Fire-fighting supervision alarm method and system based on big data

Abstract

The invention provides a fire-fighting supervision alarm method and system based on big data, and the method comprises the steps: obtaining fire-fighting supervision multi-source data, carrying out the preprocessing and feature extraction of the fire-fighting supervision multi-source data, obtaining a fire-fighting time sequence feature vector, inputting the fire-fighting time sequence feature vector to an instability early-warning strategy based on chaotic recognition, and carrying out the early-warning of the instability. The method comprises the steps of obtaining an instability early-warning strategy, generating a nonlinear instability early-warning signal based on the instability early-warning strategy, obtaining real-time environment data, carrying out deduction by adopting a pre-trained Fourier neural operator model based on the real-time environment data, obtaining a physical field deduction sequence set, and generating a fire-fighting risk alarm based on the physical field deduction sequence set. And obtaining a causal conduction path for the fire-fighting time sequence feature vectors of the time domain before and after fire-fighting risk alarm through a time-varying causal network, and generating a structured early-warning diagnosis report in combination with a nonlinear instability early-warning signal and a physical field deduction sequence. And a dynamic risk assessment fire-fighting supervision alarm method is provided in real time.

Description

Technical Field

[0001] This invention relates to the field of fire supervision and early warning technology, and more specifically, to a fire supervision and alarm method and system based on big data. Background Technology

[0002] Fire monitoring and early warning are key technologies for preventing fires and protecting life and property. With the advancement of the Internet of Things, sensor technology and information technology construction, the amount of data available in the fire protection field has increased significantly in terms of scale, dimensions and timeliness, which makes it possible to provide intelligent early warning based on big data. However, existing fire monitoring and early warning methods are still insufficient in early risk identification and dynamic simulation of fire risks, and it is difficult to provide forward-looking and dynamic fire early warning information in complex environments such as industrial manufacturing.

[0003] On the one hand, existing fire early warning methods mostly rely on single sensor threshold alarms or simple statistical models, which have limited ability to identify early signs of complex nonlinear disasters. Traditional methods, such as alarms based on fixed smoke concentration thresholds, temperature thresholds, etc., although responding quickly and directly, cannot effectively capture nonlinear precursor signals such as critical slowdown and sudden changes in dynamic characteristics before system instability, resulting in delayed early warning.

[0004] On the other hand, existing technologies lack the ability to dynamically deduce the physical nature of early warning signals. Fire warnings are easily disconnected from the actual development of fire risks. Traditional methods can only indicate the existence of risks by issuing early warning signals, but cannot simulate and dynamically deduce the physical development process of potential fire risks, such as the spatiotemporal evolution of temperature fields and toxic gas concentration fields. It is difficult to assess the scope and intensity of the impact and the threat to key protected targets, resulting in low decision support value of early warning information.

[0005] In summary, there is an urgent need for a fire monitoring and alarm method that can provide forward-looking early warning of fire risks in complex environments and offer dynamic risk assessments with high decision support value in real time. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a fire monitoring and alarm method and system based on big data.

[0007] The first aspect of this invention provides a fire monitoring and alarm method based on big data, comprising: Acquire multi-source fire supervision data, and preprocess and extract features from the multi-source fire supervision data to obtain a fire time series feature vector; The fire-fighting time-series feature vector is input into the instability early warning strategy based on chaos identification, and a nonlinear instability early warning signal is generated based on the instability early warning strategy. Real-time environmental data is obtained based on nonlinear instability early warning signals. Based on the real-time environmental data, a pre-trained Fourier neural operator model is used to perform deduction to obtain a set of physical field deduction sequences. A fire risk alarm is generated based on the set of physical field deduction sequences. The fire time-series characteristic vectors before and after the fire risk alarm are analyzed by time-varying causal networks to obtain the causal transmission path, and a structured early warning diagnosis report is generated by combining nonlinear instability early warning signals and physical field deduction sequences.

[0008] According to a preferred embodiment, multi-source fire monitoring data is acquired, and the multi-source fire monitoring data is preprocessed and feature extracted to obtain a fire time-series feature vector, including: Based on multi-source fire supervision data, statistical features, dynamic features, spectral features and nonlinear features are obtained within a sliding window; Feature selection is performed based on statistical characteristics, dynamic characteristics, spectral characteristics, and nonlinear characteristics using the maximum correlation and minimum redundancy criterion. The selected features are subjected to nonlinear dimensionality reduction using kernel principal component analysis to obtain the fire-fighting time-series feature vector.

[0009] According to a preferred embodiment, a fire-fighting time-series feature vector is input into an instability early warning strategy based on chaotic identification, and a nonlinear instability early warning signal is generated based on the instability early warning strategy, including: Extract the process disorder index sequence from the fire-fighting time-series feature vector; The critical slowdown phenomenon was monitored in the process disorder index sequence to obtain the critical slowdown index. The process disorder index sequence is reconstructed in phase space, and geometric dynamic indexes are obtained based on the phase space reconstruction results; The critical slowdown index and the geometric dynamic index are integrated into a comprehensive instability index, and a nonlinear instability early warning signal is generated based on the comprehensive instability index.

[0010] According to a preferred embodiment, a critical slowing phenomenon is monitored on a process turbulence index sequence to obtain critical slowing indices, including: The critical slowing index includes autocorrelation time, autocorrelation variance, and low-frequency noise intensity parameters. The autocorrelation function is obtained based on the process disorder index sequence, and the autocorrelation time and autocorrelation variance are obtained by fitting an exponential decay curve based on the autocorrelation function. The power spectrum is obtained based on the process disorder index sequence, and the low-frequency noise intensity parameter in the power spectrum is monitored.

[0011] According to a preferred embodiment, a phase space reconstruction is performed on the process disorder index sequence, and a geometric dynamic index is obtained based on the phase space reconstruction result, including: The geometric dynamics indicators include the maximum Lyapunov index, correlation dimension, and recursive quantification analysis indicators. Based on the process disorder index sequence, the average mutual information method is used to obtain the delay time, the false nearest neighbor method is used to obtain the embedding dimension, and the Takens embedding theorem is used to reconstruct the phase space of the process disorder index sequence based on the delay time and the embedding dimension, and the reconstructed phase space trajectory is obtained. Based on the reconstructed phase space trajectory, the Wolf algorithm is used to obtain the maximum Lyapunov exponent; Based on the reconstructed phase space trajectory, the correlation dimension is obtained by analyzing the slope of the linear region in the double logarithmic graph of correlation integral and distance; Recursive graphs are generated based on reconstructed phase space trajectories, and deterministic and laminar flow indices are extracted as recursive quantitative analysis indicators.

[0012] According to a preferred embodiment, the critical slowdown index and the geometric dynamic index are fused into a comprehensive instability index, and a nonlinear instability early warning signal is generated based on the comprehensive instability index, including: The Bayesian change point detection algorithm is applied to the comprehensive instability index to identify the starting point of the growth trend of the comprehensive instability index; Historical comprehensive instability index is obtained based on fire-fighting time-series feature vectors, and adaptive instability threshold is set based on the statistical distribution of historical comprehensive instability index; When the comprehensive instability index is greater than or equal to the adaptive instability threshold, it is determined whether the comprehensive instability index is in an upward trend at the current moment based on the starting point of the growth trend. If it is in an upward trend, a nonlinear instability warning signal is generated. The nonlinear instability early warning signal includes at least the early warning time, main contributing indicators, risk units, and equipment parameters.

[0013] According to a preferred embodiment, real-time environmental data is acquired based on a nonlinear instability early warning signal; a pre-trained Fourier neural operator model is used to perform deduction based on the real-time environmental data to obtain a set of physical field deduction sequences; and a fire risk alarm is generated based on the set of physical field deduction sequences, including: Based on the risk unit in the nonlinear instability early warning signal, the corresponding simplified model of the unit is called from the pre-stored geometric model library. Combined with the equipment parameters and real-time environmental data in the nonlinear instability early warning signal, the risk source parameters are obtained. The risk source parameters include the risk source location, risk source intensity and material properties.

[0014] Based on the simplified unit model and real-time environmental data, the inference computational domain is obtained and a mesh is generated to form an initial-boundary condition vector. Based on material properties and real-time environmental data, a matching Fourier neural operator model is selected from a pre-built inference model library. The initial-boundary condition vector is input into the selected Fourier neural operator model. Through global convolution operation in the Fourier domain, the physical field prediction sequence for multiple consecutive time steps in a single forward propagation is directly output. The physical field prediction sequence includes the concentration field prediction index and the temperature field prediction index. Perturbation sampling is performed on the parameters of potential risk sources to obtain multiple perturbation initial-boundary condition vectors. Matching Fourier neural operator models are selected to obtain multiple physical field prediction sequences. All physical field prediction sequences are merged into a physical field deduction sequence set, and fire risk consequence quantitative analysis is performed to generate fire risk alarms.

[0015] According to a preferred embodiment, all physical field prediction sequences are merged into a physical field deduction sequence set, and a fire risk consequence quantitative analysis is performed to generate a fire risk alarm, including: At each time step, based on all physical field prediction sequences in the physical field deduction sequence set, the corresponding concentration exceedance probability distribution field and temperature exceedance probability distribution field are statistically analyzed, and the corresponding concentration risk envelope and temperature risk envelope are obtained: When the concentration risk envelope or temperature risk envelope at any time step intrudes into the critical protection area, a fire risk alarm is generated. The fire risk alarm shall include at least the alarm trigger time step, concentration risk envelope, temperature risk envelope and corresponding confidence information.

