Unmanned aerial vehicle fire monitoring method based on multi-sensor information fusion technology

Abstract

This invention discloses a drone-based fire monitoring method based on multi-sensor information fusion technology. The method includes the following steps: mounting sensors on the drone to collect environmental data; performing dimensionality reduction on the collected environmental data to generate a principal component matrix; inputting the principal component matrix into an improved ANFIS model trained on historical fire environmental data, whereby the improved ANFIS model performs information fusion and obtains a fusion decision result matrix; comparing each fusion decision result in the fusion decision result matrix with a threshold; if none exceed the threshold, returning to the initial step; if any fusion decision result exceeds the threshold, a fire is identified and its specific location information is output. This invention achieves complementary advantages between data dimensionality reduction and the ANFIS model during sensor data processing, enabling fusion decision analysis of various fire monitoring data acquired by sensors, thereby improving the accuracy of fire monitoring.

Description

Technical Field

[0001] This invention relates to the field of information fusion technology, specifically to a method for monitoring fires using unmanned aerial vehicles (UAVs) based on multi-sensor information fusion technology. Background Technology

[0002] Fire is one of the major disasters threatening environmental safety and sustainable development. With the development and promotion of information technology, the use of various information technologies for effective environmental monitoring and early warning of fires is of great significance for ecological protection and property protection.

[0003] Using drones equipped with sensors for fire environmental monitoring is a flexible, mobile, intelligent, and effective monitoring method. Common methods in this field include sensor data integration and visual sensor judgment. Sensor data integration involves collecting independent information from different sensors and, based on the experience of relevant personnel, determining the fire situation. Its main problem lies in the often partial and incomplete nature of the information acquired, and the lack of objectivity in the judgment criteria. Visual sensor judgment utilizes visual sensors to extract color indicators and image data in the RGB space to determine the fire situation. However, its problem is that color and image information are difficult to accurately determine in complex situations such as interference from other objects, and it lacks information fusion processing for situations of information redundancy or conflict.

[0004] With the continuous development of computer technology, deep learning methods, represented by fuzzy neural networks, can also achieve better fusion processing of multi-sensor information. Deep learning methods utilize their self-learning capabilities to filter and mine all collected fire environment data, automatically improving the relationships between different data points. However, problems arise such as inconsistent data dimensions and representations across sensor information; complex network structures in the processing system; unclear interpretive relationships between the processed data and the output decision results; weak processing capabilities; and slow processing speed. Therefore, there is an urgent need to propose a monitoring method for fire situations, optimize this type of multi-sensor information fusion technology, and improve the efficiency and accuracy of fire detection and early warning. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned deficiencies in the prior art and provide a method for monitoring fires by unmanned aerial vehicles (UAVs) based on multi-sensor information fusion technology.

[0006] The objective of this invention can be achieved by adopting the following technical solutions:

[0007] A method for monitoring fires using unmanned aerial vehicles (UAVs) based on multi-sensor information fusion technology, the method comprising the following steps:

[0008] S1. Equip a drone with n sensors. Each sensor collects environmental data from p locations, summarizes and transmits the data, and generates an input matrix with a total of p samples and n variable dimensions. Where x ij This represents the observation data of the i-th sample and the j-th variable dimension;

[0009] S2. The input matrix obtained in step S1 Perform dimensionality reduction on the input matrix. Converted into a principal component matrix Y = [y] representing p samples and m feature variables. ik ] p×m , where y ik This represents the principal component data of the i-th sample and the k-th feature variable dimension;

[0010] S3. Input the principal component matrix Y into the improved ANFIS model to obtain the fused decision result matrix. in This represents the fusion decision result for the i-th location;

[0011] The improved ANFIS model consists of a first layer (fuzzification layer), a second layer (rule layer), a third layer (normalization layer), a fourth layer (rule output layer), and a fifth layer (output layer). The improved ANFIS model uses historical fire environment data and stops when the set maximum number of iterations (epochs) is reached. During the iteration process, data is forward-transmitted to the fourth layer of the improved ANFIS model, where the error rate of the multiple linear regression is calculated based on the language rule function, and then backward-transmitted to the first layer to optimize the membership function parameters of the first layer.

