A Method and Device for Park Hazard Awareness Based on Multi-Source Data Fusion

Abstract

This invention provides a method and device for park hazard situation perception based on multi-source data fusion, belonging to the field of park hazard perception technology. This invention grids the park; when the temperature in a sub-area exceeds the standard, it is marked as an initial risk area. Temperature diffusion condition judgment is introduced, and by analyzing the temperature gradient relationship between the initial risk area and adjacent sub-areas, temperature anomalies caused by non-fire factors are effectively filtered out, achieving false alarm filtering. After confirming the temperature diffusion trend, time-series data of temperature, smoke, and CO concentration are acquired within the observation time window, and the dynamic correlation of their change rates is analyzed. A dynamic fire risk index is calculated through a multi-parameter direct evaluation function. Based on the comparison between the dynamic fire risk index and multi-level risk thresholds, graded early warning information is output, enabling managers to take differentiated response measures according to different hazard levels, optimizing resource allocation and emergency response efficiency.

Description

Technical Field

[0001] This invention relates to the field of park hazard perception technology, specifically a park hazard situation perception method and device based on multi-source data fusion. Background Technology

[0002] In recent years, with the increasing scale and complexity of large-scale industrial parks, warehousing and logistics bases, the density of internal hazardous sources has been continuously increasing, leading to a sustained rise in the risk of safety accidents such as fires and leaks. Traditional industrial parks often rely on decentralized, isolated single-point sensor threshold alarms. This type of method can only reflect local and instantaneous parameter exceedances, making it difficult to identify, comprehensively assess, and dynamically warn of hazardous situations from the perspective of spatial linkage and temporal evolution. Especially when facing risks with concealed or rapidly evolving characteristics such as smoldering, flashover, and leakage diffusion, single threshold alarm mechanisms often suffer from frequent false alarms, delayed response, or inaccurate warnings, failing to meet the urgent needs of modern industrial parks for real-time, accurate, and intelligent safety perception.

[0003] In the prior art, CN116384255A discloses a method and system for perceiving the dangerous situation of a park based on multi-source data fusion. It constructs a virtual model of the target park through the building environment and historical data, and pre-calculates fixed hazard indicators and indicator weights through simulation and artificial intelligence training. In practical applications, the real-world map is uniformly gridded, and the measured data is weighted and calculated with the aforementioned generated indicator weights to output the hazard situation value of each grid. By combining the hazard situation values ​​of all grids, the hazard situation data of the target park is obtained.

[0004] The aforementioned existing technologies have significant inherent defects in responding to sudden and evolving dangerous situations such as fires: It relies heavily on pre-established virtual models and cannot adapt to real-time environmental changes. When the real environment deviates from the simulated environment, the weights may become invalid, leading to inaccurate evaluations. When the layout of the park, equipment, etc., changes, the entire model may need to be rebuilt and retrained, resulting in poor system adaptability and delayed updates.

[0005] While existing technologies mention multi-source data fusion, they are mainly based on pre-trained risk indicators and fixed weights using virtual models. They cannot be adjusted according to the dynamic relationships between real-time data and lack in-depth exploration of the dynamic correlations between real-time data.

[0006] The final output is the danger situation data of the target park, without mentioning how to determine the risk level based on the data, when to raise an alarm, to whom to raise the alarm, or what emergency measures to take. This is an open result that requires further human interpretation, and in an emergency, it may miss the best time to deal with the situation. Summary of the Invention

[0007] The purpose of this invention is to provide a method and device for campus hazard situation awareness based on multi-source data fusion, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for assessing potential hazards in a park based on multi-source data fusion, comprising the following steps: The park area to be sensed is divided into several sub-areas by gridding. Temperature data of each sub-area is obtained. When the temperature of any sub-area exceeds the preset first safety threshold, the sub-area is marked as the initial risk area, and real-time temperature data of all its adjacent sub-areas are obtained simultaneously. Centered on the initial risk area, and combining the temperature data of adjacent sub-areas and their distances from the initial risk area, the temperature data of each adjacent sub-area is judged to meet the temperature diffusion conditions based on the changes in the temperature data and corresponding distances of each adjacent sub-area. In response to meeting the temperature diffusion conditions, time-series data of temperature, smoke concentration, and CO gas concentration in the initial risk area are acquired within a set observation time window and labeled as a risk dataset. Based on the risk dataset, the changing patterns of temperature, smoke concentration and CO gas concentration are identified according to the time series data change rates. The dynamic correlation between the changing patterns of temperature, smoke concentration and CO gas concentration is analyzed, and the dynamic fire risk index is calculated by inputting a multi-parameter direct evaluation function. Based on the dynamic fire risk index of the initial risk area, the danger level of the initial risk area is determined according to the preset multi-level risk thresholds, and graded early warning information is output.

[0009] Furthermore, the logic for dividing the area into several sub-regions using a grid includes: Determine the center of the park area, determine the size of the square grid for the initial division, construct a central grid whose center coincides with the center of the park area, expand multiple grids outward along the central grid to ensure that all grids cover the park area, and use the grid division scheme at this time as the initial grid division scheme when all grids cover the park area. The grids that are not fully filled within the park area are merged into one grid with adjacent grids that are fully filled within the park area. Adjacent grids are those that share edges or vertices. The merging logic includes: Based on the centers of the grids not filled by the park area and the centers of the adjacent grids that are completely filled by the park area, the distance between their centers is determined. The grids not filled by the park area are merged into the nearest adjacent grid that is completely filled by the park area. If the distances are the same, then any adjacent grid that is completely filled by the park area is randomly selected for merging. The final result of grid division is that each grid is a sub-region.

[0010] Furthermore, the adjacent sub-regions of the initial risk region include directly adjacent sub-regions and indirectly adjacent sub-regions, wherein the directly adjacent sub-regions are all sub-regions that share edges or vertices with the initial risk region; The indirectly adjacent sub-regions are all sub-regions that share edges or vertices with the directly adjacent sub-regions.

[0011] Furthermore, determining whether the temperature data of adjacent sub-regions meet the temperature diffusion conditions specifically includes: Obtain the distance between the center of the initial risk region and the center of each adjacent sub-region, obtain the current temperature data and inherent temperature data of each adjacent sub-region, and calculate the temperature gradient of the adjacent sub-regions. The method for calculating the temperature gradient is as follows: The difference between the current temperature data of an adjacent sub-region and the inherent temperature data of the adjacent sub-region is the first temperature difference. The ratio of the first temperature difference to the distance between the center of the initial risk area and the center of the adjacent sub-region is represented as the temperature gradient of the adjacent sub-region.

[0012] Calculate the temperature gradient of all adjacent sub-regions, and then... Perform normalization and map to Interval, normalized temperature gradient Calculated using the following formula: in The normalized temperature gradient of the i-th neighboring sub-region is... This represents the maximum temperature gradient across all adjacent sub-regions. This represents the minimum temperature gradient across all adjacent sub-regions. When the temperature gradient of adjacent sub-regions meets the preset conditions and the proportion of adjacent sub-regions that meet the conditions exceeds the preset percentage, the initial risk area meets the temperature diffusion conditions.

[0013] Furthermore, the logic for setting the observation time window includes: when the initial risk area meets the temperature diffusion conditions, setting the observation time window with that moment as the end point, and backtracking by a preset fixed duration to determine the start and end times of the observation time window; Within the observation time window, time-series data of temperature, smoke concentration, and CO gas concentration in the initial risk area are acquired at a fixed sampling frequency.

[0014] Furthermore, the logic for generating the time-series data change rate includes: The fixed duration of the observation time window is denoted as The fixed sampling frequency is denoted as Therefore, the sampling interval is The total number of sampling points within the window is ; According to the formula: Determined at The rate of change of internal temperature over time was used to obtain a result of length [missing information]. Time series data change rate sequence: ; According to the formula: Repeat the above calculations to obtain the time-series data on the rate of change of smoke concentration and CO gas concentration. , Ultimately, three sets of time-series data change rate sequences of equal length and time synchronization were obtained.

