An AI-based intelligent safety monitoring method and system for coal bunkers in thermal power plants

Abstract

This invention discloses an AI-based intelligent safety monitoring method and system for coal bunkers in thermal power plants, belonging to the field of thermal power plant monitoring technology. The method includes: dividing the thermal power plant into coal bunker operating areas and non-operating areas based on a 3D visualization model; collecting video data from the coal bunker operating areas and preprocessing the video data; using anomaly recognition technology to analyze the preprocessed video data frame by frame to identify potential safety hazards in the coal bunker operating areas; these hazards include improper wearing of safety helmets, improper wearing of work clothes, and improper deployment of fire extinguishing equipment; constructing a risk level assessment model and using this model to assess the risks of potential safety hazards; and generating a multi-level early warning mechanism for the thermal power plant's coal bunker based on the risk assessment results. This invention achieves comprehensive and refined intelligent monitoring of the safety status of personnel operating in the thermal power plant's coal bunker and the deployment of fire extinguishing equipment.

Description

Technical Field

[0001] This invention relates to the field of thermal power plant monitoring technology, and more specifically, to an AI-based intelligent safety monitoring method and system for coal bunkers in thermal power plants. Background Technology

[0002] A coal bunker in a thermal power plant is a large, enclosed or semi-enclosed coal storage structure, typically located near the boiler or coal conveying system. It is a crucial component of the power plant's fuel system, its primary function being to provide a stable and continuous coal supply to the boiler. Safety monitoring of the coal bunker involves using technologies such as video surveillance, sensors, and image recognition to monitor and provide early warnings of risks related to personnel, equipment operation, environmental indicators, and work activities within the bunker in real time, ensuring personnel safety, stable equipment operation, and fire prevention.

[0003] Existing coal bunker safety monitoring technologies mainly rely on video image recognition and target detection models, which can basically identify whether workers are wearing safety helmets and work clothes. However, there is a key defect: it can only determine whether the safety helmet and work clothes are present in the image, but cannot further detect whether they are worn correctly. For example, it cannot determine whether the safety helmet is offset, loose, or without a chin strap. This defect can lead to misjudgment or omission when facing actual safety hazards, thus failing to effectively identify actual high-risk behaviors and reducing the practicality and safety protection capabilities of the monitoring system.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] In response to the problems in related technologies, this invention proposes an AI-based intelligent safety monitoring method and system for coal bunkers in thermal power plants, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] According to one aspect of the present invention, an AI-based intelligent safety monitoring method for coal bunkers in thermal power plants is provided, the method comprising:

[0008] S1. Construct a three-dimensional visualization model of the thermal power plant, and divide the thermal power plant into a coal bunker operating area and a coal bunker non-operating area based on the three-dimensional visualization model;

[0009] S2. Collect video data of the coal bunker operation area, preprocess the video data, and use anomaly recognition technology to analyze the preprocessed video data frame by frame to identify potential safety hazards in the coal bunker operation area.

[0010] Safety hazards include not wearing a safety helmet correctly, not wearing work clothes correctly, and not properly deploying fire extinguishing equipment;

[0011] S3. Construct a risk level assessment model and use the risk level assessment model to conduct risk assessments on safety hazard behaviors. Based on the risk assessment results, generate a multi-level early warning mechanism for coal bunkers in thermal power plants.

[0012] Preferably, video data of the coal bunker operation area is collected, and the video data is preprocessed. Anomaly detection technology is used to analyze the preprocessed video data frame by frame to identify potential safety hazards in the coal bunker operation area, including:

[0013] S21. Use pre-deployed camera equipment to collect video data of the coal bunker operation area, extract the initial target image sequence from the video data, perform defogging processing on the initial target image sequence to obtain the target image sequence, and the target image sequence includes the image sequence of the workers and the image sequence of the work environment.

[0014] S22. Use an edge fitting algorithm to extract the helmet contour features in the image sequence of workers to determine the degree of helmet offset and identify the behavior of not wearing a helmet correctly based on the degree of helmet offset.

[0015] S23. Use optical flow field technology to extract the work clothes stream points in the image sequence of workers, and identify the behavior of not wearing work clothes correctly based on the work clothes stream points;

[0016] S24. Extract smoke features from the image sequence of the working environment, locate the flammable source of the coal bunker based on the smoke features, calculate the distance from the flammable source of the coal bunker to the fire extinguishing equipment, and identify the behavior of incorrectly deploying fire extinguishing equipment.

[0017] Preferably, the method of extracting the helmet contour features from the worker's image sequence using an edge fitting algorithm to determine the degree of helmet offset and identifying the behavior of not wearing a helmet correctly based on the degree of helmet offset includes:

[0018] S221. Threshold segmentation is performed on the worker images in the coal bunker worker image sequence to separate the safety helmet region, and the outer edge contour of the safety helmet region is located using the boundary tracking algorithm.

[0019] S222. Arbitrarily select contour pixels on the outer edge contour as the contour starting point, iterate and track from the contour starting point point by point according to the preset position until returning to the contour starting point, form the safety helmet contour feature, and store the safety helmet contour points in the coordinate array.

[0020] S223. Perform vertical error matching between the helmet contour features in the continuous frame images and the standard contour to generate vertical error curves for the continuous frames, and extract the maximum offset of the vertical error curves and compare it with a preset threshold.

[0021] S224. If the vertical error curve of consecutive frames tends to be stable and the maximum offset is less than or equal to the preset threshold, it means that the current worker's safety helmet is not offset. If the vertical error curve of consecutive frames tends to be jittery and the maximum offset is greater than the preset threshold, it means that the current worker's safety helmet is offset, and step S225 is executed.

[0022] S225. Convert the helmet outline points into helmet point cloud data, reconstruct the connectivity structure of the contact area between the chin strap and the helmet shell using a geometric topology reconstruction model, and identify the behavior of not wearing the helmet correctly based on the connectivity structure.

