A building three-dimensional risk evolution evaluation and early warning method for a whole life cycle of fire

Abstract

This invention proposes a three-dimensional risk evolution assessment and early warning method for buildings throughout the entire fire lifecycle, belonging to the field of building fire safety assessment technology. The method includes: constructing a hierarchical risk index system; collecting historical building fire data and classifying and preprocessing it to obtain training datasets for different sub-dimensions; constructing corresponding support vector machine (SVM) classification models; optimizing the hyperparameters of the three sub-supplier SVM models using a particle swarm optimization algorithm; inputting real-time collected building feature data into the three optimal parameterized sub-supplier SVM models to construct real-time risk state points in a three-dimensional risk space; calculating the overall risk index; and comparing the overall risk index with a classification threshold to obtain a fire risk early warning signal. This invention achieves feature-driven self-assessment under multi-source data fusion, precisely characterizes the risks throughout the entire fire lifecycle, and realizes scientific dynamic classification early warning.

Description

Technical Field

[0001] This invention relates to the field of building fire safety assessment technology, and in particular to a three-dimensional risk evolution assessment and early warning method for buildings throughout the entire fire life cycle. Background Technology

[0002] Building fires are a major threat to urban public safety. With the acceleration of urbanization, building types are becoming increasingly diverse, and the degree of functional integration within buildings is constantly increasing, leading to a multi-source coupling and dynamic evolution of building fire risks. The occurrence and spread of building fires often involve multiple risk factors, including ignition causes, the condition of fire protection facilities, and the level of personnel management. These dimensions are interconnected and mutually influential, making it difficult for a single-dimensional static evaluation to fully reflect the overall fire risk status of a building.

[0003] Currently, building fire risk assessment methods mainly include the analytic hierarchy process (AHP), fuzzy comprehensive evaluation, and data-driven methods based on machine learning. Traditional methods rely on expert experience in determining indicator weights, resulting in strong subjectivity and insufficient consistency and repeatability of assessment results. While some machine learning methods can automatically extract features from historical data, they lack systematic mechanisms for setting graded thresholds and quantifying overall risk, making it difficult to effectively integrate individual risks into overall risk.

[0004] Chinese invention patent application number 202011517115.X discloses a method for constructing a fire risk prediction model for commercial buildings. This method trains the model on historical fire data of building sites using a random forest algorithm, adjusts the decision tree through feature importance assessment, and uses a combined score of fire frequency and duration as the target value for risk level prediction, thus achieving the classification and prediction of building fire risk levels. However, it lacks comprehensive consideration of real-time dynamic factors such as the status of building fire protection facilities and personnel management, and cannot adapt to the risk distribution drift caused by changes in building usage over time. Summary of the Invention

[0005] In view of this, the present invention provides a three-dimensional risk evolution assessment and early warning method for buildings throughout the entire fire life cycle, in order to solve the problems in existing building fire risk assessment methods where model parameters rely on manual tuning, resulting in limited assessment accuracy, and where the assessment model cannot adaptively update with real-time monitoring data after changes in building usage status. This method achieves feature-driven assessment under multi-source data fusion, accurately depicts the risks throughout the entire fire life cycle, and enables scientific dynamic hierarchical early warning.

[0006] The technical solution of this invention is implemented as follows: On the one hand, this invention provides a three-dimensional risk evolution assessment and early warning method for buildings throughout their entire fire lifecycle, including: S1. Construct a hierarchical risk indicator system based on three sub-dimensions: fire risk, spread risk, and disaster risk. Collect historical building fire data and classify, organize, and preprocess it according to the hierarchical risk indicator system to obtain sub-item training datasets for different sub-dimensions. S2. For different sub-dimensions, a one-to-one multi-classification strategy and radial basis kernel function are used to construct the corresponding support vector machine classification model, resulting in three sub-sub-dimension support vector machine models. S3. The hyperparameters of the three sub-item support vector machine models are optimized using the particle swarm optimization algorithm, and the models are trained using the sub-item training datasets to obtain three optimal parameterized sub-item support vector machine models. The models are then normalized according to the classification accuracy ratio of each optimal parameterized sub-item support vector machine model on the validation set to obtain the dynamic weight vector of each sub-item. S4. Input the real-time collected building feature data into three optimal parameterized sub-item support vector machine models respectively, obtain the decision margin of each sub-item, map the decision margin into continuous risk potential value, construct the real-time risk state point in the three-dimensional risk space with the risk potential value of the three sub-items; use the dynamic weight vector of each sub-item to calculate the distance between the real-time risk state point and the preset disaster target point, and convert the distance into the overall risk index. S5. Compare the overall risk index with the classification threshold to obtain a fire risk warning signal.

