A Precision Spray Control Method for Coal Mines Based on Point Cloud Segmentation and Dust Prediction

Abstract

The present application belongs to the technical field of intelligent mine, and specifically relates to a coal mine precise spray control method based on point cloud segmentation and dust prediction, aiming to improve the intelligent and fine level of coal mine dust control. It includes real-time collection of three-dimensional point cloud data of coal mine operation environment by laser radar, semantic segmentation of three-dimensional point cloud data by using deep learning model, identification of dust distribution area, mine air flow parameter and operation intensity information, and construction of dynamic spatial scene semantic graph. Based on the dynamic spatial scene semantic graph, an AI prediction model integrating space-time features is used to real-time predict the dust diffusion path, concentration change trend and high-risk area. According to the dust prediction result, combined with the field working condition and spray system layout, the start-stop instruction, spray direction parameter and intensity parameter of the intelligent spray device are generated, and the dust reduction effect data is obtained. Through the closed-loop feedback mechanism, the dust reduction effect data is fed back to the control end, and the spray control strategy is continuously optimized.

Description

Technical Field

[0001] This invention relates to the field of smart mining technology, and in particular to a method for precise spray control in coal mines based on point cloud segmentation and dust prediction. Background Technology

[0002] During coal mining, operations such as tunneling, transportation, and loading generate a large amount of coal dust and silica dust. If the dust concentration exceeds the safety threshold, it will not only seriously endanger the health of workers and induce occupational diseases such as pneumoconiosis, but also increase the risk of major safety accidents such as gas explosions and coal dust explosions, becoming a key issue restricting safe production and intelligent development in coal mines.

[0003] The dust suppression spraying equipment commonly used in existing mines mostly adopts traditional control methods based on fixed time intervals or fixed areas, lacking the ability to dynamically respond to actual working conditions. Existing spraying systems are mostly controlled by independent units, lacking a coordinated linkage mechanism. Parameters such as spray pressure and angle are fixed and cannot be dynamically adjusted according to dust diffusion trends, resulting in low dust suppression efficiency, serious water waste, and even localized overspraying or blind spots, failing to meet the needs of modern coal mines for green, safe, and efficient dust control. Summary of the Invention

[0004] The purpose of this invention is to provide a precise spray control method for coal mines based on point cloud segmentation and dust prediction, aiming to improve the intelligence and precision of coal mine dust control.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The present invention provides a precise spray control method for coal mines based on point cloud segmentation and dust prediction. The method includes S1: Real-time acquisition of three-dimensional point cloud data of the coal mine operating environment using lidar, semantic segmentation of the three-dimensional point cloud data using a deep learning model, identification of dust distribution areas, mine airflow parameters, and operation intensity information, and construction of a dynamic spatial scene semantic map. S2: Based on the dust distribution areas, mine airflow parameters, and operation intensity information in the dynamic spatial scene semantic map, an AI prediction model integrating spatiotemporal features is used to predict the dust diffusion path, concentration change trend, and high-risk areas in real time. The AI ​​prediction model integrating spatiotemporal features is an artificial intelligence prediction model that includes time series features and spatial distribution features, used to simulate the dynamic diffusion process of dust under the action of airflow. S3: Based on the dust prediction results, combined with the on-site working conditions and spray system layout, start / stop commands, spray direction parameters, and intensity parameters of the intelligent spray device are generated, and dust suppression effect data is obtained. The dust suppression effect data is then transmitted back to the control terminal through a closed-loop feedback mechanism to continuously optimize the spray control strategy.

[0006] In some embodiments, semantic segmentation of 3D point cloud data is performed using a deep learning model, including: S11: Feature extraction is performed on the original point set to obtain the 3D spatial coordinates, reflection intensity, and surface normal vector of each point. An adaptive voxel partitioning strategy is used to adjust the voxel scale and generate geometric feature vectors. The original point set is the unprocessed point cloud data set output by the LiDAR. S12: A feature propagation network based on dynamic graph convolution is constructed. An adjacency graph is built using local point cloud topological relationships, and the geometric feature vectors are updated through geometric attention weights. S13: The segmentation result is optimized using a composite loss function. The composite loss function includes a classification loss term and a boundary structure preservation regularization term. The classification loss term indicates the difference between the predicted label and the true label, and the boundary structure preservation regularization term is used to constrain the spatial consistency of the prediction results of adjacent points.

