Wind Farm Fire Identification Methods and Systems

Abstract

This invention relates to the field of fire safety. To improve the accuracy of fire identification, it provides a method and system for fire identification in wind farms. By integrating multi-source data, a multi-dimensional fire perception system covering physical, chemical, optical, and visual quantities is constructed, achieving cross-validation and false alarm suppression from the source. Based on this, a dual-validation architecture using Bayesian networks and Transformer models is adopted to accurately capture precursor signals of fire overheating and smoldering stages, achieving very early fire warnings. Finally, the combination of a three-level judgment logic of suspected, pending verification, and normal, along with a dynamic high-frequency sampling verification mechanism, effectively reduces the probability of false alarms.

Description

Technical Field

[0001] This invention relates to the field of fire safety, specifically a method and system for identifying fires in wind farms. Background Technology

[0002] Existing wind farm fire protection systems suffer from numerous technical deficiencies and industry pain points, failing to meet the safety requirements for high-quality development in the wind power industry: First, the sensing methods are limited and outdated, relying heavily on traditional smoke and temperature detectors, which are susceptible to strong electromagnetic interference and drastic temperature fluctuations from wind turbines, resulting in a high false alarm rate. Furthermore, these systems can only trigger alarms during the open flame or dense smoke stages, failing to detect the very early signs of overheating or smoldering fires, thus missing the optimal time for fire suppression. Second, the fire suppression devices lack specific design considerations, failing to differentiate configurations based on the fire hazard characteristics of different areas such as the nacelle, hub, and tower base. This leads to low fire suppression efficiency, and some extinguishing agents can cause secondary damage to precision electrical equipment. Additionally, the anti-misfire mechanism is inadequate, resulting in a high risk of unplanned shutdowns. Third, the operation and maintenance system is incomplete, failing to provide full-cycle real-time monitoring of the operational status of fire suppression devices, sensors, and communication links, making it prone to "operation with faults," and hindering effective equipment startup during fires. Summary of the Invention

[0003] To improve the accuracy of fire identification, this invention provides a method and system for identifying fires in wind farms.

[0004] The technical solution adopted by the present invention to solve the above problems is: Methods for identifying fires in wind farms include: Collect detection data, including: temperature detection data, smoke spectrum detection data, VOC gas detection data, electric arc light detection data, and video detection data; Feature extraction based on detection data includes: extracting features such as temperature rise rate, temperature gradient, and duration of sustained high temperature based on temperature detection data; extracting the 650nm / 940nm dual-band light intensity ratio feature based on smoke spectrum detection data; extracting the concentration change rate feature based on VOC gas detection data; extracting the pulse amplitude and duration feature based on arc light detection data; and extracting visual features of flames and smoke based on video detection data. Construct and train a Bayesian network with the occurrence of a fire as the root node and the extracted features as child nodes. Train a node conditional probability table based on historical fire incidents and false alarm samples from the wind farm. Construct and train a Transformer model. The Transformer model takes a temporal feature matrix as input and outputs the fire probability. The temporal feature matrix is ​​composed of features from N consecutive sampling periods. Acquire the current detection data and extract features, and obtain probabilities P1 and P2 based on the trained Bayesian network and Transformer model, respectively. If both P1 and P2 are greater than the preset threshold, it is judged as a suspected fire. If both P1 and P2 are less than the preset threshold, it is judged as normal. Otherwise, it is marked as pending verification: trigger the A sensors closest to the area pending verification to increase the sampling frequency and extend the sampling period, and then make a judgment again.

[0005] Furthermore, the Transformer model is trained using knowledge distillation.

[0006] Furthermore, after triggering the A sensors closest to the area to be reviewed to increase their sampling frequency and extend their sampling period, the following process is performed again: if B or more sampling points simultaneously trigger their respective preset feature thresholds, the event to be reviewed is determined to be a suspected fire; otherwise, it is determined to be normal; where B... <A。

[0007] Furthermore, after acquiring the current detection data and extracting features, P1 is first calculated based on the trained Bayesian network. P2 is only calculated if P1 is not lower than the preset judgment threshold; otherwise, it is directly judged as normal. The preset judgment threshold is less than the preset threshold.

