A laser positioning system for electrical fire monitoring and emergency response

Abstract

The application belongs to the technical field of electrical fire, and discloses a laser positioning system for electrical fire monitoring and emergency response, comprising a laser radar device, a data processing unit, a communication module and a monitoring platform.The laser positioning system for electrical fire monitoring and emergency response can accurately detect and locate the fire source by emitting laser beams and receiving reflected signals by the laser radar device, and analyzing the reflected signals by the data processing unit; can distinguish smoke particles and water mist particles according to different depolarization ratios, thereby avoiding false alarms caused by environmental factors (such as water vapor and dust); in addition, the data processing unit also uses waveform decomposition method and frequency domain analysis method to extract weak smoke signals from background signals, further improving the accuracy of fire source identification; once the location of the fire is determined, the communication module will wirelessly transmit the information to the monitoring platform, ensuring rapid and accurate positioning of the fire source.

Description

Technical Field

[0001] This invention belongs to the field of electrical fire technology, specifically relating to a laser positioning system for electrical fire monitoring and emergency response. Background Technology

[0002] With the continuous expansion of power facilities and the increasing frequency of use of electrical equipment, the incidence of electrical fires is also rising year by year, posing a serious threat to people's lives and property.

[0003] Traditional electrical fire monitoring methods mainly include temperature sensors and smoke detectors. These methods can play an early warning role to some extent, but they have many shortcomings in practical applications:

[0004] Insufficient accuracy: Traditional detection methods struggle to accurately identify the specific location of a fire source, especially in large and complex environments;

[0005] High false alarm rate: Due to environmental factors (such as water vapor and dust), traditional detectors are prone to false alarms, which affects the reliability of the system and the efficiency of emergency response.

[0006] Slow response time: The time between the discovery of a fire and the implementation of effective measures is relatively long, which is not conducive to timely control of the fire's spread;

[0007] To address the above problems, this invention proposes a laser positioning system for electrical fire monitoring and emergency response. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a laser positioning system for electrical fire monitoring and emergency response. It has advantages such as improving the accuracy of fire monitoring, reducing the false alarm rate, and accelerating the emergency response speed. It solves the problems of inaccurate positioning, susceptibility to environmental interference leading to frequent false alarms, and slow response in current electrical fire monitoring technologies.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] A laser positioning system for electrical fire monitoring and emergency response, the system comprising: a lidar device, a data processing unit, a communication module, a monitoring platform, a smoke diffusion prediction module, and an emergency decision generation module;

[0011] The lidar device is used to emit laser beams and receive reflected signals, and to detect and locate fire sources by analyzing the reflected signals;

[0012] The data processing unit is used to process the received reflected signals, identify smoke particles and water mist particles based on changes in signal intensity, and extract smoke signals from the background signals using waveform decomposition and frequency domain analysis to identify the fire source and determine the location of the fire.

[0013] The communication module is used to wirelessly transmit the location information of the fire to the monitoring platform.

[0014] The monitoring platform is used to locate the source of a fire based on its location.

[0015] The smoke diffusion prediction module is used to predict the diffusion range and speed of smoke at the fire scene based on the fire source, using the U-Net convolutional neural network architecture and SiLU activation function, combined with temperature, humidity, wind speed, CO2 concentration sensor data and building internal structure information.

[0016] The emergency decision-making module is used to extract key information from fire accident reports based on named entity recognition technology and Bayesian networks, and generate the optimal fire emergency response strategy in real time.

[0017] Preferably, the lidar device uses polarization lidar technology, which can distinguish between smoke particles and water mist particles, wherein the depolarization ratio of smoke particles is about 0.05 and the depolarization ratio of haze particles is between 0.04 and 0.08.

[0018] Preferably, the communication module supports MQTT, CoAP and HTTP communication protocols; the communication module achieves wireless communication through a GPRS module.

[0019] Preferably, the monitoring platform is developed based on the Spring Boot framework and uses the Spring Cloud toolset to build a distributed system based on a microservice architecture. It completes communication with detectors, data persistence, and management functions, and provides related services to users through two clients: a browser and an Android App. The monitoring platform also provides an alarm notification service. When the monitoring platform receives alarm information from the detector, in addition to providing corresponding prompts on the two clients, it will also promptly notify users via SMS and email.

[0020] Preferably, the system further includes: a drone monitoring component, which is used to monitor power facilities in real time using image sensors carried by the drone, and combine this with data from a ground-based lidar device to locate the fire ignition point; the drone monitoring component is also used to perform color analysis on each pixel in the image data using a comprehensive flame color model based on RGB and HSI, and divide the image into flame and non-flame areas, specifically including the following steps:

[0021] S1: Acquire image data of the power fire scene using drones;

[0022] S2: Using a comprehensive flame color model based on RGB and HSI, perform color analysis on each pixel in the image data and divide it into flame regions and non-flame regions;

[0023] S3: The distribution of statistical information in an image is described using a Gaussian function, with the specific formula as follows:

[0024]

[0025] Where, μ x and μ y σ is the mean of the flame region in the image, and σ is the standard deviation.

[0026] Preferably, the system further includes: a fire prediction model based on artificial intelligence algorithms, wherein the fire prediction model based on artificial intelligence algorithms determines the direction of smoke movement and calculates the fire spread speed and locates the coordinates of the fire center point based on the fire time trajectory matrix; the construction process of the fire prediction model based on artificial intelligence algorithms includes:

[0027] A fire time trajectory matrix is ​​constructed, and the fire signal generated in the power fire is chaotically discriminated by the principal component analysis method to obtain the UAV image feature values ​​for power fire monitoring.

[0028] Feature extraction of the image is performed using the HIS color gamut, the image is converted from the RGB color space to the HSV color space, the tonal information in the image is analyzed, and the flame area is identified.

[0029] Using the concept of the pyramid model P(v,w,s), we analyze the maximum matrix order s' in the (t+1)th layer of the pyramid to characterize the motion characteristics of the smoke and determine the diffusion direction and speed of the smoke.

[0030] By setting two parameters, i and j, the movement components of the flue gas in the horizontal and vertical directions are described. The direction of flue gas movement is determined using the absolute error and minimum standard of the flue gas flow. The specific formula is as follows:

[0031]

[0032] By measuring the distance Δd that the fire spreads per unit time, the fire spread rate V = Δd / Δt is calculated. Combining the data on the direction of smoke movement and the fire spread rate, the location of the fire center point is determined by reverse reasoning, thus locating the fire ignition point.

[0033] Preferably, the flue gas diffusion prediction module includes: a CFD simulation database integration unit, a multi-source condition adaptive algorithm unit, and a real-time update mechanism unit;

[0034] The CFD simulation database integration unit is used to establish a two-dimensional fire smoke diffusion simulation database under different conditions through the fire dynamics simulator - Smokeview, providing training samples for deep learning models.

[0035] The multi-source conditional adaptive algorithm unit is used to perform chaotic discrimination on the signals generated by power fires based on the influence of fire source location, heat release rate, and wind speed on smoke diffusion, and to obtain UAV image feature values ​​for power fire monitoring.