[0007] In some embodiments, adjusting the voxel scale using an adaptive voxel partitioning strategy means dynamically adjusting the voxel scale based on the point density of the original point set within a set radius.

[0008] In some embodiments, the geometric attention weights are calculated based on the Euclidean distance between pairs of points in the original point set, and the Euclidean distance between the pairs of points is inversely proportional to the weight coefficient.

[0009] In some embodiments, the classification loss term is the cross-entropy loss. The boundary structure preserving regularization term maintains the continuity of the semantic boundary by constraining the similarity of the prediction results of neighboring points.

[0010] In some embodiments, step S2 includes: S21: fusing semantic scene graph embedding, wind speed and direction vectors, work intensity information, and dust concentration compression features to generate a spatiotemporal state vector. The spatiotemporal state vector is a comprehensive vector representing the spatiotemporal characteristics of the coal mine environment at a given moment. Semantic scene graph embedding is an embedding representation that transforms a dynamic spatial scene semantic graph into a low-dimensional vector. S22: constructing a time-gated spatial migration network, generating a spatial migration matrix through time-gated units, simulating the dust diffusion process, and predicting the dust concentration distribution at the next moment. The time-gated spatial migration network includes a neural network with a time-gated mechanism for simulating the dust spatial migration process. The spatial migration matrix is ​​a matrix representing the probability of dust migration between different regions. S23: optimizing the prediction model using a trend consistency regularized loss function. The loss function includes a mean squared error loss term and a diffusion trend constraint term, which ensures the gradient consistency between the predicted dust diffusion direction and the actual diffusion direction.

[0011] In some embodiments, the time-gating unit normalizes the element values ​​of the spatial migration matrix to the (0,1) interval through an activation function to characterize the dust migration probability between regions.

[0012] In some embodiments, step S3 further includes coordinated control of the intelligent spray device according to the spray control strategy. The intelligent spray device includes multiple spray units. Coordinated control of the intelligent spray device according to the spray control strategy includes: S31: Establishing a coordinated control model for the spray units, using spray intensity as the control variable, constructing an optimization function that includes dust concentration control targets and resource consumption targets, and solving for the optimal spray intensity parameters under resource constraints. The coordinated control model for the spray units is an optimization model used to coordinate the operating parameters of multiple spray units. The resource constraints are restrictions on the intensity range of a single spray unit and the total spray intensity of the system. S32: Dynamically adjusting the spray pressure parameters based on the dust concentration error, and adjusting the spray angle parameters according to the dust concentration gradient direction. The dust concentration error refers to the difference between the predicted dust concentration and the target concentration threshold. S33: Introducing a deep reinforcement learning strategy, using the current dust concentration, airflow parameters, and historical spray parameters as state inputs, and the spray intensity parameters as action outputs, optimizing the decision model through a reward function. The reward function is a feedback function set according to the degree of dust concentration reduction and resource consumption.

[0013] In some embodiments, the objective of the optimization function is to minimize the weighted sum of the predicted dust concentration and the spray resource consumption.

[0014] In some embodiments, deep reinforcement learning strategies update decision network parameters by minimizing Bellman error.

[0015] Compared with the prior art, the beneficial effects of this application are as follows:

[0016] 1. This application provides a precise spray control method for coal mines based on point cloud segmentation and dust prediction. It acquires 3D point cloud data using lidar and combines it with a deep learning model for semantic segmentation, enabling accurate identification of dust distribution areas, mine airflow parameters, and operational intensity information, thus constructing a dynamic spatial scene semantic map. Based on the multi-dimensional information in the dynamic spatial scene semantic map, an AI prediction model integrating spatiotemporal features is used to simulate the dust diffusion process, enabling real-time prediction of dust diffusion paths, concentration change trends, and high-risk areas. This model overcomes the limitations of traditional models based on empirical rules or simplified physical models, accurately depicting the nonlinear and dynamic diffusion characteristics of dust under airflow, identifying potential risks in advance, providing forward-looking guidance for spray control, and enhancing the initiative and scientific nature of coal mine dust management.

