Fire hazard ground-air integrated monitoring and early warning system and early warning method

Abstract

The invention relates to the technical field of fire alarm, in particular to a fire ground-air integrated monitoring and early warning system and method, and the system comprises a flame and smoke recognition module which collects image data from a ground monitoring camera and an air unmanned aerial vehicle, carries out the feature extraction of the image data, generates a feature data set, and carries out the early warning based on the feature data set. And judging existence of flames and smog by using image analysis, and generating a fire identification result. According to the invention, multi-source heterogeneous data obtained by the ground monitoring camera, the aerial unmanned aerial vehicle and the infrared thermal imaging equipment are integrated for collaborative analysis, so that high-precision feature extraction of fire hazards and judgment of fire behaviors are realized, and the accuracy and response efficiency of fire recognition are improved; the unmanned aerial vehicle is used for dynamically collecting air thermal imaging data in real time, and an alarm is quickly formed based on temperature abnormal points, so that an abnormal temperature area can be timely positioned and pre-warned in the early stage of a disaster.

Description

Technical Field

[0001] The present invention relates to the technical field of fire alarm, and particularly to a ground-air integrated fire monitoring and early warning system and an early warning method. Background Art

[0002] The technical field of fire alarm involves a series of technical means and devices that use various detection sensors (such as smoke detectors, temperature sensors, infrared thermal imaging devices, etc.) to continuously sense fire hazards in real time, and use communication networks to quickly transmit monitoring data to the control center for data analysis, fire situation identification, and triggering of alarm linkage mechanisms.

[0003] The existing technology mainly relies on a single type of detection sensor installed at a fixed position to complete data collection. The monitoring perspective is limited and the detection area is relatively fixed, making it difficult to cover the monitoring blind spots under dynamically changing or complex environmental conditions. As a result, in some special terrains, vegetation-dense areas, or when emergencies occur, there are easily monitoring blank areas, thus delaying the early discovery of fires. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the drawbacks existing in the prior art, and to propose a ground-air integrated fire monitoring and early warning system and an early warning method.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A ground-air integrated fire monitoring and early warning system includes: A flame and smoke recognition module, which collects image data from ground monitoring cameras and aerial drones, extracts features from the image data to generate a feature data set, and based on the feature data set, uses image analysis to distinguish the presence of flames and smoke and generates a fire recognition result; A temperature monitoring module, which receives temperature data from an infrared thermal imaging camera carried by an aerial drone, performs threshold analysis on the temperature data to generate temperature analysis data, and based on the temperature analysis data, identifies abnormal temperature points and generates an abnormal temperature alarm result; A resource energy-saving regulation module, which judges the current fire risk level according to the fire recognition result and the abnormal temperature alarm result to generate a risk level assessment result, and adjusts the working frequency of the ground monitoring camera according to the risk level assessment result to generate an energy-saving control result; An event response and communication module, which formulates an emergency response strategy and generates an emergency response plan by combining the risk level assessment result and the energy-saving control result.

[0006] Preferably, the steps for obtaining the feature data set are: Collect image data from ground monitoring cameras and aerial drones to obtain a comprehensive image data set; Based on the comprehensive image dataset, the edges, textures, and color distributions of the images are extracted through high-pass filtering and grayscale processing to construct a feature dataset.

[0007] Preferably, the steps for obtaining the fire recognition result are as follows: Based on the feature dataset, the flame and smoke features in the images are respectively recognized through a convolutional neural network to obtain a preprocessed feature dataset for distinguishing flames and smoke; Based on the preprocessed feature dataset, the classification result is verified to obtain the fire recognition result.

[0008] Preferably, the steps for obtaining the temperature analysis data are as follows: Receive the temperature data from the infrared thermal imaging camera carried by the aerial drone, extract the infrared temperature information at different times and different geographical locations, eliminate the sensor errors or the reflected signals in the invalid areas, classify according to the data acquisition time, and mark the data categories of different seasons to obtain a classified temperature dataset; Based on the classified temperature dataset, calculate the season-adaptive temperature threshold, and the calculation formula is: ; Wherein, is the calculated season-adaptive temperature threshold, represents the temperature data of the th sampling point in winter, represents the temperature data of the th sampling point in summer, represents the temperature data of the th sampling point in spring and autumn, is the number of sampling points in winter, is the number of sampling points in summer, is the number of sampling points in spring and autumn; Based on the season-adaptive temperature threshold, analyze the temperature values of each monitoring point in the temperature dataset, mark the points exceeding the season-adaptive temperature threshold as abnormally high temperature areas, and generate temperature analysis data.

[0009] Preferably, the steps for obtaining the abnormal temperature alarm result are as follows: Based on the temperature analysis data, extract the temperature point data at different positions within the monitoring area to obtain a temperature change dataset; According to the temperature change dataset, calculate the temperature anomaly index, and the calculation formula is: ; Wherein, is the temperature anomaly index, is the current monitoring point temperature, is the surrounding background temperature, is the temperature gradient in the horizontal direction, is the temperature gradient in the vertical direction, is the time interval; Based on the temperature anomaly index, abnormal temperature points are screened to generate an abnormal temperature alarm result.

[0010] Preferably, the steps for obtaining the risk level assessment result are as follows: Based on the fire recognition result and the abnormal temperature alarm result, the spatial coordinates of the flame area, the smoke diffusion area, and the abnormal temperature points are extracted to obtain the basic fire risk data; According to the basic fire risk data, calculate the fire risk level index. The calculation formula is: ; where, is the fire risk level index, is the area of the flame area, is the area of the smoke diffusion area, is the average temperature of the abnormal temperature points, is the center coordinate of the flame, is the center coordinate of the abnormal temperature points; Based on the fire risk level index, set the risk level classification standard to divide areas with different risk levels to obtain the risk level assessment result.

[0011] Preferably, the steps for obtaining the energy-saving control result are as follows: Based on the risk level assessment result, calculate the adjusted working frequency of the ground monitoring cameras. The calculation formula is: ; where, is the adjusted working frequency, is the original working frequency, is the fire risk level index, is the number of times the fire alarm has been triggered in the current area in the most recent month, is the number of times of abnormal recognition of flame or smoke in the current area in the most recent month, is the average ambient temperature in the current area in the most recent month; Based on the adjusted working frequency, re-adjust the sampling intervals of each monitoring camera to obtain the energy-saving control result.

[0012] Preferably, the steps for obtaining the emergency response plan are as follows: According to the risk level assessment result and the energy-saving control result, integrate fire trucks, fire-fighting personnel, and medical support resources, and assign tasks and schedules to each resource to obtain the emergency response plan.