[0023] Preferably, the helmet contour points are converted into helmet point cloud data, and the connectivity structure of the contact area between the chin strap and the helmet shell is reconstructed using a geometric topology reconstruction model. Based on this connectivity structure, the identification of improper helmet wearing includes:

[0024] S2251. Extract the helmet contour points from the coordinate array and convert them into helmet point cloud data. Use the helmet point cloud data as input and perform topological reconstruction on the connected structure of the contact area between the chin strap and the helmet shell using geometric topological reconstruction technology.

[0025] S2252. Perform topological analysis on the connectivity structure of the contact area between the mandibular band and the cap obtained by topological reconstruction to obtain the connectivity characteristics of the contact area between the mandibular band and the cap.

[0026] S2253. The zero-order connectivity property and the first-order connectivity property are compared with the preset thresholds respectively. Based on the comparison results, the connection and buckle engagement status of the chin strap of the safety helmet are determined. Combined with the maximum offset of the corresponding vertical error curve, the behavior of not wearing the safety helmet correctly is comprehensively identified.

[0027] Preferably, the connectivity properties include zero-order connectivity properties and first-order connectivity properties;

[0028] Among them, the zero-order connectivity property is used to represent the number of branches in a connected structure and to determine whether there is a break in the mandibular band.

[0029] First-order connectivity is used to represent the number of closed voids in a connected structure and to determine whether a snap-fit ​​part forms a complete ring-shaped connection structure.

[0030] Preferably, using point cloud data of the safety helmet as input, topological reconstruction of the connectivity structure of the contact area between the chin strap and the helmet shell is performed using geometric topological reconstruction technology, including:

[0031] Extract the contour points of the bottom edge contour of the safety helmet, perform the initial construction operation, and generate the initial connected structure of the chin strap and the helmet shell contact area containing only the upper and lower basic surfaces.

[0032] Set the reconstruction parameters of the initial connected structure. Under the constraints of the reconstruction parameters, extract the triangular patch sequence between adjacent edge contours layer by layer according to the hierarchical order in the safety helmet point cloud data and perform triangular patch construction to fill the point cloud region in the initial connected structure.

[0033] Check if the vertices in the triangle construction already exist in the generated initial connected structure. If the vertices already exist, close the triangle; if the vertices do not exist, add vertices and then perform the face closure operation.

[0034] Repeat the triangular piece construction process until the contact structure between the chin strap and the cap shell is reconstructed into a three-dimensional geometric topological model.

[0035] Preferably, the process involves extracting smoke features from the image sequence of the work environment, locating the flammable source of the coal bunker based on these features, calculating the distance from the flammable source to the fire extinguishing equipment, and identifying instances of improperly deployed fire extinguishing equipment.

[0036] S241. Based on the sparse attention mechanism, the graph neural network extracts the smoke focus field in the image sequence of the working environment and reverses the smoke diffusion source. The smoke diffusion source is superimposed with the historical high-risk occurrence point to locate the flammable source point of the coal bunker in the working environment.

[0037] S242. Calculate the spatial distance between the flammable source of the coal bunker and the current deployment location of the fire extinguishing equipment, assess the response range of the fire extinguishing equipment, and identify potential blind spot locations based on the assessment results of the response range.

[0038] S243. Input the flammable source points of the coal bunker, the current deployment location of the fire extinguishing equipment, and the blind spot location codes into the event reasoning model based on fire extinguishing accessibility rules, and reason out the behavior of incorrectly deployed fire extinguishing equipment.

[0039] Preferably, the flammable source points of the coal bunker, the current deployment location of the fire extinguishing equipment, and the blind spot location are encoded and input into an event reasoning model based on fire extinguishing accessibility rules. The reasoning results in the following behaviors indicating incorrect deployment of fire extinguishing equipment:

[0040] S2431. Encode the flammable sources of coal bunkers, the current deployment locations of fire extinguishing equipment, and blind spot locations as sub-nodes to construct a multi-dimensional semantic graph;

[0041] S2432. Based on the semantic relationships between child nodes in a multidimensional semantic graph, the meta-path learning algorithm is used to analyze the association strength between child nodes.

[0042] S2433. Based on the predefined fire extinguishing accessibility rules and the association strength between child nodes, the deployment behavior of fire extinguishing equipment is determined using a fuzzy inference model to obtain the behavior probability distribution, and the behavior type of incorrectly deployed fire extinguishing equipment is identified based on the behavior probability distribution.

[0043] Preferably, analyzing the association strength between child nodes using the meta-path learning algorithm includes:

[0044] Obtain the set of paths from any child node to the other child nodes in the multidimensional semantic graph, and use any selected child node as the starting point to propagate the access probability layer by layer down along each path;

[0045] The intermediate child nodes visited during path propagation are integrated into the candidate node set. It is determined whether each intermediate child node in the candidate node set has a non-zero access probability in both paths. If so, the current intermediate child node is constructed with association features to represent the association strength between the child node and the target child node. If not, the current intermediate child node is marked as an irrelevant node.

[0046] Tracing back the complete path of intermediate child nodes that satisfy the association characteristics in the path propagation, abstracting the complete path into a meta-path, and counting the frequency of the meta-path in the path set, selecting meta-paths within a preset frequency range to construct a meta-path pattern set.

[0047] According to another aspect of the present invention, an AI-based intelligent safety monitoring system for coal bunkers in thermal power plants is also provided, the system comprising:

[0048] The work area division module is used to construct a three-dimensional visualization model of the thermal power plant, and divide the thermal power plant into coal bunker work areas and coal bunker non-work areas based on the three-dimensional visualization model;

[0049] The hazard behavior identification module is used to collect video data from the coal bunker operation area, preprocess the video data, and use anomaly recognition technology to analyze the preprocessed video data frame by frame to identify safety hazard behaviors in the coal bunker operation area. Safety hazard behaviors include not wearing a safety helmet correctly, not wearing work clothes correctly, and not deploying fire extinguishing equipment correctly.

