Emergency fire-fighting monitoring method and system based on big data

Abstract

The invention discloses an emergency fire-fighting monitoring method and system based on big data, relates to the field of fire-fighting monitoring, and specifically comprises the following steps: collecting real-time multi-modal data of a fire scene, carrying out multi-scale processing, and extracting smoke, open fire and obstacle information from image data by using a target detection model; performing time sequence analysis on the continuous frames of smoke region images by using an LSTM network, and predicting the smoke diffusion degree; extracting flame sound information from the sound data by using a sound detection model; according to the method, regional fire behavior information is obtained through fusion of various kinds of information, the danger coefficient of each region is calculated, dynamic planning of an evacuation path is carried out based on the regional danger coefficient by using a genetic algorithm, a weight dynamic adjustment mechanism is designed, and a path planning scheme is flexibly adjusted according to different needs according to actual conditions; introducing a Dijkstra algorithm to dynamically adjust a local path according to the real-time information of the fire scene; the utilization efficiency of affected data and the flexibility and reliability of evacuation path planning in the case of a fire are improved.

Description

Technical Field

[0001] This invention relates to the field of fire monitoring technology, specifically to an emergency fire monitoring method and system based on big data. Background Technology

[0002] In existing emergency fire monitoring methods, the multimodal data collected at the fire scene is often affected when a fire occurs. For example, the collected image data is affected by smoke, and the collected sound data is affected by burning objects and noise from frightened people. It is essential to determine a safe evacuation route at the fire scene based on the affected multimodal data. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by proposing an emergency fire monitoring method and system based on big data.

[0004] The technical solution to achieve the objective of this invention is as follows: The emergency fire monitoring method and system based on big data includes the following steps: Real-time data from the fire scene is collected, including image and audio data; the image data consists of regional fire scene images captured by the camera. Among them, the sound data consists of real-time environmental sounds at the fire scene collected by sensors. ; Using target detection models to analyze regional fire images Perform smoke detection, open flame detection, and obstacle detection; generate area fire scene images. Input the target detection model to perform smoke detection and obtain the smoke detection results. ,in This represents the probability that the bounding box of the smoke is identified as smoke. The number of pixels within the smoke bounding box represents the size of the smoke region. To reduce the impact of dense smoke from the fire on the quality of the acquired images, an image dehazing model is used to process the regional fire images. The defogging process yields the defogging fire scene image. Images of the fire scene after the fog was cleared The input is fed into the target detection model to perform open flame detection and obtain the open flame detection result. ,in This represents the probability that the bounding box of an open flame is identified as an open flame. The number of pixels within the open flame bounding box represents the size of the open flame region; obstacle detection is then performed to obtain the obstacle detection results. ,in This represents the probability that the bounding box of an obstacle is identified as an obstacle. This indicates the number of pixels within the bounding box of an obstacle, i.e., the size of the obstacle. A Long Short-Term Memory (LSTM) network was used to analyze the smoke diffusion level of a continuously captured set of regional fire images; firstly, based on the smoke detection results... Extracting from continuous area fire images Images of smoke-affected areas were generated and labeled using optical flow. These smoke-affected areas were then input into an LSTM network to predict the smoke diffusion level in the next fire scene image. ,in Indicates the direction of smoke diffusion. Indicates the distance the smoke has spread; Short-time Fourier transform is used to extract real-time ambient sound from the fire scene. The features were analyzed and sound was classified using CNN to extract flame sound information; the original audio data was truncated using a Hamming window function and subjected to Fourier transform, and the complete sound spectrum signal was obtained by continuously sliding the window. The sound spectrum signal is input into a CNN for classification processing to obtain the flame sound information. ,in This represents the probability of identifying the sound as a flame. Indicates the frequency of the flame sound; The smoke detection results, open flame detection results, and flame sound information are fused to generate a real-time fire intensity index for the current area. The real-time fire intensity index for a given area can be adjusted based on the size of the smoke-affected area. Flame area size and the frequency of the flame sound Weighted fusion is performed for calculation; to promote the fusion of three different modalities, the multimodal information is normalized, and the three normalized values ​​are then weighted and fused to obtain the real-time fire index for the current area. ; Based on the degree of smoke diffusion Analyze the possibility of the fire in the current area escalating. According to the distance of smoke diffusion The ratio of the image size to the smoke area size is used to calculate the likelihood of the fire in the current area increasing. ; The data analysis module is based on the real-time fire intensity index of the area. The possibility of the fire intensifying in the area and obstacle detection results The risk coefficient of the corresponding node in the current area is calculated comprehensively based on the obstacle detection results. Calculating obstacle size Images of the fire zone Size ratio ,when When the value exceeds the threshold, the node cannot pass, and its corresponding node code is set to -1; when the value exceeds the threshold, the node cannot pass. When the fire intensity is below the threshold, further adjustments are made based on the real-time fire intensity index of the area. The possibility of the fire intensifying in the area Calculate the node risk factor for the current region. ; After calculating the node hazard coefficients for each region, dynamic evacuation route planning is performed based on a genetic algorithm; a combination of global route planning and local dynamic planning strategies is used to address changes in the fire intensity in the fire area.