[0008] Furthermore, after determining a suspected fire, the process also includes: conducting a CFD fire situation simulation, generating an early warning level based on the simulation results, and executing multi-level response scheduling according to the early warning level.

[0009] Furthermore, the CFD fire situation simulation specifically includes: inputting the coordinates of suspected fire points and corresponding detection data into a pre-built wind turbine CFD model, and the CFD model outputs the fire spread path, smoke diffusion range, and time to reach adjacent units.

[0010] Furthermore, the early warning levels are divided into monitoring level, early warning level, alarm level, and action level; the multi-level response and dispatch are: autonomous handling by the site, mobilization of regional emergency forces, and joint action with local fire departments.

[0011] Wind farm fire detection system, including: The data acquisition unit is used to collect detection data, including: temperature detection data, smoke spectrum detection data, VOC gas detection data, electric arc light detection data, and video detection data; The feature extraction unit performs feature extraction based on the detection data, including: extracting features such as temperature rise rate, temperature gradient, and duration of sustained high temperature based on temperature detection data; extracting the 650nm / 940nm dual-band light intensity ratio feature based on smoke spectrum detection data; extracting the concentration change rate feature based on VOC gas detection data; extracting the pulse amplitude and duration feature based on arc light detection data; and extracting visual features of flames and smoke based on video detection data. The model building unit is used to construct a Bayesian network with the occurrence of a fire as the root node and the extracted features as child nodes. It trains the node conditional probability table based on historical fire incidents and false alarm samples from wind farms. It is also used to construct and train a Transformer model. The Transformer model takes a temporal feature matrix as input and outputs the fire probability. The temporal feature matrix is ​​composed of features from N consecutive sampling periods. The judgment unit is used to acquire the current detection data and extract features. It obtains probabilities P1 and P2 based on the trained Bayesian network and Transformer model, respectively. If both P1 and P2 are greater than the preset threshold, it is judged as a suspected fire. If both P1 and P2 are less than the preset threshold, it is judged as normal. Otherwise, it is marked as pending verification: the adjacent sensor is triggered to increase the sampling frequency and extend the sampling period before judgment is performed again.

[0012] Furthermore, the data acquisition unit includes: a distributed fiber optic temperature sensor, a multispectral smoke detector, a VOC gas sensor, an arc light monitoring unit, and a dual-band pan-tilt camera.

[0013] Furthermore, it also includes a communication unit, including a primary link, a backup link, and an emergency link; the primary link adopts an industrial-grade fiber optic ring network; the backup link integrates 5G / WiFi wireless transmission and LoRa low-power wide area network, and automatically switches to the backup link when the primary link fails; the emergency link is equipped with a satellite communication terminal and a drone relay communication interface, and automatically starts when both the primary link and the backup link are interrupted.

[0014] The advantages of this invention compared to existing technologies are as follows: By integrating multi-source data, a full-dimensional fire perception system covering physical, chemical, optical, and visual quantities is constructed, achieving cross-validation and false alarm suppression from the source; on this basis, a dual-validation architecture of Bayesian network and Transformer model is adopted, which can accurately capture the precursor signals of fire overheating and smoldering stages, realizing early warning of fires; finally, the combination of suspected-pending-normal three-level judgment logic and dynamic high-frequency sampling verification mechanism effectively reduces the probability of false alarms. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for identifying fires in wind farms. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] like Figure 1 As shown, the wind farm fire identification method includes: The collected detection data includes: temperature detection data, smoke spectrum detection data, VOC gas detection data, electric arc light detection data, and video detection data.