[0012] S4. Set a threshold for determining the occurrence of a fire, and process the fusion decision result matrix obtained in step S3. The decision is evaluated based on all fusion decision results: if none of the fusion decision results exceed the threshold, then proceed to step S1; if none of the fusion decision results at location i exceed the threshold, then proceed to step S1. If the threshold is exceeded, it is determined that a fire has occurred at that location and the location information is output.

[0013] Furthermore, in step S1, the drone and sensors fully collect environmental data of the monitored location, covering as many fire-related physical parameter values ​​as possible, and fix them into a matrix representation to reflect the actual environmental conditions to the greatest extent. The specific processing procedure is as follows:

[0014] By equipping a drone with n sensors, each sensor collects environmental data, including temperature, humidity, and carbon dioxide concentration, at p locations. Simultaneously, the collected environmental data is aggregated and transmitted, generating an input matrix with p samples and n variable dimensions. In the form of x ij represents the environmental data of the \(i\)-th sample in the \(j\)-th dimension.

[0015] Furthermore, in the step S2, through data processing and calculation of the cumulative contribution rate of the principal components, dimensionality reduction of the input matrix is achieved; while ensuring the accuracy of the results, multi-dimensional data is converted into low-dimensional data, simplifying the data form while retaining the characteristic information of the high-dimensional data, obtaining the principal component matrix input to the improved ANFIS model, reducing the complexity of the network structure and the running processing time of the subsequent improved ANFIS model. The specific processing process of the above step of obtaining the principal component matrix Y is as follows:

[0016] S201. Perform standardization processing on the input matrix to obtain the standardized matrix

[0017]

[0018] where is the standardized index variable of the \(j\)-th dimension of the \(i\)-th sample; s j are respectively the arithmetic mean and standard deviation of the \(j\)-th variable, \(i = 1, 2, \cdots, p\), \(j = 1, 2, \cdots, n\);

[0019] S202. Construct the correlation coefficient matrix \(R = [r ij n×n , where \(r ij is the correlation coefficient of the \(j\)-th variable dimension of the \(i\)-th sample;

[0020] S203. Calculate the characteristic equation of \(R\) to obtain \(n\) eigenvalues of \(R\), and arrange them in descending order as the eigenvalue matrix \(\lambda = [\lambda_1, \lambda_2, \cdots, \lambda n , and the eigenvector matrix \(W = [w_1, w_2, \cdots, w n corresponding to each eigenvalue, where \(w j = [w 1j , w 2j , \cdots, w nj T ;

[0021] S204. Select the first \(m\) eigenvalues of \(R\), \(m < n\), such that the corresponding cumulative contribution rate of the principal components and obtain the eigenvector matrix \(W M = where k = 1, 2, \cdots, m;

[0022] ​​​S205. Using the normalized matrix X obtained in step S201 and the eigenvector matrix W obtained in step S204... m Construct the principal component matrix Y = XW m =[y ik ] p×m , where y ik This represents the value of the k-th principal component of the i-th sample.

[0023] Further, in step S3, a corresponding principal component matrix is ​​constructed based on environmental data from the historical fire environment database. An improved ANFIS model is then set and trained to enable it to process multi-dimensional data of the matrix according to specific steps, thereby improving the model's processing capability. During training, a multiple linear regression method is used, setting training optimization objectives and a specific number of iterations. The language rule function of the ANFIS model is adjusted to reduce the error between theoretical and actual results, improving the accuracy of fire judgment. Finally, the current environmental data is input to obtain the fused decision result, which is then summarized into a fused decision result matrix. The specific processing procedure is as follows:

[0024] S301. Using environmental data from the historical fire environment database, construct an input matrix with p samples and n variable dimensions, the same size as in step S1. x′ ij Represents the input matrix The i-th sample, the j-th dimension of the environmental data;

[0025] S302. For the input matrix obtained in step S301 Repeat step S2 to construct the same principal component matrix Y′=[y′] with p samples and m feature variables as in step S2. ik ] p×m y′ ik This represents the principal component data of the i-th sample and the k-th feature variable dimension of the principal component matrix Y′;

[0026] S303. Input the principal component matrix Y′ obtained in step S302 into the first fuzzification layer of the improved ANFIS model. This layer consists of adaptive nodes composed of membership functions. Obtain the membership function values ​​corresponding to each element of Y′, denoted as:

[0027] Q 1,ik =μ(y′) ik )k=1,2,…,m, i=1,2,…,p;

[0028] S304, the membership function values ​​Q obtained in step S303 1,ik Entering the second rule layer, the strength calculation formula is used. Get the rule strength value Q for each row 2,i ;