[0013] The present invention provides a ground-air integrated monitoring and early warning method for fires, comprising the following steps: Collect image data from ground monitoring cameras and aerial drones, perform pixel-level processing on the collected images, extract image brightness, color, and texture features, and form a feature dataset; Based on the feature dataset, classify and judge each pixel point in the image, identify the flame and smoke patterns in the image, and generate a fire recognition result according to the recognized ratio and distribution; Receive the temperature data collected by the infrared thermal imaging camera of the drone, calculate the maximum value, minimum value, and average value of the temperature data, identify and mark the temperature points exceeding the normal range, and generate an abnormal temperature alarm result; Combine the fire recognition result and the abnormal temperature alarm result, calculate the current fire risk index, adjust the working frequency of the ground monitoring camera and the inspection frequency of the drone according to the risk level, and generate an energy-saving control result; Combine the energy-saving control result and the risk level, formulate emergency response measures for different risk levels, and generate an emergency response plan.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, through the collaborative analysis of multi-source heterogeneous data obtained by integrating ground monitoring cameras, aerial drones, and infrared thermal imaging devices, high-precision feature extraction of fire hazards and determination of fire conditions are realized, improving the accuracy and response efficiency of fire recognition; using drones to collect real-time dynamic aerial thermal imaging data and quickly forming an alarm based on temperature anomaly points enables the abnormal temperature area to be located and warned in the early stage of the disaster; according to the comprehensive risk assessment of fire conditions and temperature changes, the operating state of ground monitoring equipment is automatically adjusted to achieve a balance between monitoring accuracy and resource energy consumption; at the same time, response strategies are formulated based on the fire risk assessment, and the timeliness and effectiveness of disaster response are improved with a real-time, accurate, and highly targeted emergency plan, ultimately realizing the common improvement of the overall efficiency of the fire monitoring and early warning system and the resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a system flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0017] Please refer to Figure 1, the present invention provides a technical solution: A ground-air integrated fire monitoring and early warning system includes: A flame and smoke recognition module that collects image data from ground monitoring cameras and aerial drones, extracts features from the image data to generate a feature data set, and based on the feature data set, uses image analysis to determine the presence of flames and smoke and generates a fire recognition result; A temperature monitoring module that receives temperature data from an infrared thermal imaging camera carried by an aerial drone, performs threshold analysis on the temperature data to generate temperature analysis data, and based on the temperature analysis data, identifies abnormal temperature points and generates an abnormal temperature alarm result; A resource energy-saving regulation module that determines the current fire risk level according to the fire recognition result and the abnormal temperature alarm result, generates a risk level assessment result, and adjusts the working frequency of the ground monitoring camera based on the risk level assessment result to generate an energy-saving control result; An event response and communication module that formulates an emergency response strategy and generates an emergency response plan by combining the risk level assessment result and the energy-saving control result.

[0018] The steps for obtaining the feature data set are as follows: Collect image data from ground monitoring cameras and aerial drones to obtain a comprehensive image data set; Based on the comprehensive image data set, extract the edges, textures, and color distributions of the images through high-pass filtering and grayscale processing to construct a feature data set.

[0019] Specifically, based on the ground monitoring cameras and aerial drone shooting sources arranged in the early stage, first determine the resolution and frame rate of each ground monitoring camera, and record the parameters of the drone cameras with flight altitudes in the range of 50 meters to 200 meters. Then, read the image data output by the cameras frame by frame and perform a preliminary integrity check. Compare each frame of the image with a pre-set brightness range. For example, in empirical evaluation, the brightness range is selected as 30 to 220. When the average brightness of a certain frame of the image is lower than 30 or exceeds 220, that frame is excluded. The specific values of 30 and 220 are the results obtained by calculating the average brightness and statistically analyzing the histogram distribution after multiple samplings. In addition, check the clarity of the image and check for serious defocus. The clarity threshold can be set near an empirical value (for example, taking 10 as the benchmark) and combined with the local pixel gradient change to measure the clarity. If the local gradient mean is less than 10, it is classified as a defocused image. After confirming that the brightness and clarity requirements are met, perform a chronological sorting according to the image timestamp for subsequent multi-angle and multi-period joint analysis. For the images taken by the drone, additionally record the flight position and attitude angle and compare the wind speed data within the range of level 0 to level 5 to avoid large fluctuations in the images collected when the wind speed exceeds level 5. Integrate all the data according to this process, and finally obtain the aggregated image data as the comprehensive image dataset.

[0020] Based on the aforementioned comprehensive image dataset obtained, perform high-pass filtering on the pixel matrix of each frame of the image and set the filtering radius to 3. This value is selected after gradually testing the effects of different radii and statistically analyzing the edge clarity in the experiment. Adopt the method of retaining high-frequency components and attenuating the low-frequency region in the frequency domain to emphasize the image detail and contour information. Subsequently, convert the filtered image to grayscale format, and define a weighting coefficient for each pixel according to reference experience. The gray value can be obtained by taking the ratio of the red channel weight of 0.3, the green channel weight of 0.59, and the blue channel weight of 0.11. These weights are typical values selected by statistically analyzing the sensitivity of the human eye to the three primary colors. Next, detect the gradient amplitude change of the local area pixel by pixel and record the edge features. When extracting texture features, the entropy threshold of the texture direction can be calculated in the horizontal and vertical directions for the same image. If the texture entropy of some areas is detected to be greater than the local statistical average, mark them as high-texture areas. Finally, further perform histogram-based statistics on the color distribution of the grayscale image and compare the dispersion degree of the color channels. If the dispersion degree exceeds the empirical upper limit, such as set to 100, mark that area as a significant color difference point. Through this series of processing processes, finally obtain the edge information, texture information, and color distribution for feature analysis and obtain the feature dataset.

[0021] The steps to obtain the fire recognition result are as follows: Based on the feature dataset, the flame and smoke features in the image are respectively identified through a convolutional neural network to obtain a preprocessed feature dataset for distinguishing flames and smoke; Based on the preprocessed feature dataset, the classification results are verified to obtain the fire recognition results.

[0022] Specifically, referring to the previously obtained feature dataset, first, all the image samples are classified and sorted according to the labeled flame and smoke categories. In the software, the images related to flames are used as positive samples, and the images related to smoke are used as another type of positive samples. Then, the images without flames and without smoke are regarded as background samples. After that, these images are converted to a unified resolution size, such as and reviewed according to the preset brightness range and color distribution. For example, when the average brightness of the image exceeds or is lower than , it is marked as an abnormal sample. These values are obtained by statistically averaging the brightness of a large number of outdoor and indoor surveillance images and taking the median value with a certain floating range . After completing the sample verification, the images are divided into three parts: a training set, a validation set, and a test set. The training set is used to build a convolutional neural network and complete parameter learning. The network used contains three convolutional layers and two fully connected layers. The size of the convolutional kernel is fixed at , and the number of convolutional kernels in each layer ranges from to . The convolutional stride is set to 1, the activation function is fixed to use ReLU, and mini-batch iteration is performed with a batch size of 32. The initial learning rate is taken as . If the loss value does not decrease in ten consecutive iterations, the learning rate is multiplied by 0.1 and decreased. The total number of training iterations is set to 50 rounds, and the accuracy of the validation set is continuously monitored. When the accuracy of the validation set stops improving for more than 5 rounds, the training stops. After training, the accuracy, recall rate, and precision rate of the network in flame and smoke detection are evaluated according to the test set results. After training is completed, forward inference is performed on all images to respectively identify the flame features and smoke features in the images, and then category labels are added to the feature dataset according to the results, and coordinate or segmentation region information is added, so as to finally obtain a preprocessed feature dataset that can distinguish flames and smoke.

