CN117690088A - Forest fire image recognition on-line monitoring method and system for power transmission line - Google Patents

Abstract

The invention discloses an on-line monitoring method and system for forest mountain fire image recognition of a power transmission line, which relate to the technical field of image monitoring, and are used for integrating a dynamic threshold algorithm of a power transmission line monitoring model of multidimensional weather data, introducing meteorological grid data on the basis of adopting a YOLOv6 model for smoke and fire recognition, preprocessing the data by adopting an inverse distance weight interpolation method, selecting data such as rainfall, wind speed, temperature, humidity and the like, adopting a logistic regression algorithm after normalization, calculating to obtain the dynamic threshold of the model, and simultaneously realizing more refined smoke and fire alarm. The monitoring system adopts the Euclidean distance calculation method to match the towers with the meteorological data, so that each tower can update a smoke and fire alarm threshold value every hour, the alarm accuracy is further improved, the auditing efficiency of business personnel is greatly improved, better auxiliary effect is achieved, the monitoring accuracy is high, and the waste of disaster relief resources is effectively avoided.

Description

Forest fire image recognition on-line monitoring method and system for power transmission line

Technical Field

The invention relates to the technical field of image monitoring, in particular to an on-line monitoring method and system for forest mountain fire image recognition of a power transmission line.

Background

The power transmission line is often built in mountain areas, forests, wilderness and other areas where people smoke is rare, once mountain fires occur in the areas, large-scale power failure accidents are extremely easy to occur, social economy and people safety are seriously threatened, smoke is used as an early signal of fire occurrence, in mountain forests with complex backgrounds, the smoke is easier to observe relative to flames, is used as an early signal of fire occurrence, and in mountain forests with complex backgrounds, the smoke is easier to observe relative to flames, so that the mountain fires can be timely found by detecting the smoke, the expansion of the accidents can be effectively avoided, and in order to cope with the situations, the images of smoke and fire are distinguished and alarmed through image recognition of video image contents collected by monitoring equipment arranged on the power transmission line, and the serious accidents threatening the lives of people and social economy and safety caused by the large-scale power failure accidents and the like caused by the large-area mountain fires are avoided.

The prior art has the following defects:

the existing monitoring system has low recognition precision, if the areas with moist climate or wide areas of mountain areas, hills and water areas are monitored, because the areas are easy to generate cloud and fog and are easy to be mistakenly recognized as smoke by a power transmission channel model, a large number of smoke false alarms exist in the alarm, and the waste of disaster relief resources can be caused while the burden of the system is increased;

disclosure of Invention

Aiming at the defects of the prior art, the invention aims to provide the on-line monitoring method and system for forest mountain fire image recognition of the power transmission line, which are used for introducing weather grid data, fusing rainfall, wind speed, temperature, humidity and other data, realizing the numerical value of fire hazard forecast data, reducing the false judgment of warning caused by a simple image recognition algorithm and improving the warning accuracy.

In order to achieve the above object, the present invention provides the following technical solutions: an on-line monitoring method for forest mountain fire image recognition of a power transmission line, comprising the following steps:

s1: preprocessing the obtained meteorological grid data;

s2: four types of data are selected, wherein the four types of data comprise rainfall, wind speed, temperature and humidity;

s3: normalizing the four types of data, taking the normalized data as the input of logistic regression, and calculating the probability of forest fire occurrence to obtain a smoke and fire alarm dynamic threshold;

s4: performing target detection on smoke and fire under a real scene by using a YOLOv6 model to obtain an area where a target is located and corresponding confidence;

s5: and comparing the confidence coefficient with a dynamic threshold value of the smoke and fire alarm, and judging whether to send out the alarm according to a comparison result.

Preferably, in step S1, preprocessing the weather grid data includes the following steps:

filling the blank value by adopting a spatial interpolation method, wherein the expression of the spatial interpolation method is as follows:

wherein Z (x) _o ) Is x _o Estimated value of point, Z (x _t ) Is x _i The true value of the point, n is the number of weather stations used for interpolation, d _to For the predicted point x _o To a known site x _i P is the power of the distance;

after the full-province meteorological grid data are obtained, n stations closest to the original station are obtained through an inverse distance weight interpolation method, actual measurement values of the n stations at corresponding time points and distances between the stations and the original station are obtained, and interpolation full processing is carried out on meteorological element data.