[0029] S305, the rule strength value Q obtained in step S304 2,i Entering the third normalization layer, according to the formula Obtain the normalized strength rule value Q 3,i ;

[0030] S306. Extract all elements in the i-th row of the principal component matrix Y′ to form a completely linear language rule function, expressed as follows: Where p ij Let ε be the slope of the language rule function. i The error value between the estimated value and the actual value calculated for the language rule function is the parameter value obtained by performing multiple linear regression using the least squares method for multiple variables; the error value ε is stored and transmitted backward. i Upgrade to the first blurring layer, and simultaneously obtain the rule output value Q. 4,i =Q 3,i f i ;

[0031] S307, Rule Output Value Q 4,i Entering the fifth output layer, the results are summed and finally output as the fused decision result Q. 5,i The formula is And summarize them into a fusion decision result matrix Q = [Q 5,i ] p×1 Q 5,i This represents the fusion decision result for the i-th location;

[0032] S308. Set the training iteration count (epochs) to 600, and repeat steps S301 to S307, using the error term ε of the language rule function. i =0, Q 5,i =1 is the training objective. The iteration ends after reaching the maximum value of epochs, thus completing the optimization of the parameter c. j d j The training is optimized to obtain a well-trained, improved ANFIS model.

[0033] S309. Input the principal component matrix Y obtained in step S2 into the improved ANFIS model trained in step S308, and repeat steps S303 to S307 to complete the fusion of multi-sensor information. Output the fusion decision result matrix in probabilistic form.

[0034] Furthermore, the membership function μ(y′) ik The settings are as follows:

[0035] Where c j d jThe parameters to be optimized are those for improving the adaptive function of the ANFIS model; c j d represents the center position of the peak of the membership function. j This represents the width of the interval where the function value of the membership function is greater than zero.

[0036] Furthermore, the membership function μ(y′) ik It is set to the following standard Gaussian function form:

[0037] Where e j f j The parameters to be optimized to improve the adaptive function of the ANFIS model; e j f represents the coordinates of the center of the Gaussian function peak. j This represents the standard deviation of the Gaussian function.

[0038] Furthermore, step S4 completes the fire determination and information output, and the specific process is as follows:

[0039] The threshold for determining whether a fire has occurred is set to [0.05, 0.15], with a preferred value of 0.10. The fusion decision result matrix is ​​then used for evaluation. Check if all fusion decision results exceed the threshold; if none exceed the threshold, it is determined that no fire has occurred, the drone changes location, and returns to step S1; if the fusion decision result of the i-th location... If the threshold is exceeded, a fire is determined to have occurred at that location, and the specific location information is output.

[0040] The present invention has the following advantages and effects compared with the prior art:

[0041] (1) The present invention performs matrix processing on all environmental data collected by the sensor. By inputting the matrix and using the principal component matrix method, the representation and processing of environmental data are unified.

[0042] (2) The present invention uses a data dimensionality reduction method to obtain the principal component matrix, which simplifies the data form while retaining the high-dimensional data feature information, thereby reducing the complexity of nodes in the subsequent improved ANFIS model and improving the speed of model operation and processing.

[0043] (3) This invention uses an improved ANFIS model, which makes adaptive adjustments to the language rule function and the operation processing methods within each layer, so that the structure of each layer is adapted to the form of the input matrix, clarifies the optimization goal of training, and thus improves the processing capability of the model and the accuracy of fire prediction. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0045] Figure 1 This is a flowchart of a drone fire monitoring method based on multi-sensor information fusion technology disclosed in this invention;

[0046] Figure 2 This is a structural diagram of the improved ANFIS model used in the embodiments of the present invention;

[0047] Figure 3 This is an implementation effect diagram obtained by connecting locations with equal fusion decision results and exceeding a threshold after implementing Embodiment 1 of the present invention using curves;

[0048] Figure 4 This is an implementation effect diagram obtained by connecting locations with equal fusion decision results and exceeding a threshold after implementing Embodiment 2 of the present invention using curves. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] This embodiment will adopt Figure 1 The flowchart and Figure 2 The diagram shown illustrates a UAV fire monitoring method based on multi-sensor information fusion technology. This method monitors an area measuring 3000 meters long x 3000 meters wide, achieving information fusion from various sensors to determine the fire situation. The specific process is as follows:

[0052] S1. Equip the F450 drone platform with four sensors (temperature, humidity, smoke, and infrared sensors), fly the drone platform to a fixed altitude above the area to be measured to collect data, with each data collection point spaced 150 meters apart, and collect environmental data from a total of 400 locations within the range. After data collection and transmission, a total of 400 samples and an input matrix with four variable dimensions are generated. Where x ij This represents the observation data of the i-th sample and the j-th variable dimension;

[0053] S2. The input matrix obtained in step S1 Perform dimensionality reduction on the input matrix. Converted to a principal component matrix y = [y] representing 400 samples and 2 feature variables. ik ] 400×2 , where y ik This represents the principal component data of the i-th sample and the k-th feature variable dimension. The specific processing procedure is as follows:

[0054] S201, For the input matrix Perform standardization processing to obtain the standardization matrix.

[0055]

[0056] in, Let j be the standardized index variable of sample i; s j Then these are the arithmetic mean and standard deviation of the j-dimensional variables, i = 1, 2, ..., 400, j = 1, 2, 3, 4;

[0057] S202, According to the formula Construct the correlation coefficient matrix R = [r ij ] 4×4 , where r ij Let be the correlation coefficient of the j-th variable dimension of sample i;

[0058] S203. Calculate the characteristic equation of R, obtain the four eigenvalues ​​of R, and arrange them in descending order to form the eigenvalue matrix λ = [λ1, λ2, λ3, λ4], and the corresponding eigenvector matrix W = [w1, w2, w3, w4], where w j =[w 1j ,w 2j ,w 3j ,w 4j ] T ;

[0059] S204. Calculate the cumulative contribution rate of the corresponding principal components. We can select the first two eigenvalues ​​of R and obtain the corresponding eigenvector matrix. in k = 1, 2;

[0060] S205. Using the normalized matrix X obtained in step S201 and the eigenvector matrix W obtained in step S204... m Construct the principal component matrix Y = XW m =[y ik ] 400×2 , where yik This represents the value of the k-th principal component of the i-th sample;

[0061] S3. Input the principal component matrix Y into the improved ANFIS model to obtain the fused decision result matrix. in This represents the fusion decision result for the i-th location;

[0062] The improved ANFIS model consists of a first layer (fuzzification layer), a second layer (rule layer), a third layer (normalization layer), a fourth layer (rule output layer), and a fifth layer (output layer). The specific operation process is as follows:

[0063] S301. Using previously known fire environment data (temperature, humidity, smoke, and infrared information) obtained by drones, construct an input matrix of the same size as in step S1, containing 400 samples and four variable dimensions. x′ ij Represents the input matrix The i-th sample and the j-th dimension of the environment data are used as the training set and the test set, respectively.

[0064] S302. For the input matrix obtained in step S301 Repeat step S2 to construct the same 400 samples as in step S2, with a principal component matrix Y′ = [y′] represented by two feature variable dimensions. ik ] 400×2 y′ ik This represents the principal component data of the i-th sample and the k-th feature variable dimension of the principal component matrix Y′;

[0065] S303. Input the principal component matrix Y′ obtained in step S302 into the first fuzzification layer of the improved ANFIS model. This layer consists of adaptive nodes composed of membership functions. Obtain the membership function values ​​corresponding to each element of Y′, denoted as:

[0066] Q 1,ik =μ(y′) ik k = 1, 2, i = 1, 2, ..., 400;

[0067] One possible setting for the membership function is as follows:

[0068] Where c j d j The parameters to be optimized are those for improving the adaptive function of the ANFIS model; c j d represents the center position of the peak of the membership function. j This represents the width of the interval where the function value of the membership function is greater than zero;

[0069] S304, the membership function values ​​Q obtained in step S303 1,ik Entering the second rule layer, the strength calculation formula is used. Get the rule strength value Q for each row 2,i ;

[0070] S305, the rule strength value Q obtained in step S304 2,i Entering the third normalization layer, according to the formula Obtain the normalized strength rule value Q 3,i ;

[0071] S306. Extract all elements in the i-th row of the principal component matrix Y′ to form a completely linear language rule function, expressed as follows: Where p ij Let ε be the slope of the language rule function. i The error value between the estimated value and the actual value calculated for the language rule function is the parameter value obtained by performing multiple linear regression using the least squares method for multiple variables; the error value ε is stored and transmitted backward. i Upgrade to the first blurring layer, and simultaneously obtain the rule output value Q. 4,i =Q 3,i f i ;