[0015] Furthermore, the dynamic correlation between the changing patterns of temperature, smoke concentration, and CO gas concentration was analyzed, specifically including: Time-series data on temperature, smoke concentration, and CO gas concentration are obtained from the risk dataset, with the lag step size denoted as... According to the formula: Determine the cross-correlation coefficient between time-series data of temperature and time-series data of smoke concentration. ,in and These are the time-series averages of temperature and smoke concentration, respectively. Find the maximum value in the sequence, denoted as . Simultaneously record the lag step size corresponding to the maximum value. As a leading feature in the temperature-smoke time series; According to the formula: Repeat the above calculations, where This is the average of the time-series data for CO gas concentration. The cross-correlation coefficient between the time-series data for temperature and the time-series data for CO gas concentration is obtained. Maximum cross-correlation number The lag step size corresponding to the maximum value As a leading characteristic of temperature-CO sequence; Based on the rate of change of time-series data for temperature, smoke concentration, and CO gas concentration, a (n-1)*3 data matrix is ​​constructed by treating the three sets of time-series data rate of change sequences as column vectors. : According to the formula: For the matrix To centralize, among which, The elements are the mean of the time series change rates of temperature, smoke concentration, and CO gas concentration. According to the formula: Calculate the covariance matrix ; For covariance matrix Perform eigenvalue decomposition and solve the characteristic equation: Three eigenvalues ​​were obtained. , , Sort by size from largest to smallest: and the corresponding feature vectors , , The largest eigenvalue This is a characteristic of co-principal components.

[0016] Furthermore, the multi-parameter direct evaluation function is: This indicates a dynamic fire risk index. arrive This indicates the preset weight coefficient value.

[0017] Furthermore, the logic for determining the hazard level of the initial risk area and outputting tiered early warning information is as follows: Set threshold , , ,in The risks are divided into four levels: high risk, low risk, medium risk and high risk. like It is classified as a "watch" level item; if It is classified as low-risk; if It is classified as medium risk; if It was classified as high-risk.

[0018] This invention also provides a park hazard situation awareness device based on multi-source data fusion. The device is used to implement the above-mentioned park hazard situation awareness method based on multi-source data fusion, comprising: Grid division module: used to divide the park area to be sensed into grids to form several sub-areas, acquire temperature data of each sub-area, and when the temperature of any sub-area exceeds the preset first safety threshold, the sub-area is marked as the initial risk area, and real-time temperature data of all its adjacent sub-areas are acquired simultaneously. Diffusion judgment module: It is used to determine whether the temperature data of adjacent sub-regions meet the temperature diffusion conditions based on the temperature data of each adjacent sub-region and the distance between the adjacent sub-regions and the initial risk area, taking the initial risk area as the center and combining the temperature data of the adjacent sub-regions and the corresponding distance. Data acquisition module: In response to meeting the temperature diffusion conditions, it collects time-series data of temperature, smoke concentration and CO gas concentration in the initial risk area within a set observation time window and labels it as a risk dataset; Risk assessment module: Based on the risk dataset, it identifies the changing patterns of temperature, smoke concentration and CO gas concentration according to the time series data change rate, and analyzes the dynamic correlation between the changing patterns of temperature, smoke concentration and CO gas concentration. The multi-dimensional feature vector is input into the multi-parameter direct evaluation function to calculate the dynamic fire risk index. Hazard warning module: Based on the dynamic fire risk index of the initial risk area, and according to the preset multi-level risk thresholds, it determines the hazard level of the initial risk area and outputs graded warning information.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention divides the park into grids. When the temperature in a certain sub-area exceeds the standard, it is marked as an initial risk area. Once a local temperature anomaly is detected, the subsequent process is immediately initiated to achieve rapid response and capture weak signals in the early stages of a fire. When the temperature in a certain sub-area exceeds the standard, an alarm is not immediately triggered. Instead, a temperature diffusion condition judgment is introduced. By analyzing the temperature gradient relationship between the initial risk area and adjacent sub-areas, isolated temperature anomalies caused by non-fire factors such as local overheating of equipment and direct sunlight are effectively filtered out, thus achieving false alarm filtering. After confirming the temperature diffusion trend, this invention acquires time-series data of temperature, smoke, and CO concentration within the observation time window, analyzes the dynamic correlation of the rate of change of the three, and calculates the dynamic fire risk index through a multi-parameter direct evaluation function. This multi-source data fusion method significantly improves the accuracy of risk assessment and reduces false alarms caused by environmental interference. This invention outputs graded early warning information based on the comparison between a dynamic fire risk index and multi-level risk thresholds. This allows managers to take differentiated response measures according to different hazard levels, thereby optimizing resource allocation and emergency response efficiency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a structural block diagram of the overall device of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0022] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0023] Example: Please see Figure 1 The present invention provides a technical solution: A method for assessing potential hazards in a park based on multi-source data fusion, comprising the following steps: Step 1: Divide the park area to be sensed into a grid to form several sub-areas, obtain the temperature data of each sub-area, and when the temperature of any sub-area exceeds the preset first safety threshold, mark the sub-area as the initial risk area, and simultaneously obtain the real-time temperature data of all its adjacent sub-areas. The initial partitioning logic, which expands outward from the center, is universal and can adapt to any shape of the park area. This ensures that the final generated grid can cover every corner of the park. By merging grids that are not filled by the park area into the nearest neighboring grid that is completely filled by the park area, it effectively avoids the generation of a large number of meaningless small fragmented grids in irregular park areas. This ensures that every physical location in the park belongs to one and only one sub-region. The regular grid makes it very efficient and deterministic to calculate the geographical location of any grid, its distance from other grids, and to find its neighboring grids.

[0024] In this embodiment, the logic for dividing the area into several sub-regions by gridding includes: Determine the center of the park area, determine the size of the square grid for the initial division, construct a central grid whose center coincides with the center of the park area, expand multiple grids outward along the central grid to ensure that all grids cover the park area, and use the grid division scheme at this time as the initial grid division scheme when all grids cover the park area. The grids that are not fully filled within the park area are merged into one grid with adjacent grids that are fully filled within the park area. Adjacent grids are those that share edges or vertices. The merging logic includes: Based on the centers of the grids not filled by the park area and the centers of the adjacent grids that are completely filled by the park area, the distance between their centers is determined. The grids not filled by the park area are merged into the nearest adjacent grid that is completely filled by the park area. If the distances are the same, then any adjacent grid that is completely filled by the park area is randomly selected for merging. The final result of grid division is that each grid is a sub-region.

[0025] The shared edge is defined as a boundary line segment with a length greater than zero that completely overlaps with two grids. The shared vertex is defined as a geometric vertex with identical coordinates that exists between two grids but does not share an edge. The adjacent grids are all grids that share an edge or a vertex with the target grid, and are uniformly defined as adjacent grids of the target grid.

[0026] Static, fixed safety thresholds have fundamental flaws in complex real-world park environments. For example, they lack environmental adaptability, failing to distinguish between summer heat and winter cold. In summer, high temperatures cause frequent triggering of fixed thresholds, leading to widespread false alarms. In winter, low temperatures prevent fixed thresholds from promptly capturing abnormal temperature rises, resulting in missed alarms. They also cannot differentiate between peak production periods and nighttime shutdowns. During periods of high-intensity operation, normal fluctuations are inherently large, and fixed thresholds may be too lenient, missing early potential hazards. During quiet periods, normal fluctuations are small, and fixed thresholds may be too sensitive, leading to false alarms. To address these issues, this invention proposes a dynamic preset method for a first safety threshold, comprising the following steps: Divide the day into 24 independent time slices, for example For time slice 0, For time slice 9, for each sub-area within the park, a historical dataset is created for each time slice, with a historical learning period set, for example, N=30 days. Temperature data for that sub-area within the same time slice over the most recent N days are collected, and the average temperature within that time slice is calculated to form an average temperature series. Based on this sequence, the historical statistical features specific to this time slice are calculated: According to the formula: It is the historical average temperature. It is the average temperature of the sub-region within that time slice on the j-th learning day.