[0007] Based on the above technical solutions, preferably, the layered risk indicator system has three primary indicators: fire risk, spread risk, and disaster risk. The fire risk primary indicator has five secondary indicators: inherent building risk, seasonal risk, holiday risk, combustible gas concentration, and electrical spark alarm frequency. The spread risk primary indicator has eleven secondary indicators: fire compartments, automatic fire extinguishing systems, smoke control systems, fire hydrant systems, automatic alarm systems, personnel responsible for firefighting and evacuation, employee fire safety training, fire drills, fire prevention inspections, facility inspections, and regular maintenance of fire-fighting equipment and fire safety signs. The disaster risk primary indicator has five secondary indicators: fire control room staffing, evacuation routes, safety exits, emergency lighting, and evacuation signs. Each secondary indicator corresponds to one of four risk levels: safe, Level III risk, Level II risk, and Level I risk.

[0008] Based on the above technical solutions, preferably, step S2 specifically includes: For each sub-dimension, the number of risk level categories is determined to be four. A one-to-one strategy is used to decompose the four-class problem into 6 independent binary sub-problems, resulting in a set of 6 binary sub-problems corresponding to that sub-dimension. For each binary sub-problem, the corresponding subsets of samples of the two classes are extracted from the sub-item training dataset, and a binary support vector machine is constructed using the radial basis kernel function as the kernel function, resulting in 6 binary support vector machines in the sub-item dimension; The samples to be classified are sequentially input into 6 binary support vector machines for classification voting. The number of votes for each risk level is counted by majority voting. The level with the most votes is taken as the risk level classification result for that sub-item dimension, thus obtaining the sub-item support vector machine model for the sub-item dimension.

[0009] Based on the above technical solutions, preferably, step S3 specifically includes: S31. Using the classification accuracy on the validation set as the fitness function and the penalty factor as the fitness function. With kernel function parameters To find the optimal target, the particle swarm size is set to be... Initialize the particle swarm's position and velocity to obtain the initial particle swarm; S32. Combine the parameters corresponding to each particle in the initial particle swarm. The sub-items are substituted into the support vector machine model for training. The classification accuracy is calculated on the validation set as the fitness value, and the individual historical best position of each particle is initialized. and the global historical best position ; S33. The particle swarm optimization algorithm with shrinkage factor and adaptive inertia weight is used to iteratively update the velocity and position of the particle swarm. After each iteration, the fitness is re-evaluated and the individual historical best position and the global historical best position are updated until the convergence condition is met. The corresponding parameter combination is taken as the optimal hyperparameter combination. S34. Train the corresponding sub-item support vector machine model on the sub-item training dataset with the optimal hyperparameter combination to obtain the optimal parameterized sub-item support vector machine model. Input the validation set samples into the optimal parameterized sub-item support vector machine model, calculate the risk level label of each validation set sample, and compare the risk level label with the true label of the validation set sample one by one. Statistically record the optimal fitness value of the model on the validation set. , These correspond to ignition, spread, and disaster, respectively. S35. Normalize the optimal fitness value to obtain the dynamic weight vector.