[0016] According to a preferred embodiment, at each time step, based on all physical field prediction sequences in the physical field deduction sequence set, the corresponding concentration exceedance probability distribution field and temperature exceedance probability distribution field are statistically analyzed, and the corresponding concentration risk envelope and temperature risk envelope are obtained, including: For each time step, based on a preset safety threshold, the concentration over-threshold probability of each node in all physical field prediction sequences at that time step is identified, forming a concentration over-threshold probability distribution field; For each time step, based on the preset thermal radiation damage threshold, the temperature over-threshold probability of each node in all physical field prediction sequences at that time step is identified, forming a temperature over-threshold probability distribution field. For the concentration exceeding the threshold probability distribution field and the temperature exceeding the threshold probability distribution field, the spatial boundaries under a specified confidence level are extracted respectively to form the concentration risk envelope and temperature risk envelope corresponding to the time step. The risk envelope represents the maximum spatial range of fire risk impact under the specified confidence level.

[0017] A second aspect of the present invention also provides a fire monitoring and alarm system based on big data, comprising: The fire data acquisition module is used to collect multi-source fire supervision data, and to preprocess and extract features from the multi-source fire supervision data to obtain a fire time series feature vector. A nonlinear instability early warning module is provided, which receives the fire-fighting time-series feature vector and executes an instability early warning strategy based on chaos identification to generate a nonlinear instability early warning signal. The fire risk physical field simulation module is used to respond to the nonlinear instability warning signal, call a pre-trained Fourier neural operator model based on real-time environmental data to perform rapid simulation, obtain a set of physical field simulation sequences, and perform quantitative analysis of fire risk consequences to generate a fire risk alarm. The causal analysis module is used to perform time-varying causal network analysis on the fire time-series feature vectors before and after the fire risk alarm to obtain the causal transmission path, and combine the nonlinear instability early warning signal with the physical field deduction sequence set to generate a structured early warning diagnosis report.

[0018] Based on the above, this application embodiment obtains fire time series feature vectors by acquiring multi-source fire supervision data, preprocessing and extracting features, and inputting them into an instability early warning strategy based on chaos identification to generate nonlinear instability early warning signals. First, key dynamic features are extracted from multi-source heterogeneous data, and critical slowing monitoring and phase space reconstruction technology in chaos theory are used to capture nonlinear precursors before system instability, realizing a forward-looking early warning of fire risks in complex environments, and solving the problems of insensitivity to disaster precursors and delayed warning in traditional threshold alarm methods.

[0019] On the other hand, by triggering based on nonlinear instability early warning signals and using a pre-trained Fourier neural operator model to rapidly extrapolate the fire risk physical field driven by real-time environmental data, the future evolution sequence of key physical parameters such as temperature field and concentration field is obtained, realizing the dynamic extrapolation and consequence quantitative assessment of the real-time fire risk physical field, thereby providing dynamic risk assessment information with high decision support value in complex scenarios such as industrial manufacturing. Attached Figure Description

[0020] Figure 1 The execution flowchart of the fire monitoring and alarm method based on big data of the present invention is presented.

[0021] Figure 2 A flowchart illustrating the acquisition of nonlinear instability early warning signals in the fire monitoring and alarm method based on big data of the present invention is presented.

[0022] Figure 3 A schematic diagram of the fire monitoring and alarm system based on big data of the present invention is shown. Detailed Implementation

[0023] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0026] like Figure 1 , Figure 2 , Figure 3 As shown: The first aspect of this invention provides a fire monitoring and alarm method based on big data, comprising: S100: Acquire multi-source fire supervision data, and preprocess and extract features from the multi-source fire supervision data to obtain a fire time series feature vector.

[0027] The aforementioned multi-source fire monitoring data specifically refers to multimodal, heterogeneous time-series data used in complex industrial settings, such as petrochemical plants, to capture early instability and potential leakage risks in process production systems. This multi-source data may include process operation data, equipment parameters, weak leakage symptom data, and environmental sensing data. The process operation data may include reactor temperature, pressure, flow rate, etc., which can be acquired through a distributed control system. The equipment parameters may include vibration amplitude, spectral characteristics, and drive motor current of key moving equipment, which can be obtained through direct equipment readings. The weak leakage symptom data may be obtained through distributed monitoring systems deployed near high-risk equipment or pipelines. The raw DAS waveform signal acquired by the acoustic sensor may also include the video stream acquired by the visual sensor, the semantic event sequence and corresponding confidence score generated after processing by the video analysis model, such as the confidence score of visible non-water vapor escaping from the sealing surface and video smoke. For example, the video analysis model may be a deep learning-based image recognition model, such as a convolutional neural network trained on frameworks like YOLO and Faster R-CNN, used to identify image information highly related to fire risk, such as smoke, flames or steam. The environmental perception data may include combustible gas concentration, which can be obtained through real-time monitoring data from combustible gas detectors deployed in the monitoring area.

[0028] S100 includes: First, based on multi-source fire supervision data, statistical features, dynamic features, spectral features, and nonlinear features are obtained within a sliding window.

[0029] By deploying a PTP server, time synchronization can be ensured for all data acquisition terminals, such as distributed control systems used to acquire process operation data and distributed acoustic sensors used to acquire weak leakage signs, thus avoiding time misalignment issues for multi-source data. Anti-aliasing filtering and resampling are performed on data with different sampling frequencies, such as 1000Hz DAS raw waveform signals and 1Hz process operation data. All parameters contained in the fire monitoring multi-source data are unified to a common sampling frequency, such as 10Hz, to achieve time alignment of multi-source heterogeneous data.

[0030] For low-dimensional continuous sequences with clear physical meaning, such as temperature, pressure, and drive motor current, data cleaning is performed directly to generate corresponding original parameter sequences. For example, for temperature, pressure, and drive motor current, robust Mahalanobis distance within a sliding window can be calculated. When outliers are detected, a Kalman smoother that integrates basic physical constraints, such as simple relationships based on mass or energy balance, can be used for repair, generating corresponding original parameter sequences, including temperature, pressure, and drive motor current sequences. For high-dimensional data streams such as DAS raw waveform signals, real-time feature extraction is performed first to obtain feature sequences, and then the feature sequences are cleaned. For example, with a window of 0.1 seconds, the mean signal energy of the 20k-50kHz ultrasonic band and the 20-200Hz infrasound band is calculated, generating two 10Hz feature sequences: ultrasonic energy sequence and infrasound energy sequence. Algorithms such as isolated forest are used to detect and repair univariate outliers on the feature sequences. The original parameter sequences and feature sequences are combined into a preprocessed dataset.

[0031] Using a sliding window of preset length and a fixed step size, each time-series variable in the preprocessed dataset, such as temperature and ultrasonic energy, is traversed to calculate a set of corresponding multidimensional features, including statistical features, dynamic features, spectral features, and nonlinear features. The statistical features may include the mean, variance, and skewness within the sliding window; the dynamic features may include the variance of the first-order difference and the zero-crossing rate; the spectral features may include the dominant frequency amplitude calculated through Fast Fourier Transform, and the energy proportion of specific frequency bands, such as characteristic frequency bands related to equipment failure; the nonlinear features are used to quantify the complexity and randomness of the original parameter sequence and feature sequence. Approximate entropy, permutation entropy, etc., are used to concatenate the multidimensional features of all time-series variables in each sliding window to form an initial high-dimensional feature vector. For example, if the sliding window is set to 30 seconds and the fixed step size is 1 second, and the preprocessed dataset contains 6 time-series variables, namely temperature, pressure, drive motor current, ultrasonic energy, infrasound energy, and video smoke confidence, the mean, variance, first-order difference variance, dominant frequency amplitude, and approximate entropy of each time-series variable are calculated within a 30-second sliding window. Then, each sliding window will generate an initial high-dimensional feature vector of 6 time-series variables * 5 features = 30 dimensions.

[0032] Then, feature selection is performed based on statistical features, dynamic features, spectral features, and nonlinear features using the maximum correlation and minimum redundancy criterion.

[0033] To ensure that the initial high-dimensional feature vector can focus on key indicators, specifically data that characterize fire risk indicators, a maximum correlation minimum redundancy algorithm is used to filter the initial high-dimensional feature vector. This can be achieved by first obtaining historical normal operation data and known fault event data from historical databases, such as long-running SCADA systems or event logs. The maximum correlation minimum redundancy algorithm uses this data as a reference to evaluate the correlation between each feature and the target state, such as an oil pipeline leak event, while also considering the redundancy between features. Through iterative selection, a subset of relevant features with the highest correlation to the target state and the lowest redundancy is retained. For example, continuing the previous example, evaluating the aforementioned 30-dimensional initial high-dimensional feature vector using the maximum correlation minimum redundancy algorithm reveals that features such as temperature variance, approximate entropy of ultrasonic energy, and first-order difference variance of drive motor current are highly correlated with oil pipeline leak events, while there is high redundancy between the mean pressure and the mean temperature. The maximum correlation minimum redundancy algorithm may then select a subset of relevant features containing 15 features.

[0034] Finally, the selected features are subjected to nonlinear dimensionality reduction using kernel principal component analysis to obtain the fire-fighting time-series feature vector.

