A forest fire danger early warning system based on internet of things communication

Abstract

This invention discloses a forest fire risk early warning system based on Internet of Things (IoT) communication, belonging to the field of data processing technology. The system comprises a data acquisition module that collects forest area data from edge devices; an environmental feature module that uses an adaptive forgetting factor recursive method to obtain environmental anomaly response vectors; a feature fusion module that extracts image feature tensors, maps these vectors to scaling matrices, combines them with a forced masking weighted feature tensor, and outputs predicted bounding box parameters; a feature inversion module that obtains a penalty matrix based on the parameters, constructs a partial differential equation for canopy drag force, and solves for the fire plume feature matrix and flow field uncertainty tensor matrix by minimizing the inverse source-finding objective function; and a communication optimization module that extracts elements as fire risk mapping weights, arranges them in descending order, and allocates time resources so that the terminal only sends fire plume elements to the decision center. This invention effectively isolates meteorological interference, improves the accuracy of locating small fire points under dense smoke, and solves the communication congestion problem during extreme fire events.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a forest fire risk early warning system based on Internet of Things (IoT) communication. Background Technology

[0002] Forest fires are extremely destructive natural disasters that seriously threaten the ecological environment and human life and property. With the large-scale popularization and development of Internet of Things (IoT) technology, forest fire risk early warning systems based on IoT communication have gradually become the mainstream means of forest fire prevention. These systems typically collect environmental data such as temperature, humidity, wind speed, and images through sensing devices deployed in forest areas, and transmit this data to a decision-making center via a communication network for fire identification and risk assessment. Currently, Chinese invention patent application number 202411595630.8 discloses a real-time monitoring and early warning system for forest fires based on IoT. In terms of data transmission, this solution determines the selected transmission link and backup transmission link by calculating the success probability of different transmission paths to address the link jitter problem. However, the aforementioned existing technologies still have the following problems in practical applications: In real forest environments, wind resistance from tree canopies and natural gusts exhibit strong randomness and dynamic changes. Existing technologies lack a deep separation mechanism for the steady state and abnormal fluctuations of the natural environment, making it impossible to accurately distinguish between normal meteorological disturbances and real fire plume thermal convection characteristics. This can easily lead to misjudging turbulent winds as fire turbulence, resulting in false alarms or missed alarms in early fire identification. Furthermore, existing technologies have failed to establish a deep linkage between environmental changes and image features. In real fire scenes accompanied by dense smoke, visible light images are often filled with noise, and existing technologies lack adaptive suppression of visible light noise. The dynamic adjustment capability of concealed heat sources is highlighted, but the accuracy of fire point location in complex scenarios is reduced. In terms of communication transmission mechanism, the existing solution mainly focuses on path optimization under normal conditions. However, when extreme fire risks break out, the core area of ​​the fire site is covered with dense smoke. If the sensing terminal continues to send a large amount of image matrix data to the monitoring center according to the normal logic, not only will these images that have been completely obscured by smoke become invalid noise, but the concurrent transmission of massive invalid data will instantly crowd out the extremely limited narrowband channel bandwidth, which is very likely to cause data packet collisions and broadcast storms. As a result, the most core and critical fire characteristic indicators cannot be reported in time, delaying the emergency response speed of the decision-making center. Summary of the Invention

[0003] The technical problem solved by this invention is that existing technologies lack a deep separation mechanism for the steady state and abnormal fluctuations of the natural environment, making it impossible to accurately distinguish between normal meteorological disturbances and real fire plume thermal convection characteristics. This can easily lead to misjudging turbulent winds as fire turbulence, resulting in false alarms or missed alarms in early fire identification. Existing technologies have failed to establish a deep linkage between environmental changes and image features. In real fire scenes accompanied by a large amount of dense smoke, visible light images are often filled with noise. Existing technologies lack the ability to dynamically adjust to adaptively suppress visible light noise and highlight hidden heat sources, reducing the accuracy of fire point location in complex scenarios.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a forest fire risk early warning system based on Internet of Things communication, comprising a data acquisition module, an environmental feature module, a feature fusion module, a feature inversion module, and a communication optimization module; The acquisition module is used to acquire forest area data from the edge sensing device; The environmental feature module is used to obtain the environmental change response vector based on the environmental temperature in the forest area data, combined with a recursive method with an adaptive forgetting factor. The feature fusion module is used to extract feature tensors from images in forest area data, map environmental change response vectors into scaling matrices, and perform weighted processing on feature tensors based on scaling matrices and forced masking, and output predicted bounding box parameters through a pre-trained detection head. The feature inversion module is used to obtain the penalty matrix based on the predicted bounding box parameters, construct the partial differential equation of canopy drag force and the expected observation state matrix, and iteratively solve the partial differential equation of canopy drag force by minimizing the inverse source-finding objective function including the penalty matrix to obtain the fire plume feature matrix and the flow field uncertainty tensor matrix. The communication optimization module is used to extract the element values ​​corresponding to the edge sensing terminal from the flow field uncertainty tensor matrix as fire risk mapping weights, sort them in descending order and allocate time resources, so that the edge sensing terminal can extract the element at the corresponding position from the fire plume feature matrix and transmit it to the decision center.

[0005] Preferably, the forest area data includes ambient temperature, image data, wind speed, weather forecast data, and installation and configuration data of edge sensing terminals; The meteorological forecast data includes macro wind speed values ​​and macro wind direction angles; The installation configuration data includes the camera intrinsic parameter matrix, physical installation height, and pitch angle of the edge sensing terminal; The image data includes visible light images and infrared images; The visible light image is a visible light image matrix including a red channel, a green channel, and a blue channel; The infrared image is an infrared image matrix comprising a single thermal radiation channel.

[0006] Preferably, the process of obtaining the environmental change response vector specifically includes: Construct an autoregressive time series equation, the mathematical expression of which is: ; in, The actual temperature at the current time step. Represents a matrix of historical observation data; The parameter vector represents the parameter to be identified; This represents the inherent white noise interference in the environment; A transpose matrix representing the historical observation data matrix; Based on the parameter vector identified at the previous moment, calculate the theoretical predicted temperature at the current moment, subtract the collected ambient temperature from the theoretical predicted temperature to obtain the predicted temperature residual, and calculate the mean square value of the predicted temperature residual within the sliding time window. Substitute the mean square value of the predicted temperature residual into the exponential forgetting factor update function to calculate the adaptive forgetting factor; The parameter vector is updated using a recursive least squares algorithm with an adaptive forgetting factor, and the constant term elements in the parameter vector are extracted as steady-state environment values. The predicted temperature residuals within the sliding time window are stacked in chronological order to form an environmental anomaly response vector.

[0007] Preferably, the feature fusion module includes a dual-modal feature extraction unit, a priori mapping unit, a channel recalibration unit, and a micro-fire detection head; The dual-modal feature extraction unit includes two parallel feature extraction branches, each consisting of multiple two-dimensional convolutional layers, used to map the visible light image matrix into a visible light feature tensor and the infrared image matrix into an infrared feature tensor. The prior mapping unit includes a two-layer fully connected network. The first fully connected layer maps the environmental change response vector to a hidden vector, and the second fully connected layer maps the hidden vector to a scaling matrix. The scaling matrix includes visible light initial attention weights and infrared initial attention weights.

[0008] Preferably, the channel recalibration unit is used to generate a forced masking matrix based on the signal-to-noise ratio and temperature rise gradient, multiply the scaling matrix and the forced masking matrix element by element to obtain an updated scaling matrix, and use the updated scaling matrix to perform weighted processing on the visible light feature tensor and the infrared feature tensor respectively to obtain the recalibrated visible light feature tensor and the recalibrated infrared feature tensor. The signal-to-noise ratio is the global variance of the grayscale values ​​of all pixels in the visible light image matrix; The temperature gradient is the sum of the absolute values ​​of the elements in the environmental change response vector; The process of generating the forced masking matrix specifically includes: If the signal-to-noise ratio is less than the preset occlusion threshold and the temperature rise gradient is greater than the preset gradient threshold, then the forced occlusion matrix includes a preset visible light penalty coefficient and a preset infrared gain coefficient. If the signal-to-noise ratio is greater than or equal to the occlusion threshold, or the temperature rise gradient is less than or equal to the gradient threshold, then the masking matrix is ​​forced to include the preset weight preservation coefficients. The micro-fire detection head is used to stitch together the recalibrated visible light feature tensor and the recalibrated infrared feature tensor in the channel dimension, and output the predicted bounding box through convolution classification and position regression.

[0009] Preferably, the feature fusion module further includes an adaptive loss calculation unit, which is used to calculate the localization loss during the offline training phase; The process of calculating the localization loss during the offline training phase specifically includes: The true bounding box parameters of the labeled sample and the predicted bounding box parameters output by the microfire detection head are obtained. The true bounding box parameters include the true center point x-coordinate, the true center point y-coordinate, the true bounding box width, and the true bounding box height. The predicted bounding box parameters include the predicted center point x-coordinate, the predicted center point y-coordinate, the predicted width, and the predicted height. The first deviation index is the deviation of the center point Euclidean distance, and the second deviation index is the deviation of the aspect ratio difference. Extract the aspect ratio of the actual border, and set the first regression weight and the second regression weight based on the aspect ratio; Substituting the first regression weight, the second regression weight, the first deviation index, and the second deviation index into the total positioning loss function, the mathematical expression of the total positioning loss function is as follows: ; in, For the total positioning loss, As the first regression weight, As the first deviation index, As the second regression weight, This is the second deviation index.

