Fire risk dynamic assessment and disposal method based on multi-source data

Abstract

The invention relates to the technical field of fire safety monitoring and early warning, and discloses a fire risk dynamic assessment and disposal method based on multi-source data. The method comprises the following steps: acquiring temperature, humidity, smoke concentration, wind speed, population density and building structure parameters with space-time marks in real time; cleaning and standardizing the multi-source monitoring data flow to generate a standard data sequence; dynamically constructing and updating a fire risk factor causal network based on data statistical association and a time sequence, and identifying a risk evolution mechanism; deducing a future risk state by using the network, generating a dynamic risk level map and triggering a corresponding disposal plan; and taking plan execution effect data as feedback, and optimizing paths and weights in the causal network in real time. According to the method, the assessment model is discovered from static rules to dynamic causal and is changed from open-loop early warning to closed-loop optimization, and the risk perception accuracy and the emergency response effectiveness are improved through continuous autonomous learning.

Description

Technical Field

[0001] This invention relates to the field of fire safety monitoring and early warning technology, specifically to a method for dynamic assessment and handling of fire risks based on multi-source data. Background Technology

[0002] Currently, fire risk early warning systems based on multi-source sensor data are widely used. These systems typically rely on pre-set physical models or static rule bases to determine risk levels by comparing real-time monitoring data with fixed thresholds. Another common approach is to use machine learning models, trained on historical data, to identify characteristic patterns associated with fire risk. These technological solutions form the foundation of existing fire early warning capabilities.

[0003] Existing technologies have limitations. Methods based on fixed rules or thresholds cannot adapt to the dynamic changes in the interaction between the environment and risk factors; their judgment logic is rigid and difficult to cope with complex and ever-changing real-world scenarios. While machine learning models that rely on the correlation of historical data can discover statistical patterns, they often struggle to distinguish between simple correlations and true causal relationships, which may lead to false alarms, misinterpreting coincidental changes in sensor readings not caused by fire as danger signals. Most existing systems operate in an open-loop mode, meaning a one-way process from data input to contingency plan output, lacking the evaluation and utilization of the actual effectiveness of response actions.

[0004] These shortcomings render existing assessment models static and outdated, unable to learn and self-correct from actual risk evolution and response practices. When faced with new building layouts, unknown risk coupling patterns, or unverified response strategies, the system's accuracy and adaptability significantly decline. Enabling risk assessment models to autonomously discover causal relationships from real-time data streams and continuously optimize based on response feedback is a key challenge in improving the intelligence level of fire early warning systems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamic assessment and handling of fire risks based on multi-source data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for dynamic fire risk assessment and response based on multi-source data, the method comprising: Acquire a monitoring data stream with spatiotemporal markers, the monitoring data stream including temperature readings, humidity percentage, smoke concentration values, wind speed values, real-time population density values, and building structure parameters; The monitoring data stream is cleaned and standardized in real time to remove outliers and align timestamps, generating a standardized multi-source spatiotemporal data sequence. The standardized multi-source spatiotemporal data sequence is input into a dynamic causal network construction process. This process gradually constructs and updates a dynamic causal network that reflects the causal relationship between fire risk factors based on the statistical correlation strength and temporal order between the data. Using the dynamic causal network, the current and future fire risk status is extrapolated, and a risk level map is generated; Based on the aforementioned risk level map, a strategy generation mechanism is activated to generate contingency plans for different risk levels. The execution effect data of the aforementioned contingency plan is used as a feedback signal and sent back to the dynamic causal network construction process to adjust the causal relationship paths and connection weights in the dynamic causal network.

[0007] Preferably, the dynamic causal network construction process specifically includes: Sliding window analysis is performed on the normalized multi-source spatiotemporal data sequence to calculate the correlation matrix between different data dimensions within the window; Based on the correlation pairs in the correlation matrix that exceed a preset threshold, and combined with the lead-lag relationship in the time series, a preliminary candidate causal relationship edge set is established; For each edge in the candidate causal relationship edge set, a Granger causality test is performed to confirm its statistical significance; The edges that pass the test are added to the dynamic causal network, and each edge is assigned an initial confidence weight. As new normalized multi-source spatiotemporal data sequences are continuously input, the sliding window analysis and Granger causality test are periodically re-executed to incrementally update the structure and edge confidence weights of the dynamic causal network.

[0008] Preferably, the simulation of current and future fire risk status includes: The normalized multi-source spatiotemporal data sequence at the current moment is used as the input node state of the dynamic causal network; Based on the causal relationship paths and confidence weights defined in the dynamic causal network, the propagation process of risk factor states in the network is simulated; through multiple rounds of state propagation iterations, the stable state values ​​of all risk factors are calculated, and the fire risk level map at the current moment is generated. Based on the trend extrapolation of historical data sequences, and combined with the dynamic causal network to simulate the state of risk factors at different future time points, a predictive fire risk level map for multiple future time points is generated.

[0009] Preferably, the strategy generation mechanism operates in the following manner: By analyzing the current fire risk level map and the predicted fire risk level map, spatial areas and time points where the risk exceeds the critical value can be identified. Match and identify basic handling action instructions corresponding to the risk area, risk level, and risk factor combination from the pre-set handling rule base; The matched basic handling action instructions are combined and arranged according to time sequence and logical dependencies to form a preliminary handling plan flow; The initial contingency plan flow is sent to the conflict detection and resolution module to check whether there are conflicts between different contingency plan actions in terms of resources, time, and space. The detected conflicts are coordinated and optimized, and finally an executable set of contingency plans is output.

