Intelligent forest fire prevention risk monitoring system based on digital twinning

Abstract

The application discloses a kind of wisdom forest fire prevention risk monitoring systems based on digital twinning, comprising: environment acquisition module, for collecting the multi-source environment state data in forest area, constructs environment state dataset;Risk modeling module, for environment state dataset is input to improved T-GNN model, generates risk evolution trajectory sequence;Mutation identification module, for performing fuzzy segmentation operation, constructs initial mutation point set;Early warning generation module, for calculating the trigger frequency and weighted score value of mutation point, filters and generates early warning trigger point set;Delay response index table is constructed;Trajectory calibration module, for error evaluation, complete calibration operation.The application can realize the accurate identification and dynamic early warning of forest fire risk evolution trend, with the advantages of timely response, controllable error, strong robustness, etc., suitable for intelligent fire prevention application scenarios in complex forest environment.

Description

Technical Field

[0001] This invention relates to the field of forest fire prevention and intelligent monitoring technology, and in particular to an intelligent forest fire risk monitoring system based on digital twins. Background Technology

[0002] With the increasing complexity of dynamic changes in forest ecosystems and the frequent occurrence of extreme weather events, the demand for forest fire risk monitoring and early warning is showing a significant upward trend. Most existing forest fire monitoring methods rely on static threshold judgments, rule-matching analysis, or coarse-grained models based on historical data, generally lacking the ability to dynamically perceive the evolutionary trends of real-time environmental states, making it difficult to meet the requirements for accurately capturing high-frequency mutations and continuous risk evolution. In practical deployments, traditional risk models often use fixed areas as analysis units, ignoring the unevenness and local coupling effects of spatiotemporal risk diffusion in the forest environment, leading to frequent accumulation of prediction errors and difficulty in timely tracking of risk evolution trajectories. Most risk assessment methods lack in-depth modeling of fire risk level changes over continuous time periods, relying solely on single-point index fluctuations for judgment, making it difficult to identify potential abrupt early warning signals. Furthermore, existing methods extract mutation information mostly based on static statistical thresholds, unable to adaptively adjust for error shifts occurring during risk evolution, resulting in problems such as delayed, false, or missed mutation point identification. Faced with the multi-source heterogeneity, dynamic mutation and delayed response characteristics of the forest environment, the lack of an intelligent fire risk monitoring system that can integrate spatial structure correlation, state evolution trend and confidence feedback mechanism limits the practicality and accuracy of existing methods in complex scenarios.

[0003] Therefore, how to provide a smart forest fire risk monitoring system based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a smart forest fire risk monitoring system based on digital twins. This invention constructs an environmental acquisition module, a risk modeling module, a mutation identification module, an early warning generation module, and a trajectory calibration module. It employs an improved T-GNN model to generate risk evolution trajectory sequences, combines a sliding window with risk gradient changes to perform fuzzy segmentation, extracts mutation points, and performs confidence accumulation calculations to generate transition early warning information. Based on a delayed response index table, it performs risk trajectory tracking and error assessment, and adjusts time-weighted parameters for mutation points whose historical cumulative error values ​​exceed the deviation calibration threshold to complete the confidence calibration operation. This system possesses advantages such as fast response speed, high identification accuracy, strong error calibration capability, and high environmental adaptability.

[0005] A smart forest fire risk monitoring system based on digital twins according to an embodiment of the present invention includes the following steps:

[0006] The environmental data acquisition module is used to collect multi-source environmental status data in forest areas and construct environmental status datasets in time series format.

[0007] The risk modeling module is used to input the environmental state dataset into the improved T-GNN model to generate a risk evolution trajectory sequence;

[0008] The mutation identification module is used to perform fuzzy segmentation on the risk evolution trajectory sequence, divide the state into segments based on the sliding window and risk gradient changes, and construct an initial mutation point set.

[0009] The early warning generation module is used to perform confidence accumulation calculation based on the initial set of mutation points, calculate the trigger frequency and weighted score value of mutation points in multiple consecutive time periods, and filter to generate a set of early warning trigger points; determine whether there are nodes in the set of early warning trigger points that meet the transition early warning conditions, and if so, generate transition early warning information; record the transition early warning information in time sequence and construct a delayed response index table;

[0010] The trajectory calibration module is used to evaluate the error between the prediction result of the mutation point and the fire risk confirmation information, and calculate the historical cumulative error value. If the historical cumulative error value is greater than the preset deviation calibration threshold, the time weighting parameter in the confidence cumulative calculation is adjusted to complete the calibration operation.

[0011] Preferably, the steps between modules include:

[0012] Step 1: Collect multi-source environmental status data in the forest area and construct an environmental status dataset in time series format;

[0013] Step 2: Input the environmental state dataset into the improved T-GNN model to generate a risk evolution trajectory sequence;

[0014] Step 3: Perform fuzzy segmentation on the risk evolution trajectory sequence, divide the state into segments based on the sliding window and risk gradient changes, and construct an initial set of mutation points;

[0015] Step 4: Based on the initial set of mutation points, perform confidence accumulation calculation to calculate the trigger frequency and weighted score of mutation points in multiple consecutive time periods, and filter to generate a set of early warning trigger points; determine whether there are nodes in the set of early warning trigger points that meet the transition early warning conditions, and if so, generate transition early warning information;

[0016] Step 5: Record the transition warning information in time sequence and construct a delayed response index table;

[0017] Step Six: Perform a risk trajectory tracking operation on the delayed response index table, extract the risk value change trend within the corresponding time period of the index, and calculate the actual risk increase. If the actual risk increase is greater than the preset correction threshold, perform a response correction operation, adjust the risk level prediction value of the corresponding mutation point, and output the corrected warning information.

[0018] Step 7: Evaluate the error between the predicted result of the mutation point and the fire risk confirmation information, and calculate the historical cumulative error value; if the historical cumulative error value is greater than the preset deviation calibration threshold, adjust the time weighting parameter in the confidence cumulative calculation to complete the calibration operation.

[0019] Preferably, step one specifically comprises:

[0020] An environmental data acquisition unit is set up within the monitoring area; the environmental data acquisition unit includes a temperature acquisition sensor, a humidity acquisition sensor, a wind speed acquisition sensor, a smoke concentration acquisition sensor, and a vegetation humidity acquisition sensor;

[0021] The environmental data acquisition unit synchronously collects environmental status information of the monitoring area according to a preset sampling period, generating raw environmental data;

[0022] Perform data validation on the raw environmental data, remove outlier and missing value records, and complete data cleaning;

[0023] The cleaned environmental data is serialized and arranged in chronological order to construct a time series data table containing sampling time index, spatial location information and environmental parameter values;

[0024] Perform formatting and encoding operations on the time series data table to generate an environmental state dataset in time series format according to a unified data structure template;

[0025] Perform block grouping operations on the environmental status dataset in time series format, and build a data grouping index based on forest partition identifiers and geographic coordinate information.