[0027] According to the formula: It is the historical temperature standard deviation, which quantifies the inherent normal fluctuation range of temperature in this sub-region during a specific time period, such as 9 a.m. every day.

[0028] To predict the normal temperature level that this sub-region should have under current environmental conditions in real time, a multiple linear regression model is established. Data from all time slices of this sub-region during the learning period, totaling n = 24*N samples, are used to train the model for this sub-region. Each sample contains: the average temperature of the sub-region at that time slice. The ambient temperature at that time point Time slice .

[0029] The model is in the following form: For the predicted dynamic temperature baseline, It is the regression coefficient.

[0030] coefficient This characterizes the combined sensitivity and hysteresis effect of the sub-region's temperature response to changes in ambient temperature. When the ambient temperature changes, the regional temperature does not change immediately and synchronously; for example, under sunlight, indoor temperatures rise more slowly than outdoor temperatures. When the ambient temperature changes by 1 degree, it means that the temperature in this sub-region is expected to change by 0.9 degrees, indicating poor insulation and almost following environmental changes; when If the temperature is low, it indicates that the area's temperature is not sensitive to environmental changes, which may be due to good building insulation.

[0031] coefficient Overall, it describes the average trend of temperature change over time throughout the day, such as warming in the morning due to increased sunshine and activity, and gradually cooling in the afternoon and evening. Introducing it as a continuous variable enables the model to learn and predict this smooth, periodic trend, thereby distinguishing normal standards at different times, which is the key to achieving refined perception of the time dimension.

[0032] constant term This represents the baseline temperature level of the area when the ambient temperature is 0°C, without considering the effect of time, such as the temperature rise caused by geothermal energy, waste heat from equipment, or building insulation. It's about capturing these temperature shifts.

[0033] Choosing multiple linear regression can balance complexity and practicality. Although temperature changes in the park are affected by many factors, linear relationships can sufficiently approximate the actual situation in local areas and short time scales. Nonlinear models, such as neural networks, are more flexible, but require a large amount of data and are computationally complex.

[0034] To establish a model capable of predicting the normal temperature baseline of a sub-region in real time, the multiple linear regression model needs to be trained. The core of this training is to determine the regression coefficients in the model. .

[0035] Using the least squares method, according to the formula: The coefficient matrix is ​​what we are looking for. , It is The matrix, where each row corresponds to one sample. It is A column vector, where each element is the average temperature of the subregion over a time slice.

[0036] These coefficients were obtained through multiple linear regression fitting and are applicable to all time slices in this region. The goal of the least squares method is to optimize the model's predicted values. Compared with the actual observed value The overall sum of squared errors is minimized, thus obtaining the statistically optimal fit coefficient.

[0037] Within a limited learning period, the number of effective historical samples available for each sub-region's time slice is limited. Under these conditions, the least squares method can provide stable and statistically significant fitting results, effectively avoiding the overfitting problem that is prone to occur when using complex algorithms on small datasets, and ensuring the long-term stability of the model in actual operation.

[0038] The first security threshold ultimately used for real-time comparison It is not simply determined by the forecast baseline, but rather synthesized through an adaptive formula that intelligently adjusts the tightness of the warning based on the inherent volatility of the current period. The calculation formula is as follows: The first safety threshold, It is the historical standard deviation corresponding to the current time slice. The real-time dynamic baseline at the current moment is calculated using the following formula: For real-time ambient temperature, This represents the current hour. The real-time dynamic baseline is the most likely normal temperature value for the region under current conditions, predicted by a regression model. This allows the threshold to automatically fluctuate according to weather, season, and diurnal patterns, overcoming the shortcomings of fixed thresholds that frequently result in false alarms in summer and are prone to missed alarms in winter.

[0039] This is an adaptive coefficient responsible for automatically adjusting in reverse based on the inherent fluctuation level of the temperature data in the current time slice of the current sub-region. The inherent fluctuation level is the temperature fluctuation of the sub-region in the current time slice due to normal operation, and its value is determined through the following steps: Calculate the historical standard deviation of all time slices for this sub-region. The median, used as the benchmark volatility level for the region. Set the base coefficient This coefficient is based on a preset baseline value according to the risk level of the sub-region. For example, it is set in the production area. Non-production area setting Based on the historical standard deviation corresponding to the current time slice Calculate the real-time value according to the following formula value: When the temperature fluctuates drastically in the current time slice, that is when When it is large, the calculated The value decreases, leading to Closer This is equivalent to tightening early warning standards and improving sensitivity to address the challenge of weak risk signals being easily drowned out in complex contexts, aiming to reduce missed detections; when the current time slice is stable, i.e. when When smaller, The larger the value, the wider the warning boundary is intended to reduce false alarms.

[0040] parameter It is a constant preset based on the inherent risk level of the target sub-region. For example, in a key monitoring area, it is set to... In general monitoring areas, set The parameters To mitigate residual risks and errors that cannot be perfectly predicted using historical data or fully absorbed by dynamic models, including minor deviations in model predictions, inherent errors in the sensor system, and combinations of extreme operating conditions never before encountered, if the system relies entirely on dynamic adjustment, i.e., only parameters... In the most unfavorable boundary conditions, the overall warning threshold may be too low, resulting in a lack of sufficient early warning lead time when real risks occur, thus endangering the reliability of the system.

[0041] The acquisition of temperature data for each sub-region refers to acquiring the real-time highest temperature of each sub-region at short intervals, such as 10 seconds. Real-time ambient temperature The system time is then used; the time slice is determined based on the current date and time, and the parameters of that sub-region in this time slice are read from the pre-stored model. , , and regression coefficients Based on the above formula, calculate the current situation in real time. ;Compare and ,like If so, immediately execute step 1 to mark the sub-region as the initial risk region and initiate the subsequent multi-source data fusion and situation assessment process.

[0042] Step 2: Taking the initial risk area as the center, and combining the temperature data of adjacent sub-areas and the distance between them, determine whether the temperature data of adjacent sub-areas meet the temperature diffusion conditions based on the changes in the temperature data and corresponding distances of each adjacent sub-area. The adjacent sub-regions of the initial risk region include directly adjacent sub-regions and indirectly adjacent sub-regions. The directly adjacent sub-regions are all sub-regions that share edges or vertices with the initial risk region. The indirectly adjacent sub-regions are all sub-regions that share edges or vertices with the directly adjacent sub-regions.

[0043] The spatial topological definition of shared edges and shared vertices is exactly the same as the definition criteria for grid cells in step 1, that is: if two sub-regions have a completely overlapping boundary line segment, then the boundary line segment is determined to be a shared edge; if two sub-regions do not have a shared edge, but have a geometric vertex with completely identical coordinates, then the geometric vertex is determined to be a shared vertex.

[0044] Determining whether the temperature data of adjacent sub-regions meet the temperature diffusion conditions specifically includes: Obtain the distance between the center of the initial risk region and the center of each adjacent sub-region, obtain the current temperature data and inherent temperature data of each adjacent sub-region, and calculate the temperature gradient of the adjacent sub-regions. The method for calculating the temperature gradient is as follows: The difference between the current temperature data of an adjacent sub-region and the inherent temperature data of the adjacent sub-region is the first temperature difference. The ratio of the first temperature difference to the distance between the center of the initial risk area and the center of the adjacent sub-region is represented as the temperature gradient of the adjacent sub-region.

[0045] The inherent temperature data of the adjacent sub-regions serve as the real-time dynamic baseline of the adjacent sub-regions. Calculate the temperature gradient of all adjacent sub-regions, and then... Perform normalization and map to Interval, normalized temperature gradient Calculated using the following formula: in The normalized temperature gradient of the i-th neighboring sub-region is... This represents the maximum temperature gradient across all adjacent sub-regions. This represents the minimum temperature gradient across all adjacent sub-regions. When the temperature gradient of adjacent sub-regions meets the preset conditions and the proportion of adjacent sub-regions that meet the conditions exceeds the preset percentage, the initial risk area meets the temperature diffusion conditions.