[0050] The early warning mechanism construction module is used to build a risk level assessment model and use the risk level assessment model to conduct risk assessment on safety hazard behaviors. Based on the risk assessment results, a multi-level early warning mechanism for coal bunkers in thermal power plants is generated.

[0051] The beneficial effects of this invention are as follows:

[0052] 1. This invention utilizes frame-by-frame analysis and anomaly recognition technology of video data to identify potential safety hazards in coal bunker work areas in real time, such as not wearing safety helmets correctly, not wearing work clothes, or not deploying fire extinguishing equipment. This effectively improves the detection capability of safety hazards. By constructing a risk level assessment model, the identified safety hazard behaviors are assessed for risk. This not only accurately assesses potential safety risks but also generates a multi-level early warning mechanism based on the risk assessment results. This provides a scientific and accurate basis for decision-making in the safety management of coal bunkers in thermal power plants, enabling early warning and prevention of accidents, and improving the overall level of safety protection.

[0053] 2. This invention combines multiple AI technologies such as image processing, optical flow analysis, geometric modeling, and topology reconstruction to achieve comprehensive and refined intelligent monitoring of the safety status of coal bunker workers and the deployment of fire extinguishing equipment in thermal power plants. It can not only identify whether safety helmets and work clothes are worn, but also determine whether they are worn correctly, such as detecting whether the safety helmet is offset or whether the chin strap is connected completely. This significantly makes up for the shortcomings of traditional visual recognition technology, which can only identify whether they are worn but cannot determine the correctness of the wearing.

[0054] 3. This invention extracts the smoke focusing field in the working environment and reverses it to deduce the smoke diffusion source, achieving high-precision positioning of flammable sources in coal bunkers. Then, combined with the deployment location and response range calculation of fire extinguishing equipment, it identifies potential fire extinguishing blind spots, enhancing the spatial perception capability of fire risks. By encoding flammable sources, fire extinguishing equipment locations, and blind spot locations to construct a multi-dimensional semantic map, and combining predefined fire extinguishing accessibility rules and fuzzy inference models to determine deployment behavior, it helps to achieve automatic perception, intelligent diagnosis, and risk warning of fire extinguishing capability coverage blind spots in complex coal bunker environments. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of an AI-based intelligent safety monitoring method for coal bunkers in thermal power plants, according to an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram illustrating a specific implementation of an AI-based intelligent safety monitoring method for coal bunkers in thermal power plants, according to an embodiment of the present invention.

[0058] Figure 3 This is a schematic diagram of an AI-based intelligent safety monitoring system for coal bunkers in thermal power plants, according to an embodiment of the present invention.

[0059] In the picture:

[0060] 1. Work area division module; 2. Hazardous behavior identification module; 3. Early warning mechanism construction module. Detailed Implementation

[0061] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0062] According to an embodiment of the present invention, an AI-based intelligent safety monitoring method and system for coal bunkers in thermal power plants is provided.

[0063] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-2 As shown, according to an embodiment of the present invention, an AI-based intelligent safety monitoring method for coal bunkers in thermal power plants includes:

[0064] S1. Construct a three-dimensional visualization model of the thermal power plant, and divide the thermal power plant into coal bunker operating areas and coal bunker non-operating areas based on the three-dimensional visualization model.

[0065] It should be noted that constructing a 3D visualization model of a thermal power plant, and dividing the plant into coal bunker operating areas and non-operating areas based on this model, includes:

[0066] Step 1: Use lidar, UAV multi-view image acquisition or RGB-D camera to acquire point cloud data of the thermal power plant site and unify the coordinate system to complete the spatial registration and noise reduction of the original point cloud data;

[0067] Step 2: Extract key targets such as building structure, equipment, passage, and coal bunker, and perform semantic annotation based on spatial location, structural shape, and attribute features to form a labeled 3D semantic point cloud;

[0068] Step 3: Convert the segmented point cloud into a visualized 3D mesh model, and simultaneously map RGB textures to generate a realistic model;

[0069] Step 4: Combining the coal bunker structural design drawings and operating specifications, the coal bunker area in the model is divided into a coal bunker operating area and a coal bunker non-operating area according to function using geometric space analysis technology.

[0070] The coal bunker operation area refers to the area where personnel carry out daily operations, inspections, maintenance, and patrols. It typically includes the coal bunker entrance, coal loading and unloading area, control room, coal conveying system, monitoring equipment location, and personnel operation passages. It is characterized by frequent personnel activities and equipment operations, and has a high risk of safety hazards.

[0071] Non-operational areas of a coal bunker refer to areas not involved in daily operations and personnel activities. These mainly include coal storage areas, enclosed areas, equipment storage areas, and equipment rooms. These areas are primarily used for coal storage, equipment placement, and auxiliary functions.

[0072] S2. Collect video data of the coal bunker operation area and preprocess the video data. Use anomaly recognition technology to analyze the preprocessed video data frame by frame to identify safety hazards in the coal bunker operation area. Safety hazards include not wearing a safety helmet correctly, not wearing work clothes correctly, and not deploying fire extinguishing equipment correctly.

[0073] It should be noted that improper helmet wearing includes: the helmet is not properly secured, resulting in it shifting or becoming loose and unable to provide the intended protection; the helmet does not cover the entire head, such as not covering the forehead or back of the head, leaving some areas exposed; and the helmet is not securely fastened, such as the helmet strap being loose or not tightened properly.

[0074] Incorrect wearing of work clothes includes: not wearing the work clothes completely, such as not fastening the zipper or buttons, resulting in body exposure; not wearing the work clothes properly, leaving protective gaps, etc.