[0023] Based on the above-obtained preprocessed feature dataset, first, the category labels and coordinate information of all flame or smoke information are read item by item and compared with the original images. Then, the true labels are accurately compared with the network's judgment results to record the samples with judgment errors. Next, the possible misjudgments in these error samples are reviewed one by one and their true categories are redefined. If the color distribution or shape features of some smoke samples are highly confused with the features of flames, these samples will be added to the training set and the network will be retrained. During the setting of the threshold, if it is found that the recall rate is lower than If it indicates that the discrimination standard is too strict, the judgment threshold can be changed from the original to and tested again. These values are obtained from the classification result statistics of the early demonstration dataset and combined with the shooting materials in the real environment for evaluation. After completing the above verification and retraining steps, the classification accuracy of fire and smoke is comprehensively statistically analyzed. When both the accuracy rate and the recall rate reach over 80%, it is regarded as the completion of classification verification, and then the final fire recognition result is output.

[0024] The steps for obtaining temperature analysis data are as follows: Receive the temperature data from the infrared thermal imaging camera carried by the aerial drone, extract the infrared temperature information at different times and different geographical locations, eliminate the sensor errors or the reflected signals in the invalid areas, classify according to the data acquisition time and mark the data categories in different seasons to obtain the classified temperature dataset; Based on the classified temperature dataset, calculate the season-adaptive temperature threshold. The calculation formula is: ; Among them, is the calculated season-adaptive temperature threshold, represents the temperature data of the th sampling point in winter, represents the temperature data of the th sampling point in summer, represents the temperature data of the th sampling point in spring and autumn, is the number of sampling points in winter, is the number of sampling points in summer, is the number of sampling points in spring and autumn; Based on the season-adaptive temperature threshold, analyze the temperature values of each monitoring point in the temperature dataset, mark the points exceeding the season-adaptive temperature threshold as abnormally high temperature areas, and generate temperature analysis data.

[0025] Specifically, the temperature data of the infrared thermal imaging camera carried by the aerial drone is received. First, the infrared image stream taken by the drone when the flight altitude is between 50 meters and 200 meters in actual operation is captured frame by frame and the temperature value of each pixel is extracted. The flight time, geographic coordinates and the nominal accuracy coefficient inside the infrared equipment are recorded one by one. Then, the effective value of the infrared temperature measurement is distinguished in the range of 40℃ to 800℃ in an empirical manner and compared to determine whether the sensor is overloaded or the signal is distorted. In order to determine the conditions for the error of the sensor, a reference standard needs to be set in advance. For example, the maximum allowable error at 400℃ is not more than ±5℃. This standard comes from the long-term comparison of the difference between the actual temperature measured data and the infrared temperature measurement data and is obtained through multiple calibration processes. When it is found that the temperature deviation of a certain point is greater than this allowable error range It is defined as an abnormality and the relevant records are excluded. At the same time, the area with obvious reflection in the infrared image is compared with the material information of the corresponding position in the visible light image. If it is a mirror material and the light intensity is above 500lx, it is determined to be an invalid reflection area and the corresponding temperature data is eliminated. After the elimination, it is grouped and classified according to the acquisition time and assigned different time period labels such as morning, noon, and night. The date is corresponded to the seasonal interval, such as December to February of the following year is winter, June to August is summer, March to May and September to November are spring and autumn, and the meteorological environment category of these data is marked to distinguish the background temperature distribution in different seasons. If a certain day is near the change of seasons, the actual temperature statistics of the local meteorological station are used to determine whether it is classified as winter, spring, autumn or summer, and then a classified temperature data set is formed.

[0026] The benefit of the formula is that it combines the temperature sampling information of winter, summer, spring and autumn to perform difference and average calculations, thereby taking into account the differences and average levels between sampling points in different seasons under the same calculation framework.

[0027] The steps to obtain the parameters are as follows: This item represents the sum of the temperatures of the winter sampling points. All the valid values ​​in the previous steps are screened and recorded in the database. After confirmation, the value-by-value addition operation is performed, and then the repeated temperature records in the adjacent time period are merged. It should be noted that the time range determined in winter is strictly divided according to the monthly division given by the local meteorological department. For example, from mid-December to mid-February of the following year, each temperature point that meets the winter range is uniformly included in the calculation for superposition. In order to obtain the actual usable value, it is necessary to first exclude the infrared temperature measurement distortion caused by highly reflective objects, and then perform linear interpolation on the remaining temperatures to make up for a small number of missing measurement points. Finally, all the corrected values ​​are added. For example, when p=5, the corresponding 5 valid records can be read in the database, such as 3℃, -1℃, 2℃, 0℃ and 4℃, and they are added to get .

[0028] The steps for obtaining the parameter are as follows: This item represents the total temperature of the sampling points in summer. The summer range is based on June to August defined by the local meteorological station. All infrared temperature measurement data within this range are collected daily and summarized. After excluding the temperature spikes or extreme outliers caused by abnormal situations, the remaining temperatures are added up in sequence to obtain , for example, when q = 4, 4 records such as 30°C, 31°C, 29°C, and 32°C can be retrieved from the database, and adding them up gives 30 + 31 + 29 + 32 = 122.

[0029] The steps for obtaining the parameter are as follows: This item represents the total temperature of the sampling points in spring and autumn. To distinguish between spring and autumn, in accordance with the local meteorological station's classification, the temperature measurement data from March to May and from September to November are collected daily. After excluding the data that cannot be measured normally, the temperature integration is carried out one by one to obtain r effective sample values, and then these sample values are added up to get , for example, when r = 6, six data points such as 14°C, 10°C, 18°C, 15°C, 16°C, and 12°C can be selected from the collection records, and the result after addition is 14 + 10 + 18 + 15 + 16 + 12 = 85.

[0030] The steps for obtaining the parameter are as follows: This parameter represents the number of sampling points in winter. The specific value is the number of sampling records remaining after comparing the temperature readings of all infrared image frames during the winter period and excluding duplicates or invalid ones. Assuming that a total of 130 valid infrared temperature records are collected from December of a certain year to February of the following year, and 98 of them belong to the same period and have almost the same temperature values, after merging and only retaining the essential differences, there are finally 5 left, then .

[0031] The steps for obtaining the parameter are as follows: This parameter represents the number of sampling points in summer. Using the same method as p but based on the detection data from June to August for statistics. The common point is that abnormal extreme spikes and overlapping sampling records also need to be excluded, and the valid entries representative of the total temperature distribution are retained to determine the final number of sampling points. Assuming that a total of 200 temperature records are captured during the summer collection period and after excluding some duplicate measurements, 4 valid differences remain, then .