[0035] Since precursors to fire and explosion risks may manifest in the nonlinear structure of data, kernel principal component analysis (KPCA) can be used to perform nonlinear dimensionality reduction on relevant feature subsets. This involves mapping the relevant feature subsets to a high-dimensional feature space using kernel functions such as radial basis functions and kernels, and then performing linear principal component analysis in the high-dimensional feature space to capture the nonlinear relationships within the relevant feature subsets. The top N kernel principal components, which retain a certain proportion of the total variance of the relevant feature subsets (e.g., more than 95%), are selected to ultimately output a low-dimensional, high-information-density fire-related time-series feature vector. For example, continuing with the previous example, performing KPCA on the aforementioned 15-dimensional relevant feature subsets shows that the first 8 kernel principal components can retain more than 95% of the total variance. Therefore, the 15-dimensional feature subset corresponding to each sliding window is dimensionality reduced and transformed into an 8-dimensional fire-related time-series feature vector.

[0036] In summary, the S100 effectively integrates multi-source heterogeneous data through time synchronization and multi-path processing strategies, ensuring the quality of the preprocessed dataset. By integrating multi-dimensional feature extraction and data dimensionality reduction based on the maximum correlation minimum redundancy algorithm and kernel principal component analysis, it extracts high-dimensional continuous feature vectors from massive amounts of raw data that can comprehensively characterize the dynamic evolution of fire risk. The fire time series feature vector integrates the statistical, dynamic, spectral, and nonlinear features of multi-source data, providing high-quality input for subsequent proactive fire early warning.

[0037] S200: Input the fire-fighting time-series feature vector into the instability early warning strategy based on chaos identification, and generate a nonlinear instability early warning signal based on the instability early warning strategy.

[0038] S200 includes: First, the process disorder index sequence is extracted from the fire-fighting time-series feature vector.

[0039] Historical normal operation data can be obtained through historical databases, such as long-running SCADA systems or event logs. This data is then preprocessed and feature extracted using the S100 algorithm to obtain historical fire-fighting time-series feature vectors. These vectors can be used to train a single-class support vector machine (SVM) model. This model learns the distribution boundary of the historical fire-fighting time-series feature vectors in the feature space under normal operating conditions, defining a normal state region. For the fire-fighting time-series feature vectors obtained in S100, the signed distance to the decision boundary in the trained SVM model is calculated. This signed distance characterizes the degree to which the current process production system deviates from the normal state region; a larger signed distance indicates a more unstable process production system. After standardizing this signed distance using methods such as Z-score, a process disorder index sequence is obtained. This process disorder index sequence, as a comprehensive, single-dimensional index, represents the multi-dimensional abnormal information contained in the fire-fighting time-series feature vectors. The upward trend of its value directly reflects the degree to which the process production system is deviating from its inherent stable operating mode and evolving towards a disordered or unstable state.

[0040] Then, the critical slowing phenomenon is monitored on the process turbulence index sequence to obtain the critical slowing index, which includes autocorrelation time, autocorrelation variance, and low-frequency noise intensity parameters. When the process production system approaches the critical transition point of fire risk abrupt change, such as leakage or fire, its recovery speed from minor disturbances will slow down. This phenomenon is called critical slowing. Therefore, by monitoring the critical slowing phenomenon of the process turbulence index sequence, early evidence of the decline in the resilience of the process production system can be provided.

[0041] Obtain the autocorrelation function based on the process disorder degree index sequence, and obtain the autocorrelation time and autocorrelation variance by fitting an exponential decay curve based on the autocorrelation function. Specifically, calculate the autocorrelation function of the process disorder degree index sequence within a sliding window, such as a sliding window with a length of 5 minutes. The autocorrelation function describes the similarity of the process disorder degree index sequence with itself at different time delays and is a key tool for analyzing the internal dependence structure of the process disorder degree index sequence. For example, for a time series sample {x_1, x_2,..., x_M} with zero mean and a finite length of M within a sliding window, the estimation formula for the autocorrelation function ρ(τ) with a time delay of τ is ρ(τ) = (Σ_{t = 1}^{M - τ}(x_t * x_{t + τ})) / (Σ_{t = 1}^{M}x_t^2), where M represents the total number of data points within the sliding window, x_t represents the value of the process disorder degree index sequence at time t, x_{t + τ} represents the value of the process disorder degree index sequence at time t + τ, τ is an integer, and 0 ≤ τ < M. This formula quantifies the similarity of the process disorder degree index sequence with itself when the time delay is τ. When the critical slowing down phenomenon occurs, the memory of the process disorder degree index sequence is enhanced, and the decay rate of the autocorrelation function ρ(τ) will slow down. The data points of the calculated ρ(τ) from τ = 0 to a preset maximum value, such as M / 4, can be fitted to an exponential decay curve by nonlinear least squares method. The decay time constant extracted from this exponential decay curve is the autocorrelation time. The increase in the autocorrelation time characterizes the decline of the resilience of the process production system and the tendency of the process production system state to become unstable. At the same time, the sequence variance within this sliding window can be obtained from the fitting residuals or directly calculated and used as the autocorrelation variance. The increase in the autocorrelation variance is also a characteristic of critical slowing down.

[0042] Obtain the power spectrum based on the process disorder degree index sequence and monitor the low-frequency band noise intensity parameter in the power spectrum. Specifically, perform a Fourier transform on the process disorder degree index sequence within the same sliding window to obtain its power spectrum. Near the critical transition point, the response of the process production system to disturbances will increase in the low-frequency part, specifically manifested as an increase in the intensity of the power spectrum in the low-frequency band. The low-frequency band can be a frequency band with a frequency lower than 1 / 10 of the main frequency. The main frequency is the frequency corresponding to the peak in the power spectrum. By calculating this low-frequency band noise intensity parameter, such as integrating the power spectrum of the frequency band from 0 to 0.1 Hz, this phenomenon can be quantified. The increase in the low-frequency band noise intensity parameter is a spectral characteristic of the impending instability of the process production system state.

[0043] Simultaneously, the process disorder index sequence is reconstructed in phase space, and geometric dynamics indices are obtained based on the phase space reconstruction results. These geometric dynamics indices include the maximum Lyapunov exponent, correlation dimension, and recursive quantification analysis indices. Phase space reconstruction and critical slowdown monitoring are performed concurrently, aiming to capture instability precursors from the perspective of the deep geometric structure of the dynamic evolution of the process production system. The catastrophic process of a complex industrial system can be regarded as a sudden change in the morphology of its phase space attractor. By reconstructing the phase space of a one-dimensional process disorder index sequence, the potential high-dimensional dynamic structure of the process production system can be recovered. Furthermore, the instability tendency of the process production system can be quantified by analyzing the geometric characteristics of the trajectories in the reconstructed phase space, such as divergence, complexity, and recursive patterns.

[0044] Based on the process disorder index sequence, the average mutual information method is used to obtain the delay time, and the false nearest neighbor method is used to obtain the embedding dimension. Based on the delay time and the embedding dimension, Takens embedding theorem is used to reconstruct the phase space of the process disorder index sequence and obtain the reconstructed phase space trajectory. The goal of phase space reconstruction is to find suitable parameters to embed the one-dimensional process disorder index sequence into a high-dimensional space, so that the reconstructed trajectory in the high-dimensional space, i.e., the reconstructed phase space trajectory, can truly reflect the topological characteristics of the process production system.

[0045] In the process of reconstructing the phase space trajectory, the average mutual information method is first used to obtain the delay time S. By evaluating the statistical dependence between the process disorder index sequence and its delayed process disorder index sequence after the delay time S, the average mutual information under different delay times S is calculated. The delay time S that makes the average mutual information reach its first local minimum can be selected as the delay time S, so that the process disorder index sequence and the delayed process disorder index sequence maintain a certain correlation and have the greatest independence, avoiding information redundancy. Then, the false nearest neighbor method is used to obtain the embedding dimension m. The false nearest neighbor method determines the embedding dimension by judging whether the embedding dimension is sufficient. If the embedding dimension m is large enough, then the points in the m-dimensional phase space should be true nearest neighbors. If the embedding dimension m is too small, through projection, points that were not originally adjacent will appear as neighbors in the lower-dimensional phase space. If the points appear to be adjacent, they become false nearest neighbors. Starting from m=1, the embedding dimension is gradually increased. It is checked whether the nearest neighbor of each point in the m-dimensional space is still its nearest neighbor when the dimension is increased to m+1. If not, it is considered a false nearest neighbor. When the proportion of false nearest neighbors is lower than a preset false nearest neighbor threshold, such as 5%, the corresponding m is the minimum sufficient embedding dimension, and the embedding dimension m is obtained. The m-dimensional space at this time is fixed as the reconstructed phase space. According to Takens' embedding theorem, based on the determined delay time S, embedding dimension m, and reconstructed phase space, the one-dimensional process disorder index sequence is mapped to a trajectory point sequence in an m-dimensional space. Each trajectory point represents the multi-dimensional state of the process production system at a certain moment, and the entire trajectory point sequence describes the continuous evolution path of the process production system state in the reconstructed phase space, that is, the reconstructed phase space trajectory.