[0046] A temperature gradient represents the temperature change per unit distance. During heat diffusion, the closer to the heat source, the more significant the temperature rise; therefore, the gradient value reflects the intensity of heat propagation. Using the inherent temperature, rather than simply comparing the current temperature with the initial risk area temperature, eliminates the influence of inherent temperature differences between different areas, focusing instead on the abnormal temperature rise caused by the initial risk area. This design allows the system to distinguish between isolated localized high-temperature points and potential spreading risks, thereby ensuring timely early warnings while reducing the probability of false triggers for subsequent in-depth analysis and optimizing system resource allocation. Since gradient values ​​in different directions can vary significantly, directly setting an absolute threshold to determine whether diffusion has occurred may be inaccurate. Normalization maps the gradients in each direction to... The interval makes the gradient values ​​in each direction comparable, which facilitates the setting of unified conditions. Normalization makes the judgment conditions adaptable to different scenarios, such as different seasons and different park layouts, because it is based on the relative performance of the current gradient in each direction, rather than absolute values.

[0047] When the normalized temperature gradient exceeds the normalized temperature gradient threshold In order to meet the preset conditions, The value ranges from 0.6 to 0.8, with 0.7 being the preferred value. The value reflects the stringency of the response to abnormal temperature rises. Below 0.6, the system will classify many moderate temperature fluctuations, such as those caused by changes in sunlight, air conditioning start-up and shutdown, and normal equipment heat dissipation, as abnormal. While this can improve monitoring sensitivity, it will drastically increase the risk of false alarms, leading to unnecessary consumption of subsequent in-depth monitoring resources and potentially causing warning fatigue. 0.6 is a balance point, ensuring that only temperature increases exceeding normal environmental fluctuations are detected; when... At values ​​above 0.8, only a few areas directly adjacent to the fire source and experiencing the most intense heating may meet the criteria. While this nearly eliminates false alarms, it also delays the identification of early-stage diffusion signals. In the early stages of a fire, the diffusion of heat to more distant areas may be attenuated, leading to excessively high values. This will filter out these important early signals, leading to delayed warnings. By... Setting it within this range effectively filters out the majority of isolated, non-diffusion temperature fluctuations, while ensuring sufficient capture capability for potential, destructive initial signals of thermal diffusion. The preferred value of 0.7 achieves an optimal balance between sensitivity and robustness, making it suitable for most general scenarios.

[0048] When the proportion of adjacent sub-regions that meet the conditions exceeds the spatial anomaly proportion threshold When this condition is met, the initial risk area meets the temperature diffusion condition. The value ranges from 50% to 70%, with 60% being the preferred value. This requirement dictates that adjacent sub-regions of anomalies must exhibit a certain degree of spatial prevalence, meaning they must account for a proportion exceeding [a certain percentage]. From a spatial statistical perspective, if the proportion of anomalous areas is less than 50%, it means that more than half of the observed directions did not detect a significant temperature rise. This pattern is more consistent with a localized heat source than with true heat diffusion. Because real fire heat diffusion is affected by building structure, wind direction, and obstacles, it is not necessarily absolutely uniform and may be blocked or weakened in some directions. Setting the upper limit to 70%, rather than higher figures like 90% or 100%, reflects a reasonable tolerance for the complexity of the park. It allows for weaker diffusion signals in a few directions, such as the upwind direction or the other side of the insulation wall, but as long as significant anomalies are observed in most directions, temperature diffusion can be considered valid, thus avoiding underreporting caused by overly stringent spatial uniformity requirements. By... By setting this range, it is possible to reliably distinguish between "anomaly at a single point" and "spread across a region," fundamentally eliminating the possibility of isolated local heat sources triggering the entire monitoring and assessment process, and greatly improving the accuracy of the risk triggering mechanism.

[0049] This invention creatively combines the two to produce a synergistic and multiplying technical effect: through dual filtering, it greatly suppresses false alarms caused by occasional sensor failures, local equipment overheating, and brief environmental interference, and realizes intelligent resource scheduling; it accurately captures the actual initial heat diffusion of a fire.

[0050] Step 3: In response to meeting the temperature diffusion conditions, acquire time-series data of temperature, smoke concentration and CO gas concentration in the initial risk area within the set observation time window, and label it as a risk dataset; The logic for setting the observation time window includes: when the initial risk area meets the temperature diffusion conditions, setting the observation time window with that moment as the end point, and backtracking by a preset fixed duration to determine the start and end times of the observation time window.

[0051] In actual deployment, to ensure real-time response, the temperature, smoke concentration, and CO gas concentration data of the initial risk area are continuously measured by default.

[0052] The purpose of setting the observation time window is to obtain time-series data on temperature, smoke concentration, and CO gas concentration in the initial risk area over a continuous period before the triggering moment. Its direct function is to form a risk dataset for subsequent in-depth analysis. If only instantaneous data at the triggering moment is used, the system can only perform simple threshold comparisons and cannot identify patterns of change or analyze dynamic correlations. This continuous time-series data is the sole data foundation supporting all subsequent advanced analyses.

[0053] The approach of looking back is intended to obtain data that is a record of the process immediately before the spatial spread is identified. This data directly describes how the current event developed from quantitative change to qualitative change. This approach can reflect the essential characteristics of the risk more directly and purely than analyzing data from an arbitrary time period or data from the trigger point into the future.

[0054] Within the observation time window, time-series data of temperature, smoke concentration, and CO gas concentration in the initial risk area are acquired at a fixed sampling frequency.

[0055] In this embodiment, the preset fixed duration and fixed sampling frequency are set to 180 seconds and 0.2Hz, respectively. This combination achieves an optimal balance between capturing the complete early evolution trend and controlling the data scale for real-time analysis, ensuring that subsequent time-series analysis is both reliable and efficient. Those skilled in the art will understand that the above-mentioned preferred values ​​can be adjusted according to the actual monitoring scenario. A reasonable preset range for the window duration is suggested to be 120 to 300 seconds. The lower limit of the range is based on the fact that if the duration is too short, there will be insufficient sequence data points, making it impossible to reliably calculate the rate of change and statistical correlation, leading to unstable feature extraction. The upper limit of the range is based on the fact that if the duration is too long, too much early stable data with low correlation to the current risk will be included, diluting key features and increasing unnecessary processing delays and storage burdens. The recommended reasonable range for sampling frequency is 0.2Hz to 1Hz. The lower limit is based on the fact that temperature, smoke, and CO concentration are typical slowly varying signals, and their highest effective frequency components are usually below 0.1Hz. According to the Nyquist sampling theorem, the sampling frequency must be greater than twice the highest frequency of the signal to reproduce the signal without distortion. Therefore, choosing a sampling frequency greater than or equal to 0.2Hz is a technically necessary condition to meet information integrity. Below this value, it may be impossible to reliably capture the dynamic changes of parameters. The upper limit is based on the fact that, under the premise of meeting the above information integrity, setting the sampling frequency to a fixed value of no more than 1Hz is based on engineering optimization considerations. Since fire signals change slowly, when the sampling frequency is much higher than its necessary frequency, the data difference between adjacent sampling points is minimal, resulting in a large amount of information redundancy, which is not helpful for assessing fire risk. An excessively high sampling rate will linearly increase the computational load of data transmission, storage, and subsequent time-series analysis.

