Forest fire monitoring method based on edge intelligence

Abstract

The invention relates to the technical field of environment monitoring, in particular to a forest fire monitoring method based on edge intelligence, which comprises the following steps: S1, deploying edge nodes and sensing equipment; s2, collecting data acquired by the sensing equipment in real time, and performing primary processing on the data; s3, fire risk intelligent identification and decision making are carried out through the edge nodes; s4, emergency response and event management; and S5, user interaction and feedback. The edge intelligence technology is introduced, the architecture of the forest fire monitoring system is optimized, and local processing and real-time analysis of data are realized, so that the response speed and the detection precision of fire monitoring are improved, and the method is suitable for further popularization and application.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and in particular to a forest fire monitoring method based on edge intelligence. Background Art

[0002] With global climate change and the increase in human activities, the frequent occurrence of forest fires has posed a serious threat to the ecological environment and human society. Existing forest fire monitoring technologies mainly rely on satellite remote sensing, drone inspections, and ground sensor networks. Although satellite remote sensing can cover a large area, it has problems such as long response time and limited resolution; drone inspections are affected by endurance and flight restrictions when covering a wide area.

[0003] In addition, traditional monitoring systems rely on centralized data processing models, which leads to high data transmission delays and makes it difficult to achieve real-time monitoring and rapid response. These defects limit the application effect and efficiency of existing technologies in actual fire monitoring. Summary of the invention

[0004] To solve the above problems, the present invention provides a forest fire monitoring method based on edge intelligence, which is used to detect, identify and warn of fires in real time in forest areas, thereby improving the timeliness and accuracy of fire response.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a forest fire monitoring method based on edge intelligence, comprising the following steps:

[0006] S1: Deploy edge nodes and sensor devices;

[0007] S2: collect data obtained by sensor equipment in real time and perform preliminary processing on it;

[0008] S3: Intelligent identification and decision-making of fire risks through edge nodes;

[0009] S4: Emergency response and incident management.

[0010] As a possible implementation, further, S1 includes the following steps:

[0011] S11: Select the edge node deployment location, including:

[0012] Use GIS system combined with historical fire data to establish fire risk assessment model and identify high-risk areas;

[0013] A grid layout scheme is adopted to optimize node distribution according to terrain characteristics and vegetation density;

[0014] Analyze signal coverage through communication simulation software to ensure the communication quality between nodes;

[0015] Consider power supply and maintenance convenience to select an appropriate installation location.

[0016] S12: Edge computing device deployment, including:

[0017] Install and configure edge computing devices and sensor networks;

[0018] Deploy multiple types of sensors, including temperature sensors, smoke sensors, and infrared thermal imagers;

[0019] Establish a hierarchical edge computing architecture: sensor node layer, data aggregation layer, edge computing layer;

[0020] Configure device network connection: support 4G / 5G, LoRa and other communication methods;

[0021] Implement environmental adaptability protection measures, including: waterproof, dustproof, and lightning protection;

[0022] S13: lightweight model deployment, wherein the lightweight model is a model optimized based on the lightweight neural network model MobileNet V3, which has improved the model and network structure;

[0023] The model improvement is specifically model pruning, which analyzes the importance of model weights, removes redundant parameters, and reduces computation and storage requirements; uses structured pruning to maintain the integrity of the model structure, and retrains the model after pruning to restore performance; model quantization: uses INT8 quantization technology to convert the model's weights and activation values ​​from floating point numbers to 8-bit integers; uses calibration data to statically quantize the model before deployment, and dynamically quantizes the activation values ​​during inference to further reduce computing resource usage; knowledge distillation: uses large models to guide the training of lightweight models and improve the performance of small models;

[0024] The specific improvements to the network structure are: adding specific fire feature extraction modules (such as dedicated channels for temperature change rate and smoke concentration characteristics); introducing a multi-task learning mechanism so that the model can handle multiple related tasks (such as fire detection and environmental anomaly analysis) at the same time.

[0025] As a possible implementation, further, S2 includes the following steps:

[0026] S21: Real-time data collection, specifically: Establish a stable and reliable data collection mechanism, which includes:

[0027] Design a multi-frequency sampling strategy: 5 minutes / time in normal state and 1 minute / time in abnormal state;

[0028] Realize data time synchronization: Use NTP protocol to ensure system clock synchronization;

[0029] Establish a data cache mechanism: local storage of 24-hour data, and support for continued transmission after network disconnection;

[0030] Realize data integrity check: CRC check ensures data accuracy;

[0031] S22: Edge data preprocessing, specifically: performing data preprocessing and preliminary analysis at the edge, which includes:

[0032] Data cleaning: remove outliers and fill in missing values;

[0033] Data standardization: unify different sensor data formats and units;

[0034] Data compression: Use lossless compression algorithm to reduce transmission bandwidth;

[0035] Preliminary analysis: Calculate key indicators to determine whether they exceed warning thresholds.

[0036] As a possible implementation, further, S3 includes the following steps:

[0037] S31: Fire risk detection, specifically: using AI algorithms to achieve intelligent identification of fire risks, including:

[0038] The key features are extracted in the feature engineering stage, including temperature change rate and smoke concentration;

[0039] Use historical data to train deep learning models to ensure recognition accuracy, and deploy lightweight models on the edge for real-time reasoning and calculations; continuously optimize model parameters based on actual application results to improve detection performance.

[0040] S32: Intelligent decision-making control, specifically: formulate intelligent response strategies based on detection and identification results, and establish a three-level warning mechanism of yellow, orange and red, corresponding to different risk levels;

[0041] Develop differentiated response strategies, initiate corresponding disposal plans at different levels, realize automated control, link alarm equipment and fire-fighting equipment, and generate decision-making suggestions for managers to assist in emergency response.

[0042] As a possible implementation method, further, the training method of the deep learning model includes the following steps:

[0043] 1) Data preparation, which includes:

[0044] Data collection: Collect multimodal data, including temperature, smoke concentration, infrared thermal imaging images, etc., combined with historical fire records and sensor equipment to collect data.

[0045] Data labeling: Label the fire area, fire level, environmental characteristics, etc. to form a high-quality training data set.

[0046] Data preprocessing: data cleaning (removing outliers, supplementing missing values), standardization (unifying data format and units) and enhancement (image rotation, cropping, brightness adjustment, etc.);

[0047] 2) Model selection, which includes:

[0048] The lightweight model MobileNet V3 is selected as the basic model because of its high computational efficiency and suitability for edge device deployment. According to the fire feature requirements, key features are extracted, including temperature change rate, smoke concentration gradient, and hot spot area ratio.

[0049] 3) Model training, which includes:

[0050] Training method: Use transfer learning, load pre-trained weights (ImageNet dataset), and fine-tune on the fire dataset;

[0051] Loss function: Use a combination of classification loss (cross entropy) and regression loss (mean square error) to optimize the classification and positioning capabilities of fire detection;

[0052] Optimization algorithm: Use the Adam optimizer, set the initial learning rate to 0.001, and adopt the learning rate decay strategy;

[0053] Training strategy: Dynamically adjust sampling weights, balance the ratio of fire and non-fire samples, and use data enhancement technology to improve the generalization ability of the model;

[0054] 4) Model validation and optimization, which includes:

[0055] Evaluate model performance on the validation set, using metrics such as accuracy, recall, F1 score, etc.