[0075] The improper deployment of fire extinguishing equipment includes: fire extinguishers or fire extinguishing equipment not being deployed in the prescribed locations and quantities, resulting in the inability to cover potential fire source areas; and the fire extinguishing equipment having insufficient effective distance or spray range to cover the area where the fire has occurred.

[0076] This process involves collecting video data from the coal bunker operation area, preprocessing the video data, and then using anomaly detection technology to analyze the preprocessed video data frame by frame to identify potential safety hazards in the coal bunker operation area, including:

[0077] S21. Use pre-deployed camera equipment to collect video data of the coal bunker operation area, extract the initial target image sequence from the video data, perform defogging processing on the initial target image sequence to obtain the target image sequence, and the target image sequence includes the image sequence of the workers and the image sequence of the work environment.

[0078] S22. Use an edge fitting algorithm to extract the helmet contour features from the worker's image sequence to determine the degree of helmet offset and identify the behavior of not wearing a helmet correctly based on the degree of helmet offset.

[0079] Among these methods, edge fitting algorithms are used to extract the contour features of safety helmets from the image sequences of workers to determine the degree of helmet offset, and the identification of improper helmet wearing based on the degree of helmet offset includes:

[0080] S221. Threshold segmentation is performed on the worker images in the coal bunker worker image sequence to separate the safety helmet region, and the outer edge contour of the safety helmet region is located using a boundary tracking algorithm.

[0081] It should be noted that the boundary tracking algorithm is the Freeman chain code algorithm. The Freeman chain code algorithm is a digital algorithm used to represent boundary shapes. It records the outer edge contour by encoding the boundary pixels in an image. The specific rules are:

[0082] Step 1: Starting from a known boundary pixel (usually the top-left boundary pixel), track the movement in a predefined direction. The direction of movement for each pixel is represented by a numerical code. In the Freeman chain code algorithm, eight directions are typically used to represent adjacent pixels. These eight directions are represented by numbers from 0 to 7: 0 represents upward, 1 represents upward to the right, 2 represents right, 3 represents downward to the right, 4 represents downward, 5 represents downward to the left, 6 represents left, and 7 represents upward to the left.

[0083] Step 2: Based on the positions of the boundary pixels in the image, trace the boundaries sequentially in these directions until returning to the starting point. For each step, record the direction number of the current pixel relative to the previous pixel, ultimately forming a chain code sequence of numbers that represents the shape of the outer boundary of the object in the image.

[0084] S222. Arbitrarily select contour pixels on the outer edge contour as the contour starting point, iterate and track from the contour starting point point by point according to the preset position until returning to the contour starting point, form the safety helmet contour feature, and store the safety helmet contour points in the coordinate array.

[0085] S223. Perform vertical error matching between the helmet contour features in the continuous frame images and the standard contour to generate vertical error curves for the continuous frames, and extract the maximum offset of the vertical error curves and compare it with a preset threshold.

[0086] S224. If the vertical error curve of consecutive frames tends to be stable and the maximum offset is less than or equal to the preset threshold, it means that the current worker's safety helmet is not offset. If the vertical error curve of consecutive frames tends to be jittery and the maximum offset is greater than the preset threshold, it means that the current worker's safety helmet is offset, and step S225 is executed.

[0087] It should be noted that the purpose of judging the vertical error curve trend and maximum offset of consecutive frames is to improve the accuracy of helmet offset detection and avoid misjudgment caused by single-frame errors. The error of a single frame may be affected by instantaneous external disturbances, data noise, or error sources. Therefore, by analyzing the error trend of consecutive frames, the helmet offset can be identified more stably and accurately.

[0088] Specifically, the vertical error curve of consecutive frames can reflect whether there is a continuous offset of the helmet. If the offset is continuous and regular, it indicates that the helmet has indeed shifted. However, if it is only an occasional fluctuation or a short-term error change, it can be determined that this is not caused by the helmet offset, but by a brief fluctuation caused by other factors.

[0089] S225. Convert the helmet outline points into helmet point cloud data, reconstruct the connectivity structure of the contact area between the chin strap and the helmet shell using a geometric topology reconstruction model, and identify the behavior of not wearing the helmet correctly based on the connectivity structure.

[0090] It should be noted that step S224 is a preliminary assessment of the worker's safety helmet status, used to determine whether to proceed to a more in-depth analysis process. In step S224, the safety helmet is evaluated for misalignment by analyzing the vertical error curves of consecutive frames. If the error curve tends to be stable and the maximum misalignment is less than or equal to a preset threshold, it indicates that the safety helmet is not misaligned, and the system considers the worker to be wearing it correctly, requiring no further action; if the error curve shows fluctuations and the misalignment is greater than the threshold, it is determined that the safety helmet is misaligned, at which point step S225 is triggered.

[0091] In step S225, the helmet contour points are converted into point cloud data, and a geometric topology reconstruction model is used to further analyze and reconstruct the connectivity structure of the contact area between the chin strap and the helmet shell to identify whether the helmet is worn correctly. In fact, step S224 acts as a threshold, designed to avoid performing complex geometric topology reconstruction for every frame.

[0092] The process involves converting helmet contour points into helmet point cloud data, reconstructing the connectivity structure of the chin strap contact area with the helmet shell using a geometric topology reconstruction model, and identifying instances of improper helmet wearing based on this connectivity structure.

[0093] S2251. Extract the helmet contour points from the coordinate array and convert them into helmet point cloud data. Use the helmet point cloud data as input and perform topological reconstruction on the connected structure of the contact area between the chin strap and the helmet shell using geometric topological reconstruction technology.

[0094] The process involves using point cloud data of the safety helmet as input and performing topological reconstruction on the connectivity structure of the contact area between the chin strap and the helmet shell using geometric topological reconstruction technology.