[0010] Preferably, the process of obtaining the penalty matrix specifically includes: Based on the predicted bounding box parameters and camera intrinsic parameter matrix, physical installation height and pitch angle output by each edge sensing terminal; By using perspective projection inverse transformation, the coordinates of the four pixel vertices of the predicted bounding box parameters are projected onto a local physical ground coordinate system with the vertical projection point of the corresponding edge sensing terminal as the origin, and the first coordinate pair corresponding to the pixel vertex is obtained. Obtain the actual installation coordinate pairs and installation orientation angles of each edge sensing terminal in the local spatial grid, perform rotation and translation operations on the first coordinate pair using a two-dimensional affine transformation matrix to obtain the second coordinate pair, and connect the second coordinate pairs sequentially to generate a global projection polygon of the grid system. Traverse the grid coordinates of the nodes of the local spatial grid. For any node's grid coordinates, emit a ray in the positive direction of the horizontal axis of the node's grid coordinates, and count the number of times the ray intersects with the boundaries of all global projection polygons of the grid system. Construct a penalty matrix. If the number of intersections is odd, determine that the node falls inside the projected polygon, and overwrite the elements in the penalty matrix corresponding to the node with minimal relaxation weights. If the number of intersections is not odd, it is determined that the node does not fall inside the projected polygon, and the element at the position corresponding to the node in the penalty matrix is ​​overwritten with a maximal relaxation weight.

[0011] Preferably, the mathematical expression of the partial differential equation for the canopy drag force is: ; in, These are divergence and gradient differential operators used to calculate the rate of change of numerical values ​​between adjacent nodes within a spatial grid. Let be a vector consisting of the spatial coordinates of nodes on a two-dimensional local spatial grid. Let the first random variable be a two-dimensional random vector composed of the second random variable. This is a preset canopy drag coefficient containing random disturbances. For space state function, As a background of normal wind conditions, Let be the characteristic distribution function of the fire plume to be determined; The spatial gradient vector of the spatial state function; The state function, background normal wind field and fire plume characteristic distribution function are numerically discretized on the two-dimensional local space grid and then mapped and stored as state matrix, background normal wind field matrix and fire plume characteristic matrix, respectively. The mathematical expression for the canopy drag coefficient is: ; in, The basic drag force constant representing the forest area under windless conditions; The coordinates of the nodes on the local spatial grid ( , The vector formed by ) For the first random variable With the second random variable The vector formed This represents the total physical length of the grid horizontally. This represents the total physical width of the grid vertically; The process of obtaining the desired observation state matrix specifically includes: A blank matrix is ​​constructed as the initial expected observation state matrix. For any target element in the expected observation state matrix, the Euclidean distance between the node grid coordinates corresponding to the target element and the grid coordinates of the edge sensing terminal is calculated. Based on the wind speed of the edge sensing terminal, the wind speed assignment of the target element is calculated using the inverse distance weight interpolation algorithm. The expected observation state matrix is ​​obtained by traversing and calculating the wind speed assignment.

[0012] Preferably, the process of obtaining the fire plume feature matrix and the flow field uncertainty tensor matrix specifically includes: Construct a reverse source tracing objective function, the mathematical expression of which is: ; in, For reverse source tracing objective function, For the mathematical expectation of the error, For the penalty matrix, This represents the control variable matrix. The new matrix is ​​obtained by squaring each of the one hundred elements in the matrix. Represents the Hadamard product operation; For the mathematical variance of the error, The first preset weight; For the two random variables in the partial differential equation of canopy drag force, the Smoljak sparse grid algorithm is introduced to determine the coordinates of multiple configuration points and their corresponding weight coefficients. Initialize the fire plume feature matrix, and substitute the fire plume feature matrix and each configuration point into the partial differential equation of canopy drag force to obtain multiple theoretical state matrices. The sum of squared differences between the corresponding elements of each theoretical state matrix and the expected observed state matrix is ​​calculated to obtain multiple total error values; Multiply each total error value by its corresponding weighting coefficient, and sum all the products to obtain the expected value of the error. The square of the difference between each total error value and the expected value of the error is taken as the deviation. Each deviation is multiplied by the corresponding weighting coefficient, and all the product results are added together to obtain the error variance. By combining the gradient descent optimization algorithm, the element values ​​in the fire plume feature matrix are iteratively adjusted until the inverse source-finding objective function converges to a minimum value, and the converged fire plume feature matrix is ​​obtained. Based on the converged fire plume feature matrix and each configuration point, multiple theoretical state matrices are obtained; Extract multiple element values ​​corresponding to the same node from each theoretical state matrix, and calculate the variance of the multiple element values; Extract the variance of all nodes and construct the flow field uncertainty tensor matrix according to the node position order.

[0013] Preferably, the operation process of the communication optimization module specifically includes: Collect the number of failed transmissions that did not receive an acknowledgment frame and the total number of transmission attempts within the current communication cycle, and divide the number of failed transmissions by the total number of transmission attempts to obtain the data packet collision rate; If the packet collision rate is greater than the preset congestion threshold, a broadcast storm is determined to have occurred, and the directed graph routing protocol based on the shortest geographical path is shut down. Calculate the Euclidean distance between nodes in the local spatial grid and the edge sensing terminal, and bind the edge sensing terminal to the node with the closest Euclidean distance; Extract the element values ​​corresponding to the nodes from the flow field uncertainty tensor matrix and use them as the fire risk mapping weights for the edge sensing terminal. If there is a fire risk mapping weight greater than the preset extreme value threshold, it is determined that there is an extreme fire risk, a local communication mask instruction is generated and broadcast to the edge sensing terminal, and the edge sensing terminal clears the image data queued in the local transmission queue. The fire hazard mapping weights of the edge sensing terminals are sorted in descending order of their numerical values. The time division multiple access mechanism is then activated, and time resources are allocated in descending order. The beneficial effects of this invention are as follows: This invention uses the environmental change response vector as prior knowledge to intervene in the channel recalibration process at the bottom layer of image processing. It can adaptively generate a forced mask based on the temperature rise gradient, actively suppress the visible light feature weights containing a large amount of invalid noise, and at the same time amplify the infrared thermal radiation feature weights that hide heat sources by several times. This invention constructs a partial differential equation of canopy drag force containing spatially distributed random perturbations, and introduces a sparse grid algorithm and an inverse source-finding objective function for dimensionality reduction and minimization. It accurately projects the two-dimensional pixel-level fire point prediction bounding box and transforms it into a spatial penalty matrix within the physical space grid, filtering out the random flow field perturbations caused by uneven tree distribution density and natural gusts, and inversely solves for the pure fire plume heat convection intensity and flow field uncertainty tension. This invention constructs a partial differential equation for canopy drag force containing spatially distributed random perturbations, and introduces the Smojack sparse grid algorithm and inverse source-finding objective function for dimensionality reduction and minimization. This inversion mechanism accurately projects the two-dimensional pixel-level fire point prediction bounding box and transforms it into a spatial penalty matrix within the physical space grid. Then, through rigorous mathematical and physical equation derivation, it filters out the random flow field perturbations caused by uneven tree distribution density and natural gusts, and solves inversely the pure fire plume thermal convection intensity and flow field uncertainty tensor. This not only realizes high-precision physical source tracing of small fire sources in the macro-geographic coordinate system, but also provides objective quantitative indicators for fire risk assessment that are free from meteorological artifacts, and solves the coordinate drift problem of traditional positioning algorithms under strong wind conditions. Attached Figure Description

[0014] Figure 1 This is a basic flowchart of a forest fire risk early warning system based on Internet of Things (IoT) communication, provided as an embodiment of the present invention. Detailed Implementation

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Example, refer to Figure 1 This paper provides a forest fire risk early warning system based on Internet of Things (IoT) communication, including a data acquisition module, an environmental feature module, a feature fusion module, a feature inversion module, and a communication optimization module. The data acquisition module is used to collect forest area data from edge sensing devices; The environmental feature module is used to obtain the environmental change response vector based on the ambient temperature in the forest area data, combined with a recursive method with an adaptive forgetting factor. The feature fusion module is used to extract the feature tensors of images in the forest area data, and map the environmental change response vector into a scaling matrix; the feature tensors are weighted based on the scaling matrix and the forced masking, and the predicted bounding box parameters are output through the pre-trained detection head. The feature inversion module is used to obtain the penalty matrix based on the predicted bounding box parameters, construct the partial differential equation of canopy drag force and the expected observation state matrix, and iteratively solve the partial differential equation of canopy drag force by minimizing the inverse source-finding objective function including the penalty matrix to obtain the fire plume feature matrix and the flow field uncertainty tensor matrix. The communication optimization module is used to extract the element values ​​corresponding to the edge sensing terminals from the flow field uncertainty tensor matrix as fire risk mapping weights, sort them in descending order and allocate time resources, so that the edge sensing terminals can extract the elements at the corresponding positions from the fire plume feature matrix and transmit them to the decision center.