Preferably, in step S3, the normalization of the four types of data is performed as an input of logistic regression, and the calculation of the probability of occurrence of the forest fire includes the following steps:

the rainfall wind speed, temperature and humidity data are selected as parameters, and a multiple linear regression method is adopted to establish a linear equation for forest fire occurrence;

and normalizing the four types of data to serve as input variables of a logistic regression equation, and solving a multiple linear regression equation to obtain the coefficient of each variable shadow so as to obtain the forest fire occurrence probability alpha.

The rainfall wind speed, temperature and humidity data are selected as parameters, a linear equation for forest fire occurrence is established by adopting a multiple linear regression method, four types of data are normalized and then used as input variables of a logistic regression equation, and coefficients of shadows of each variable are obtained by solving the multiple linear regression equation, so that the forest fire occurrence probability alpha is obtained, specifically:

collecting historical data including rainfall, wind speed, temperature and humidity and whether forest fires happen or not, ensuring data quality and integrity, cleaning the data, processing missing values and abnormal values, checking the distribution situation of the data, possibly needing to normalize or standardize to ensure the consistent scale of different variables, establishing a model by using a multiple linear regression method, taking the rainfall, the wind speed, the temperature and the humidity as independent variables, establishing a linear equation by taking whether forest fires happen or not as dependent variables, solving the multiple linear regression equation by using a statistical tool or programming language to obtain coefficients of the respective variables, converting the result of multiple linear regression into probability by using a logistic regression method, training a logistic regression model by using the historical data, adjusting model parameters, enabling the model to fit the historical data better, evaluating the performance of the model by using a test set or cross verification method and the like, checking whether the model can accurately predict the occurrence of the forest fires, and applying the trained logistic regression model to new data to predict the probability of the forest fires.

Preferably, in step S3, the calculation expression of the dynamic threshold of pyrotechnic warning is:

I＝α×U×C _r ×C _S ；

U＝f(V)+f(T)+f(r _RH )+f(M)；

wherein I represents a dynamic threshold of smoke and fire warning, alpha is the occurrence probability of forest fires, U is a functional expression of forest fire risk meteorological indexes, and C _r And C _S Respectively refer to a precipitation correction coefficient and a snow correction coefficient.

Preferably, in step S4, the target detection of the smoke in the real scene by using the YOLOv6 model includes the following steps:

after training the YOLOv6 model by adopting the smoke and fire related data set, inputting the picture data of the real scene into the YOLOv6 model for target detection, and returning the detected target candidate area and the confidence coefficient by the YOLOv6 model.

Preferably, the performance index of the YOLOv6 model is evaluated by selecting Precision and Recall ratio recapture, and the expression is as follows:

where TP is the case where positive sample prediction is correct, TN is the case where positive sample prediction is incorrect, FP is the case where counterexample prediction is positive, and FN is the case where counterexample prediction is counterexample.

Preferably, in step S5, comparing the confidence level with a dynamic threshold of the pyrotechnic alarm, and determining whether to send out the alarm according to the comparison result includes the following steps:

after the confidence coefficient of the monitoring area is obtained through the YOLOv6 model, the confidence coefficient is compared with the dynamic threshold value of the smoke and fire alarm, if the confidence coefficient is larger than or equal to the dynamic threshold value of the smoke and fire alarm, the monitoring area is judged to be in fire and alarm, and if the confidence coefficient is smaller than the dynamic threshold value of the smoke and fire alarm, the monitoring area is judged to be in no fire and no alarm is sent.

The invention also provides an on-line monitoring system for forest mountain fire image recognition of the power transmission line, which comprises a processing module, a feature selection module, a threshold calculation module, a target detection module and a comparison module:

the processing module is used for: preprocessing the obtained weather grid data, and sending the processed weather grid data to a computer;

the characteristic selection module: four types of data in the weather grid data are selected, wherein the four types of data comprise rainfall, wind speed, temperature and humidity, and the four types of data are sent to the weather grid data;

a threshold calculating module: normalizing the four types of data and then taking the normalized four types of data as the input of logistic regression, calculating the probability of forest fire occurrence, obtaining a smoke fire alarm dynamic threshold value, and transmitting the smoke fire alarm dynamic threshold value to a comparison module;

the target detection module: carrying out target detection on smoke and fire under a real scene by adopting a YOLOv6 model, obtaining a region where the target is located and corresponding confidence coefficient, and sending the confidence coefficient to a comparison module;

and a comparison module: and comparing the confidence coefficient with a dynamic threshold value of the smoke and fire alarm, and judging whether to send out the alarm according to a comparison result.