[0095] Extract the contour points of the bottom edge contour of the safety helmet, perform the initial construction operation, and generate the initial connected structure of the chin strap and the helmet shell contact area containing only the upper and lower basic surfaces.

[0096] Set the reconstruction parameters of the initial connected structure. Under the constraints of the reconstruction parameters, extract the triangular patch sequence between adjacent edge contours layer by layer according to the hierarchical order in the safety helmet point cloud data and perform triangular patch construction to fill the point cloud region in the initial connected structure.

[0097] Check if the vertices in the triangle construction already exist in the generated initial connected structure. If the vertices already exist, close the triangle; if the vertices do not exist, add vertices and then perform the face closure operation.

[0098] Repeat the triangular piece construction process until the contact structure between the chin strap and the cap shell is reconstructed into a three-dimensional geometric topological model.

[0099] It should be noted that by extracting the bottom edge contours of the safety helmet and performing initial construction operations, a reliable basic structure can be provided for subsequent topology reconstruction. By setting reasonable reconstruction parameters and progressively extracting edge contours and constructing triangular patches in a hierarchical order, gaps in the point cloud data can be effectively filled, ensuring the detail and integrity of the geometric model. Vertex existence detection and patch closure operations can prevent the generation of redundant data, ensuring the connectivity and rationality of the topology.

[0100] S2252. Perform topological analysis on the connectivity structure of the contact area between the mandibular band and the cap obtained by topological reconstruction to obtain the connectivity characteristics of the contact area between the mandibular band and the cap.

[0101] The connectivity properties include zero-order connectivity properties and first-order connectivity properties. Zero-order connectivity properties are used to represent the number of branches in the connected structure and to determine whether there is a break in the mandibular band. First-order connectivity properties are used to represent the number of closed cavities in the connected structure and to determine whether the snap-fit ​​part forms a complete ring-shaped connection structure.

[0102] It should be noted that the zero-order connectivity property is used to represent the number of branches in a connected structure. The number of branches is obtained by calculating the relationship between nodes and edges in the topology of the contact area between the mandibular girdle and the cap shell. If the number of branches is zero, it indicates that the mandibular girdle is continuous; if the number of branches is greater than zero, it indicates that the mandibular girdle has a break or incomplete connection.

[0103] First-order connectivity indicates the number of closed voids in a connected structure. A void refers to a region in the topology that is completely enclosed but not connected. When analyzing the contact area between the chin strap and the cap shell, the number of voids is calculated by checking for the existence of closed loop structures. If the number of voids is greater than zero, it indicates that the buckle area has not formed a complete loop connection structure.

[0104] S2253. The zero-order connectivity property and the first-order connectivity property are compared with the preset thresholds respectively. Based on the comparison results, the connection and buckle engagement status of the chin strap of the safety helmet are determined. Combined with the maximum offset of the corresponding vertical error curve, the behavior of not wearing the safety helmet correctly is comprehensively identified.

[0105] It should be noted that the zero-order connectivity and first-order connectivity properties are compared with preset thresholds. These preset thresholds are set according to the actual helmet design specifications and tolerances used in practical applications, such as thresholds for the number of branches and the number of cavities. These thresholds are used to determine whether the chin strap is broken or whether the buckle connection is incomplete. If the zero-order connectivity property (number of branches) exceeds the threshold, it indicates that the chin strap is broken; if the first-order connectivity property (number of cavities) exceeds the threshold, it indicates that the buckle connection is incomplete.

[0106] S23. Use optical flow field technology to extract the work clothes streamers in the image sequence of the workers, and identify the behavior of not wearing work clothes correctly based on the work clothes streamers.

[0107] It should be noted that the image sequence is processed using optical flow field technology. The Lucas-Kanade optical flow method is used to calculate the optical flow of each pixel in the image, i.e., to calculate its motion vector between consecutive frames. By extracting the flow points associated with the work clothes, tracking is performed based on the changes in the position of these flow points in the image. The motion patterns of the work clothes flow points, such as speed, acceleration, and direction, are calculated to determine whether they conform to the work clothes wearing specifications. If the motion of the flow points is abnormal, such as unstable direction, excessive speed, or large acceleration, it may indicate that the worker is not wearing the work clothes correctly. The anomaly identification and motion pattern analysis of the optical flow points, combined with the standard position and shape of the work clothes, determine whether there is any behavior of incorrect work clothes wearing.

[0108] In the image sequence, the optical flow vector of the work clothes streamer points refers to the pixel-level motion information obtained by estimating the motion of pixels in the work clothes region in consecutive frame images, rather than the physical meaning of the work clothes themselves having vector attributes.

[0109] Specifically, the pixel displacement of workwear areas (such as red or blue fabric, reflective strips, patterns, etc.) in a continuous frame as the worker moves is called the optical flow vector, which is a two-dimensional motion vector. Because the workwear has color or texture features (such as bright areas, reflective strips, name tags, patterns), these areas appear as feature regions with significant changes in pixel intensity in the image. When the worker moves, these feature regions change position in the image, and the optical flow algorithm can capture these changes and calculate the corresponding optical flow vector. In the specific implementation, the optical flow calculation formula includes:

[0110] The basic formula for calculating the optical flow field is:

[0111] I(x,y,t)=I(x+Δx,y+Δy,t+Δt);

[0112] In the formula, I(x,y,t) represents the pixel intensity of the image at position (x,y) and time t, Δx and Δy represent the displacement of the position between two time points, and Δt represents the time interval.

[0113] Motion model of optical flow field:

[0114] ;

[0115] In the formula, v(x,y) represents the optical flow vector at position (x,y). and These represent the gradients of image intensity in the spatial direction, respectively. This represents the change in image intensity over time.

[0116] S24. Extract smoke features from the image sequence of the working environment, locate the flammable source of the coal bunker based on the smoke features, calculate the distance from the flammable source of the coal bunker to the fire extinguishing equipment, and identify the behavior of incorrectly deploying fire extinguishing equipment.