[0010] Based on the above technical solution, preferably, in step S32, the training of each binary classification support vector machine includes: For each binary sub-problem, the parameter combination corresponding to the current particle is used. The optimal classification hyperplane is determined by solving the following dual convex quadratic programming problem to find the hyperparameters: in, The number of samples in the binary classification training subset; , The first The and the first Feature vectors of historical building samples; For the first Category labels for each historical building sample; , For Lagrange multipliers; The penalty factor corresponding to the current particle; For radial basis kernel functions; The optimal Lagrange multiplier vector is obtained by solving the problem. Afterwards, based on satisfying Support vector set The risk level label is calculated as follows: in, Indicates the sample to be predicted The risk level label is output after calculation by the decision function. Represents a symbolic function. The expression represents the result obtained by solving the dual problem. The optimal Lagrange multiplier corresponding to each training sample This represents the index of the historical building samples in the binary classification training subset. Indicates satisfaction The set of support vectors composed of historical building samples. Indicates the first Category labels for each historical building sample. For classification bias, Represent the radial basis function kernel and calculate the support vectors. With the sample to be predicted Nonlinear similarity in high-dimensional feature spaces Indicates satisfaction The set of free support vectors, Represents the set of free support vectors The sample index in Represents free support vectors Corresponding category tags, The sample to be predicted; The validation set samples are input into the decision functions. The outputs of the six binary classification decision functions are integrated using a majority voting method, and the classification accuracy of the validation set is calculated as the current parameter combination. The fitness value is updated accordingly. and .

[0011] Based on the above technical solutions, preferably, step S33 specifically includes: S331. The preset particle swarm size is N=30. The position vector of each particle p∈{1,2,…,N} in the optimization space represents a set of candidate hyperparameter combinations for the support vector machine, denoted as... Wherein, the first dimension of the position vector is the coordinate. The real-time representation of the penalty factor value corresponding to the particle, and the second dimension of the position vector. The current nuclear parameter value of the particle is displayed in real time. S332. At the current iteration number t, the particle swarm optimization algorithm calculates the velocity vector of the p-th particle at the next moment using a velocity update formula with a shrinkage factor, based on the p-th particle's current individual historical best position and the swarm's current global historical best position. And update its position vector to obtain ; S333, Set the penalty factor corresponding to the current position of each particle. With kernel parameters The numerical values ​​are substituted into the corresponding support vector machine model, the model is trained using the training dataset of each item, and the classification accuracy of the trained model on the validation set is used as the fitness value of the particle; the historical best position is dynamically updated based on the fitness value. S334. Repeat steps S332 to S333 until the final globally historical best position is obtained. The first dimension value is determined as the optimal penalty factor. The second dimension value is determined as the optimal kernel parameter. In this way, the optimal parameterized support vector machine classification model is constructed.

[0012] Based on the above technical solutions, preferably, the convergence condition is to satisfy one of the following three conditions: First, the number of iterations reaches the preset maximum number of iterations. ; Second, continuous Global optimal fitness in the next iteration Maximum relative rate of change Continuously below the accuracy threshold : in, Indicates continuity The maximum relative rate of change of the global optimal fitness in the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The global optimal fitness value at the next iteration; Third, the globally optimal particle in parameter space continuous The Euclidean distance of each iteration is less than the length of the diagonal of the search interval. The search area is , .

[0013] Based on the above technical solutions, preferably, the speed update formula is: in, Indicates the first The particle in the first During the nth iteration Velocity components in one parameter dimension Indicates the contractility factor. Indicates the first The adaptive inertia weights for the next iteration. Indicates the first The particle in the first During the nth iteration Velocity components in one parameter dimension Indicates the individual's cognitive acceleration coefficient. express Uniformly distributed random numbers within the range. Indicates the first The particle in the first During the nth iteration The individual's historical best position in each parameter dimension. Indicates the first The particle in the first During the nth iteration The current position of each parameter dimension. Indicates the social learning acceleration factor. express Uniformly distributed random numbers within the range. Indicates the first In the nth iteration, the global historical best position is at the . Components of each parameter dimension Indicates the first The particle in the first During the nth iteration The updated position of each parameter dimension. This represents the total acceleration constant determined by the individual cognitive acceleration coefficient c1 and the social learning acceleration coefficient c2, satisfying the following condition: and >4.