[0017] 2. Based on dust prediction results, the system generates start / stop commands, direction, and intensity parameters for the intelligent spray device. Through a closed-loop feedback mechanism, it continuously optimizes the control strategy, enabling precise, on-demand, targeted, and dynamic spraying. Compared to existing spraying methods with fixed parameters and independent control, this method combines on-site conditions and spray system layout to achieve coordinated operation of multiple devices. While ensuring dust suppression effectiveness, it reduces water waste, avoids localized overspraying or blind spots, significantly improving the efficiency and economy of spray dust suppression, and promoting the intelligent and refined upgrading of coal mine dust control. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a coal mine precision spray control method based on point cloud segmentation and dust prediction provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of the invention, it should be understood that the terms "upper," "lower," "left," "right," "front," "rear," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or relative positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Unless otherwise specified, the above-mentioned orientational descriptions can be flexibly set in practical applications, provided that the relative positional relationships shown in the accompanying drawings are satisfied.

[0021] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0022] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "communication" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection. They can refer to a direct connection or an indirect connection through an intermediate medium, or a communication between the internal components of two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0023] In embodiments of the invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, article, or apparatus 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, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, article, or apparatus that includes that element.

[0024] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0025] As mentioned in the background section, there is an urgent need for an integrated technology that can achieve precise perception of the working environment, dynamic prediction of dust diffusion, and intelligent collaborative control of the spray system. This technology would address issues such as insufficient perception accuracy, prediction lag, and rigid control strategies in existing technologies, thereby improving the intelligence and precision of coal mine dust control.

[0026] In view of this, embodiments of this application provide a method for precise spray control in coal mines based on point cloud segmentation and dust prediction, exemplarily, such as... Figure 1 As shown. The method includes:

[0027] S1. Perceiving and modeling the working environment: Real-time acquisition of 3D point cloud data of the coal mine working environment through LiDAR, semantic segmentation of the 3D point cloud data using a deep learning model, identification of dust distribution areas, mine airflow parameters and work intensity information, and construction of dynamic spatial scene semantic map.

[0028] Among them, 3D point cloud data refers to a dataset containing the 3D spatial coordinates and reflection characteristics of each point in the coal mine operating environment, obtained by emitting laser beams from a lidar and receiving reflected signals. Dynamic spatial scene semantic graph refers to a spatial scene model that integrates semantic information such as dust distribution, airflow parameters, and operating intensity, and is dynamically updated with the operating environment.

[0029] In some embodiments, a deep learning model is used to perform semantic segmentation on 3D point cloud data, including...

[0030] S11: Extract features from the original point set to obtain the three-dimensional spatial coordinates, reflection intensity, and surface normal vector of each point. Adjust the voxel scale using an adaptive voxel partitioning strategy to generate geometric feature vectors.

[0031] The original point set is the unprocessed point cloud data set output by the LiDAR. The geometric feature vector refers to a feature vector containing geometric attributes such as the three-dimensional coordinates of a point, reflection intensity, and surface normal vector. In some embodiments, adjusting the voxel scale using an adaptive voxel partitioning strategy means dynamically adjusting the voxel scale based on the point density within a set radius in the original point set to balance the impact of uneven point cloud density on feature extraction.

[0032] As one possible approach, to improve the expressive power of point cloud data in complex coal mine scenarios, the original point set is first processed. Preprocessing is performed; each point Represented as three-dimensional spatial coordinates And accompanied by its reflection intensity and surface normal vector To avoid the interference of uneven density on learning, an adaptive voxel partitioning strategy is adopted to dynamically adjust the scale of the voxel containing each point. The definition is as follows:

[0033]

[0034] in, For point In radius Point density within, Indicated by point The spatial volume of a three-dimensional spherical neighborhood centered at r. To adjust the hyperparameters at the voxel scale, It is a small constant that prevents division by zero. For point The points surrounding the points; the geometric feature vector of the extracted points is:

[0035]

[0036] Used as input for deep semantic modeling, ensuring that both structural and material information are encoded.

[0037] S12: Construct a feature propagation network based on dynamic graph convolution, build an adjacency graph using local point cloud topology, and update the geometric feature vector through geometric attention weights.

[0038] Dynamic graph convolution refers to a graph convolution operation that dynamically adjusts the kernel parameters according to the local point cloud topology; geometric attention weights are weight coefficients calculated based on the spatial distance between point pairs to characterize the importance of neighboring points. In some embodiments, geometric attention weights are calculated based on the Euclidean distance between point pairs in the original point set, and the Euclidean distance between point pairs is inversely proportional to the weight coefficients.