[0010] Preferably, the processing of the feedback signal specifically includes: Monitor the actual execution of each action in the executable emergency response plan, and collect fire risk-related monitoring data for the corresponding areas after execution; The actual implementation effect data collected is compared and analyzed with the risk prediction data before the implementation of the plan, and the prediction deviation is calculated. Based on the magnitude and direction of the prediction deviation, the confidence weights of the relevant causal relationship edges in the dynamic causal network are adjusted; for edges that consistently exhibit large prediction deviations, a re-evaluation and correction of the dynamic causal network structure is triggered.

[0011] Preferably, the propagation process of the simulated risk factor state in the network includes: Map the current normalized multi-source spatiotemporal data sequence to the initial state values ​​of nodes in a dynamic causal network; Based on the direction and confidence weight of each causal relationship edge in the dynamic causal network, calculate the state increment of each node affected by its predecessor node; An iterative propagation algorithm is used to update the state values ​​of all nodes in each iteration. The update rule is based on a linear combination of the state increment and the current state. The convergence condition is set to the point that the change in the node state value is less than the specified tolerance. When the convergence condition is met, the iteration stops and the stable state values ​​of all risk factors are output. Spatial interpolation is performed using steady-state values ​​to generate a fire risk level map for the current moment.

[0012] Preferably, the generation of the fire risk level map further includes: The risk factor state values ​​calculated through the dynamic causal network are normalized. Map the normalized risk factor state values ​​to a predefined risk level range; By combining the spatial markers, the risk level information of discrete points is used to generate a continuous spatial risk distribution map through spatial interpolation technology; The spatial risk distribution map is overlaid with time dimension information to form a spatiotemporally integrated fire risk level map.

[0013] Preferably, the workflow of the conflict detection and resolution module includes: A resource constraint model is established, which defines the total amount, location, and deployment time limits of available fire-fighting resources; Map each action in the initial contingency plan flow to the resource constraint model and check for resource over-allocation or time window overlap conflicts. If a conflict is detected, the actions in the contingency plan will be sorted according to their priority scores, with priority given to ensuring the resource requirements of higher priority actions. For action sequences with time conflicts, a time offset or resource replacement strategy is used for rescheduling until all conflicts are resolved.

[0014] Preferably, the re-evaluation and modification of the dynamic causal network structure includes: Suspend the use of the current version of the dynamic causal network for risk simulation; Extract the complete normalized multi-source spatiotemporal data sequence and corresponding feedback signals within a historical period; The causal relationship network structure is relearned on the historical data using a causal discovery algorithm based on transfer entropy. The newly learned network structure is compared with the existing network structure, the consistent parts are merged, and new causal relationship edges are adopted or existing edges are deleted based on statistical significance. The predictive performance of the corrected network structure was verified using the latest data, and once the performance improvement was confirmed, it was deployed as the new dynamic causal network.

[0015] Preferably, the basic handling action instructions corresponding to the risk area, risk level, and risk factor combination matched and identified from the pre-set handling rule base include: Construct a rule base for handling situations, with each rule containing a combination of risk factor thresholds, spatial region type, and corresponding basic handling action instructions; Analyze the current fire risk level map and the predicted fire risk level map, and extract the risk factor numerical combination of spatial areas where the risk exceeds the critical value; The extracted risk factor values ​​are combined and matched one by one with the rules in the disposal rule base. The matching degree is calculated based on the Euclidean distance between the risk factor values ​​and the rule thresholds. Select the rule with the highest matching degree and extract its basic processing action instructions; If multiple rules have the same degree of matching, the basic handling action instruction is selected based on the predefined urgency level indicator in the rule.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By dynamically constructing and updating causal networks based on statistical correlations and temporal sequences between data streams, the foundation of risk assessment is shifted from static rules or shallow correlations to the exploration of dynamic causal mechanisms among risk factors. This method can identify temporally sequential causal chains, distinguish between the root causes and accompanying phenomena of risk, and thus gain a deeper understanding of the evolutionary mechanism of fire risk. It improves the accuracy and foresight of risk inference from the source, reducing false alarm rates caused by accidental correlations. The dynamic updating characteristic of the causal network enables the model to adapt to specific risk patterns under different environments and building structures, discover previously unknown risk paths, and enhance the system's robustness and interpretability in complex and ever-changing scenarios.

[0017] The execution effectiveness data of the emergency response plan is used as feedback signals to feed back into the causal network construction process, adjusting the paths and weights within the network to form a closed-loop learning and optimization system. This mechanism directly links the model's evolution with the effectiveness of actual actions, using practical results to verify and correct its understanding of causal relationships. The effect of each response action is transformed into knowledge for optimizing the model, driving the causal network to more accurately reflect the operating rules of the real world. This feedback-based online optimization ensures that the risk assessment model is no longer static but can continuously learn from successful or unsuccessful experiences, allowing its decision-making intelligence and emergency response plan generation capabilities to evolve over time, gradually approaching the optimal emergency response strategy. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the fire risk dynamic assessment and response method based on multi-source data described in this invention. Figure 2 A flowchart for simulating fire risk status; Figure 3 A flowchart for feedback signal processing; Figure 4 A graph showing the results of the strategy generation mechanism. Figure 5 The graph shows the results of the simulation of risk factor propagation. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a method for dynamic fire risk assessment and response based on multi-source data. The method includes: acquiring monitoring data streams with spatiotemporal markers, including temperature readings, humidity percentages, smoke concentration values, wind speed values, real-time population density values, and building structure parameters; performing real-time cleaning and standardization on these data streams, removing outliers and aligning timestamps to generate a standardized multi-source spatiotemporal data sequence; inputting this sequence into a dynamic causal network construction process, and gradually constructing and updating a dynamic causal network reflecting the causal relationships between fire risk factors based on the statistical correlation strength and temporal order of the data; using this network to extrapolate the current and future fire risk status and generate a risk level map; initiating a strategy generation mechanism based on the risk level map to produce response plans for different risk levels; and transmitting the execution effect data of the response plans as feedback signals back to the dynamic causal network construction process to adjust the causal relationship paths and connection weights in the network.