[0017] This invention accurately captures environmental anomaly responses by introducing a time-series algorithm with an adaptive forgetting factor, and uses this as a cross-modal prior condition to dynamically intervene in the feature recalibration process of bispectral images. This overcomes the interference of smoke obscuring and meteorological artifacts on single visual recognition in complex forest areas. Furthermore, it combines a rigorous partial differential equation of canopy drag force with an inverse source-finding objective function for dimensionality reduction and inversion, removing flow field noise caused by natural random gusts. Based on the physical fire risk weight obtained from the inversion, it dynamically schedules underlying communication resources, effectively resolving the problems of IoT channel congestion and broadcast storms caused by the concurrent generation of massive invalid images during extreme fire outbreaks. This enables very early and accurate detection of small fire sources and reliable reporting with extremely low latency.

[0018] The forest area data includes ambient temperature, image data, wind speed, weather forecast data, and installation and configuration data of edge sensing terminals; Weather forecast data includes macro wind speed values ​​and macro wind direction angles; Installation configuration data includes the camera intrinsic parameter matrix, physical installation height, and pitch angle of the edge sensing terminal; Image data includes visible light images and infrared images; The visible light image is a visible light image matrix that includes red, green, and blue channels; The infrared image is an infrared image matrix containing a single thermal radiation channel.

[0019] In one specific embodiment of the present invention, multiple edge computing gateways are deployed in the original forest area. Each edge computing gateway serves as the computing core and topological origin of a local area. Each gateway defines and constructs a 100m × 100m two-dimensional local spatial grid in physical space. The entire forest area is covered by splicing together these multiple independently operating but collaborative two-dimensional local spatial grids. Within the coverage area of ​​each two-dimensional local spatial grid, several edge sensing terminals are distributed and deployed. Unlike traditional drones that fly above the tree canopy for inspection, the edge sensing terminals are statically fixed and installed on monitoring poles below the canopy layer. The edge sensing terminals integrate temperature sensors, three-dimensional ultrasonic anemometers, and lightweight dual-spectrum cameras.

[0020] The process of obtaining the environmental change response vector specifically includes: Construct an autoregressive time series equation. The mathematical expression of the autoregressive time series equation is as follows: ; in, The actual temperature at the current time step. This represents a historical observation data matrix, with its internal elements arranged as follows: , representing the actual ambient temperature at the previous moment, the two moments before, and the three moments before, respectively, and the fixed value 1 used to match the constant term; This represents the parameter vector to be identified, with its internal elements arranged as follows: The first three elements represent the weighting coefficients of historical data on the current result, and the fourth element... Physically, it represents the DC bias of the environment, which is the steady-state environmental value after removing gusts and short-term fluctuations. This represents the inherent white noise interference in the environment; Represents the transpose of a matrix; Based on the parameter vector identified at the previous moment, calculate the theoretical predicted temperature at the current moment, subtract the collected ambient temperature from the theoretical predicted temperature to obtain the predicted temperature residual, and calculate the mean square value of the predicted temperature residual within the sliding time window. Substitute the mean square value of the predicted temperature residual into the exponential forgetting factor update function to calculate the adaptive forgetting factor; The parameter vector is updated using a recursive least squares algorithm with an adaptive forgetting factor, and the constant term elements in the parameter vector are extracted as steady-state environment values. The predicted temperature residuals within the sliding time window are stacked in chronological order to form an environmental anomaly response vector.

[0021] In one specific embodiment of the present invention, For the current edge sensing terminal, the ambient temperature is collected at a fixed sampling frequency with a preset sampling period of two seconds; In order to isolate abnormal changes in the environment, it is necessary to predict the normal temperature trend at the current moment; A third-order autoregressive time series equation was constructed. The physical meaning of this equation is to derive the theoretical temperature of the current time step using the historical real temperature data of the past three time steps. The mathematical expression for the theoretically predicted temperature is: ; in, The theoretically predicted temperature for the current moment. For historical observation data matrix, This is the parameter vector identified at the previous time step; To prevent misjudgment caused by single sensor noise, the mean square value of the predicted residual within the current sliding time window is calculated in real time. The length of the sliding time window is preset to 10, which corresponds to a time span of 20 seconds, and is used to cover and smooth out one forest gust cycle. The calculation process of the mean square value of the predicted residual is to obtain the arithmetic mean of the squares of the residuals at the current time and the past nine times. After obtaining the mean square value, a nonlinear factor update operation is performed to adjust the algorithm's trust in historical data. Specifically, the mean square value is substituted into the exponential forgetting factor update function, which is mathematically expressed as follows: ; in, This represents the adaptive forgetting factor calculated at the current time step; This represents the set lower threshold of the forgetting factor, which is set to 0.95 in this embodiment; This represents the upper limit threshold of the forgetting factor, with a value of 0.99; This represents sensitivity, with a value of 2.0. To predict the mean square value of the residuals; Utilizing an adaptive forgetting factor The recursive least squares algorithm is used for iterative identification of environmental parameters. Specifically, by calculating the gain matrix and covariance matrix and combining them with the current residual, the parameter vector at the current time step is updated. ; The ten predicted temperature residuals within the sliding time window are stacked in chronological order to form a column vector of ten rows and one column, which is the environmental change response vector.

[0022] This invention constructs a recursive least squares algorithm based on a third-order autoregressive time series equation and an adaptive forgetting factor. It can dynamically adjust the algorithm's trust metric for historical environmental temperature data according to the mean square value of the predicted residuals within the sliding time window, thereby keenly and accurately capturing the abnormal and sudden trends of the physical environment. By extracting the steady-state environmental bias and residual sequence to form an environmental change response vector, it removes the inherent white noise of the sensor and the short-term temperature fluctuations caused by normal gusts in forest areas, providing a clean and highly sensitive numerical dimension for subsequent cross-modal image processing as an early fire prior indicator.

[0023] The feature fusion module includes a dual-modal feature extraction unit, a priori mapping unit, a channel recalibration unit, and a microfire detection head; The dual-modal feature extraction unit includes two parallel feature extraction branches, each consisting of multiple two-dimensional convolutional layers, used to map the visible light image matrix into a visible light feature tensor and the infrared image matrix into an infrared feature tensor. The prior mapping unit consists of two fully connected layers. The first fully connected layer maps the environmental change response vector to a hidden vector, and the second fully connected layer maps the hidden vector to a scaling matrix, which contains the initial attention weights for visible light and infrared light.

[0024] In one specific embodiment of the present invention, the visible light image matrix and the infrared image matrix are input to a dual-modal feature extraction unit. The dual-modal feature extraction unit includes two parallel independent feature extraction branches, each consisting of five standard two-dimensional convolutional layers with a stride of 2. After sliding multiply-accumulate and downsampling calculations of the convolutional kernels, the visible light image matrix is ​​mapped to a dimension of... The visible light feature tensor and the infrared image matrix are mapped to a dimension of The infrared feature tensor, with 256 channels, represents the number of different texture and edge features extracted by the network in high-dimensional space; To enable environmental temperature data to influence image features, a dimensionality transformation of the environmental change response vector is required. This vector is then input into a priori mapping unit, and the first fully connected layer multiplies the environmental change response vector by... The weight matrix is ​​then added with a bias, mapping it to a 16-dimensional hidden vector. The second fully connected layer multiplies this hidden vector by... The weight matrix is ​​processed by the Sigmoid activation function, resulting in a 2×1 scaling matrix. ; and These represent the initial attention weights for visible light and infrared light, respectively, calculated by the network in the preliminary calculations. Their values ​​are both between [0,1].

[0025] This invention extracts high-dimensional features of visible light and infrared light through parallel dual-modal convolutional branches, and then uses a two-layer fully connected network to reduce the dimensionality of the independently calculated environmental change response vector and map it into a pair of scaling matrices containing initial attention weights. This establishes a mathematical bridge between purely numerical meteorological time-series features and two-dimensional visual spatial features, enabling small temperature rise changes in the physical environment to be directly transformed into attention guidance instructions for visible light and infrared image channels at the bottom layer of the neural network. This significantly enhances the network's sensitivity and extraction accuracy for latent fire source features.

[0026] The channel recalibration unit is used to generate a forced masking matrix based on the signal-to-noise ratio and temperature rise gradient. The scaling matrix is ​​multiplied element-wise with the forced masking matrix to obtain the updated scaling matrix. The updated scaling matrix is ​​then used to weight the visible light feature tensor and the infrared feature tensor to obtain the recalibrated visible light feature tensor and the recalibrated infrared feature tensor. The signal-to-noise ratio is the global variance of the grayscale values ​​of all pixels in the visible light image matrix; The temperature gradient is the sum of the absolute values ​​of the elements in the environmental change response vector; The process of generating the forced masking matrix specifically includes: If the signal-to-noise ratio is less than the preset occlusion threshold and the temperature rise gradient is greater than the preset gradient threshold, then the forced occlusion matrix includes a preset visible light penalty coefficient and a preset infrared gain coefficient. If the signal-to-noise ratio is greater than or equal to the occlusion threshold, or the temperature rise gradient is less than or equal to the gradient threshold, then the masking matrix is ​​forced to include the preset weight preservation coefficients. The micro-fire detection head is used to concatenate recalibrated visible light feature tensors and recalibrated infrared feature tensors in the channel dimension, and output predicted bounding boxes through convolution classification and position regression.