[0039] As one possible implementation, in the deep semantic modeling stage, a feature propagation structure based on dynamic graph convolution is constructed, and an adjacency graph is dynamically built using local point cloud topology information. The update formula is:

[0040]

[0041] in, For the first midpoint Embedded representation, For the first Adjacent points in the layer feature, For the first Adjacent points in the layer feature, For shared sensing machines. For point The local neighborhood point set; The geometric attention weights between point pairs are used to enhance geometric correlation modeling, and are defined as follows:

[0042]

[0043] in, This represents the Euclidean distance between two points. This is a distance attenuation control parameter to prevent interference from distant points.

[0044] S13: Optimize the segmentation results using a composite loss function.

[0045] The composite loss function includes a classification loss term and a boundary structure preservation regularization term. The classification loss term indicates the difference between the predicted label and the true label, while the boundary structure preservation regularization term is used to constrain the spatial consistency of the prediction results of neighboring points.

[0046] In some embodiments, the classification loss term is the cross-entropy loss. The boundary structure preserving regularization term maintains the continuity of the semantic boundary by constraining the similarity of the prediction results of neighboring points.

[0047] More specifically, to improve the semantic accuracy and spatial consistency of the segmentation results, a composite loss function is designed. It consists of classification loss and boundary structure preservation regularization term:

[0048]

[0049] in Standard cross-entropy loss measures the predicted label. With real labels The difference; the second term is the spatial smoothing regularization term, used to constrain adjacent points. and Prediction results and Similarity, maintaining semantic boundary continuity The weighting factor for the regularization term; To control the degree of influence of spatial distance on the regularization term.

[0050] S2: Based on dust distribution areas, mine airflow parameters, and work intensity information in the dynamic spatial scene semantic map, an AI prediction model that integrates spatiotemporal features is used to make real-time predictions of dust diffusion paths, concentration change trends, and high-risk areas.

[0051] The purpose of step S2 is to dynamically predict dust diffusion, including dynamic simulation and real-time prediction of dust diffusion paths, concentration changes, and high-risk areas, providing a forward-looking assessment of pollution trends. Specifically, the AI ​​prediction model integrating spatiotemporal features is an artificial intelligence prediction model that includes time series features and spatial distribution features, used to simulate the dynamic diffusion process of dust under the influence of airflow.

[0052] In some embodiments, step S2 includes:

[0053] S21: A spatiotemporal state vector is generated by integrating semantic scene graph embedding, wind speed and direction vectors, work intensity information, and dust concentration compression features. The spatiotemporal state vector is a comprehensive vector representing the spatiotemporal characteristics of the coal mine environment at a given moment. Semantic scene graph embedding is an embedding representation that transforms a dynamic spatial scene semantic graph into a low-dimensional vector.

[0054] For example, a spatiotemporal state vector is constructed by fusing features from multiple input sources. Depicting moments Coal mine environmental conditions:

[0055]

[0056] in, It is a semantic scene graph embedding that represents the semantic distribution and boundary structure of point clouds; It is a wind speed and direction vector, used to model the diffusion driving force; This is information on the workload of equipment and personnel; It is a low-dimensional compressed representation of the current dust concentration distribution map; It is a vector concatenation operation used to unify the modeling input; this state vector comprehensively considers spatial layout, physical driving force and pollution source intensity, and is the basis for AI prediction model prediction.

[0057] S22: Construct a time-gated spatial migration network. A spatial migration matrix is ​​generated through time-gated units to simulate the dust diffusion process and predict the dust concentration distribution at the next time step. The time-gated spatial migration network includes a neural network with a time-gated mechanism to simulate the dust spatial migration process. The spatial migration matrix is ​​a matrix representing the probability of dust migration between different regions.

[0058] In some embodiments, the time-gating unit normalizes the element values ​​of the spatial migration matrix to the (0,1) interval through an activation function to characterize the dust migration probability between regions.

[0059] The Time-Gated Spatial Migration Network (TGSM-Net) is essentially a network that uses time-gated units to control the spatial migration matrix. Dynamic modeling of diffusion weights:

[0060]

[0061] in, Predicted time The diffusion and migration matrix represents the diffusion and migration of pollution from... The probability of flow between regions; It is the Sigmoid function, used to normalize the transfer weights to... interval; These are the learnable parameters of an AI prediction model, used to transform the state vector Mapped to a diffusion structure; It is about predicting the next moment. The dust concentration distribution vector is obtained; the spatial migration matrix is ​​dynamically adjusted through a time gating mechanism to simulate the dust diffusion process under complex airflow and environment.