[0056] The model is optimized through pruning and quantization (INT8) to reduce computing resource usage and ensure that the inference delay is controlled within 50ms.

[0057] As a possible implementation method, further, an intelligent response strategy is formulated based on the detection and identification results, and a three-level warning mechanism of yellow, orange and red is established, corresponding to different risk levels, specifically:

[0058] Build a rule-based rapid decision-making mechanism, which includes:

[0059] Set multi-level trigger thresholds to achieve warning within seconds;

[0060] Dynamically adjust judgment parameters according to environmental factors and output standardized early warning information;

[0061] Establish feature combination judgment logic to reduce the false alarm rate; the combination of feature combinations includes:

[0062] Key features extracted: temperature change rate, temperature acceleration, smoke concentration gradient change, and hot spot area ratio; the area ratio of the hot spot area is detected by infrared thermal imaging to determine the fire scale, wind direction and speed, and environmental background characteristics;

[0063] Feature combination method: Multimodal fusion: Fusion of feature data collected by different sensors to form a multi-dimensional feature vector; Time series analysis: Modeling of feature data in time series to analyze the changing trend of features over time; Feature weight allocation: Different weights are allocated according to the importance of key features of fire occurrence; Feature interaction combination: New combination features are constructed through feature interaction analysis;

[0064] The judgment logic is specifically established by: using a rule engine or a machine learning model to perform intelligent identification and graded warning of fire risks; establishing rule-based judgment logic, machine learning-based judgment logic, and risk grading and response logic;

[0065] Rule-based judgment logic, including: setting multi-level trigger thresholds (such as temperature change rate > 5°C / minute; smoke concentration gradient change > 10ppm / minute; hot spot area ratio > 20%) and dynamically adjusting the judgment thresholds according to environmental background characteristics (such as humidity, vegetation density); when a single feature exceeds the threshold, a preliminary warning is triggered;

[0066] The judgment logic based on machine learning includes: inputting multimodal feature vectors into lightweight deep learning models to classify and predict fire risks; the model outputs fire risk levels (low risk, medium risk, high risk) or fire probability; reducing false alarm rates through interactive logic of feature combinations; and using new data to conduct online learning and parameter optimization of the model based on actual application results to improve judgment accuracy;

[0067] Risk classification and response logic, specifically: three-level early warning mechanism:

[0068] Yellow warning indicates low risk: a single feature is close to the threshold, but the combined condition is not met, and the system records and continuously monitors it;

[0069] Orange warning represents medium risk: the combination of multiple features meets the medium risk condition, triggering an alarm and notifying the management personnel;

[0070] Red warning indicates high risk: if multiple features meet the high risk condition, the system will automatically link the fire extinguishing equipment and start the emergency plan;

[0071] Prioritize high-weight features (such as temperature change rate and smoke concentration gradient), and make a comprehensive judgment in combination with low-weight features (such as humidity and wind speed).

[0072] As a possible implementation, further, S4 includes the following steps:

[0073] S41: Intelligent emergency response, specifically: building an intelligent emergency response system, establishing a forest fire emergency plan library based on knowledge graph technology, and automatically matching the optimal plan according to the fire level and environmental conditions; establishing a cross-departmental collaborative response platform to achieve information sharing and joint response among firefighting, forestry, emergency and other departments; using blockchain technology to record the response process to ensure that the entire emergency response process is traceable; integrating intelligent equipment such as drones and robots to enhance emergency response capabilities;

[0074] S42: Event analysis and experience learning, specifically: establish a systematic event analysis and continuous improvement mechanism; use big data analysis technology to conduct multi-dimensional analysis of fire events, evaluate the response effect of the early warning system, mine event characteristics through machine learning methods, and optimize early warning models and handling processes; establish a standardized experience review mechanism to form a complete case library; regularly organize technical seminars to promote the continuous evolution and upgrading of the system.

[0075] As a possible implementation method, further, the forest fire emergency plan library includes fire level classification standards: according to the severity of the fire, the speed of fire spread, the scope of the affected area, etc., the fire situation is divided into different levels (yellow warning, orange warning, red warning), and each level corresponds to a specific response strategy and resource scheduling plan; environmental condition data, which includes environmental parameters such as humidity, temperature, wind speed, wind direction, vegetation density, and terrain characteristics; emergency resource information, which includes equipment resources, human resources, and material reserves; fire handling process; cross-departmental coordination mechanism; historical cases and experience library; dynamic adjustment rules;

[0076] The dynamic adjustment rules dynamically adjust the contents of the plan according to the fire level and environmental conditions, and include:

[0077] Humidity: When humidity is low, fire spreads faster and more resources need to be mobilized first;

[0078] Wind speed and direction: When the wind speed is high and the wind direction is toward high-risk areas, the deployment location of the fire-fighting equipment needs to be adjusted;

[0079] Vegetation density: Areas with high vegetation density have a greater risk of fire and require priority dispatch of firefighting resources;

[0080] Terrain features: In mountainous or complex terrains, it is necessary to give priority to the deployment of intelligent devices such as drones or robots;

[0081] If the fire level is orange warning and the wind speed is greater than the threshold, the drone monitoring plan will be given priority;

[0082] If the fire level is red warning and the humidity is less than the threshold, the dispatch amount of fire extinguishing agent will be increased;

[0083] The system continuously updates the knowledge graph based on real-time data and optimizes the matching logic of the rule engine. It uses machine learning technology to mine successful experiences from historical cases and improve the intelligence level of plan matching.

[0084] As a possible implementation method, big data analysis technology is further used to conduct multi-dimensional analysis of fire events to evaluate the response effect of the early warning system; the analysis method specifically includes:

[0085] A) Multi-dimensional analysis, including:

[0086] Time series analysis: analyzing the changing trends of key features before and after a fire: temperature change rate, smoke concentration gradient; identifying the key time points of a fire: the initial stage of a fire, the fire spread stage, and the fire extinguishing stage;

[0087] Spatial distribution analysis: using GIS mapping technology to analyze the geographical characteristics of the fire area: distribution of high-risk areas, fire spread paths, etc.; combining the area proportion of hot spots and wind direction to evaluate the spatial pattern of fire spread;

[0088] Feature correlation analysis: using data mining technology to analyze the correlation between different features (such as the relationship between temperature change rate and smoke concentration gradient; identifying key triggering factors for fire);

[0089] Emergency response analysis, evaluate the dispatch efficiency of emergency resources: equipment arrival time, fire extinguishing equipment start-up time, etc.; analyze delays or unreasonable resource allocation problems in the emergency response process;

[0090] B) Model optimization analysis, which includes:

[0091] Model performance evaluation: Use fire event data to verify the accuracy and false alarm rate of the early warning model; evaluate the response speed of the model by comparing the actual fire occurrence time with the model prediction time;

[0092] Feature importance analysis: Use machine learning models (random forest, XGBoost) to calculate the contribution of each feature to fire prediction; optimize feature selection and improve model performance;