[0036] The real-time update mechanism unit is used to automatically update the model parameters as new data is added, maintain optimal prediction performance, and automatically adjust the anchor box size during each training process to adapt to the target size in a specific dataset.

[0037] Preferably, the emergency decision-making scheme generation module includes: a fire emergency knowledge graph construction unit, a Bayesian network inference engine unit, and a sensitivity analysis and similarity calculation unit;

[0038] The fire emergency knowledge graph construction unit is used to extract key information from fire accident reports based on named entity recognition technology, and to form a fire emergency ontology together with pre-built knowledge of fire emergencies and emergency management, so as to realize the processing and utilization of unstructured information data.

[0039] The Bayesian network inference engine unit is used to select key features of fire accidents as nodes, transform entity types into node states through rule inference, and perform structure and parameter learning to establish a fire emergency Bayesian network, thereby supporting real-time decision-making.

[0040] The sensitivity analysis and similarity calculation unit is used to infer key features and the most likely accident outcome from partial information of the target fire accident, design fire accident similarity and fitness calculation methods, and obtain corresponding emergency decision-making schemes.

[0041] Preferably, the system further integrates video fire detection technology and an improved YOLOv5-Augmentation model for early-stage fire monitoring and alarm, as well as mid-stage fire smoke temperature field distribution analysis, providing information support for rescue work.

[0042] The video fire detection technology introduces the CBAM attention mechanism, replaces PANet with BiFPN to enhance the feature extraction network, replaces nearest neighbor interpolation with transposed convolution, and uses lightweight networks MobileNetV3, ShuffleNetV2, and GhostNet to lighten the YOLOv5s model.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. This laser positioning system for electrical fire monitoring and emergency response utilizes a lidar device to emit laser beams and receive reflected signals. Combined with a data processing unit that analyzes the reflected signals, it can accurately detect and locate fire sources. The lidar device employs polarized lidar technology, which can distinguish between smoke particles and water mist particles based on their different depolarization ratios, thus avoiding false alarms caused by environmental factors (such as water vapor and dust). Furthermore, the data processing unit uses waveform decomposition and frequency domain analysis to extract weak smoke signals from the background signal, further improving the accuracy of fire source identification. Once the location of the fire is determined, the communication module wirelessly transmits the information to the monitoring center, ensuring rapid and accurate location of the fire source.

[0045] 2. This laser positioning system for electrical fire monitoring and emergency response, upon receiving fire location information, immediately activates the emergency response procedure via a monitoring platform developed based on the Spring Boot framework. It notifies relevant personnel through multiple channels (including browser interface, Android application, SMS, and email), significantly reducing the time lag between fire detection and action. Simultaneously, the system includes a drone monitoring component equipped with image sensors for real-time monitoring of power facilities, combining this data with ground-based lidar to accurately pinpoint the fire's origin. This approach not only quickly confirms the exact location of the fire but also assesses the speed and direction of its spread, supporting timely and effective firefighting measures. Furthermore, the integrated artificial intelligence fire prediction model, through learning from historical data, improves the accuracy of future fire risk assessments, providing early warnings of potential risk areas and facilitating earlier preventative measures.

[0046] 2. This laser positioning system for electrical fire monitoring and emergency response combines an advanced U-Net convolutional neural network architecture and SiLU activation function with named entity recognition technology and Bayesian networks in its smoke diffusion prediction module and emergency decision generation module. This enables accurate prediction of the smoke diffusion range and speed at the fire scene and real-time generation of optimal emergency response strategies. By integrating sensor data such as temperature, humidity, wind speed, and CO2 concentration with building internal structural information, the system effectively overcomes the accuracy deficiencies of traditional methods, providing more precise smoke movement path analysis. Simultaneously, its ability to extract key information from unstructured data makes emergency decision-making more scientific and rational, significantly shortening the time lag between fire detection and action, ensuring rapid and accurate location of the fire ignition point and assessment of fire spread trends. This provides strong support for timely and effective firefighting measures and significantly improves the reliability and efficiency of the electrical fire monitoring system. Attached Figure Description

[0047] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are 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.

[0048] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0049] Figure 2 This is a structural diagram of the polarization lidar system of the present invention;

[0050] Figure 3 This is a general framework diagram of the monitoring platform of the present invention;

[0051] Figure 4 This is a diagram of the YOLOv5 network structure of the present invention;

[0052] Figure 5 This is a schematic diagram of the named entity recognition model of the present invention;

[0053] Figure 6 This is a flowchart illustrating the selection process for the fire emergency decision-making scheme of the present invention.

[0054] In the figure, 1—laser, 2—polarizer, 3—beam expander, 4—atmosphere, 5—backscattered light, 6—telescope, 7—aperture stop, 8—collimating lens, 9—polarizing beam splitter, 10—focusing lens, 11—PMT, 12—fiber optic cable. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, the present invention provides a laser positioning system for electrical fire monitoring and emergency response, characterized in that it includes:

[0058] LiDAR device: used to emit laser beams and receive reflected signals, and to detect and locate fire sources by analyzing the reflected signals; specifically, based on polarization lidar technology, according to the different depolarization ratios of smoke particles and water mist particles, the approximate direction and distance of the fire source can be preliminarily determined, and basic data collection and preliminary analysis can be completed to provide raw data for subsequent processing;

[0059] Data processing unit: Connected to the lidar device, used to process the received reflected signals and identify smoke particles and water mist particles based on changes in signal intensity to determine the location of the fire; specifically, it performs in-depth processing on the reflected signals from the lidar device, using waveform decomposition and frequency domain analysis to extract smoke signals from complex background signals and remove noise interference, thereby more accurately determining the location of the fire.

[0060] Communication module: Connected to the data processing unit, it transmits fire location information to the monitoring center;

[0061] Monitoring platform: Used to receive fire location information sent by the communication module and initiate emergency response procedures;

[0062] Smoke diffusion prediction module: Utilizing the U-Net convolutional neural network architecture and SiLU activation function, combined with sensor data on temperature, humidity, wind speed, CO2 concentration, and building internal structure information, it accurately predicts the diffusion range and speed of smoke at the fire scene.

[0063] Emergency Decision Generation Module: Based on named entity recognition technology and Bayesian networks, it extracts key information from unstructured data and generates the optimal fire emergency response strategy in real time. The unstructured data mainly refers to fire accident reports.

[0064] Example 1

[0065] The lidar device of this invention is a polarization lidar. This technology provides key support for effectively distinguishing between smoke particles and water mist particles. In actual operation, the lidar device emits a laser beam and then receives the reflected signal. Based on this, by deeply analyzing the reflected signal, the power information of the S-beam and P-beam can be obtained, and the power ratio between the two, i.e., the depolarization ratio, can be calculated. Different types of particles have significantly different depolarization ratios. For example, the depolarization ratio of smoke particles is usually around 0.05, while the depolarization ratio of haze particles is between 0.04 and 0.08. It is based on these different depolarization ratio characteristics that the lidar device can initially distinguish between smoke particles and water mist particles.