[0062] S23: The prediction optimization model is performed using a trend consistency regularization loss function. The loss function includes a mean squared error loss term and a diffusion trend constraint term. The diffusion trend constraint term is used to ensure the gradient consistency between the predicted dust diffusion direction and the actual diffusion direction.

[0063] Setting a loss function can effectively enhance the sensitivity of AI prediction models to diffusion trends. For example, a trend consistency regularization loss is proposed. This encourages AI prediction models to maintain consistency with the actual diffusion direction in terms of gradient:

[0064]

[0065] in, It is the standard mean square error loss, and the predicted concentration is consistent with the actual concentration. For a moment The actual dust concentration; and These represent spatial locations. The predicted and actual concentration gradients reflect the direction of diffusion. It is a trade-off term used to adjust the influence of trend normalization. It represents the number of intelligent spray devices; this loss enhances the alignment of the diffusion structure based on the regression accuracy, making the AI ​​prediction model robust and practical in predicting the evolution path of pollution hotspots.

[0066] S3: Based on the dust prediction results, combined with the on-site working conditions and the layout of the spray system, generate start and stop commands, spray direction parameters and intensity parameters for the intelligent spray device, and obtain dust suppression effect data. Through a closed-loop feedback mechanism, the dust suppression effect data is transmitted back to the control terminal to continuously optimize the spray control strategy.

[0067] In some embodiments, step S3 further includes coordinating control of the intelligent spraying device according to a spray control strategy. Coordinating control of the intelligent spraying device according to a spray control strategy includes:

[0068] S31: Establish a collaborative control model for the spray unit, using spray intensity as the control variable, construct an optimization function that includes dust concentration control target and resource consumption target, and solve for the optimal spray intensity parameters under the condition of satisfying resource constraints.

[0069] The spray unit coordinated control model is an optimization model used to coordinate the operating parameters of multiple spray units. Resource constraints are limitations on the intensity range of individual spray units and the total spray intensity of the system. In some embodiments, the objective of the optimization function is to minimize the weighted sum of predicted dust concentration and spray resource consumption.

[0070] The intelligent spraying device includes multiple spraying units. For example, a collaborative control model for these spraying units is established. A smart spray device, defining the time of each spray unit. Spray intensity is the control variable ∈[0,1], used to adjust the opening of the spray valve, where 0 represents closed and 1 represents maximum spray intensity; the goal is to reduce resource consumption while meeting dust control requirements, and the following optimization objective function is constructed. :

[0071]

[0072] in, It is the predicted number in step S2 The region in the next moment The dust concentration distribution vector; It is the current number Spray intensity control quantity for each spray unit; , It is a weighted coefficient that balances pollution control with resource expenditure; It is a comprehensive objective function value, the smaller the better; at the same time, resource constraints are introduced:

[0073]

[0074] in, It represents the maximum total spray intensity that the system can withstand, limiting excessive spraying and resource waste; the spray unit collaborative control model can be solved online in real time, outputting the optimal spray allocation scheme.

[0075] S32: Dynamically adjust spray pressure parameters based on dust concentration error, and adjust spray angle parameters according to the direction of dust concentration gradient. Here, dust concentration error refers to the difference between the predicted dust concentration and the target concentration threshold.

[0076] Considering the dynamic changes in mine operations, dust generation and diffusion exhibit nonlinear time-varying characteristics. Therefore, an adaptive control mechanism is introduced into the pressure and direction parameters of the spray system to achieve precise control. For example, taking the first... Taking a single spray unit as an example, the spray pressure is defined as... This affects the spray distance and coverage area; the spray angle is... Determine the spray direction and area; set the target concentration threshold. Then the first The concentration error for each region is:

[0077]

[0078] Update rules with proportional-derivative control are adopted:

[0079]

[0080] in, It is the ratio and differential gain coefficient of the spray pressure that adjusts the response rate and stability; It represents the gradient direction of dust concentration in space, indicating the diffusion trend; sign(·) is the sign function, used to adjust the spray direction to tend towards the high concentration area; It is an angle adjustment step size; this mechanism can fine-tune the spray physical parameters in real time according to the current forecast, achieving a dual dynamic response.

[0081] S33: Introduce a deep reinforcement learning strategy, using the current dust concentration, airflow parameters, and historical spray parameters as state inputs, and the spray intensity parameters as action outputs, to optimize the decision model through a reward function.