[0018] This embodiment employs distributed fiber optic temperature sensors, multispectral smoke detectors, VOC gas sensors, arc light monitoring units, and dual-band pan-tilt cameras for data acquisition. The distributed fiber optic temperature sensors are deployed along the cable and mounted against the wall on the surfaces of high-temperature-prone equipment such as cable trays and gearboxes from the nacelle to the tower base. With a high accuracy of ±0.5℃, they achieve continuous real-time monitoring of equipment temperature, accurately capturing abnormal temperature rise trends. Multispectral smoke detectors are deployed at the top of enclosed spaces such as the nacelle and substation power distribution room, as well as in areas prone to smoke accumulation. Through multispectral recognition technology, they effectively avoid interference from dust and water vapor generated during wind turbine operation, accurately identifying fire smoke. The arc light monitoring unit is specifically installed inside high-voltage electrical equipment such as tower base electrical cabinets and substation switchgear to quickly capture arc light signals generated by electrical short circuits and arcing, enabling early detection of electrical fires. VOC gas sensors are deployed around equipment prone to flammable medium leaks, such as gearboxes and hydraulic stations, to detect volatile organic compounds generated by leaks of lubricating oil and hydraulic oil in real time, providing early warning of fires before open flames occur. Dual-band (infrared / visible light) pan-tilt cameras are installed in open areas such as the top of the engine room and the substation area, capturing areas with abnormal temperatures through infrared thermal imaging and visually verifying the fire situation through visible light video. They also support pan-tilt rotation scanning to expand the monitoring range.

[0019] In addition, aspirating early-stage smoke detectors and linear temperature-sensing cables can be installed. Aspirating early-stage smoke detectors are deployed not only in key protection areas above core equipment such as gearboxes and generators in the engine room, but also extend to cover enclosed spaces with dense electrical equipment and prone to smoldering, such as engine room control cabinets, hydraulic stations, and converter cabinets. A distributed, multi-point sampling pipeline layout is used, with 8mm diameter flame-retardant sampling pipes evenly arranged along the top of the equipment and the corners of the engine room. A φ2mm sampling hole is opened every 1.5m to achieve comprehensive air sampling coverage within the engine room. It uses a built-in high-sensitivity laser scattering detection unit and an active air pump to extract air samples from the protected area at a constant flow rate of 0.8L / min. After two-stage dust filtration to remove sand, oil mist particles, and other contaminants from the engine room, the samples are then filtered to remove particles with a diameter of 0.01mm. This detector can accurately detect invisible smoke particles ranging from μm to 10μm, and can identify smoke concentrations as low as 0.005%obs / m. It can capture precursor signals during the overheating and smoldering stages of a fire (30 to 60 minutes before the open flame occurs), providing warnings 2 to 3 orders of magnitude earlier than traditional smoke detectors. The detector is equipped with automatic pipe blockage detection and self-cleaning functions, and can monitor the airflow pressure of the sampling pipe in real time. When the pipe becomes blocked due to dust accumulation, it automatically triggers an alarm and starts a reverse blowing cleaning program. At the same time, the whole unit adopts an all-metal sealed shell, and the internal circuit board is coated with a three-proof coating to withstand continuous high temperatures of up to 70°C, high-frequency vibrations, and oil mist corrosion in the nacelle, ensuring detection stability under long-term operation of wind turbine units. Linear temperature-sensing cables are continuously laid in narrow, densely packed areas that are difficult for personnel to inspect, such as vertical cable channels inside towers, underground cable trenches in substations, cable interlayers inside box-type transformers, and cable trays for wind farm collector lines. Employing recoverable differential and constant temperature composite detection technology, these cables combine differential temperature (triggered by alarm at a temperature rise rate ≥ 5℃ / min) and constant temperature (alarm thresholds can be flexibly set within the range of 60℃~140℃ according to area characteristics; default 85℃ for cable channels and 90℃ for cable trenches) dual alarm functions. This enables continuous temperature monitoring at any point along the cable line with no blind spots. The maximum laying length of a single detection cable can reach 200m. To meet the fire detection needs of long-distance cable tunnels in wind farms, the outer layer is made of high-strength flame-retardant polyvinyl chloride sheath, with an additional steel wire braided tensile layer and aluminum foil electromagnetic shielding layer in the middle. It has excellent tensile, compressive, and electromagnetic interference resistance, and can withstand the mechanical stress during cable laying and the electromagnetic radiation generated by the operation of wind turbines. At the same time, the surface of the sheath is treated with anti-corrosion and anti-rodent-gnawing treatment, making it suitable for the humid underground cable trenches, corrosive media, and complex outdoor environments. When a local temperature abnormality is triggered by non-fire factors, it is not necessary to replace the entire cable. Only the alarm point needs to be checked and reset to restore normal use, which greatly reduces the later operation and maintenance costs.