[0066] However, the signals received by lidar are not pure; they contain noise from the solar background light and noise generated by the detector itself. This noise can interfere with the accurate identification of particles. Therefore, the data processing unit plays an important role in subsequent processing. It first uses wavelet filtering to denoise the noisy signal, effectively reducing the impact of noise. Then, the data processing unit uses waveform decomposition and frequency domain analysis to extract the weak smoke signal from the complex background signal. In this process, the data processing unit performs a more detailed analysis and judgment of the particle type based on changes in signal strength, thereby more accurately identifying smoke particles and water mist particles, further improving the accuracy of fire source identification.

[0067] like Figure 2 As shown, the light emitted from laser 1 first passes through polarizer 2 and then through beam expander 3 to obtain linearly polarized light with magnified and collimated spot. After that, the laser enters the atmosphere 4. During the process of interacting with matter, the polarization state of the light changes constantly. The backscattered light signal 5 is received by telescope 6 and passes through aperture stop 7, collimating lens 8, and then polarizing beam splitter 9. The transmitted light is the component parallel to the original polarization state (P light), and the light reflected at 90° is the component perpendicular to the original polarization state (S light). Finally, the parallel light passes through focusing lens 10 and is then converted into an electrical signal by photomultiplier tube 11 for processing.

[0068] Based on the lidar equations, the powers of the S-beam and P-beam can be derived as follows:

[0069]

[0070] In the formula: k p ,k s —System constants for the P-channel and S-channel optical paths; P L —Laser single-pulse energy; Y r — Geometric overlap factor of the lidar system; c — speed of light; τ — laser pulse width; A r — Telescope receiving area; r — Detection distance; β p (r),β s (r) — the parallel and perpendicular components of the aerosol backscattering coefficient at height r; α p (r),α S (r) — Parallel and perpendicular components of the aerosol extinction coefficient at height r, where α is the extinction coefficient for randomly oriented aerosol particles. p (r) and α S (r) values ​​are equal;

[0071] In simple terms, the power ratio of the S-beam and the P-beam is the depolarization ratio, which can be used to examine whether the shape of the detected particles is regular, thus allowing for a rough determination of the particle type. By comparing the two formulas of the lidar equation and eliminating identical parameters, we can obtain:

[0072]

[0073] As defined, the deflection ratio is shown in the following formula:

[0074]

[0075] Therefore, the formula for calculating the deflection ratio is derived from the two formulas above:

[0076]

[0077] k P and k S The ratio of these two values ​​is called the system's gain ratio. Ideally, it is 1. However, due to the inconsistencies between optical and photoelectric conversion elements, it must be calibrated experimentally. Finally, the P value can be directly measured by the system. S (r) and P P (r) is calculated.

[0078] Example 2:

[0079] The lidar signal contains solar background noise and detector noise, and has nonlinear, non-stationary, and discrete characteristics. Therefore, the data processing unit uses wavelet filtering to denoise the weak signal. The principle of wavelet denoising is to map the entire signal from the time domain to the wavelet domain. It takes advantage of the fact that the wavelet coefficients of the noise signal decay rapidly with the increase of the decomposition scale, while the effective signal basically does not change with the decomposition scale. By removing the wavelet coefficients of the noise signal and reconstructing the signal, the denoised lidar signal can be accurately restored.

[0080] Assume the detected noisy Mie scattering echo signal is:

[0081] f(t) = p(t) + e(t) + s(t)

[0082] Where: p(t) – clean echo signal; e(t) – detector noise; s(t) – background light noise;

[0083] Based on the definition of wavelet analysis, its V is given. j The standard equations in space are as follows:

[0084]

[0085] The decomposition and reconstruction formulas for the wavelet coefficients are as follows:

[0086]

[0087] In the formula: —Approximation coefficient; — Detail coefficients; h — Filter function; j — Number of decomposition levels; k — Signal length; l — Positive integer related to discrete grid points.

[0088] Among them, formula The two parts are the low-frequency and high-frequency components of the signal, respectively. (Formula) For the scaling function, the formula is... It is a wavelet function.

[0089] In the calculation process, the low-frequency part mainly consists of the detected aerosol scattering signal, while the high-frequency part consists of various noise signals to be removed. The specific operation is as follows: First, determine the wavelet basis type and the number of decomposition layers. For discrete lidar signals, wavelet bases such as Daubechies and Symlets A are commonly used. The choice of the number of decomposition layers is very important. If the number of layers is too small, the noise cannot be eliminated to the maximum extent, and if the number of layers is too large, the signal-to-noise ratio will decrease. Therefore, the number of decomposition layers is generally selected to be 3 to 6. Second, threshold quantization is performed on the high-frequency coefficients to remove system noise and solar background light noise. Finally, the lowest-level low-frequency coefficients and the high-frequency coefficients of each layer are reconstructed into the denoised echo signal.

[0090] Example 3:

[0091] The overall framework of the monitoring platform is as follows: Figure 3 As shown, it mainly consists of two parts: the server and the client. The server is divided into microservice architecture components, a database, and microservice application modules. The microservice architecture components are provided by Spring Cloud and mainly include API gateway, service discovery, load balancing (not shown in the diagram), and service fault tolerance protection. These components are the foundation for implementing the microservice architecture. The server uses the relational database MySQL and the non-relational database Redis for data persistence and caching, respectively. The database is an important external dependency for many service functions of the application modules. The microservice application modules implement specific business logic, which is the main purpose of server development. The client includes a browser client and an Android mobile client, which exchange information with the server via the HTTP protocol.

[0092] Example 4:

[0093] This study applies drone monitoring technology to pinpoint the ignition point of power fires and utilizes drone-captured images to effectively segment suspected fire sources. In power fire scenarios, the color distribution of flames exhibits specific patterns. In-depth research into these patterns, combined with real-time drone monitoring, high-altitude aerial views of the fire scene, and image processing and data analysis techniques, enables precise extraction of fire zones. Furthermore, automated operation and information sharing improve firefighting efficiency and reduce personnel risks, bringing new possibilities to power fire emergency management and rescue efforts. To more accurately identify flames, a comprehensive flame color model based on RGB and HSI is adopted. The RGB model is based on red, blue, and black pixels. The combination of three basic colors, green and blue, describes the color information in an image, while the HSI model characterizes color properties from three dimensions: hue, saturation, and brightness. Both models have their advantages in flame color recognition, so combining them can more effectively extract suspected flame areas. In practical applications, image data of a power fire scene is first acquired using a drone. Then, a comprehensive flame color model based on RGB and HSI is used to analyze the color of each pixel in the image, dividing it into flame and non-flame areas. Through fine processing of color information, most non-flame areas can be effectively filtered out, retaining suspected flame areas.