[0021] Example 1: See Figure 2 The causal network construction process, as a core component of dynamic fire risk assessment, strictly adheres to standardized multi-source spatiotemporal data sequences. The process begins with sliding window analysis of the standardized multi-source spatiotemporal data sequences. Sliding window analysis uses windows with fixed time spans to segment the data stream. The window size is set based on the data sampling frequency and business requirements; for example, a 30-minute window is used to balance real-time performance and stability. In practice, each window contains data points across multiple dimensions, including temperature readings, humidity percentages, smoke concentration, wind speed, real-time population density, and building structural parameters. The system calculates the correlation matrix between different data dimensions within the window. The correlation calculation is based on either the Pearson correlation coefficient or mutual information methods. The Pearson correlation coefficient is used to capture linear relationships, while mutual information methods are suitable for nonlinear relationships, ensuring a comprehensive assessment of the dependencies between variables. After the correlation matrix is ​​generated, the system scans all elements in the matrix, identifying correlation pairs exceeding a preset threshold. This threshold is calibrated using historical fire event data to filter out weak or accidental correlations, retaining only significantly relevant relationships for subsequent analysis.

[0022] In some embodiments, based on the correlation pairs exceeding a preset threshold in the correlation matrix and combined with the lead-lag relationship in the time series, the system initially establishes a candidate causal relationship edge set. Lead-lag relationship analysis is implemented through a cross-correlation function, calculating the lag correlation between the time series of different variables to determine which variable's change leads another variable in time, thereby inferring the causal direction. For example, if the change in temperature readings leads the change in smoke concentration values, and the correlation exceeds the threshold, a candidate causal relationship edge is generated from the temperature node to the smoke node. The candidate causal relationship edge set is temporarily stored in a graph structure, containing the source node, target node, and preliminary correlation strength information of the edge. In a specific implementation, the system performs a Granger causality test on each edge in the candidate causal relationship edge set to confirm its statistical significance. The Granger causality test is based on a vector autoregression model, testing whether the historical value of one variable can significantly predict the current value of another variable. The significance level is set to 0.05; only edges with p-values ​​below this level are considered statistically significant. The test process uses a subset of data within a sliding window to ensure robust results.

[0023] It is understandable that causal relationship edges that pass the Granger causality test are added to the dynamic causal network, and each edge is assigned an initial confidence weight. The initial confidence weight is calculated based on the p-value or correlation value of the test, for example, using 1 minus the p-value as the weight basis, and then normalized to the range of 0 to 1 to reflect the reliability of the causal relationship. The dynamic causal network is maintained in the form of a directed graph, where nodes represent risk factors, edges represent causal relationships, and weights are dynamically adjusted. In specific implementation, as new normalized multi-source spatiotemporal data sequences are continuously input into the system, the dynamic causal network construction process periodically re-executes the sliding window analysis and Granger causality test. The update cycle can be configured according to the data flow rate, such as every 10 minutes or triggered when the amount of new data accumulates to a certain scale. The update process is performed incrementally; the system only processes the data from the latest sliding window, recalculates the correlation matrix and Granger causality test, thereby adjusting the structure of the dynamic causal network and the confidence weights of the edges. For example, if new data indicates that the causal relationship of a certain edge has weakened, its weight will be adjusted accordingly; if a new significant relationship is found, a new edge is added. The incremental update mechanism uses version control and retains the network's historical state to support rollback and auditing.

[0024] In some embodiments, the implementation of sliding window analysis involves the optimized selection of window size and sliding step size. The window size is typically related to the temporal characteristics of the data; for example, in fire risk scenarios, a window covering the typical fire development cycle, such as 30 minutes to 1 hour, is selected. The sliding step size is set as a fraction of the window size, such as 50% overlap, to ensure continuity and capture rapid changes. The correlation matrix is ​​calculated using efficient algorithms, such as parallel processing based on a distributed computing framework, to handle high-dimensional data streams. For lead-lag relationship analysis, the system uses the lag time corresponding to the maximum cross-correlation value as the causal direction, and the lag time range is set according to domain knowledge to avoid overfitting.

[0025] Optionally, the initial assignment of confidence weights can consider various factors, such as historical consistency or expert knowledge, but primarily relies on statistical indicators. During periodic updates, the system introduces a decay factor to handle the impact of older data; for example, using an exponential decay model to give higher weight to recent data, enhancing the network's responsiveness to dynamic changes. The overall implementation of the dynamic causal network construction process is embedded in the stream processing platform, supporting high throughput and low latency. It can be understood that the output of this process is a continuously evolving dynamic causal network, providing a foundation for fire risk projection. Changes in the network structure are logged for monitoring and debugging. All computational modules undergo unit testing and integration verification to ensure reliability in a real-world environment. Data stream processing follows fault-tolerant mechanisms, such as checkpoint settings, to address system failures.

[0026] Example 2: See Figure 3In its implementation, the process of extrapolating current and future fire risk states begins by using a standardized multi-source spatiotemporal data sequence at the current moment as the input node state of a dynamic causal network. This standardized multi-source spatiotemporal data sequence includes dimensions such as temperature readings, humidity percentages, smoke concentration, wind speed, real-time population density, and building structural parameters. Each data dimension corresponds to a specific node in the dynamic causal network. The node state values ​​are mapped to a predefined numerical range using a linear transformation method; for example, temperature readings are converted from a Celsius scale to a normalized value between 0 and 1 to ensure comparability of data with different dimensions. After mapping, each node in the dynamic causal network is assigned an initial state value, representing the instantaneous level of each risk factor at the current moment. The system then simulates the propagation process of risk factor states in the network based on the causal relationship paths and confidence weights defined in the dynamic causal network. The propagation process is based on directed graph theory, where nodes represent risk factors, directed edges represent causal relationships, the direction of the edges is determined by the causal leadership relationship, and the weights are assigned by the confidence values. The simulation process uses an iterative algorithm for state propagation. In each iteration, the state value of each node is updated based on the state of its predecessor node and the connection weight. The update rule aims to reflect the cumulative effect of causal influence until the network reaches a stable state.