[0093] Using big data analysis technology, we conduct multi-dimensional analysis of fire events and evaluate the response effect of the early warning system. The evaluation methods include:

[0094] a) Detection performance evaluation, which includes:

[0095] Accuracy: Calculate the accuracy of fire risk detection, that is, the proportion of correctly detected fires; the calculation formula is: Accuracy = (TP+TN) / (TP+TN+FP+FN)

[0096] TP: the number of correctly detected fires; TN: the number of correctly detected non-fires; FP: the number of false fire alarms; FN: the number of missed fire alarms;

[0097] Recall rate: evaluates the system's ability to capture fire events and avoid underreporting; its calculation formula is: Recall rate = TP / (TP+FN);

[0098] False Positive Rate: Evaluate the false alarm situation of the system and reduce the false alarm rate to reduce unnecessary resource waste; the calculation formula is: False Positive Rate = FP / (FP+TN);

[0099] b) Response speed assessment, which includes:

[0100] Warning time: Calculate the time interval from the occurrence of fire to the issuance of warning by the system. Keep the warning time within the shortest range.

[0101] Emergency response time: Evaluate the time from when an early warning is issued to when emergency resources arrive at the scene. Optimize resource scheduling processes and shorten response time.

[0102] c) Resource scheduling efficiency evaluation, which includes:

[0103] Equipment utilization: Statistics on the use of resources such as fire-fighting equipment, drones, and sensors to improve resource utilization and avoid resource waste;

[0104] Reasonable resource allocation: Analyze whether resource scheduling matches the fire level and environmental conditions, optimize resource allocation strategies, and ensure efficient response;

[0105] Through machine learning methods, event characteristics are mined to optimize early warning models and handling processes. Among them, the optimization methods based on event characteristics include:

[0106] Key feature extraction, including: temperature change rate: analyze the temperature change per unit time and extract the temperature rise characteristics at the initial stage of the fire; smoke concentration gradient: calculate the rate of change of smoke concentration and identify potential fire sources; hot spot area ratio: use infrared thermal imagers to detect the area ratio of hot spots and assess the scale of the fire; environmental background characteristics: including humidity, wind speed, wind direction, vegetation density, etc., used to correct fire risk assessment; construct composite features through feature combination (such as the interaction between temperature change rate and smoke concentration gradient) to enhance the model's ability to identify fire risks; use time series analysis technology to extract dynamic change trends of features (such as the temperature change rate in the past 5 minutes);

[0107] Model optimization, which includes: model training and updating: using historical fire data and real-time collected data, regularly updating model weights to ensure that the model adapts to new environmental characteristics; using transfer learning methods to fine-tune the pre-trained model to accelerate model convergence; false alarm and missed alarm optimization: adjusting the decision threshold of the model through feature importance analysis to reduce false alarm and missed alarm rates; building rule-based auxiliary judgment logic (such as dynamically adjusting the warning threshold based on environmental humidity and wind speed); dynamic adjustment strategy: dynamically adjusting model parameters (such as learning rate, regularization parameters) according to real-time data to improve the adaptability of the model; in abnormal conditions (such as in the early stages of a fire), increasing the data sampling frequency (from 5 minutes / time to 1 minute / time) to enhance the real-time performance of the model;

[0108] Process optimization, including: classifying fire conditions into three levels of warning: yellow, orange and red, based on event characteristics (such as temperature change rate, smoke concentration gradient, hotspot area ratio, etc.); different fire levels correspond to different response strategies:

[0109] Yellow warning: mainly monitoring, recording abnormal data;

[0110] Orange warning: dispatch some resources for preliminary disposal;

[0111] Red alert: fully activate emergency response plans and mobilize all available resources;

[0112] Combine the fire level and environmental conditions (such as humidity, wind speed, and vegetation density) to automatically match the optimal emergency plan; dynamically adjust the plan content (such as resource allocation priority and equipment scheduling sequence) to ensure the applicability of the plan; use event characteristics (such as fire spread direction and hot spot area distribution) to optimize resource scheduling strategies;

[0113] Post-event review: Review the entire process of the fire incident and analyze the characteristic changes at key time nodes (such as the initial stage of the fire, the fire spread stage, and the fire extinguishing stage); evaluate the detection effect of the early warning model (such as accuracy, recall rate, and false alarm rate); evaluate the execution efficiency of the disposal process (such as resource scheduling time, fire extinguishing equipment startup time); identify deficiencies in the early warning model and disposal process (such as high false alarm rate, response delay, etc.);

[0114] Continuous optimization, which includes: using event review data to optimize the model's feature selection and parameter settings; introducing new data features (such as smoke diffusion characteristics in the later stages of a fire) to improve the model's predictive capabilities; optimizing emergency plan content and resource scheduling strategies based on review results; dynamically adjusting warning thresholds to improve system adaptability; organizing fire events into standardized cases, recording fire types, triggering factors, handling processes and results; and using machine learning technology to mine common features in the case library to optimize warning models and handling processes.

[0115] As a possible implementation mode, further, the forest fire monitoring method based on edge intelligence also includes user interaction and feedback, specifically including the following steps:

[0116] Real-time monitoring interface, providing an intuitive visual monitoring platform; real-time display of monitoring data and analysis results through data visualization technology; integrated GIS map function, intuitive display of equipment location distribution and fire situation; multi-terminal adaptation, support PC and mobile terminal access at any time; authority management according to user roles to ensure information security;

[0117] The manual adjustment function provides flexible system parameter configuration functions; supports dynamic adjustment of various warning thresholds to adapt to different monitoring needs; allows modification of data collection frequency to optimize system resource utilization; can customize alarm rules and notification methods to meet personalized needs; realizes remote configuration and restart of equipment to improve operation and maintenance efficiency.

[0118] The present invention aims to optimize the architecture of the forest fire monitoring system by introducing edge intelligence technology, realize local processing and real-time analysis of data, and thus improve the response speed and detection accuracy of fire monitoring.

[0119] The beneficial effects of the present invention are:

[0120] 1) Real-time data processing - edge computing: By deploying edge computing devices in the forest, the present invention can perform real-time data processing near the data source, significantly reducing data transmission delays. This makes the detection and response of fire risks faster, and measures can be taken in time to reduce fire losses.

[0121] 2) Efficient fire risk detection - machine learning algorithm: The present invention adopts a lightweight convolutional neural network model for fire risk detection, which can more accurately identify environmental anomalies (such as high temperature, smoke, etc.); compared with the traditional threshold detection method, the machine learning algorithm can adapt to complex environmental changes and improve the accuracy and reliability of detection.

[0122] 3) Intelligent decision support - automated decision making: The present invention can intelligently decide whether to activate fire-fighting equipment or sound an alarm based on real-time analysis results, reducing the need for human intervention and improving the efficiency of responding to fires; this intelligent decision support system can respond quickly in emergency situations and reduce the risk of human error.

[0123] 4) User-friendly interactive interface - real-time monitoring and manual adjustment: Through the user APP, users can view forest environment data and monitoring images in real time, and have manual adjustment functions. This interactive design enables users to flexibly manage the monitoring system, enhancing the system's operability and user experience.