[0026] Preferably, step two specifically includes:

[0027] Perform modality label appending operations on the time-series formatted environmental state dataset to assign modality labels to data from different sources and generate labeled input tensors;

[0028] Perform temporal and positional encoding on the labeled input tensor to construct temporal and spatial embedding vectors, generating a multidimensional embedded input structure;

[0029] Based on the multidimensional embedded input structure, the state evolution features of each monitoring point in a continuous time period are extracted to generate a forest state object sequence;

[0030] Perform graph structure construction operations on the forest state object sequence, construct physical adjacency matrix and state adjacency matrix based on spatial adjacency relationship and state similarity relationship respectively, and generate a dual-channel graph structure representation;

[0031] The dual-channel graph structure representation is input into the graph convolution computation unit in the improved T-GNN model to calculate multidimensional attention weights, perform graph convolution operations, and extract state propagation features.

[0032] Perform temporal modeling operations on the state propagation characteristics to generate a risk prediction vector sequence based on the temporal evolution path of the nodes in the graph;

[0033] The risk prediction vector sequence is concatenated to form a risk evolution trajectory sequence; the risk evolution trajectory sequence includes the risk level estimate, state change trend value and confidence score value within a continuous time period.

[0034] Preferably, the improved T-GNN model includes an input construction unit, a state object construction unit, a dual-channel graph structure generation unit, a graph convolution calculation unit, a risk attention unit, and a temporal modeling unit, specifically:

[0035] The input building unit receives multi-source data from the environmental state dataset, adds modal labels to each type of data, and generates a time embedding representation structure by encoding all data according to timestamps. The output is an input tensor containing the time embedding dimension and the modal label dimension.

[0036] The state object construction unit receives the input tensor, extracts continuous evolution segments according to spatial location and time window, encapsulates time index, spatial coordinates and risk-related features for each segment, and generates a forest state object set.

[0037] The dual-channel graph structure generation unit receives a set of forest state objects, generates a spatial adjacency matrix based on the physical distance relationship between spatial coordinates, generates a state adjacency matrix based on the similarity of evolutionary trends between state features, and weights and fuses the spatial adjacency matrix and the state adjacency matrix to construct a graph structure representation tensor containing node indexes, connection weights and connection labels.

[0038] The graph convolutional computation unit receives the graph structure representation tensor, performs node feature aggregation and propagation operations according to the computation process of the graph neural network, and outputs multi-layer graph convolutional feature representation.

[0039] The risk attention unit introduces a state sensitivity factor and a risk gradient change factor during graph convolution, respectively calculates the risk correlation score and propagation weight adjustment coefficient between nodes, adjusts the propagation intensity in the graph convolution path according to the risk correlation score, and generates a weighted graph feature representation with risk attention weight.

[0040] The time-series modeling unit receives the weighted graph feature representation, performs sequence modeling operations according to the node time index, outputs a risk prediction vector sequence within a continuous time period, and concatenates the risk prediction vector sequence into a risk evolution trajectory sequence.

[0041] Preferably, step three specifically includes:

[0042] A fixed-length sliding time window is set for the risk evolution trajectory sequence, and the entire time period is traversed in chronological order to extract the risk level estimate sequence and the state change trend sequence within each window.

[0043] The average gradient change value is calculated for the risk level estimate sequence within each sliding time window, the mutation sensitivity feature is extracted, and the risk fluctuation intensity index is calculated by combining the state change trend value sequence.

[0044] The risk volatility intensity index is compared with the set mutation identification threshold. If the volatility intensity is greater than the mutation identification threshold, it is marked as a candidate mutation point at the end of the window.

[0045] All candidate mutation points are merged in chronological order. Fuzzy clustering is performed by combining the continuity of risk gradient changes in adjacent time periods to remove duplicate labels and short-term abnormal fluctuation labels, and an initial mutation point set is output. Each mutation point in the initial mutation point set is labeled with a mutation direction label and a local risk increase value label.

[0046] Preferably, step four specifically includes:

[0047] The frequency of occurrence of each mutation point in the initial mutation point set within a continuous time period is statistically analyzed to construct a mutation trigger frequency sequence.

[0048] The confidence-weighted score is calculated by combining the risk level increase value of each mutation point over a continuous time period, the consistency of mutation direction, and historical evolution background information.

[0049] Set a mutation confidence screening threshold, filter mutation points with a trigger frequency higher than the set threshold and a weighted score higher than the screening threshold, and construct a set of early warning trigger points;

[0050] Each mutation point in the set of early warning trigger points is arranged in chronological order and spatial location, and combined with the historical risk evolution trajectory, continuous surge patterns and clustered mutation patterns are identified;

[0051] Determine whether there is a mutation point that simultaneously satisfies the trigger frequency threshold condition, the weighted score threshold condition, and the continuous surge pattern condition. If so, mark it as a transition warning trigger node. Generate transition warning information for each transition warning trigger node. The transition warning information includes a mutation time index, spatial location code, confidence level score, and risk improvement trend label.

[0052] Preferably, step five specifically includes:

[0053] The mutation time index and spatial location code in the transition warning information are arranged in chronological order to construct a transition warning sequence; a unique serial number is assigned to each transition warning information in the transition warning sequence, and the corresponding risk level score, risk improvement trend label and historical trigger frequency are recorded;

[0054] Perform time period labeling operation on the transition warning sequence, map each transition warning information to a sampling time window, and label the window start time, end time and span value;

[0055] The time difference between adjacent transition warning information is calculated to generate a time interval vector; for each transition warning information, the corresponding forest state object within the time period is matched, and the state change trend data within the time period in the risk evolution trajectory is extracted.

[0056] By combining the time interval vector and state change trend data, a delayed response index table is constructed; the delayed response index table includes the transition warning sequence number, time span parameter, state change characteristics, risk prediction residual, and trend offset index.

[0057] Preferably, step six specifically includes:

[0058] Perform trajectory backtracking operation on the time period corresponding to each time span parameter in the delayed response index table, and extract the risk level estimate sequence within the corresponding time period from the risk evolution trajectory sequence;

[0059] Perform difference calculation and smoothing on the risk level estimate sequence to construct a risk change trend vector;

[0060] Calculate the risk increase magnitude within a time period based on the risk change trend vector to form an actual risk increase magnitude sequence;

[0061] Set a correction threshold range, compare the magnitude values ​​in the actual risk increase sequence item by item, and if the magnitude value is greater than the correction threshold, mark the corresponding time period and locate the associated mutation point number.

[0062] Extract the original risk level prediction value and confidence score value for the marked mutation point number, perform the correction coefficient calculation operation, and calculate the correction gain based on the product of the risk increase and the confidence score value;

[0063] The modified gain is added to the original risk level prediction value to generate the modified risk level prediction value.

[0064] The corrected warning information is output, which includes the mutation point number, time index, spatial location code, corrected risk level prediction value, risk increase value, and corrected gain weight.

[0065] Preferably, step seven specifically comprises:

[0066] For each mutation point number, the predicted risk level value corresponding to the risk level is calculated to differ from the actual risk level data in the fire risk confirmation information, generating a sequence of mutation point prediction deviation values.

[0067] The prediction deviation value sequence of mutation points is aggregated according to the time index to construct a historical cumulative error table. The historical cumulative error table records the cumulative prediction deviation value of each mutation point number in multiple consecutive time periods.