[0027] In one specific embodiment of the present invention, if the signal-to-noise ratio is less than the occlusion threshold and the temperature rise gradient is greater than the gradient threshold, it indicates that a severe temperature rise accompanied by dense smoke has occurred at the physical site. At this time, a two-row, one-column forced occlusion matrix is ​​generated. The first element of the forced occlusion matrix is ​​a preset visible light penalty coefficient, specifically set to 0.1; the second element is a preset infrared gain coefficient, specifically set to 5.0; the forced occlusion matrix is... ; If the signal-to-noise ratio is greater than or equal to the occlusion threshold, or the temperature rise gradient is less than or equal to the gradient threshold, then both elements of the generated forced masking matrix are preset weight preservation coefficients, specifically 1.0. The forced masking matrix is ​​as follows: ; In this embodiment, the occlusion threshold is set to 50.0, and the gradient threshold is set to 15.0; The scaling matrix is ​​multiplied element-wise with the forced masking matrix to obtain the updated scaling matrix. The reason for this operation is to forcibly reduce the weight of the visible light channel to one-tenth of the original weight to suppress noise, and to forcibly amplify the weight of the infrared channel by five times to highlight the hidden heat source. Use the updated weights Multiply by each element of the visible light feature tensor to obtain the recalibrated visible light feature tensor, and use... Multiply by each element of the infrared feature tensor to obtain the recalibrated infrared feature tensor. Then, concatenate and superimpose the recalibrated visible light feature tensor and the recalibrated infrared feature tensor along the channel dimension. Feed them into the micro fire detection head for convolution classification and position regression to output the final predicted bounding box.

[0028] To address the challenge of dense smoke obstruction in extreme fire scenarios, this solution proposes a forced masking intervention mechanism based on dynamic evaluation of signal-to-noise ratio and temperature gradient. When it is determined that the global variance of the visible light image at the scene drops sharply due to smoke, accompanied by a dramatic increase in local temperature, the channel recalibration unit actively generates a forced masking matrix containing a large penalty and an extremely high gain coefficient. This suppresses the visible light feature weights filled with invalid noise, while simultaneously amplifying the infrared thermal radiation feature weights that can penetrate the smoke by several times. This dynamic and forced channel-level feature fusion strategy eliminates the overfitting phenomenon of the detection network to dense smoke artifacts and ensures a high detection rate for small fire points in severe vision obstruction scenarios.

[0029] The feature fusion module also includes an adaptive loss calculation unit, which is used to calculate the localization loss during the offline training phase. The process of calculating the localization loss during the offline training phase specifically includes: Obtain the true bounding box parameters of the labeled samples and the predicted bounding box parameters output by the microfire detection head. The true bounding box parameters include the true center point x-coordinate, the true center point y-coordinate, the true bounding box width, and the true bounding box height. The predicted bounding box parameters include the predicted center point x-coordinate, the predicted center point y-coordinate, the predicted width, and the predicted height. The first deviation index is the Euclidean distance deviation from the center point. The mathematical expression for the Euclidean distance deviation from the center point is: ; in, As the first deviation index, and These are the x-coordinates of the center points of the ground truth bounding box and the predicted bounding box, respectively. and These are the ordinates of the center points of the ground truth bounding box and the predicted bounding box, respectively. The first deviation metric is used to measure the absolute deviation of the predicted position, which is the diagonal length of the smallest bounding rectangle that can simultaneously enclose the actual bounding rectangle and the predicted bounding rectangle. The second deviation index is the aspect ratio difference deviation, and its mathematical expression is: ; in, The second deviation index, and These are the width and height of the actual border, respectively; and These are the predicted width and predicted height of the predicted bounding box output by the micro-fire detection head, respectively. Extract the aspect ratio of the actual border, and set the first regression weight and the second regression weight based on the aspect ratio; Substituting the first regression weight, the second regression weight, the first deviation index, and the second deviation index into the total localization loss function, the mathematical expression of the total localization loss function is: ; in, For the total positioning loss, As the first regression weight, As the first deviation index, As the second regression weight, This is the second deviation index.

[0030] In a specific embodiment of the present invention, before the feature fusion module is put into operation, offline supervised learning training needs to be performed. Specifically, a training set containing several discrete samples is pre-constructed, wherein each labeled sample is strictly paired with multi-source input data and real label data. The multi-source input data includes visible light image matrices, infrared image matrices, and environmental change response vectors acquired synchronously at the same historical time step; the real label data includes manually verified real target category vectors, and the real bounding box parameters provided by the labeled samples include four values: the x-coordinate of the center point. , center point ordinate Border width Border height ; The predicted bounding box parameters output by the microfire detection head also include four values: the x-coordinate of the predicted center point. Predict the ordinate of the center point Predicted width Predicted height ; During the offline training phase of the multidimensional feature penetration perception module model, it is necessary to optimize the network parameters by calculating the error through the loss function, namely the weights of the fully connected layers and the convolutional kernel elements of each convolutional layer. At this time, the training data consists of known labeled data and network prediction data. The left boundary of the smallest bounding rectangle is the left boundary of the actual border. ) and the left boundary of the predicted bounding box ( The minimum value of ), the right boundary is the right boundary of the actual border ( ) and the right boundary of the predicted bounding box ( The maximum value of ), with the upper boundary being the upper boundary of the actual border ( ) and the upper boundary of the predicted bounding box ( The minimum value of ), the lower boundary is the lower boundary of the actual border () ) and the lower boundary of the predicted bounding box ( The maximum value of ) is obtained by finding the diagonal length of the minimum bounding rectangle based on its left, right, top, and bottom boundaries. The second deviation index uses the inverse trigonometric function to calculate the angle difference, which is used to measure the geometric similarity between the predicted box and the ground box in non-rigid shapes. In real forest fire scenarios, if strong winds occur, the flames will be severely elongated and torn apart by the wind, and the aspect ratio of the actual bounding box will become extremely narrow. At this time, forcing the predicted box to be exactly the same as the actual box shape will cause the network to overfit the noise boundary, thus ignoring the tracking of the core location of the fire. Therefore, a dynamic weight adjustment operation is performed, specifically, the aspect ratio of the actual border is extracted. ; If the aspect ratio of the true border is greater than 3.0, or less than 0.33, then the first regression weight of the Euclidean distance of the center point will be adjusted. Set to 1.5; adjust the second regression weight of the aspect ratio difference. Set to 0.5; If the aspect ratio of the actual border is greater than or equal to 0.33 and less than or equal to 3.0, then the first regression weight of the Euclidean distance of the center point will be adjusted. The second regression weights for the aspect ratio difference were all set to 1.0; The intersection-union ratio of the ground truth bounding box and the predicted bounding box is calculated by dividing the overlapping area of ​​the two rectangles by the total area occupied. The area represents the pixel area occupied on the image. The overlapping area refers to the total number of pixels in the intersection and overlap of the predicted and ground truth bounding boxes. The union area is the sum of the predicted bounding box area and the ground truth bounding box area, minus the overlapping area. The backpropagation algorithm is used to perform gradient descent and numerical updates on all unknown weight matrix elements in the feature extraction unit and the prior mapping unit until the loss converges and training is completed. The micro-fire detection head is a pure convolutional prediction layer branch deployed at the end of the multi-dimensional feature penetration perception module. After obtaining the spliced ​​tensor output by the front-channel recalibration unit, the micro-fire detection head uses a preset kernel size. The two-dimensional convolutional layer performs feature smoothing, followed by a kernel size of [missing information]. Dimensionality reduction and channel number adjustment are performed on the two-dimensional convolutional layer; The number of output channels of a convolutional layer strictly corresponds to the number of parameters required for prediction; that is, the output has one dimension. The three-dimensional prediction tensor, where the spatial dimension is... The representative divides the original image into A dense detection grid; The channel dimension includes the center point coordinate offset of the preset anchor box within each grid, the scaling of the border width and height, the confidence level, and the class probability. The 3D prediction tensor output by the micro fire detection head will be processed by a non-maximum suppression algorithm to remove overlapping redundant bounding boxes, outputting the final uniquely located prediction bounding box parameters, which will then be sent to the adaptive loss calculation unit to calculate the positioning error.

[0031] This invention extracts the aspect ratio of the actual labeled border in real time, which can keenly detect abnormal flame shapes that are severely elongated or torn by strong winds. It also dynamically adjusts the regression weights of the Euclidean distance of the center point and the aspect ratio difference in the total loss function. This flexible error penalty mechanism allows the prediction network to abandon the mechanical fitting of non-rigid flame edges when facing extremely narrow firestorms, thereby locking the core origin coordinates of the fire and greatly improving the spatial regression robustness of the model under extreme wind conditions.