[0117] This includes extracting smoke features from the image sequence of the work environment, locating the flammable source of the coal bunker based on the smoke features, calculating the distance from the flammable source of the coal bunker to the fire extinguishing equipment, and identifying behaviors such as improper deployment of fire extinguishing equipment.

[0118] S241. A graph neural network based on a sparse attention mechanism extracts the smoke focus field in the image sequence of the work environment and reverses the smoke diffusion source. The smoke diffusion source is superimposed with historical high-risk occurrence points to locate the flammable source point of the coal bunker in the work environment.

[0119] It should be noted that the graph neural network based on the sparse attention mechanism extracts the smoke focus field from the image sequence of the work environment and inversely infers the smoke diffusion source. By overlaying the smoke diffusion source with historical high-risk occurrence points, the flammable source points of the coal bunker in the work environment are located, including:

[0120] Step 1: Input the image frames of the work environment image sequence into the graph neural network model using the sparse attention mechanism. The model organizes key region nodes in the image in a graph structure, such as smoke areas, ventilation openings, heat source equipment, and structural boundaries, and represents spatial correlation by constructing edge weights between nodes. The sparse attention mechanism operates in the information transfer stage between nodes. It adaptively selects neighboring nodes with high information contribution for feature aggregation, thereby effectively avoiding information interference caused by redundant connections and enhancing sensitivity to sparse features (such as smoke edges and abrupt changes in concentration gradients).

[0121] Step 2: After multi-layer iteration of the network, the smoke saliency weight of each node is obtained, forming a smoke focusing field. Regions with high focusing values ​​are considered high-density smoke centers. Based on the gradient distribution of the smoke focusing field, the initial diffusion source location of the smoke is calculated using a spatial inverse vector field, serving as a potential ignition source or flammable trigger point.

[0122] Step 3: Spatially overlay the identified diffusion source locations with the heat map of historical high-risk events in the coal bunker, and use spatial statistical matching (e.g., based on KL divergence, mutual information, or cosine similarity) to measure the spatial overlap between the current diffusion source and historical high-risk points. High-matching areas are marked as potential flammable sources in the coal bunker in the current working environment.

[0123] S242. Calculate the spatial distance between the flammable source of the coal bunker and the current deployment location of the fire extinguishing equipment, assess the response range of the fire extinguishing equipment, and identify potential blind spot locations based on the assessment results of the response range.

[0124] It should be noted that the Euclidean distance formula is used to calculate the spatial distance between the flammable source and each fire extinguishing device. By calculating the spatial distance between multiple fire extinguishing devices and the flammable source, the coverage area of ​​each device can be determined, and the response range of the fire extinguishing devices can be further evaluated.

[0125] The response range is typically set as a specific maximum coverage radius based on the technical parameters of the fire extinguishing equipment, such as the spray range and the diffusion radius of the extinguishing agent. If the response range of a fire extinguishing device includes a flammable source, the device is considered to cover that flammable source; otherwise, it indicates the existence of a fire extinguishing blind spot. By calculating the response range of each fire extinguishing device (usually a circular or spherical area), the influence area of ​​each device is determined, thereby identifying potential blind spots in the work environment not covered by fire extinguishing equipment.

[0126] S243. Input the flammable source points of the coal bunker, the current deployment location of the fire extinguishing equipment, and the blind spot location codes into the event reasoning model based on fire extinguishing accessibility rules, and reason out the behavior of incorrectly deployed fire extinguishing equipment.

[0127] Specifically, the flammable source points of the coal bunker, the current deployment location of fire extinguishing equipment, and the location of blind spots are encoded and input into an event reasoning model based on fire extinguishing accessibility rules. The reasoning results in the following behaviors indicating incorrect deployment of fire extinguishing equipment:

[0128] S2431. Encode the flammable sources of coal bunkers, the current deployment locations of fire extinguishing equipment, and blind spot locations as sub-nodes to construct a multi-dimensional semantic graph;

[0129] S2432. Based on the semantic relationships between child nodes in a multidimensional semantic graph, the meta-path learning algorithm is used to analyze the association strength between child nodes.

[0130] The analysis of the association strength between child nodes using the meta-path learning algorithm includes:

[0131] Obtain the set of paths from any child node to the other child nodes in the multidimensional semantic graph, and use any selected child node as the starting point to propagate the access probability layer by layer down along each path;

[0132] The intermediate child nodes visited during path propagation are integrated into the candidate node set. It is determined whether each intermediate child node in the candidate node set has a non-zero access probability in both paths. If so, the current intermediate child node is constructed with association features to represent the association strength between the child node and the target child node. If not, the current intermediate child node is marked as an irrelevant node.

[0133] Tracing back the complete path of intermediate child nodes that satisfy the association characteristics in the path propagation, abstracting the complete path into a meta-path, and counting the frequency of the meta-path in the path set, selecting meta-paths within a preset frequency range to construct a meta-path pattern set.

[0134] It should be noted that the meta-path learning algorithm can efficiently mine deep semantic relationships between child nodes in a multidimensional semantic graph, avoiding the loss of association information caused by one-sided connections. By analyzing the access probability in path propagation, intermediate nodes with non-zero access probability in multiple paths are extracted as strongly associated feature nodes, which helps to accurately identify structural node pairs with semantic co-occurrence relationships, thereby enhancing the accuracy and robustness of node relationship modeling. By backtracking and statistically analyzing frequently occurring meta-paths, a stable semantic propagation structure can be formed, and a representative set of meta-path patterns can be constructed, providing interpretability for downstream tasks such as relationship prediction, risk propagation modeling, and behavioral reasoning.

