonlinear mapping was performed on the dominance of each parameter on rock mass stability at different levels, resulting in a weight bias towards dip angle and connectivity for low-level structural planes (with good integrity), and a weight bias towards fracture width and connectivity for high-level structural planes (with well-developed fractures); As voxels The static risk cost is stored in the corresponding node of the semantic octree map.

[0032] Furthermore, step five specifically includes: Inter-frame non-rigid registration is performed on key crack regions to calculate the crack width variation. Changes in the principal directions of the displacement field on both sides of the fracture and local microseismic energy accumulation rate Construct a time series of fracture state variables over the past L time steps. , where X(t) is the fracture state quantity at time t, including the calculation of the fracture width change, the change of the principal direction of the displacement field on both sides of the fracture, the change acceleration, and the local microseismic energy accumulation rate.

[0033] Inputting state time series data into a lightweight time series model to predict the future. Predicted value of crack width change at time step Compared with predicted energy release values .

[0034] The dynamic risk cost is calculated as follows: ,in This represents the quantified value of the uncertainty in time series prediction; where α is the weight of the crack width variation, with a value ranging from [value missing]. β represents the energy release weight, with a range of values ​​of [value missing]. , The weight is the quantification value of uncertainty, and its value range is... ,and .

[0035] when When the preset prior threshold for rock mass instability is exceeded, the dynamic risk cost coefficient... According to nonlinear coefficient Magnification, the nonlinear coefficient and The magnitude of the exceedance of the threshold is positively correlated.

[0036] Furthermore, in step six, the formula for calculating the total risk cost is as follows: in, Static risk weights For dynamic risk weights, .

[0037] Furthermore, step seven specifically includes: S71. Obtain basic parameters: The rated operating radius of the rescue robotic arm, The radius of influence of the unstable rock mass. To ensure equipment safety, d represents the building risk coefficient, and d represents the real-time distance between the rescue robotic arm and the unstable rock mass. The minimum threshold for the danger radius. This represents the maximum threshold for the danger radius.

[0038] S72. Calculate the basic hazard radius Among them, the equipment safety factor With building risk coefficient The risk level is dynamically adjusted based on the distribution density of high-risk areas in the overall risk cost map.

[0039] S73. Dynamically adjust to obtain the dynamic danger radius. Where λ is the dynamic adjustment coefficient, and is related to the rate of change of dynamic risk of the fracture. Positive correlation.

[0040] S74. Obtain the initial danger radius by performing threshold truncation. .

[0041] S75. Using the initial danger radius as the decision variable, construct a dual optimization objective: maximize rescue efficiency. Where Coverage(R) is the search and rescue coverage area constrained by the danger radius R; minimizing rescue risk. , where Risk(R) is the cumulative total risk cost under the constraint of the danger radius R.

[0042] S76. The dual optimization problem is decomposed into multiple sub-problems by using an adaptive weight vector for the rescue phase. Then, a multi-objective evolutionary algorithm is used to obtain the Pareto optimal solution set through crossover and mutation iteration. Based on the preset efficiency-risk preference weight, the final rescue danger radius is selected from the Pareto optimal solution set.

[0043] Furthermore, in step eight, the cost function used for path planning is: in, The cumulative total risk cost along the path is represented by Length(path), where Length(path) is the total path length, and γ is the path length weighting coefficient, ranging from [0.1, 0.3]. When the crack width changes... When the preset safety threshold is exceeded, local dynamic replanning is triggered to generate an avoidance trajectory and control the robotic arm to retract to a safe area; the visual inertial SLAM adopts a tightly coupled optimization framework, integrates binocular visual images with IMU data from gyroscopes and accelerometers, and outputs the six-degree-of-freedom pose of the rescue robotic arm with a positioning accuracy better than 5cm.