[0032] The steps for obtaining the parameter are as follows: This parameter represents the number of sampling points in spring and autumn. It is necessary to distinguish the data according to the intervals from March to May and from September to November. All the temperature readings in the above two periods are read and grouped by the same type. If the number of valid records remaining after excluding abnormal data is 6, then .

[0033] Calculation process: Set p = 5, q = 4, r = 6, and obtain from the previous steps respectively , , . When substituting these values into the formula: ; ; ; ; Therefore: ; This result indicates that the value of is approximately 52.17. This threshold is obtained by comprehensively calculating the sampling information of each season of winter, summer, spring, and autumn. It shows that as long as the temperature at the monitoring point exceeds 52.17 °C, it can be regarded as significantly higher than the background range of these seasons. This is of reference significance for the subsequent identification of abnormal temperature regions. If the infrared temperature collected is higher than 52.17 °C, it will be recorded as an abnormal high temperature region and included in the subsequent analysis.

[0034] Based on the season-adaptive temperature threshold, first compare the actual temperature values of each monitoring point already recorded in the classified temperature dataset, and use the obtained as the judgment criterion. Then, read the collected temperature values one by one and compare them with . When the temperature of a certain monitoring point is greater than the threshold, it is counted as an abnormal high temperature point. If the temperature of the monitoring point is below the threshold, it is regarded as a normal point. At the same time, establish a temperature comparison table arranged in chronological order, mark the geographical coordinates and specific temperature values of each high temperature point, and record the temperature at the relevant moment in the additional data sequence to observe whether there is a continuous over-limit situation. If it is found that a certain area exceeds the limit multiple times, the area can be directly locked during subsequent retrieval and further investigation measures can be prepared. In addition, monitoring points close to the threshold boundary and with a temperature fluctuation range within ±2 °C will also be classified as critical and specially marked. The above temperature comparison and classification process traverses all monitoring point data in sequence through a computer program to form a final summary. Finally, all monitoring points with temperature values higher than the threshold are uniformly marked as abnormal high temperature regions to obtain temperature analysis data.

[0035] The steps to obtain the abnormal temperature alarm result are as follows: Based on the temperature analysis data, extract the temperature point data at different positions within the monitoring area to obtain a temperature change dataset; According to the temperature change dataset, calculate the temperature anomaly index. The calculation formula is: ; Among them, is the temperature anomaly index, is the temperature of the current monitoring point, is the ambient background temperature, is the temperature gradient in the horizontal direction, is the temperature gradient in the vertical direction, is the time interval; Based on the temperature anomaly index, abnormal temperature points are screened to generate the abnormal temperature alarm result.

[0036] Specifically, based on the previously obtained temperature analysis data, first completely read the temperature records of each monitoring point in different time periods in the temperature analysis data, and uniformly organize the basic information such as the monitoring point location, acquisition time, and temperature reading in the same database or the same information storage. Then, match and compare the geographical coordinates of the monitoring points to determine whether there are sufficient temperature data samples at the same location in several consecutive time periods. If there are multiple temperature records at the same point in a certain time period, they are arranged in chronological order, and the temperature change amount between adjacent time points is registered in a comparison list, and the specific time difference and coordinate annotation are attached to the list. At the same time, in order to further distinguish significant changes from minor changes, a judgment criterion can be set within the scope of experience. For example, compare the temperature change amount with the range of 0°C to 10°C. If it exceeds 10°C, it is recorded in the significant change sub-table, otherwise it is classified into the ordinary change sub-table. This 10°C value is the median obtained by referring to the daily temperature fluctuations in the monitoring area in the past month and statistically analyzing the peak-to-valley value difference, and then adding 2°C as a floating value to prevent missed classification. Next, continue to integrate the monitoring information at different times in this hierarchical list. For example, if the temperature fluctuates greatly multiple times within the same hour, it is counted as one time series mutation point. Use this method to aggregate and summarize each monitoring point separately and form an initial table with a multi-column structure. In this table, fields such as location number, time index, temperature difference, and flag label are listed in sequence. Then, splice the information of all monitoring points and sort them globally according to the acquisition time, so that the temperature change process of each location over time can be presented from beginning to end. Then, interpolate and complete these ordered records at a unified interval. Especially when there are scattered breakpoints in individual time periods, the temperature mean value in the adjacent time period can be used as a reference for linear interpolation. After interpolation, assign a retrievable index to each monitoring point in the final result for cross-record association and comparison in the future. After summarizing the above steps, these temperature records with time series evolution information can be aggregated and arranged into a temperature change data set.

[0037] The advantage of the formula is that it comprehensively measures the temperature difference between the current monitoring point and the background temperature by combining the temperature gradients in the spatial and time dimensions, taking into account both the difference in temperature distribution between the monitoring point and the surrounding area and the rate of temperature change within the specified time interval, and can more accurately reflect the degree to which the monitoring point temperature deviates from the normal state.

[0038] The steps for obtaining the parameter are as follows: This parameter represents the temperature at the current monitoring point, specifically from the real-time readings of an infrared thermometer or other temperature acquisition devices. When obtaining it, situations of faulty sensors or distorted temperature reflections need to be excluded. In order to accurately extract , when performing on-site operation and maintenance, a normal temperature measurement range of 0°C to 200°C is usually set, and for the outdoor environment, the upper limit is allowed to be appropriately extended to 300°C under summer or high-temperature conditions. Abnormal temperature measurement records are selected through the temperature data screening principle defined above, and are distinguished according to the monitoring point coordinates and timestamps and written into the database. If there are multiple temperature readings at the same location and the same time, the one with higher monitoring device accuracy is selected as the valid value and duplicates are removed. For example, during a certain flight mission, several temperature data with the same coordinates are obtained at different flight altitudes, and the result with better resolution should be preferentially retained. Finally, after indexing all the valid temperature data, it can be retrieved and obtained when calculating. For example, if the monitored temperature at a certain monitoring point is 78°C at 18:30 on July 12, 2025, and it is found through sensor comparison on-site that the device with the highest accuracy also outputs 78°C, then 78°C is recorded as .

[0039] The steps for obtaining the parameter are as follows: This parameter represents the ambient background temperature. Usually, a representative value is selected from the temperature samples within a certain number of meters near the monitoring point, and it needs to be obtained through spatial interpolation in combination with temperature measurements at multiple points. During acquisition, the background temperature of the same area is first continuously scanned through a ground temperature measurement device, or supplementary shooting is carried out by a drone in adjacent areas. The obtained data is merged based on geographical coordinates. If the temperature data of adjacent points in space differ little, the arithmetic mean can be used to obtain a background temperature. Otherwise, outliers need to be removed during merging and the linear interpolation method is used to form a smooth background temperature field. Then, the corresponding grid point of the monitoring point position in this plane is matched and its temperature is extracted as the value of. For example, a certain monitoring point is located at X = 100m, Y = 200m. A background temperature field with a 5-meter grid resolution has been stored in the background. If the temperature at this point can be obtained as 23°C after interpolation, then this 23°C is recorded as .