[0135] S2433. Based on the predefined fire extinguishing accessibility rules and the association strength between child nodes, the deployment behavior of fire extinguishing equipment is determined using a fuzzy inference model to obtain the behavior probability distribution, and the behavior type of incorrectly deployed fire extinguishing equipment is identified based on the behavior probability distribution.

[0136] It should be noted that predefined fire suppression accessibility rules typically include requirements for the deployment location and coverage area of ​​fire suppression equipment. For example, fire suppression equipment should be able to cover all potential fire source areas to avoid fire suppression blind spots. The rules take into account factors such as the equipment's response range, spatial layout, environmental conditions, and potential obstacles. For instance, the rules may specify the maximum effective spray range of fire suppression equipment, requiring that the equipment be deployed within a certain range of the fire source and should not overlap excessively with other equipment to ensure maximum coverage.

[0137] S3. Construct a risk level assessment model and use the risk level assessment model to conduct risk assessments on safety hazard behaviors. Based on the risk assessment results, generate a multi-level early warning mechanism for coal bunkers in thermal power plants.

[0138] It should be noted that the key factors affecting the risk level are identified, and a risk level assessment model is built using machine learning methods (such as decision trees, support vector machines, or neural networks). This model can quantitatively assess different types of safety hazard behaviors.

[0139] The risk level of each safety hazard behavior is calculated using a risk level assessment model, typically employing a quantitative risk scoring system such as low, medium, and high risk levels. Based on the risk assessment results, different warning thresholds are defined, classifying the risk level into multiple levels (e.g., low risk, medium risk, high risk, and extremely high risk), and different response measures are set.

[0140] For example, for low-risk levels, the system may only provide warning information; for high-risk levels, it will activate emergency response procedures; based on the risk assessment results at each level, a multi-level early warning mechanism is generated to ensure that corresponding safety measures can be taken under different risk levels, and to automatically trigger operations such as alarms, personnel evacuation or equipment shutdown to ensure the safe operation of the coal bunker.

[0141] According to another embodiment of the invention, such as Figure 3 As shown, an AI-based intelligent safety monitoring system for coal bunkers in thermal power plants is also provided. This system includes:

[0142] The work area division module 1 is used to construct a three-dimensional visualization model of the thermal power plant, and divide the thermal power plant into a coal bunker work area and a coal bunker non-work area based on the three-dimensional visualization model;

[0143] The hazard behavior identification module 2 is used to collect video data of the coal bunker operation area, preprocess the video data, and use anomaly identification technology to analyze the preprocessed video data frame by frame to identify safety hazard behaviors in the coal bunker operation area. Safety hazard behaviors include not wearing a safety helmet correctly, not wearing work clothes correctly, and not deploying fire extinguishing equipment correctly.

[0144] The early warning mechanism construction module 3 is used to build a risk level assessment model and use the risk level assessment model to conduct risk assessment on safety hazard behaviors, and generate a multi-level early warning mechanism for coal bunkers in thermal power plants based on the risk assessment results.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based intelligent safety monitoring method for coal bunkers in thermal power plants, characterized in that, include: Construct a 3D visualization model of the thermal power plant, and divide the thermal power plant into a coal bunker operating area and a coal bunker non-operating area based on the 3D visualization model; Video data of the coal bunker operation area is collected using pre-deployed camera equipment. The initial target image sequence is extracted from the video data and defogging is performed on the initial target image sequence to obtain the target image sequence, including the image sequence of the workers and the image sequence of the work environment. Threshold segmentation is performed on the images of workers in the coal bunker image sequence to separate the safety helmet region, and the outer edge contour of the safety helmet region is located using a boundary tracking algorithm; Arbitrarily select a contour pixel on the outer edge contour as the contour starting point, iterate and track point by point from the contour starting point according to the preset position until returning to the contour starting point, form the safety helmet contour feature, and store the safety helmet contour points in the coordinate array; The helmet contour features in consecutive frame images are matched with the standard contour in the vertical direction to generate vertical error curves for consecutive frames. The maximum offset of the vertical error curve is extracted and compared with a preset threshold. If the vertical error curves of consecutive frames tend to be stable and the maximum offset is less than or equal to the preset threshold, it means that the current worker does not have a safety helmet offset. If the vertical error curve of consecutive frames tends to jitter and the maximum offset is greater than the preset threshold, it indicates that the safety helmet of the current worker is offset. The safety helmet contour points are converted into safety helmet point cloud data. The geometric topology reconstruction model is used to reconstruct the connection structure of the contact area between the chin strap and the helmet shell. The behavior of not wearing the safety helmet correctly is identified based on the connection structure. Optical flow field technology is used to extract the flow points of work clothes in the image sequence of workers, and the behavior of not wearing work clothes correctly is identified based on the flow points of work clothes. Extract smoke features from the image sequence of the working environment, locate the flammable source of the coal bunker based on the smoke features, calculate the distance from the flammable source of the coal bunker to the fire extinguishing equipment, and identify the behavior of incorrectly deploying fire extinguishing equipment; A risk level assessment model is constructed, and the model is used to assess the risks of potential safety hazards. Based on the risk assessment results, a multi-level early warning mechanism for coal bunkers in thermal power plants is generated.

2. The intelligent safety monitoring method for coal bunkers in thermal power plants based on AI according to claim 1, characterized in that, The process of converting helmet contour points into helmet point cloud data, reconstructing the connectivity structure of the chin strap and helmet shell contact area using a geometric topology reconstruction model, and identifying improper helmet wearing behavior based on the connectivity structure includes: Extract the helmet contour points from the coordinate array and convert them into helmet point cloud data. Use the helmet point cloud data as input and perform topological reconstruction on the connectivity structure of the contact area between the chin strap and the helmet shell using geometric topological reconstruction technology. Topological analysis was performed on the connectivity structure of the contact area between the mandibular band and the cap obtained by topological reconstruction to obtain the connectivity characteristics of the contact area between the mandibular band and the cap. The zero-order connectivity property and the first-order connectivity property are compared with preset thresholds respectively. Based on the comparison results, the connection and buckle engagement status of the chin strap of the safety helmet are determined. Combined with the maximum offset of the corresponding vertical error curve, the behavior of not wearing the safety helmet correctly is comprehensively identified.