[0044] Compared with the prior art, the present invention has the following advantages: 1. This invention designs a 10-dimensional geometry-texture joint feature extraction scheme, which integrates the geometric and texture features of the three-dimensional space of the rock mass. It is optimized for the harsh environment of low light and dust in caves. Compared with the existing single feature recognition scheme, the accuracy of identifying rock fissures and potential dangerous rock masses is improved by more than 40%, and the false judgment rate of rock mass stability is reduced by 35%, which greatly improves the accuracy of three-dimensional semantic perception in complex cave environments.

[0045] 2. This invention constructs a total risk cost map that deeply integrates static risk and dynamic time-series prediction, realizing the precise binding of the three-dimensional spatial location of the rock mass with the temporal evolution trend of the fracture. It can predict the instability trend of accelerated fracture expansion 15 to 30 minutes in advance, leaving sufficient response time for emergency avoidance and solving the problem of the disconnect between static and dynamic information in the existing technology.

[0046] 3. This invention designs a dynamic optimization scheme for the rescue danger radius based on a dual-objective and multi-objective evolutionary algorithm. With the optimization objectives of maximizing rescue efficiency and minimizing rescue risk, the danger radius of the rescue operation can be adaptively adjusted according to the real-time risk status on site, achieving the optimal balance between rescue efficiency and operational safety, and solving the problem of delayed response when manually setting danger areas in existing technologies.

[0047] 4. This invention establishes a closed-loop data link from front-end environmental perception of the UAV, multi-source data fusion, intelligent risk assessment, and operation radius optimization to autonomous execution at the back end of the robotic arm. It realizes the automation and intelligence of the entire process of cave rescue, and the path replanning response time reaches the millisecond level. This significantly reduces the risk of equipment damage caused by secondary disasters and significantly improves the response speed and success rate of emergency rescue in caves.

[0048] 5. The system and method of this invention have strong adaptability and are compatible with existing mainstream rescue drones, rescue robotic arms and other equipment. They do not require large-scale modification of existing equipment and can be quickly deployed in various underground space emergency rescue scenarios such as cave collapses, mine roadway accidents and underground pipe gallery disasters, and have extremely strong engineering application value. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Detailed Implementation

[0050] The present invention will be further described below.

[0051] like Figure 1As shown in the figure, the 3D semantic perception and dynamic risk adaptive avoidance system for cave rescue provided in this embodiment is deployed at the emergency rescue site of a cave collapse. The system hardware includes: a quadcopter reconnaissance drone equipped with a 16-line LiDAR, a binocular global shutter camera, and an IMU; a tracked rescue robot equipped with a binocular vision camera, a 6-axis IMU, and a 6-DOF demolition rescue robotic arm; and an on-site emergency command host computer equipped with a GPU computing card. The drone and rescue robot are connected to the on-site emergency rescue network via fiber optic Ethernet to achieve low-latency data transmission with the host computer. The system includes a multi-source data access module, a 3D point cloud semantic mapping module, a fracture dynamic risk assessment module, a rescue danger radius dynamic adjustment module, and a robotic arm autonomous positioning and navigation module, which are connected in sequence. Each module is deployed on the host computer, and a lightweight SLAM node and motion control node are deployed at the rescue robotic arm end.

[0052] The overall process of the three-dimensional semantic perception and dynamic risk adaptive avoidance method for cave rescue in this embodiment is as follows: Figure 1 As shown, the specific steps include: Step 1: Multi-source data access and preprocessing: The multi-source data access module acquires 3D point cloud data (acquisition frequency 10Hz) from a UAV-borne 16-line LiDAR, RGB image data (acquisition frequency 30Hz) from a binocular vision camera, and disaster environment parameter data from temperature, humidity, and microseismic monitoring instruments through the on-site emergency rescue network. It uses the PTP precise time protocol to perform hard synchronization of all multi-source data, ensuring that the timestamp alignment error between the point cloud and image data is less than 1ms. After receiving the raw point cloud data, the host computer first performs motion distortion correction to eliminate motion distortion of the point cloud during UAV flight. Then, it uses statistical filtering to remove outliers, with the filtering parameters set to 50 neighboring points and a standard deviation factor of 1.5. Finally, a clean, color point cloud data stream with color information is obtained, which serves as the input for subsequent modules.