[0094] The method for locating the origin of a fire includes the following steps:

[0095] Step 1: First, acquire real-time image data of power facilities using the image sensor carried by the drone;

[0096] Step 2: Then, a comprehensive flame color model based on RGB and HSI color spaces is used to analyze each pixel in these images, dividing the images into flame regions and non-flame regions. This color model based on RGB and HSI fully utilizes the advantages of the two color spaces. The RGB model describes the color information in the image based on the combination of the three basic colors of red, green and blue, while the HSI model describes the color characteristics from three dimensions: hue, saturation and brightness. The combination of the two can more effectively identify suspected flame regions.

[0097] A comprehensive flame color model based on RGB and HSI:

[0098] A two-layer parallel feature extraction structure was constructed based on the comprehensive flame color model. In the first layer, an RGB feature extraction sub-network based on convolutional neural network was designed for the RGB color space. This sub-network contains multiple convolutional layers and pooling layers. By using convolutional kernels of different sizes, the flame color features in the RGB image are extracted at multiple scales. For example, a 3×3 convolutional kernel is used to capture the local color details of the flame, and a 5×5 convolutional kernel is used to extract color distribution features over a larger range. At the same time, an attention mechanism is introduced to enable the network to pay more attention to the key color features of the flame area and suppress the interference of background noise.

[0099] In the second layer, a Transformer-based HSI feature extraction sub-network is constructed for the HSI color space. The Transformer's self-attention mechanism can effectively capture the global dependencies between hue, saturation, and brightness in the HSI color space. Through a multi-head attention mechanism, flame color features, such as hue variation trends, saturation distribution characteristics, and brightness differences, are extracted from different feature representation sub-spaces. Then, the features extracted by the RGB feature extraction sub-network and the HSI feature extraction sub-network are fused. The fusion method combines feature concatenation and weighted fusion. The fusion weights are dynamically adjusted according to the importance of RGB and HSI features in different scenarios, thereby obtaining a more comprehensive and representative flame color feature representation. This two-layer parallel feature extraction structure makes full use of the different characteristics of the RGB and HSI color spaces, improving the accuracy and robustness of flame color feature extraction.

[0100] This model introduces an adaptive threshold dynamic adjustment module. This module, based on environmental sensing sensors such as light and smoke sensors, acquires environmental information in real time. Simultaneously, it utilizes historical image data and statistical information from the current image, such as color histograms and brightness distribution, and employs machine learning algorithms, such as a regression model based on Support Vector Machines (SVM), to dynamically adjust the threshold for flame color recognition. For instance, in environments with strong light, the model automatically increases the brightness threshold to avoid misjudgments due to excessive ambient light. In environments with heavy smoke, it adjusts the hue and saturation threshold ranges based on the influence of smoke on flame color, ensuring accurate flame color identification. Through this adaptive threshold dynamic adjustment module, the model can maintain high flame recognition accuracy under different environmental conditions, enhancing its adaptability and stability.

[0101] Insufficient accuracy in flame core extraction can affect the overall quality of flame segmentation in power fire images, potentially impacting the accuracy of disaster assessment and the timeliness of rescue operations. To address this issue, a flame image segmentation algorithm based on an RGB color probability model is proposed. This method fully utilizes the characteristics of the RGB color space, constructing a color probability model to meticulously analyze each pixel in the image, thereby more accurately identifying flame regions, especially the flame core. The specific steps include:

[0102] Step 1: Data Preparation and Image Preprocessing

[0103] Image collection: Collect a wide range of images including various fire scenes, such as fires caused by different environments (indoor and outdoor), different combustible materials, and different lighting conditions (direct sunlight and dim light);

[0104] Image adjustment: Adjust the image size uniformly to ensure that all images have the same dimensions for easier subsequent processing. Then, normalize the RGB values ​​in the images, adjusting the value range to between 0 and 1.

[0105] Step 2: Marking the Flame Area

[0106] Using image annotation tools, manually and carefully annotate the flame area and flame core in the image, assigning one mark to the flame area, another different mark to the flame core, and another mark to other non-flame areas;

[0107] Step 3: Construct a color probability model

[0108] Statistical RGB characteristics: Extract RGB values ​​from the labeled flame area and flame core, and calculate their average value and fluctuation (similar to the concept of standard deviation, reflecting the dispersion of the data);

[0109] Establish a distribution model: Based on the statistical average and fluctuation, establish color distribution models for the flame area and the flame core respectively; this can be understood as creating a "file" for the color characteristics of the flame area and the flame core, recording their respective color features.

[0110] Step 4: Pixel Analysis and Classification

[0111] Calculate the probability: For each pixel in the image, substitute its RGB value into the previously established color distribution model of the flame region and flame core to calculate the probability that this pixel belongs to the flame region and flame core.

[0112] Classification determination: Set two judgment criteria values; if the probability of a pixel belonging to the flame area is greater than the first criterion value, it is initially considered to belong to the flame area; based on belonging to the flame area, if the probability of it belonging to the flame core is greater than the second criterion value, and this probability is also greater than the probability of belonging to the flame area, then the pixel is determined to belong to the flame core.

[0113] Step 5: Optimize classification results

[0114] Morphological processing: Morphological operations such as dilation and erosion are used; dilation can appropriately expand the range of the flame area and fill in some small holes that may exist; erosion can remove some isolated small pixels that may be noise, making the edges of the flame area clearer and more accurate;

[0115] Connected component analysis: Connected component analysis is performed on the morphologically processed image to remove small, discontinuous regions as noise, retaining only larger regions that match the characteristics of a flame, thus further optimizing the identification results of the flame region and the flame core.

[0116] Step Six: Model Evaluation and Optimization

[0117] Evaluate model performance: Test the model using images that were not previously used in model building; measure the accuracy of the model in identifying flame regions and the flame core by calculating metrics such as accuracy, recall, and F1 score.

[0118] Adjusting model parameters: Based on the model evaluation results, the parameters in the color distribution model, such as the average value, fluctuation, and the standard values ​​for classification, are adjusted and optimized to improve the model's performance and make the recognition more accurate.

[0119] This method improves the accuracy of flame segmentation and better preserves the detailed information of the flame core. As shown in the formula below, the Gaussian function is used to describe the distribution of statistical information in the image.

[0120]

[0121] Based on the significant characteristics of electrical fire flames, a series of key indicators are selected as the core basis for fire identification; high-frequency energy information of sub-images is accurately calculated to achieve accurate identification of electrical fire flames; this process is represented as follows:

[0122] e(x,y)=∣LH N | 2 +∣HL N | 2 +∣HH N | 2

[0123] In the formula, e(x,y) represents the high-frequency energy information of the power fire scene image collected by the UAV; LH N HL N HH N These are high-frequency energy information in the vertical, diagonal, and horizontal directions of the flame surface.