In the technical scheme, the invention has the technical effects and advantages that:

according to the invention, a dynamic threshold algorithm of a transmission line monitoring model integrating multidimensional weather data is introduced, meteorological grid data is introduced on the basis of smoke and fire identification by using a YOLOv6 model, a reverse distance weight interpolation method is adopted to preprocess the data, rainfall, wind speed, temperature, humidity and other data are selected, a logistic regression algorithm is adopted after normalization, a dynamic threshold of the model is calculated, and meanwhile, in order to realize more refined smoke and fire alarm. The monitoring system adopts the Euclidean distance calculation method to match the towers with the meteorological data, so that each tower can update a smoke and fire alarm threshold value every hour, the alarm accuracy is further improved, the auditing efficiency of business personnel is greatly improved, better auxiliary effect is achieved, the monitoring accuracy is high, and the waste of disaster relief resources is effectively avoided.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings may be obtained according to these drawings for a person having ordinary skill in the art.

FIG. 1 is a flow chart of the method of the present invention.

Fig. 2 is a schematic diagram of a network structure of the YOLOv6 model of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Example 1: referring to fig. 1, the method for on-line monitoring forest mountain fire image recognition of a power transmission line according to the embodiment includes the following steps:

firstly, preprocessing acquired meteorological grid data, filling a vacancy value by adopting a spatial interpolation method, then selecting four types of data (rainfall, wind speed, temperature and humidity), normalizing the data to be used as input of logistic regression, thereby calculating the probability of forest fire occurrence and further obtaining a dynamic threshold value of smoke and fire warning, meanwhile, adopting a YOLOv6 model to detect smoke and fire in a real scene to obtain a region where a target is located and a corresponding confidence level, and finally comparing the confidence level with the dynamic threshold value of smoke and fire warning to be used as a evidence for warning whether the follow-up warning is carried out.

In actual operation, the meteorological observation station is subjected to various interferences such as lightning strike and the like to cause the loss of meteorological element data, which directly affects the accuracy of the smoke and fire alarm model, so that preprocessing the meteorological grid data comprises the following steps:

according to the first law of geography, the closer the distance in space is, the higher the similarity of sites is, so that the inverse distance weight interpolation method in the spatial interpolation method is one of the most commonly used spatial interpolation methods, the inverse distance weight weighting belongs to accurate interpolation, the accurate interpolation means that the surface must pass through each known sample point, based on the similar principle, the closer the distance from a test point is, the larger the weight is, and the method is more suitable for mountain areas or areas where precipitation sites are not very dense, and the principle is shown as the following formula:

wherein Z (x) _o ) Is x _o Estimated value of point, Z (x _i ) Is x _i The true value of the point, n is the number of weather stations used for interpolation, d _to For the predicted point x _o To a known site x _i P is the power of the distance.

Therefore, after the full-province weather grid data are obtained, the method adopts an inverse distance weight interpolation method to obtain n stations closest to the original station, obtains actual measurement values of the n stations at the time point and the distances between the stations and the original station, and brings the actual measurement values into the above steps, so that the weather element data are interpolated completely, and the calculation processing of subsequent steps is facilitated.