[0053] Step 2: Geometric-texture joint feature extraction for crack geometric abrupt changes: The 3D point cloud semantic mapping module performs geometric-texture joint feature extraction on the preprocessed point cloud data to address crack geometric abrupt changes. For any point p, a K-nearest neighbor set N(p) is selected, with the number of neighboring points K set to 30. The specific extraction process is as follows: Calculate the 3D centroid of the neighborhood point set .

[0054] Constructing the neighborhood covariance matrix Eigenvalue decomposition is performed on the covariance matrix C to obtain the result that satisfies eigenvalues , , , and the corresponding feature vectors.

[0055] Calculate roughness Roughness characterizes the degree of unevenness of a local surface; the greater the roughness, the higher the likelihood of a crack edge.

[0056] Calculate normal vector consistency ,in Let p be the normal vector of point p, and let the minimum eigenvalue be the eigenvector. The corresponding eigenvectors are determined. Let q be the normal vector of the neighborhood point q. The larger this value is, the more drastic the change in the local normal vector is, and the higher the probability of a fracture surface.

[0057] Calculate curvature change The value approaches 0 in flat areas and increases significantly at crack edges and corners.

[0058] Convert the RGB image to the LAB color space and calculate the color covariance matrix of the neighborhood of point p. ,in Let q be the LAB color vector. Given the color mean vector within the neighborhood, take the trace of the color covariance matrix. As a texture feature, the larger the trace, the more drastic the local texture change.

[0059] The above features are concatenated with the original 3D coordinates and RGB color values ​​of point p to construct a 10-dimensional multi-channel input feature vector. This serves as the input for the subsequent semantic segmentation network.

[0060] Step 3: Training and inference of the 3D point cloud semantic model: Construct a Res16UNet sparse convolutional semantic segmentation network based on the MinkowskiEngine deep learning library. The network input is the 10-dimensional feature vector obtained in Step 2, and the output is a point-by-point semantic label. The semantic categories include three types: intact rock mass, fracture surface, and potentially dangerous rock mass.

[0061] The model training process is as follows: 1. Dataset Construction: Point cloud data of real cave environments from 5 typical karst landforms and 3 metal mine tunnels in China were collected. Pixel-level annotations were performed on fracture surfaces and potential unstable rock masses to obtain 2,000 sets of real-labeled samples. At the same time, a high-fidelity cave simulation environment was built using Unreal Engine, and 10,000 sets of precisely labeled synthetic fracture simulation data were generated using DomainRandomization technology for model pre-training and data augmentation.

[0062] 2. Model pre-training: On large-scale unlabeled cave point cloud data, a contrastive learning method is used for self-supervised pre-training to learn the general geometric representation of cave rock masses.

[0063] 3. Model fine-tuning: A semi-supervised learning method is adopted, combining 2000 sets of real labeled samples with 10000 sets of samples to fine-tune the pre-trained model. The optimizer is AdamW, the initial learning rate is set to 1e-3, the batch size is set to 8, the training epochs are 100, and an early stopping mechanism is adopted to prevent overfitting.

[0064] 4. Model Inference: Input the 10-dimensional feature vector of the point cloud obtained in step 2 into the trained semantic segmentation network to infer the semantic label of each point. In the low-light and dusty environment of the cave, the average recognition accuracy mIoU of the model reached 89.2%.

[0065] Step 4: Static Risk Assessment of Rock Mass: The static risk cost is obtained by quantifying the storage category label and stability probability of each voxel in the map through geometric analysis. The stability probability is calculated using an adaptive weighted fusion method based on the rock mass structural surface classification rules. Specifically, the weighting coefficients are dynamically determined based on the rock mass structural surface classification rules to calculate the stability probability. : in, This is the normalized value of the angle between the fracture orientation and the gravity direction corresponding to voxel v. The steeper the dip angle, the lower the stability. This is the normalized value of the crack width corresponding to voxel v; This is the normalized value for the crack connectivity corresponding to voxel v.