[0056] The algorithms in step 4, such as the calculation of the time-series rate of change, cross-correlation analysis, and principal component analysis, mathematically require that the input data be a sequence sampled at equal intervals with a uniform time length. The direct purpose of fixing the duration and sampling frequency is to generate such a standardized risk dataset, which is the prerequisite and foundation for the correct execution of all subsequent mathematical processing. The fixed duration and frequency ensure that the number of data points and the total time span processed by the core algorithm are exactly the same for each triggered risk assessment. This eliminates fluctuations in algorithm output caused by variations in data length or density. For example, autocorrelation calculation is sensitive to sequence length; different lengths can lead to results that cannot be directly compared. Fixed input ensures the stability and reliability of the algorithm output, avoiding misclassification of the same risk at different levels due to accidental differences in the analysis period. Since the analysis of all risk events is based on the exact same observation duration and sampling interval, the dynamic fire risk index calculated at different times and locations is comparable.

[0057] After confirming an abnormal temperature and its spread, acquiring time-series data on temperature, smoke, and CO concentration is crucial for a final diagnosis and assessment of the nature of the risk. An abnormal temperature indicates heat release, smoke confirms visible combustion products, and elevated CO concentration detects early, invisible combustion. By analyzing whether these changes occur synergistically, the system can reliably distinguish a real fire from other disturbances, such as equipment overheating or dust, thus preventing false alarms. Simultaneously, CO, as a very early indicator, allows the system to identify risks before the fire has fully developed, enabling earlier warnings.

[0058] Step 4: Based on the risk dataset, identify the changing patterns of temperature, smoke concentration, and CO gas concentration according to the time series data change rates, analyze the dynamic correlation between the changing patterns of temperature, smoke concentration, and CO gas concentration, and input the multi-parameter direct evaluation function to calculate the dynamic fire risk index. The logic for generating the time-series data change rate includes: The fixed duration of the observation time window is denoted as The fixed sampling frequency is denoted as Therefore, the sampling interval is The total number of sampling points within the window is ; According to the formula: Determined at The rate of change of internal temperature over time was used to obtain a result of length [missing information]. Time series data change rate sequence: ; According to the formula: Repeat the above calculations to obtain the time-series data on the rate of change of smoke concentration and CO gas concentration. , Ultimately, three sets of time-series data change rate sequences of equal length and time synchronization were obtained.

[0059] Based on the fixed duration of the observation time window and fixed sampling frequency The sampling interval can be determined. The total number of sampling points within the window is The "+1" here is because the sampling includes two data points: the start time and the end time of the window. For example, when T=180 seconds, f=0.2 Hz, Δt=5 seconds, n=180*0.2+1=36+1=37, meaning there are 37 sampling times from the start to the end of the window, corresponding to 37 raw data points for temperature, smoke concentration, and CO gas concentration.

[0060] In engineering, the formula is a discrete approximation of the instantaneous rate of change; it calculates the rate of change over a fixed, very short sampling interval. The average rate of change of intrinsic parameters is calculated simply and efficiently, requiring only the values ​​of two adjacent points and a fixed... It is well-suited for real-time processing, as it quantifies and extracts the important attribute of the rate of continuous, time-dependent change from the raw absolute value data.

[0061] Since raw temperature, smoke, and CO concentration values ​​only reflect instantaneous conditions and cannot quantify the speed and urgency of risk development, calculating the rate of change of time-series data can capture the core risk characteristic of how quickly parameters change, thus effectively distinguishing between slow environmental fluctuations and rapidly deteriorating real fire situations; by dividing by a fixed sampling interval... The rate of change was unified into a physical quantity with the change per unit time. This established a unified common benchmark for measuring the intensity of dynamic changes of the three parameters with completely different dimensions: temperature, smoke concentration, and CO concentration. This is the basis for subsequent mathematical operations such as cross-correlation and principal component analysis. This step directly supports the calculation of the dynamic fire risk index and ultimately realizes the leap from monitoring data to risk assessment, significantly improving the accuracy and timeliness of early warning.

[0062] The dynamic correlation between the changes in temperature, smoke concentration, and CO gas concentration was analyzed, specifically including: Time-series data on temperature, smoke concentration, and CO gas concentration are obtained from the risk dataset, with the lag step size denoted as... According to the formula: Determine the cross-correlation coefficient between time-series data of temperature and time-series data of smoke concentration. ,in and These are the time-series averages of temperature and smoke concentration, respectively. Find the maximum value in the sequence, denoted as . Simultaneously record the lag step size corresponding to the maximum value. As a leading feature in the temperature-smoke time series; According to the formula: Repeat the above calculations, where This is the average of the time-series data for CO gas concentration. The cross-correlation coefficient between the time-series data for temperature and the time-series data for CO gas concentration is obtained. Maximum cross-correlation number The lag step size corresponding to the maximum value As a leading characteristic of temperature-CO sequence; Calculating the lag step size in cross-correlation analysis This is to capture the temporal relationship between temperature, smoke, and CO concentration; in the aforementioned cross-correlation formula, the lag step size... The theoretical calculation range is However, in engineering implementation, directly calculating this entire range presents two problems: when When it is very large, the significant number of points involved in the calculation Very few, and the results are greatly affected by noise and endpoint effects, with weak statistical significance, and are considered invalid calculations; in the early development stage of a fire, the causal influence between parameters exists within a limited time window, and calculating these correlation coefficients with huge lags does not help to reveal the dynamic nature of the current risk, but may instead introduce false signals.

[0063] For example, in the early stages of a fire, temperature changes have a significant, identifiable, and quantifiable impact on smoke concentration. The maximum time span that might be required is defined as the maximum physical lag time. This time must be shorter than the observation time window to ensure sufficient data for computation. This is based on a preset maximum physical lag time, such as 120 seconds, and a known sampling interval. For example, if the time is 5 seconds, the maximum lag step size is calculated to be 120 / 5 = 24. Therefore, in the embodiment, the lag step size traversed by the system when actually calculating the cross-correlation is... The range is This range is entirely determined by the maximum physical lag time and the sampling interval.

[0064] The formula is the standard form of the standardized cross-correlation function, and the cross-correlation coefficient is... and The value is restricted to the range Within the range, a value of 1 indicates a perfect positive correlation, meaning that within a lag step of 1... Below, the two fluctuation curves completely overlap; when the value is 0, it indicates no linear correlation, meaning the fluctuations of the two curves have no coordinated pattern; when the value is -1, it indicates a perfect negative correlation, meaning that at a lag step of 0, the two curves have no coordinated pattern. Below, the peak of one curve corresponds exactly to the trough of another, completely opposite.

[0065] Maximum cross-correlation number and The value is in the interval Within this system, the coefficient is the core input of the dynamic fire risk index, used to quantify the intensity of the coordinated changes in temperature and smoke, and temperature and CO. If the coefficient is close to 0 or negative, it indicates that the increase in both has no coordinated pattern in its temporal evolution, and is merely a coincidence, which will be judged as low risk or a false alarm; if the coefficient is close to 1, it indicates that the changes in both are highly coordinated, which strongly suggests that both are driven by the same ordered process, namely combustion, and the probability of confirming it as a real fire is extremely high.

[0066] This invention uses an algorithm to actively mine dynamic features that characterize the inherent physical coupling relationship of a fire process from multi-source time-series data, and uses these features as the core basis for risk decision-making. This design effectively solves long-standing technical problems such as high false alarm rate and inaccurate early warning.

[0067] During the development of a fire, changes in temperature, smoke, and carbon monoxide typically exhibit time lags. For example, CO may increase before a significant temperature rise, while smoke may appear after the temperature has risen. The lag step corresponding to the maximum value should be recorded. and This is to quantify this time-leading characteristic, which is key to assessing the stage of fire development.

[0068] Lag step This characterizes the time shift between temperature change and smoke concentration change when the synergy between them reaches its strongest. When, it indicates that the temperature sequence is advanced. At a time step of 1, the correlation with the smoke sequence is strongest, indicating that temperature changes precede smoke changes. In fire scenarios, this is typically a characteristic of open flame or rapidly developing flaming fires: the ignition source first releases a large amount of heat, causing a sharp rise in temperature, followed by the gradual accumulation and detection of solid particles produced by combustion; when When, it indicates that the smoke sequence is advanced At a time step of 1, the correlation with the temperature sequence is strongest, indicating that changes in smoke lead changes in temperature. This strongly points to smoldering, burning, or incomplete combustion without an open flame: the substance pyrolyzes at a lower temperature, initially producing a large amount of visible smoke, but the heat release is slow and accumulates, resulting in a significant temperature rise that lags behind the formation of smoke.