[0124] 5) Resource optimization configuration - dynamic regulation: The present invention can dynamically adjust monitoring parameters and equipment settings according to real-time data to optimize resource configuration; this flexibility enables the system to maintain efficient operation under different environmental conditions, thereby improving the overall monitoring effect.

[0125] 6) Reduced network dependence - local processing power: Since edge computing devices have a certain computing power, the system can continue to work even when the network is unstable or interrupted, ensuring the continuity of data processing and fire monitoring; this feature is particularly important in remote areas and can effectively deal with the problem of insufficient network coverage.

[0126] 7) Complete emergency response mechanism: The edge has 24-hour data storage and independent alarm push capabilities; supports on-site activation of emergency plans without relying on cloud-based decision-making; records the disposal process through blockchain technology to ensure traceability; integrates multi-party resources to achieve cross-departmental collaborative response. BRIEF DESCRIPTION OF THE DRAWINGS

[0127] Figure 1 It is a simplified flow chart of the present invention. DETAILED DESCRIPTION

[0128] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0129] See attached Figure 1 As shown, this embodiment provides a forest fire monitoring method based on edge intelligence, comprising the following steps:

[0130] S1: Deploy edge nodes and sensor devices; the system deployment steps include:

[0131] S11: Select edge node deployment location

[0132] Rationally lay out the monitoring system according to the characteristics of the forest area to achieve optimal monitoring coverage. Use the GIS system combined with historical fire data to establish a fire risk assessment model to identify high-risk areas. Adopt a grid layout solution to optimize node distribution based on terrain characteristics and vegetation density. Analyze signal coverage through communication simulation software to ensure the quality of communication between nodes. Consider power supply and maintenance convenience to select an appropriate installation location.

[0133] S12: Edge computing device deployment

[0134] Install and configure edge computing devices and sensor networks. Deploy multiple types of sensors: temperature sensors (accuracy ±0.5℃), smoke sensors (adjustable sensitivity), infrared thermal imagers, etc. Establish a hierarchical edge computing architecture: sensor node layer, data aggregation layer, edge computing layer. Configure device network connection: support multiple communication methods such as 4G / 5G and LoRa. Furthermore, different communication methods can be replaced according to different usage scenarios, and NB-IoT (narrowband Internet of Things) can be used instead of LoRa communication method, which is particularly suitable for low-power and long-distance sensor networks. Or introduce Wi-SUN (wireless smart utility network) technology to support a wider range of device connections and efficient communication. Implement protective measures: environmental adaptability protection such as waterproof, dustproof, and lightning protection.

[0135] S13: Lightweight model deployment

[0136] The model optimized based on the existing lightweight neural network model (MobileNet V3) is adopted to ensure its high computing efficiency and low resource consumption. Among them, the model and network structure have been improved, including:

[0137] Model pruning: By analyzing the importance of model weights, redundant parameters are removed to reduce computational and storage requirements. Structured pruning (such as channel pruning) is used to maintain the integrity of the model structure. After pruning, the model is retrained to restore performance; Model quantization: Using INT8 quantization technology, the model's weights and activation values ​​are converted from floating point numbers (FP32) to 8-bit integers. The model is statically quantized using calibration data before deployment. The activation values ​​are dynamically quantized during the inference process to further reduce computing resource usage; Knowledge distillation: The training of lightweight models is guided by large models (ResNet and EfficientNet) to improve the performance of small models.

[0138] Improvements to the network structure: Adding specific fire feature extraction modules (such as dedicated channels for temperature change rate and smoke concentration characteristics); introducing a multi-task learning mechanism so that the model can handle multiple related tasks (such as fire detection and environmental anomaly analysis) at the same time.

[0139] S2: real-time collection of data obtained by the sensor equipment and preliminary processing of the data, including the following steps:

[0140] S21: Real-time data collection

[0141] Establish a stable and reliable data collection mechanism. Design a multi-frequency sampling strategy: 5 minutes / time in normal state, and 1 minute / time in abnormal state. Realize data time synchronization: Use NTP protocol to ensure system clock synchronization. Establish a data cache mechanism: Locally store 24 hours of data and support continuous transmission when the network is disconnected. Implement data integrity verification: CRC verification ensures data accuracy.

[0142] S22: Edge Data Preprocessing

[0143] Data preprocessing and preliminary analysis are performed at the edge. Data cleaning: remove outliers and supplement missing values. Data standardization: unify the data formats and units of different sensors. Data compression: use lossless compression algorithms to reduce transmission bandwidth. Preliminary analysis: calculate key indicators and determine whether they exceed the warning threshold.

[0144] S3: Intelligent identification and decision-making of fire risks through edge nodes, including the following steps:

[0145] S31: Fire risk detection

[0146] Use AI algorithms to achieve intelligent identification of fire risks. Extract key features such as temperature change rate and smoke concentration during the feature engineering stage. Use historical data to train deep learning models to ensure identification accuracy. Deploy lightweight models on the edge for real-time reasoning and calculation. Continuously optimize model parameters based on actual application results to improve detection performance.

[0147] Among them, the training method of the deep learning model includes the following steps:

[0148] 1) Data preparation

[0149] Data collection: Collect multimodal data, including temperature, smoke concentration, infrared thermal imaging images, etc., combined with historical fire records and sensor data.

[0150] Data labeling: Label the fire area, fire level, environmental characteristics, etc. to form a high-quality training data set.

[0151] Data preprocessing: data cleaning (removing outliers and supplementing missing values), standardization (unifying data format and units) and enhancement (image rotation, cropping, brightness adjustment, etc.).

[0152] 2) Model selection

[0153] The lightweight model MobileNet V3 is selected as the basic model because of its high computational efficiency and suitability for edge device deployment. According to the fire feature requirements, key features (such as temperature change rate, smoke concentration gradient, hot spot area ratio, etc.) are extracted. Furthermore, other lightweight models (such as Tiny-YOLO, SqueezeNet) can be used to replace MobileNet V3, and a more suitable model architecture can be selected according to the actual scenario. The knowledge distillation technology is introduced to guide the training of small models through large models to further improve the performance of lightweight models.

[0154] 3) Model training

[0155] Training method: Use transfer learning, load pre-trained weights (ImageNet dataset), and fine-tune on the fire dataset.

[0156] Loss function: A combination of classification loss (cross entropy) and regression loss (mean square error) is used to optimize the classification and positioning capabilities of fire detection.

[0157] Optimization algorithm: Use the Adam optimizer, set the initial learning rate to 0.001, and adopt the learning rate decay strategy.

[0158] Training strategy: Dynamically adjust sampling weights to balance the ratio of fire and non-fire samples. Use data enhancement technology to improve the generalization ability of the model.

[0159] 4) Model verification and optimization

[0160] Evaluate model performance on the validation set using metrics such as precision, recall, F1 score, etc.

[0161] The model is optimized through pruning and quantization (INT8) to reduce computing resource usage and ensure that the inference delay is controlled within 50ms.