[0027] In some embodiments, the propagation of the simulated risk factor state in the network is specifically implemented through an iterative propagation algorithm. The algorithm sets a maximum number of iterations to avoid infinite loops and defines a convergence condition where the change in node state values ​​is less than a specified tolerance, for example, a tolerance threshold of 0.001. The core of the iterative propagation algorithm involves a state update formula, which is used to calculate the new state value of each node in each iteration. The formula is expressed as follows:

[0028] in: This represents the state value of node i in the t-th iteration. This represents the state value of node i in the (t+1)th iteration. Let represent the set of all predecessor nodes of node i, that is, the set of nodes j that have causal edges pointing from node j to node i. This represents the confidence weight of the causal edge from node j to node i. The retention coefficient represents the current state value and is used to control the inertia of the node's own state. This represents the scaling factor for the influence of preceding nodes, used to adjust the strength of causal propagation. (Coefficient) and Calibrate settings through pre-experiments or historical data, for example, by taking... and To ensure the stability and convergence of the iterative process, the system computes the state updates of all nodes in parallel during each iteration, optimizes performance using matrix operations, and continues iterating until the changes in the state values ​​of all nodes are less than the tolerance threshold. At this point, the stable state values ​​of all risk factors are output as the deduction result. It can be understood that the stable state values ​​represent the equilibrium level of risk factors under the interaction of the causal network.

[0029] Optionally, the process of generating the fire risk level map for the current moment includes normalizing the stable state values ​​of risk factors calculated through a dynamic causal network. The normalization process uses a minimum-maximum scaling method to linearly transform the state values ​​to a unified range of 0 to 1, for example, by subtracting the minimum value from each state value and then dividing by the range. The normalized risk factor state values ​​are then mapped to predefined risk level intervals, which are based on historical fire statistics. For example, low risk intervals are set as [0, 0.3], medium risk intervals as (0.3, 0.6], high risk intervals as (0.6, 0.9], and extremely high risk intervals as (0.9, 1.0]. The mapping process is achieved by comparing the normalized values ​​with interval thresholds. Combining the spatial markers in the standardized multi-source spatiotemporal data sequence, the system generates a continuous spatial risk distribution map from the risk level information of discrete monitoring points using spatial interpolation technology. The spatial interpolation technology uses an inverse distance weighting method, which weights the risk values ​​of neighboring points based on geographical distance to generate a high-resolution raster map. Time dimension information is then overlaid onto the spatial risk distribution map. This process creates a spatiotemporally integrated fire risk level map. Time information is embedded in each grid cell as a timestamp, and the map output is in a spatiotemporal data cube format, supporting multidimensional queries and visualization. In some embodiments, trend extrapolation is used to generate predictive fire risk level maps for multiple future time points. Trend extrapolation is based on historically normalized multi-source spatiotemporal data sequences, using an autoregressive integral moving average model to predict the future values ​​of each risk factor. The prediction time interval is set to a fixed step size, such as 5 minutes, 10 minutes, and 15 minutes in the future. After the predicted values ​​are input into a dynamic causal network, the state propagation iteration process is repeated to calculate the stable state values ​​of the risk factors at future time points. The same normalization, level mapping, and spatial interpolation steps are then applied to generate the predictive fire risk level map.

[0030] Optionally, the specific implementation of spatial interpolation technology involves setting interpolation parameters, such as the power parameter and search radius in inverse distance weighting. The power parameter is usually set to 2, and the search radius is adaptively adjusted according to the monitoring point density to ensure interpolation accuracy. The generation of the spatiotemporal integrated fire risk level map also includes data compression and index optimization to support efficient storage and real-time access. The system uses a hierarchical grid structure to manage map data, with different resolution levels corresponding to different levels of detail. In specific implementation, the selection of the trend extrapolation model can be adjusted according to data characteristics. For example, for risk factors with strong nonlinearity, a long short-term memory network can be used instead of an autoregressive integral moving average model. Model training uses historical data for cross-validation to avoid overfitting. The results of future risk projections are integrated with current projection results, displaying dynamic changes in risk through time-series animations to assist decision-makers in understanding trends. The entire projection process is deployed on a distributed computing platform, utilizing parallel processing capabilities to handle large-scale spatiotemporal data, ensuring projection efficiency. It can be understood that the fire risk level map generation process is modularly designed, with each step independently verifiable and optimized. The normalization module supports multiple scaling methods to adapt to data with different distribution characteristics. The risk level range definition supports dynamic updates, adjusting thresholds based on real-time feedback to improve the adaptability of the map. The spatial interpolation module integrates a geographic information system library to ensure the geographic accuracy of the interpolation results.

[0031] Example 3: In specific implementation, the strategy generation mechanism begins by analyzing the current fire risk level map and the predictive fire risk level map. The map data is input in the form of a spatiotemporal grid. The system identifies spatial areas and time points where the risk exceeds the threshold by scanning the risk value of each grid cell and comparing it with a preset threshold. The threshold is dynamically set based on factors such as building type and historical accident data. For example, a lower smoke concentration threshold is set for commercial areas, while a higher threshold is allowed for industrial areas. The identification process uses a connected component analysis algorithm in image processing to merge adjacent high-risk grids into continuous risk areas and mark the spatial boundary, peak risk level, and risk factor combination for each area. Time point information is directly extracted from the timestamp attribute of the map. After identification, the system matches basic disposal action instructions from a preset disposal rule base. The disposal rule base is stored in a relational database table structure. Each record contains a rule number, applicable risk factor threshold combination, target spatial area type, expected risk level range, and corresponding basic disposal action instruction text. The matching process requires comparing the actual risk factor values ​​of the identified risk areas with the threshold combinations of each rule in the rule base to calculate the degree of matching between the actual values ​​and the rule thresholds.