[0068] Set a deviation calibration threshold, compare the cumulative values ​​of each mutation point number in the historical cumulative error table item by item, and filter out the set of mutation point numbers whose cumulative values ​​are greater than the deviation calibration threshold;

[0069] Extract the confidence accumulation calculation path corresponding to the mutation point number set, and obtain the time-weighted parameters and scoring weight records;

[0070] A correction factor is calculated based on the cumulative error value and the scoring weight; the correction factor includes a time decay coefficient and a scoring smoothing coefficient.

[0071] The correction factor is applied to the original time-weighted parameters to update the weight distribution structure in the confidence accumulation calculation, generating a calibrated set of time-weighted parameters, thus completing the calibration operation.

[0072] The beneficial effects of this invention are:

[0073] This invention constructs an environmental acquisition module, a risk modeling module, a mutation identification module, an early warning generation module, and a trajectory calibration module. Combining an improved T-GNN model, it performs graph convolution calculations and temporal modeling on multi-source environmental state data, generating a risk evolution trajectory sequence containing risk level estimates and state change trend values. Addressing the problems of low mutation identification accuracy, delayed early warning triggering, and uncontrollable accumulation of risk level errors in existing forest fire monitoring, this invention proposes a fuzzy segmentation method based on sliding windows and risk gradient changes. It dynamically constructs an initial mutation point set and generates transition early warning information by combining confidence accumulation calculation and transition pattern recognition. Furthermore, it establishes a delayed response index table to track actual risk change trends and correct mutation point prediction results, improving the accuracy of early warning levels. To address the problem of mutation point prediction error accumulation, it sets a deviation calibration threshold and dynamically adjusts time weighting parameters to complete adaptive calibration of the confidence scoring path, effectively reducing prediction bias. Ultimately, it achieves high-precision evolution modeling of fire risk, sudden situation identification, and continuous early warning output under complex environmental conditions in forest areas, improving the real-time performance, accuracy, and controllability of forest fire response. Attached Figure Description

[0074] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0075] Figure 1 This is a schematic diagram of a smart forest fire risk monitoring system based on digital twins proposed in this invention;

[0076] Figure 2 This is a flowchart of a smart forest fire risk monitoring system based on digital twins proposed in this invention;

[0077] Figure 3 This is a schematic diagram of the improved T-GNN model proposed in this invention. Detailed Implementation

[0078] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0079] refer to Figure 1-3 A smart forest fire risk monitoring system based on digital twins includes:

[0080] The environmental data acquisition module is used to collect multi-source environmental status data in forest areas and construct environmental status datasets in time series format.

[0081] The risk modeling module is used to input the environmental state dataset into the improved T-GNN model to generate a risk evolution trajectory sequence;

[0082] The mutation identification module is used to perform fuzzy segmentation on the risk evolution trajectory sequence, divide the state into segments based on the sliding window and risk gradient changes, and construct an initial mutation point set.

[0083] The early warning generation module is used to perform confidence accumulation calculation based on the initial set of mutation points, calculate the trigger frequency and weighted score value of mutation points in multiple consecutive time periods, and filter to generate a set of early warning trigger points; determine whether there are nodes in the set of early warning trigger points that meet the transition early warning conditions, and if so, generate transition early warning information; record the transition early warning information in time sequence and construct a delayed response index table;

[0084] The trajectory calibration module is used to evaluate the error between the prediction result of the mutation point and the fire risk confirmation information, and calculate the historical cumulative error value. If the historical cumulative error value is greater than the preset deviation calibration threshold, the time weighting parameter in the confidence cumulative calculation is adjusted to complete the calibration operation.

[0085] In this embodiment, the steps between modules are as follows:

[0086] Step 1: Collect multi-source environmental status data in the forest area and construct an environmental status dataset in time series format;

[0087] Step 2: Input the environmental state dataset into the improved T-GNN model to generate a risk evolution trajectory sequence;

[0088] Step 3: Perform fuzzy segmentation on the risk evolution trajectory sequence, divide the state into segments based on the sliding window and risk gradient changes, and construct an initial set of mutation points;

[0089] Step 4: Based on the initial set of mutation points, perform confidence accumulation calculation to calculate the trigger frequency and weighted score of mutation points in multiple consecutive time periods, and filter to generate a set of early warning trigger points; determine whether there are nodes in the set of early warning trigger points that meet the transition early warning conditions, and if so, generate transition early warning information;

[0090] Step 5: Record the transition warning information in time sequence and construct a delayed response index table;

[0091] Step Six: Perform a risk trajectory tracking operation on the delayed response index table, extract the risk value change trend within the corresponding time period of the index, and calculate the actual risk increase. If the actual risk increase is greater than the preset correction threshold, perform a response correction operation, adjust the risk level prediction value of the corresponding mutation point, and output the corrected warning information.

[0092] Step 7: Evaluate the error between the predicted result of the mutation point and the fire risk confirmation information, and calculate the historical cumulative error value; if the historical cumulative error value is greater than the preset deviation calibration threshold, adjust the time weighting parameter in the confidence cumulative calculation to complete the calibration operation.

[0093] This implementation method sets up an environmental acquisition module, a risk modeling module, a mutation identification module, an early warning generation module, and a trajectory calibration module, which sequentially perform operations such as forest environmental status data acquisition, risk modeling, risk mutation identification, early warning information generation, and prediction calibration, and has a clear processing flow and stable operating logic. By collecting multi-source environmental status data from forest areas and constructing a time-series dataset, the dynamic environmental changes within the monitoring area can be comprehensively reflected. Inputting the environmental status dataset into an improved T-GNN model generates a risk evolution trajectory sequence, enabling high-precision modeling of the spatiotemporal evolution of forest risks. Performing fuzzy segmentation on the risk evolution trajectory sequence to construct an initial set of abrupt change points helps in the timely detection of potential risk transition trends. Based on the initial set of abrupt change points, confidence accumulation calculation is performed to generate a set of early warning trigger points and determine the conditions for transition warnings, further improving the accuracy and reliability of early warning judgments. Time-series recording of transition warning information and construction of a delayed response index table provide a data foundation for risk posterior tracking. Performing risk trajectory tracking on the delayed response index table extracts the trend of risk value changes and determines whether the predicted risk level needs correction, further enhancing the dynamic adjustment capability of the early warning results. By evaluating the error between the prediction results and fire risk confirmation information and adjusting the time weighting parameters in the confidence accumulation calculation, model calibration and accuracy optimization can be achieved, thereby improving the overall stability and real-time response effect of forest fire risk monitoring.

[0094] In this embodiment, step one specifically includes:

[0095] An environmental data acquisition unit is set up within the monitoring area; the environmental data acquisition unit includes a temperature acquisition sensor, a humidity acquisition sensor, a wind speed acquisition sensor, a smoke concentration acquisition sensor, and a vegetation humidity acquisition sensor. The location and distribution density of the environmental data acquisition unit are obtained by fitting an optimization model based on the risk level of the forest zone and the historical frequency of fire occurrence.