[0162] S32: Fire risk detection

[0163] Develop intelligent response strategies based on analysis results. Establish a three-level warning mechanism of yellow, orange, and red, corresponding to different risk levels. Develop differentiated response strategies and initiate corresponding disposal plans at different levels. Implement automated control and link alarm equipment and fire extinguishing equipment. Generate decision suggestions for managers to assist in emergency response. The decision rule engine is specifically: build a rule-based rapid decision mechanism, which includes: setting multi-level trigger thresholds to achieve second-level warnings; dynamically adjusting judgment parameters according to environmental factors; outputting standardized warning information;

[0164] Establish feature combination judgment logic to reduce the false alarm rate; the combination of feature combinations includes: extracting key features: temperature change rate, temperature acceleration, smoke concentration gradient change, hot spot area ratio: using infrared thermal imagers to detect the area ratio of hot spots, judge the scale of the fire, wind direction and speed, and environmental background characteristics.

[0165] Feature combination methods: Multimodal fusion: Fusion of feature data collected by different sensors to form a multi-dimensional feature vector; Time series analysis: Time series modeling of feature data to analyze the changing trend of features over time; Feature weight allocation: Different weights are allocated according to the importance of key features of fire occurrence; Feature interaction combination: New combination features are constructed through feature interaction analysis.

[0166] The judgment logic is established by using a rule engine or machine learning model to intelligently identify fire risks and provide graded warnings. Establish rule-based judgment logic, machine learning-based judgment logic, and risk grading and response logic.

[0167] Rule-based judgment logic:

[0168] Set multiple trigger thresholds (such as temperature change rate > 5℃ / minute; smoke concentration gradient change

[0169] >10ppm / minute; hotspot area ratio>20%. ) and dynamically adjust the judgment threshold according to environmental background characteristics (such as humidity, vegetation density). When a single feature exceeds the threshold, a preliminary warning is triggered.

[0170] Judgment logic based on machine learning:

[0171] Input the multimodal feature vector into the lightweight deep learning model to classify and predict fire risks. The model outputs the fire risk level (low risk, medium risk, high risk) or the probability of fire. Reduce the false alarm rate through the interactive logic of feature combination. According to the actual application effect, use new data to perform online learning and parameter optimization on the model to improve the judgment accuracy.

[0172] Risk classification and response logic: three-level early warning mechanism:

[0173] Yellow warning (low risk): A single feature is close to the threshold, but the combination condition is not met. The system records and continuously monitors it.

[0174] Orange warning (medium risk): The combination of multiple features meets the medium risk condition, triggering an alarm and notifying the management personnel.

[0175] Red warning (high risk): When the combination of multiple features meets the high-risk conditions, the system automatically links the fire-fighting equipment and activates the emergency plan.

[0176] Prioritize high-weight features (such as temperature change rate and smoke concentration gradient), and make a comprehensive judgment in combination with low-weight features (such as humidity and wind speed).

[0177] S4: Emergency response and incident management, including the following steps:

[0178] S41: Intelligent emergency response

[0179] Build an intelligent emergency response system. Establish a forest fire emergency plan library based on knowledge graph technology, and automatically match the optimal plan according to the fire level and environmental conditions. Establish a cross-departmental collaborative disposal platform to achieve information sharing and joint response among firefighting, forestry, emergency and other departments. Use blockchain technology to record the disposal process to ensure that the entire emergency disposal process is traceable. Integrate intelligent equipment such as drones and robots to improve emergency response capabilities.

[0180] Among them, the forest fire emergency plan library includes fire level classification standards: according to the severity of the fire, the speed of fire spread, the scope of the affected area, etc., the fire situation is divided into different levels (yellow warning, orange warning, red warning); each level corresponds to a specific response strategy and resource scheduling plan; environmental condition data (humidity, temperature, wind speed, wind direction, vegetation density, terrain characteristics and other environmental parameters); emergency resource information (equipment resources, human resources, material reserves); fire handling process; cross-departmental coordination mechanism; historical case and experience library; dynamic adjustment rules.

[0181] Among them, the dynamic adjustment rules dynamically adjust the contents of the plan according to the fire level and environmental conditions, including:

[0182] Humidity: When humidity is low, fire spreads faster and more resources need to be mobilized as a priority.

[0183] Wind speed and direction: When the wind speed is high and the wind direction is pointing towards high-risk areas, the deployment location of the fire-fighting equipment needs to be adjusted.

[0184] Vegetation density: Areas with high vegetation density have a greater risk of fire and fire-fighting resources need to be prioritized.

[0185] Terrain features: In mountainous or complex terrains, it is necessary to deploy drones, robots and other intelligent devices first:

[0186] If the fire level is orange warning and the wind speed is greater than the threshold, the drone monitoring plan will be given priority.

[0187] If the fire level is red warning and the humidity is less than the threshold, the dispatch amount of fire extinguishing agent will be increased.

[0188] The system continuously updates the knowledge graph based on real-time data and optimizes the matching logic of the rule engine. It also uses machine learning technology to mine successful experiences from historical cases and improve the intelligence level of plan matching.

[0189] S42: Event Analysis and Experience Learning

[0190] Establish a systematic event analysis and continuous improvement mechanism. Use big data analysis technology to conduct multi-dimensional analysis of fire events, evaluate the response effect of the early warning system, and use machine learning methods to mine event characteristics, optimize early warning models and disposal processes; establish a standardized experience review mechanism to form a complete case library. Organize technical seminars regularly to promote the continuous evolution and upgrade of the system.

[0191] In this embodiment, big data analysis technology is used to perform multi-dimensional analysis on fire events; the analysis method specifically includes:

[0192] Multi-dimensional analysis:

[0193] Time series analysis (analyzing the changing trends of key features before and after a fire: temperature change rate, smoke concentration gradient; identifying key time points for a fire: the initial stage of a fire, the fire spread stage, and the extinguishing stage); spatial distribution analysis (using GIS map technology to analyze the geographical characteristics of the fire area: the distribution of high-risk areas, the path of fire spread, etc.; combining the area share of hot spots and wind direction to evaluate the spatial pattern of fire spread); feature association analysis (using data mining technology to analyze the correlation between different features: such as the relationship between the temperature change rate and the smoke concentration gradient; identifying key triggering factors for a fire); emergency response analysis (evaluating the dispatch efficiency of emergency resources: equipment arrival time, fire extinguishing equipment start-up time, etc.; analyzing delays or unreasonable resource allocation in the emergency response process)

[0194] Model optimization analysis:

[0195] Model performance evaluation: Use fire event data to verify the accuracy and false alarm rate of the early warning model; evaluate the response speed of the model by comparing the actual fire occurrence time with the model prediction time.

[0196] Feature importance analysis: Use machine learning models (random forest, XGBoost) to calculate the contribution of each feature to fire prediction; optimize feature selection and improve model performance.