[0032] The process of obtaining the penalty matrix specifically includes: Based on the predicted bounding box parameters and camera intrinsic parameter matrix, physical installation height and pitch angle output by each edge sensing terminal; By using inverse perspective projection transformation, the coordinates of the four pixel vertices of the predicted bounding box parameters are projected onto the local physical ground coordinate system with the vertical projection point of the corresponding edge sensing terminal as the origin, and the first coordinate pair corresponding to the pixel vertex is obtained. Obtain the actual installation coordinate pairs and installation orientation angles of each edge sensing terminal in the local spatial grid. Use a two-dimensional affine transformation matrix to perform rotation and translation operations on the first coordinate pair to obtain the second coordinate pair. Connect the second coordinate pairs sequentially to generate a global projection polygon of the grid system. Traverse the grid coordinates of the nodes of the local spatial grid. For any node's grid coordinates, emit a ray in the positive direction of the horizontal axis of the node's grid coordinates, and count the number of times the ray intersects with the boundaries of all global projected polygons of the grid system. Construct a penalty matrix. If the number of intersections is odd, determine that the node falls inside the projected polygon and overwrite the element in the penalty matrix corresponding to the node with the minimum relaxation weight. If the number of intersections is not odd, it is determined that the node does not fall inside the projected polygon, and the element at the position corresponding to the node in the penalty matrix is ​​overwritten as the maximum relaxation weight. In one specific embodiment of the present invention, during actual deployment, a distributed edge computing architecture is adopted to cover the entire forest area. For each independent IoT edge gateway, a two-dimensional local spatial grid is constructed within its physical monitoring coverage area, with the physical mounting point of the edge gateway as the origin. The total horizontal physical length of the physical monitoring coverage area is preset. The total physical width is 100 meters in length. For a 100-meter micro-environment monitoring field, the local spatial grid divides the monitoring area into a 10×10 discrete grid, with each grid mapped to a node. A fixed step size of 10 meters is set for the spatial resolution, a grid coordinate system is constructed, and the coordinates of each node are obtained. The coordinates of the node are the grid coordinates of the corresponding grid center point. The edge gateway receives the parameters of the predicted bounding box from the edge sensing terminal. The data format of the predicted bounding box parameters is the center point pixel coordinates, pixel width and height in the pixel coordinate system. It also collects the intrinsic parameter matrix, physical installation height and pitch angle of the lightweight dual-spectrum camera of the edge sensing terminal. The projection positions of these four vertices on the real physical ground are calculated using the inverse perspective projection transformation. The four projection positions are obtained with the ground directly below the terminal as the origin of the coordinate system, and are output as the first coordinate pair. Perform a two-dimensional affine transformation operation across coordinate systems: Collect the actual installation coordinate pairs of the edge sensing terminal in the local spatial grid. These coordinate pairs represent the translation distance of the edge sensing terminal relative to the origin at the lower left corner of the entire grid. Read the installation orientation angle of the edge sensing terminal's camera optical axis on the horizontal plane relative to the grid's vertical axis. Construct a two-dimensional affine transformation formula, and rotate and translate the first two-dimensional coordinate pair in sequence. Let any one of the first coordinate pairs be... Calculate the second coordinate pair transformed to the local space grid. The mathematical expressions for the x and y coordinates of the second coordinate pair are: ; ; in, For the first coordinate pair, For the second coordinate pair, The x-coordinate represents the actual installation location of the edge sensing terminal within the local spatial grid. The vertical coordinates of the actual installation location of the edge sensing terminal in the local spatial grid. The mounting orientation angle of the edge sensing terminal's camera relative to the grid's vertical axis on the horizontal plane; and This represents the pure rotation matrix operation components after eliminating the installation offset; plus and That is, the translation is completed; After calculating the four vertices of the predicted bounding box, the resulting second coordinate pairs are connected sequentially using line segments in the same coordinate system to generate a closed geometric shape, named the mesh system global projection polygon. The edge computing gateway processes data from multiple edge sensing terminals in parallel, and finally generates multiple global projection polygons of the mesh system in the same coordinate system. Construct a penalty matrix with a dimension of 10×10. Initialize the penalty matrix by uniformly assigning all elements in the penalty matrix the preset maximum penalty weight, which is 10.0. Traverse the physical coordinates of the nodes in the local spatial mesh. For any node, let its coordinates be... The determination is made using ray cross-detection logic, specifically, with Starting from the x-axis, a horizontal ray is emitted in the positive x-axis direction. The equation is solved using the intersection points of line segments, and the number of intersections between the horizontal ray and all boundary line segments of the global projection polygons of multiple mesh systems is counted. If the number of intersections is odd, according to the geometric topology theorem of closed polygons, it is determined that the node falls inside at least one globally projected polygon. This physically indicates that the node is identified as a fire source by the visual features of at least one edge-sensing terminal. In this case, the row and column indices of the node are extracted, and the element at that fixed position in the penalty matrix is ​​overwritten with a preset minimum relaxation weight, which is 0.001. When the number of intersections is even (including 0), it is uniquely determined that the center point of the grid is outside all globally projected polygons. The penalty matrix element is kept unchanged at 10.0. After traversing all nodes, the final penalty matrix is ​​output.

[0033] This invention reads the camera's intrinsic parameters, installation height, and azimuth angle, and uses inverse perspective projection transformation to accurately project and restore the two-dimensional fire point prediction box in the pixel coordinate system to the absolute physical ground coordinate system. Based on the polygon ray topology theorem, the global projection polygon of the grid system and the overwriting mechanism of extreme relaxation and penalty weights delineate high-confidence fire-prone physical areas in the discrete mathematical matrix. This mechanism seamlessly transforms the inference results of the front-end vision into a strongly constrained prior boundary in the process of solving the subsequent partial differential physical equations, which greatly narrows the solution space of the subsequent physical inversion algorithm and accelerates the calculation process.

[0034] The mathematical expression for the partial differential equation of the canopy drag force is: ; in, These are divergence and gradient differential operators used to calculate the rate of change of numerical values ​​between adjacent nodes within a spatial grid. Let be a vector consisting of the spatial coordinates of nodes on a two-dimensional local spatial grid. Let the first random variable be a two-dimensional random vector composed of the second random variable. This is a preset canopy drag coefficient containing random disturbances. For space state function, As a background of normal wind conditions, Let be the characteristic distribution function of the fire plume to be determined; The spatial gradient vector of the spatial state function; After numerical discretization on a two-dimensional local spatial grid, the state function, the background normal wind field, and the fire plume characteristic distribution function are mapped and stored as the state matrix, the background normal wind field matrix, and the fire plume characteristic matrix, respectively. The first random variable follows a uniform distribution in the interval -0.1 to 0.1; the second random variable follows a standard normal distribution. The mathematical expression for the canopy drag coefficient is: ; in, The value represents the basic drag force constant under windless conditions in the forest area; in this embodiment, the value is taken as 0.05. The coordinates of the nodes on the local spatial grid ( , The vector formed by ) For the first random variable With the second random variable The vector formed This represents the total physical length of the grid horizontally. This represents the total physical width of the grid vertically.

[0035] Used to characterize the periodic spatial impact of forest density disturbances along the horizontal axis of a two-dimensional grid. This is used to characterize the spatial impact of natural gust intensity disturbances along the vertical axis of a two-dimensional grid, where the canopy drag coefficient is used to represent random variables. This is converted into a physical drag coefficient that is linked to spatial coordinates; In a specific embodiment of the present invention, in a forest environment, the wind resistance brought by the tree canopy and the natural gusts have a strong randomness. If this randomness is not handled, conventional algorithms will misjudge the chaotic wind as fire turbulence. The present invention writes this randomness into the partial differential equation and reverse-derives and solves the pure fire characteristics after removing the natural random disturbance. Background normal wind force matrix Generated locally by the edge gateway, the specific generation process includes: The edge gateway receives macro weather forecast data from the superior meteorological station through a wireless communication module. The macro weather forecast data includes macro wind speed values ​​and macro wind direction angles. The edge gateway extracts the preset north angle of the two-dimensional local spatial grid, that is, the angle between the longitudinal north directions of the local spatial grid, and uses trigonometric functions to calculate the projection component of the macroscopic wind speed on the horizontal axis of the grid. The mathematical expression of the projection component is: ; in, For projected components, This refers to the macroscopic wind speed value. From a macro perspective, The preset north-facing angle; Construct a 10×10 blank two-dimensional matrix and apply the unique projection components calculated above. Assign values ​​to all one hundred elements in the two-dimensional matrix to generate the background normal wind pressure matrix, which represents the background of the uniform environmental wind field that the area should have under the condition of no local topography and forest interference.

[0036] The process of obtaining the desired observation state matrix specifically includes: A blank matrix is ​​constructed as the initial expected observation state matrix. For any target element in the expected observation state matrix, the Euclidean distance between the node grid coordinates corresponding to the target element and the grid coordinates of the edge sensing terminal is calculated. Based on the wind speed of the edge sensing terminal, the wind speed assignment of the target element is calculated using the inverse distance weight interpolation algorithm. The expected observation state matrix is ​​obtained by traversing and calculating the wind speed assignment.