[0069] Lag step This characterizes the time shift between temperature change and CO concentration change when the synergy between the two reaches its strongest point. When this occurs, it indicates that the temperature change leads the CO concentration change. This can occur in the rapid open flame of some clean fuels because combustion is complete, and the CO formation rate may be highly synchronized with or even slightly lag behind the heat release rate. When CO concentration changes precede temperature changes, this indicates that CO concentration changes are the most significant characteristic in the very early stages of most fires, especially smoldering and organic fires. As an early gaseous product of incomplete combustion, the significant increase in CO concentration often precedes the perceptible temperature rise. A large positive... The value is a key signal that a fire is in its early smoldering or latent stage.

[0070] Through quantitative extraction and This invention enables the differentiation between smoldering and open flame modes based on the leading characteristics of the time sequence before a fire has fully developed. This provides crucial information for taking targeted early response measures. The specific numerical values ​​of the leading-lag relationship reflect the evolution of a fire from latent to developing to intense, enabling the dynamic fire risk index to more accurately reflect the urgency of the current situation.

[0071] Based on the rate of change of time-series data for temperature, smoke concentration, and CO gas concentration, a (n-1)*3 data matrix is ​​constructed by treating the three sets of time-series data rate of change sequences as column vectors. : According to the formula: For the matrix To centralize, among which, The elements are the mean of the time series change rates of temperature, smoke concentration, and CO gas concentration. According to the formula: Calculate the covariance matrix ; For covariance matrix Perform eigenvalue decomposition and solve the characteristic equation: Three eigenvalues ​​were obtained. , , Sort by size from largest to smallest: and the corresponding feature vectors , , The largest eigenvalue This refers to the characteristics of co-principal components; The construction matrix Centering, calculating the covariance matrix The process of eigenvalue decomposition is the standard mathematical procedure of principal component analysis (PCA). PCA is a standard and mature method in statistics used to extract the most significant directions of change and the strength of cooperation in multivariate datasets. Its output is the largest eigenvalue. It naturally and unambiguously measures the magnitude of change in multidimensional data along the most important common direction.

[0072] The three rate-of-change sequences of temperature, smoke concentration, and CO gas concentration constitute a three-dimensional dataset. Principal component analysis can accurately identify the main direction of this data in three-dimensional space. If fire development is the dominant factor, then the data points will be mainly distributed along this direction. The distribution will be very large; if the changes are random and disordered, the data point distribution will tend to be spherical. The size will be relatively small.

[0073] Introducing co-principal component features The goal is to provide a direct indicator that can quantify the overall synergistic strength of changes in temperature, smoke, and CO gas concentration. This feature, through principal component analysis, extracts the variance of the principal components representing the common direction of change among the three parameters from multivariate data. The magnitude of this variance directly characterizes the degree of consistency in the synergistic deterioration of multiple parameters during fire dynamics. This design further filters false alarms under complex interference scenarios, effectively distinguishing between the highly synergistic patterns exhibited by multiple parameters driven by the same fire source in real fires and the chaotic change patterns caused by accidental or localized anomalies. Furthermore, As a continuous variable reflecting the overall energy release intensity of the situation, it complements the paired time-series features extracted by cross-correlation analysis, together constituting the dynamic fire risk index. The feature inputs enable the final risk assessment to be comprehensively judged from two dimensions: pairwise relationship analysis and overall system synergy, which significantly improves the accuracy and reliability of graded early warning.

[0074] The multi-parameter direct evaluation function is: This indicates a dynamic fire risk index. arrive This indicates the preset weight coefficient value.

[0075] The multi-parameter direct evaluation function involves only a few multiplications and additions, resulting in extremely fast computation and high determinism. This ensures the system can output risk assessment results within seconds or even milliseconds, meeting the core requirement of timely early warning. In contrast, nonlinear models often have computational overhead and latency an order of magnitude higher, making them unsuitable for the resource constraints of this scenario. Linear transformations are deterministic, eliminating the numerical instability or convergence issues that may occur with nonlinear models. This guarantees high reliability and repeatability of system behavior, meeting the basic requirements of functional safety standards for critical system logic. The five dynamic features input to this function are not raw data, but rather advanced diagnostic features that have undergone complex nonlinear transformation preprocessing and deep extraction, such as cross-correlation analysis and principal component analysis in step 4. These features themselves already contain nonlinear mapping relationships. Although the model is linear, the weight coefficients w_1 to w_5 can be automatically learned from historical data through a weight training method. They can be optimized according to different application scenarios, thus achieving flexibility in accurate assessment while ensuring the simplicity and reliability of the core logic.

[0076] Weighting coefficient value arrive The methods for determining this include: Collect time-series data from Q historical scenarios, where Q should exceed 100. Process the data using a method completely simulating the above approach to obtain the maximum cross-correlation coefficient between the time-series data of temperature and the time-series data of smoke concentration. The maximum cross-correlation coefficient between time-series temperature data and time-series CO gas concentration data. Temperature-smoke timing leading characteristics Temperature-CO time-leading characteristics Synergistic principal component features Assign a realistic fire risk reference value to each historical scenario. This value is typically determined by security experts through post-incident assessments based on the severity and speed of the incident's consequences, according to the formula... , Determine the target arrive ,in: According to the formula: The specific weight coefficient values ​​are obtained by solving. ,in It is a Q-dimensional column vector containing the true risk values ​​of all samples; The dynamic fire risk index can be obtained by substituting the calculated weight coefficient values ​​into the multi-parameter direct evaluation function.

[0077] The matrix This refers to the historical dynamic feature data matrix assembled for weight coefficient training. Each row of this matrix represents an independent historical risk scenario, and each column represents a dynamic feature. The dynamic features with clear physical meaning obtained through steps 1-4 of this invention from Q dispersed historical scenarios are systematically and structurally organized into a unified data set.

[0078] The formula It is the normal equation for solving the optimal weight vector in multiple linear regression. It provides this invention with a mathematical tool for optimally linearly fitting five heterogeneous dynamic features to expert-defined fire risk reference values. This ensures that the weight coefficients are derived from historical statistical patterns rather than subjective experience, improving the scientific rigor and consistency of the assessment. Secondly, the weights... It can be retrained with new data, enabling the model to continuously approximate real-world risk patterns in specific scenarios and achieve adaptive optimization; finally, this formula is directly evaluated using a multi-parameter evaluation function. This constitutes a closed, efficient, and interpretable optimization system. While ensuring real-time computing performance, it also ensures the transparency and verifiability of the entire risk assessment logic from feature extraction to final decision-making, thereby fundamentally supporting the technical goal of this invention to achieve high-precision, low-false-alarm, and graded early warning.

[0079] Step 5: Based on the dynamic fire risk index of the initial risk area, determine the hazard level of the initial risk area according to the preset multi-level risk thresholds, and output graded early warning information.

[0080] The logic for determining the hazard level of the initial risk area and outputting tiered early warning information is as follows: Set threshold , , ,in The risks are divided into four levels: high risk, low risk, medium risk and high risk. like It is classified as a "watch" level item; if It is classified as low-risk; if It is classified as medium risk; if It was classified as high-risk.