[0124] The energy within the combustion zone is quantified into a one-dimensional random sequence. After selecting a random sequence with higher regularity, wavelet analysis is performed to effectively distinguish different frequencies. During combustion, low-frequency coefficients can better reflect the overall trend of energy change within the combustion zone. During a disaster, the energy within the fire zone shows a significant upward trend due to continuous accumulation. The zero-crossing rate of high-frequency coefficients is used to characterize the flashing frequency of flame energy during a fire. When a fire occurs, due to the flashing characteristics of the fire, its frequency coefficient changes continuously with the disaster. When the target has a slight change in boundary shape, its high-frequency coefficient tends to zero. In addition, as time passes, the affected area of ​​the fire continues to increase under the influence of high temperature. The change in the area of ​​suspected fire zones in images acquired by UAVs is calculated to characterize the changes in flame morphology. The calculation is as follows:

[0125]

[0126] In the formula, A is the relative area change rate of the suspected fire area; V1V2 is the area of ​​the suspected fire area in the adjacent images; this indicator can intuitively reflect the speed and range of flame spread, providing an important basis for fire monitoring and response.

[0127] By combining the above methods with drone monitoring and artificial intelligence technology, the area of ​​power fire can be identified.

[0128] Example 5:

[0129] The structure of the fire time trajectory matrix includes data composition and matrix layout.

[0130] Data composition design

[0131] Time-series recording: In order to accurately capture the development process of electrical fires, the fire is continuously monitored and recorded at certain time intervals. At the beginning of the fire, due to the rapid changes in the fire intensity, the time interval is set very short, such as recording relevant data once every 10 seconds. As the fire gradually stabilizes, the time interval can be appropriately lengthened, adjusted to record once every 30 seconds. This dynamic time interval setting can obtain more intensive data during the critical change phase of the fire and avoid excessive redundant recording during the relatively stable phase, effectively improving the efficiency and quality of data collection.

[0132] Multi-feature fusion recording: Comprehensive collection of various feature data related to electrical fires. Temperature is a key feature. Multiple high-precision temperature sensors distributed at different locations at the fire scene are used to acquire temperature data at each location in real time. At the same time, smoke concentration is also an important feature. Professional smoke concentration detection equipment is used to continuously monitor changes in smoke concentration in different areas. In addition, flame brightness information is recorded. With the help of special optical sensors, changes in flame brightness are measured. These multi-faceted feature data reflect the development trend of the fire from different perspectives.

[0133] Spatial zoning recording: The fire scene is divided into multiple appropriately sized areas, each serving as an independent recording unit. For large power facility sites, the area is divided into small grid areas based on equipment distribution and layout. Within each grid area, characteristic data such as temperature, smoke concentration, and flame brightness are recorded. This not only provides an overall understanding of the fire scene but also clearly reveals the differences in fire characteristics in each local area, providing detailed spatial data support for subsequent precise analysis.

[0134] Matrix layout design includes row structure design and column structure design:

[0135] Row structure design: Each row of the matrix represents a specific point in time. Starting from the start of fire monitoring, data is recorded at preset time intervals. For example, if the monitoring starts at 9:00 AM, and the initial setting is to record data every 10 minutes, then the first row records the fire-related data at 9:00 AM, the second row records the data at 9:10 AM, the third row records the data at 9:20 AM, and so on. As the fire develops, if the fire intensity tends to stabilize, the recording interval can be adjusted to once every 30 minutes. Subsequent rows still record data sequentially according to the adjusted time intervals. Each row of data fully covers all aspects of the fire information at that moment, making it convenient to observe the changes in the fire situation over time.

[0136] Column structure design: The columns of the matrix are divided into two main categories: fire characteristics and spatial regions.

[0137] Fire Characteristics Columns: Different fire characteristics are grouped and arranged. First, temperature-related columns are arranged, such as a "Equipment A Temperature" column to record the real-time temperature at equipment A; a "Equipment B Temperature" column to record the temperature at equipment B. After the temperature columns, smoke-related columns are set, such as a "Area 1 Smoke Concentration" column to record the smoke concentration in area 1; a "Area 2 Smoke Concentration" column to record the smoke concentration in area 2. In addition to the directly measured data columns, trend columns are also set, such as a "Temperature Trend" column to record whether the temperature rises, falls, or remains stable at adjacent time points; and a "Smoke Concentration Trend" column to record the changing trend of smoke concentration.

[0138] Spatial Region Column: Following the Fire Characteristics Column, a spatial region column is set up. If the fire scene is divided into multiple regions, such as regions A, B, C, etc., corresponding columns are set up to clarify the information of each region. For example, the "Equipment Status in Region A" column records the operating status of the equipment in region A, whether it is normal, faulty, or damaged; the "Combustion Status in Region B" column records information such as whether there is an open flame in region B and the size of the burning area. These columns work together with the preceding fire characteristics column so that each fire characteristic data can be mapped to a specific spatial region, which facilitates the analysis of the development of the fire in different regions.

[0139] Constructing a time trajectory matrix for electrical fires and a model for determining the direction of smoke movement (a fire prediction model based on artificial intelligence algorithms) aims to accurately track the path and speed of fire spread, helping to predict and control the spread of fire. Determining the direction of smoke movement helps commanders make correct response decisions and improves emergency response efficiency. Simultaneously, it provides scientific support for optimizing the allocation of firefighting resources, maximizing the control of fire spread and ensuring the safety of personnel and property. First, principal component analysis is used to perform chaotic discrimination on fire signals generated in electrical fires. It is assumed that a one-dimensional time series can be represented as:

[0140] x(0),…,x(k),…,x(N-1)

[0141] Where N represents the length of data collected by the drone during monitoring, as shown in the following formula, the time trajectory matrix of the power fire is constructed:

[0142]

[0143] In the formula, X is the time trajectory matrix; l is the coefficient; m is the embedding dimension. Using the above formula, the feature values ​​of UAV images used for power fire monitoring are obtained, and a principal component analysis map of the risk of forest fires along power transmission lines is established. Through in-depth analysis of this map, a fine analysis of the flame signal in the image is achieved, providing strong support for accurate identification and emergency response. In images of power transmission line fires taken by UAVs, smoke often contains fine particles and water vapor, exhibiting significant blackening characteristics. Using the HSI domain for feature extraction yields more accurate analysis results. This method can more comprehensively capture color information in the image, especially those features closely related to smoke and flames. Next, the normalized RGB vector (r, g, b) is substituted into the conversion formula to obtain the h component of the HSV color domain. This step is crucial, as it converts the image from the RGB color space to the HSV space, allowing for more intuitive analysis of the hue information in the image. Through this color domain conversion and feature extraction method, smoke features in power transmission line fire images can be more accurately identified and analyzed.

[0144] In the context of electrical fires, as heat accumulates, the density of smoke gradually decreases and then increases. To accurately describe the neighborhood movement direction of smoke in such disasters, the concept of a pyramid model P(v,w,s) is introduced. By analyzing the order s′ of the largest square matrix in the (t+1)th layer of the pyramid, the movement characteristics of smoke can be characterized in more detail.