Aiming at the problem that partial false alarm exists in the alarm of the transmission line monitoring application, so that auditing pressure is brought to basic service personnel, the invention provides a dynamic threshold algorithm of a transmission line monitoring model fused with multidimensional weather data, full-saving weather grid data is introduced on the basis of adopting a YOLOv6 model for smoke and fire identification, a reverse distance weight interpolation method is adopted for preprocessing the data, rainfall, wind speed, temperature, humidity and other data are selected, and a logistic regression algorithm is adopted after normalization is carried out, so that a dynamic threshold of the model is calculated; meanwhile, in order to realize finer smoke and fire alarming, the European distance calculation method is adopted, the towers are matched with 1.2 kilowatt-hour weather data in the whole province, so that one smoke and fire alarming threshold value is updated every hour for each tower, the alarming accuracy is further improved, the recognition accuracy of the algorithm in a real scene reaches 0.864, after the algorithm is put into use, the alarming times of the whole province power transmission line are reduced from 10 kilowatt-hour to 6 kilowatt-hour, the auditing efficiency of business personnel is greatly improved, and better auxiliary effects are achieved.

Example 2: the influence of a single meteorological factor on the occurrence of a forest fire has the characteristic of poor prediction precision, and the forest fire prediction requires the comprehensive effect of multiple meteorological factors;

therefore, after preprocessing the all-province weather grid data, carrying out correlation analysis, selecting 24-hour rainfall, 14:00 wind speed on the same day, temperature and humidity data as parameters, adopting a multiple linear regression method to establish a linear equation for forest fire occurrence, normalizing the data of four weather factors, using the normalized data as input variables of a logistic regression equation, and obtaining a coefficient of each variable shadow by solving the multiple linear regression equation so as to obtain the forest fire occurrence probability alpha;

the rainfall wind speed, temperature and humidity data are selected as parameters, a linear equation for forest fire occurrence is established by adopting a multiple linear regression method, four types of data are normalized and then used as input variables of a logistic regression equation, and coefficients of shadows of each variable are obtained by solving the multiple linear regression equation, so that the forest fire occurrence probability alpha is obtained, specifically:

collecting historical data including rainfall, wind speed, temperature and humidity and whether forest fires happen or not, ensuring data quality and integrity, cleaning the data, processing missing values and abnormal values, checking the distribution situation of the data, possibly needing to normalize or standardize to ensure the consistent scale of different variables, establishing a model by using a multiple linear regression method, taking the rainfall, the wind speed, the temperature and the humidity as independent variables, establishing a linear equation by taking whether forest fires happen or not as dependent variables, solving the multiple linear regression equation by using a statistical tool or programming language to obtain coefficients of the respective variables, converting the result of multiple linear regression into probability by using a logistic regression method, training a logistic regression model by using the historical data, adjusting model parameters, enabling the model to fit the historical data better, evaluating the performance of the model by using a test set or cross verification method and the like, checking whether the model can accurately predict the occurrence of the forest fires, and applying the trained logistic regression model to new data to predict the probability of the forest fires.

Taking 2018-2022 as an example, the present application calculates the alpha value by adopting the method as shown in table 1, and takes the alpha value into the following formula to calculate the current smoke alarm dynamic threshold value I.

I＝α×U×C _r ×C _S ；

U＝f(V)+f(T)+f(r _RH )+f(M)；

Wherein alpha is the occurrence probability of forest fires, U is a functional expression of forest fire weather index in forest fire weather class, C _r And C _S All refer to meteorological drought grade and refer to precipitation correction coefficient and snow correction coefficient respectively;

table 1: alpha value calculation results of 2018-2022

Year of year	2018	2019	2020	2021	2022
						Alpha value	0.74	0.73	0.77	0.8	0.7

Based on the superior performance of the YOLOv6 model in terms of detection precision and speed, the application adopts the YOLOv6 model to carry out target detection on smoke;

after training a model by adopting a smoke and fire related data set, inputting image data of a real scene into the model for target detection, returning the detected target candidate area and confidence coefficient by an algorithm, and judging whether the scene has fire or not and carrying out subsequent alarm processing by comparing the smoke and fire detection confidence coefficient with a smoke and fire alarm dynamic threshold value;

the yellow 6 model was trained by selecting the kaullage 2022 dataset, which includes MODIS, VIIRS, copernicusSentinel-2 and Landsat-8 geosynchronous satellites that are used throughout the world for flame detection due to their excellent time accuracy and ability to remotely detect flames, and which contains satellite images of forests in addition to images from Google and kagle, the dataset classifying the images into 4 classes: fire, no fire, smoke and fire, the interior of which contained 4800 images, of which 80% of the data was used for training and 10% for testing and validation.