[0066] Grade(v) is the structural plane grade of the region where voxel v is located, determined based on the rock mass structural plane classification rules. According to the "Engineering Rock Mass Classification Standard" (GB / T50218-2014) and the five-level classification system of rock mass structural planes, the structural plane grades are divided into grades I to V (grade I is intact rock mass, and grade V is extremely fractured and developed zone) by comprehensively considering fracture density, roughness, filling material characteristics and spatial relationship with the main structural plane. w1(·), w2(·), and w3(·) are adaptive weighting functions related to the structural plane level, satisfying w1+w2+w3=1, and employing a nonlinear mapping mechanism: for low-level structural planes (Level I~II, relatively good integrity), the weights are biased towards dip angle and connectivity (w1=0.35~0.40, w3=0.35~0.40, w2=0.20~0.30); for high-level structural planes (Level IV~V, well-developed fractures), the weights are biased towards fracture width and connectivity (w2=0.40~0.50, w3=0.35~0.40, w1=0.15~0.25); Level III structural planes use a balanced weight (w1=0.30, w2=0.40, w3=0.30). The adaptive weights are dynamically adjusted based on the results of field geological surveys and structural plane statistical analysis, reflecting the control effect of geomechanical background on rock mass stability. The static risk cost, as a voxel v, is stored in the corresponding node of the semantic octree map, forming a static risk map.

[0067] Step 5: Dynamic Risk Assessment of Fractures: During the rescue operation, the UAV periodically inspects the key fracture areas along a preset route, continuously transmitting time-series point cloud data. The dynamic risk assessment module uses FlowNet3D for non-rigid registration of the time-series point cloud frames of the key fracture areas to accurately calculate the relative displacement field of the rock masses on both sides of the fracture, obtaining the fracture width change Δd, displacement rate, and acceleration, and constructing a time series of fracture state variables X(t-9),...,X(t) for the past L=10 time points. Simultaneously, data from a microseismic monitoring instrument is integrated, and the local microseismic energy accumulation rate E_acc is extracted as a characterization parameter of the fracture activity within the rock mass.

[0068] Input the state time series into a pre-trained lightweight temporal convolutional network (TCN) to predict the change in crack width within the next Δt = 30 minutes. Compared with predicted energy release values In this embodiment, the TCN model has a convolution kernel size of 3 and a number of layers of 4. Causal convolution is used to avoid information leakage, and the model can predict the trend of accelerated crack propagation 15-30 minutes in advance.

[0069] The dynamic risk cost calculation introduces a penalty term for prediction uncertainty, reflecting a quantitative consideration of the reliability of time series predictions: Wherein, σ(t+Δt) is the quantified value of uncertainty in the time series prediction (obtained by Monte Carlo dropout or ensemble learning methods to obtain the variance of the prediction distribution), α is the weight of crack width variation (ranging from 0.2 to 0.4), β is the weight of energy release (ranging from 0.3 to 0.5), and γ is the weight of uncertainty quantification (ranging from 0.2 to 0.4), and α+β+γ=1. In this embodiment, α=0.35, β=0.40, and γ=0.25, reflecting the importance attached to the energy release index and prediction reliability.

[0070] when When the preset a priori threshold for rock mass instability is exceeded (the rate of change of fracture width exceeds 5 mm / h or the cumulative change exceeds 10 mm), the dynamic risk cost is amplified by a nonlinear coefficient η: Where k is the amplification factor (k=2.0 in this embodiment), the nonlinear coefficient is positively correlated with the magnitude of the predicted value exceeding the threshold, so that the risk of exceeding the threshold is given an exponential warning, ensuring a rapid response under high-risk conditions.