[0014] Based on the above technical solutions, preferably, step S4 specifically includes: The real-time collected building feature data is preprocessed and then input into three optimal parameterized component support vector machine models. The decision margins output by each model are then extracted. , This yields the decision margins for the three components; The Sigmoid mapping function is used to determine the decision margins of each component. Convert to range Continuous risk potential value : in, This is the steepness parameter of the mapping curve. This is the offset of the mapping center; The larger the value, the higher the risk level of the corresponding sub-item; Risk potential values ​​of three sub-items Construct real-time risk status points in a three-dimensional risk space for coordinates. The upper limit of the risk potential energy corresponding to when all three sub-items are in an extremely dangerous state is defined as the disaster target point. ; Using dynamic weight vectors Real-time risk status points Disaster target point The weighted Euclidean distance is calculated by weighting the differences of each component's coordinates. and will Normalization converted into overall risk index .

[0015] Further preferred intelligent assessment and early warning methods for building fire risks include: Sample labels are generated for the risk levels of warning signals confirmed through on-site verification, and the confirmed risk levels are used as the actual labels. Write the sample record at the corresponding monitoring time to obtain labeled samples with real labels; For the three optimal parameterized component support vector machine models, establish a capacity of [missing information]. sliding window Continuously compare the support vector machine models of each component in the sliding window. Prediction results within Calculate the inconsistency rate between the fire risk level prediction label and the prediction. : in, For indicator functions, For sliding windows The sample index within, Let represent the fire risk level prediction label calculated by the optimal parameterized support vector machine model of the uth sub-item on the nth validation set sample, u∈{I,S,C}, which correspond to ignition, spread, and disaster, respectively; like Continuously exceeding the preset drift threshold If this occurs, incremental retraining of the corresponding sub-item support vector machine model will be triggered; For the component support vector machine models that trigger retraining, construct an incremental training set. : in, This refers to the set of support vectors obtained from training the optimal parameterized component support vector machine model. For sliding windows The set of all labeled samples whose internal prediction results are inconsistent with the true labels; To from the sliding window A subset of labeled samples whose internal prediction results are consistent with the true labels is randomly selected to maintain the class distribution balance of the training set. The optimal combination of hyperparameters for the optimal parameterized component support vector machine model that triggers incremental retraining. Using prior information, initialize the particle swarm in its neighborhood, and then use it in the incremental training set. The particle swarm optimization algorithm is re-executed to obtain the updated optimal hyperparameter combination. and in After completing the full training, the updated optimal parameterized component support vector machine model is obtained; The updated optimal parameterized sub-item support vector machine model is replaced by the optimal parameterized sub-item support vector machine model that was triggered for incremental retraining using a hot deployment method. Simultaneously, the dynamic weight vector of each sub-item is recalculated using the optimal fitness value of each updated optimal parameterized sub-item support vector machine model on the validation set. The updated three optimal parameterized sub-item support vector machine models re-enter the sliding window prediction inconsistency rate monitoring.

[0016] In addition, the present invention also provides a regional wetland carbon sequestration collaborative management system to implement the above-described method.