[0032] In some embodiments, the matching degree is calculated based on the Euclidean distance formula between the risk factor value and the rule threshold, which measures the degree of difference between the actual risk situation and the preset conditions of the rule. The matching degree Euclidean distance formula is expressed as follows:

[0033] in: The Euclidean distance represents the degree of matching; the smaller the value, the higher the degree of matching. This represents the total number of risk factors involved in the matching. It is an index of risk factors. This represents the actual value of the k-th risk factor extracted from the risk level map. This represents the threshold set by a rule in the rule base for the k-th risk factor. The system iterates through all rules in the rule base, calculating the Euclidean distance between each rule and the current risk situation. The rule with the smallest Euclidean distance is selected as the best match, and its associated basic action command is extracted. The basic action command is an atomic operation command, such as "turn on the sprinkler system in area A" or "send inspection command B2 to the patrol personnel's PDA". If multiple rules have the same calculated Euclidean distance, the system performs a secondary selection based on the predefined urgency indicator in the rule record. The urgency indicator is usually an integer priority code, with lower values ​​indicating higher urgency. The system selects the basic action command corresponding to the rule with the smallest urgency indicator value.

[0034] Understandably, the matched basic emergency response instructions are combined and arranged according to chronological order and logical dependencies to form a preliminary emergency response plan flow. The arrangement process uses a directed acyclic graph (DAG) model to represent the order and dependencies between actions. Nodes represent basic emergency response actions, and directed edges represent the sequential relationship or triggering conditions between actions. For example, the action "cut off power" must precede the action "start emergency lighting." The preliminary emergency response plan flow is then sent to a conflict detection and resolution module. This module establishes a resource constraint model, which defines the total amount, spatial location, and time constraints required for deployment of available fire-fighting resources. Resource types include fire trucks, rescue personnel, and fire-fighting equipment. The module maps each action in the preliminary emergency response plan flow to the resource constraint model, checking for resource over-allocation or overlapping time windows. Resource over-allocation checks are performed by comparing the total demand for the same type of resource at the same time point with the total available resources; overlapping time window checks are performed by comparing whether the time intervals for different actions using the same resource overlap. If a conflict is detected, the module sorts the actions according to their priority scores. The priority score is calculated based on factors such as the risk level of the area where the action is located and the expected effect coefficient of the action. Priority is given to ensuring the resource needs of high-priority actions. For action sequences with time conflicts, a time offset strategy or a resource replacement strategy is used for rescheduling. The time offset strategy adjusts the start time of low-priority actions to avoid conflict, while the resource replacement strategy allocates similar alternative resources to conflicting actions. This process continues until all conflicts are resolved, and finally, an executable set of contingency plans is output.

[0035] Optionally, the processing of feedback signals specifically involves monitoring the actual execution of various actions within the executable emergency response plan. The system collects data such as the start time, completion status, and resource consumption of actions through an IoT execution status interface, and simultaneously collects fire risk-related monitoring data for the corresponding area after the plan's execution, such as deploying a sensor network to remeasure indicators like temperature and smoke concentration. The collected actual execution effect data is compared and analyzed with risk prediction data based on a dynamic causal network before the plan's execution, calculating the prediction deviation. The prediction deviation is calculated for each affected grid cell and major risk factor, using, for example, mean absolute deviation or root mean square deviation as quantitative indicators. Based on the magnitude and direction of the prediction deviation, the confidence weights of relevant causal relationship edges in the dynamic causal network are adjusted. The adjustment strategy is: if the prediction deviation is consistently positive, the weight of the relevant edge is appropriately reduced; if the prediction deviation is consistently negative, the weight is appropriately increased. Weight adjustment employs an incremental learning approach, such as using gradient descent for fine-tuning. For causal relationship edges that consistently exhibit significant prediction deviations as detected through feedback monitoring, the system triggers a reassessment and correction process for the dynamic causal network structure. This reassessment process may involve suspending the current network usage and initiating a causal discovery process based on historical data, as described in Example 5. In some embodiments, the workflow of the conflict detection and resolution module can be further refined. The resource constraint model is maintained in the form of a two-dimensional table, with one dimension being the resource list and the other a timeline. Conflicts are detected by simulating the resource allocation process. It is understood that the feedback signal acquisition and processing cycle and the dynamic causal network update cycle can be synchronized, for example, both set to a 15-minute cycle, thereby forming a closed-loop assessment-disposal-optimization loop. This ensures that the system can continuously learn from actual execution results, improving the accuracy of dynamic fire risk assessment and disposal.

[0036] See Figure 4 This diagram demonstrates the complete process of dynamic fire risk assessment and response strategy generation. The diagram clearly shows the risk level distribution of different spatial areas through color gradients, presented sequentially from low to high risk. Identified high-risk areas are clearly marked with red borders, and each high-risk area is automatically matched with a corresponding response strategy based on its risk characteristics and scale. Blue arrows in the diagram indicate the specific response measures recommended for each risk area, including personnel deployment, equipment activation, and emergency procedures. This visualization method effectively supports command personnel in quickly understanding the risk distribution pattern and formulating corresponding response plans, demonstrating the practical application effect of intelligent response strategy generation.