[0071] Step Six: Constructing the Total Risk Cost Map: Integrating static and dynamic risk costs, a two-layer total risk cost map that varies with time and space is constructed. The total risk cost of each voxel at time t is: In this embodiment, static risk weights Dynamic risk weights , The total risk cost map is updated in real time with on-site data, with an update frequency of 1Hz.

[0072] Step 7: Dynamic Optimization of Rescue Danger Radius: The dynamic adjustment module for the rescue danger radius, based on the total risk cost map, uses the MOEA / D multi-objective evolutionary algorithm to dynamically optimize the rescue danger radius. The specific process is as follows: S71. Obtain basic parameters: rated working radius of the rescue robotic arm. Radius of the affected area of ​​the unstable rock mass Equipment safety factor Building risk coefficient Real-time distance between the rescue robotic arm and the unstable rock mass Minimum threshold for danger radius Maximum threshold of danger radius .

[0073] S72. Calculate the basic hazard radius Among them, the equipment safety factor With rock mass risk coefficient Dynamically adjusted based on the distribution density of high-risk areas in the overall risk cost map: when a high-risk area (C... total When the proportion of ≥0.7) exceeds 30%, Upgraded to 1.8 Upgraded to 1.5 to achieve risk linkage response.

[0074] S73. Dynamically adjust to obtain the dynamic danger radius. Where λ is the dynamic adjustment coefficient, and is related to the rate of change of dynamic risk of the fracture. Positive correlation. In this embodiment, This ensures that the greater the rate of change in risk, the more sensitive the radius adjustment becomes.

[0075] S74. Obtain the initial danger radius by performing threshold truncation. Using the danger radius R as the decision variable, a dual optimization objective is constructed: maximizing rescue efficiency. (Search and rescue coverage area under the constraint of danger radius R), minimizing rescue risk. (Cumulative total risk cost under the constraint of danger radius R).

[0076] S75. The MOEA / D algorithm is used to decompose the dual optimization problem into N=100 subproblems, with a population size of 100 and 50 iterations. An adaptive preference vector is introduced for each rescue phase: in the initial life detection phase, the efficiency-risk preference vector is set to (0.7, 0.3), focusing on expanding the search and rescue area; after discovering signs of rock instability, the preference vector dynamically switches to (0.3, 0.7), focusing on operational safety. Pareto optimal solutions are obtained through crossover and mutation iterations. Based on the efficiency-risk preference weights of the current phase, the final rescue danger radius is selected from the Pareto optimal solution set.

[0077] Step 8: Autonomous Localization and Path Planning of the Robotic Arm: The rescue robotic arm, equipped with a binocular vision camera and a 6-axis IMU, employs a tightly coupled visual-inertial SLAM system. By fusing binocular images and IMU data, it outputs the robotic arm's six-degree-of-freedom pose in real-time in a rocky cave environment without GPS, achieving a positioning accuracy better than 5cm. The RRT* path planning algorithm is used to search for the lowest-risk work path from the robotic arm's current position to the target work point on the total risk cost map. The path cost function is: In this embodiment, the path length weighting coefficient .

[0078] When the crack width change predicted by the crack dynamic risk assessment module When the speed exceeds the preset safety threshold by 10mm, local dynamic replanning is immediately triggered. The local planner of the robotic arm, based on model predictive control (MPC), generates a local avoidance trajectory that avoids high-risk areas within 200ms, controlling the robotic arm to quickly retract to a safe area, effectively avoiding equipment damage caused by sudden rock collapse.

[0079] The system and method in this embodiment fully establish a closed-loop data link from the front-end perception of the UAV to the back-end execution of the robotic arm. This enables accurate three-dimensional semantic perception, dynamic and static risk fusion assessment, adaptive optimization of the danger radius, and dynamic risk avoidance in complex cave environments, significantly improving the intelligence level, safety, and success rate of emergency rescue in caves.