[0096] The environmental data acquisition unit synchronously collects environmental status information of the monitoring area according to a preset sampling period to generate raw environmental data. The preset sampling period is a dynamic sampling period ranging from minutes to hours. The sampling period is adjusted according to the environmental fluctuations and sensor stability of the monitoring area. The adjustment function is obtained by jointly fitting the sensor stability weight function and the sampling delay model.

[0097] Perform data validation operations on the raw environmental data, including format validity validation, sampling interval integrity validation, and spatial coordinate consistency validation; perform anomaly removal processing on the data validation results, mark and remove duplicate sampling records, missing field records, and logical conflict records, and generate a structured clean dataset;

[0098] The structured cleaning dataset was sequentially arranged according to time order to construct a time series data table containing sampling time index, spatial location information and five types of environmental parameter values. The sampling time index was expressed in a unified timestamp format, the spatial location information included latitude and longitude coordinates and monitoring point number, and the environmental parameter values ​​included temperature, humidity, wind speed, smoke concentration and vegetation humidity.

[0099] The time series data table is formatted and encoded to generate an environmental state dataset in time series format according to a unified data structure template. The data structure template includes field encoding specifications, field order rules, value range specifications, and default processing strategies. All field values ​​are mapped to a standardized interval through a normalization function, which is obtained by fitting the historical maximum and minimum value range.

[0100] A block-based grouping operation is performed on the environmental state dataset in time series format, and a data grouping index is established based on forest partition identifiers and geographic coordinate information. The partition identifiers and coordinate information constitute grouping key values. The dataset is organized according to a tree-like hierarchical structure to construct a hierarchical data partitioning index table, and it is bound one-to-one with the node initialization operation in the subsequent graph structure generation process.

[0101] In this embodiment, step two specifically includes:

[0102] Modal labeling is applied to the time-series environmental state dataset to assign modal labels to data from different sources, generating labeled input tensors. Modal labels include temperature, humidity, wind speed, smoke concentration, and vegetation humidity modes. The numerical weights corresponding to each modal label are obtained by linear fitting based on the correlation coefficient matrix of historical environmental data. Modal weights are used to balance the importance distribution of different environmental parameters during the input phase.

[0103] A temporal location encoding operation is performed on the labeled input tensor to construct temporal and spatial embedding vectors, generating a multidimensional embedded input structure. The temporal embedding vector is generated by mapping the periodic function of the time index, and the spatial embedding vector is generated by the hash encoding function of geographic coordinates and partition identifiers. The temporal and spatial embeddings are normalized and fused to form a multidimensional embedded input structure. The parameters of the embedding function are obtained by fitting the minimum mean square error of the historical spatiotemporal coupling features.

[0104] Based on the multidimensional embedded input structure, the state evolution features of each monitoring point in a continuous time period are extracted to generate a forest state object sequence. The state evolution features include parameter change rate, fluctuation amplitude and time stability index. The fluctuation amplitude is calculated by the sliding window mean square error function and the time stability is obtained by fitting the autocorrelation function. Each forest state object contains a time index, spatial coordinates and state evolution vector.

[0105] A graph structure construction operation is performed on the forest state object sequence. Physical adjacency matrices and state adjacency matrices are constructed based on spatial adjacency and state similarity relationships, respectively, generating a dual-channel graph structure representation. The physical adjacency matrix is ​​constructed based on the geographical distance threshold of monitoring points, which is obtained by fitting the regional distribution density. The state adjacency matrix is ​​constructed based on the cosine similarity of state vectors, and the similarity threshold is determined by fitting the historical risk synergy coefficient curve. The dual-channel graph structure is synthesized into a graph structure tensor through a weighted fusion function, and the weight ratio is obtained by fitting the optimal feature preservation criterion.

[0106] The dual-channel graph structure representation is input into the graph convolution computation unit in the improved T-GNN model to calculate multidimensional attention weights, perform graph convolution operations, and extract state propagation features. The graph convolution computation is performed based on node aggregation and edge weight propagation rules, and the multidimensional attention weights are obtained by jointly fitting the risk sensitivity function and the node influence function. The risk propagation path between nodes is dynamically adjusted through the attention distribution matrix.

[0107] Temporal modeling is performed on the state propagation characteristics to generate a risk prediction vector sequence based on the temporal evolution path of the nodes in the graph. The temporal modeling uses recursive temporal convolution to process continuous time nodes, and the time weight coefficients are obtained by fitting the historical state response delay curve. The risk prediction vector sequence includes risk intensity components, propagation trend components, and time stability components.

[0108] The risk prediction vector sequence is concatenated to form the risk evolution trajectory sequence; the risk evolution trajectory sequence includes the risk level estimate, state change trend value and confidence score value within a continuous time period; the risk level estimate is calculated through the risk mapping function, and the confidence score value is dynamically updated through the output error inverse correction function.

[0109] This implementation method further enhances the spatiotemporal correlation modeling capability and risk evolution prediction accuracy of multi-source forest environment data by introducing modality identification, dual-channel graph structure and dynamic attention weight mechanism into the improved T-GNN model.

[0110] In this embodiment, the improved T-GNN model includes an input construction unit, a state object construction unit, a dual-channel graph structure generation unit, a graph convolution calculation unit, a risk attention unit, and a temporal modeling unit, specifically:

[0111] The input building unit receives multi-source data from the environmental state dataset, adds modal labels to each data type, and generates modal labels by mapping data source type to sampling device number; it encodes all data according to timestamp to generate a time embedding representation structure, which is generated by sine and cosine position encoding functions, and the time step is set according to the linear relationship between sampling period and change response delay; it concatenates the time embedding vector and modal label vector to generate an input tensor containing time embedding dimension and modal label dimension; each dimension in the input tensor represents the environmental state expression vector at a single time and in a single modality;

[0112] The state object construction unit receives an input tensor and extracts continuous evolution segments according to spatial location and time window. The length of the time window is obtained by fitting the statistical median of the risk evolution cycle. Each segment is encapsulated with a time index, spatial coordinates, and risk-related features. The risk-related features include the rate of change of environmental variables, the amplitude of periodic fluctuations, and the difference between adjacent time points. The rate of change is calculated by moving difference, and the amplitude of periodic fluctuations is obtained by fitting the variance function within the window. Finally, a set of forest state objects is generated, and each object is bound to a unique identifier, a location label, and a state representation vector.

[0113] The dual-channel graph structure generation unit receives a set of forest state objects and generates a spatial adjacency matrix based on the physical distance relationship between spatial coordinates. The distance threshold is set by fitting the sensor density distribution and response influence radius within the forest partition. A state adjacency matrix is ​​generated based on the evolutionary trend similarity between state features. The state similarity is calculated using a cosine similarity function, and the similarity threshold is fitted by fitting the minimum similarity between state sequences before and after historical abrupt changes. The spatial adjacency matrix and the state adjacency matrix are weighted and fused. The fusion weight parameters are fitted by fitting a joint optimization function that maximizes graph structure connectivity and state resolution. A graph structure representation tensor containing node indices, connection weights, and connection labels is constructed.

[0114] The graph convolutional computation unit receives the graph structure representation tensor and performs node feature aggregation and propagation operations according to the computation process of the graph neural network. The propagation function includes weighted summation of adjacent nodes and nonlinear activation mapping. The graph convolution depth is fitted according to the optimal balance between the average path length and state aggregation capability in the graph. The output is a multi-layer graph convolutional feature representation, with each layer representing the local structure propagation features at different scales.