[0037] Example 4: In specific implementation, the simulation of the propagation process of risk factor states in the network begins by mapping the current time-of-flight normalized multi-source spatiotemporal data sequence to the initial state values ​​of nodes in the dynamic causal network. The mapping process uses a linear scaling function to transform the original values ​​of each monitoring data point to a uniform numerical range. For example, temperature readings are mapped from 20-100 degrees Celsius to the range of 0-1, and humidity percentages are normalized by dividing by 100 from 0-100%. Each risk factor corresponds to a node in the dynamic causal network, and the initial state value of the node is directly assigned by the mapped data, forming the initial state vector of the network. Based on the direction and confidence weight of each causal relationship edge in the dynamic causal network, the state increment of each node affected by its predecessor nodes is calculated. The state increment reflects the amount of node state change caused by the causal relationship in a single propagation step. The calculation process traverses each node in the network. For each node, all incoming edges pointing to it are found, i.e., the set of its predecessor nodes. Then, a weighted sum is performed based on the current state values ​​of these predecessor nodes and the confidence weights of the connecting edges.

[0038] In some embodiments, an iterative propagation algorithm is used to update the node state values. In each iteration, the algorithm updates the state values ​​of all nodes, and the update rule is based on a linear combination of the state increment and the current state. The state update formula is expressed as follows:

[0039] in: This represents the state value of node v in the iterth iteration. This represents the state value of node v in the (iter+1)th iteration. This represents the state increment of node v calculated at the iter-th iteration. The retention factor represents the current state value of a node, used to maintain the continuity of the state. The absorption coefficient represents the state increment and is used to control the rate at which new influences are incorporated. and Must meet The conditions are usually set as follows: , To ensure the stability of the iteration, the system computes the state updates of all nodes in parallel during the iteration process, utilizing a graph computing framework to optimize performance. The convergence condition is set so that the change in node state values ​​is less than a specified tolerance, with a tolerance threshold of 0.0001. When the maximum difference between the state values ​​of all nodes in two consecutive iterations is less than this tolerance, the network is considered to have reached a stable state, and the iteration stops. At this point, the stable state values ​​of all risk factors output represent the equilibrium risk level under the influence of the current causal network. Spatial interpolation is performed using the stable state values ​​to generate a fire risk level map for the current moment. The Kriging interpolation method is used, which can consider spatial autocorrelation and generate a smooth, continuous risk surface.

[0040] The conflict detection and resolution module's workflow includes establishing a resource constraint model, which defines the total amount, spatial location, and deployment time limits of available fire-fighting resources in a structured manner. The core of the model is a resource inventory and a timeline planning table. The resource inventory details the quantity, current location status, and operational characteristics of each type of resource, while the timeline planning table defines the availability status of resources at different points in time. Each action in the initial emergency response plan flow is mapped to the resource constraint model to check for resource over-allocation or overlapping time windows. The mapping process parses each plan action into a request for the use of a specific type of resource within a specific time interval. The system maintains a resource allocation status table to track the reservation status of all resources. See Table 1, Resource Allocation Status Table.

[0041] Table 1: Resource Allocation Status Table Resource ID Resource types Current location Scheduled start time Scheduled end time Occupy Contingency Plan Action ID state R001 ladder fire truck Coordinates (X1, Y1) 2024-08-2014:00 2024-08-2015:30 A101 Allocated R002 Water tanker fire truck Coordinates (X2, Y2) 2024-08-2014:20 2024-08-2015:00 A103 Allocated R003 Rescue command vehicle Coordinates (X3, Y3) 2024-08-2013:50 2024-08-2016:10 A105 Allocated R004 ladder fire truck Coordinates (X4, Y4) 2024-08-2014:00 2024-08-2014:45 A102 Allocated Understandably, conflict detection is performed by querying the resource allocation status table. For new contingency plans, the system checks whether the requested resource type has been fully occupied within a predetermined time interval, and whether the same resource has overlapping requests from different actions. If a conflict is detected, the contingency plans are sorted according to their priority scores. The priority score is calculated based on factors such as the risk level, scope of impact, and urgency of the action. The calculation function comprehensively considers parameters such as risk level, affected population, and asset value. Higher-priority actions have priority in resource allocation. For action sequences with time conflicts, a time offset strategy or a resource replacement strategy is used for rescheduling. The time offset strategy adjusts the start time of the action, for example, delaying the execution of low-priority actions to avoid peak resource usage periods. The resource replacement strategy finds similar available resources for conflicting actions, for example, dispatching the next closest backup vehicle when the nearest fire truck is occupied. The conflict resolution process may require multiple iterations until the resource requirements of all actions can be met under constraints, ultimately outputting a set of conflict-free, executable contingency plans. In practical implementation, the resource constraint model supports dynamic updates. When new resources are added or the status of existing resources changes, the model can adjust the resource allocation plan in real time. It can be understood that the entire conflict detection and resolution process is implemented using a constraint programming algorithm, which can find a feasible scheduling solution within a reasonable timeframe. Optionally, for particularly complex resource conflict scenarios, the system can introduce a manual review process, where command personnel confirm or fine-tune the automatically generated resolution solutions.

[0042] See Figure 5 This paper presents the simulation results of risk factor propagation based on dynamic causal networks. A heatmap is used to illustrate the spatial risk distribution after multiple rounds of iterative propagation to a stable state. The contour lines clearly depict the continuous changing trend of risk levels, and white lines mark the boundaries of different risk levels. High-risk core areas are specifically marked, showing the central location of risk aggregation. This simulation method considers the causal relationship network between various risk factors and achieves dynamic extrapolation of risk impact through a state propagation algorithm. The results show that risk propagation has significant spatial correlation, with high-risk areas often forming contiguous clusters. This provides important decision-making basis for preventative resource deployment and early intervention. The propagation model based on causal networks can more accurately reflect the development pattern of real fire risks, improving the predictability and accuracy of risk assessment.