[0080] 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 three-dimensional semantic perception and dynamic risk adaptive avoidance system for cave rescue, characterized in that, It includes a multi-source data access module, a 3D point cloud semantic mapping module, a crack dynamic risk assessment module, a rescue danger radius dynamic adjustment module, and a robotic arm autonomous positioning and navigation module; The multi-source data access module is used to access the emergency rescue network at the cave rescue site and obtain multi-source data synchronized in time. The three-dimensional point cloud semantic mapping module is connected to the multi-source data access module. It is used to extract geometric-texture joint features from the original point cloud to target the sensitivity of rock mass fractures to geometric abrupt changes. It also constructs a semantic octree map with fracture stability probability through a semantic segmentation network as a static risk map. The crack dynamic risk assessment module is used to perform non-rigid registration between time frames of key crack regions obtained by semantic recognition, calculate the time series of crack state quantities, predict the future evolution trend of cracks through a lightweight time series model, and output dynamic risk cost. The rescue danger radius dynamic adjustment module is connected to the three-dimensional point cloud semantic mapping module and the crack dynamic risk assessment module, respectively. It is used to integrate the static risk map and the dynamic risk cost to construct a total risk cost map, and dynamically optimize the safe operating radius of the rescue equipment based on the multi-objective evolutionary algorithm of adaptive preference adjustment in the rescue stage. The robotic arm autonomous positioning and navigation module is used to achieve precise six-degree-of-freedom positioning of the rescue robotic arm in a rock cave environment without GPS through visual inertial SLAM. Based on the total risk cost map and dynamically optimized safe operation hazard radius, it plans the lowest risk operation path and performs local dynamic replanning when the dynamic risk trigger threshold is reached.

2. The cave rescue three-dimensional semantic perception and dynamic risk adaptive avoidance system according to claim 1, characterized in that, The multi-source data includes 3D point cloud data collected by UAV-borne LiDAR, RGB image data collected by binocular vision cameras, and environmental parameter data of the cave disaster. The multi-source data access module performs hard synchronization of the multi-source data through the PTP precise time protocol, completes the timestamp alignment of the point cloud and image data, and performs motion distortion correction and outlier filtering preprocessing on the original point cloud data. The semantic mapping module of the 3D point cloud uses a sparse convolutional network based on a deep learning library for semantic segmentation. The dynamic risk assessment module of the fracture uses a lightweight temporal model, which is a temporal convolutional network or a long short-term memory network. The dynamic adjustment module of the rescue danger radius uses the MOEA / D algorithm for multi-objective evolution. The path planning algorithm used by the robotic arm autonomous positioning and navigation module is the RRT* algorithm.

3. An avoidance method for the cave rescue three-dimensional semantic perception and dynamic risk adaptive avoidance system according to claim 1 or 2, characterized in that, Includes the following steps: Step 1: Multi-source data access and preprocessing: Through the emergency rescue network at the cave rescue site, acquire lidar point cloud data, binocular visual image data, and disaster environment parameter data collected by drones, perform time synchronization and preprocessing on the multi-source data to obtain a regular point cloud data stream with color information; Step 2: Geometric-texture joint feature extraction for crack geometrical abrupt change sensitivity: On the preprocessed 3D point cloud data, extract local geometrical abrupt change features to characterize the crack edge and the boundary of the intact rock mass, as well as texture features to distinguish the lithological interface and the crack filling material, to form a multi-channel input feature vector; Step 3, 3D point cloud semantic model training and inference: Construct a semantic segmentation network based on sparse convolution, use real cave annotation data and synthetic simulation data to complete model training, input the feature vector obtained in step 2 into the trained model, and infer the point-by-point semantic labels of the rock mass; Step 4: Static risk assessment of rock mass: Construct a semantic octree map based on semantic tags, obtain the storage category label and stability probability of each voxel in the map, and obtain the static risk cost; The stability probability is calculated using an adaptive weighted fusion method based on the rock mass structure surface classification rules; Step 5: Dynamic Risk Assessment of Fractures: Using the time-series data continuously transmitted back by UAVs, non-rigid registration is performed between frames in key fracture areas to calculate the changes in fracture width, the changes in the main direction of displacement fields on both sides of the fracture, and the accumulation rate of local microseismic energy. A time series of multi-physics state variables is constructed and input into a lightweight time-series model to predict future fracture change trends. The dynamic risk cost of the fusion prediction uncertainty penalty is obtained, and when the predicted value exceeds the preset rock mass instability prior threshold, the risk cost is amplified by a nonlinear coefficient. Step 6: Construction of Total Risk Cost Map: Integrate the static risk cost obtained from static risk assessment and the dynamic risk cost obtained from dynamic risk assessment to construct a two-layer total risk cost map that varies with time and space. Step 7: Dynamic optimization of rescue danger radius: Based on the total risk cost map, calculate the dynamic basic danger radius that is linked to the risk distribution density, set a dynamic adjustment coefficient that is positively correlated with the dynamic risk change rate of the fracture for adaptive adjustment, and use a multi-objective evolutionary algorithm for adaptive preference adjustment in the rescue stage to select the Pareto optimal solution as the final rescue danger radius. Step 8: Autonomous localization and path planning of the robotic arm: Visual inertial SLAM is used to achieve accurate localization of the rescue robotic arm in the absence of GPS. Based on the total risk cost map and dynamically optimized safe operation hazard radius, the lowest risk operation path is planned. When the dynamic risk triggers the threshold, local dynamic replanning is performed to complete the adaptive risk avoidance of the rescue operation.