[0040] The steps for obtaining the parameter are as follows: This parameter represents the temperature gradient in the horizontal direction. The temperature distribution along the horizontal direction around the monitoring point is required, and it is obtained by dividing the temperature difference between two adjacent reference points by their distance on the X coordinate axis. If multiple temperature acquisition stations are deployed on-site or multiple temperature values at the same height are marked in the drone shooting, then it can be obtained from the coordinate differences and temperature differences of these adjacent points , in the case of rich data volume, more segmented sampling can be done in the horizontal direction, and then the gradients of adjacent segments can be combined into an average gradient to obtain a more accurate . When obtaining it, attention should be paid to the unification of the coordinate system and skipping the discontinuous areas in the interpolation algorithm to ensure that the obtained gradient represents the true temperature change rate. For example, if the temperature at X = 95m in a certain area is 25°C and the temperature at X = 105m is 27°C, then °C / m.

[0041] The acquisition steps of the parameter are as follows: This parameter represents the temperature gradient in the vertical direction and needs to be calculated in combination with the temperature distribution above and below the monitoring point. In actual operation and maintenance, the temperature can be measured at the same geographical location but different altitudes, or the temperature change values at different flight heights can be recorded during the UAV cruise. For areas with complex terrain, altitude correction of the terrain needs to be done well. The processing flow is similar to , and it is also obtained by dividing the temperature difference between adjacent points by their distance on the Y coordinate axis to get . The average value of multiple measurement results can be taken and abnormal temperature measurement records can be excluded to obtain the final vertical gradient. For example, if the temperature measured at a certain point at an elevation of Y = 200m is 25°C and the temperature measured at an elevation of Y = 205m is 26.5°C, then °C / m.

[0042] The acquisition steps of the parameter are as follows: This parameter represents the time interval and is usually obtained from the time difference between two consecutive data acquisitions at the same monitoring point or the same spatial position. If the temperature measurement device samples at a fixed frequency, this time interval can be stabilized between a few seconds and dozens of seconds. If there are batch records during sampling, the time difference between adjacent record timestamps can be calculated to get . It is necessary to ensure the clock synchronization of the data in advance to avoid time dislocation. If the monitoring points are distributed in the UAV cruise path, the image timestamps of each shot also need to be unified to the same time zone and date before performing the difference operation. For example, if a certain monitoring point acquires a temperature record at 18:30:10 and acquires the next record at 18:30:25, then seconds.

[0043] Calculation process: Select a monitoring point in the monitoring area and obtain °C, °C, °C / m, °C / m, seconds. First, calculate the sum of squares term in the denominator: ; ; Adding the two gives: ; Then perform the square root operation: ; After that, the denominator part is: ; The numerator part is: ; Therefore: ; The result shows that the temperature anomaly index is 53.72. When the value is greater than 30, this point can be classified as an obvious anomaly area, indicating that the current monitoring point has a significant temperature deviation in the case where the temperature difference from the background temperature is large and the horizontal and vertical temperature gradients increase relatively slowly, and it also indicates that this monitoring point may be in a high heat source or other areas that cause temperature rise.

[0044] Based on the temperature anomaly index, first pair all the temperature anomaly index results calculated by the above formula with the corresponding monitoring points, and then call the spatial coordinates and time indexes of each monitoring point one by one and compare them. To identify abnormal temperature points, a reference critical value can be set to distinguish the normal and abnormal ranges. This critical value can be set through historical statistical data. For example, after sorting out the distribution of temperature anomaly indexes monitored in the past six months, it is found that the average value fluctuates in the range of 15 to 20, and combined with the on-site measurement results of sudden high-temperature scenarios, it is decided to use 30 as the judgment criterion. Subsequently, compare the temperature anomaly index of each record with 30. If the value is greater than 30, mark the corresponding monitoring point as an abnormal temperature point and record the specific time and temperature reading. If the value is between 20 and 30, classify it as a moderately abnormal range and mark it for reminder in the system. If it is less than 20, classify it as the normal level and file it according to the daily inspection process. To make the retrieval process more convenient, these marks can be sorted according to the coordinate grid and stored in an index table. Further count the occurrence frequency of the situation where the value of the same monitoring point is greater than 30 multiple times, and conduct multi-dimensional investigation in combination with external environmental information such as humidity, wind speed, and air pressure data. When all the screening is completed, a comprehensive set of abnormal temperature points is obtained. Finally, form alarm information for these abnormal points and output them in the form of monitoring point ID, coordinates, temperature reading, and anomaly index, so as to generate the abnormal temperature alarm result.

[0045] The steps to obtain the risk level assessment result are as follows: Based on the fire recognition result and the abnormal temperature alarm result, extract the spatial coordinates of the flame area, smoke diffusion area, and abnormal temperature points to obtain the basic fire risk data; According to the basic fire risk data, calculate the fire risk level index. The calculation formula is: ; wherein, is the fire risk level index, is the area of the flame region, is the area of the smoke diffusion region, is the average temperature of the abnormal temperature points, is the center coordinate of the flame, is the center coordinate of the abnormal temperature points; Based on the fire risk level index, set the risk level classification standard, divide the regions with different risk levels, and obtain the risk level assessment results.

[0046] Specifically, based on the previously obtained fire recognition results and abnormal temperature alarm results, first read all the image information with flame category marks and smoke category marks in the database, and associate the confirmed abnormal temperature monitoring points among them. Under the same index, extract the coordinate ranges of the flame region and the smoke diffusion region one by one and summarize them into a regional coordinate list. At the same time, for all the high-temperature monitoring points recorded in the abnormal temperature alarm results, extract their geographical location coordinates and corresponding temperature values. In order to make these coordinate information comparable in the same reference system, it is necessary to unify the ground coordinate system and the aerial image acquisition coordinate system in advance. By searching for flight track information and camera geographical calibration parameters, analyze the relative coordinates during the UAV shooting and convert them to the ground coordinate system. List the processed coordinate data in layers according to the flame region, the smoke diffusion region, and the abnormal temperature points, and then save them in the form of a data table to mark corresponding position identifiers, coordinate values, and area or temperature values and other fields. Subsequently, through the area calculation program, perform contour tracking and grid integration on the pixel plane for the flame region and the smoke diffusion region respectively, multiply the total number of pixels within the contour by the scale factor between the camera and the ground truth, obtain the actual area values corresponding to the flame region and the smoke diffusion region, and cooperate with the previous temperature summary of the abnormal temperature points to perform further arithmetic averaging. Finally, integrate the spatial coordinates and numerical information of the flame region, the smoke diffusion region, and the abnormal temperature points to generate the basic fire risk data.