[0005] Furthermore, this application also proposes an emergency fire monitoring system based on big data, including: a data acquisition module, a target detection module, a diffusion detection module, a sound detection module, a data analysis module, and a path planning module; the data acquisition module collects environmental data of the fire scene, including image data and sound data; the target detection module performs smoke detection on the image data using a target detection model, and performs open flame detection and obstacle detection on the defogging image data processed by the image defogging model; the diffusion detection module uses an LSTM network to analyze the degree of smoke diffusion; the sound detection module detects flame sound using short-time Fourier transform and CNN; the data analysis module analyzes the fire situation in various areas of the fire scene based on the results of the target detection module, diffusion detection module, and sound detection module; the path planning module performs global path planning using a genetic algorithm and dynamically adjusts local paths using Dijkstra's algorithm based on real-time fire scene data; Compared with existing technologies, the beneficial effects of this invention are as follows: This invention performs multi-scale processing on real-time multimodal data of the fire scene, extracts smoke, open flame, and obstacle information from image data using a target detection model, performs temporal analysis on continuous frame smoke area images using an LSTM network to predict the degree of smoke diffusion, extracts flame sound information from sound data using a sound detection model, fuses multiple information sources to obtain regional fire information to calculate the hazard coefficient of each region, and uses a genetic algorithm to dynamically plan evacuation routes based on the regional hazard coefficient. A dynamic weight adjustment mechanism is designed to flexibly adjust the route planning scheme according to different needs based on the actual situation. The Dijkstra algorithm is introduced to dynamically adjust local routes based on real-time fire information. This improves the efficiency of utilizing affected data and the flexibility and reliability of evacuation route planning during a fire. Attached Figure Description

[0006] Figure 1 This is a flowchart of an emergency fire monitoring method based on big data, as described in an embodiment of this application. Figure 2This is a flowchart illustrating the smoke detection, open flame detection, and obstacle detection processes of the target detection model in an embodiment of this application. Figure 3 The flowchart of dynamic path planning based on genetic algorithm in the embodiments of this application is shown. Figure 4 This is a block diagram of an emergency fire monitoring system based on big data, as described in an embodiment of this application. Detailed Implementation

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0008] like Figure 1 As shown in the figure, a specific embodiment of the present invention discloses an emergency fire monitoring method based on big data, comprising the following steps: Collect real-time multimodal data from the fire scene, including image and sound data; The image data was processed using a target detection model to detect smoke, open flame, and obstacles, respectively, and the smoke detection results, open flame detection results, and obstacle detection results were obtained. The degree of smoke diffusion is obtained by analyzing the image data using a diffusion analysis model; The sound data is detected using a sound detection model to obtain sound detection results; The fire intensity index of the current area is calculated based on the smoke detection results, open flame detection results, and sound detection results. The possibility of the fire in the area increasing is analyzed using the degree of smoke diffusion. The danger coefficient of the corresponding node in the current area is calculated based on the fire intensity index of the current area, the possibility of the fire in the area increasing, and the obstacle detection results. A genetic algorithm is used to dynamically plan evacuation routes based on the hazard coefficient of the nodes. The combination of global path planning and local dynamic planning strategies is used to deal with changes in the fire intensity in the fire area.

[0009] This embodiment provides an emergency fire monitoring method based on big data. It performs multi-scale processing on real-time multimodal fire data, extracts smoke, open flame, and obstacle information from image data using a target detection model, performs temporal analysis on consecutive frames of smoke area images using an LSTM network to predict the degree of smoke diffusion, extracts flame sound information from sound data using a sound detection model, fuses multiple information sources to obtain regional fire information, calculates the hazard coefficient of each region, and uses a genetic algorithm to dynamically plan evacuation routes based on the regional hazard coefficient. A dynamic weight adjustment mechanism is designed to flexibly adjust the route planning scheme according to different needs based on actual conditions. The Dijkstra algorithm is introduced to dynamically adjust local routes based on real-time fire information. This improves the efficiency of utilizing affected data during a fire and enhances the flexibility and reliability of evacuation route planning.