3. The intelligent safety monitoring method for coal bunkers in thermal power plants based on AI according to claim 2, characterized in that, The connectivity properties include zero-order connectivity properties and first-order connectivity properties; Among them, the zero-order connectivity property is used to represent the number of branches in a connected structure and to determine whether there is a break in the mandibular band. First-order connectivity is used to represent the number of closed voids in a connected structure and to determine whether a snap-fit ​​part forms a complete ring-shaped connection structure.

4. The intelligent safety monitoring method for coal bunkers in thermal power plants based on AI according to claim 3, characterized in that, The step of using point cloud data of the safety helmet as input and performing topological reconstruction of the connectivity structure of the contact area between the chin strap and the helmet shell through geometric topological reconstruction technology includes: Extract the contour points of the bottom edge contour of the safety helmet, perform the initial construction operation, and generate the initial connected structure of the chin strap and the helmet shell contact area containing only the upper and lower basic surfaces. Set the reconstruction parameters of the initial connected structure. Under the constraints of the reconstruction parameters, extract the triangular patch sequence between adjacent edge contours layer by layer according to the hierarchical order in the safety helmet point cloud data and perform triangular patch construction to fill the point cloud region in the initial connected structure. Check if the vertices in the triangle construction already exist in the generated initial connected structure. If the vertices already exist, close the triangle; if the vertices do not exist, add vertices and then perform the face closure operation. Repeat the triangular piece construction process until the contact structure between the chin strap and the cap shell is reconstructed into a three-dimensional geometric topological model.

5. The intelligent safety monitoring method for coal bunkers in thermal power plants based on AI according to claim 3, characterized in that, The process of extracting smoke features from the image sequence of the working environment, locating the flammable source of the coal bunker based on the smoke features, calculating the distance from the flammable source of the coal bunker to the fire extinguishing equipment, and identifying the behavior of improperly deployed fire extinguishing equipment includes: A graph neural network based on sparse attention mechanism extracts the smoke focus field in the image sequence of the work environment and reverses the smoke diffusion source. The smoke diffusion source is superimposed with historical high-risk occurrence points to locate the flammable source of coal bunker in the work environment. Calculate the spatial distance between the flammable source of the coal bunker and the current deployment location of the fire extinguishing equipment, assess the response range of the fire extinguishing equipment, and identify potential blind spot locations based on the assessment results of the response range; The flammable source points of the coal bunker, the current deployment location of the fire extinguishing equipment, and the blind spot location are encoded and input into an event reasoning model based on fire extinguishing accessibility rules to infer the behavior of incorrectly deployed fire extinguishing equipment.

6. The intelligent safety monitoring method for coal bunkers in thermal power plants based on AI according to claim 5, characterized in that, The process of inputting the flammable source points of the coal bunker, the current deployment location of the fire extinguishing equipment, and the blind spot location codes into an event reasoning model based on fire extinguishing accessibility rules, and reasoning out the behavior of incorrectly deployed fire extinguishing equipment, includes: The flammable sources of coal bunkers, the current deployment locations of fire extinguishing equipment, and blind spot locations are encoded as sub-nodes to construct a multi-dimensional semantic graph. Based on the semantic relationships between child nodes in a multidimensional semantic graph, the meta-path learning algorithm is used to analyze the association strength between child nodes. Based on the predefined fire extinguishing accessibility rules and the association strength between child nodes, the deployment behavior of fire extinguishing equipment is determined using a fuzzy inference model to obtain the behavior probability distribution. Based on the behavior probability distribution, the behavior type of incorrectly deployed fire extinguishing equipment is identified.

7. The intelligent safety monitoring method for coal bunkers in thermal power plants based on AI according to claim 6, characterized in that, The analysis of the association strength between child nodes using the meta-path learning algorithm includes: Obtain the set of paths from any child node to the other child nodes in the multidimensional semantic graph, and use any selected child node as the starting point to propagate the access probability layer by layer down along each path; The intermediate child nodes visited during path propagation are integrated into the candidate node set. It is determined whether each intermediate child node in the candidate node set has a non-zero access probability in both paths. If so, the current intermediate child node is constructed with association features to represent the association strength between the child node and the target child node. If not, the current intermediate child node is marked as an irrelevant node. Tracing back the complete path of intermediate child nodes that satisfy the association characteristics in the path propagation, abstracting the complete path into a meta-path, and counting the frequency of the meta-path in the path set, selecting meta-paths within a preset frequency range to construct a meta-path pattern set.

8. An AI-based intelligent safety monitoring system for coal bunkers in thermal power plants, used to implement the AI-based intelligent safety monitoring method for coal bunkers in thermal power plants as described in any one of claims 1-7, characterized in that, The system includes: The work area division module is used to construct a three-dimensional visualization model of the thermal power plant, and divide the thermal power plant into coal bunker work areas and coal bunker non-work areas based on the three-dimensional visualization model; The hazard behavior identification module is used to collect video data from the coal bunker operation area, preprocess the video data, and use anomaly recognition technology to analyze the preprocessed video data frame by frame to identify safety hazard behaviors in the coal bunker operation area. The safety hazard behaviors include not wearing a safety helmet correctly, not wearing work clothes correctly, and not deploying fire extinguishing equipment correctly. The early warning mechanism construction module is used to build a risk level assessment model and use the risk level assessment model to conduct risk assessment on safety hazard behaviors. Based on the risk assessment results, a multi-level early warning mechanism for coal bunkers in thermal power plants is generated.

Images