[0047] The advantage of the formula is that it simultaneously incorporates the area of the flame region, the area of the smoke diffusion region, and the average temperature of the abnormal temperature points, and introduces the relative distance between the flame center and the abnormal temperature point center, comprehensively reflecting the relationship between the flame scale, the smoke impact, and the temperature distribution in space.

[0048] The steps to obtain the parameter are as follows: This parameter represents the area of the flame region. First, the flame region in the flame recognition result needs to be contour-segmented, and the coordinate grid method is used to calculate the actual area of each pixel on the ground projection. If the shooting angle between the camera and the target ground is between 30 degrees and 80 degrees, then trigonometric calculations are performed by combining information such as the camera internal parameters, pitch angle, and flight altitude to determine the actual ground size corresponding to each pixel. Then, the sum of the pixel areas within all flame contours is added up to obtain , whose unit can be square meters. In actual measurement, a calibration board is often arranged outdoors first to obtain the correspondence between camera imaging and actual size, and the calibration board is calibrated in the captured image through an automatic matching algorithm. Then, the final accumulated area is summarized into the database. For example, during a certain acquisition, it is obtained through image comparison that the flame region occupies a number of pixel sets, and after conversion, it is square meters.

[0049] The steps to obtain the parameter are as follows: This parameter represents the area of the smoke diffusion region, which is similar to the method of obtaining . After identifying the smoke distribution contour, it needs to be projected onto the same ground coordinate system and the area is statistically calculated. To eliminate the area error that may be caused by the fuzzy smoke edge band, convolution detection is usually performed on the smoke contour during recognition and an empirically set gray threshold is used for segmentation. Then, the blocks exceeding this threshold are included in the smoke contour. The contour pixels are uniformly projected onto the actual ground and the coordinate system is converted, and then the areas are accumulated. The overlapping regions are removed through multi-view image cross-validation, and the final area value is recorded. For example square meters.

[0050] The steps to obtain the parameter are as follows: This parameter represents the average temperature of abnormal temperature points. First, all the temperatures marked as abnormal high temperature points need to be read from the previously established abnormal temperature monitoring results. Then, after removing redundant data collected repeatedly at the same location in a very short time through duplicate checking, arithmetic averaging is performed. If there are multiple temperature values in the same region that are significantly greater than other records, they can be separately included in additional statistics and then differential analysis is carried out to ensure that the final obtained average temperature can truly reflect the overall level. This average value should be stored in the database in units of °C. For example, the temperatures extracted from 10 abnormal temperature points are 85°C, 88°C, 92°C, 94°C, 90°C, 93°C, 91°C, 87°C, 96°C, 89°C respectively. After adding them up and dividing by 10, we get °C.

[0051] The steps to obtain the parameters are as follows: These two parameters represent the coordinates of the flame center. Usually, it is necessary to first perform aggregation calculation on the flame contour in the geographic coordinate system, find the outermost boundary of the contour, and determine the center position by the centroid method or the extreme value method. When using the centroid method, the coordinates of each pixel point can be multiplied by its brightness or a certain weighting value and then accumulated, and then divided by the total weighting value to obtain the centroid coordinates. Mapping these coordinates to the ground reference system gives , and then retain the values in the same coordinate system. If the flame is distributed in multiple adjacent regions, it can be regarded as a whole, and the weighted centroid method is used for merging to obtain a unique center. For example, after calculation in a synthetic image, 、 indicates that the flame center is at the coordinates (50, 40).

[0052] The steps to obtain the parameters are as follows: These two parameters represent the center coordinates of the abnormal temperature points. When obtaining them, it is necessary to find the average or geometric center among all the identified abnormal temperature coordinates. If there are multiple abnormal temperature points closely distributed, the X values of all their coordinate points can be added up and then divided by the total number to obtain , and similarly, perform operations on the Y values to obtain . If the distribution range of these points is large, regional clustering can also be adopted and the centers are calculated separately for each region. Each center can form a pair of and then participate in the subsequent risk assessment respectively. For example, among 10 abnormal temperature points, the X coordinates are 52, 55, 56, 54, 59, 49, 57, 51, 53, 60. Adding them up and dividing by 10 gives 54.6. After rounding or retaining decimals, it is the center X coordinate. Similarly, Y can be calculated to be 41.8, and finally it can be recorded as 、 .

[0053] Calculation process: When square meters, square meters, °C, 、 、 、 , then: ; ; The numerator part is: ; The denominator part is: ; Therefore: ; The result shows that the fire risk level index is approximately 5656.25. A higher value means a larger area covered by flames and smoke or a higher average abnormal temperature. Also, when the distance between the center of the flame and the center of the abnormal temperature point is relatively close, the risk index will increase accordingly. If it is found through repeated calculations during different monitoring periods that the R value continues to grow, it indicates that the fire situation has worsened and more urgent measures need to be taken.

[0054] Based on the above fire risk level index, multiple levels of risk thresholds can be preset in the system for comparison. For example, areas with a value less than 1000 are marked as low - risk level, areas between 1000 and 3000 are medium - risk level, areas between 3000 and 8000 are high - risk level, and when it is greater than 8000, it is regarded as a severe - risk level. These thresholds can be obtained from the average indicators in the statistics of previous - year fire samples and compared with safety specifications. Then, the R values in different monitoring areas are compared item - by - item with these graded values. If the R value corresponding to a certain monitoring area exceeds 3000, it is classified into the high - risk or higher - level category for registration and listed in the system in sequence according to the coordinate index and time information. To facilitate daily inspections, an automatic screening program can be set in the background. Once it is detected that the R value of an area is greater than 8000, a first - level prompt flag is issued, and the area can be highlighted on the map interface. Finally, a risk - grading table is formed, indicating the risk level and the corresponding R value for each area, which is used to generate the final risk - level assessment result.

[0055] The steps to obtain the energy - saving control result are as follows: Based on the risk - level assessment result, calculate the adjusted working frequency of the ground monitoring cameras. The calculation formula is: ; Where, is the adjusted working frequency, is the original working frequency, is the fire risk level index, is the number of times the fire alarm has been triggered in the current area in the most recent month, is the number of abnormal identifications of flames or smoke in the current area in the most recent month, is the average ambient temperature in the current area; Based on the adjusted working frequency, readjust the sampling interval of each monitoring camera to obtain the energy - saving control result.

[0056] Specifically, the advantage of the formula lies in simultaneously considering multiple factors such as the fire risk level index, the number of times the fire alarm is triggered, the number of abnormal identifications, and the ambient temperature. Through multiplicative regulation of the original working frequency in the form of an index, the monitoring cameras can increase the frequency when the risk is high and decrease the frequency when the risk is low.