[0020] Distributed fiber optic temperature measurement captures temperature anomalies during the overheating stage; VOC gas sensors detect flammable medium leaks 10-30 minutes in advance; aspirating early-stage smoke detectors capture submicron-sized smoke particles during the smoldering stage; arc light monitoring units provide millisecond-level responses to electrical short circuits; and dual-band vision enables visual verification of the open flame stage, forming a progressive sensing chain of "leakage warning - overheat warning - smoldering warning - open flame confirmation." Different modal data exhibit significantly different sensitivities to interference: temperature data is easily affected by normal equipment temperature rises but is unaffected by dust interference; multispectral smoke data can distinguish between water vapor and smoke but is susceptible to electromagnetic interference; and arc light data is only sensitive to electrical discharges but cannot identify oil fires. Through cross-validation of multimodal features, false alarms from single sensors can be automatically eliminated. For example, a normal temperature rise in the gearbox will only trigger abnormal temperature characteristics without accompanying an increase in VOC concentration or changes in smoke characteristics, allowing the system to directly determine it as a normal operating condition, reducing the probability of false alarms from the data source.

[0021] Feature extraction is performed based on detection data, including: extracting features such as temperature rise rate, temperature gradient, and duration of sustained high temperature based on temperature detection data; extracting the 650nm / 940nm dual-band light intensity ratio feature based on smoke spectral detection data; extracting the concentration change rate feature based on VOC gas detection data; extracting the pulse amplitude and duration feature based on arc light detection data; and extracting visual features of flames and smoke based on video detection data.

[0022] For continuous temperature sequences from distributed fiber optic temperature measurements, a 5-second sliding window is used for median filtering to remove instantaneous temperature spikes caused by lightning surges and equipment vibrations, extracting three types of time-domain features: temperature rise rate, temperature gradient, and duration of sustained high temperature. For multispectral smoke detectors, the 650nm / 940nm dual-band light intensity ratio is extracted from spectral data, effectively distinguishing fire smoke from dust and water vapor interference generated by wind turbine operation. Concentration change rate features are extracted from VOC gas sensor data. Pulse amplitude and duration features are extracted from arc light monitoring unit data. The video stream from a dual-band pan-tilt camera is first downsampled by 16 times and then visual features of flames and smoke are extracted using a lightweight YOLOv8n model. For smoke particle concentration data collected by aspirating early-stage smoke detectors, absolute concentration value and concentration rise rate features can be extracted. For linear temperature-sensing cables, temperature rise rate, temperature gradient, and duration of sustained high temperature can also be extracted. In this embodiment, a temperature exceeding a high-temperature threshold is defined as high temperature; the high-temperature threshold can be set according to actual conditions.

[0023] A Bayesian network is constructed with the occurrence of a fire as the root node and the extracted features as child nodes. The node conditional probability table is trained based on historical fire incidents and false alarm samples from wind farms.