[0145] To quantify the direction of smoke movement, two parameters, i and j, are defined, describing the horizontal and vertical components of smoke movement during a fire. Based on this, the direction of smoke movement is accurately determined using the absolute error and minimum standard of the smoke flow, as shown in the following formula:

[0146]

[0147] This method not only considers the continuity of smoke movement but also takes into account its rate of change in different directions, thus ensuring the accuracy and reliability of the results. To obtain the discrete smoke movement direction values ​​(i,j), the following formula is used for calculation. This step transforms the continuous smoke movement direction into discrete values, making it easier to analyze and predict the smoke movement.

[0148]

[0149] PCA-based video image processing algorithms can not only accurately extract the feature values ​​of the measured images, but also construct principal component analysis maps of power grid fires based on the obtained feature values.

[0150] The specific process of determining the location of the fire's epicenter through reverse reasoning includes:

[0151] Step 1: Data Collection and Organization: During a fire, data is collected comprehensively using various monitoring devices. Through thermal imagers, smoke sensors, and monitoring equipment equipped with image recognition functions, information such as the fire spread distance, smoke movement direction, and fire spread speed at different times is continuously collected. This data is then organized in chronological order to form a systematic dataset, providing accurate and complete data support for subsequent reverse reasoning.

[0152] Step 2: Construct a fire spread model: Based on the collected data, construct a dynamic fire spread model. This model takes into account the influence of different environmental factors on fire spread, such as the structure and material of the building, the surrounding temperature, humidity, and wind speed and direction. For different types of fire scenarios, the model will adjust the parameters according to the actual situation to more accurately simulate the development process of the fire. Through this model, you can intuitively see the spread of the fire at different times, as well as the changes in the diffusion path and speed of the smoke.

[0153] Step 3: Identify key time points and characteristic data: During the fire's spread, select several representative key time points. These time points are selected based on significant changes in the fire, such as a sudden acceleration in the fire's spread or a change in the direction of smoke diffusion. For each key time point, extract corresponding key characteristic data such as the fire's spread distance, the direction of smoke movement, and the speed of fire spread. This data will serve as an important basis for reverse reasoning, helping us to trace the origin of the fire.

[0154] Step 4: Reverse Reasoning Process: Starting from the current state of the fire, and using key time points as a benchmark, reverse reasoning is performed based on the fire spread speed and smoke movement direction. For example, if at a certain key time point, the fire spreads northeastward at a speed of 2 meters per second, then during reverse reasoning, the fire is traced back in the opposite direction (southwest) at the same speed. During the tracing process, the reasoning path is continuously adjusted based on changes in the smoke movement direction. If the smoke shows an abnormal diffusion direction within a certain period of time, it indicates that there may be special environmental factors affecting the fire in that area. At this time, the reasoning path needs to be corrected to more accurately approach the center of the fire.

[0155] Step 5: Cross-validation and Precise Location: Through multiple reverse reasoning steps, starting from different key time points, multiple possible backtracking paths to the fire's origin are obtained. These paths are cross-compared and validated to identify their intersection area. This intersection area is very likely the location of the fire's epicenter. To further improve the accuracy of the location, information such as the terrain and building layout at the fire scene is analyzed. If the intersection area happens to be located in a specific location within the building, such as an area with concentrated electrical equipment or a flammable material storage area, then it can be more certain that this area is the fire's epicenter. At the same time, other auxiliary information, such as residual traces at the fire scene and equipment failure records, is used to further validate and adjust the location results, ultimately achieving rapid and accurate location of the fire's epicenter.

[0156] Example 6:

[0157] In this embodiment, the operation process of the laser positioning system for electrical fire monitoring and emergency response includes:

[0158] Step 1: System Initialization: Activate the lidar device, data processing unit, communication module and monitoring platform, and calibrate the lidar device;

[0159] Step 2: Real-time monitoring: The lidar device emits a laser beam and receives the reflected signal. The data processing unit analyzes the signal to identify smoke particles and water mist particles and determine the location of the fire.

[0160] Step 3: Data transmission: Once signs of fire are detected, the fire location information is wirelessly transmitted to the monitoring center via a communication module that supports MQTT, CoAP, and HTTP protocols;

[0161] Step 4: Emergency Response: After receiving the fire location information, the monitoring platform activates the emergency procedure, notifies relevant personnel via SMS and email, and displays alarm information on browsers and Android apps; at the same time, it deploys drones equipped with image sensors to the scene as needed, and combines the data from ground-based lidar to accurately locate the fire point; it uses a fire prediction model based on artificial intelligence algorithms to conduct risk assessment and provide information on the direction and speed of fire spread;

[0162] Step 5: Follow-up: Organize appropriate firefighting operations, assess the extent of damage and plan repairs after the fire is extinguished, and collect event data to optimize the system and train more advanced fire prediction models.

[0163] In this embodiment, the flue gas diffusion prediction module further includes:

[0164] CFD simulation database integration unit: Establishes a two-dimensional smoke diffusion simulation database under different conditions using the fire dynamics simulator - Smokeview (FDS-SMV) to provide training samples for deep learning models;

[0165] Multi-source conditional adaptive algorithm unit: Considering the influence of fire source location, heat release rate, and wind speed on smoke diffusion, the principal component analysis method is used to perform chaotic discrimination on the signals generated by power fires, and obtain UAV image feature values ​​for power fire monitoring;

[0166] Real-time update mechanism unit: As new data is added, the model can automatically update its parameters to maintain optimal prediction performance. During each training process, the anchor box size is automatically adjusted to better adapt to the target size in a specific dataset.

[0167] Example 7:

[0168] The smoke diffusion prediction module is part of the electrical fire monitoring and emergency response laser positioning system. It aims to accurately predict the range and velocity of smoke diffusion at a fire scene using deep learning methods. This module combines the U-Net convolutional neural network architecture, the SiLU activation function, and a multi-source conditional adaptive algorithm.

[0169] I. Data Collection and Preprocessing

[0170] A two-dimensional smoke diffusion simulation database under different conditions was established using the fire dynamics simulator Smokeview (FDS-SMV) to provide training samples for the deep learning model. Specifically, three operating conditions were set: fire source location, longitudinal ventilation velocity, and fire source heat release rate, to construct a basic dataset containing 17,280 smoke temperature field images in 36 groups. Considering the influence of factors such as fire source location, heat release rate, and wind speed on smoke diffusion, principal component analysis was used to perform chaotic discrimination on the signals generated by power fires, obtaining UAV image feature values ​​for power fire monitoring. As new data is added, the model can automatically update parameters to maintain optimal prediction performance, and automatically adjust the anchor frame size during each training process to better adapt to the target size in the specific dataset.