4. The evasion method according to claim 3, characterized in that, In step one, the PTP precise time protocol is used to perform hard synchronization of multi-source data to achieve timestamp alignment between point cloud and image data; preprocessing includes motion distortion correction and statistical filtering to remove outliers from the original point cloud data, resulting in a preprocessed color point cloud data stream with color information.

5. The evasion method according to claim 3, characterized in that, In step two, the geometry-texture joint feature extraction for crack geometry abrupt change sensitivity specifically involves: For any point p, take its neighborhood point set N(p) and calculate the three-dimensional centroid of the neighborhood point set. Construct the covariance matrix: Performing eigenvalue decomposition on the covariance matrix C yields the following results: eigenvalues , , , and the corresponding feature vectors; Calculate roughness ; Calculate normal vector consistency ,in Let p be the normal vector of point p, and let the minimum eigenvalue be the eigenvector. The corresponding eigenvectors are determined. Let be the normal vector of the neighboring point q; Calculate curvature change ; In the LAB color space, calculate the color covariance matrix of the neighborhood of point p. ,in Let q be the LAB color vector. Given the color mean vector within the neighborhood, take the trace of the color covariance matrix. As a texture feature; The above features are concatenated with the original 3D coordinates and RGB color values ​​of point p to construct a 10-dimensional multi-channel input feature vector: Where x, y, z are the three-dimensional spatial coordinates of point p, and R, G, B are the color channel values ​​of point p. Let be the roughness of the local surface where point p is located. For the normal vector of point p to be consistent, Let p be the curvature change of the local surface where point p is located. Let be the trace of the color covariance matrix of the neighborhood of point p in the LAB region.

6. The evasion method according to claim 3, characterized in that, In step three, the semantic segmentation network is constructed using a deep learning library combined with a sparse convolutional network. During model training, self-supervised comparative learning pre-training is first performed on large-scale unlabeled cave point cloud data, and then fine-tuning is performed using a semi-supervised method combined with real labeled samples. The real labeled data are labeled data of fissures, intact rock masses, and potentially dangerous rock masses collected from real cave environments, and the synthetic simulation data are cave scene data with fissure labels generated by a 3D simulation engine.