The basic theory involved in the method is as follows:

1) Logistic regression algorithm

Logistic regression is one of the most popular mathematical modeling methods that can be used to determine the relationship between several independent variables and two-class dependent variables and predict the probability of occurrence of an event, in a regression model, an equation is established that predicts the value of the dependent variable based on one or more independent variables, which will input the variable x ₁ ，x ₂ ，...，x _n Mapping to a predicted output variable y, the value of which is between 0 and 1, representing the probability of an event occurring;

in fire probability modeling, the goal of logistic regression is to find the best model describing the relationship between the occurrence or non-occurrence (i.e., dependent variables) of a fire and the fire influencing factors (i.e., a set of independent variables), the general form of the logistic regression equation is:

wherein z=w ₀ +w ₁ x ₁ +w ₂ x ₂ +…+w _n x _n Is a linear combination of input features and their corresponding weights, w ₀ Is a bias term. The goal of the logistic regression algorithm is to find the weight w ₀ 、w ₁ 、…、w _n Thereby minimizing the error between the predicted output variable and the true output variable, which is estimated by a maximum likelihood method;

one of the main advantages of logistic regression is its simplicity and ease of implementation, which is a simple algorithm that can be easily implemented using standard statistical software, and moreover logistic regression is highly interpretable, enabling the user to understand the contribution of each independent variable to the prediction probability, which is also robust in terms of multiple collinearity, meaning that it can handle the related independent variable without producing biased estimates;

however, logical regression is not without drawbacks, the main limitation being that it only works for binary classification problems, meaning that it can only predict the likelihood of an event occurring or not.

Real-time object detection has become a key component in many applications, covering various fields of automatic driving automobiles, robots, video monitoring, augmented reality and the like, and in various object detection algorithms, a YOLO frame stands out with excellent speed and accuracy balance, and can rapidly and reliably identify objects in images;

Joseph-Redmon et al, in 2016, first proposed a real-time end-to-end target detection method YOLO (You Only Look Once), in which all candidate frames can be predicted only by once image transmission through the network, so that compared with the two-stage algorithm, YOLO has a strong advantage in detection speed and is widely applied to industrial-level target detection applications;

according to the method, the YOLOv6 network is adopted as a backbone network for target identification, the whole network structure of the YOLOv6 is shown in fig. 2, the network consists of a first part of high-efficiency backbone network, a second part of PAN topology neck and a third part of high-efficiency decoupling head with a mixed channel strategy, the high-efficiency backbone network structure improves the characteristic extraction performance on the basis of accelerating the calculation speed and saving the operation cost, and the decoupling head structure enables the network to be more efficient, so that compared with YOLOv5, YOLOv6 is improved in multiple aspects such as loss functions, deployment and training strategies.

Meanwhile, in order to adapt to different use scenes, the Yolov6 adopts a plurality of latest technologies including self-distillation, TAL and SimOTA label distribution technologies, SIoU regression loss, anchor-free, repOptimizer-QAT and the like, and in general, a plurality of pre-training models are provided, and the Yolov6 exceeds other target detection algorithms in terms of detection precision and speed;

the Precision and Recall rate are selected as algorithm performance evaluation indexes, and are common measures for evaluating the performance of a target detection model, and the formula is as follows:

where TP is the case where positive sample prediction is correct, TN is the case where positive sample prediction is incorrect, FP is the case where counterexample prediction is positive, and FN is the case where counterexample prediction is counterexample.

1. Algorithm for fusing multidimensional weather data: the invention has the core of realizing the numerical value of the fire hazard forecast by integrating the multidimensional weather data such as rainfall, wind speed, temperature, humidity and the like. The algorithm can more accurately predict the fire risk situation and early warn the mountain fire risk in advance.

2. The matching method of the pole tower and the meteorological data comprises the following steps: by means of the European distance calculation method, 10 tens of thousands of towers are matched with 1.2 tens of thousands of meteorological data in the whole province, and false alarm is avoided.

3. Dynamic threshold update mechanism: according to the smoke condition monitored in real time, each tower updates a smoke and fire alarm threshold value every hour, so that the real-time performance and accuracy of alarm are improved.