[0115] The risk attention unit introduces a state sensitivity factor and a risk gradient change factor during graph convolution. The state sensitivity factor is calculated based on the degree of synergy between risk volatility and the propagation path, while the risk gradient change factor is obtained by fitting the magnitude of risk change and the direction of propagation over a continuous time period. Risk correlation scores and propagation weight adjustment coefficients between nodes are calculated separately. The risk correlation score is calculated using a Gaussian weighting function, and the propagation weight adjustment coefficient is dynamically adjusted through an error back-update mechanism. The propagation intensity in the graph convolution path is adjusted according to the risk correlation score, generating a weighted graph feature representation with risk attention weights.

[0116] The temporal modeling unit receives the weighted graph feature representation and performs sequence modeling operations according to the node time index. The sequence modeling adopts a joint design of positional recursion structure and graph sequence nesting structure. The positional recursion structure introduces a time gating factor, while the graph sequence nesting structure retains the adjacency propagation path information. It outputs a risk prediction vector sequence within a continuous time period, which includes the risk level prediction value, confidence score value, and state fluctuation trend value. The risk prediction vector sequence is then concatenated into a risk evolution trajectory sequence, and the concatenation rule is obtained by optimizing the fit using the minimum prediction error criterion.

[0117] This implementation method constructs an improved T-GNN model, which integrates multi-source state representation, dual-channel graph representation, risk-sensitive propagation, and graph sequence modeling mechanisms. This further enhances the spatiotemporal identification capability of potential state mutations in forest fire risk data, improves the continuity and accuracy of risk evolution trajectory generation, and provides a key model foundation for accurate monitoring and intelligent early warning of forest fire risk.

[0118] In this embodiment, step three specifically includes:

[0119] A fixed-length sliding time window is set for the risk evolution trajectory sequence. The length of the sliding time window is obtained by fitting a linear regression model based on the ratio of the average response period of forest risk status changes to the sampling frequency. The entire time period is traversed in chronological order, and the risk level estimate sequence and the status change trend sequence within each window are extracted. Each time window contains multiple continuous risk sample points, and a time index sequence is formed by mapping through timestamps.

[0120] For each sliding time window, the average gradient change value of the risk level estimate sequence is calculated. The average gradient change value is obtained by the difference mean function between continuous risk samples. Sudden change sensitive features are extracted, including gradient change rate, local variance, and trend fluctuation amplitude. The local variance is obtained by fitting the sliding variance function. The risk fluctuation intensity index is calculated by combining the state change trend value sequence. The risk fluctuation intensity index is calculated by the weighted superposition function of gradient change rate and trend fluctuation amplitude. The weight ratio is obtained by fitting the optimal classification boundary of historical risk sudden change samples.

[0121] The risk volatility intensity index is compared with the set mutation identification threshold, which is obtained by fitting the empirical relationship between the standard deviation range of historical risk samples and the average volatility. If the volatility intensity is greater than the mutation identification threshold, it is marked as a candidate mutation point at the end of the window.

[0122] All candidate mutation points are merged in chronological order, and candidate mutation points with a time interval less than twice the window step size are considered to be in the same risk interval. Fuzzy clustering is performed based on the continuity of risk gradient changes in adjacent time periods. The clustering membership function is obtained by fitting a two-parameter Gaussian function of risk fluctuation intensity and time interval. Duplicate labels and short-term abnormal fluctuation labels are removed from the clustering results. Short-term abnormal fluctuations are screened out using a short-term fluctuation threshold judgment function. An initial set of mutation points is output, and each mutation point records the time index, risk level increase value, and state change rate.

[0123] Each mutation point in the initial mutation point set is labeled with a mutation direction label and a local risk increase value label. The mutation direction label is generated based on the sign relationship of the risk gradient before and after, and the local risk increase value is calculated using the risk mean difference function within the window. The mutation direction and increase value information serve as input features for subsequent confidence accumulation calculation operations.

[0124] This implementation method further improves the accuracy and stability of risk mutation detection by introducing a sliding time window, multi-index fluctuation feature extraction and fuzzy clustering fusion mechanism, and enhances the ability to identify small mutations in forest fire risk trajectories, providing a high-confidence mutation feature basis for subsequent early warning trigger point screening and transition risk judgment.

[0125] In this embodiment, step four specifically includes:

[0126] The number of occurrences of each mutation point in the initial mutation point set within a continuous time period is statistically analyzed. The length of the continuous time period is set as the shortest stable period of the mutation, which is obtained by statistical fitting of the minimum stable duration of mutation points in historical samples. The statistical results are used to construct a mutation trigger frequency sequence, and the mutation trigger frequency value is obtained by fitting the cumulative number of times in the sliding window by the window step size.

[0127] The confidence-weighted score is calculated by combining the risk level increase value, mutation direction consistency, and historical evolution background information of each mutation point over a continuous time period. The mutation direction consistency value is calculated by the consistency rate of mutation direction labels in adjacent windows. The historical evolution background information includes the stability index of the previous stage and the difference between adjacent risk levels. The stability index is calculated using the moving variance evaluation function. The confidence-weighted score is based on the weighted combination of mutation intensity, direction consistency, and evolution background. The weight parameters are obtained by fitting the mutation confidence function in the risk backtracking analysis.

[0128] A mutation confidence screening threshold is set, which is obtained by constructing a function based on the difference between the median of the mutation validity distribution and the tail value of the confidence distribution in the multi-region forest fire risk evolution sample. Mutation points with a trigger frequency higher than the set number threshold and a weighted score higher than the screening threshold are screened to construct a set of early warning trigger points. During the screening process, a spatial non-isolation judgment rule is added to remove mutation points that are spatially isolated or have broken risk transmission chains.

[0129] Each mutation point in the set of early warning trigger points is arranged according to time sequence and spatial location to construct an early warning trigger sorting matrix. The rows of the matrix represent the time sequence order, and the columns represent the spatial coordinate indices. Combined with the historical risk evolution trajectory, a pattern recognition operation is performed to identify continuous surge patterns and clustered mutation patterns. The continuous surge pattern is obtained by fitting a time window in which the risk level increases positive for three or more consecutive periods. The clustered mutation pattern is obtained by calculating the clustering degree function on the spatial adjacency graph.

[0130] Determine whether there is a mutation point that simultaneously satisfies the trigger frequency threshold condition, the weighted score threshold condition, and the continuous surge pattern condition; if so, mark it as a transition warning trigger node, and the transition warning trigger node satisfies the triple conditions of high confidence, high frequency triggering, and pattern significance;

[0131] A transition warning information is generated for each transition warning trigger node; the transition warning information includes a mutation time index, spatial location code, confidence level score and risk improvement trend label; the confidence level score is obtained by mapping the normalized weighted score value to a risk level range; the risk improvement trend label is output after judging the direction of risk increase by a trend discriminant function.