[0072] S307, Rule Output Value Q 4,i Entering the fifth output layer, the results are summed and finally output as the fused decision result Q. 5,i The formula is And summarize them into a fusion decision result matrix Q = [Q 5,i ] 400×1 Q 5,i This represents the fusion decision result for the i-th location;

[0073] S308. Set the training iteration count (epochs) to 600, and repeat steps S301 to S307, using the error term ε of the language rule function. i =0, Q 5,i =1 is the training objective. The iteration ends after 600 epochs, completing the optimization of the parameter c. j d j The training is optimized to obtain a well-trained, improved ANFIS model.

[0074] S309. Input the principal component matrix Y obtained in step S2 into the improved ANFIS model trained in step S308, and repeat steps S303 to S307 to complete the fusion of multi-sensor information. Output the fusion decision result matrix in probabilistic form.

[0075] S4. Set the threshold for determining the occurrence of a fire to 0.10, and apply the fusion decision result matrix obtained in step S3. The decision is evaluated based on the fusion decision results: if all fusion decision results are less than or equal to 0.10, then proceed to step S1; if the fusion decision result of the i-th location is less than or equal to 0.10, then proceed to step S1. If the value exceeds 0.10, it is determined that a fire has occurred at that location, and the location information is sent to the flight control platform.

[0076] This embodiment uses an F450 UAV platform equipped with four sensors to collect environmental data such as temperature and humidity. The ANFIS model is trained and improved using data from a historical fire environment database, and the final fire determination is output in probabilistic form. The UAV location is adjusted or transmitted based on a threshold of 0.10. To facilitate presentation of the results of this embodiment, the obtained fusion decision result matrix is ​​shown. In the diagram, positions where the fusion decision results are equal and exceed a threshold are connected by a curve to obtain, as shown below. Figure 3 The implementation effect diagram is shown.

[0077] Example 2

[0078] This embodiment will adopt Figure 1 The flowchart and Figure 2 The structural diagram shown illustrates a UAV fire monitoring method based on multi-sensor information fusion technology. This method monitors an area of ​​3000 meters long × 3000 meters wide, identical to that in Example 1, achieving information fusion from various sensors to determine the fire situation. The specific process is as follows:

[0079] S1. Equip the F450 drone platform with 5 sensors (temperature, humidity, smoke, carbon dioxide, and infrared sensors), fly the drone platform to a fixed altitude above the area to be measured to collect data, with each data collection point spaced 200 meters apart, and collect environmental data at a total of 225 locations within the range, generating a total of 225 samples and an input matrix with 5 variable dimensions. Where x ij This represents the observation data of the i-th sample and the j-th variable dimension;

[0080] S2. Referring to step S2 in Example 1, process the input matrix obtained in step S1. Perform dimensionality reduction on the input matrix. Converted to a principal component matrix Y = [y] representing 225 samples and 3 feature variables. ik ] 225×3 , where y ik This represents the principal component data of the i-th sample and the k-th feature variable dimension;

[0081] S3. Input the principal component matrix Y into the improved ANFIS model to obtain the fused decision result matrix. in This represents the fusion decision result for the i-th location;

[0082] The improved ANFIS model consists of a first layer (fuzzification layer), a second layer (rule layer), a third layer (normalization layer), a fourth layer (rule output layer), and a fifth layer (output layer). The specific operation process is as follows:

[0083] S301. Using previously known fire environment data (temperature, humidity, smoke, carbon dioxide, and infrared information) obtained by drones, construct an input matrix of the same size as in step S1, containing 225 samples and 5 variable dimensions. x′ ij Represents the input matrix The i-th sample and the j-th dimension of the environment data are used as the training set and the test set, respectively.