The invention aims to improve the accuracy and reliability of fire hazard prediction, prevent misjudgment and missing report, and ensure the accuracy of data matching by adopting an accurate Euclidean distance calculation method when processing and matching the towers with meteorological data, and update a smoke and fire alarm threshold value per hour of each tower so as to ensure the timeliness and accuracy of alarm.

In summary, the technical key point of the invention is that an algorithm for fusing multidimensional weather data, a matching method of a tower and weather data and a dynamic threshold updating mechanism are integrated; the point to be protected is the accuracy and reliability of the algorithm, the accuracy of data processing and matching, and the timeliness of dynamic threshold updating.

Comparing the confidence coefficient with a dynamic threshold value of the smoke and fire alarm, and judging whether to send out an alarm according to a comparison result comprises the following steps:

after the confidence coefficient of the monitoring area is obtained through the YOLOv6 model, the confidence coefficient is compared with the dynamic threshold value of the smoke and fire alarm, if the confidence coefficient is larger than or equal to the dynamic threshold value of the smoke and fire alarm, the monitoring area is judged to be in fire and alarm, and if the confidence coefficient is smaller than the dynamic threshold value of the smoke and fire alarm, the monitoring area is judged to be in no fire and no alarm is sent.

Example 3: the embodiment of the system for on-line monitoring of forest mountain fire image recognition of the power transmission line comprises a processing module, a feature selection module, a threshold calculation module, a target detection module and a comparison module:

the processing module is used for: preprocessing the obtained weather grid data, and sending the processed weather grid data to a computer;

the characteristic selection module: four types of data in the weather grid data are selected, wherein the four types of data comprise rainfall, wind speed, temperature and humidity, and the four types of data are sent to the weather grid data;

a threshold calculating module: normalizing the four types of data and then taking the normalized four types of data as the input of logistic regression, calculating the probability of forest fire occurrence, obtaining a smoke fire alarm dynamic threshold value, and transmitting the smoke fire alarm dynamic threshold value to a comparison module;

the target detection module: carrying out target detection on smoke and fire under a real scene by adopting a YOLOv6 model, obtaining a region where the target is located and corresponding confidence coefficient, and sending the confidence coefficient to a comparison module;

and a comparison module: and comparing the confidence coefficient with a dynamic threshold value of the smoke and fire alarm, and judging whether to send out the alarm according to a comparison result.

The above formulas are all formulas with dimensions removed and numerical values calculated, the formulas are formulas with a large amount of data collected for software simulation to obtain the latest real situation, and preset parameters in the formulas are set by those skilled in the art according to the actual situation.

In the description of the present specification, the descriptions of the terms "one embodiment," "example," "specific example," and the like, mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The preferred embodiments of the invention disclosed above are intended only to assist in the explanation of the invention. The preferred embodiments are not intended to be exhaustive or to limit the invention to the precise form disclosed. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best understand and utilize the invention. The invention is limited only by the claims and the full scope and equivalents thereof.

Claims (8)

1. An on-line monitoring method for forest mountain fire image recognition of a power transmission line is characterized by comprising the following steps of: the monitoring method comprises the following steps:

s1: preprocessing the obtained meteorological grid data;

s2: four types of data are selected, wherein the four types of data comprise rainfall, wind speed, temperature and humidity;

s3: normalizing the four types of data, taking the normalized data as the input of logistic regression, and calculating the probability of forest fire occurrence to obtain a smoke and fire alarm dynamic threshold;

s4: performing target detection on smoke and fire under a real scene by using a YOLOv6 model to obtain an area where a target is located and corresponding confidence;

s5: and comparing the confidence coefficient with a dynamic threshold value of the smoke and fire alarm, and judging whether to send out the alarm according to a comparison result.

2. The method for on-line monitoring of forest fire image recognition of the power transmission line according to claim 1, wherein the method comprises the following steps: in step S1, preprocessing weather grid data includes the following steps:

filling the blank value by adopting a spatial interpolation method, wherein the expression of the spatial interpolation method is as follows:

wherein Z (x) _v ) Is x _v Estimated value of point, Z (x _i ) Is x _i The true value of the point, n is the number of weather stations used for interpolation, d _to For the predicted point x _o To a known site x _i P is the power of the distance;

after the full-province meteorological grid data are obtained, n stations closest to the original station are obtained through an inverse distance weight interpolation method, actual measurement values of the n stations at corresponding time points and distances between the stations and the original station are obtained, and interpolation full processing is carried out on meteorological element data.