[0017] The present invention has the following advantages over the prior art: (1) By decomposing building fire risk into three sub-dimensions: fire risk, spread risk and disaster risk, support vector machine classification models are constructed respectively, and dynamic weight vectors are automatically calibrated based on the accuracy of the validation set of each model. Then, the overall risk index is generated through weighted distance mapping in the three-dimensional risk space, realizing feature self-driven assessment under multi-source data fusion, finely depicting the risk of the entire life cycle of fire, and realizing scientific dynamic hierarchical early warning. (2) Based on the classification accuracy of risk weight and support vector machine sub-model, the weights of each dimension are automatically labeled by the optimization results of particle swarm optimization algorithm, so that the evaluation system is completely driven by data features, which significantly improves the objectivity and credibility of the evaluation results; (3) By mapping the risks of fire, spread, and disaster to three-dimensional coordinates, and using weighted Euclidean distance to characterize the degree of building approaching the "disaster point", it has an exponential nonlinear amplification effect when dealing with concurrent multi-dimensional risks, which greatly improves the system's early warning sensitivity to extreme coupled risks. (4) By mapping the decision margin to a continuous risk potential value, the evolution of the overall risk index TRI exhibits excellent mathematical continuity and smoothness, avoiding the cliff-like jump in the warning level caused by a small change in a single indicator, which is more in line with the objective physical laws of fire risk evolution. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a three-dimensional risk evolution assessment and early warning method for buildings throughout the entire fire life cycle, as described in this invention. Figure 2 This is a schematic diagram of the index system for a three-dimensional risk evolution assessment and early warning method for buildings throughout the entire fire life cycle, as proposed in this invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 and Figure 2As shown, this invention provides a three-dimensional risk evolution assessment and early warning method for buildings throughout their entire fire lifecycle, comprising: S1. A hierarchical risk indicator system is constructed based on three sub-dimensions: fire risk, spread risk, and disaster risk. Historical building fire data is collected and categorized and preprocessed according to the hierarchical risk indicator system to obtain training datasets for different sub-dimensions. The hierarchical risk indicator system has three primary indicators: fire risk, spread risk, and disaster risk. The fire risk primary indicator has five secondary indicators: inherent building risk, seasonal risk, holiday risk, combustible gas concentration, and number of electrical spark alarms. The spread risk primary indicator has five secondary indicators. The system includes 11 secondary indicators: fire compartmentation, automatic fire extinguishing system, smoke control system, fire hydrant system, automatic alarm system, personnel responsible for fire fighting and evacuation, employee fire safety training, fire drills, fire prevention inspections, facility inspections, and regular maintenance of fire-fighting equipment and fire safety signs. The primary indicator of disaster risk includes 5 secondary indicators: staffing of fire control room personnel, evacuation routes, safety exits, emergency lighting, and evacuation signs. Each secondary indicator corresponds to one of four risk levels: safe, Level III risk, Level II risk, and Level I risk, as shown in Table 1. Table 1. Hierarchical Risk Indicator System The classification is based on the "Code for Fire Protection Design of Buildings" (GB50016-2014), the "Unified Standard for Design of Civil Buildings" (GB50352-2019) 3.1, and the "Standard for Delineation of Key Fire Safety Units in Hubei Province". Relevant data is input through regular fire safety inspections.

[0022] The 21 secondary indicators and their grading standards selected in this invention are only examples. Depending on the building type (such as large shopping malls, industrial plants, transportation hubs, etc.), the secondary indicators can be added, subtracted, replaced, or their weights fine-tuned.

[0023] In one embodiment of the present invention, the risk level is determined based on the overall risk probability: Level I Risk: The highest level of warning is triggered when the overall risk probability is ≥0.90, or when any sub-item risk level is Level I. Level II Risk: When the overall risk probability is ≥0.80 and <0.90, or when at least two sub-risks are Level II, a second-highest level warning is triggered. Level III risk: Triggered when the overall risk probability is ≥0.60 and <0.80; Safety: Overall risk probability < 0.60.

[0024] Understandably, historical data sources include records uploaded by fire safety inspection apps, building basic information databases, fire protection facility IoT monitoring systems, historical fire statistics archives, and meteorological calendar data.

[0025] Following the rule that "the larger the value, the lower the risk," all secondary indicator risk levels are mapped to discrete numerical features, forming the sample vector input. The mapping function is as follows: in, This indicates the risk level of the secondary indicator.

[0026] In one embodiment of the present invention, the preprocessing includes: For the mapped discrete risk level data (such as 1, 2, 3, 4), the validity of the value range should be verified based on the business rules. According to the preset mapping function, check whether the input value belongs to the set {1, 2, 3, 4}: If a value that does not belong to the set appears (such as 0, 5 or non-numeric characters), it is determined to be a data transmission error or a manual input error. For such errors, the mode imputation method is used to correct the error directly based on the historical distribution of the corresponding indicator. The distribution pattern of the specific secondary indicator is retrieved from the historical database, the frequency of occurrence of each level (level 1 to 4) is counted, and the level value with the highest frequency of occurrence is selected as the "mode" of the indicator. When the current sample is missing the indicator, the historical mode is automatically imputed to the corresponding dimension of the feature vector according to the preset rules. Based on the risk level, the status of all building sub-risk secondary indicators is converted into numerical features (e.g., safe → 4, Level III risk → 3, Level II risk → 2, Level I risk → 1). The numerical features of each building's sub-risk secondary indicators are combined into a set of data. All numerical features are then subjected to min-max normalization to ensure they fall within the [0,1] interval. In the formula, x is the original eigenvalue. and These are the lower limit minimum and upper limit maximum values ​​of the feature, respectively. ∈[0,1] represents the normalized result; For the secondary indicators, their states have been mapped to discrete values ​​in a fixed set {1,2,3,4} according to rules. To ensure the absolute consistency of the normalization results across different evaluation batches, the theoretical extreme value of the domain of this mapping is taken, i.e., a constant setting. =4, =1; All processed data were divided into training, validation, and test sets in a 7:2:1 ratio.