7. The evasion method according to claim 3, characterized in that, In step four, for each voxel in the semantic octree map... The weighting coefficients are dynamically determined based on the rock mass structure plane classification rules, and the stability probability is calculated. : in, voxels The normalized value corresponding to the angle between the fracture orientation and the direction of gravity. voxels The normalized value corresponding to the crack width, voxels Normalized value corresponding to the connectivity of the fracture; Voxels determined based on rock mass structural plane classification rules The structural class of the area is comprehensively evaluated based on the crack density, roughness, filling material characteristics, and spatial relationship with the main structural surface; An adaptive weighting function related to the structural plane level, satisfying Furthermore, a nonlinear mapping was performed on the dominance of each parameter on rock mass stability at different levels, resulting in a weighting bias towards dip angle and connectivity at low-level structural planes, and a weighting bias towards fracture width and connectivity at high-level structural planes; As voxels The static risk cost is stored in the corresponding node of the semantic octree map.

8. The evasion method according to claim 3, characterized in that, Step five specifically involves: Inter-frame non-rigid registration is performed on key crack regions to calculate the crack width variation. Changes in the principal directions of the displacement field on both sides of the fracture In addition to the local microseismic energy accumulation rate, a time series of fracture state variables over the past L time periods was constructed. , where X(t) is the fracture state quantity at time t, including the calculation of the fracture width change, the change of the principal direction of the displacement field on both sides of the fracture, the change acceleration, and the local microseismic energy accumulation rate; Inputting state time series data into a lightweight time series model to predict the future. Predicted value of crack width change at time step Compared with predicted energy release values ; The dynamic risk cost is calculated as follows: ,in This represents the quantification value of the uncertainty in time series prediction; where α is the weight of the crack width variation; and β is the weight of energy release. The weights are quantified values ​​for uncertainty, and ; when When the preset prior threshold for rock mass instability is exceeded, the dynamic risk cost coefficient... According to nonlinear coefficient Magnification, the nonlinear coefficient and The magnitude exceeding the threshold is positively correlated.

9. The evasion method according to claim 3, characterized in that, In step six, the formula for calculating the total risk cost is as follows: in, Static risk weights For dynamic risk weights, .

10. The evasion method according to claim 3, characterized in that, Step seven specifically involves: S71. Obtain basic parameters: The rated operating radius of the rescue robotic arm, The radius of influence of the unstable rock mass. To ensure equipment safety, d represents the building risk coefficient, and d represents the real-time distance between the rescue robotic arm and the unstable rock mass. The minimum threshold for the danger radius. This represents the maximum threshold for the danger radius. S72. Calculate the basic hazard radius Among them, the equipment safety factor With building risk coefficient The risk level is dynamically adjusted based on the distribution density of high-risk areas in the overall risk cost map. S73. Dynamically adjust to obtain the dynamic danger radius. Where λ is the dynamic adjustment coefficient, and is related to the rate of change of dynamic risk of the fracture. Positive correlation; S74. Obtain the initial danger radius by performing threshold truncation. ; S75. Using the initial danger radius as the decision variable, construct a dual optimization objective: maximize rescue efficiency. Where Coverage(R) is the search and rescue coverage area constrained by the danger radius R; minimizing rescue risk. , where Risk(R) is the cumulative total risk cost under the constraint of the danger radius R; S76. The dual optimization problem is decomposed into multiple sub-problems by using an adaptive weight vector for the rescue phase. Then, a multi-objective evolutionary algorithm is used to obtain the Pareto optimal solution set through crossover and mutation iteration. Based on the preset efficiency-risk preference weight, the final rescue danger radius is selected from the Pareto optimal solution set.

11. The evasion method according to claim 3, characterized in that, In step eight, the cost function used for path planning is: in, The cumulative total risk cost along the path is represented by Length(path), where Length(path) is the total path length and γ is the path length weighting coefficient. This is applied when the crack width changes... When the preset safety threshold is exceeded, local dynamic replanning is triggered to generate an avoidance trajectory and control the robotic arm to retract to a safe area; the visual inertial SLAM adopts a tightly coupled optimization framework, which integrates binocular visual images with IMU data from gyroscopes and accelerometers to output the six-degree-of-freedom pose of the rescue robotic arm.

Images