[0046] Based on the reconstructed phase space trajectory, the Wolf algorithm is used to obtain the maximum Lyapunov exponent. The maximum Lyapunov exponent characterizes the average rate at which two infinitely close trajectories in the reconstructed phase space diverge exponentially with time. It can be used to quantify the degree of dynamic chaos and the limit of predictive ability in a process production system. Using the Wolf algorithm, an initial point and a neighboring point are selected on the reconstructed phase space trajectory. The distance between the two points is recorded as the first distance. The evolution of these two points is tracked as the trajectory point index increases with the evolution step size. After k evolution steps, the latest distance between the two points is recorded as the second distance. To avoid the distance increasing too rapidly and causing inaccuracy in the linear region, when the second distance exceeds a preset upper limit threshold, a new, closer neighboring point is found near the evolved initial point along the current evolution direction, and this new distance is recorded. The distance growth rate is iteratively repeated until the total evolutionary step size covered a sufficiently long preset time range, such as at least 500 evolutionary steps, to obtain stable statistics, or until a preset maximum number of iterations, such as 1000 neighbor search operations, is reached. The distance growth rate before each evolution or neighbor search is recorded, and the logarithmic mean of all distance growth rates is calculated as the maximum Lyapunov exponent. When the maximum Lyapunov exponent is greater than zero, it indicates that the process production system has chaotic characteristics that are sensitive to initial conditions. In the instability precursor stage, the maximum Lyapunov exponent may start from a small positive value close to zero and continue to increase. This indicates that the instability and unpredictability of the process production system dynamics are enhanced. Small disturbances, such as equipment micro-vibration and flow pulsation, may be rapidly amplified, which is a strong dynamic signal that the process production system is about to experience a sudden change in fire risk.

[0047] Based on the reconstructed phase space trajectory, the correlation dimension is obtained by analyzing the slope of the linear region in the double logarithmic graph of the correlation integral and distance. The correlation dimension is used to describe the complexity and fractal characteristics of the phase space attractor geometry, and characterizes the effective degrees of freedom in the process production system. For all trajectory points in the reconstructed phase space trajectory {Y_i}, all possible trajectory point pairs are considered, and the Euclidean distance between each pair of points is calculated. The correlation integral C(r) is defined, and its value is specifically the proportion of trajectory point pairs whose distance is less than a given radius r. The given radius r is gradually increased from the minimum value to the maximum value, and the corresponding correlation integral C(r) is calculated for each pair. The minimum value can be the minimum distance of all trajectory point pairs or a very small positive number, and the maximum value can be the maximum size of the reconstructed phase space trajectory or the median distance of all point pairs, to ensure that the variation of C(r) in the interval between the maximum and minimum values ​​can be captured. Log(r) is used as the horizontal axis, and log[ Using C(r)] as the ordinate, a double logarithmic coordinate system diagram is constructed, i.e., a double logarithmic graph. A linear region can be identified in this double logarithmic graph. By fitting a straight line to the data points within this restricted region, the slope of the obtained straight line is the critical dimension. In the process of the process production system evolving from stable operation to unstable state, the dynamics of the process production system may change from a high-dimensional, complex chaotic attractor to a lower-dimensional, simpler deterministic behavior, i.e., the correlation dimension continuously decreases. For example, assuming the process production system is specifically a catalytic cracking unit, about 20 minutes before the leakage event of the catalytic cracking unit, the correlation dimension continuously decreases from about 2.1 during stable operation to about 1.6. The decrease in the correlation dimension means that the effective degrees of freedom of the process production system dynamics are decreasing, and the behavior of the process production system tends to a simpler but also less stable operating mode. The continuous decrease in the correlation dimension is the geometric representation of the process production system approaching the critical transition point.

[0048] A recursive graph is generated based on the reconstructed phase space trajectory, and deterministic and laminar flow indicators are extracted as recursive quantitative analysis indicators. Specifically, an A×A matrix graph R, i.e., the recursive graph, is generated based on the reconstructed phase space trajectory {Y_i}, where A is the total number of trajectory points. For a pixel with coordinates (i, j) in the recursive graph, if the Euclidean distance between its corresponding trajectory point pair (Y_i, Y_j) in the reconstructed phase space is less than a preset critical distance, then the process system state is considered revisitable within this critical distance, and the pixel is assigned a value of 1; otherwise, it is assigned a value of 0. Pixels with a value of 1 are used as recursive points. The critical distance is usually set as a fixed proportion of the diameter or standard deviation of the reconstructed phase space trajectory, for example... To ensure the recursion point density is within a reasonable range, such as 1% to 10%, for the assigned recursion graph, calculate the number of all recursion points that form continuous diagonals and whose length is greater than the minimum length threshold, such as 2. Calculate the proportion of the number of determined recursion points to the total number of all recursion points and use this as determinism. High determinism indicates deterministic dynamics of the process production system, such as strong periodicity. Calculate the number of all laminar recursion points that form continuous vertical line segments and whose length is greater than the minimum length threshold in the recursion graph. Calculate the proportion of the number of laminar recursion points to the total number of all recursion points and use this as laminarity. High laminarity indicates that the process production system is stuck in a state of stagnation, i.e., frequently falling into temporary steady or metastable states.

[0049] Understandably, before a process production system becomes unstable, its deterministic dynamic behavior may be disrupted, and it may frequently fall into state stagnation. Therefore, the decrease in determinism and the increase in laminarity can serve as auxiliary judgment criteria for the loss of dynamic regularity and the stagnation of evolution in the process production system. Thus, determinism and laminarity can be mutually verified by the changes in the maximum Lyapunov index and correlation dimension, and together they constitute geometric dynamic indicators.

[0050] Finally, the critical slowdown index and the geometric dynamic index are merged into a comprehensive instability index, and a nonlinear instability early warning signal is generated based on the comprehensive instability index.

[0051] For the comprehensive instability index, a Bayesian change point detection algorithm is applied to identify the starting point of the growth trend of the comprehensive instability index. Specifically, the critical slowing index, including autocorrelation time, autocorrelation variance, and low-frequency noise intensity, and the geometric dynamic index, including the maximum Lyapunov exponent, correlation dimension, determinism, and laminarity, are fused together. Since the physical meaning and dimensions of the above index sequences are different, each index sequence is first standardized, for example, using Z-score standardization based on recent historical windows, such as the mean and standard deviation of the past 24 hours, so that all index sequences are numerically compatible. The system is comparable and can adapt to the slow drift of operating conditions. The greater the numerical fluctuation of a certain indicator sequence within the recent historical window, the smaller its information entropy, indicating that the indicator sequence is more abnormal at present. Therefore, at the current decision moment, the indicator sequence should be given a higher weight. Thus, the information entropy of each indicator sequence within the sliding window is calculated first, and then its corresponding weight is obtained. All standardized indicator sequences are weighted and summed to generate a comprehensive instability index. The comprehensive instability index is a real-time, dimensionless scalar sequence. Its absolute value and trend of change jointly characterize the overall instability risk level of the process production system.

[0052] To distinguish between short-term fluctuations and continuous trends in the comprehensive instability index, a Bayesian change point detection algorithm is applied. This algorithm models the comprehensive instability index as a piecewise constant process and assumes that the mean of the comprehensive instability index undergoes a sudden change at an unknown change point. By calculating the posterior probability of each time t given all comprehensive instability indices, the Bayesian change point detection algorithm can identify the change point most likely to exhibit a continuous trend. The change point with the highest posterior probability is identified as the starting point of the current growth trend of the comprehensive instability index. This starting point provides an objective time reference for subsequent judgments on whether the current trend is continuous.

[0053] Historical comprehensive instability index is obtained based on fire-fighting time-series feature vectors. An adaptive instability threshold is set based on the statistical distribution of the historical comprehensive instability index. Specifically, historical normal operation data is selected from the historical fire-fighting time-series feature vectors, and the historical comprehensive instability index is repeatedly obtained based on the historical normal operation data. The historical comprehensive instability index represents the inherent and normal fluctuation range of the process production system under normal operation. Statistical analysis is performed on the historical comprehensive instability index to calculate a specific high quantile of its distribution, and the value of this high quantile is used as the adaptive instability threshold. For example, the 99.5th percentile of the historical comprehensive instability index is used as the adaptive instability threshold. That is, under historical normal conditions, only 0.5% of the time points will the comprehensive instability index naturally fluctuate to exceed the adaptive instability threshold. The adaptive instability threshold can filter out false alarms caused by normal fluctuations, while maintaining high sensitivity to true abnormal instability.

[0054] When the comprehensive instability index is greater than or equal to the adaptive instability threshold, it is determined whether the comprehensive instability index is in an upward trend based on the starting point of the growth trend. If it is in an upward trend, a nonlinear instability early warning signal is generated. This nonlinear instability early warning signal includes at least the early warning time, major contributing indicators, risk units, and equipment parameters. Specifically, the final early warning trigger follows a dual-judgment logic to further improve the reliability of the early warning. The first condition of the dual-judgment logic is a threshold criterion, specifically, the real-time monitored comprehensive instability index being greater than or equal to the adaptive instability threshold. This indicates that the current instability risk level of the process production system has exceeded the upper limit of its historical normal fluctuation range. The second condition of the decision logic is the trend criterion. Specifically, based on the most recent starting point of the growth trend identified by Bayesian change point detection, the evolution of the comprehensive instability index within the interval from the starting point of the growth trend to the current time t is analyzed. By performing linear regression on this interval, the trend slope is calculated. If the trend slope is greater than 0, it is determined that the current comprehensive instability index is in an growth trend, rather than a short-term normal fluctuation. When both conditions one and two are satisfied, the process production system is determined to be in a critical state of nonlinear instability, and a structured nonlinear instability early warning signal is automatically generated immediately. The nonlinear instability early warning signal includes at least the warning time, main contributing indicators, risk units, and equipment parameters.