[0197] In this embodiment, big data analysis technology is used to conduct multi-dimensional analysis of fire events to evaluate the response effect of the early warning system; the evaluation method specifically includes:

[0198] Detection performance evaluation, which includes:

[0199] Accuracy: Calculate the accuracy of fire risk detection, that is, the proportion of correctly detected fires; the calculation formula is: Accuracy = (TP+TN) / (TP+TN+FP+FN)

[0200] TP: the number of correctly detected fires; TN: the number of correctly detected non-fires; FP: the number of false fire alarms; FN: the number of missed fire alarms;

[0201] Recall rate: evaluates the system's ability to capture fire events and avoid underreporting; its calculation formula is: Recall rate = TP / (TP+FN);

[0202] False Positive Rate: Evaluate the false alarm situation of the system and reduce the false alarm rate to reduce unnecessary resource waste; the calculation formula is: False Positive Rate = FP / (FP+TN);

[0203] Responsiveness assessment, which includes:

[0204] Warning time: Calculate the time interval from the occurrence of fire to the issuance of warning by the system. Keep the warning time within the shortest range.

[0205] Emergency response time: Evaluate the time from when an early warning is issued to when emergency resources arrive at the scene. Optimize resource scheduling processes and shorten response time.

[0206] Resource scheduling efficiency evaluation, which includes:

[0207] Equipment utilization: Statistics on the usage of resources such as fire-fighting equipment, drones, and sensors to improve resource utilization and avoid resource waste.

[0208] Reasonable resource allocation: Analyze whether resource scheduling matches the fire level and environmental conditions, optimize resource allocation strategies, and ensure efficient response.

[0209] In this embodiment, the event characteristics are mined through machine learning methods to optimize the early warning model and the handling process; the optimization method based on the event characteristics specifically includes:

[0210] I) Key feature extraction:

[0211] Temperature change rate: Analyze the temperature change per unit time and extract the temperature rise characteristics at the initial stage of fire.

[0212] Smoke Gradient: Calculates the rate of change of smoke concentration and identifies potential fire sources.

[0213] Hot spot area ratio: Use infrared thermal imaging cameras to detect the hot spot area ratio and assess the scale of the fire.

[0214] Environmental background characteristics: including humidity, wind speed, wind direction, vegetation density, etc., used to correct fire risk assessment.

[0215] By combining features (such as the interaction between the temperature change rate and the smoke concentration gradient), composite features are constructed to improve the model's ability to identify fire risks. Time series analysis technology is used to extract the dynamic change trend of features (such as the temperature change rate in the past 5 minutes).

[0216] II) Model Optimization:

[0217] Model training and updating: Use historical fire data and real-time collected data to regularly update model weights to ensure that the model adapts to new environmental characteristics; use transfer learning methods to fine-tune the pre-trained model to accelerate model convergence.

[0218] Optimization of false alarms and missed alarms: Through feature importance analysis, adjust the decision threshold of the model to reduce the false alarm rate and missed alarm rate; build rule-based auxiliary judgment logic (such as dynamically adjusting the warning threshold based on environmental humidity and wind speed)

[0219] Dynamic adjustment strategy: Dynamically adjust model parameters (such as learning rate and regularization parameters) according to real-time data to improve the adaptability of the model; in abnormal conditions (such as the early stage of a fire), increase the data sampling frequency (from 5 minutes / time to 1 minute / time) to enhance the real-time performance of the model.

[0220] III) Process Optimization:

[0221] Based on the characteristics of the event (such as temperature change rate, smoke concentration gradient, hot spot area ratio, etc.), the fire situation is divided into three levels of warning: yellow, orange and red.

[0222] Different fire levels correspond to different response strategies:

[0223] Yellow warning: Mainly monitor and record abnormal data.

[0224] Orange alert: Dispatch some resources for preliminary disposal.

[0225] Red alert: fully activate emergency plans and mobilize all available resources.

[0226] Combine the fire level and environmental conditions (such as humidity, wind speed, and vegetation density) to automatically match the optimal emergency plan; dynamically adjust the plan content (such as resource allocation priority and equipment scheduling sequence) to ensure the applicability of the plan. Use event characteristics (such as the direction of fire spread and the distribution of hot spots) to optimize resource scheduling strategies.

[0227] Post-event review: Review the entire process of the fire incident and analyze the characteristic changes at key time nodes (such as the initial stage of the fire, the fire spread stage, and the fire extinguishing stage); evaluate the detection effect of the early warning model (such as accuracy, recall rate, and false alarm rate); evaluate the execution efficiency of the disposal process (such as resource scheduling time, fire extinguishing equipment startup time); identify deficiencies in the early warning model and disposal process (such as high false alarm rate, response delay, etc.).

[0228] IV) Continuous Optimization:

[0229] Utilize event review data to optimize the model’s feature selection and parameter settings; introduce new data features (such as smoke diffusion characteristics in the later stages of a fire) to improve the model’s predictive capabilities; optimize emergency plan content and resource scheduling strategies based on review results; dynamically adjust warning thresholds to improve system adaptability; organize fire events into standardized cases, record fire types, triggering factors, handling procedures, and results; use machine learning technology to mine common features in the case library to optimize warning models and handling procedures.

[0230] S5: User interaction and feedback, including the following steps:

[0231] S51: Real-time monitoring interface

[0232] Provides an intuitive visual monitoring platform. Real-time display of monitoring data and analysis results through data visualization technology. Integrates GIS map function to intuitively display equipment location distribution and fire situation. Achieve multi-terminal adaptation and support PC and mobile terminal access at any time. Manage permissions according to user roles to ensure information security.

[0233] S52: Manual adjustment function

[0234] Provides flexible system parameter configuration functions. Supports dynamic adjustment of various warning thresholds to meet different monitoring needs. Allows modification of data collection frequency to optimize system resource utilization. Customizable alarm rules and notification methods to meet personalized needs. Realizes remote configuration and restart of equipment to improve operation and maintenance efficiency.

[0235] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A forest fire monitoring method based on edge intelligence, characterized in that: The steps include: S1: Deploy edge nodes and sensor devices; S2: collect data obtained by sensor equipment in real time and perform preliminary processing on it; S3: Intelligent identification and decision-making of fire risks through edge nodes; S4: Emergency response and incident management.

2. The forest fire monitoring method based on edge intelligence according to claim 1 is characterized in that: S1 includes the following steps: S11: Select the edge node deployment location, including: Use GIS system combined with historical fire data to establish fire risk assessment model and identify high-risk areas; A grid layout scheme is adopted to optimize node distribution according to terrain characteristics and vegetation density; Analyze signal coverage through communication simulation software to ensure the communication quality between nodes; Consider power supply and maintenance convenience and choose an appropriate installation location; S12: Edge computing equipment deployment, including: Install and configure edge computing devices and sensor networks; Deploy multiple types of sensors, including temperature sensors, smoke sensors, and infrared thermal imagers; Establish a hierarchical edge computing architecture: sensor node layer, data aggregation layer, edge computing layer; Configure device network connection: support 4G / 5G, LoRa and other communication methods; Implement environmental adaptability protection measures, including: waterproof, dustproof, and lightning protection; S13: lightweight model deployment, wherein the lightweight model is a model optimized based on the lightweight neural network model MobileNet V3, which has improved the model and network structure; The model improvement is specifically model pruning, which analyzes the importance of model weights, removes redundant parameters, and reduces computation and storage requirements; uses structured pruning to maintain the integrity of the model structure, and retrains the model after pruning to restore performance; model quantization: uses INT8 quantization technology to convert the model's weights and activation values ​​from floating point numbers to 8-bit integers; uses calibration data to statically quantize the model before deployment, and dynamically quantizes the activation values ​​during inference to further reduce computing resource usage; knowledge distillation: uses large models to guide the training of lightweight models and improve the performance of small models; The specific improvements to the network structure are: adding a specific fire feature extraction module; introducing a multi-task learning mechanism so that the model can handle multiple related tasks at the same time.