[0081] The fundamental reason for setting three thresholds and classifying risks into four levels is to more accurately map the continuous process of fire risk evolution and match differentiated emergency response strategies, thus addressing the inherent shortcomings of traditional methods that only focus on whether an alarm is triggered. The three thresholds enable refined, segmented perception of the entire process of risk development, escalation, and crisis; enabling tiered responses and optimized resource allocation. Different levels trigger response actions with completely different costs and impacts: **Attention Level:** The system may only automatically record and mark the risk, notifying security personnel to strengthen video patrols or on-site inspections, without triggering audible and visual alarms, avoiding unnecessary panic and resource consumption; **Low Risk Level:** Triggers a system warning, notifying area managers and security personnel to verify the situation and initiating preliminary preparations, such as checking fire exits; **Medium Risk Level:** Immediately triggers local audible and visual alarms, automatically notifying the park's fire station and emergency response team, and initiating preliminary steps of the emergency plan; **High Risk Level:** Triggers the highest-level alarm across the entire area, automatically activating smoke extraction, sprinkler systems, and other fire-fighting equipment, and issuing mandatory evacuation orders. These three thresholds precisely separate these four response levels, ensuring that the system response matches the actual risk level and achieving optimal resource allocation.

[0082] Traditional single high thresholds are prone to alarm delays, and by the time an alarm is triggered, the fire may be difficult to control, posing a risk of missed alarms. Single low thresholds are highly susceptible to frequent false alarms due to environmental interference, leading to a "boy who cried wolf" effect and causing personnel to become desensitized to alarms, also posing a risk of false alarms. The multi-level thresholds of this invention may only trigger a level of concern under brief, weak risk fluctuations, without immediately eliciting a large-scale response; however, continuous and intensifying risk signals will gradually escalate the alarm, significantly improving the system's reliability and usability.

[0083] First risk threshold Second risk threshold and the third risk threshold The determination of the risk level is achieved through an ordered classification statistical model trained on historical risk event data, ensuring the objectivity and optimality of the hierarchical decision-making. The specific steps are as follows: Collect complete data records for M completed historical risk events, where M is a sufficient number of positive integers. Each record should contain: The time-series data of temperature, smoke concentration, and CO gas concentration during the incident; the actual hazard level label assigned to the incident by safety experts based on factors such as the severity of the eventual consequences, the scope of spread, and the difficulty of handling the situation, denoted as [label name missing]. The risk levels are 0, 1, 2, and 3 for the "Attention" level, "Low Risk" level, "Medium Risk" level, and "High Risk" level, respectively. The higher the value, the higher the risk level and the more dangerous the event.

[0084] For the m-th historical event, the method described in this invention is applied to calculate the dynamic fire risk index from the start to the end of the event. The time series sequence, from which the maximum value is extracted as the value. , obtain the training dataset ,in This represents the actual value of the true risk level observed in the m-th sample.

[0085] Other feasible extraction methods include, but are not limited to: taking the time of risk level transition. The value is taken from the stable period of the event's development. In practical applications, we usually use the maximum value because the maximum value reflects the peak risk of an event and is strongly correlated with the severity of the event. The maximum value is easy to obtain and does not require additional annotations, such as the time of level transition, or definitions, such as the start and end times of the stable period.

[0086] Considering different parks or different periods The values ​​may have scale differences, therefore... Standardize the values: calculate all mean and standard deviation For each Standardize: The standardized dynamic fire risk index is .

[0087] Since risk levels have a natural order relationship (0 < 1 < 2 < 3), this invention uses an ordered logistic regression model for training, which can simultaneously learn classification thresholds and feature coefficients.

[0088] The basic form of the ordered logistic regression model is: This formula is the core definition of ordered logistic regression, where =0,1,2 Let be the threshold parameter to be learned, and satisfy . ,when Less than At that time, the focus tends to be on the level of attention; in and In between, it tends to be low-risk; and When between these levels, the risk tends to be moderate; when greater than this level, the risk is higher. At that time, it tends to be high-risk. for The regression coefficient is usually positive, indicating that... The larger the size, the higher the risk level. Indicates that in a given Under the condition that the actual risk level is less than or equal to The cumulative probability, where, for the boundary case, we have: and .

[0089] Model parameter set The maximum likelihood estimation method is used to solve this problem. Maximum likelihood estimation is a classic method for parameter estimation in statistics. This solution process automatically and objectively learns the optimal parameters from the data, completely replacing the manual parameter adjustment mode that relies on expert experience. It ensures that the obtained threshold parameters are the statistically optimal solution supported by current historical data. The specific steps are as follows: Constructing the likelihood function: For M independent samples in the training dataset, the likelihood function... The product of the observation probabilities of all samples: To simplify the calculation, we take the natural logarithm to obtain the log-likelihood function: Maximizing the log-likelihood function using a numerical optimization algorithm Simultaneously satisfying the ordered constraint of the threshold In practice, mature statistical modeling tool libraries are directly used. For example, in the Python environment, the `OrderedModel` class from the `statsmodels` library is used; in the R language environment, the `polr` function from the `MASS` package is used. The standardized feature vectors are then... As the independent variable, the true rating label As an ordered dependent variable, the parameters can be automatically estimated and the optimal parameter values ​​can be output by calling the corresponding fitting function. .

[0090] The standardized threshold is destandardized, which is the reverse of standardization, to obtain the original threshold. Final threshold at scale: The calculated , , For real-time risk classification, new tagged risk event data can be automatically entered into the historical database. The above steps are periodically re-executed to iteratively optimize the thresholds, enabling the system to continuously learn and adapt to environmental changes.

[0091] Using the threshold calculation method described above, firstly, the scientific and objective nature of the threshold is ensured, as it learns directly from historical patterns, eliminating human bias and laying a theoretical foundation for reducing false alarms and missed alarms in the system. Secondly, this method endows the system with strong adaptive capabilities, enabling it to automatically generate the most suitable threshold based on the inherent risk characteristics of different parks. More importantly, the system thus possesses continuous evolution capabilities, and can periodically iterate and optimize the threshold by absorbing new risk event data, allowing the risk perception capability to continuously improve over time and with data accumulation.

[0092] Please see Figure 2 A park hazard situation awareness device based on multi-source data fusion, the device being used to implement the aforementioned park hazard situation awareness method based on multi-source data fusion, comprising: Grid division module: Divide the park area to be sensed into a grid to form several sub-areas, acquire the temperature data of each sub-area, and when the temperature of any sub-area exceeds the preset first safety threshold, mark the sub-area as the initial risk area and simultaneously acquire the real-time temperature data of all its adjacent sub-areas. Diffusion judgment module: Taking the initial risk area as the center, and combining the temperature data of adjacent sub-areas and the distance between them and the initial risk area, the module judges whether the temperature data of adjacent sub-areas meets the temperature diffusion conditions based on the changes in the temperature data and corresponding distances of each adjacent sub-area. Data acquisition module: In response to meeting the temperature diffusion conditions, it collects time-series data of temperature, smoke concentration and CO gas concentration in the initial risk area within the set observation time window and labels it as a risk dataset; Risk assessment module: Based on the risk dataset, it identifies the changing patterns of temperature, smoke concentration and CO gas concentration according to the time series data change rate, and analyzes the dynamic correlation between the changing patterns of temperature, smoke concentration and CO gas concentration. The multi-dimensional feature vector is input into the multi-parameter direct evaluation function to calculate the dynamic fire risk index. Hazard warning module: Based on the dynamic fire risk index of the initial risk area, it determines the hazard level of the initial risk area according to the preset multi-level risk thresholds and outputs graded warning information.

[0093] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0094] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0095] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for perceiving hazardous situations in a park based on multi-source data fusion, characterized in that, The specific steps include: Step 1: Divide the park area to be sensed into a grid to form several sub-areas, obtain the temperature data of each sub-area, and when the temperature of any sub-area exceeds the preset first safety threshold, mark the sub-area as the initial risk area, and simultaneously obtain the real-time temperature data of all its adjacent sub-areas. Step 2: Taking the initial risk area as the center, and combining the temperature data of adjacent sub-areas and the distance between them, determine whether the temperature data of adjacent sub-areas meet the temperature diffusion conditions based on the changes in the temperature data and corresponding distances of each adjacent sub-area. Step 3: In response to meeting the temperature diffusion conditions, acquire time-series data of temperature, smoke concentration and CO gas concentration in the initial risk area within the set observation time window, and label it as a risk dataset; Step 4: Based on the risk dataset, identify the changing patterns of temperature, smoke concentration, and CO gas concentration according to the time series data change rates, analyze the dynamic correlation between the changing patterns of temperature, smoke concentration, and CO gas concentration, and input the multi-parameter direct evaluation function to calculate the dynamic fire risk index. Step 5: Based on the dynamic fire risk index of the initial risk area, determine the hazard level of the initial risk area according to the preset multi-level risk thresholds, and output graded early warning information.