[0024] A Bayesian network adapted to wind farm operating conditions was constructed, with "whether a fire has occurred" as the root node and extracted features as child nodes. The node conditional probability table was trained based on over 1000 historical real fire samples from wind farms (covering typical wind farm fire types such as electrical short circuits, lubricating oil leaks, and lightning strikes) and false alarm samples (covering common false alarm scenarios such as electromagnetic interference, normal gearbox temperature rise, dust storms, and nacelle condensation). After training, the single-frame fire identification accuracy reached 92.3%. The node conditional probability table can be automatically updated quarterly based on newly added operational data from the wind farm, continuously improving the model's adaptability. Compared to purely data-driven AI models, the Bayesian network significantly improves the identification accuracy in rare fire scenarios (such as instantaneous fires caused by lightning strikes), and model updates only require adjustments to the conditional probability table, without retraining, adapting to the dynamic changes in wind farm operating conditions.

[0025] Construct and train a Transformer model. The Transformer model takes the temporal feature matrix as input and outputs the fire probability. The temporal feature matrix is ​​composed of features from N consecutive sampling periods.

[0026] In this embodiment, feature vectors from three consecutive sampling periods (1 second each) are concatenated into a temporal feature matrix, which is then input into a trained lightweight Transformer model, outputting the temporal correlation fire probability P2. The lightweight Transformer model employs knowledge distillation training, using a large-scale fire temporal recognition model with 50M parameters as the teacher model, distilling to obtain a student model with 4.2M parameters. The model input is a 3×8-dimensional temporal feature matrix, which can capture the changing trends of multi-source data within 3 consecutive seconds, effectively distinguishing between slow temperature rise in equipment and rapid temperature rise in fire. For example, the temperature rise rate of a gearbox due to load changes is typically below 2℃ / min and fluctuates continuously, while the temperature rise rate caused by a fire can reach over 10℃ / min and continues to rise. This temporal difference cannot be distinguished by features at a single moment. The introduction of the Transformer model effectively improves the recognition accuracy of such easily confused scenarios.

[0027] Acquire the current detection data and extract features, and obtain probabilities P1 and P2 based on the trained Bayesian network and Transformer model, respectively. If both P1 and P2 are greater than the preset threshold, it is judged as a suspected fire. If both P1 and P2 are less than the preset threshold, it is judged as normal. Otherwise, it is marked as pending verification: trigger the A sensors closest to the area pending verification to increase the sampling frequency and extend the sampling period, and then make a judgment again.

[0028] During the review process, a second assessment can be made based on the trained model. In order to reduce computing power, this embodiment uses a preset feature threshold method for review: the system will automatically retrieve the three nearest sensors around the area to be reviewed and perform high-frequency sampling at 10Hz (originally 1Hz), and extend the observation period to 5s; if two or more sampling points trigger their respective preset feature thresholds simultaneously within the extended observation period of 5s, the event to be reviewed will be judged as a suspected fire.

[0029] Through the aforementioned dual verification logic, the false alarm rate caused by strong electromagnetic interference, drastic temperature fluctuations, and dust and water vapor interference in wind farms is reduced to below 0.1%, while retaining the ability to capture precursor signals of overheating and smoldering stages. Field tests show that this dual verification logic reduces the false alarm rate by 87.2% compared to a single Bayesian network algorithm; and compared to a single AI model, with the same false alarm rate, the fire warning time is advanced by 12-18 seconds.

[0030] Furthermore, after acquiring the current detection data and extracting features, P1 is first calculated based on the trained Bayesian network. P2 is only calculated if P1 is not lower than a preset judgment threshold; otherwise, it is directly judged as normal. The preset judgment threshold is less than a preset threshold. In this embodiment, the preset threshold was determined to be 0.7 after 3 months of trial operation optimization at the wind farm site, and the preset judgment threshold is 0.5. The Bayesian network first performs a rapid initial screening, filtering out most of the normal operating condition data, and only sending the suspected abnormal data into the Transformer model for time-series verification, which reduces the average computing power consumption of the edge computing gateway by 62% and stabilizes the single-frame inference latency within 78ms.

[0031] After determining a suspected fire, the process also includes: conducting a CFD fire situation simulation, generating an early warning level based on the simulation results, and executing multi-level response scheduling according to the early warning level.