[0171] II. Deep Learning Model Design

[0172] U-Net was chosen as the basic architecture due to its excellent performance in medical image segmentation tasks, making it suitable for tasks requiring fine boundary definitions, such as smoke diffusion prediction. The U-Net model employs an encoder-decoder framework. The encoder uses techniques such as convolution and pooling to extract spatial information from the image. The decoder, based on the features extracted by the encoder and combined with information obtained from upsampling during decoding, utilizes multi-scale feature fusion techniques to reconstruct image details. SiLU (Sigmoid Linear Unit) is introduced as the activation function, which possesses good nonlinear expressive power and stability, helping to improve the model's learning performance. The SiLU activation function formula is:

[0173]

[0174] By utilizing the concept of a pyramid model, the maximum matrix order s' in each layer of the pyramid is analyzed to characterize the motion characteristics of smoke, thereby determining the diffusion direction and speed of the smoke. In addition, BiFPN (Bidirectional Feature Pyramid Network) is used to enhance the feature extraction network, and features of different resolutions are unified through a resize operation. Then, multiple top-down and bottom-up multi-scale feature fusions are performed.

[0175] III. Model Training and Validation

[0176] Based on the aforementioned dataset, the collected data was divided into training and testing sets. The cross-entropy loss function was used to guide model learning, and an early stopping mechanism was implemented to prevent overfitting. For the U-Net model, validation results showed that, based on the tunnel mid-section, a heat release rate of 5MW, and a windless condition, the SiLU activation function exhibited better nonlinear expression and stability when training the neural network. The model's performance was evaluated by comparing the actual flue gas diffusion with the model's predictions. Key metrics included, but were not limited to, mean absolute error (MAE) and root mean square error (RMSE). Experiments demonstrated that the U-Net model can meet the flue gas diffusion prediction requirements of different scenarios and exhibits excellent prediction accuracy.

[0177] IV. Integration and Deployment

[0178] The trained smoke diffusion prediction model is integrated into the overall electrical fire monitoring and emergency response laser positioning system, enabling it to activate immediately and output prediction information upon fire detection. This process involves inputting the data stream collected by the sensors into the trained U-Net model, performing inference calculations, and outputting prediction results. A user-friendly interface is developed so that firefighters and other relevant personnel can intuitively view the smoke diffusion prediction results, assisting them in making quick and accurate decisions.

[0179] In this embodiment, the emergency decision-making scheme generation module further includes:

[0180] Fire Emergency Knowledge Graph Construction Unit: Based on named entity recognition technology, key information is extracted from fire accident reports and combined with pre-built knowledge of fire emergencies and emergency management to form a fire emergency ontology, realizing the processing and utilization of unstructured information data;

[0181] Bayesian network inference engine unit: Selects key features of fire accidents as nodes, transforms entity types into node states through rule inference, and performs structure and parameter learning to establish a fire emergency Bayesian network, thereby supporting real-time decision-making;

[0182] Sensitivity analysis and similarity calculation unit: By inferring key features and the most likely accident outcome from partial information of the target fire accident, designing methods for calculating fire accident similarity and fitness, and obtaining corresponding emergency decision-making schemes.

[0183] Example 8:

[0184] like Figure 5As shown, the emergency decision-making module first constructs a fire emergency knowledge graph. It uses Named Entity Recognition (NER) technology to extract key information from a large number of unstructured fire accident reports, such as the location of the fire source, the speed of fire spread, and the direction of smoke diffusion. This information is then combined with pre-built knowledge of fire emergencies and emergency management to form a fire emergency ontology. This ontology serves as the top-level architecture, guiding the construction of the knowledge graph and enabling effective processing and utilization of unstructured data. Specifically, by parsing the sentence structure in the text, entities and their relationships in the fire event are extracted, such as the time and location of the fire, the type of building involved, and the fire extinguishing equipment used. This information is then mapped to a predefined knowledge framework, thereby establishing a rich fire emergency knowledge graph. This invention uses Protégé-5.6.1 as an ontology construction tool, defining the domain name as "fire emergency." Based on the ontology classes and class hierarchy determined earlier, the Classes module in Protégé is used to define the defined concept classes and their subclasses. Then, the relationships between concept classes are edited using the Object properties module, and the relationships are further refined in the Data... In the properties module, edit the data properties of the class to determine the data type; finally, use the OWL language to save the fire emergency ontology model for subsequent development, maintenance, reasoning, and querying.

[0185] like Figure 6 As shown, next, based on the aforementioned knowledge graph, a Bayesian network inference engine is used to generate emergency decision-making solutions. In this process, key features of the fire accident are selected as nodes in the Bayesian network, such as fire scale, fire source type, and surrounding environmental factors. Through learning from historical fire data, rule-based inference transforms entity types into node states, i.e., setting the probability distribution of nodes according to different input conditions. Simultaneously, structure and parameter learning is performed to determine the dependencies and weights between nodes, ultimately establishing a complete fire emergency Bayesian network model. This step not only realizes the transformation from a knowledge graph model to the basic structure of a Bayesian network but also uses the entity data in the knowledge graph as a dataset for Bayesian network learning. This invention enhances the model's predictive ability and adaptability. It constructs a Bayesian network structure learning method based on pre-edited knowledge. The initial network structure derived from the fire emergency ontology improves the efficiency and accuracy of the Bayesian network structure learning method. The resulting Bayesian network structure learning method first uses the K2 scoring function as the sampling basis based on the given initial network structure, and then uses the MCMC search algorithm to obtain a set of superior network structures. Finally, it statistically analyzes the network structure that appears most frequently as the final result. Since the fire emergency entities corresponding to node states in the fire emergency knowledge graph cannot be completely extracted, missing data is filled in based on the mode of the corresponding node in the dataset.

[0186] To improve the quality of decision-making schemes, sensitivity analysis and similarity calculation mechanisms were introduced. By reasoning from partially known information about the target fire accident, key characteristics and the most likely outcomes can be derived. Based on these inferences, a set of methods for calculating fire accident similarity and fitness was designed to evaluate the effectiveness and applicability of different emergency measures. For example, by comparing the similarity between the current fire scenario and other historical cases, the most suitable response strategy for the current situation can be selected; or, based on the fitness score under a specific environment, resource allocation or action priorities can be adjusted to ensure that every measure taken minimizes losses and protects personnel safety. The similarity threshold was set to 0.8, resulting in a candidate set array of similar fire accidents, simSet. The fitness calculation formula is:

[0187] fit(P,H)=ω·res(P,H)+(1-ω)ris(P,H)

[0188] Where ω is the weight of the accident result in the fitness calculation compared with the accident risk, which is set to 0.5 in this paper; when the fitness is greater than or equal to 0, it means that the fire emergency decision-making scheme for this similar fire accident is applicable to the current target fire accident, and the larger the fitness value, the better the applicability; when the fitness is less than 0, it means that the fire emergency decision-making scheme for this similar fire accident is not applicable to the current target fire accident.

[0189] Example 9:

[0190] like Figure 4 As shown, the system further integrates video fire detection technology and an improved YOLOv5-Augmentation model for early-stage fire monitoring and alarm, as well as mid-stage fire smoke temperature field distribution analysis, providing important information support for rescue work.