[0010] In some preferred embodiments, real-time multimodal fire data is acquired and processed; the acquired real-time multimodal fire data includes image data and audio data, wherein the image data is regional fire scene images. The sound data is the real-time environmental sound of the fire scene. ; Calculate the regional fire scene image Image size is the number of pixels in the image. .

[0011] like Figure 2 As shown, a target detection model is used to analyze regional fire images. Smoke detection, open flame detection, and obstacle detection are performed; image data collected at a fire scene can be affected by dense smoke, so directly analyzing regional fire scene images... When performing open flame and obstacle detection, the dense smoke from the fire in the image can affect the accuracy of these detections, impacting subsequent fire evacuation route planning. Therefore, it is necessary to first analyze the regional fire scene image. Smoke detection was performed, and smoke detection results were obtained. ,in This represents the probability of identifying it as smoke. The number of pixels within the smoke bounding box represents the size of the smoke region; subsequently, the regional fire image... Defogging was performed to obtain a defogging fire scene image. Then, examine the defogging fire scene images. Open flame detection and obstacle detection are performed, and the results of open flame detection and obstacle detection are obtained; among them, the open flame detection result is... ,in This represents the probability that the bounding box of an open flame is identified as an open flame. The number of pixels within the open flame bounding box indicates the size of the open flame area; obstacle detection results are... ,in This represents the probability that the bounding box of an obstacle is identified as an obstacle. This indicates the number of pixels within the obstacle's bounding box, which represents the obstacle's size.

[0012] Using target detection models to analyze regional fire images Smoke detection results were obtained. .

[0013] In some preferred embodiments, the YOLOv8 model is selected as the target detection model for smoke detection; specifically, the regional fire scene image is first standardized to convert the regional fire scene image into a standard format. Dividing the pixel values ​​in the image by 255 yields a standardized area fire image. Then, standardized regional fire scene images were generated. The data is input into a YOLOv8 model for prediction; the YOLOv8 model extracts standardized regional fire images through a backbone network. The YOLOv8 model transforms images into feature representations. This feature extraction process consists of a convolutional module, a C2F module, and an SPPF module. The convolutional module performs initial feature extraction, followed by further feature extraction and gradient information acquisition through the C2F module. The SPPF module pools the features to capture receptive field information of different sizes, improving the object detection model's ability to identify objects of different sizes. Smoke features are then acquired through this backbone network. After acquiring the smoke features, to enhance their representational power, the YOLOv8 model performs bidirectional feature aggregation through intermediate layers to obtain fused smoke features. These fused smoke features are then input into the YOLOv8 model's detection head for smoke detection, outputting the smoke detection results. ,in, This represents the probability that the bounding box of the smoke is identified as smoke. This indicates the number of pixels within the smoke bounding box, i.e., the size of the smoke region.

[0014] Image dehazing model for regional fire images Defogging is performed to obtain a defogging fire scene image. .

[0015] In some preferred embodiments, the FFA-Net model is used as the image dehazing model. The FFA-Net model incorporates a feature attention module that combines channel attention and pixel attention mechanisms, enabling it to focus more on smoke pixel information during dehazing, resulting in a clearer dehazed image. Dehazed fire scene images are obtained using the FFA-Net model. Then, images of the fire scene will be removed from the fog. The data is input into the target detection model for open flame detection and obstacle detection.

[0016] Target detection model is used to analyze defogging fire images. Open flame detection and obstacle detection were performed to obtain open flame detection results. and obstacle detection results .

[0017] In some preferred embodiments, the YOLOv8 model is selected as the target detection model for open flame detection and obstacle detection; similar to the smoke detection described above, the defogging fire scene image is... After inputting into the YOLOv8 model, the open flame detection results are obtained. and obstacle detection results ,in, This represents the probability that the bounding box of an open flame is identified as an open flame. This indicates the number of pixels within the open flame bounding box, which represents the size of the open flame area. This represents the probability that the bounding box of an obstacle is identified as an obstacle. This indicates the number of pixels within the obstacle's bounding box, which represents the obstacle's size.

[0018] The degree of smoke diffusion was obtained by analyzing the image data using a diffusion analysis model.