[0043] Example 5: In specific implementation, the reassessment and correction of the dynamic causal network structure is triggered when the system detects a persistent and significant prediction deviation. The process begins by suspending the use of the current version of the dynamic causal network for fire risk simulation. The system temporarily switches the risk assessment task to a standby mode, for example, using a simple assessment method based on fixed thresholds to ensure the continuity of system functionality. After suspending the current version of the dynamic causal network, the system extracts a complete and standardized multi-source spatiotemporal data sequence and corresponding feedback signals from the historical database. The length of the historical period is set according to the data frequency and business needs; for example, data from the most recent 30 days is extracted. The data sequence includes time series of temperature readings, humidity percentages, smoke concentration values, wind speed values, real-time population density values, and building structural parameters. The feedback signals include data on the implementation effect of emergency response plans collected during the same period, such as records of actual fire occurrences or risk mitigation status. The extracted data needs to be verified for completeness, and missing or invalid records are removed to ensure the data quality for reassessment.

[0044] In some embodiments, a causal discovery algorithm based on transfer entropy is used to relearn the causal network structure on historical data. Transfer entropy is an information-theoretic causal measure that can capture nonlinear causal relationships between variables and is not limited by data distribution assumptions. The core calculation formula of the transfer entropy causal discovery algorithm is as follows:

[0045] in: This represents the transfer entropy value from variable X to variable Y, used to quantify the strength of the causal influence of X on Y. This represents the state of variable Y at time t+1. This represents the k-th order historical state vector of variable Y at time t. This represents the l-th order historical state vector of variable X at time t. This represents the probability distribution function, estimated using frequency statistics from historical data. Parameters and Let Y and X represent the embedding dimensions, respectively, typically determined through trial and error or information criteria. The algorithm calculates the transition entropy values ​​between all variable pairs, classifying variable pairs with transition entropy values ​​exceeding a significance threshold as having a causal relationship. This significance threshold is determined using a permutation test. The newly learned network structure is compared with the existing network structure based on graph edit distance or the difference in the node adjacency matrix. Consistent parts are merged, and new causal edges are adopted or existing edges are deleted based on statistical significance. The merging process employs a conservative strategy, adopting only new edges that are significant in multiple bootstrap samplings, while deleting existing edges requires their transition entropy values ​​to remain below the threshold.

[0046] Understandably, the latest data is used to verify the predictive performance of the corrected network structure. The verification process employs the hold-out method, using recent data as the test set to calculate the prediction accuracy of the corrected network for future risk states. For example, root mean square error or mean absolute percentage error is used as a performance indicator. After confirming the performance improvement, the system deploys the corrected network structure as a new dynamic causal network and reactivates the risk inference function based on the causal network. Matching basic response action instructions from a pre-set response rule base involves constructing the response rule base, which is implemented in the form of database tables. Each rule includes a rule ID, rule name, applicable risk factor threshold combination, target spatial area type identifier, rule priority weight, and corresponding basic response action instruction code. The system parses the current fire risk level map and the predictive fire risk level map, extracting risk factor numerical combinations for spatial areas where the risk exceeds the critical value. The extraction process uses a spatial query language to traverse the high-value grids in the risk map, reading the normalized temperature, smoke concentration, and other factor values ​​within each grid. The extracted risk factor values ​​are matched against each rule in the rule base. The matching degree is calculated based on the Euclidean distance between the risk factor values ​​and the rule thresholds. The Euclidean distance is calculated as the square root of the sum of the squares of the differences between each factor value and its corresponding threshold. The rule with the highest matching degree, i.e., the rule with the smallest Euclidean distance, is selected, and its basic handling action instruction is extracted. If multiple rules have the same matching degree, i.e., the calculated Euclidean distances are equal, the basic handling action instruction is selected according to the predefined urgency level identifier in the rule. The urgency level identifier is usually an integer code; the smaller the value, the higher the urgency level. The system selects the instruction corresponding to the rule with the smallest urgency level identifier value.

[0047] Optionally, the rule base for handling issues supports a dynamic update mechanism, allowing administrators to add, modify, or disable rules based on new fire safety regulations or incident experience. Rule updates must be managed through version control. The execution efficiency of the matching process is optimized by establishing a rule index, such as creating a multi-dimensional index for risk factor threshold combinations to accelerate similarity searches. In some embodiments, the implementation of the transfer entropy causal discovery algorithm employs an efficient computing framework, such as using parallel computing to accelerate the probability distribution estimation process, with historical data being divided into time blocks for distributed processing. It is understood that the re-evaluation and correction process of the dynamic causal network structure has an upper limit on execution frequency, such as a maximum of once per month, to avoid system instability caused by overly frequent network structure changes. During rule matching, for boundary cases, such as when the risk factor value happens to fall within the threshold boundaries of multiple rules, the system introduces fuzzy matching logic, considering the weight coefficients of each factor for comprehensive evaluation.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fire risk dynamic assessment and disposal method based on multi-source data, characterized in that, The method comprises the following steps: acquiring a monitoring data stream with a space-time label, the monitoring data stream containing temperature readings, humidity percentages, smoke concentration values, wind speed values, real-time population density values, and building structure parameters; performing real-time cleaning and standardization processing on the monitoring data stream, eliminating abnormal values and aligning timestamps to generate a standardized multi-source space-time data sequence; inputting the standardized multi-source space-time data sequence into a dynamic causal network construction process, which gradually constructs and updates a dynamic causal network reflecting the causal relationships among fire risk factors based on the statistical correlation strength and time sequence among the data; using the dynamic causal network to deduce the current and future fire risk states and generate a risk level atlas; based on the risk level atlas, starting a strategy generation mechanism to output disposal plans for different risk levels; using the execution effect data of the disposal plans as feedback signals to return to the dynamic causal network construction process to adjust the causal relationship paths and connection weights in the dynamic causal network.