[0191] The proposed video fire detection technology introduces the CBAM attention mechanism, replaces PANet with BiFPN to enhance the feature extraction network, replaces nearest neighbor interpolation with transposed convolution, and employs lightweight networks MobileNetV3, ShuffleNetV2, and GhostNet to streamline the YOLOv5s model. Validation results show that the improved YOLOv5s model achieves an average accuracy of 82.1%, with 5.9M parameters and a computational cost of 8.1 GFLOPs.

[0192] Finally, to ensure the system's real-time performance and dynamic update capabilities, when a new fire accident occurs, the system automatically collects relevant information and adds it to the existing knowledge base, continuously optimizing the Bayesian network model. In addition, for each training process, the system can automatically adjust the anchor box size, enabling the model to better adapt to new target size changes and maintain optimal predictive performance. This continuous learning approach helps the system become more intelligent and efficient over time, providing more accurate and reliable decision support for future fire emergencies.

[0193] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A laser positioning system for electrical fire monitoring and emergency response, characterized in that, The system includes: a lidar device, a data processing unit, a communication module, a monitoring platform, a smoke diffusion prediction module, an emergency decision-making scheme generation module, and a drone monitoring component; The lidar device is used to emit laser beams and receive reflected signals, and to detect and locate fire sources by analyzing the reflected signals; The data processing unit is used to process the received reflected signals, identify smoke particles and water mist particles based on changes in signal intensity, and extract smoke signals from the background signals using waveform decomposition and frequency domain analysis to identify the fire source and determine the location of the fire. The communication module is used to wirelessly transmit the location information of the fire to the monitoring platform. The monitoring platform is used to locate the source of a fire based on its location. The smoke diffusion prediction module is used to predict the diffusion range and speed of smoke at the fire scene based on the fire source, using the U-Net convolutional neural network architecture and SiLU activation function, combined with temperature, humidity, wind speed, CO2 concentration sensor data and building internal structure information. The emergency decision-making scheme generation module is used to extract key information from fire accident reports based on named entity recognition technology and Bayesian networks, and generate the optimal fire emergency response strategy in real time. The system also includes: the UAV monitoring component, used to monitor power facilities in real time using image sensors carried by the UAV, and combined with data from a ground-based lidar device to locate the fire ignition point; a fire prediction model based on artificial intelligence algorithms, which determines the direction of smoke movement and calculates the fire spread speed based on the fire time trajectory matrix, and locates the coordinates of the fire center point; the construction process of the fire prediction model based on artificial intelligence algorithms includes: A fire time trajectory matrix is ​​constructed, and the fire signal generated in the power fire is chaotically discriminated by the principal component analysis method to obtain the UAV image feature values ​​for power fire monitoring. Feature extraction is performed on the image using the HSI color gamut, the image is converted from the RGB color space to the HSV color space, the tonal information in the image is analyzed, and the flame area is identified. Using the pyramid model The concept of analyzing the pyramids The order s' of the largest square matrix in the layer is used to characterize the motion characteristics of the smoke and determine the direction and speed of smoke diffusion. By setting Two parameters describe the movement of the flue gas in the horizontal and vertical directions. The direction of flue gas movement is determined using the absolute error and minimum standard of the flue gas flow. The specific formula is as follows: ； By measuring the distance the fire spreads per unit time Calculate the fire spread rate By combining data on the direction of smoke movement and the speed of fire spread, a reverse reasoning method was used to determine the location of the fire's center point and pinpoint the fire's origin.

2. The system according to claim 1, characterized in that, The lidar device uses polarization lidar technology, which can distinguish between smoke particles and water mist particles. The depolarization ratio of smoke particles is about 0.05, and the depolarization ratio of haze particles is between 0.04 and 0.

08.

3. The system according to claim 1, characterized in that, The communication module supports MQTT, CoAP and HTTP communication protocols; the communication module achieves wireless communication through a GPRS module.

4. The system according to claim 1, characterized in that, The monitoring platform is developed based on the Spring Boot framework and uses the Spring Cloud toolset to build a distributed system based on a microservice architecture. It completes communication with detectors, data persistence, and management functions, and provides related services to users through two clients: a browser and an Android app. The monitoring platform also provides an alarm notification service. When the monitoring platform receives alarm information from the detector, in addition to providing corresponding prompts on the two clients, it will also promptly notify users via SMS and email.

5. The system according to claim 1, characterized in that, The system further includes the UAV monitoring component, which is also used to perform color analysis on each pixel in the image data using a comprehensive flame color model based on RGB and HSI, and divide the image into flame and non-flame regions. Specifically, this includes the following steps: S1: Acquire image data of the power fire scene using drones; S2: Using a comprehensive flame color model based on RGB and HSI, perform color analysis on each pixel in the image data and divide it into flame regions and non-flame regions; S3: The distribution of statistical information in an image is described using a Gaussian function, with the specific formula as follows: ； in, and It is the mean value of the flame region in the image. σ That is the standard deviation.

6. The system according to claim 1, characterized in that, The flue gas diffusion prediction module includes: a CFD simulation database integration unit, a multi-source conditional adaptive algorithm unit, and a real-time update mechanism unit; The CFD simulation database integration unit is used to establish a two-dimensional fire smoke diffusion simulation database under different conditions through the fire dynamics simulator - Smokeview, providing training samples for deep learning models. The multi-source conditional adaptive algorithm unit is used to perform chaotic discrimination on the signals generated by power fires based on the influence of fire source location, heat release rate, and wind speed on smoke diffusion, and to obtain UAV image feature values ​​for power fire monitoring. The real-time update mechanism unit is used to automatically update the model parameters as new data is added, maintain optimal prediction performance, and automatically adjust the anchor box size during each training process to adapt to the target size in a specific dataset.

7. The system according to claim 1, characterized in that, The emergency decision-making scheme generation module includes: a fire emergency knowledge graph construction unit, a Bayesian network inference engine unit, and a sensitivity analysis and similarity calculation unit; The fire emergency knowledge graph construction unit is used to extract key information from fire accident reports based on named entity recognition technology, and to form a fire emergency ontology together with pre-built knowledge of fire emergencies and emergency management, so as to realize the processing and utilization of unstructured information data. The Bayesian network inference engine unit is used to select key features of fire accidents as nodes, transform entity types into node states through rule inference, and perform structure and parameter learning to establish a fire emergency Bayesian network, thereby supporting real-time decision-making. The sensitivity analysis and similarity calculation unit is used to infer key features and the most likely accident outcome from partial information of the target fire accident, design fire accident similarity and fitness calculation methods, and obtain corresponding emergency decision-making schemes.

8. The system according to claim 1, characterized in that, The system further integrates video fire detection technology and an improved YOLOv5-Augmentation model for early-stage fire monitoring and alarm, as well as mid-stage fire smoke temperature field distribution analysis, providing information support for rescue work. The video fire detection technology introduces the CBAM attention mechanism, replaces PANet with BiFPN to enhance the feature extraction network, replaces nearest neighbor interpolation with transposed convolution, and uses lightweight networks MobileNetV3, ShuffleNetV2, and GhostNet to lighten the YOLOv5s model.

Images