[0032] The edge computing gateway uploads the coordinates of suspected fire locations, initial temperature field, smoke concentration, arc light characteristics, and high-frequency sampling dynamic data to the cloud, triggering the cloud-based intelligent decision-making layer to perform the following operations: First, CFD fire situation simulation: The cloud-based CFD fire situation simulation module is based on the principle of computational fluid dynamics. It uses the received fire location and dynamic data as boundary conditions, combined with real-time meteorological information (wind direction and wind speed), to simulate and simulate the fire smoke diffusion path, temperature field distribution, and the arrival time required for the fire to spread to adjacent units.

[0033] Secondly, the warning level is generated: The four-level warning judgment unit divides the fire warning into four progressive levels based on the CFD simulation results (such as the fire spread speed, estimated arrival time, and affected area) and the probability of edge reporting P1 and P2: monitoring level (parameter abnormality, continuous observation), warning level (preliminary alarm from multiple sensors, increasing monitoring frequency), alarm level (confirmation of fire, initiation of pre-action), and action level (confirmation of fire, immediate initiation of fire extinguishing and evacuation).

[0034] Finally, a multi-level response scheduling is performed. In this embodiment, a three-level response is used.

[0035] Level 1 Response (Local Automatic Response): When the warning level reaches the alarm level or action level, the edge computing gateway will locally execute at least one of the following actions: triggering a local audible and visual alarm, cutting off the electrical circuit of the corresponding wind turbine, executing an emergency feathering shutdown, and activating the automatic fire extinguishing device in the nacelle. In this embodiment, the nacelle is equipped with a non-pressurized ultrafine dry powder fire extinguishing unit with a spray time ≤ 5 seconds, an extinguishing time ≤ 20 seconds, and a filling extinguishing agent weight ≥ 3 kg, capable of extinguishing Class A, B, C, and E fires. The hub is equipped with a miniaturized aerosol fire extinguishing unit, the tower base electrical cabinet is equipped with a fire detection tube type fire extinguishing unit, the booster station is equipped with a high-pressure fine water mist fire extinguishing unit, the offshore booster station is equipped with 100% extinguishing agent reserve, the nacelle can also be equipped with a perfluorohexanone gas fire extinguishing device, and the tower is equipped with a fine water mist spray system.

[0036] Level 2 Response (Site-level Dispatch): When the warning level reaches the action level, the site control center will execute the following: display the fire spread path simulated by CFD on the digital twin screen in real time, dispatch fire-fighting mobile equipment (fire trucks, mobile fire extinguishing devices) within the site to the fire location, notify the site operation and maintenance personnel and safety officers, and evacuate surrounding personnel through the broadcast system.

[0037] Level 3 Response (Social Collaboration): When the warning level reaches the action level and CFD analysis indicates that the fire may get out of control, the cloud will automatically execute the following: automatically contact the fire department and forward thermal imaging images, upload the predicted fire path and the best access route based on CFD analysis, etc.

[0038] Throughout the entire process, the digital twin model unit synchronously receives multimodal features and equipment status data uploaded from the edge, drives the 1:1 three-dimensional model to map the physical wind farm in real time, and visualizes and renders the CFD-inspired fire spread animation on the large screen of the central control center, providing operation and maintenance personnel with global situational awareness and auxiliary decision support.

[0039] Furthermore, a full-cycle operation and maintenance diagnostic layer can be set up, including a pressure monitoring unit, a chemical agent status detection unit, a communication link detection unit, a sensing accuracy calibration unit, and an intelligent maintenance suggestion unit. The full-cycle operation and maintenance diagnostic layer collects and monitors the pressure, chemical agent status, communication link stability, and sensor detection accuracy of each fire extinguishing device in real time. When it detects that the internal pressure of the fire extinguishing device is lower than the alarm value, the chemical agent caking, the communication link is interrupted, or the sensing accuracy is off, it will alert the operation and maintenance personnel through audible and visual alarms and remote push. It will also generate targeted maintenance suggestions based on the type of anomaly. At the same time, it is equipped with a "black box" recording unit. The black box recording unit of a single wind turbine can store no less than 1,000 core data such as fire alarm information, fault information, and system operation data. The data of the entire wind farm is synchronized to the cloud for permanent storage.