[0019] In some preferred embodiments, a Long Short-Term Memory (LSTM) network is used to analyze the smoke diffusion level of a continuously captured set of regional fire scene images; based on the smoke detection results... Zhang's continuous area fire images Extract Let there be a set of smoke region images, denoted as [image set]. The optical flow method was used to annotate the smoke region image, and the smoke region set was input into an LSTM to predict the smoke diffusion level of the fire area in the next moment. Specifically, based on smoke detection results... Zhang's continuous area fire images The location of the smoke will be continuously located and will The smoke location is from This was extracted from images of the fire scene in the Zhang area. The set of smoke region images is denoted as _____, which reduces the interference of the background on the prediction of the smoke diffusion degree. ;Calculated based on optical flow method Zhang's smoke area image The direction and distance of smoke diffusion are shown in the image of the smoke area. , For example, calculate the image of the smoke area. Each pixel in the image is relative to its position in the smoke area. The displacement vectors on the image are then used to calculate the average direction and average distance of all unique vectors to form the smoke region image. Labels indicating the direction and distance of smoke spread; images of continuous fire zones. The data is input into an LSTM, and the smoke diffusion direction and distance labels from each smoke region image calculated by the optical flow method are used to optimize the LSTM, ultimately yielding the predicted smoke diffusion level for the next moment. ,in Indicates the direction of smoke diffusion. Indicates the distance the smoke has spread.

[0020] The sound data is detected using a sound detection model to obtain sound detection results.

[0021] In some preferred embodiments, short-time Fourier transform is used to extract sound features and CNN is used for sound classification to extract flame sound information. Since fire sound data is not a stationary signal, it is not suitable to directly use Fourier transform to process the original fire sound data. This application uses a time-frequency analysis method to process the fire sound data. Short-time Fourier transform divides the signal into multiple time windows of equal length. Within a single time window, the signal can be considered a stationary signal. Therefore, Fourier transform can be used to calculate the spectral characteristics of the fire sound data within a single time window, and then the window is continuously slid to obtain the spectral characteristics of the complete original fire sound data. Specifically, a Hamming window is used as the time window to truncate the original fire sound data to effectively suppress spectral leakage. After short-time Fourier transform processing of the original fire sound data, the fire sound spectrum signal is obtained. The sound spectrum signal of the fire scene The input is fed into a CNN network for binary classification to obtain the voiceprint detection results. ,in This indicates the probability of detecting the sound of flames in that area. This indicates the frequency of the sound of the flame.

[0022] The fire intensity index of the current area is calculated based on the smoke detection results, open flame detection results, and sound detection results. The possibility of the fire in the area increasing is analyzed using the degree of smoke diffusion. The danger coefficient of the corresponding node in the current area is calculated based on the fire intensity index of the current area, the possibility of the fire in the area increasing, and the obstacle detection results.

[0023] Specifically, the multimodal information smoke detection results Open flame test results Voiceprint detection results The data is then integrated, and a weighted summation method is used to calculate the real-time fire intensity index for the current area. Smoke detection results middle Indicates the size of the smoke area and the results of open flame detection. middle This indicates the size of the open flame area and the results of the voiceprint detection. middle These three values, representing the frequency of the flames, can all reflect the real-time fire intensity in the current area. A weighted sum of these three values ​​can be used to calculate the real-time fire intensity index for the current area. First, the data is standardized, including the size of the smoke area. Size of the open flame area Each represents the number of pixels within the bounding box, and the size of the smoke region can be adjusted based on the size of the input image. Size of the open flame area The standardization process is performed using the following formula:

[0024]

[0025] Calculated This indicates the extent to which the smoke is affecting the real-time fire situation in the current area. This indicates the degree of impact of the open flame zone on the real-time fire situation in the current area. and The values ​​range from 0 to 100; for the frequency of flame sound Logarithmic processing is performed, and the formula is as follows:

[0026] Calculated This indicates the degree of impact of the sound of flames on the real-time fire situation in the current area; subsequently, it indicates... , and Perform weighted summation:

[0027] Among them, the probability of classifying a smoke bounding box as smoke. As The weights determine the probability of classifying an open flame bounding box as an open flame. As The weight of the probability that a fire sound exists in that area. As Weighting; this takes into account the hazards of fire, even , , The values ​​are all low, indicating that the probability of smoke, open flames, and fire sounds is low. However, this cannot be completely ignored. Therefore, it is included as a weight in the calculation of the current area fire index, which can also serve as an alert when the probability of smoke, open flames, and fire sounds is low.

[0028] Based on the degree of smoke diffusion Analyze the possibility of the fire in the current area escalating. Smoke diffusion degree In Indicates the direction of smoke diffusion. This indicates the distance smoke has traveled; the smoke spread distance represents the distance smoke has moved between two adjacent images of a fire area, which can reflect the trend of fire development; the smoke spread distance can be used to analyze the likelihood of the fire in the current area intensifying. The fire in the current area is likely to intensify. The calculation method is as follows:

[0029] in, This represents the square root operation. This represents the side length of the regional fire scene image. Since the input images of the target recognition model are all square, the side length of the image can be obtained by taking the square root of the number of image pixels.