[0082] Here, the reward function refers to a feedback function set according to the degree of dust concentration reduction and resource consumption. In some embodiments, the deep reinforcement learning strategy updates the decision network parameters by minimizing the Bellman error.

[0083] For example, in dynamic operating conditions and uncertain environments, a collaborative control strategy based on deep reinforcement learning is introduced to improve the system's intelligent decision-making capabilities and system state. Defined as:

[0084]

[0085] in, This is the current predicted dust concentration in the area; It is the current wind speed and direction vector; This refers to the previous spray control decision; defining the action. for:

[0086]

[0087] The control commands for each spray unit, and the instant reward function, are designed as follows:

[0088]

[0089] in, It is the dust residue penalty coefficient; It is the spray resource penalty coefficient; reward value. A larger value indicates a better system, with the goal of minimizing concentration and resource overhead. A deep reinforcement learning framework is used to minimize the Bellman error update policy network loss function. :

[0090]

[0091] in, This is a discount factor used to weigh the importance of current rewards against future long-term rewards; its value ranges from 0 to 1. These are the current network parameters; The state sampled during experience playback ,action ,award Next state Expectations These are the parameters of the target reinforcement learning network; This refers to the value of actions under the current strategy. In the next state In, all possible actions The maximum action value represents the expected future return under the optimal strategy; through reinforcement learning strategies, spray behavior can be autonomously optimized through continuous interactive learning, achieving multi-unit linkage control and enhancing the system's adaptability and robustness in complex scenarios.

[0092] This application addresses key issues in coal mine operations, such as high dust concentration, difficult diffusion control, and delayed spray response. It proposes an intelligent dust suppression method that integrates real-time segmentation of lidar point clouds, AI prediction of dust diffusion, and collaborative control of intelligent spray devices, and constructs an integrated system architecture of "spatial perception, dynamic prediction, intelligent decision-making, and closed-loop control".

[0093] Three-dimensional point cloud data is collected by lidar and semantic segmentation is performed using a deep learning model to accurately identify the work area, equipment distribution, and dust cloud location, thereby perceiving the mine's spatial structure and pollution status in real time (corresponding to the work environment perception and modeling in step S1). Furthermore, a spatiotemporal prediction model integrating multi-source data such as wind speed, direction, and work intensity is introduced, and a deep neural network is used to dynamically simulate and provide risk warnings for dust concentration evolution trends (corresponding to the dynamic prediction of dust diffusion in step S2). Finally, a spray collaborative control strategy based on optimization algorithms and reinforcement learning is designed to intelligently adjust the opening intensity, angle, and timing of spray units, achieving coordinated response among multiple devices and dynamic optimal allocation of spray resources (corresponding to the intelligent spray collaborative control in step S3).

[0094] The method provided in this application possesses high real-time performance, adaptability, and intelligence, significantly improving the spatial coverage efficiency and resource utilization of dust suppression. In practical coal mine applications, it can effectively reduce peak dust concentrations and duration of exceedances, decrease blind spot spraying and resource waste, improve mine operation safety and environmental quality, and provide key technical support for promoting the intelligent upgrading and green low-carbon transformation of coal mine ventilation and dust removal systems.