[0040] An automatic accident analysis and report generation module can also be set up in the cloud. After the fire is extinguished and there is no risk of reignition, a standardized analysis report is automatically generated based on the time-series data records of the entire chain, which includes fire source location, response time, handling measures, and loss assessment. It also supports data archiving and review analysis.

[0041] Correspondingly, the present invention also provides a wind farm fire identification system, including: The data acquisition unit is used to collect detection data, including: temperature detection data, smoke spectrum detection data, VOC gas detection data, electric arc light detection data, and video detection data; The feature extraction unit performs feature extraction based on the detection data, including: extracting features such as temperature rise rate, temperature gradient, and duration of sustained high temperature based on temperature detection data; extracting the 650nm / 940nm dual-band light intensity ratio feature based on smoke spectrum detection data; extracting the concentration change rate feature based on VOC gas detection data; extracting the pulse amplitude and duration feature based on arc light detection data; and extracting visual features of flames and smoke based on video detection data. The model building unit is used to construct a Bayesian network with the occurrence of a fire as the root node and the extracted features as child nodes. It trains the node conditional probability table based on historical fire incidents and false alarm samples from wind farms. It is also used to construct and train a Transformer model. The Transformer model takes a temporal feature matrix as input and outputs the fire probability. The temporal feature matrix is ​​composed of features from N consecutive sampling periods. The judgment unit is used to acquire the current detection data and extract features. It obtains probabilities P1 and P2 based on the trained Bayesian network and Transformer model, respectively. If both P1 and P2 are greater than the preset threshold, it is judged as a suspected fire. If both P1 and P2 are less than the preset threshold, it is judged as normal. Otherwise, it is marked as pending verification: the adjacent sensor is triggered to increase the sampling frequency and extend the sampling period before judgment is performed again.

[0042] Furthermore, it also includes communication units, including a primary link, a backup link, and an emergency link, to ensure the continuity and reliability of data transmission. Specifically, the primary link adopts an industrial-grade fiber optic ring network with a transmission rate of ≥1Gbps, enabling high-speed, low-latency data interaction between all wind turbines in the wind farm and the central control center; the backup link integrates 5G / WiFi wireless transmission and LoRa low-power wide area network, automatically switching to the backup link when the primary link fails; the emergency link is equipped with a satellite communication terminal and a drone relay communication interface, automatically activating when both the primary and backup links are interrupted.

Claims

1. A method for identifying fires in wind farms, characterized in that, include: Collect detection data, including: temperature detection data, smoke spectrum detection data, VOC gas detection data, electric arc light detection data, and video detection data; Feature extraction based on detection data includes: extracting features such as temperature rise rate, temperature gradient, and duration of sustained high temperature based on temperature detection data; extracting the 650nm / 940nm dual-band light intensity ratio feature based on smoke spectrum detection data; extracting the concentration change rate feature based on VOC gas detection data; extracting the pulse amplitude and duration feature based on arc light detection data; and extracting visual features of flames and smoke based on video detection data. Construct a Bayesian network with the occurrence of a fire as the root node and the extracted features as child nodes, and train the node conditional probability table based on historical fire incidents and false alarm samples from the wind farm. Construct and train a Transformer model. The Transformer model takes a temporal feature matrix as input and outputs the fire probability. The temporal feature matrix is ​​composed of features from N consecutive sampling periods. Acquire the current detection data and extract features, and obtain probabilities P1 and P2 based on the trained Bayesian network and Transformer model, respectively. If both P1 and P2 are greater than the preset threshold, it is judged as a suspected fire. If both P1 and P2 are less than the preset threshold, it is judged as normal. Otherwise, it is marked as pending verification: trigger the A sensors closest to the area pending verification to increase the sampling frequency and extend the sampling period, and then make a judgment again.