[0030] According to the regional real-time fire intensity index The possibility of the fire intensifying in the area and obstacle detection results The risk coefficient of the corresponding node in the current area is calculated comprehensively based on the obstacle detection results. Calculating obstacle size Images of the fire zone Size ratio The calculation formula is as follows:

[0031] when When the value is greater than the threshold, it indicates that an obstacle at that node is affecting passage, and the node is impassable; the corresponding node danger coefficient is set to -1. When the fire intensity is below the threshold, further adjustments are made based on the real-time fire intensity index of the area. The possibility of the fire intensifying in the area and obstacle ratio Calculate the node risk factor for the current region. The calculation method is as follows:

[0032] like Figure 3 As shown, a genetic algorithm is used to plan dynamic evacuation routes based on the hazard coefficient of the nodes. The combination of global path planning and local dynamic planning strategies is used to deal with changes in the fire intensity in the fire area.

[0033] Calculated according to the above formula Each region corresponds to Risk level of each node Then, dynamic evacuation route planning is performed based on genetic algorithms; two strategies, global route planning and local dynamic planning, are combined to deal with changes in the fire intensity in the fire area; global route planning takes the weighted sum of route length and hazard coefficient as the optimization objective and can dynamically adjust the weights according to different situations; after global route planning, local route optimization is performed in combination with real-time fire information.

[0034] Evacuation paths are represented using a node sequence encoding method, with the starting point of the evacuation path denoted as 0 and the exit as 0. Code 1 to The intermediate node is as described above. Each region encodes the path from the origin to the exit as a chromosome; for example, a chromosome... This represents a path that starts from point 0, passes through nodes 2 and 7 in sequence, and reaches the safe exit. Evacuation routes.

[0035] The initial population, or initial path, can be generated using a breadth-first search approach, or a heuristic method can be employed, prioritizing nodes with lower risk coefficients to form the initial path. When generating the initial population using breadth-first search, a risk coefficient check for nodes needs to be added. If the risk coefficient of a node is -1, it indicates that there is an obstacle affecting passage at that node, and the node should be bypassed. The population size... This will directly affect the optimization efficiency and results of the genetic algorithm, so it is necessary to determine a reasonable population size based on the complexity of the fire area.

[0036] After generating the initial population, a fitness function for calculating individual fitness is designed, taking into account both path length and node risk coefficient. A cost function-based approach is adopted, defining the cost function as a weighted sum of path length and node risk coefficient. The cost function is defined as follows:

[0037] in, Indicates the path length. This represents the sum of the danger coefficients of all nodes traversed. , They represent and Adaptive weights, , The sum of these values ​​is 1, and this weight is dynamically adjusted based on the different objectives of the path planning. Different evacuation path planning strategies are selected based on the overall fire intensity of the area. When the overall fire intensity of the area is relatively low, the evacuation path planning should choose a rapid escape strategy, prioritizing the shortest evacuation path and increasing the weight of path length in the cost function. At this time, the path length weight Weight greater than the node risk coefficient Furthermore, the difference between the two values ​​should be greater than or equal to 0.4. When the fire begins to spread throughout the area, evacuation route planning should adopt a balanced strategy, balancing the weights of route length and node hazard coefficients in the cost function, and reducing the weight of route length. And increase the weight of node risk coefficient. At this time, the path length weight and node risk coefficient weight The difference between the two is less than 0.2; when the overall fire intensity is high, evacuation route planning should prioritize safety, completely avoiding dangerous areas, and further increasing the weight of the node hazard coefficient in the cost function. Weight greater than path length And the difference between the two is greater than or equal to 0.4.

[0038] After determining the cost function, the fitness function is defined as the reciprocal of the cost function, resulting in a negative correlation between the cost function and the fitness function. That is, the lower the cost of a path, the higher its fitness. The fitness function is defined as follows:

[0039] in, It is a small constant to prevent division by zero.

[0040] The initial population is processed based on genetic operations, including three steps: selection, crossover, and mutation. The selection operator is the tournament selection operator, which randomly selects from the population. Individuals, compared to this The fitness of each individual is used to select the one with the highest fitness as the parent. The tournament selection operator is simple to operate and can improve the optimization speed of global path planning. When selecting the crossover operator, it is necessary to consider that there should be no duplicate nodes in the path during evacuation path planning. Therefore, a partial crossover mapping operator is selected to recombine the parent genes. The mutation operator introduces new information by randomly changing some nodes in the individual. In path planning, the node replacement mutation operator is used. A node that is not the start or end point is selected in the individual and replaced with a node that is not present in the individual's adjacent nodes. The population is iterated continuously according to the above steps until convergence, and the individual with the highest fitness is output as the global evacuation path.