[0055] The warning time is specifically defined as the moment when the comprehensive instability index first simultaneously meets both conditions one and two. The main contributing indicators can be analyzed at the warning time by considering the contribution of various indicator sequences, including autocorrelation time, autocorrelation variance, low-frequency noise intensity, maximum Lyapunov exponent, correlation dimension, determinism, and laminarity, to the comprehensive instability index. Based on the deviation of each indicator sequence's standardized value from its historical normal mean, and combined with its corresponding weights, one or two indicator sequences with the largest contribution and their changing directions can be listed. For example, an increase in the maximum Lyapunov exponent and an increase in autocorrelation time provide direct physical or dynamic clues for manual assessment. The location of risk units is a multi-factor fusion reasoning process. By analyzing the anomaly types indicated by the main contributing indicators—for example, an increase in the Lyapunov exponent suggesting dynamic system instability, or a sudden increase in acoustic energy in a specific frequency band suggesting possible leakage or cracking—and combining the high-dimensional time-series feature vector generated by S100, the source of the original variable or feature sequence that has the greatest impact on the main contributing indicators can be extracted. For example, the feature with the largest contribution originates from the installation of... The 20-50kHz ultrasonic energy sequence generated by the distributed acoustic sensor at the bottom flange of reactor R-101 is then queried in a pre-built fire supervision process knowledge base. For example, assuming a fire supervision scenario in a petrochemical plant, the fire supervision process knowledge base can be constructed by integrating the digital twin model of the monitoring plant, EAM data from the asset management system, and historical maintenance records. This clarifies the topological relationships between monitoring points, equipment, and process units throughout the plant, as well as typical fault-symptom mapping rules. The anomaly type and characteristic source are matched and inferred in the fire supervision process knowledge base, thereby mapping abstract indicator anomalies to specific process units. For example, the matching and inference result of the bottom flange and ultrasonic energy anomaly of reactor R-101 is mapped to the fault mode of minor leakage in the reactor feed pipeline, which is associated with the specific process unit of the catalytic cracking unit reactor feed unit. The specific equipment parameters are further associated with specific equipment identifiers, such as the quench water pump P-101A, based on the risk unit location, which can be used to guide maintenance personnel to the specific site location for verification.

[0056] In summary, by integrating multiple heterogeneous indices obtained from two different theoretical perspectives—critical slowdown and geometric dynamics—into a single, comprehensive instability index with clear decision-making significance, and based on dual statistical judgments of threshold exceeding and trend confirmation, a structured early warning signal containing early warning time, main contributing indicators, risk units, and equipment parameters is automatically generated. This transforms the complex analytical conclusions of the preceding steps into intuitive early warning outputs, enabling the forward-looking identification of nonlinear instability states in the process production system.

[0057] S300: Based on nonlinear instability early warning signals, real-time environmental data is acquired. Based on the real-time environmental data, a pre-trained Fourier neural operator model is used to perform deduction to obtain a set of physical field deduction sequences. Based on the set of physical field deduction sequences, a fire risk alarm is generated.

[0058] The S300 includes: Based on the risk unit in the nonlinear instability early warning signal, the corresponding simplified model of the unit is called from the pre-stored geometric model library. Combined with the equipment parameters and real-time environmental data in the nonlinear instability early warning signal, the risk source parameters are obtained. The risk source parameters include the risk source location, risk source intensity and material properties.

[0059] Based on the risk units in the nonlinear instability early warning signal, such as the reactor feed unit of a catalytic cracking unit, the simplified unit model corresponding to the risk unit is called from the pre-stored geometric model library. The geometric model library stores the three-dimensional simplified geometric representation of each functional unit of the plant. It can be obtained by dividing and simplifying the outline of the digital twin model of the plant according to the functional unit and other standards. It can be stored in STL or OBJ format. The simplified unit model retains the shape of the main obstacles such as equipment, pipe racks, and buildings, and is used to define the physical boundary of fluid flow.

[0060] Combining the equipment parameters in the warning signal, such as the P-101A quench water pump, the preset three-dimensional coordinates bound to the equipment parameters are queried in the coordinate system of the called simplified unit model and used as the location of the risk source. If the warning signal is triggered by a data source with spatial positioning capabilities, such as a distributed acoustic sensor, the location of the risk source can be further calibrated using sound source positioning algorithms such as time difference positioning.

[0061] Combining the anomaly types indicated by the main contributing indicators in the early warning signal, such as an increase in the maximum Lyapunov exponent suggesting systemic instability, a sudden increase in acoustic energy in a specific frequency band suggesting a possible leak, and real-time environmental data and environmental perception data, the risk source intensity Q is obtained through an inversion estimation strategy. This inversion estimation strategy is a rapid estimation logic based on early signs and a simplified physical model. For example, based on this rapid estimation logic, if the main contributing indicators are associated with leak signs, such as specific acoustic characteristics, one or more combustible gas detectors located downwind of the risk source are selected, and their concentration readings over time are obtained as a sequence and used as the initial concentration ramp-up rate. Based on wind speed, wind direction, and atmospheric stability category in the real-time environmental data, the atmospheric diffusion parameters required for the simplified Gaussian point source model are determined. σ_y, σ_z), and back-calculated using the formula C(x, y, z) ≈ [Q / (2πσ_yσ_zu)] * exp[-y² / (2σ_y²) - z² / (2σ_z²)], where C is the detector concentration, x is the distance of the detector from the risk source in the downwind direction, u is the wind speed, and y and z are the downwind and vertical coordinates of the detector relative to the source. By solving this equation, the point source leakage intensity Q can be obtained. For example, for a hydrocarbon mixture of 50 g / s, if the main contributing indicators are related to abnormal thermal parameters, such as unstable temperature and pressure trends, then the design conditions and material properties of the equipment are queried according to the equipment parameters, and combined with the relevant time-series features in the fire-fighting time-series feature vector obtained in S100, such as the amplitude of the temperature change rate, a simplified model based on energy or mass conservation is used for estimation.

[0062] Based on risk units and equipment parameters, a pre-built material attribute database is used to determine the most likely leaking or combustion-involving dominant substance and its corresponding material attributes. This database is pre-built based on information such as the plant's process material balance sheet, piping and instrumentation diagrams, and safety data sheets, establishing a mapping relationship between equipment units and potentially hazardous substances. The material attributes include at least molecular weight, gas phase density under standard conditions, specific heat capacity at constant pressure, heat of combustion, lower explosive limit, and critical temperature and critical pressure used for equation of state calculations. For example, for a leak in the feed system of a catalytic cracking unit, the associated dominant substance is a C3 / C4 light hydrocarbon mixture, whose material attributes include an average molecular weight of approximately 50 g / mol, a gas phase density of approximately 2.0 kg / m³, a lower explosive limit of approximately 1.5% vol, and a heat of combustion of approximately 46 MJ / kg.

[0063] Based on the simplified unit model and real-time environmental data, the inference computational domain is obtained and a mesh is generated to form an initial-boundary condition vector.

[0064] Centered on the location of the risk source, and based on a preset warning range (e.g., a radius of 200 meters and the prevailing wind direction), a rectangular computational domain is defined in three-dimensional space to ensure coverage of critical areas potentially affected by the disaster, such as adjacent equipment and control rooms. Within this computational domain, an unstructured, body-fitted computational mesh is automatically generated based on a simplified unit model. The mesh cells are densified near the risk source and around key obstacles to capture gradients and flow around them. Simultaneously, real-time environmental data is acquired, including parameters such as current wind speed, wind direction, ambient temperature, atmospheric pressure, and atmospheric stability. This real-time environmental data can be obtained through internet-based meteorological platforms, etc. Initial values ​​for the physical field are set at all nodes of this body-fitted computational mesh. The initial concentration field value of all nodes in the entire computational domain can be set to 0, and the initial temperature value can be set to the ambient temperature in the real-time environmental data. Risk source values ​​are assigned to the nodes where the risk source is located. Based on the risk source intensity and material properties, the initial release conditions are calculated by a simplified model. For example, for point source leakage, the initial concentration value C_0 can be estimated according to the formula C_0≈Q / (ρ*A*v), where ρ is the gas phase density of the leaked substance under the release conditions, A is the equivalent release port area estimated based on equipment parameters such as flange size and crack assumptions, and v is the initial release rate estimated based on Bernoulli's principle. C_0 is assigned as the initial concentration field value of the node. At the same time, based on the characteristics of the leaked substance, such as whether it is a high-temperature material or whether flash evaporation may occur, an initial temperature value higher than the ambient temperature is set for the node.

[0065] Based on real-time environmental data, velocity inlet boundary conditions are applied to the inlet boundary of the computational domain. The inlet boundary is determined based on the wind direction in the real-time environmental data. The velocity inlet boundary conditions may include wind speed and wind direction. Pressure outlet boundary conditions are set at the outlet boundary of the computational domain. No-slip wall boundary conditions are set on the obstacle surface and ground represented by the simplified unit model. Free flow boundary conditions are set on other open sides or top surfaces of the computational domain. The velocity inlet boundary conditions, pressure outlet boundary conditions, no-slip wall boundary conditions, and free flow boundary conditions together constitute the boundary conditions. The boundary conditions and the initial physical field conditions are jointly encoded into a high-dimensional, structured data volume, namely the initial-boundary condition vector. The initial-boundary condition vector completely represents the initial state of the physical field and its external constraints at the start of the deduction.

[0066] Based on material properties and real-time environmental data, a matching Fourier neural operator model is selected from a pre-built inference model library.