3. The forest fire monitoring method based on edge intelligence according to claim 1 is characterized in that: S2 includes the following steps: S21: Real-time data collection, specifically: Establish a stable and reliable data collection mechanism, which includes: Design a multi-frequency sampling strategy: 5 minutes / time in normal state and 1 minute / time in abnormal state; Realize data time synchronization: Use NTP protocol to ensure system clock synchronization; Establish a data cache mechanism: local storage of 24-hour data, and support for continued transmission after network disconnection; Realize data integrity check: CRC check ensures data accuracy; S22: Edge data preprocessing, specifically: performing data preprocessing and preliminary analysis at the edge, which includes: Data cleaning: remove outliers and fill in missing values; Data standardization: unify different sensor data formats and units; Data compression: Use lossless compression algorithm to reduce transmission bandwidth; Preliminary analysis: Calculate key indicators to determine whether they exceed warning thresholds.

4. The forest fire monitoring method based on edge intelligence according to claim 1 is characterized in that: S3 includes the following steps: S31: Fire risk detection, specifically: using AI algorithms to achieve intelligent identification of fire risks, including: The key features are extracted in the feature engineering stage, including temperature change rate and smoke concentration; Use historical data to train deep learning models to ensure recognition accuracy, and deploy lightweight models on the edge for real-time reasoning and calculation; continuously optimize model parameters based on actual application results to improve detection performance; S32: Intelligent decision-making control, specifically: formulate intelligent response strategies based on detection and identification results, and establish a three-level warning mechanism of yellow, orange and red, corresponding to different risk levels; Develop differentiated response strategies, initiate corresponding disposal plans at different levels, realize automated control, link alarm equipment and fire-fighting equipment, and generate decision-making suggestions for managers to assist in emergency response.

5. The forest fire monitoring method based on edge intelligence according to claim 4 is characterized in that: The training method of deep learning model includes the following steps: 1) Data preparation, which includes: Data collection: Collect multimodal data, including temperature, smoke concentration, infrared thermal imaging images, etc., combined with historical fire records and sensor equipment to collect data; Data annotation: Annotate the fire area, fire level, and environmental characteristics to form a high-quality training data set; Data preprocessing: data cleaning, standardization and enhancement; 2) Model selection, which includes: The lightweight model MobileNet V3 is selected as the basic model. According to the fire feature requirements, key features are extracted, including temperature change rate, smoke concentration gradient, and hot spot area ratio. 3) Model training, which includes: Training method: Use transfer learning, load pre-trained weights, and fine-tune on the fire dataset; Loss function: Use a combination of classification loss and regression loss to optimize the classification and positioning capabilities of fire detection; Optimization algorithm: Use the Adam optimizer, set the initial learning rate to 0.001, and adopt the learning rate decay strategy; Training strategy: Dynamically adjust sampling weights, balance the ratio of fire and non-fire samples, and use data enhancement technology to improve the generalization ability of the model; 4) Model validation and optimization, which includes: Evaluate model performance on the validation set using metrics such as precision, recall, and F1 score; By optimizing the model through pruning and quantization, computing resources are reduced and the inference delay is kept within 50ms.

6. The forest fire monitoring method based on edge intelligence according to claim 4 is characterized in that: Formulate intelligent response strategies based on detection and identification results, and establish a three-level warning mechanism of yellow, orange and red, corresponding to different risk levels, specifically: Build a rule-based rapid decision-making mechanism, which includes: Set multi-level trigger thresholds to achieve warning within seconds; Dynamically adjust judgment parameters according to environmental factors and output standardized early warning information; Establish feature combination judgment logic to reduce the false alarm rate; the combination of feature combinations includes: Key features extracted: temperature change rate, temperature acceleration, smoke concentration gradient change, and hot spot area ratio; the area ratio of the hot spot area is detected by infrared thermal imaging to determine the fire scale, wind direction and speed, and environmental background characteristics; Feature combination method: Multimodal fusion: Fusion of feature data collected by different sensors to form a multi-dimensional feature vector; Time series analysis: Modeling of feature data in time series to analyze the changing trend of features over time; Feature weight allocation: Different weights are allocated according to the importance of key features of fire occurrence; Feature interaction combination: New combination features are constructed through feature interaction analysis; The judgment logic is specifically established by: using a rule engine or a machine learning model to perform intelligent identification and graded warning of fire risks; establishing rule-based judgment logic, machine learning-based judgment logic, and risk grading and response logic; Rule-based judgment logic, including: setting multi-level trigger thresholds and dynamically adjusting the judgment thresholds according to environmental background characteristics; triggering a preliminary warning when a single feature exceeds the threshold; The judgment logic based on machine learning includes: inputting multimodal feature vectors into lightweight deep learning models to classify and predict fire risks; the model outputs the fire risk level or the probability of fire occurrence; the interactive logic of feature combination is used to reduce the false alarm rate; based on the actual application effect, new data is used to conduct online learning and parameter optimization of the model to improve the judgment accuracy; Risk classification and response logic, specifically: three-level early warning mechanism: Yellow warning represents low risk: a single feature is close to the threshold, but the combined condition is not met, and the system records and continuously monitors; Orange warning indicates medium risk: the combination of multiple features meets the medium risk condition, triggering an alarm and notifying the management personnel; Red warning indicates high risk: if multiple features meet the high risk condition, the system will automatically link the fire extinguishing equipment and start the emergency plan; Prioritize high-weight features and make comprehensive judgments based on low-weight features.

7. The forest fire monitoring method based on edge intelligence according to claim 1 is characterized in that: S4 includes the following steps: S41: Intelligent emergency response, specifically: building an intelligent emergency response system, establishing a forest fire emergency plan library based on knowledge graph technology, and automatically matching the optimal plan according to the fire level and environmental conditions; establishing a cross-departmental collaborative response platform to achieve information sharing and joint response among firefighting, forestry, and emergency departments; using blockchain technology to record the response process to ensure that the entire emergency response process is traceable; integrating drones and robotic equipment to enhance emergency response capabilities; S42: Event analysis and experience learning, specifically: establish a systematic event analysis and continuous improvement mechanism; use big data analysis technology to conduct multi-dimensional analysis of fire events, evaluate the response effect of the early warning system, mine event characteristics through machine learning methods, and optimize early warning models and handling processes; establish a standardized experience review mechanism to form a complete case library; regularly organize technical seminars to promote the continuous evolution and upgrading of the system.