2. The wind farm fire identification method according to claim 1, characterized in that, The Transformer model is trained using knowledge distillation.

3. The wind farm fire identification method according to claim 1, characterized in that, The process involves triggering the A sensors closest to the area to be reviewed to increase their sampling frequency and extend their sampling period, followed by a second assessment: if B or more sampling points simultaneously trigger their respective preset feature thresholds, the event to be reviewed is classified as a suspected fire; otherwise, it is classified as normal. Specifically, B... <A。 4. The wind farm fire identification method according to claim 1, characterized in that, After acquiring the current detection data and extracting features, P1 is first calculated based on the trained Bayesian network. P2 is only calculated if P1 is not lower than the preset judgment threshold; otherwise, it is directly judged as normal. The preset judgment threshold is less than the preset threshold.

5. The wind farm fire identification method according to claim 1, characterized in that, After determining a suspected fire, the process also includes: conducting a CFD fire situation simulation, generating an early warning level based on the simulation results, and executing multi-level response scheduling according to the early warning level.

6. The wind farm fire identification method according to claim 5, characterized in that, The CFD fire situation simulation includes inputting the coordinates of suspected fire points and corresponding detection data into a pre-built wind turbine CFD model. The CFD model then outputs the fire spread path, smoke diffusion range, and time to reach adjacent units.

7. The wind farm fire identification method according to claim 6, characterized in that, The warning levels are monitoring level, early warning level, alarm level and action level; the multi-level response and dispatch are: autonomous handling by the site, mobilization of regional emergency forces and linkage with local fire departments.

8. A wind farm fire detection system, characterized in that, include: The data acquisition unit is used to collect detection data, including: temperature detection data, smoke spectrum detection data, VOC gas detection data, electric arc light detection data, and video detection data; The feature extraction unit performs feature extraction based on the detection data, including: extracting features such as temperature rise rate, temperature gradient, and duration of sustained high temperature based on temperature detection data; extracting the 650nm / 940nm dual-band light intensity ratio feature based on smoke spectrum detection data; extracting the concentration change rate feature based on VOC gas detection data; extracting the pulse amplitude and duration feature based on arc light detection data; and extracting visual features of flames and smoke based on video detection data. The model building unit is used to construct a Bayesian network with the occurrence of a fire as the root node and the extracted features as child nodes. It trains the node conditional probability table based on historical fire incidents and false alarm samples from wind farms. It is also used to construct and train a Transformer model. The Transformer model takes a temporal feature matrix as input and outputs the fire probability. The temporal feature matrix is ​​composed of features from N consecutive sampling periods. The judgment unit is used to acquire the current detection data and extract features. It obtains probabilities P1 and P2 based on the trained Bayesian network and Transformer model, respectively. If both P1 and P2 are greater than the preset threshold, it is judged as a suspected fire. If both P1 and P2 are less than the preset threshold, it is judged as normal. Otherwise, it is marked as pending verification: the adjacent sensor is triggered to increase the sampling frequency and extend the sampling period before judgment is performed again.

9. The wind farm fire identification system according to claim 8, characterized in that, The data acquisition unit includes: a distributed fiber optic temperature sensor, a multispectral smoke detector, a VOC gas sensor, an arc light monitoring unit, and a dual-band pan-tilt camera.

10. The wind farm fire identification system according to claim 8, characterized in that, It also includes a communication unit, including a primary link, a backup link, and an emergency link; the primary link adopts an industrial-grade fiber optic ring network; the backup link integrates 5G / WiFi wireless transmission and LoRa low-power wide area network, and automatically switches to the backup link when the primary link fails; the emergency link is equipped with a satellite communication terminal and a drone relay communication interface, and automatically starts when both the primary link and the backup link are interrupted.

Images