3. The method for on-line monitoring of forest fire image recognition of the power transmission line according to claim 2, wherein the method comprises the following steps: in step S3, the four types of data are normalized and then used as the input of logistic regression, and the calculation of the probability of forest fire occurrence comprises the following steps:

the rainfall wind speed, temperature and humidity data are selected as parameters, and a multiple linear regression method is adopted to establish a linear equation for forest fire occurrence;

and normalizing the four types of data to serve as input variables of a logistic regression equation, and solving a multiple linear regression equation to obtain the coefficient of each variable shadow so as to obtain the forest fire occurrence probability alpha.

4. The method for on-line monitoring of forest fire image recognition of a power transmission line according to claim 3, wherein the method comprises the following steps: in step S3, the calculation expression of the dynamic threshold of the smoke and fire alarm is:

I＝α×U×C _r ×C _S ；

U＝f(V)+f(T)+f(r _RH )+f(M)；

wherein I represents a dynamic threshold of smoke and fire warning, alpha is the occurrence probability of forest fires, U is a functional expression of forest fire risk meteorological indexes, and C _r And C _S Respectively refer to a precipitation correction coefficient and a snow correction coefficient.

5. The on-line monitoring method for forest fire image recognition of the power transmission line according to claim 4, wherein the method comprises the following steps: in step S4, the target detection of the smoke in the real scene by using the YOLOv6 model includes the following steps:

after training the YOLOv6 model by adopting the smoke and fire related data set, inputting the picture data of the real scene into the YOLOv6 model for target detection, and returning the detected target candidate area and the confidence coefficient by the YOLOv6 model.

6. The on-line monitoring method for forest fire image recognition of the power transmission line according to claim 5, wherein the method comprises the following steps of: selecting Precision and Recall rate Recall to evaluate the performance index of the YOLOv6 model, wherein the expression is as follows:

where TP is the case where positive sample prediction is correct, TN is the case where positive sample prediction is incorrect, FP is the case where counterexample prediction is positive, and FN is the case where counterexample prediction is counterexample.

7. The on-line monitoring method for forest fire image recognition of the power transmission line according to claim 6, wherein the method comprises the following steps: in step S5, comparing the confidence level with a dynamic threshold of smoke and fire alarm, and judging whether to send out an alarm according to the comparison result includes the following steps:

after the confidence coefficient of the monitoring area is obtained through the YOLOv6 model, the confidence coefficient is compared with the dynamic threshold value of the smoke and fire alarm, if the confidence coefficient is larger than or equal to the dynamic threshold value of the smoke and fire alarm, the monitoring area is judged to be in fire and alarm, and if the confidence coefficient is smaller than the dynamic threshold value of the smoke and fire alarm, the monitoring area is judged to be in no fire and no alarm is sent.

8. An on-line monitoring system for forest mountain fire image recognition of a power transmission line, which is realized by the monitoring method as claimed in any one of claims 1 to 7, and is characterized in that: the device comprises a processing module, a feature selection module, a threshold calculation module, a target detection module and a comparison module:

the processing module is used for: preprocessing the obtained weather grid data, and sending the processed weather grid data to a computer;

the characteristic selection module: four types of data in the weather grid data are selected, wherein the four types of data comprise rainfall, wind speed, temperature and humidity, and the four types of data are sent to the weather grid data;

a threshold calculating module: normalizing the four types of data and then taking the normalized four types of data as the input of logistic regression, calculating the probability of forest fire occurrence, obtaining a smoke fire alarm dynamic threshold value, and transmitting the smoke fire alarm dynamic threshold value to a comparison module;

the target detection module: carrying out target detection on smoke and fire under a real scene by adopting a YOLOv6 model, obtaining a region where the target is located and corresponding confidence coefficient, and sending the confidence coefficient to a comparison module;

and a comparison module: and comparing the confidence coefficient with a dynamic threshold value of the smoke and fire alarm, and judging whether to send out the alarm according to a comparison result.