[0132] This implementation method further enhances the ability to screen for significant forest fire risk mutation points by constructing a multi-factor weighted confidence scoring mechanism, introducing a spatial temporal mutation pattern recognition method, and integrating clustering rule judgment, thereby improving the accuracy of transition warning nodes and the foresight of risk response.

[0133] In this embodiment, step five specifically includes:

[0134] The mutation time index and spatial location code in the transition warning information are arranged in chronological order to construct a transition warning sequence. The transition warning sequence is a time-ordered set of risk mutation nodes, and each node includes a time index, spatial location, confidence level score, and trend label. A unique sequence number is assigned to each transition warning information in the transition warning sequence, and the sequence number is set to increment in chronological order. The corresponding risk level score, risk improvement trend label, and historical trigger frequency are recorded. The historical trigger frequency is obtained by fitting the cumulative number of mutations at the same location within a sliding time window.

[0135] A time period labeling operation is performed on the transition warning sequence, mapping each transition warning information to a sampling time window. The sampling time window is constructed based on the data sampling period set in the monitoring system. The start time, end time, and span value of the window are also labeled, with the span value being the time difference between the start time and the end time.

[0136] The time difference between adjacent transition warning messages is calculated to generate a time interval vector; the time interval vector represents the response interval between adjacent warning nodes and is used to analyze the system's sensitivity to continuous sudden changes and response delay characteristics.

[0137] For each transition warning information, match the forest state object within the corresponding time period, and extract the state change trend data within the time period in the risk evolution trajectory; the state change trend data includes the slope of the risk level estimate, trend inflection point and fluctuation amplitude at continuous time points; the state object is located by spatial index and time index, and the risk attribute subset under the corresponding time segment is extracted.

[0138] By combining the time interval vector and state change trend data, a delayed response index table is constructed. The delayed response index table includes the transition warning sequence number, time span parameter, state change characteristics, risk prediction residual and trend offset index. The risk prediction residual is the difference between the actual risk level of the transition node and the historical prediction value. The trend offset index is calculated and fitted based on the difference in the abrupt change amplitude of the state change rate before and after.

[0139] This implementation method constructs a delayed response index table to achieve the association and error archiving of transition warning information in the time and state dimensions, thereby further enhancing the ability to trace the source of the warning trigger delay phenomenon.

[0140] In this embodiment, step six specifically includes:

[0141] For each time span parameter in the delayed response index table, a trajectory backtracking operation is performed to extract the risk level estimate sequence within the corresponding time period from the risk evolution trajectory sequence. The risk level estimate sequence represents the risk estimate values ​​at consecutive time points, which is used to characterize the risk evolution trend within that time period.

[0142] The risk level estimate sequence is subjected to difference calculation and smoothing. The difference calculation operation performs first-order difference on the risk estimates between adjacent time points to obtain the risk change rate sequence. The smoothing process adopts the weighted moving average method, which calculates the weighted mean of each time point based on a set smoothing window to generate a stable risk change trend vector.

[0143] The risk increase magnitude value within a time period is calculated based on the risk change trend vector. The risk increase magnitude value is the cumulative increment value of the trend vector, which is obtained by fitting all positive difference values ​​within the time period. The risk increase magnitude values ​​of all time periods are used to form a sequence to form the actual risk increase magnitude sequence.

[0144] Set a correction threshold range and compare the magnitude values ​​in the actual risk increase sequence item by item; the correction threshold is a percentile parameter obtained by fitting the distribution of historical error samples, which represents the sensitivity threshold for the system to trigger correction operations; if the magnitude value is greater than the correction threshold, mark the corresponding time period and locate the associated mutation point number, which is located based on the transition warning sequence number index in the delayed response index table;

[0145] The original risk level prediction value and confidence score value are extracted from the labeled mutation point numbers. The original risk level prediction value is the output result of the initial graph neural network. The correction coefficient calculation operation is performed, and the correction gain is calculated based on the product of the risk increase and the confidence score value. The correction gain is obtained by fitting a normalized product function to enhance the correction response to high-confidence mutation points.

[0146] The modified gain is added to the original risk level forecast to generate the modified risk level forecast; the modified risk level forecast represents an adjusted estimate that is closer to the actual risk trend and has higher accuracy.

[0147] The corrected warning information is output, which includes the mutation point number, time index, spatial location code, corrected risk level prediction value, risk increase value, and corrected gain weight.

[0148] This implementation method effectively enhances the system's adaptive adjustment capability to underestimated risk situations by constructing a sequence of actual risk increases and superimposing a correction gain mechanism, thereby further improving the sensitivity and response accuracy of forest fire risk prediction.

[0149] In this embodiment, step seven specifically includes:

[0150] For each mutation point number, the predicted risk level is calculated by subtracting the actual risk level data from the fire risk confirmation information. The difference is calculated by subtracting the actual value from the predicted value at each time point to generate a sequence of mutation point prediction deviation values. The prediction deviation value sequence is used to characterize the error magnitude between the model output and the actual observation.

[0151] The prediction deviation value sequence of mutation points is aggregated according to the time index to construct a historical cumulative error table. The historical cumulative error table records the sum of prediction deviations for each mutation point number in multiple consecutive time periods. The absolute error values ​​within the time period are superimposed using a weighted cumulative method, and timestamp information and spatial location codes are added.

[0152] A deviation calibration threshold is set, and the cumulative values ​​of each mutation point number in the historical cumulative error table are compared item by item. The deviation calibration threshold is a stable threshold range obtained by fitting the error convergence interval during the system training process, which is used to screen abnormal prediction points whose long-term prediction deviation exceeds the normal fluctuation range. A set of mutation point numbers with cumulative values ​​greater than the deviation calibration threshold is selected.

[0153] Extract the confidence cumulative calculation path corresponding to the mutation point number set. The confidence cumulative calculation path records the trigger frequency, weighted score, time weight and confidence score combination information of each mutation point within a time period; obtain the time weighted parameter and score weight record to form the original confidence cumulative parameter set;

[0154] The correction factor is calculated based on the cumulative error value and the scoring weight; the correction factor includes a time depreciation coefficient and a scoring smoothing coefficient; the time depreciation coefficient is used to penalize and adjust the original time weighting function through the cumulative error value, and the scoring smoothing coefficient is used to suppress the volatility of the score based on the error; the correction factor is obtained by fitting an exponential function and a smoothed normal function to form a set of error-driven weight adjustment factors.

[0155] The correction factor is applied to the original time-weighted parameters to update the weight distribution structure in the confidence accumulation calculation, generating a calibrated time-weighted parameter set; the calibrated time-weighted parameter set has a weighted configuration that is more sensitive to historical bias; the calibration operation is completed.

[0156] This implementation improves the dynamic adjustment capability of confidence accumulation calculation through a historical error-driven correction factor construction mechanism, further enhancing the long-term stability and error convergence performance of the model in forest fire risk prediction tasks.

[0157] Example 1:

[0158] To verify the feasibility of this invention in practice, it was applied to a smart forestry management project in a hilly forest area, with a monitoring range covering approximately 120 square kilometers of ecological protection forest. The forest type in this area is mostly mixed coniferous and broad-leaved forest with high vegetation coverage, and the dry season from April to June is historically a high-risk period for forest fires. In the early stages of traditional fire risk warnings, management departments mainly relied on manual patrols and simple meteorological indicators, which resulted in problems such as delayed warnings, high false alarm rates, and difficulties in tracking the spread of risk.