[0037] In a specific embodiment of the present invention, this embodiment uses four edge sensing terminals, namely ultrasonic anemometers, to acquire four discrete horizontal wind speeds and the grid coordinates corresponding to the ultrasonic anemometers. Construct a 10×10 blank matrix as the initial desired observation state matrix. For any target element in the desired observation state matrix, let the coordinates of its corresponding node be... The inverse distance weighted interpolation algorithm is used to calculate the value of the element; Specifically, the Euclidean distance between the center point coordinates of the target element and the grid coordinates of each ultrasonic anemometer is calculated, and the wind speed value of the target element is calculated. The mathematical expression for the wind speed value is as follows: ; in, Assign a value to the wind speed of the target element. , , , These represent the wind speeds at various horizontal levels. , , , These are the Euclidean distances between the center point coordinates of the target element and the fixed physical coordinates of each ultrasonic anemometer; The wind speed assignment uses the square of the distance as an attenuation penalty. The closer the anemometer is to the target node, the greater the impact of the actual wind speed it collects on the weight of that node.

[0038] This invention constructs a partial differential equation for canopy drag force containing continuous two-dimensional random variables. By utilizing uniform distribution functions and normal distribution functions, it quantifies the random physical disturbances of uneven spatial distribution density of trees and the intensity of natural gusts, respectively. It also combines macro-meteorological wind direction projection and local inverse distance weighted interpolation techniques to construct the expected observation state and background normal wind drag matrix.

[0039] The process of obtaining the fire plume characteristic matrix and the flow field uncertainty tensor matrix specifically includes: Construct the reverse source tracing objective function. The mathematical expression of the reverse source tracing objective function is as follows: ; in, For reverse source tracing objective function, For the mathematical expectation of the error, For the penalty matrix, This represents the control variable matrix. The new matrix is ​​obtained by squaring each of the one hundred elements in the matrix. Represents the Hadamard product operation; For the mathematical variance of the error, The first preset weight is set to 0.05 in this embodiment; In one specific embodiment of the present invention, in traditional deterministic equations, the theoretical matrix is ​​usually directly approximated to the observation matrix to inversely deduce the fire plume characteristic matrix. However, in the present invention, because... Including random perturbations, the theoretical matrix results also fluctuate randomly. Therefore, it is necessary to construct an inverse source-finding objective function to combat uncertainty. By iteratively solving the inverse source-finding objective function, the fire plume flow characteristic matrix can be obtained. The process of calculating the expected value of the error specifically includes: Calculate the sum of squared differences between the corresponding elements of several theoretical state matrices and the expected observed state matrices, and then sum the sum of these sums of squared differences by weight to obtain the expected error. The mathematical variance of the error is the discrete variance of the sum of squares of several differences.

[0040] For the two random variables in the partial differential equation of canopy drag force, the Smoljak sparse grid algorithm is introduced to determine the coordinates of multiple configuration points and their corresponding weight coefficients. Initialize the fire plume feature matrix, and substitute the fire plume feature matrix and each configuration point into the partial differential equation of canopy drag force to obtain multiple theoretical state matrices. The sum of squared differences between the corresponding elements of each theoretical state matrix and the expected observed state matrix is ​​calculated to obtain multiple total error values; Multiply each total error value by its corresponding weighting coefficient, and sum all the products to obtain the expected value of the error. The square of the difference between each total error value and the expected value of the error is taken as the deviation. Each deviation is multiplied by the corresponding weighting coefficient, and all the product results are added together to obtain the error variance. By combining the gradient descent optimization algorithm, the element values ​​in the fire plume feature matrix are iteratively adjusted until the inverse source-finding objective function converges to a minimum value, and the converged fire plume feature matrix is ​​obtained. Based on the converged fire plume feature matrix and each configuration point, multiple theoretical state matrices are obtained; Extract multiple element values ​​corresponding to the same node from each theoretical state matrix, and calculate the variance of the multiple element values; Extract the variance of all nodes and construct the flow field uncertainty tensor matrix according to the node position order.

[0041] In one specific embodiment of the present invention, since the partial differential equation of canopy drag force contains continuous random variables, continuous integration cannot be directly performed when optimizing the inverse source-finding objective function. Therefore, a sparse grid random configuration method is adopted to transform the infinite derivation of the partial differential equation of canopy drag force into a lightweight finite-order deterministic calculation, specifically including: The partial differential equation for canopy drag includes tree density perturbation. With gust intensity disturbance Two random variables constitute a two-dimensional probability space. This invention introduces the Smojack sparse grid algorithm to prune the four diagonal edge nodes with extremely low probability of joint occurrence, retaining only the five core nodes along the cross axis, thus balancing computational accuracy and edge computing power. Regarding the determination of specific numerical values, based on the physical probability distribution characteristics of the random variable, the orthogonal polynomial root-finding method is used to calculate the node coordinates, specifically for variables following a standard normal distribution. The mathematical roots are extracted by solving the third-order Gauss-Hermitian orthogonal polynomial equation. , and As the ordinate, for a uniformly distributed system... Normalized roots were extracted using Gauss-Legend de Gauss orthogonal polynomials. , and As the horizontal axis coordinate; Based on Gauss's quadrature theorem, by calculating the derivative values ​​of the aforementioned orthogonal polynomial at each root, i.e. at the coordinates of the placement point, the weight coefficient corresponding to each placement point coordinate is obtained. The weight coefficient represents the discrete probability contribution of this specific perturbation state in the global continuous probability distribution. Finally, the tree density perturbation is set. With gust intensity disturbance The coordinate pairs formed are The coordinates of these five sets of configuration points are constant as follows: center point and their corresponding initial weights symmetrical point , The corresponding initial weights are all ,as well as , The corresponding initial weights are all ; After obtaining the initial weight coefficients corresponding to the coordinates of each configuration point according to Gauss's quadrature theorem, in order to meet the calculation requirements of mathematical expectation in probability theory, all initial weight coefficients are normalized. Specifically, each initial weight coefficient is divided by the sum of all initial weight coefficients to obtain the final normalized weight coefficients. Initialize the fire plume feature matrix, with all elements in the fire plume feature matrix being 0. Substitute the all-zero fire plume feature matrix and the five configuration points into the partial differential equation of canopy drag force to obtain five theoretical state matrices. Calculate the sum of squares of the differences between corresponding elements of the five theoretical state matrices and the observation matrices to obtain five total error values. Multiply each of the five total error values ​​by its corresponding normalized weight, and sum all the products to obtain the expected error. ; The square of the difference between the total error value and the expected value of the error is taken as the deviation. Each of the five deviations is multiplied by its corresponding weight, and the results of all the products are added together to obtain the mathematical variance of the error. ; Combining the gradient descent optimization algorithm, the values ​​of one hundred elements in the fire plume feature matrix are continuously adjusted iteratively until the calculation result of the inverse source-finding objective function decreases and converges to a minimum value. The iteration is stopped, and the converged fire plume feature matrix is ​​obtained. The elements of the fire plume feature matrix quantify the intensity of the abnormal wind field generated by the high temperature convection of the fire at each point in the local spatial grid, and this intensity has been completely stripped of the random interference of natural gusts. Based on the obtained fire feather flow feature matrix and five configuration points, five theoretical state matrices are obtained. Five element values ​​corresponding to the same node are extracted from the five theoretical state matrices, and the variance of the five element values ​​is obtained. Extract the variance corresponding to all nodes, and construct a 10×10 flow field uncertainty tensor matrix according to the position order of the nodes. The larger the value in the flow field uncertainty tensor matrix, the more severe the unpredictable weather caused by the destruction of the wind field by the fire storm at the actual physical location corresponding to the node.

[0042] This invention utilizes a very small number of core configuration points and normalized weight coefficients to transform the dimensionality reduction of continuous probability space integral operations into lightweight finite-time deterministic deductions. By constructing an inverse source-finding objective function that includes error mathematical expectation, Hadamard product penalty constraints, and error variance, and performing gradient descent optimization, it achieves powerful denoising of meteorological observation data. It also solves the pure fire plume feature matrix and flow field uncertainty tensor that are completely free from the interference of natural random wind fields, achieving extremely high-precision physical quantification and source tracing of fire thermodynamic characteristics in macroscopic space.

[0043] The operation process of the communication optimization module specifically includes: Collect the number of failed transmissions that did not receive an acknowledgment frame and the total number of transmission attempts within the current communication cycle, and divide the number of failed transmissions by the total number of transmission attempts to obtain the data packet collision rate; If the packet collision rate is greater than the preset congestion threshold, a broadcast storm is determined to have occurred, and the directed graph routing protocol based on the shortest geographical path is shut down. Calculate the Euclidean distance between nodes in the local spatial grid and the edge sensing terminal, and bind the edge sensing terminal to the node with the closest Euclidean distance; Extract the element values ​​corresponding to the nodes from the flow field uncertainty tensor matrix and use them as the fire risk mapping weights for the edge sensing terminal. If there is a fire risk mapping weight greater than the preset extreme value threshold, it is determined that there is an extreme fire risk, a local communication mask instruction is generated and broadcast to the edge sensing terminal, and the edge sensing terminal clears the image data queued in the local transmission queue. The fire hazard mapping weights of the edge sensing terminals are sorted in descending order of their numerical values. The time division multiple access mechanism is then activated, and time resources are allocated in descending order. The edge sensing terminal that obtains the time slot extracts the element at the corresponding position from the fire plume feature matrix and transmits it to the decision center.