[0057] The steps to obtain the parameter are as follows: This parameter represents the original operating frequency, which is usually obtained through on-site testing and recording during the camera installation and debugging phase. The unit is the same as that of and can be used to compare the change situation. When obtaining this value, the operation and maintenance team will limit the frequency to the range of 5 to 30 frames per second according to the clarity requirement and system load capacity during the first debugging, and select a suitable value to store in the monitoring and management platform. For example, when testing a certain model of camera, it is found that when using 20 frames per second, the picture can be kept stable while taking into account the smoothness of the picture and the network bandwidth occupancy. Then 20 is defined as the of this camera.

[0058] The steps to obtain the parameter are as follows: This parameter represents the fire risk level index, which is obtained through monitoring by calculating with the previous formula .

[0059] The steps to obtain the parameter are as follows: This parameter represents the number of times the fire alarm has been triggered in the current area in the most recent month. Usually, it is automatically counted in the background of the fire alarm system and written into the database. Each time the system triggers a fire alarm based on the flame or smoke recognition result and the temperature anomaly index, the current time and geographical location will be recorded in the table and the alarm count will be incremented. When 30 days or a natural month has passed, a new round of counting will start. When counting the total number of times in this month, it is necessary to read the time stamps one by one and ensure that the records belong to this area to avoid cross-regional data interfering with the results, and finally obtain an integer value . For example, if a certain area has 10 alarm triggers within 30 days, then the background will write into the database.

[0060] The steps to obtain the parameter are as follows: This parameter is the number of times of abnormal flame or smoke recognition in the most recent month. In the system, during the image analysis stage, it will be counted when there is a situation where normal identification cannot be carried out or there is a large deviation. Each time such a situation occurs, a record will be generated and bound to the date and geographical location of the day. After a cycle of one month, the total number of times will be formed. For example, if a monitoring point detects three deviations or missed detections in flame recognition and two abnormal smoke recognitions within 30 days, then these five times will be combined and counted into .

[0061] The steps to obtain the parameter are as follows: This parameter represents the average ambient temperature in the current area in the most recent month. When obtaining it, it is necessary to collect the full-day temperature readings in this area within 30 days, which are collected through the ground monitoring station and the temperature measurement equipment carried by the drone, and are distinguished in the system according to the date and coordinates. After putting all the qualified data into a temperature summary table, first eliminate the sensor failures or extreme abnormal values, and then calculate the arithmetic mean of the remaining data to obtain , in degrees Celsius. If the amount of data is large during the statistical period, the daily average can be calculated first and then the secondary average of the daily average can be taken. For example, for a total of 900 valid readings of daytime and nighttime temperatures over 30 days, the sum is added up and divided by 900 to obtain °C.

[0062] Calculation process: In the example scenario, assume that through the previously obtained fire risk level index , and the number of fire alarm triggers this month , the number of abnormal flame or smoke identifications , the average ambient temperature in the most recent month , the original working frequency of a certain camera , substitute all these values into the formula: ; ; ; The denominator has a negative sign: ; Perform the exponential operation: ; So we get: ; This result indicates that the value is extremely close to zero, indicating that the fire risk-related indicators are very high at this time, and the camera working frequency can hardly continue to decrease to achieve energy conservation. It can be determined that the risk factors in this area are much greater than the normal range. If a minimum executable working frequency needs to be set for the same risk level, such as 1 frame per second, it can be truncated at the lower limit after calculation to make not less than 1 frame per second, thus avoiding the situation of complete stoppage of shooting. When subsequent , , these terms are small, the exponential value will increase, making rise closer to or even the same as it.

[0063] Based on the adjusted working frequency calculated above, first retrieve the camera IDs corresponding to all cameras and the Read out information such as the corresponding installation location, etc. Then compare the maximum and minimum operating frequency ranges that the camera itself can execute. For example, there may be a hardware limit of 5 frames per second to 30 frames per second. By comparing these boundary values for correction to avoid exceeding the bandwidth pressure that the device can withstand. When the frequency is lower than 5 frames per second or higher than 30 frames per second, the results are corrected to the feasible range according to the segmented constraints. Then record these corrected results and list a camera sampling interval comparison table. The sampling interval can be obtained by simple reciprocal operation according to the operating frequency. For example, divide one second into the corresponding number of frame times, or a more accurate mapping can be performed according to the actual transmission and processing duration of each frame. When the comparison and correction of all cameras are completed, a list will be obtained, which indicates the new operating frequency and the deduced sampling interval line by line. To better allocate system resources, these sampling intervals will also be linked to the risk levels of the corresponding monitoring areas. For example, set a more compact sampling interval for high-risk areas, so that more dense image data can be obtained for key areas in the same round of detection. In areas with low risk or fewer alarm triggers in the past month, the system will set the sampling interval more loosely. After the operating parameters of all cameras are updated, they will be uniformly transmitted to the front-end control terminal and automatically issued. After the system checks and confirms at the control terminal, a new sampling cycle will be started and take effect immediately. After summarizing these update operations, the new operating frequency configuration and the corresponding energy-saving control results can be output.

[0064] The steps to obtain the emergency response plan are as follows: Integrate fire trucks, fire-fighting personnel, and medical support resources according to the risk level assessment results and energy-saving control results, and assign tasks and schedules to each resource to obtain the emergency response plan.

[0065] Specifically, according to the risk level assessment results and energy-saving control results obtained previously, the risk level of the relevant area and the corrected working frequency of each surveillance camera are first read in the background, and the classification information such as risk area, low-risk area and non-alarm triggered area is recorded. Then, the available number of firefighting vehicles, vehicle type and vehicle equipment parameters are adjusted one by one. For example, the foam fire extinguishing agent capacity that the vehicle can carry is in the range of 200 liters to 500 liters and the performance of the equipment that can withstand the highest temperature. The current number of firefighting personnel on duty, professional division of labor and available deployment time period are retrieved. For example, 20 regular firefighters, 5 full-time high-altitude rescue personnel and 3 mobile special service personnel are registered on site. The vehicle and personnel reserves of medical support resources are retrieved, and the number of teams capable of performing emergency medical rescue and the qualification level of doctors are listed. For example, 2 ambulances with first aid equipment and 4 doctors with advanced trauma treatment qualifications are registered in nearby hospitals. According to this information combined with the risk level assessment results, The high-risk areas or medium-risk areas indicated are allocated and route schedules are set for fire trucks. The departure time, driving distance, estimated time taken in the traffic route and the coordinates of the places to be arrived at are clearly specified for each vehicle. High-risk areas are listed as the priority destinations and then the vehicles in the medium-risk areas are arranged. Firefighting personnel are scheduled according to their professional skills and working time intervals. For example, high-altitude rescue personnel who have participated in hazardous chemical fires are deployed to higher terrain or buildings first. The handover period of each member is recorded and a detailed rotation sequence is arranged. Medical support resources are divided into emergency rescue groups and general treatment groups according to the degree of medical emergency needs, and the equipment carried and the crew establishment of the ambulance are marked. The order of action is given according to the distance and route planning, and the hospital or attendance ratio of each doctor and the time of arrival at the scene are confirmed. The tasks are arranged in a table with hours as the basic unit and marked with execution numbers for easy on-site or remote command and viewing. After summary, an emergency response plan is obtained.