[0159] In this scenario, the system deployed a multi-source environmental data acquisition unit, including temperature and humidity sensors, anemometers, smoke concentration sensors, and vegetation humidity sensors. All devices were connected to the forest digital twin platform via wireless communication nodes. Based on the seven-step monitoring process proposed by this invention, the system analyzed environmental data using an improved T-GNN model, generated risk evolution trajectories, identified initial abrupt change points, and dynamically updated the fire risk level through a confidence cumulative scoring and response correction mechanism. It generated transition warning information and final corrected warning information, and performed error calibration. The entire experimental period lasted three months.

[0160] During actual operation, the system collected approximately 2.75 million environmental sample data points, initially identified 3,282 abrupt change points, generated 295 transition warning messages, and ultimately output 172 corrected warning messages. To verify the technical effectiveness, the system proposed in this invention was compared and analyzed with traditional warning mechanisms (represented by temperature and humidity level + smoke concentration threshold triggering alarm mechanisms), focusing on indicators such as warning timeliness, false alarm rate, risk level prediction deviation, and consistency with actual fire conditions. Statistical data are shown in Table 1.

[0161] Table 1. Comparison of early warning performance of the present invention system and traditional mechanisms over three months.

[0162]

[0163] As shown in Table 1, in terms of early warning response efficiency, the system of this invention, through its risk trajectory tracking and jump identification mechanism for abrupt change points, significantly shortens the response delay between risk perception and information output, reducing the average early warning response time to 2.8 hours. This allows for faster deployment of manual or technical inspection resources during dry seasons. Simultaneously, by filtering short-term abnormal fluctuations through a confidence accumulation mechanism, the number of invalid early warnings is significantly reduced, making the total number of early warnings more reasonable. After introducing a risk prediction correction module, the system significantly reduces the risk level judgment bias, with the average prediction error controlled within ±0.6 levels, demonstrating higher reliability compared to the ±2.1 levels of traditional mechanisms. The effective early warning hit rate has increased from 63.8% to 89.5%, and the false alarm rate has decreased by 15.7 percentage points, showing a clear technological advantage.

[0164] To further verify the stability and risk evolution understanding capabilities of this invention in actual fire monitoring, we statistically analyzed the detailed deviations between some actual fire events and the system's prediction records, comparing the differences in fire number, mutation point identification time, first warning time, corrected warning time, and actual fire time. The data are detailed in Table 2.

[0165] Table 2. Comparison of Typical Fire Events and System Prediction Records

[0166]

[0167] As can be seen from the analysis in Table 2, in all recorded typical fire events, the system of this invention was able to complete the corrected early warning output 2 to 3 hours before the fire occurred, and the average error between the corrected early warning time and the actual fire time was controlled within 2 hours. This effect is attributed to the composite intelligent prediction mechanism of "mutation point + risk trajectory tracking + confidence accumulation + risk correction + error calibration" adopted in this system. Even when there is some noise and bias in the actual mutation triggering, the system can track the real risk escalation process through trend analysis and a delayed response index table, effectively eliminating invalid fluctuations and improving the overall prediction accuracy and timeliness.

[0168] This embodiment fully demonstrates the system's ability to be implemented and its risk identification capabilities in large-scale forest scenarios. It possesses significant technical advantages in improving early warning accuracy, reducing false alarm rates, and detecting fire trends in advance. It can effectively replace traditional single-point meteorological threshold triggering mechanisms and is particularly suitable for application in ecological protection areas, state-owned forest farms, and hilly areas. In the future, combining it with data sources such as satellite remote sensing and high-altitude monitoring image recognition can further enhance the system's effectiveness, demonstrating good scalability and practical prospects.

[0169] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart forest fire risk monitoring system based on digital twins, characterized in that, include: The environmental data acquisition module is used to collect multi-source environmental status data in forest areas and construct environmental status datasets in time series format. The risk modeling module is used to input the environmental state dataset into the improved T-GNN model to generate a risk evolution trajectory sequence; The mutation identification module is used to perform fuzzy segmentation on the risk evolution trajectory sequence, divide the state into segments based on the sliding window and risk gradient changes, and construct an initial mutation point set. The early warning generation module is used to perform confidence accumulation calculation based on the initial set of mutation points, calculate the trigger frequency and weighted score value of mutation points in multiple consecutive time periods, and filter to generate a set of early warning trigger points; determine whether there are nodes in the set of early warning trigger points that meet the transition early warning conditions, and if so, generate transition early warning information; The transition warning information is recorded in time sequence, and a delayed response index table is constructed; The trajectory calibration module is used to evaluate the error between the prediction results of the mutation point and the fire risk confirmation information, and to calculate the historical cumulative error value. If the historical cumulative error value is greater than the preset deviation calibration threshold, the time weighting parameter in the confidence cumulative calculation is adjusted to complete the calibration operation.

2. A smart forest fire risk monitoring method based on digital twins, characterized in that, Includes the following steps: Step 1: Collect multi-source environmental status data in the forest area and construct an environmental status dataset in time series format; Step 2: Input the environmental state dataset into the improved T-GNN model to generate a risk evolution trajectory sequence; Step 3: Perform fuzzy segmentation on the risk evolution trajectory sequence, divide the state into segments based on the sliding window and risk gradient changes, and construct an initial set of mutation points; Step 4: Based on the initial set of mutation points, perform confidence accumulation calculation to calculate the trigger frequency and weighted score of mutation points in multiple consecutive time periods, and filter to generate a set of early warning trigger points; determine whether there are nodes in the set of early warning trigger points that meet the transition early warning conditions, and if so, generate transition early warning information; Step 5: Record the transition warning information in time sequence and construct a delayed response index table; Step Six: Perform a risk trajectory tracking operation on the delayed response index table, extract the risk value change trend within the corresponding time period of the index, and calculate the actual risk increase. If the actual risk increase is greater than the preset correction threshold, then the corresponding correction operation is performed to adjust the risk level prediction value of the corresponding mutation point and output the corrected warning information. Step 7: Evaluate the error between the predicted result of the mutation point and the fire risk confirmation information, and calculate the historical cumulative error value; if the historical cumulative error value is greater than the preset deviation calibration threshold, adjust the time weighting parameter in the confidence cumulative calculation to complete the calibration operation.

3. The intelligent forest fire risk monitoring method based on digital twins according to claim 2, characterized in that, Step one specifically involves: An environmental data acquisition unit is set up within the monitoring area; the environmental data acquisition unit includes a temperature acquisition sensor, a humidity acquisition sensor, a wind speed acquisition sensor, a smoke concentration acquisition sensor, and a vegetation humidity acquisition sensor; The environmental data acquisition unit synchronously collects environmental status information of the monitoring area according to a preset sampling period, generating raw environmental data; Perform data validation on the raw environmental data, remove outlier and missing value records, and complete data cleaning; The cleaned environmental data is serialized and arranged in chronological order to construct a time series data table containing sampling time index, spatial location information and environmental parameter values; Perform formatting and encoding operations on the time series data table to generate an environmental state dataset in time series format according to a unified data structure template; Perform block grouping operations on the environmental status dataset in time series format, and build a data grouping index based on forest partition identifiers and geographic coordinate information.