[0044] In one specific embodiment of the present invention, the congestion threshold is set to 30%, and the preset extreme value threshold is set to 5.0. Immediately generate a local communication mask instruction and broadcast it to these 4 edge sensing terminals. When the edge sensing terminals receive the local communication mask instruction, they directly clear all the three-channel visible light image matrices and single-channel infrared image matrices queued in the local memory transmission queue. The core area of ​​the fire is covered with thick smoke, and the image matrix data is huge and has been completely blocked by smoke, becoming invalid noise. Clearing it can free up narrowband channel bandwidth. The four extracted fire risk mapping weights are sorted in descending order of their numerical values. The underlying MAC layer starts the time division multiple access mechanism and allocates time resources according to the gradient order of the descending arrangement. The edge sensing terminal with the largest fire risk mapping weight is allocated the first transmission time slot. The edge sensing terminal that obtains the time slot extracts the corresponding element from the fire plume feature matrix based solely on its mapped row and column indices and transmits it to the decision center. The edge sensing terminal that obtains the time slot transmits only a single-dimensional fire field disturbance scalar, which represents the pure fire plume thermal convection intensity obtained by inverse solving after rigorously stripping away natural gusts through partial differential equations. After receiving these extremely lightweight scalars that occupy only a few bytes, the decision center can directly map the set of coordinate points representing the extreme values ​​of the actual fire destructive force on the macro-geographic information system.

[0045] To address the problem of IoT link congestion and paralysis caused by massive concurrent data reporting from nodes during sudden fires, this invention designs an extreme communication preemption and optimized scheduling protocol driven by the physical characteristics of flow field uncertainty. When a broadcast storm is detected in the network and extreme fire risk is encountered, the conventional routing protocol is decisively shut down, and a local communication mask command is issued to forcibly clear the massive image payloads that have been blocked and rendered ineffective by smoke in each terminal queue. At the same time, the underlying MAC layer directly extracts the fire risk mapping weight obtained from physical inversion and starts the time-division multiple access descending order scheduling mechanism, so that the terminal in the most dangerous core area obtains the highest communication priority and only needs to send a few bytes of pure fire field feature scalar to complete the alarm.

[0046] This invention incorporates environmental change response vectors as prior knowledge into the channel recalibration process at the bottom layer of image processing. It can adaptively generate forced masking based on temperature rise gradients, actively suppressing visible light feature weights containing significant invalid noise, while simultaneously amplifying infrared thermal radiation feature weights that hide heat sources. This cross-modal dynamic fusion mechanism of physical environmental abrupt changes and dual-spectral image features, combined with an adaptive localization loss function that dynamically adjusts regression weights, effectively avoids false alarms and missed alarms of early-stage fires in harsh environments. Furthermore, this invention constructs a partial differential equation for canopy drag force containing spatially distributed random perturbations. The equations are solved by introducing a sparse grid algorithm and an inverse source-finding objective function for dimensionality reduction and minimization. The two-dimensional pixel-level fire point prediction bounding box is accurately projected and transformed into a spatial penalty matrix within the physical space grid. This filters out the random flow field disturbances caused by uneven forest distribution density and natural gusts. The pure fire plume thermal convection intensity and flow field uncertainty tensor are solved in reverse. This not only realizes high-precision physical source tracing of small fire sources in the macro-geographic coordinate system, but also provides objective quantitative indicators for fire risk assessment that are free from meteorological artifacts, and solves the coordinate drift problem of traditional positioning algorithms under strong wind conditions.

[0047] This invention utilizes a recursive least squares algorithm with an adaptive forgetting factor to construct a time-series equation, which can accurately remove inherent white noise interference from the environment, such as gusts of wind, and extract a pure environmental change response vector. Furthermore, this change response is used as prior knowledge to intervene in the channel recalibration process at the bottom layer of image processing. When facing extreme scenarios where dense smoke from a fire causes a sharp drop in the signal-to-noise ratio, it can adaptively generate a forced mask based on the temperature rise gradient, actively suppress the visible light feature weights containing a large amount of invalid noise, and simultaneously amplify the infrared thermal radiation feature weights that hide the heat source by several times. This cross-modal dynamic fusion mechanism of physical environmental change and dual-spectral image features, combined with an adaptive localization loss function that dynamically adjusts the regression weights, completely breaks the limitations of traditional visual monitoring that is susceptible to smoke interference and flame deformation, and effectively avoids false alarms and missed alarms of early fires in harsh scenarios. This invention possesses strong resistance to natural random interference in spatial positioning and fire feature analysis. Traditional solutions often struggle to distinguish between genuine fire plume turbulence and natural canopy winds. This invention constructs a partial differential equation for canopy drag force containing spatially distributed random disturbances and introduces the Smojack sparse grid algorithm and an inverse source-finding objective function for dimensionality reduction and minimization. This inversion mechanism accurately projects the two-dimensional pixel-level fire point prediction bounding box and transforms it into a spatial penalty matrix within a physical space grid. Then, through rigorous mathematical and physical equation deduction, it filters out random flow field disturbances caused by uneven tree distribution density and natural gusts, and solves inversely for the pure fire plume thermal convection intensity and flow field uncertainty tensor. This not only achieves high-precision physical source tracing of small fire sources in a macro-geographic coordinate system but also provides objective quantitative indicators for fire risk assessment that are free from meteorological artifacts, solving the coordinate drift problem of traditional positioning algorithms under strong wind conditions. This invention addresses the problem of IoT communication paralysis that is easily caused by extreme fire outbreaks. It proposes a deterministic congestion control and communication optimization strategy based on fire hazard physical feature mapping. When a surge in the collision rate of communication network data packets is detected and there are extreme fire hazard values, the system can intelligently determine the risk of a broadcast storm and break the conventional routing protocol. It directly issues a local communication mask command, forcing edge sensing terminals to clear the massive backlog of image matrix data that has been obscured and rendered invalid by smoke in their local transmission queues. This instantly releases valuable narrowband channel bandwidth. The system directly extracts the flow field uncertainty tensor as the scheduling weight for the time division multiple access mechanism, prioritizing the allocation of time resources to the terminals with the most critical fire conditions. This allows them to send only byte-level lightweight scalar data extracted from the fire plume feature matrix to the decision center. This eliminates the communication bottleneck caused by the concurrent transmission of massive sensing data during sudden fires, ensuring that the most critical fire damage coordinate set can be reported with minimal delay. This improves communication survivability and early warning response efficiency under extreme conditions.

[0048] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A forest fire risk early warning system based on Internet of Things (IoT) communication, characterized in that, It includes a data acquisition module, an environmental feature module, a feature fusion module, a feature inversion module, and a communication optimization module; The acquisition module is used to acquire forest area data from the edge sensing device; The environmental feature module is used to obtain the environmental change response vector based on the environmental temperature in the forest area data, combined with a recursive method with an adaptive forgetting factor. The feature fusion module is used to extract feature tensors from images in forest area data, map environmental change response vectors into scaling matrices, and perform weighted processing on feature tensors based on scaling matrices and forced masking, and output predicted bounding box parameters through a pre-trained detection head. The feature inversion module is used to obtain the penalty matrix based on the predicted bounding box parameters, construct the partial differential equation of canopy drag force and the expected observation state matrix, and iteratively solve the partial differential equation of canopy drag force by minimizing the inverse source-finding objective function including the penalty matrix to obtain the fire plume feature matrix and the flow field uncertainty tensor matrix. The communication optimization module is used to extract the element values ​​corresponding to the edge sensing terminal from the flow field uncertainty tensor matrix as fire risk mapping weights, sort them in descending order and allocate time resources, so that the edge sensing terminal can extract the element at the corresponding position from the fire plume feature matrix and transmit it to the decision center.

2. The forest fire risk early warning system based on Internet of Things communication as described in claim 1, characterized in that, The forest area data includes ambient temperature, image data, wind speed, weather forecast data, and installation and configuration data of edge sensing terminals; The meteorological forecast data includes macro wind speed values ​​and macro wind direction angles; The installation configuration data includes the camera intrinsic parameter matrix, physical installation height, and pitch angle of the edge sensing terminal; The image data includes visible light images and infrared images; The visible light image is a visible light image matrix including a red channel, a green channel, and a blue channel; The infrared image is an infrared image matrix comprising a single thermal radiation channel.

3. The forest fire risk early warning system based on Internet of Things communication as described in claim 2, characterized in that, The process of obtaining the environmental change response vector specifically includes: Construct an autoregressive time series equation, the mathematical expression of which is: ； in, The actual temperature at the current time step. Represents a matrix of historical observation data; The parameter vector represents the parameter to be identified; This represents the inherent white noise interference in the environment; A transpose matrix representing the historical observation data matrix; Based on the parameter vector identified at the previous moment, calculate the theoretical predicted temperature at the current moment, subtract the collected ambient temperature from the theoretical predicted temperature to obtain the predicted temperature residual, and calculate the mean square value of the predicted temperature residual within the sliding time window. Substitute the mean square value of the predicted temperature residual into the exponential forgetting factor update function to calculate the adaptive forgetting factor; The parameter vector is updated using a recursive least squares algorithm with an adaptive forgetting factor, and the constant term elements in the parameter vector are extracted as steady-state environment values. The predicted temperature residuals within the sliding time window are stacked in chronological order to form an environmental anomaly response vector.