8. The forest fire monitoring method based on edge intelligence according to claim 7 is characterized in that: The forest fire emergency plan library includes fire level classification standards: fire levels are divided into different levels according to the severity of the fire, the speed of fire spread, and the scope of the affected area. Each level corresponds to a specific response strategy and resource scheduling plan; environmental condition data, including environmental parameters such as humidity, temperature, wind speed, wind direction, vegetation density, and terrain characteristics; emergency resource information, including equipment resources, human resources, and material reserves; fire handling procedures; Cross-departmental coordination mechanism; historical cases and experience database; dynamic adjustment rules; The dynamic adjustment rules dynamically adjust the contents of the plan according to the fire level and environmental conditions, and include: Humidity: When humidity is low, fire spreads faster and more resources need to be mobilized first; Wind speed and direction: When the wind speed is high and the wind direction is toward high-risk areas, the deployment location of the fire-fighting equipment needs to be adjusted; Vegetation density: Areas with high vegetation density have a greater risk of fire and require priority dispatch of firefighting resources; Terrain features: UAVs or robotic equipment should be deployed preferentially in mountainous or complex terrains; If the fire level is orange warning and the wind speed is greater than the threshold, the drone monitoring plan will be given priority; If the fire level is red warning and the humidity is less than the threshold, the dispatch amount of fire extinguishing agent will be increased; The system continuously updates the knowledge graph based on real-time data and optimizes the matching logic of the rule engine. It uses machine learning technology to mine successful experiences from historical cases and improve the intelligence level of plan matching.

9. The forest fire monitoring method based on edge intelligence according to claim 7 is characterized in that: Using big data analysis technology, we conduct multi-dimensional analysis of fire events and evaluate the response effect of the early warning system. The analysis methods include: A) Multi-dimensional analysis, including: Time series analysis: analyzing the changing trends of key features before and after a fire: temperature change rate, smoke concentration gradient; identifying the key time points of a fire: the initial stage of a fire, the fire spread stage, and the fire extinguishing stage; Spatial distribution analysis: using GIS mapping technology to analyze the geographical characteristics of the fire area: the distribution of high-risk areas and the path of fire spread; combining the area proportion of hot spots and wind direction to evaluate the spatial pattern of fire spread; Feature association analysis, using data mining technology to analyze the association between different features; Emergency response analysis, evaluate the dispatch efficiency of emergency resources: equipment arrival time, fire extinguishing equipment start-up time; analyze delays or unreasonable resource allocation problems in the emergency response process; B) Model optimization analysis, which includes: Model performance evaluation: Use fire event data to verify the accuracy and false alarm rate of the early warning model; evaluate the response speed of the model by comparing the actual fire occurrence time with the model prediction time; Feature importance analysis: Use machine learning models to calculate the contribution of each feature to fire prediction; optimize feature selection and improve model performance; Using big data analysis technology, we conduct multi-dimensional analysis of fire events and evaluate the response effect of the early warning system. The evaluation methods include: a) Detection performance evaluation, which includes: Accuracy: Calculate the accuracy of fire risk detection, that is, the proportion of correctly detected fires; the calculation formula is: Accuracy = (TP+TN) / (TP+TN+FP+FN) TP: the number of correctly detected fires; TN: the number of correctly detected non-fires; FP: the number of false fire alarms; FN: the number of missed fire alarms; Recall rate: evaluates the system's ability to capture fire events and avoid underreporting; its calculation formula is: Recall rate = TP / (TP+FN); False alarm rate: Evaluate the false alarm situation of the system and reduce the false alarm rate to reduce unnecessary waste of resources; the calculation formula is: false alarm rate = FP / (FP+TN); b) Response speed assessment, which includes: Warning time: calculate the time interval from the occurrence of fire to the issuance of warning by the system; control the warning time to the shortest range; Emergency response time: evaluate the time from the issuance of an early warning to the arrival of emergency resources at the scene; optimize the resource scheduling process and shorten the response time; c) Resource scheduling efficiency evaluation, which includes: Equipment utilization: Statistics on the usage of fire-fighting equipment, drones, and sensor resources to improve resource utilization and avoid resource waste; Reasonable resource allocation: Analyze whether resource scheduling matches the fire level and environmental conditions, optimize resource allocation strategies, and ensure efficient response; Through machine learning methods, event characteristics are mined to optimize early warning models and handling processes. Among them, the optimization methods based on event characteristics include: Key feature extraction, including: temperature change rate: analyze the temperature change per unit time and extract the temperature rise characteristics at the initial stage of the fire; smoke concentration gradient: calculate the rate of change of smoke concentration and identify potential fire sources; hot spot area ratio: use infrared thermal imagers to detect the area ratio of hot spots and assess the scale of the fire; environmental background characteristics: including humidity, wind speed, wind direction, vegetation density, etc., used to correct fire risk assessment; construct composite features through feature combination to improve the model's ability to identify fire risks; use time series analysis technology to extract dynamic change trends of features; Model optimization, which includes: model training and updating: using historical fire data and real-time collected data, regularly updating model weights to ensure that the model adapts to new environmental characteristics; using transfer learning methods to fine-tune the pre-trained model to accelerate model convergence; false alarm and missed alarm optimization: adjusting the decision threshold of the model through feature importance analysis to reduce the false alarm rate and missed alarm rate; building rule-based auxiliary judgment logic; dynamic adjustment strategy: dynamically adjusting model parameters according to real-time data to improve the adaptability of the model; in abnormal conditions, increasing the data sampling frequency to enhance the real-time performance of the model; Process optimization, including: Classifying fire conditions into three levels of warning: yellow, orange and red according to event characteristics; different fire levels correspond to different response strategies: Yellow warning: mainly monitoring, recording abnormal data; Orange warning: dispatch some resources for preliminary disposal; Red alert: fully activate emergency response plans and mobilize all available resources; Automatically match the optimal emergency plan based on fire level and environmental conditions; dynamically adjust the plan content to ensure the applicability of the plan; use event characteristics to optimize resource scheduling strategies; Post-event review: review the entire process of the fire incident and analyze the characteristic changes at key time nodes; evaluate the detection effect of the early warning model; evaluate the execution efficiency of the disposal process; identify deficiencies in the early warning model and disposal process; Continuous optimization, which includes: using event review data to optimize the model's feature selection and parameter settings; introducing new data features to improve the model's predictive capabilities; optimizing emergency plan content and resource scheduling strategies based on review results; dynamically adjusting warning thresholds to improve system adaptability; organizing fire incidents into standardized cases, recording fire types, triggering factors, handling processes and results; using machine learning technology to mine common features in the case library and optimize warning models and handling processes.

10. The forest fire monitoring method based on edge intelligence according to claim 1, characterized in that: It also includes user interaction and feedback, which specifically includes the following steps: Real-time monitoring interface, providing an intuitive visual monitoring platform; real-time display of monitoring data and analysis results through data visualization technology; integrated GIS map function, intuitive display of equipment location distribution and fire situation; multi-terminal adaptation, support PC and mobile terminal access at any time; authority management according to user roles to ensure information security; The manual adjustment function provides flexible system parameter configuration functions; supports dynamic adjustment of various warning thresholds to adapt to different monitoring needs; allows modification of data collection frequency to optimize system resource utilization; can customize alarm rules and notification methods to meet personalized needs; realizes remote configuration and restart of equipment to improve operation and maintenance efficiency.

Images