[0041] The planning process of the above-mentioned global evacuation route is based on the information of each node collected in the early stage of the fire. Considering that the fire environment will change over time, the evacuation route planning process needs to take into account that when people reach a certain node of the global evacuation route, the danger coefficient of the next node and adjacent nodes has changed. At this time, the route needs to be dynamically adjusted locally.

[0042] The system calculates the risk coefficient of each node in real time and the change in risk coefficient between the node and the last monitoring time. Nodes whose risk coefficient increases by more than a preset threshold are identified as dangerous nodes, and the path formed by the node and its surrounding nodes is identified as a dangerous path. Nodes whose risk coefficient decreases by more than a preset threshold are identified as safe nodes, and the path formed by the node and its adjacent nodes is identified as a safe path.

[0043] When a node in the global evacuation path is identified as a dangerous node, the dangerous node and its surrounding nodes forming a dangerous path are excluded from the path planning. Dijkstra's algorithm is used, with the node preceding it in the global evacuation path as the starting point and the node following it as the ending point, to search for the optimal alternative path in the area. When analyzing dangerous paths, since there is a possibility of further fire spread, dangerous nodes not only affect their adjacent nodes but also the adjacent nodes of those adjacent nodes. Therefore, the concept of a dangerous spread radius is introduced. Based on the change in the dangerous node's node hazard coefficient, the weighted hazard coefficient change of the surrounding nodes under its influence is calculated by distance, thereby determining the dangerous path. Specifically, the increase in the dangerous node's node hazard coefficient is denoted as... The increase in the node danger coefficient of a node surrounding a dangerous node is... The weighted hazard coefficient change of the node under the influence of the dangerous node is calculated according to the exponential decay formula. The calculation formula is as follows:

[0044] in, This indicates the magnitude of change in the weighted risk coefficient of that node. Indicates the distance between the node and the dangerous node; when If the value exceeds a preset threshold, the node will be included in the dangerous path corresponding to the dangerous node.

[0045] When an adjacent node of a node in the global evacuation path is identified as a safe node, Dijkstra's algorithm is used to search for the optimal alternative path in the region, starting from the node before the node and ending at the node after the node.

[0046] Since Dijkstra's algorithm searches for the shortest path, to ensure that the search yields a regionally optimal alternative path rather than a regionally shortest alternative path, the danger coefficient of the endpoint of the directed edge is used as the weight of that edge. For example, the node... To the node The weight of the edge is set to the node. The risk coefficient is determined, and the optimal alternative path for the region is searched using Dijkstra's algorithm. The optimal alternative path replaces the original path, completing the local adjustment of the evacuation path. When replacing the original path with the optimal alternative path, the impact of path switching on the stability of the global path needs to be further considered. A stability judgment mechanism is introduced to avoid evacuation chaos caused by frequent path adjustments. Specifically, local adjustments to evacuation paths made to ensure safety due to the increase of dangerous nodes do not undergo stability judgment. Local path adjustments made due to the increase of safe nodes require stability judgment. After finding the optimal alternative path for the region using Dijkstra's algorithm, the cost function value of the optimal alternative path is calculated, and the reduction in cost function value compared to the original path is calculated. If the reduction is greater than or equal to 15%, the alternative path replaces the original path; if the reduction is less than 15%, the original path is kept.

[0047] like Figure 4 As shown, the big data-based emergency fire monitoring system includes: a data acquisition module, a target detection module, a diffusion detection module, a sound detection module, a data analysis module, and a path planning module. The data acquisition module collects environmental data from the fire scene, including image and sound data. The target detection module uses a target detection model to detect smoke in the image data and performs open flame and obstacle detection on the defogging image data processed by the image defogging model. The diffusion detection module uses an LSTM network to analyze the degree of smoke diffusion. The sound detection module uses short-time Fourier transform and CNN to detect flame sounds. The data analysis module analyzes the fire situation in various areas of the fire scene based on the results from the target detection, diffusion detection, and sound detection modules. The path planning module uses a genetic algorithm for global path planning and uses Dijkstra's algorithm to dynamically adjust local paths based on real-time fire scene data.

[0048] Specifically, the data acquisition module collects real-time environmental data of the fire scene through cameras and sound sensors, including regional fire scene image data and real-time environmental sound data of the fire scene.

[0049] Specifically, the target detection module first performs smoke detection on the regional fire scene image using the target detection model to obtain smoke detection results. Then, to improve the accuracy of open flame detection and obstacle detection, it uses an image defogging model to perform defogging operation on the regional fire scene image to obtain a defogging fire scene image. Finally, it uses the target detection model to perform open flame detection and obstacle detection on the defogging fire scene image to obtain open flame detection results and obstacle detection results.