4. The forest fire risk early warning system based on Internet of Things communication as described in claim 3, characterized in that, The feature fusion module includes a dual-modal feature extraction unit, a priori mapping unit, a channel recalibration unit, and a micro-fire detection head; The dual-modal feature extraction unit includes two parallel feature extraction branches, each consisting of multiple two-dimensional convolutional layers, used to map the visible light image matrix into a visible light feature tensor and the infrared image matrix into an infrared feature tensor. The prior mapping unit includes a two-layer fully connected network. The first fully connected layer maps the environmental change response vector to a hidden vector, and the second fully connected layer maps the hidden vector to a scaling matrix. The scaling matrix includes visible light initial attention weights and infrared initial attention weights.

5. The forest fire risk early warning system based on Internet of Things communication as described in claim 4, characterized in that, The channel recalibration unit is used to generate a forced masking matrix based on the signal-to-noise ratio and temperature rise gradient, multiply the scaling matrix and the forced masking matrix element by element to obtain an updated scaling matrix, and use the updated scaling matrix to perform weighted processing on the visible light feature tensor and the infrared feature tensor respectively to obtain the recalibrated visible light feature tensor and the recalibrated infrared feature tensor. The signal-to-noise ratio is the global variance of the grayscale values ​​of all pixels in the visible light image matrix; The temperature gradient is the sum of the absolute values ​​of the elements in the environmental change response vector; The process of generating the forced masking matrix specifically includes: If the signal-to-noise ratio is less than the preset occlusion threshold and the temperature rise gradient is greater than the preset gradient threshold, then the forced occlusion matrix includes a preset visible light penalty coefficient and a preset infrared gain coefficient. If the signal-to-noise ratio is greater than or equal to the occlusion threshold, or the temperature rise gradient is less than or equal to the gradient threshold, then the masking matrix is ​​forced to include the preset weight preservation coefficients. The micro-fire detection head is used to stitch together the recalibrated visible light feature tensor and the recalibrated infrared feature tensor in the channel dimension, and output the predicted bounding box through convolution classification and position regression.

6. The forest fire risk early warning system based on Internet of Things communication as described in claim 3, characterized in that, The feature fusion module also includes an adaptive loss calculation unit, which is used to calculate the localization loss during the offline training phase. The process of calculating the localization loss during the offline training phase specifically includes: The true bounding box parameters of the labeled sample and the predicted bounding box parameters output by the microfire detection head are obtained. The true bounding box parameters include the true center point x-coordinate, the true center point y-coordinate, the true bounding box width, and the true bounding box height. The predicted bounding box parameters include the predicted center point x-coordinate, the predicted center point y-coordinate, the predicted width, and the predicted height. The first deviation index is the deviation of the center point Euclidean distance, and the second deviation index is the deviation of the aspect ratio difference. Extract the aspect ratio of the actual border, and set the first regression weight and the second regression weight based on the aspect ratio; Substituting the first regression weight, the second regression weight, the first deviation index, and the second deviation index into the total positioning loss function, the mathematical expression of the total positioning loss function is as follows: ； in, For the total positioning loss, As the first regression weight, As the first deviation index, As the second regression weight, The second deviation index, This represents the intersection-union ratio (IoU) between the actual bounding box and the predicted bounding box.

7. The forest fire risk early warning system based on Internet of Things communication as described in claim 6, characterized in that, The process of obtaining the penalty matrix specifically includes: Based on the predicted bounding box parameters and camera intrinsic parameter matrix, physical installation height and pitch angle output by each edge sensing terminal; By using perspective projection inverse transformation, the coordinates of the four pixel vertices of the predicted bounding box parameters are projected onto a local physical ground coordinate system with the vertical projection point of the corresponding edge sensing terminal as the origin, and the first coordinate pair corresponding to the pixel vertex is obtained. Obtain the actual installation coordinate pairs and installation orientation angles of each edge sensing terminal in the local spatial grid, perform rotation and translation operations on the first coordinate pair using a two-dimensional affine transformation matrix to obtain the second coordinate pair, and connect the second coordinate pairs sequentially to generate a global projection polygon of the grid system. Traverse the grid coordinates of the nodes of the local spatial grid. For any node's grid coordinates, emit a ray in the positive direction of the horizontal axis of the node's grid coordinates, and count the number of times the ray intersects with the boundaries of all global projection polygons of the grid system. Construct a penalty matrix. If the number of intersections is odd, determine that the node falls inside the projected polygon, and overwrite the elements in the penalty matrix corresponding to the node with minimal relaxation weights. If the number of intersections is not odd, it is determined that the node does not fall inside the projected polygon, and the element at the position corresponding to the node in the penalty matrix is ​​overwritten with a maximal relaxation weight.

8. The forest fire risk early warning system based on Internet of Things communication as described in claim 7, characterized in that, The mathematical expression for the partial differential equation of the canopy drag force is: ； in, These are divergence and gradient differential operators used to calculate the rate of change of numerical values ​​between adjacent nodes within a spatial grid. Let be a vector consisting of the spatial coordinates of nodes on a two-dimensional local spatial grid. Let the first random variable be a two-dimensional random vector composed of the second random variable. This is a preset canopy drag coefficient containing random disturbances. For space state function, As a background of normal wind conditions, Let be the characteristic distribution function of the fire plume to be determined; The spatial gradient vector of the spatial state function; The state function, background normal wind field and fire plume characteristic distribution function are numerically discretized on the two-dimensional local space grid and then mapped and stored as state matrix, background normal wind field matrix and fire plume characteristic matrix, respectively. The mathematical expression for the canopy drag coefficient is: ； in, The basic drag force constant representing the forest area under windless conditions; The coordinates of the nodes on the local spatial grid ( , The vector formed by ) For the first random variable With the second random variable The vector formed This represents the total physical length of the grid horizontally. This represents the total physical width of the grid vertically; The process of obtaining the desired observation state matrix specifically includes: A blank matrix is ​​constructed as the initial expected observation state matrix. For any target element in the expected observation state matrix, the Euclidean distance between the node grid coordinates corresponding to the target element and the grid coordinates of the edge sensing terminal is calculated. Based on the wind speed of the edge sensing terminal, the wind speed assignment of the target element is calculated using the inverse distance weight interpolation algorithm. The expected observation state matrix is ​​obtained by traversing and calculating the wind speed assignment.

9. The forest fire risk early warning system based on Internet of Things communication as described in claim 8, characterized in that, The process of obtaining the fire plume feature matrix and the flow field uncertainty tensor matrix specifically includes: Construct a reverse source tracing objective function, the mathematical expression of which is: ； in, For reverse source tracing objective function, For the mathematical expectation of the error, For the penalty matrix, This represents the new matrix obtained by squaring all elements in the Firefeather Flow characteristic matrix; Represents the Hadamard product operation; For the mathematical variance of the error, The first preset weight; For the two random variables in the partial differential equation of canopy drag force, the Smoljak sparse grid algorithm is introduced to determine the coordinates of multiple configuration points and their corresponding weight coefficients. Initialize the fire plume feature matrix, and substitute the fire plume feature matrix and each configuration point into the partial differential equation of canopy drag force to obtain multiple theoretical state matrices. The sum of squared differences between the corresponding elements of each theoretical state matrix and the expected observed state matrix is ​​calculated to obtain multiple total error values; Multiply each total error value by its corresponding weighting coefficient, and sum all the products to obtain the expected value of the error. The square of the difference between each total error value and the expected value of the error is taken as the deviation. Each deviation is multiplied by the corresponding weighting coefficient, and all the product results are added together to obtain the error variance. By combining the gradient descent optimization algorithm, the element values ​​in the fire plume feature matrix are iteratively adjusted until the inverse source-finding objective function converges to a minimum value, and the converged fire plume feature matrix is ​​obtained. Based on the converged fire plume feature matrix and each configuration point, multiple theoretical state matrices are obtained; Extract multiple element values ​​corresponding to the same node from each theoretical state matrix, and calculate the variance of the multiple element values; Extract the variance of all nodes and construct the flow field uncertainty tensor matrix according to the node position order.

10. The forest fire risk early warning system based on Internet of Things communication as described in claim 9, characterized in that, The operation process of the communication optimization module specifically includes: Collect the number of failed transmissions that did not receive an acknowledgment frame and the total number of transmission attempts within the current communication cycle, and divide the number of failed transmissions by the total number of transmission attempts to obtain the data packet collision rate; If the packet collision rate is greater than the preset congestion threshold, a broadcast storm is determined to have occurred, and the directed graph routing protocol based on the shortest geographical path is shut down. Calculate the Euclidean distance between nodes in the local spatial grid and the edge sensing terminal, and bind the edge sensing terminal to the node with the closest Euclidean distance; Extract the element values ​​corresponding to the nodes from the flow field uncertainty tensor matrix and use them as the fire risk mapping weights for the edge sensing terminal. If there is a fire risk mapping weight greater than the preset extreme value threshold, it is determined that there is an extreme fire risk, a local communication mask instruction is generated and broadcast to the edge sensing terminal, and the edge sensing terminal clears the image data queued in the local transmission queue. The fire risk mapping weights of the edge sensing terminals are sorted in descending order of their numerical values, and time resources are allocated in descending order of their numerical values. The edge sensing terminal that obtains the time slot extracts the element at the corresponding position from the fire plume feature matrix and transmits it to the decision center.

Images