[0084] S302. For the input matrix obtained in step S301 Repeat step S2 to construct the same 225 samples as in step S2, with the principal component matrix Y′=[y′] represented by the three feature variable dimensions. ik ] 225×3 y′ ik This represents the principal component data of the i-th sample and the k-th feature variable dimension of the principal component matrix Y′;

[0085] S303. Input the principal component matrix Y′ obtained in step S302 into the first fuzzification layer of the improved ANFIS model. This layer consists of adaptive nodes composed of membership functions. Obtain the membership function values ​​corresponding to each element of Y′, denoted as:

[0086] Q 1,ik =μ(y′) ik )k=1,2,3, i=1,2,…,225;

[0087] In this embodiment, the membership function is set to the following standard Gaussian function form:

[0088] Where e j f j The parameters to be optimized to improve the adaptive function of the ANFIS model; e j f represents the coordinates of the center of the Gaussian function peak. j This represents the standard deviation of the Gaussian function;

[0089] The subsequent steps are as described in steps S304 to S309 of Example 1. The ANFIS model is improved by training, and the principal component matrix Y obtained in step S2 is input. Finally, the fusion decision result matrix is ​​output in probabilistic form.

[0090] S4. Set the threshold for determining the occurrence of a fire to 0.15, and apply the fusion decision result matrix obtained in step S3. The decision is evaluated based on the fusion decision results: if all fusion decision results are less than or equal to 0.15, then proceed to step S1; if the fusion decision result of the i-th location is less than or equal to 0.15, then proceed to step S1. If the value exceeds 0.15, it is determined that a fire has occurred at that location, and the location information is sent to the flight control platform.

[0091] This embodiment uses an F450 UAV platform equipped with five sensors to collect environmental data such as temperature, humidity, and carbon dioxide. Data collected from a historical fire environment database is used to train an improved ANFIS model using a standard Gaussian function as the membership function. The model ultimately outputs the fire determination in probabilistic form, adjusting or transmitting the UAV's location based on a threshold of 0.15. To facilitate presentation of the results of this embodiment, the obtained fusion decision result matrix is ​​shown. In the diagram, positions where the fusion decision results are equal and exceed a threshold are connected by a curve to obtain, as shown below. Figure 4 The implementation effect diagram is shown.

[0092] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for monitoring fires using unmanned aerial vehicles (UAVs) based on multi-sensor information fusion technology, characterized in that, The drone fire monitoring method includes the following steps: S1. Equip a drone with n sensors. Each sensor collects environmental data from p locations, summarizes and transmits the data, and generates an input matrix with a total of p samples and n variable dimensions. Where x ij This represents the observation data of the i-th sample and the j-th variable dimension; S2. The input matrix obtained in step S1 Perform dimensionality reduction on the input matrix. Transform into a principal component matrix Y = [y] representing p samples and m feature variables. ik ] p×m , where y ik This represents the principal component data of the i-th sample and the k-th feature variable dimension; S3. Input the principal component matrix Y into the improved ANFIS model to obtain the fused decision result matrix. in This represents the fusion decision result for the i-th location; The improved ANFIS model consists of a first layer (fuzzification layer), a second layer (rule layer), a third layer (normalization layer), a fourth layer (rule output layer), and a fifth layer (output layer). The improved ANFIS model uses historical fire environment data and stops when the set maximum number of iterations (epochs) is reached. During the iteration process, data is forward-transmitted to the fourth layer of the improved ANFIS model, where the error rate of the multiple linear regression is calculated based on the language rule function, and then backward-transmitted to the first layer to optimize the membership function parameters of the first layer. S4. Set a threshold for determining the occurrence of a fire, and process the fusion decision result matrix obtained in step S3. The decision is evaluated based on all fusion decision results: if none of the fusion decision results exceed the threshold, then proceed to step S1; if none of the fusion decision results at location i exceed the threshold, then proceed to step S1. If the threshold is exceeded, it is determined that a fire has occurred at that location and the location information is output.

2. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 1, characterized in that, In step S1, n sensors are mounted on the drone. Each sensor collects environmental data, including temperature, humidity, and carbon dioxide concentration, at p locations. The collected environmental data is then aggregated and transmitted to generate an input matrix with p samples and n variable dimensions. In the form of x ij This represents the environmental data of the i-th sample and the j-th dimension.

3. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 2, characterized in that, The process of step S2 is as follows: S201, For the input matrix Perform standardization processing to obtain the standardization matrix. in, Let j be the standardized index variable of sample i; Then these are the arithmetic mean and standard deviation of the j-dimensional variable, i = 1, 2, ..., p, j = 1, 2, ..., n; S202, According to the formula Construct the correlation coefficient matrix R = [r ij ] n×n , where r ij Let be the correlation coefficient of the j-th variable dimension of sample i; S203. Calculate the characteristic equation of R, obtain the n eigenvalues ​​of R, and arrange them in descending order to form the eigenvalue matrix λ = [λ1, λ2, ..., λn]. n ], and the eigenvector matrix W = [w1, w2, ..., w] corresponding to each eigenvalue. n ], where w j =[w 1j ,w 2j ,…,w nj ] T ; S204. Select the first m eigenvalues of R, where m < n, such that the cumulative contribution rate of the corresponding principal components and obtain the eigenvector matrix corresponding to the eigenvalues where k = 1, 2, …, m; S205. Using the normalized matrix X obtained in step S201 and the eigenvector matrix W obtained in step S204... m Construct the principal component matrix Y = XW m =[y ik ] p×m , where y ik This represents the value of the k-th principal component of the i-th sample.

4. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 3, characterized in that, The process of step S3 is as follows: S301. Using environmental data from the historical fire environment database, construct an input matrix with p samples and n variable dimensions, the same size as in step S1. x′ ij Represents the input matrix The i-th sample, the j-th dimension of the environmental data; S302. For the input matrix obtained in step S301 Repeat step S2 to construct the same principal component matrix Y′=[y′] with p samples and m feature variables as in step S2. ik ] p×m y′ ik This represents the principal component data of the i-th sample and the k-th feature variable dimension of the principal component matrix Y′; S303. Input the principal component matrix Y′ obtained in step S302 into the first fuzzification layer of the improved ANFIS model. This layer consists of adaptive nodes composed of membership functions. Obtain the membership function values ​​corresponding to each element of Y′, denoted as: Q 1,ik =μ(y′ ik ),k=1,2,…,m,i=1,2,…,p; S304, the membership function values ​​Q obtained in step S303 1,ik Entering the second rule layer, the strength calculation formula is used. Get the rule strength value Q for each row 2,i ; S305, the rule strength value Q obtained in step S304 2,i Entering the third normalization layer, according to the formula Obtain the normalized strength rule value Q 3,i ; S306. Extract all elements in the i-th row of the principal component matrix Y′ to form a completely linear language rule function, expressed as follows: Where p ij Let ε be the slope of the language rule function. i The error value between the estimated value and the actual value calculated for the language rule function is the parameter value obtained by performing multiple linear regression using the least squares method for multiple variables; the error value ε is stored and transmitted backward. i Upgrade to the first blurring layer, and simultaneously obtain the rule output value Q. 4,i =Q 3,i f i ; S307, Rule Output Value Q 4,i Entering the fifth output layer, the results are summed and finally output as the fused decision result Q. 5,i The formula is And summarized into a fusion decision result matrix Q = [Q 5,i ] p×1 Q 5,i This represents the fusion decision result for the i-th location; S308. Set the training iteration count (epochs) to 600, and repeat steps S301 to S307, using the error term ε of the language rule function. i =0, Q 5,i =1 is the training objective. The iteration ends after reaching the maximum value of epochs, thus completing the optimization of the parameter c. j d j The training is optimized to obtain a well-trained, improved ANFIS model. S309. Input the principal component matrix Y obtained in step S2 into the improved ANFIS model trained in step S308, and repeat steps S303 to S307 to complete the fusion of multi-sensor information. Output the fusion decision result matrix in probabilistic form.

5. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 4, characterized in that, The process of step S4 is as follows: Set a threshold for whether a fire occurs, and then determine the fusion decision result matrix. Check if all fusion decision results exceed the threshold; if none exceed the threshold, it is determined that no fire has occurred, the drone changes location, and returns to step S1; if the fusion decision result of the i-th location... If the threshold is exceeded, a fire is determined to have occurred at that location, and the specific location information is output.

6. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 4, characterized in that, The membership function μ(y′) ik The settings are as follows: Where c j d j The parameters to be optimized are those for improving the adaptive function of the ANFIS model; c j d represents the center position of the peak of the membership function. j This represents the width of the interval where the function value of the membership function is greater than zero.

7. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 4, characterized in that, The membership function μ(y′) ik It is set to the following standard Gaussian function form: Where e j f j The parameters to be optimized to improve the adaptive function of the ANFIS model; e j f represents the coordinates of the center of the Gaussian function peak. j This represents the standard deviation of the Gaussian function.

8. The UAV fire monitoring method based on multi-sensor information fusion technology according to claim 5, characterized in that, The threshold for whether a fire occurs is set to [0.05, 0.15].

Images