[0050] Specifically, the diffusion detection module extracts images of continuous smoke regions based on smoke detection results, and uses optical flow to calculate the smoke diffusion direction and distance to annotate the smoke region images. Then, it uses an LSTM network to predict the smoke diffusion distance at the next moment.

[0051] Specifically, the sound detection module extracts sound features from real-time environmental sound data of the fire scene through short-time Fourier transform and uses CNN to detect flame sounds to obtain sound detection results.

[0052] Specifically, the data analysis module calculates the regional fire intensity index by weighting and summing the smoke detection results, open flame detection results, and sound detection results. Then, it calculates the probability of the regional fire intensifying based on the smoke diffusion distance. Finally, it uses the regional fire intensity index, the probability of the regional fire intensifying, and obstacle detection results to calculate the node hazard coefficient corresponding to the region.

[0053] Specifically, the path planning module uses a genetic algorithm to plan global paths based on the node hazard coefficients calculated by the data analysis module. It calculates the path cost based on the weighted sum of path length and node hazard coefficients, and dynamically adjusts the weights of path length and node hazard coefficients according to the actual situation to plan different paths. At the same time, it uses the Dijkstra algorithm to deal with real-time changes in the fire scene. When the node hazard coefficient changes due to the fire situation, the Dijkstra algorithm replans local routes after eliminating dangerous nodes and dangerous paths to replace the original planned routes, thus realizing the dynamic adjustment of local paths.

[0054] In this embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of an emergency fire monitoring method.

[0055] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented by different modules for different steps.

[0056] The aforementioned program can run on a processor or be stored in memory (or computer-readable medium). Computer-readable medium includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable medium does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0057] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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. A big data-based emergency fire monitoring method, characterized in that, Includes the following steps: Collect real-time multimodal data of the fire scene, including regional fire scene images and real-time environmental sound of the fire scene; The target detection model was used to perform smoke detection, open flame detection, and obstacle detection on the fire scene image of the area, and the smoke detection results, open flame detection results, and obstacle detection results were obtained respectively. The degree of smoke diffusion was obtained by analyzing the fire scene images in the area using a diffusion analysis model. The sound detection model was used to detect the real-time environmental sound at the fire scene, and the sound detection results were obtained. The fire intensity index of the current area is calculated based on the smoke detection results, open flame detection results, and sound detection results. The possibility of the fire in the area increasing is analyzed using the degree of smoke diffusion. The danger coefficient of the corresponding node in the current area is calculated based on the fire intensity index of the current area, the possibility of the fire in the area increasing, and the obstacle detection results. A genetic algorithm is used to dynamically plan evacuation routes based on the hazard coefficient of the nodes. The combination of global path planning and local dynamic planning strategies is used to deal with changes in the fire intensity in the fire area.

2. The emergency fire monitoring method based on big data as described in claim 1, characterized in that, The process involves calculating the current area fire intensity index based on smoke detection results, open flame detection results, and flame sound detection results; analyzing the likelihood of the fire intensifying using the degree of smoke diffusion; and calculating the danger coefficient of the corresponding node in the current area based on the current area fire intensity index, the likelihood of the fire intensifying, and obstacle detection results. This includes: The smoke detection result includes the size of the smoke area, the open flame detection result includes the size of the open flame area, and the flame sound detection result includes the frequency of the flame sound. The real-time fire intensity index for the current area is obtained by standardizing three values: smoke area size, open flame area size, and flame sound frequency, and then weighting and fusing them. ; The smoke diffusion level includes the smoke diffusion distance. A smoke area image is obtained based on the smoke detection results, and the ratio between the smoke diffusion distance and the size of the smoke area image is considered the likelihood that the fire in the current area will intensify. ; The obstacle detection results include the obstacle size, and the proportion of the obstacle size to the area of ​​the fire scene image is calculated. ,when When the value exceeds the threshold, the node cannot pass, and its corresponding node code is set to -1; when the value exceeds the threshold, the node cannot pass. When the fire intensity is below the threshold, the real-time regional fire intensity index is used. The possibility of the fire intensifying in the area and obstacle ratio Calculate the node risk factor for the current region. Node risk factor It is positively correlated with the three values ​​mentioned above.