[0067] The inference model library contains multiple Fourier neural operator models that have been trained offline and are specifically designed for different types of fire disaster scenarios. Each Fourier neural operator model is implemented by a deep learning network. Its core function is to learn and replace the numerical solver of complex partial differential equations describing fluid flow, heat transfer, and mass transport, such as the Navier-Stokes equations and convection-diffusion equations, and then establish a nonlinear mapping relationship between the initial-boundary condition vector and the predicted sequence of physical fields for multiple consecutive time steps in the future.

[0068] For different types of fire disaster scenarios, such as light gas leakage and diffusion, heavy gas deposition, and pool fire combustion, risk source parameters and environmental data corresponding to different types of fire disaster scenarios can be obtained from industry standard materials, such as typical accident scenarios in the "Fire Protection Design Standard for Petrochemical Enterprises" and the set scenarios in safety evaluation reports. High-fidelity computational fluid dynamics software, such as OpenFOAM and ANSYS Fluent, can be used to conduct large-scale numerical simulations by changing the risk source parameters and environmental data, generating simulation datasets. Each training sample in the simulation dataset is a data pair consisting of training input data and training output data. The input data is the initial-boundary condition vector under a specific scenario, and the output data is the concentration field sequence and temperature field sequence output in the computational domain at a fixed time step, such as 1 second, within a future period of time under that scenario, such as 60 minutes. The simulation dataset is randomly divided into training set, validation set, and test set, for example, according to a ratio of 70%:15%:15%.

[0069] A network architecture for a Fourier neural operator model is constructed, specifically comprising an input projection layer, multiple stacked Fourier neural operator layers, a linear transform layer, a residual connection layer, and an output projection layer. Using data pairs from the simulation training set, the network weights are iteratively adjusted through backpropagation and an optimizer such as Adam to minimize the loss function (e.g., mean squared error) between the model's predicted physics sequence and the output data of the simulation training set. After each training iteration, the performance of the current Fourier neural operator model on unseen simulation data is evaluated using a validation set. This is specifically achieved by calculating the loss function value on the validation set. When the validation set loss exceeds the value of the loss function on consecutive training iterations... If the Fourier neural operator model remains unchanged or increases during the training cycle, it indicates that the model may be overfitting. At this point, training should be terminated, and the model should be rolled back to the Fourier neural operator model with the lowest loss on the validation set. The trained Fourier neural operator model should be independently evaluated using the test set. If its prediction accuracy on the test set, such as the spatial correlation coefficient of the concentration field and the arrival time error, meets the preset engineering application requirements, then the Fourier neural operator model is successfully trained and can be added to the inference model library for later use. For example, the engineering application requirements can be set as follows: the spatial correlation coefficient of the concentration field prediction should be greater than 0.85; the time prediction error of the disaster cloud reaching the key location should be within ±1 minute; and the relative error of the peak concentration should be less than 20%.

[0070] The Fourier neural operator models in the inference model library are selected based on the dominant substance and atmospheric stability category. The dominant substance determines the fluid properties and combustion model, while the atmospheric stability category determines the turbulent diffusion characteristics. For example, assuming the dominant substance is C3 / C4 light hydrocarbons, and the real-time environmental data shows that the atmospheric stability is neutral type D, a matching light hydrocarbon-neutral Fourier neural operator model is selected from the inference model library to ensure the rationality of the physical basis of the inference.

[0071] The initial-boundary condition vector is input into the selected Fourier neural operator model. Through global convolution operation in the Fourier domain, the physical field prediction sequence for multiple consecutive time steps in a single forward propagation is directly output. The physical field prediction sequence includes the concentration field prediction index and the temperature field prediction index.

[0072] The initial-boundary condition vector, specifically a spatial discrete function defined on the nodes, is input into the selected Fourier neural operator model. In this model, an input projection layer (which can be a fully connected neural network) first maps the initial-boundary condition scalar at each node to a higher-dimensional latent feature space. This input projection layer aims to transform the original physical parameters, such as initial concentration, velocity, temperature, and boundary type encoding, into higher-dimensional features more conducive to the learning and expression of the Fourier neural operator model, forming an initial spatial latent feature field. This initial spatial latent feature field is then input into multiple consecutively stacked Fourier neural operator layers, for example, 4 to 8 layers, with each layer performing the same core operation. While operating independently, it possesses independent learnable parameters, enabling it to extract and fuse spatiotemporal features at different scales and levels of abstraction in a layered and progressive manner. Specifically, in each Fourier neural operator layer, a Fast Fourier Transform is performed on the initial spatial latent feature field of the input, transforming it from the spatial domain to the Fourier domain. In the Fourier domain, each frequency component is multiplied by a set of learnable complex convolutional kernel weights. This multiplication operation is mathematically equivalent to performing a global convolution operation in the spatial domain, which can model physical interactions spanning long distances between any two points in the computational domain, such as the instantaneous propagation of pressure waves and the long-distance diffusion effect of pollutants in turbulence. This is something that traditional convolutional neural networks, limited by their local receptive fields, cannot achieve.

[0073] Simultaneously, in the spatial domain, a local linear transformation is applied to the initial spatial latent feature field of the input. This can be achieved through a 1x1 convolution or a small fully connected network. This local linear transformation is used to capture and learn fine-grained feature changes and interactions within a local range. Subsequently, the global convolution result, recovered back to the spatial domain by inverse Fourier transform, is added to the result of the local linear transformation to form the total output of this Fourier neural operator layer. Finally, a nonlinear activation function, such as ReLU, is applied to the total output to introduce nonlinear expressive power, generating the latest spatial latent feature field for the next layer. The convolution result is added to a local linear transformation result to obtain the total output of the Fourier neural operator layer. This allows the Fourier neural operator model to simultaneously integrate global long-range correlations and local proximity effects, thereby more completely describing the multi-scale characteristics of physical field evolution. By stacking multiple Fourier neural operator layers, the Fourier neural operator model constructs a deep feature processing pipeline. Shallow layers may capture basic flow structures and diffusion patterns, while deeper layers can learn more complex and nonlinear spatiotemporal evolution laws, such as the propagation of flame fronts, the generation and dissipation of vortices, etc., thereby achieving a high-precision approximation of physical processes.

[0074] After processing by the final Fourier neural operator layer, a high-level latent feature field containing complex spatiotemporal evolution laws is obtained. In the output projection layer, this high-level latent feature field is mapped back from the high-dimensional latent feature space to specific physical quantity dimensions. The output projection layer directly outputs the physical field prediction sequence at each node within the entire computational domain, starting from the current moment and extending for T consecutive time steps. This physical field prediction sequence specifically includes concentration field prediction indices and temperature field prediction indices. It is important to emphasize that the specific physical meaning and units of these prediction indices are determined by the annotations of the simulation dataset. During training, the Fourier... The Fourier neural operator model learns the mapping relationship from an initial-boundary condition vector of a specific format to a concentration field sequence and a temperature field sequence of a specific format. Therefore, in the application of the simulation, as long as the input initial-boundary condition vector has the same physical meaning and data structure as the training data, the Fourier neural operator model can output a prediction sequence with consistent physical meaning. For example, assuming that the output data in the simulation dataset is volume concentration and thermal radiation flux, the Fourier neural operator model will output the predicted volume concentration and predicted thermal radiation flux at various spatial locations at multiple time steps in the future.

[0075] Perturbation sampling is performed on the parameters of potential risk sources to obtain multiple perturbation initial-boundary condition vectors. Matching Fourier neural operator models are selected to obtain multiple physical field prediction sequences. All physical field prediction sequences are merged into a physical field deduction sequence set, and fire risk consequence quantitative analysis is performed to generate fire risk alarms.

[0076] In the initial stages of fire hazards such as leaks and fires, the intensity of the risk source, Q, is a crucial uncertainty parameter. Based on engineering experience and historical data, a probability distribution can be assigned to the intensity Q. For example, Q can be assumed to follow a log-normal distribution with an estimated mean and a certain standard deviation to reflect its positive value characteristics and uncertainty range. Random sampling can be performed from this probability distribution to obtain multiple, such as 100, disturbance risk source intensity samples. For each disturbance risk source intensity sample, its initial-boundary condition vector is obtained, and its corresponding physical field prediction sequence is obtained through its matching Fourier neural operator model. These sequences are then combined into a physical field deduction sequence set, which contains multiple possible fire accident evolution paths.

[0077] Subsequently, a quantitative analysis of fire risk consequences is performed on the set of physical field deduction sequences. At each time step, based on all physical field prediction sequences in the set of physical field deduction sequences, the corresponding concentration exceedance probability distribution field and temperature exceedance probability distribution field are statistically analyzed, and the corresponding concentration risk envelope and temperature risk envelope are obtained.

[0078] Specifically, for each time step, based on a preset safety threshold, the concentration over-threshold probability of each node in all physical field prediction sequences at that time step is identified, forming a concentration over-threshold probability distribution field.

[0079] A corresponding safety threshold can be set based on the safety limit threshold of the leaked substance. For example, the safety threshold can be set as the lower explosive limit of the leaked substance to characterize the flammability risk, or as the concentration that immediately threatens life and health to characterize the health toxicity risk. For example, if the dominant substance in the leak is a C3 / C4 light hydrocarbon mixture with a lower explosive limit of 1.5% vol, the safety threshold can be set to 1.5% vol. When the concentration field prediction index of a certain node is greater than or equal to this safety threshold, there is a risk of forming a flammable gas cloud in the space corresponding to that node. The number of sequences exceeding the safety threshold in the concentration field prediction index of that node is counted, and the ratio of the number of sequences to the total number of physical field prediction sequences in the physical field inference sequence set is calculated. This ratio is used as the concentration exceeding the threshold probability of that node at time step t. By obtaining the corresponding concentration exceeding the threshold probability for all nodes, the concentration exceeding the threshold probability distribution field of the entire three-dimensional space at time step t can be obtained. The concentration exceeding the threshold probability distribution field characterizes the probability that any node in the space is within the flammable hazard concentration range at time step t.