4. The intelligent forest fire risk monitoring method based on digital twins according to claim 3, characterized in that, Step two specifically involves: Perform modality label appending operations on the time-series formatted environmental state dataset to assign modality labels to data from different sources and generate labeled input tensors; Perform temporal and positional encoding on the labeled input tensor to construct temporal and spatial embedding vectors, generating a multidimensional embedded input structure; Based on the multidimensional embedded input structure, the state evolution features of each monitoring point in a continuous time period are extracted to generate a forest state object sequence; Perform graph structure construction operations on the forest state object sequence, construct physical adjacency matrix and state adjacency matrix based on spatial adjacency relationship and state similarity relationship respectively, and generate a dual-channel graph structure representation; The dual-channel graph structure representation is input into the graph convolution computation unit in the improved T-GNN model to calculate multidimensional attention weights, perform graph convolution operations, and extract state propagation features. Perform temporal modeling operations on the state propagation characteristics to generate a risk prediction vector sequence based on the temporal evolution path of the nodes in the graph; The risk prediction vector sequence is concatenated to form a risk evolution trajectory sequence; the risk evolution trajectory sequence includes the risk level estimate, state change trend value and confidence score value within a continuous time period.

5. The intelligent forest fire risk monitoring method based on digital twins according to claim 4, characterized in that, The improved T-GNN model includes an input building unit, a state object building unit, a dual-channel graph structure generation unit, a graph convolution calculation unit, a risk attention unit, and a temporal modeling unit, specifically: The input building unit receives multi-source data from the environmental state dataset, adds modal labels to each type of data, and generates a time embedding representation structure by encoding all data according to timestamps. The output is an input tensor containing the time embedding dimension and the modal label dimension. The state object construction unit receives the input tensor, extracts continuous evolution segments according to spatial location and time window, encapsulates time index, spatial coordinates and risk-related features for each segment, and generates a forest state object set. The dual-channel graph structure generation unit receives a set of forest state objects, generates a spatial adjacency matrix based on the physical distance relationship between spatial coordinates, generates a state adjacency matrix based on the similarity of evolutionary trends between state features, and weights and fuses the spatial adjacency matrix and the state adjacency matrix to construct a graph structure representation tensor containing node indexes, connection weights and connection labels. The graph convolutional computation unit receives the graph structure representation tensor, performs node feature aggregation and propagation operations according to the computation process of the graph neural network, and outputs multi-layer graph convolutional feature representation. The risk attention unit introduces a state sensitivity factor and a risk gradient change factor during graph convolution, respectively calculates the risk correlation score and propagation weight adjustment coefficient between nodes, adjusts the propagation intensity in the graph convolution path according to the risk correlation score, and generates a weighted graph feature representation with risk attention weight. The time-series modeling unit receives the weighted graph feature representation, performs sequence modeling operations according to the node time index, outputs a risk prediction vector sequence within a continuous time period, and concatenates the risk prediction vector sequence into a risk evolution trajectory sequence.

6. The intelligent forest fire risk monitoring method based on digital twins according to claim 5, characterized in that, Step three specifically involves: A fixed-length sliding time window is set for the risk evolution trajectory sequence, and the entire time period is traversed in chronological order to extract the risk level estimate sequence and the state change trend sequence within each window. The average gradient change value is calculated for the risk level estimate sequence within each sliding time window, the mutation sensitivity feature is extracted, and the risk fluctuation intensity index is calculated by combining the state change trend value sequence. The risk volatility intensity index is compared with the set mutation identification threshold. If the volatility intensity is greater than the mutation identification threshold, it is marked as a candidate mutation point at the end of the window. All candidate mutation points are merged in chronological order. Fuzzy clustering is performed by combining the continuity of risk gradient changes in adjacent time periods to remove duplicate labels and short-term abnormal fluctuation labels, and an initial mutation point set is output. Each mutation point in the initial mutation point set is labeled with a mutation direction label and a local risk increase value label.

7. The intelligent forest fire risk monitoring method based on digital twins according to claim 6, characterized in that, Step four specifically involves: The frequency of occurrence of each mutation point in the initial mutation point set within a continuous time period is statistically analyzed to construct a mutation trigger frequency sequence. The confidence-weighted score is calculated by combining the risk level increase value of each mutation point over a continuous time period, the consistency of mutation direction, and historical evolution background information. Set a mutation confidence screening threshold, filter mutation points with a trigger frequency higher than the set threshold and a weighted score higher than the screening threshold, and construct a set of early warning trigger points; Each mutation point in the set of early warning trigger points is arranged in chronological order and spatial location, and combined with the historical risk evolution trajectory, continuous surge patterns and clustered mutation patterns are identified; Determine if there is a mutation point that simultaneously satisfies the trigger frequency threshold condition, the weighted score threshold condition, and the continuous surge mode condition. If it exists, mark it as a transition warning trigger node. A transition warning information is generated for each transition warning trigger node; the transition warning information includes a mutation time index, spatial location code, confidence level score and risk improvement trend label.

8. The intelligent forest fire risk monitoring method based on digital twins according to claim 7, characterized in that, Step five specifically involves: The mutation time index and spatial location code in the transition warning information are arranged in chronological order to construct a transition warning sequence; a unique serial number is assigned to each transition warning information in the transition warning sequence, and the corresponding risk level score, risk improvement trend label and historical trigger frequency are recorded; Perform time period labeling operation on the transition warning sequence, map each transition warning information to a sampling time window, and label the window start time, end time and span value; The time difference between adjacent transition warning messages is calculated to generate a time interval vector. For each transition warning information, match the forest state object within the corresponding time period and extract the state change trend data within that time period from the risk evolution trajectory; By combining the time interval vector and state change trend data, a delayed response index table is constructed; the delayed response index table includes the transition warning sequence number, time span parameter, state change characteristics, risk prediction residual, and trend offset index.

9. A smart forest fire risk monitoring method based on digital twins according to claim 8, characterized in that, Step six specifically involves: Perform trajectory backtracking operation on the time period corresponding to each time span parameter in the delayed response index table, and extract the risk level estimate sequence within the corresponding time period from the risk evolution trajectory sequence; Perform difference calculation and smoothing on the risk level estimate sequence to construct a risk change trend vector; Calculate the risk increase magnitude within a time period based on the risk change trend vector to form an actual risk increase magnitude sequence; Set a correction threshold range, compare the magnitude values ​​in the actual risk increase sequence item by item, and if the magnitude value is greater than the correction threshold, mark the corresponding time period and locate the associated mutation point number. Extract the original risk level prediction value and confidence score value for the marked mutation point number, perform the correction coefficient calculation operation, and calculate the correction gain based on the product of the risk increase and the confidence score value; The modified gain is added to the original risk level prediction value to generate the modified risk level prediction value. The corrected warning information is output, which includes the mutation point number, time index, spatial location code, corrected risk level prediction value, risk increase magnitude and correction gain weight.

Images