[0095] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0096] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A coal mine precision spraying control method based on point cloud segmentation and dust prediction, characterized in that, The application comprises the following steps: S1: collecting three-dimensional point cloud data of the coal mine operation environment in real time by laser radar, performing semantic segmentation on the three-dimensional point cloud data by using a deep learning model, the deep learning model comprising a feature propagation network based on dynamic graph convolution and a composite loss function, the composite loss function comprising a classification loss term and a boundary structure preservation regularization term; identifying dust distribution areas, mine air flow parameters and work intensity information by the semantic segmentation, and constructing a dynamic spatial scene semantic graph; S2: based on the dust distribution areas, mine air flow parameters and work intensity information in the dynamic spatial scene semantic graph, using an AI prediction model that fuses spatio-temporal features, the AI prediction model being a time-gated spatial migration network, generating a spatial migration matrix by a time-gated unit to simulate the dust diffusion process, and using a trend consistency regularization loss function to optimize the prediction results, to predict the dust diffusion path, concentration change trend and high-risk area in real time; S3: according to the dust prediction results, combining the field conditions and the layout of the spraying system, generating start-stop instructions, spraying direction parameters and intensity parameters of an intelligent spraying device, the intelligent spraying device comprising a plurality of spraying units, and realizing collaborative control of the plurality of spraying units by a spraying unit collaborative control model, adaptive parameter adjustment based on dust concentration error and a deep reinforcement learning strategy; obtaining dust reduction effect data, and feeding back the dust reduction effect data to the control end through a closed-loop feedback mechanism to continuously optimize the spraying control strategy; wherein the semantic segmentation of the three-dimensional point cloud data by using the deep learning model comprises: S11: performing feature extraction on the original point set to obtain the three-dimensional spatial coordinates, reflection intensity and surface normal vector of each point, adjusting the voxel size by using an adaptive voxel division strategy, and generating a geometric feature vector; wherein the original point set is an untreated point cloud data set output by the laser radar; S12: constructing a feature propagation network based on dynamic graph convolution, constructing an adjacency graph by using local point cloud topological relations, and updating the geometric feature vector by a geometric attention weight; S13: optimizing the segmentation results by using a composite loss function; wherein the composite loss function comprises a classification loss term and a boundary structure preservation regularization term; the classification loss term indicates the difference between the predicted label and the true label, and the boundary structure preservation regularization term is used to constrain the spatial consistency of the prediction results of adjacent points; the adaptive voxel division strategy is used to dynamically adjust the voxel size according to the point density of the points in the original point set within a certain radius range; the geometric attention weight is calculated based on the Euclidean distance between the points in the original point set, and the Euclidean distance between the points is inversely proportional to the weight coefficient; the classification loss term is cross-entropy loss; the boundary structure preservation regularization term preserves the continuity of the semantic boundary by constraining the similarity of the prediction results of adjacent points; Step S2 specifically comprises: S21: fuse semantic scene graph embedding, wind speed and direction vector, work intensity information and dust concentration compression features to generate a spatio-temporal state vector; wherein the spatio-temporal state vector is a comprehensive vector of the spatio-temporal characteristics of the coal mine environment at a certain moment; the semantic scene graph embedding is an embedding representation of converting a dynamic spatial scene semantic graph into a low-dimensional vector; S22: build a time-gated spatial migration network to generate a spatial migration matrix through a time-gated unit, simulate the dust diffusion process, and predict the dust concentration distribution at the next moment; wherein the time-gated spatial migration network includes a neural network with a time-gated mechanism for simulating the dust spatial migration process; the spatial migration matrix is a matrix representing the migration probability of dust between different regions; S23: use a trend consistency regularization loss function to optimize the model for prediction and optimization; the loss function includes a mean square error loss term and a diffusion trend constraint term, and the diffusion trend constraint term is used to ensure the gradient consistency of the predicted dust diffusion direction and the true diffusion direction.

2. The coal mine precision spraying control method based on point cloud segmentation and dust prediction according to claim 1, characterized in that, The time-gated unit normalizes the element values of the spatial migration matrix to the interval (0, 1) through an activation function to represent the dust migration probability between regions.

3. The coal mine precision spraying control method based on point cloud segmentation and dust prediction according to claim 1, characterized in that, The coordinated control of multiple spray units includes: S31: establish a spray unit coordinated control model, take spray intensity as a control variable, build an optimization function containing dust concentration control target and resource consumption target, and solve the optimal spray intensity parameter under the condition of meeting the resource constraint; wherein the spray unit coordinated control model is an optimization model for coordinating the working parameters of multiple spray units; the resource constraint condition is the limitation condition for the intensity range of a single spray unit and the total spray intensity of the system; S32: dynamically adjust the spray pressure parameter based on the dust concentration error, and adjust the spray angle parameter according to the dust concentration gradient direction; wherein the dust concentration error refers to the difference between the predicted dust concentration and the target concentration threshold; S33: introduce a deep reinforcement learning strategy, take the current dust concentration, wind flow parameters and historical spray parameters as state input, take the spray intensity parameter as action output, and optimize the decision model through a reward function; wherein the reward function is a feedback function set according to the dust concentration reduction degree and resource consumption.

4. The coal mine precision spraying control method based on point cloud segmentation and dust prediction according to claim 3, characterized in that, The objective of the optimization function is to minimize the weighted sum of the predicted dust concentration and the spray resource consumption.

5. The coal mine precision spraying control method based on point cloud segmentation and dust prediction according to claim 3, characterized in that, The deep reinforcement learning strategy updates the decision network parameters by minimizing the Bellman error.

Images