3. The emergency fire monitoring method based on big data as described in claim 2, characterized in that, The dynamic evacuation route planning based on genetic algorithms combines global route planning and local dynamic planning strategies to address changes in fire intensity within the fire area, including: The initial population, i.e. the initial path, is generated based on the breadth-first search approach. When generating the initial population using the breadth-first search approach, a judgment on the danger coefficient of nodes is added. When the danger coefficient of a node is -1, it means that there is an obstacle at that node affecting passage, and the node should be bypassed. After generating the initial population, a cost function-based approach is adopted, whereby the cost function is defined as a weighted sum of path length and node risk coefficient. The weights of path length and node risk coefficient are dynamically adjusted according to different path planning objectives. The fitness function is defined as the reciprocal of the cost function, which results in a negative correlation between the cost function and the fitness function. The initial population is processed based on genetic operations, including three steps: selection, crossover, and mutation. A tournament selection operator is used. A partial crossover mapping operator is selected to recombine parent genes. A node replacement mutation operator is selected. The population is iterated continuously according to the above steps until convergence, and the individual with the highest fitness is output as the global evacuation path. The risk coefficient of each node is calculated based on the real-time risk coefficient, and dangerous and safe nodes are screened. Based on the elimination of dangerous nodes and the addition of safe nodes, the Dijkstra algorithm is used to search for the optimal alternative path in the region. The optimal alternative path in the region replaces the original path, and the local adjustment of the evacuation path is completed.

4. The emergency fire monitoring method based on big data as described in claim 3, characterized in that, The weights of the path length and node risk coefficient are dynamically adjusted according to different path planning objectives, including: Different evacuation route planning strategies are selected based on the overall fire situation in the area, with the sum of the route length weight and the node hazard coefficient weight being 1; When the overall fire intensity in the area is relatively low, the shortest evacuation route is selected first, and the weight of the route length in the cost function is increased. At this time, the weight of the route length is greater than the weight of the node hazard coefficient, and the difference between the two is greater than or equal to 0.

4. When the fire begins to spread throughout the area, the weight of path length in the cost function is reduced and the weight of node hazard coefficient is increased. At this time, the difference between the weight of path length and the weight of node hazard coefficient is less than 0.

2. When the overall fire intensity is high, the weight of the node hazard coefficient in the cost function is further increased. At this time, the weight of the node hazard coefficient is greater than the weight of the path length and the difference between the two is greater than or equal to 0.

4.

5. The emergency fire monitoring method based on big data as described in claim 4, characterized in that, The process of using Dijkstra's algorithm to find the optimal alternative path in the region, based on eliminating dangerous nodes and adding safe nodes, and replacing the original path with the optimal alternative path to complete the local adjustment of the evacuation route includes: When a node in the global evacuation path is identified as a dangerous node, the dangerous node and the dangerous path formed by it and its surrounding nodes are excluded from the path planning. The Dijkstra algorithm is used to search for the optimal alternative path in the region to replace the dangerous node, with the previous node in the global evacuation path as the starting point and the next node as the ending point. The local adjustment of the evacuation path is completed by replacing the original path with the optimal alternative path in the region. When an adjacent node of a node in the global evacuation path is identified as a safe node, Dijkstra's algorithm is used to search for the optimal alternative path in the region, starting from the node before the node and ending at the node after the node. The original path is then replaced by the optimal alternative path in the region, thus completing the local adjustment of the evacuation path.

6. The emergency fire monitoring method based on big data as described in claim 5, characterized in that, The method of calculating the change range of the risk coefficient of each node based on the real-time node risk coefficient, and screening dangerous nodes and safe nodes, is characterized by: The risk coefficient of each node is calculated in real time, and the change in risk coefficient between the node and the last monitoring time is calculated; nodes whose increase exceeds the preset threshold are identified as dangerous nodes. Nodes whose reduction exceeds a preset threshold are identified as safe nodes.

7. An emergency fire monitoring system based on big data, characterized in that, include: The data acquisition module collects real-time multimodal data from the fire scene, including image and sound data. The target detection module is used to perform smoke detection, open flame detection, and obstacle detection on the image data using a target detection model, and obtain smoke detection results, open flame detection results, and obstacle detection results respectively. The diffusion detection module is used to analyze the image data using a diffusion analysis model to obtain the degree of smoke diffusion; The sound detection module is used to detect the sound data using a sound detection model to obtain sound detection results; The data analysis module is used to calculate the current area fire index based on the smoke detection results, open flame detection results, and sound detection results; analyze the possibility of the fire in the area increasing based on the degree of smoke diffusion; and calculate the danger coefficient of the corresponding node in the current area based on the current area fire index, the possibility of the fire in the area increasing, and obstacle detection results. The path planning module is used to perform dynamic evacuation path planning based on the danger coefficient of the nodes using a genetic algorithm. It combines global path planning and local dynamic planning strategies to cope with changes in the fire intensity in the fire area.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images