[0066] The present invention provides a fire ground-air integrated monitoring and early warning method, comprising the following steps: Collect image data from ground surveillance cameras and aerial drones, perform pixel-level processing on the collected images, extract image brightness, color, and texture features, and form a feature data set; Based on the feature data set, the pixels in each image are classified and judged to identify the flame and smoke patterns in the image. The fire identification results are generated according to the identified proportion and distribution. Receive temperature data collected by the drone's infrared thermal imaging camera, calculate the maximum, minimum and average values ​​of the temperature data, identify and mark temperature points that are out of the normal range, and generate abnormal temperature alarm results; Combine the fire recognition results with the abnormal temperature alarm results, calculate the current fire risk index, adjust the working frequency of the ground monitoring cameras and the inspection frequency of the drones according to the risk level, and generate an energy-saving control result; Combine the energy-saving control result with the risk level, formulate emergency response measures for different risk levels, and generate an emergency response plan.

Claims

1. A ground-to-air integrated fire monitoring and early warning system, characterized in that: The system comprises: The flame and smoke recognition module collects image data from ground surveillance cameras and aerial drones, extracts features from the image data, generates a feature data set, and uses image analysis based on the feature data set to identify the presence of flames and smoke and generate a fire recognition result; The temperature monitoring module receives temperature data from the infrared thermal imaging camera carried by the aerial drone, performs threshold analysis on the temperature data, generates temperature analysis data, identifies abnormal temperature points based on the temperature analysis data, and generates abnormal temperature alarm results; A resource energy-saving control module determines the current fire risk level according to the fire identification result and the abnormal temperature alarm result, generates a risk level assessment result, adjusts the working frequency of the ground monitoring camera according to the risk level assessment result, and generates an energy-saving control result; The event response and communication module formulates an emergency response strategy and generates an emergency response plan based on the risk level assessment result and the energy-saving control result.

2. The ground-to-air integrated fire monitoring and early warning system according to claim 1 is characterized in that: The steps for obtaining the feature data set are: Collect image data from ground surveillance cameras and aerial drones to obtain a comprehensive image dataset; Based on the comprehensive image data set, the edge, texture and color distribution of the image are extracted through high-pass filter and grayscale processing to construct a feature data set.

3. The ground-to-air integrated fire monitoring and early warning system according to claim 1 is characterized in that: The steps for obtaining the fire identification result are: Based on the feature data set, the flame and smoke features in the image are respectively identified by a convolutional neural network to obtain a preprocessed feature data set for distinguishing flame and smoke; Based on the preprocessed feature data set, the classification result is verified to obtain a fire identification result.

4. The ground-to-air integrated fire monitoring and early warning system according to claim 1 is characterized in that: The steps for obtaining the temperature analysis data are as follows: Receive temperature data from infrared thermal imaging cameras carried by aerial drones, extract infrared temperature information at different time periods and geographical locations, remove sensor errors or reflected signals from invalid areas, classify data according to data collection time, and mark data categories for different seasons to obtain classified temperature data sets; Based on the classified temperature data set, the seasonal adaptive temperature threshold is calculated using the following formula: ； in, is the calculated seasonal adaptive temperature threshold, Representing Winter The temperature data of the sampling points, Representing the summer The temperature data of the sampling points, Represents spring and autumn The temperature data of the sampling points, is the number of sampling points in winter, is the number of sampling points in summer, is the number of sampling points in spring and autumn; Based on the seasonal adaptive temperature threshold, the temperature values ​​of each monitoring point in the temperature data set are analyzed, and the points exceeding the seasonal adaptive temperature threshold are marked as abnormally high temperature areas to generate temperature analysis data.

5. The ground-to-air integrated fire monitoring and early warning system according to claim 1 is characterized in that: The steps for obtaining the abnormal temperature alarm result are: Based on the temperature analysis data, extract temperature point data at different locations within the monitoring area to obtain a temperature change data set; According to the temperature change data set, the temperature anomaly index is calculated, and the calculation formula is: ； in, is the temperature anomaly index, is the current monitoring point temperature, is the ambient background temperature, is the temperature gradient in the horizontal direction, is the temperature gradient in the vertical direction, is the time interval; Based on the temperature anomaly index, abnormal temperature points are screened and abnormal temperature alarm results are generated.

6. The ground-to-air integrated fire monitoring and early warning system according to claim 1, characterized in that: The steps for obtaining the risk level assessment result are: Based on the fire identification result and the abnormal temperature alarm result, the spatial coordinates of the flame area, the smoke diffusion area and the abnormal temperature point are extracted to obtain the basic data of fire risk; According to the fire risk basic data, the fire risk level index is calculated using the following formula: ； in, is the fire risk level index, is the flame area, is the smoke diffusion area, is the average temperature of the abnormal temperature point, is the flame center coordinate, is the center coordinate of the abnormal temperature point; Based on the fire risk level index, a risk level classification standard is set, and areas of different risk levels are divided to obtain a risk level assessment result.

7. The ground-to-air integrated fire monitoring and early warning system according to claim 1, characterized in that: The steps for obtaining the energy-saving control result are: Based on the risk level assessment results, the adjusted working frequency of the ground surveillance camera is calculated using the following formula: ； in, is the adjusted operating frequency, is the original operating frequency, is the fire risk level index, The number of fire alarms triggered in the current area in the past month. Identify the number of fire or smoke anomalies in the current area in the past month. The average ambient temperature of the current area in the last month; Based on the adjusted working frequency, the sampling interval of each monitoring camera is readjusted to obtain an energy-saving control result.

8. The ground-to-air integrated fire monitoring and early warning system according to claim 1, characterized in that: The steps for obtaining the emergency response plan are: According to the risk level assessment result and the energy-saving control result, fire trucks, firefighters and medical support resources are integrated, tasks and schedules are assigned to each resource, and an emergency response plan is obtained.

9. The ground-to-air integrated fire monitoring and early warning method of the ground-to-air integrated fire monitoring and early warning system according to any one of claims 1 to 8, characterized in that: The following steps are involved: Collect image data from ground surveillance cameras and aerial drones, perform pixel-level processing on the collected images, extract image brightness, color, and texture features, and form a feature data set; Based on the feature data set, the pixels in each image are classified and judged to identify the flame and smoke patterns in the image. The fire identification results are generated according to the identified proportion and distribution. Receive temperature data collected by the drone's infrared thermal imaging camera, calculate the maximum, minimum and average values ​​of the temperature data, identify and mark temperature points that are out of the normal range, and generate abnormal temperature alarm results; Combine the fire identification results with the abnormal temperature alarm results to calculate the current fire risk index, adjust the working frequency of ground surveillance cameras and the inspection frequency of drones according to the risk level, and generate energy-saving control results; Combine the energy-saving control results with the risk level, formulate emergency response measures for different risk levels, and generate an emergency response plan.

Images