[0080] Similarly, for each time step, based on the preset thermal radiation damage threshold, the temperature over-threshold probability of each node in all physical field prediction sequences at that time step is identified, forming a temperature over-threshold probability distribution field.

[0081] The thermal radiation damage threshold can be set based on fire safety standards. For example, for spaces where people may be present, the thermal radiation damage threshold can be set to 4 kW / m². At this radiation level, exposure for more than 20 seconds may cause second-degree burns. For spaces containing only equipment, the thermal radiation damage threshold can be set to 12.5 kW / m². At this radiation level, the equipment may be damaged. Similar to the calculation process of the concentration exceeding the threshold probability, for each node in the computational domain, at time step t, the temperature field prediction index of all physical field prediction sequences in the physical field extrapolation sequence set at that node is checked. The number of sequences that exceed the thermal radiation damage threshold is counted, and the ratio of this number to the total number of physical field prediction sequences in the physical field extrapolation sequence set is calculated. This ratio is used as the temperature exceeding the threshold probability at that node at time step t, thereby obtaining the temperature exceeding the threshold probability distribution field.

[0082] For the concentration exceeding the threshold probability distribution field and the temperature exceeding the threshold probability distribution field, the spatial boundaries under a specified confidence level are extracted respectively to form the concentration risk envelope and temperature risk envelope corresponding to the time step. The risk envelope represents the maximum spatial range of fire risk impact under the specified confidence level.

[0083] The confidence level setting is based on the risk acceptability criterion and the conservative requirements of emergency decision-making. In practical engineering applications, industry safety standards can be referenced, such as using the personal risk contour line as the boundary of the protected area and the response threshold of the emergency plan, and setting the confidence level to a relatively high fixed value, such as 90% or 95%. The core role of the confidence level is to make risk decisions. For example, a 90% confidence level indicates that, based on all current physical field prediction sequences, a spatial boundary can be defined, and it is believed that there is a 90% probability that the impact range of the fire disaster will not exceed this spatial boundary. By setting the confidence level, a statistically significant basis is provided for the delineation of the emergency evacuation range and the decision to shut down key facilities.

[0084] For each time step t, the concentration exceedance probability distribution field stores a concentration exceedance probability at each node, representing the likelihood that the predicted concentration field index at that node exceeds the safety threshold. Based on the confidence level, the corresponding probability threshold is calculated. For example, if the confidence level is 90%, the probability threshold can be 1-90%=10%. Then, all nodes with concentration exceedance probabilities greater than or equal to this probability threshold are found in the computational domain. All the found nodes form one or more connected point sets in three-dimensional space. The outermost continuous boundary surfaces of these point sets can be obtained using isosurface extraction algorithms such as MarchingCubes. These boundary surfaces are the concentration wind at the current time step t. The concentration risk envelope defines a three-dimensional spatial region where the probability of concentration exceeding the threshold is greater than or equal to 10%, indicating a significant risk that must be included in key monitoring areas. Outside this three-dimensional spatial region, the probability of concentration exceeding the threshold is less than 10%, considered a relatively safe zone at the current confidence level. By setting a high confidence level, such as 90%, and selecting a lower probability threshold, such as 10%, to delineate the spatial boundary, setting a high confidence level represents a conservative decision-making strategy, ensuring that most possible risk scenarios are included under uncertainty, thus avoiding losses due to underestimating the risk. In practical applications, this can be adjusted as needed.

[0085] Similarly, the temperature risk envelope is obtained based on the temperature over-threshold probability distribution field at each time step t.

[0086] When the concentration risk envelope or temperature risk envelope at any time step intrudes into the critical protection area, a fire risk alarm is generated. The fire risk alarm includes at least the alarm trigger time step, the concentration risk envelope, the temperature risk envelope, and the corresponding confidence information.

[0087] In the digital twin model of the monitored factory, define areas requiring priority protection, such as areas with open flame equipment, densely populated control rooms, plant boundaries, or sensitive environmental targets, and designate these as critical protection areas. The system monitors the concentration risk envelope and temperature risk envelope at each time step t in real time. When the concentration risk envelope intersects any critical protection area in space, or when the temperature field prediction index of the node covered by the temperature risk envelope exceeds the corresponding thermal radiation damage threshold within any critical protection area, a fire risk alarm is immediately generated. The fire risk alarm is a structured decision support information set, which includes at least the alarm trigger time step, the concentration risk envelope, the temperature risk envelope, and the corresponding confidence information. For example, suppose a fire risk alarm is generated at this time, the warning trigger time step is 8 minutes after the leak begins, the predicted disaster type is the risk of combustible gas cloud spreading to the ignition source and the threat of thermal radiation to personnel, the affected critical protection area is the heating furnace F-101 area, and the confidence level is 90%. At the same time, the concentration risk envelope and temperature risk envelope are provided in the form of three-dimensional graphics or two-dimensional projection diagrams to intuitively describe the specific range that the fire disaster may affect.

[0088] S400: The fire time-series feature vectors before and after the fire risk alarm are analyzed by time-varying causal network to obtain the causal transmission path, and a structured early warning diagnosis report is generated by combining nonlinear instability early warning signals and physical field deduction sequences.

[0089] Extract a preset time window before and after the fire risk alarm triggering time, such as the fire risk alarm timing feature vector within 30 minutes before and 10 minutes after the triggering time. These fire risk alarm timing feature vectors contain multi-dimensional feature variables of the entire fire risk alarm process. Causal inference analysis can be performed using time-varying causal network analysis algorithms such as PCMCI. The PCMCI algorithm can identify the causal driving relationship between variables over time through the PC stage and the MCI stage.

[0090] In the PC phase, the PCMCI algorithm examines all possible time delays within a preset maximum time delay. The time delay refers to the time span required for the causal variable to influence the outcome variable. The maximum time delay can be set to 5-10 time steps according to the time scale of the physical process. For each time delay, the PCMCI algorithm performs a conditional independence test on each pair of feature variables. By iteratively adding a condition set, i.e. the values ​​of other feature variables at historical moments, it determines whether two feature variables are independent under a given condition set. The PC phase is used to identify and eliminate spurious associations and initially construct a network framework containing potential causal connections.

[0091] In the MCI phase, on the constructed network framework, for each possible causal connection, such as from feature variable 1 at time t-τ to feature variable 2 at time t, an instantaneous conditional independence test is performed. After controlling for the influence of all other variables, including the historical value of feature variable 2 itself, the instantaneous conditional independence test determines whether the historical value of feature variable 1 still contributes independently to the current value of feature variable 2. If the instantaneous conditional independence test passes, a Granger causal relationship with a delay of τ from feature variable 1 to feature variable 2 is confirmed. Its causal strength can be quantified by test statistics, such as the partial correlation coefficient. By performing the above analysis on all feature variable pairs, the PCMCI algorithm finally outputs a dynamic causal network, where the causal nodes are the feature variables in each fire-fighting time-series feature vector, and their directed edges represent the direction of causal influence, the time delay of causal influence, and the strength of causal influence between feature variables.

[0092] In some possible embodiments, assuming that the PCMCI algorithm is used to analyze the fire-fighting time sequence feature vector before and after the fire risk alarm triggering time, a causal transmission path may be identified, specifically vibration feature (pump P-101A, t-3) - temperature feature (reactor bottom flange, t-1) - ultrasonic energy feature (20-50kHz, t) - gas concentration feature (detector GDT-01, t+2). This path indicates that the pump vibration abnormality at time t-3 is the early cause of a series of subsequent state changes. It first causes local temperature fluctuations at time t-1, which may then induce a small leak at time t to generate a specific frequency ultrasonic signal, and finally cause the accumulation of combustible gas at time t+2. The causal transmission path describes the logic and temporal sequence of the fire event transmission, and quantifies the causal strength of each link through the partial correlation coefficient of each edge.

[0093] Based on a predefined report template, a structured early warning diagnosis report is generated based on causal transmission paths, nonlinear instability early warning signals, and physical field deduction sequences. This report transforms multi-source analysis results into data with clear hierarchical structure. For example, the structured early warning diagnosis report may include a time summary, causal source analysis, and risk visualization and quantification. The time summary may include the early warning time of S200 and the fire risk alarm trigger time in S300. The causal source analysis may be a causal transmission path presented in the form of text and causal graphs, such as: increased vibration of quench pump P-101A (3 minutes ago) - temperature fluctuation of reactor bottom flange (1 minute ago) - sudden increase of 20-50kHz ultrasonic energy (current) - increase in reading of nearby combustible gas detector (2 minutes later). The risk visualization and quantification may refer to the concentration risk envelope and temperature risk envelope obtained by S300, provided in the form of three-dimensional graphics or two-dimensional projection diagrams.