2. The method for campus hazard situation awareness based on multi-source data fusion according to claim 1, characterized in that: The logic for dividing the data into several sub-regions using a grid includes: Determine the center of the park area, determine the size of the square grid for the initial division, construct a central grid whose center coincides with the center of the park area, expand multiple grids outward along the central grid to ensure that all grids cover the park area, and use the grid division scheme at this time as the initial grid division scheme when all grids cover the park area. The grids that are not fully filled within the park area are merged into one grid with adjacent grids that are fully filled within the park area. Adjacent grids are those that share edges or vertices. The merging logic includes: Based on the centers of the grids not filled by the park area and the centers of the adjacent grids that are completely filled by the park area, the distance between their centers is determined. The grids not filled by the park area are merged into the nearest adjacent grid that is completely filled by the park area. If the distances are the same, then any adjacent grid that is completely filled by the park area is randomly selected for merging. The final result of grid division is that each grid is a sub-region.

3. The method for park hazard situation awareness based on multi-source data fusion according to claim 1, characterized in that: The adjacent sub-regions of the initial risk region include directly adjacent sub-regions and indirectly adjacent sub-regions. The directly adjacent sub-regions are all sub-regions that share edges or vertices with the initial risk region. The indirectly adjacent sub-regions are all sub-regions that share edges or vertices with the directly adjacent sub-regions.

4. The method for park hazard situation awareness based on multi-source data fusion according to claim 3, characterized in that: Determining whether the temperature data of adjacent sub-regions meet the temperature diffusion conditions specifically includes: Obtain the distance between the center of the initial risk region and the center of each adjacent sub-region, obtain the current temperature data and inherent temperature data of each adjacent sub-region, and calculate the temperature gradient of the adjacent sub-regions. The method for calculating the temperature gradient is as follows: The difference between the current temperature data of an adjacent sub-region and the inherent temperature data of the adjacent sub-region is the first temperature difference. The ratio of the first temperature difference to the distance between the center of the initial risk area and the center of the adjacent sub-region is represented as the temperature gradient of the adjacent sub-region. Calculate the temperature gradient of all adjacent sub-regions, and then... Perform normalization and map to Interval, normalized temperature gradient Calculated using the following formula: in The normalized temperature gradient of the i-th neighboring sub-region is... This represents the maximum temperature gradient across all adjacent sub-regions. This represents the minimum temperature gradient across all adjacent sub-regions. When the temperature gradient of adjacent sub-regions meets the preset conditions and the proportion of adjacent sub-regions that meet the conditions exceeds the preset percentage, the initial risk area meets the temperature diffusion conditions.

5. The method for campus hazard situation awareness based on multi-source data fusion according to claim 1, characterized in that: The logic for setting the observation time window includes: when the initial risk area meets the temperature diffusion conditions, setting the observation time window with that moment as the end point, and backtracking by a preset fixed duration to determine the start and end times of the observation time window; Within the observation time window, time-series data of temperature, smoke concentration, and CO gas concentration in the initial risk area are acquired at a fixed sampling frequency.

6. The method for campus hazard situation awareness based on multi-source data fusion according to claim 1, characterized in that: The logic for generating the time-series data change rate includes: The fixed duration of the observation time window is denoted as The fixed sampling frequency is denoted as Therefore, the sampling interval is The total number of sampling points within the window is ; According to the formula: Determined at The rate of change of internal temperature over time was used to obtain a result of length [missing information]. Time series data change rate sequence: ; According to the formula: Repeat the above calculations to obtain the time-series data on the rate of change of smoke concentration and CO gas concentration. , Ultimately, three sets of time-series data change rate sequences of equal length and time synchronization were obtained.

7. The method for park hazard situation awareness based on multi-source data fusion according to claim 1, characterized in that: The dynamic correlation between the changes in temperature, smoke concentration, and CO gas concentration was analyzed, specifically including: Time-series data on temperature, smoke concentration, and CO gas concentration are obtained from the risk dataset, with the lag step size denoted as... According to the formula: Determine the cross-correlation coefficient between time-series data of temperature and time-series data of smoke concentration. ,in and These are the time-series averages of temperature and smoke concentration, respectively. Find the maximum value in the sequence, denoted as . Simultaneously record the lag step size corresponding to the maximum value. As a leading feature in the temperature-smoke time series; According to the formula: Repeat the above calculations, where This is the average of the time-series data for CO gas concentration. The cross-correlation coefficient between the time-series data for temperature and the time-series data for CO gas concentration is obtained. Maximum cross-correlation number The lag step size corresponding to the maximum value As a leading characteristic of temperature-CO sequence; Based on the rate of change of time-series data for temperature, smoke concentration, and CO gas concentration, a (n-1)*3 data matrix is ​​constructed by treating the three sets of time-series data rate of change sequences as column vectors. : According to the formula: For the matrix To centralize, among which, The elements are the mean of the time series change rates of temperature, smoke concentration, and CO gas concentration. According to the formula: Calculate the covariance matrix ; For covariance matrix Perform eigenvalue decomposition and solve the characteristic equation: Three eigenvalues ​​were obtained. , , Sort by size from largest to smallest: and the corresponding feature vectors , , The largest eigenvalue This is a characteristic of co-principal components.

8. The method for campus hazard situation awareness based on multi-source data fusion according to claim 1, characterized in that: The multi-parameter direct evaluation function is: This indicates a dynamic fire risk index. arrive This indicates the preset weight coefficient value.

9. A method for perceiving hazardous situations in a park based on multi-source data fusion as described in claim 8, characterized in that: The logic for determining the hazard level of the initial risk area and outputting tiered early warning information is as follows: Set threshold , , ,in The risks are divided into four levels: high risk, low risk, medium risk and high risk. like It is classified as a "watch" level item; if It is classified as low-risk; if It is classified as medium risk; if It was classified as high-risk.

10. A park hazard situation awareness device based on multi-source data fusion, the device being used to implement the park hazard situation awareness method based on multi-source data fusion as described in any one of claims 1-9, characterized in that, include: Grid division module: used to divide the park area to be sensed into grids to form several sub-areas, acquire temperature data of each sub-area, and when the temperature of any sub-area exceeds the preset first safety threshold, the sub-area is marked as the initial risk area, and real-time temperature data of all its adjacent sub-areas are acquired simultaneously. Diffusion judgment module: It is used to determine whether the temperature data of adjacent sub-regions meet the temperature diffusion conditions based on the temperature data of each adjacent sub-region and the distance between the adjacent sub-regions and the initial risk area, taking the initial risk area as the center and combining the temperature data of the adjacent sub-regions and the corresponding distance. Data acquisition module: In response to meeting the temperature diffusion conditions, it collects time-series data of temperature, smoke concentration and CO gas concentration in the initial risk area within a set observation time window and labels it as a risk dataset; Risk assessment module: Based on the risk dataset, it identifies the changing patterns of temperature, smoke concentration and CO gas concentration according to the time series data change rate, and analyzes the dynamic correlation between the changing patterns of temperature, smoke concentration and CO gas concentration. The multi-dimensional feature vector is input into the multi-parameter direct evaluation function to calculate the dynamic fire risk index. Hazard warning module: Based on the dynamic fire risk index of the initial risk area, and according to the preset multi-level risk thresholds, it determines the hazard level of the initial risk area and outputs graded warning information.

Images