2. The method of claim 1, wherein, The dynamic causal network construction process specifically comprises: performing sliding window analysis on the standardized multi-source space-time data sequence to calculate the correlation matrix among different data dimensions within the window; preliminarily establishing a candidate causal relationship edge set according to the correlation pairs in the correlation matrix that exceed a preset threshold and combining the lead-lag relationship on the time sequence; performing Granger causality test on each edge in the candidate causal relationship edge set to confirm its statistical significance; adding the causal relationship edges that pass the test to the dynamic causal network and assigning an initial confidence weight to each edge; periodically re-executing the sliding window analysis and Granger causality test to incrementally update the structure of the dynamic causal network and the confidence weights of the edges as new standardized multi-source space-time data sequences are continuously input.

3. The method of claim 2, wherein, The deduction of the current and future fire risk states comprises: using the current time's standardized multi-source space-time data sequence as the input node state of the dynamic causal network; simulating the propagation process of the risk factor states in the network according to the causal relationship paths and confidence weights defined in the dynamic causal network; through multiple rounds of state propagation iteration, calculating the stable state values of all risk factors to generate the fire risk level atlas at the current time; based on the trend extrapolation of historical data sequences, simulating the risk factor states at different future time points using the dynamic causal network to generate predictive fire risk level atlases at multiple future time points.

4. The method of claim 3, wherein, The strategy generation mechanism operates in the following manner: analyzing the current time's fire risk level atlas and the predictive fire risk level atlases to identify spatial regions and time points with risk exceeding a critical value; matching the identified risk regions, risk levels, and risk factor combinations with corresponding basic disposal action instructions from a preset disposal rule library; combining and arranging the matched basic disposal action instructions according to the time sequence and logical dependency relationship to form a preliminary disposal plan flow; The preliminary treatment plan is streamed into a conflict detection and resolution module, which checks whether there is a conflict between different plan actions in terms of resources, time and space, and coordinates and optimizes the detected conflicts, and finally outputs an executable treatment plan set.

5. The method of claim 4, wherein, The processing of the feedback signal is specifically: The actual execution of each action in the executable treatment plan set is monitored, and fire risk related monitoring data of the corresponding area after execution is collected; The actual execution effect data collected is compared and analyzed with the risk prediction data before plan execution, and a prediction deviation amount is calculated; According to the size and direction of the prediction deviation amount, the confidence weight of the related causal relationship edge in the dynamic causal network is adjusted; for the edge with a large prediction deviation that continues to appear, the reevaluation and correction of the structure of the dynamic causal network are triggered.

6. The method of claim 1, wherein, The propagation process of the simulated risk factor state in the network includes: The normalized multi-source spatio-temporal data sequence at the current time is mapped to the initial state value of the node in the dynamic causal network; According to the direction and confidence weight of each causal relationship edge in the dynamic causal network, the state increment of each node affected by its preceding node is calculated; An iterative propagation algorithm is used to update the state value of all nodes in each iteration, and the update rule is based on the linear combination of the state increment and the current state; The convergence condition is set as the change amplitude of the node state value being less than a specified tolerance, and the iteration is stopped when the convergence condition is met, and the stable state value of all risk factors is output; The stable state value is used for spatial interpolation to generate a fire risk level map at the current time.

7. The method of claim 3, wherein, The generation of the fire risk level map also includes: The risk factor state value calculated by the dynamic causal network is normalized; The normalized risk factor state value is mapped to a predefined risk level interval; The risk level information of the discrete points is generated into a continuous spatial risk distribution map through spatial interpolation technology in combination with the spatial marker; The time dimension information is superimposed on the spatial risk distribution map to form a spatio-temporal integrated fire risk level map. 8.The method of claim 4, wherein, The workflow of the conflict detection and resolution module includes: A resource constraint model is established, which defines the total amount, location and deployment time limit of available fire resources; Each action in the preliminary treatment plan stream is mapped into the resource constraint model to check whether there is a resource over-allocation or time window overlap conflict; If a conflict is detected, the actions are sorted according to the priority score of the plan actions to ensure the resource demand of high priority actions first; For action sequences with time conflicts, time offset or resource replacement strategies are used for rescheduling until all conflicts are resolved. 9.The method of claim 5, wherein, The reevaluation and correction of the structure of the dynamic causal network includes: The current version of the dynamic causal network is suspended for risk inference; The complete normalized multi-source spatio-temporal data sequence and the corresponding feedback signal in a historical period are extracted; A causal relationship network structure is relearned on the historical data using a transfer entropy based causal discovery algorithm; Differences between the newly learned network structure and the existing network structure are compared, consistent parts are merged, and new causal relationship edges or existing edges are deleted based on statistical significance; The prediction performance of the modified network structure is verified using the latest data, and the performance is improved to deploy it as a new dynamic causal network.

10. The method of dynamic assessment and treatment of fire risk based on multi-source data according to claim 4, characterized in that, The matching and identifying of the basic treatment action instruction corresponding to the combination of the identified risk area, risk level and risk factor from the preset treatment rule library includes: A treatment rule library is constructed, each rule contains a risk factor threshold combination, a spatial region type and a corresponding basic treatment action instruction; The fire risk level atlas at the current time and the predictive fire risk level atlas are analyzed, and the risk factor value combination of the spatial region whose risk exceeds the critical value is extracted; The extracted risk factor value combination is matched with each rule in the treatment rule library, and the matching degree is calculated based on the Euclidean distance between the risk factor value and the rule threshold; The rule with the highest matching degree is selected, and its basic treatment action instruction is extracted; If the matching degrees of multiple rules are the same, the basic treatment action instruction is selected according to the pre-defined emergency level identifier in the rule.

Images