[0094] A second aspect of the present invention provides a fire monitoring and alarm system based on big data, comprising: The fire data acquisition module is used to collect multi-source fire supervision data, and to preprocess and extract features from the multi-source fire supervision data to obtain a fire time series feature vector. A nonlinear instability early warning module is provided, which receives the fire-fighting time-series feature vector and executes an instability early warning strategy based on chaos identification to generate a nonlinear instability early warning signal. The fire risk physical field simulation module is used to respond to the nonlinear instability warning signal, call a pre-trained Fourier neural operator model based on real-time environmental data to perform rapid simulation, obtain a set of physical field simulation sequences, and perform quantitative analysis of fire risk consequences to generate a fire risk alarm. The causal analysis module is used to perform time-varying causal network analysis on the fire time-series feature vectors before and after the fire risk alarm to obtain the causal transmission path, and combine the nonlinear instability early warning signal with the physical field deduction sequence set to generate a structured early warning diagnosis report.

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fire monitoring and alarm method based on big data, characterized in that, The method includes: Acquire multi-source fire supervision data, and preprocess and extract features from the multi-source fire supervision data to obtain a fire time series feature vector; The fire-fighting time-series feature vector is input into the instability early warning strategy based on chaos identification, and a nonlinear instability early warning signal is generated based on the instability early warning strategy. Real-time environmental data is obtained based on nonlinear instability early warning signals. Based on the real-time environmental data, a pre-trained Fourier neural operator model is used to perform deduction to obtain a set of physical field deduction sequences. A fire risk alarm is generated based on the set of physical field deduction sequences. The fire risk alarm time-series feature vectors before and after the alarm are analyzed by time-varying causal networks to obtain the causal transmission path, and a structured early warning diagnosis report is generated by combining nonlinear instability early warning signals and physical field deduction sequences.

2. The fire monitoring and alarm method based on big data according to claim 1, characterized in that, Acquire multi-source fire supervision data, and preprocess and extract features from the multi-source fire supervision data to obtain a fire time-series feature vector, including: Based on multi-source fire supervision data, statistical features, dynamic features, spectral features and nonlinear features are obtained within a sliding window; Feature selection is performed based on statistical characteristics, dynamic characteristics, spectral characteristics, and nonlinear characteristics using the maximum correlation and minimum redundancy criterion. The selected features are subjected to nonlinear dimensionality reduction using kernel principal component analysis to obtain the fire-fighting time-series feature vector.

3. The fire monitoring and alarm method based on big data according to claim 1, characterized in that, The fire-fighting time-series feature vector is input into the instability early warning strategy based on chaos identification, and a nonlinear instability early warning signal is generated based on the instability early warning strategy, including: Extract the process disorder index sequence from the fire-fighting time-series feature vector; The critical slowdown phenomenon was monitored in the process disorder index sequence to obtain the critical slowdown index. The process disorder index sequence is reconstructed in phase space, and geometric dynamic indexes are obtained based on the phase space reconstruction results; The critical slowdown index and the geometric dynamic index are integrated into a comprehensive instability index, and a nonlinear instability early warning signal is generated based on the comprehensive instability index.

4. The fire monitoring and alarm method based on big data according to claim 3, characterized in that, The critical slowdown phenomenon is monitored in the process disorder index sequence to obtain the critical slowdown index. include: The critical slowing index includes autocorrelation time, autocorrelation variance, and low-frequency noise intensity parameters. The autocorrelation function is obtained based on the process disorder index sequence, and the autocorrelation time and autocorrelation variance are obtained by fitting an exponential decay curve based on the autocorrelation function. The power spectrum is obtained based on the process disorder index sequence, and the low-frequency noise intensity parameter in the power spectrum is monitored.

5. The fire monitoring and alarm method based on big data according to claim 3, characterized in that, The process disorder index sequence is reconstructed in phase space, and geometric dynamic indices are obtained based on the phase space reconstruction results. include: The geometric dynamics indicators include the maximum Lyapunov index, correlation dimension, and recursive quantification analysis indicators. Based on the process disorder index sequence, the average mutual information method is used to obtain the delay time, the false nearest neighbor method is used to obtain the embedding dimension, and the Takens embedding theorem is used to reconstruct the phase space of the process disorder index sequence based on the delay time and the embedding dimension, and the reconstructed phase space trajectory is obtained. Based on the reconstructed phase space trajectory, the Wolf algorithm is used to obtain the maximum Lyapunov exponent; Based on the reconstructed phase space trajectory, the correlation dimension is obtained by analyzing the slope of the linear region in the double logarithmic graph of correlation integral and distance; Recursive graphs are generated based on reconstructed phase space trajectories, and deterministic and laminar flow indices are extracted as recursive quantitative analysis indicators.

6. The fire monitoring and alarm method based on big data according to claim 3, characterized in that, The critical slowdown index and the geometric dynamics index are integrated into a comprehensive instability index. Based on this comprehensive instability index, a nonlinear instability early warning signal is generated, including: The Bayesian change point detection algorithm is applied to the comprehensive instability index to identify the starting point of the growth trend of the comprehensive instability index; Historical comprehensive instability index is obtained based on fire-fighting time-series feature vectors, and adaptive instability threshold is set based on the statistical distribution of historical comprehensive instability index; When the comprehensive instability index is greater than or equal to the adaptive instability threshold, it is determined whether the comprehensive instability index is in an upward trend at the current moment based on the starting point of the growth trend. If it is in an upward trend, a nonlinear instability warning signal is generated. The nonlinear instability early warning signal includes at least the early warning time, main contributing indicators, risk units, and equipment parameters.

7. The fire monitoring and alarm method based on big data according to claim 1, characterized in that, Real-time environmental data is acquired based on nonlinear instability early warning signals. A pre-trained Fourier neural operator model is used to extrapolate from this data, resulting in a set of physical field extrapolation sequences. A fire risk alarm is then generated based on this set of sequences, including: Based on the risk unit in the nonlinear instability early warning signal, the corresponding simplified model of the unit is called from the pre-stored geometric model library. Combined with the equipment parameters and real-time environmental data in the nonlinear instability early warning signal, the risk source parameters are obtained. The risk source parameters include the risk source location, risk source intensity and material properties. Based on the simplified unit model and real-time environmental data, the inference computational domain is obtained and a mesh is generated to form an initial-boundary condition vector. Based on material properties and real-time environmental data, a matching Fourier neural operator model is selected from a pre-built inference model library. The initial-boundary condition vector is input into the selected Fourier neural operator model. Through global convolution operation in the Fourier domain, the physical field prediction sequence for multiple consecutive time steps in a single forward propagation is directly output. The physical field prediction sequence includes the concentration field prediction index and the temperature field prediction index. Perturbation sampling is performed on the parameters of potential risk sources to obtain multiple perturbation initial-boundary condition vectors. Matching Fourier neural operator models are selected to obtain multiple physical field prediction sequences. All physical field prediction sequences are merged into a physical field deduction sequence set, and fire risk consequence quantitative analysis is performed to generate fire risk alarms.

8. The fire monitoring and alarm method based on big data according to claim 7, characterized in that, All physical field prediction sequences are merged into a physical field derivation sequence set, and a quantitative analysis of fire risk consequences is performed to generate a fire risk alarm, including: At each time step, based on all physical field prediction sequences in the physical field deduction sequence set, the corresponding concentration exceedance probability distribution field and temperature exceedance probability distribution field are statistically analyzed, and the corresponding concentration risk envelope and temperature risk envelope are obtained: When the concentration risk envelope or temperature risk envelope at any time step intrudes into the critical protection area, a fire risk alarm is generated. The fire risk alarm shall include at least the alarm trigger time step, concentration risk envelope, temperature risk envelope and corresponding confidence information.

9. The fire monitoring and alarm method based on big data according to claim 8, characterized in that, At each time step, based on all physical field prediction sequences in the physical field deduction sequence set, the corresponding concentration exceedance probability distribution field and temperature exceedance probability distribution field are statistically analyzed, and the corresponding concentration risk envelope and temperature risk envelope are obtained, including: For each time step, based on a preset safety threshold, the concentration over-threshold probability of each node in all physical field prediction sequences at that time step is identified, forming a concentration over-threshold probability distribution field; For each time step, based on the preset thermal radiation damage threshold, the temperature over-threshold probability of each node in all physical field prediction sequences at that time step is identified, forming a temperature over-threshold probability distribution field. For the concentration exceeding the threshold probability distribution field and the temperature exceeding the threshold probability distribution field, the spatial boundaries under a specified confidence level are extracted respectively to form the concentration risk envelope and temperature risk envelope corresponding to the time step. The risk envelope represents the maximum spatial range of fire risk impact under the specified confidence level.

10. A fire monitoring and alarm system based on big data, applied to the method described in any one of claims 1 to 9, characterized in that, include: The fire data acquisition module is used to collect multi-source fire supervision data, and to preprocess and extract features from the multi-source fire supervision data to obtain a fire time series feature vector. A nonlinear instability early warning module is provided, which receives the fire-fighting time-series feature vector and executes an instability early warning strategy based on chaos identification to generate a nonlinear instability early warning signal. The fire risk physical field simulation module is used to respond to the nonlinear instability warning signal, call a pre-trained Fourier neural operator model based on real-time environmental data to perform rapid simulation, obtain a set of physical field simulation sequences, and perform quantitative analysis of fire risk consequences to generate a fire risk alarm. The causal analysis module is used to perform time-varying causal network analysis on the fire time-series feature vectors before and after the fire risk alarm to obtain the causal transmission path, and combine the nonlinear instability early warning signal with the physical field deduction sequence set to generate a structured early warning diagnosis report.

Images