[0027] S2. For different sub-dimensions, a one-to-one multi-classification strategy and radial basis kernel function are used to construct the corresponding support vector machine classification model, resulting in three sub-sub-dimension support vector machine models. Specifically, step S2 includes: For each sub-dimensional risk level, four risk categories are defined (Level I risk, Level II risk, Level III risk, and safe). A one-to-one strategy is used to decompose the four-class classification problem into six independent binary classification sub-problems, resulting in a set of six binary classification sub-problems corresponding to that sub-dimensional risk level. The set of classifiers trained using the one-to-one strategy is as follows: in, Let be the decision function, and be the original output value of the binary classifier. This represents the Lagrange coefficient corresponding to the i-th support vector. This represents the classification target label of the i-th support vector. Indicates the bias term. This indicates the first risk level category number used for a one-to-one classification comparison. This indicates the second risk level category number used for a one-to-one classification comparison. Let Ω = {1, 2, 3, 4} represent the four categories, corresponding to Level I risk, Level II risk, Level III risk, and safety, respectively. For each binary classification sub-problem, the corresponding subsets of samples from the two classes are extracted from the sub-item training dataset. A binary classification support vector machine is constructed using the radial basis function (RBF) kernel, resulting in six binary classification support vector machines for each sub-item dimension. The expression for the RBF kernel is: in, These are kernel function parameters that control the scope of the function's operation. , These are the feature vectors of the two samples, This represents the risk sensitivity coefficient of the k-th secondary indicator. This represents the k-th secondary indicator feature value of the i-th historical building sample in the sub-item training dataset. This represents the k-th secondary indicator feature value of the j-th historical building sample in the sub-item training dataset; The samples to be classified are sequentially input into six binary support vector machines for classification voting. The majority voting method is used to count the number of votes for each risk level. The level with the most votes is taken as the risk level classification result for that sub-item dimension, thus obtaining the sub-item support vector machine model. The formula for the majority voting method is: in, Indicates the final risk level label. This represents the class mapping function, which outputs the corresponding class number for the binary classifier, II( () indicates an indicator function, which means that 1 point is awarded if the condition inside the parentheses is true, and 0 points are awarded otherwise; Let Ω = {1, 2, 3, 4} represent the four categories, corresponding to Level I risk, Level II risk, Level III risk, and safety, respectively. It is represented by the candidate risk level category number in the four-category set Ω.

[0028] Understandably, radial basis function kernels can map low-dimensional features to high-dimensional space, effectively handling nonlinear classification boundaries. The sign indicates which class sample x is biased towards, and the magnitude of the value represents the degree of bias. Class mapping function. Will The positive and negative signs are translated back to specific category labels. It is a transformation logic that ensures that the calculation results can be converted into statistically verifiable votes.

[0029] Through category mapping function Binary decision function The continuous output values ​​are converted into discrete category label votes, if If it is a positive value, then Output Labels p Otherwise, output the label. q， Ultimately, the overall risk assessment module evaluates all sub-classifiers. The output results are used for majority voting to determine the final risk level label of the target building under a specific sub-item. .

[0030] By training on historical samples with known risk levels, multiple sets of binary classification discriminant functions are constructed, where the variables... p The category objective (i.e., a specific risk level label) of the discriminant function is defined, while X i With X j This constitutes the feature vector set for the corresponding category. The model calculates X using a kernel function. i With X j By studying the spatial distribution patterns between them, the optimal classification hyperplane can be determined, thereby enabling automated classification and assessment of buildings with unknown risk states.

[0031] This invention employs a kernel function as the core of the sub-item risk assessment model, which achieves high-dimensional nonlinear mapping of the feature space. The kernel function calculates the feature vector X of the building to be assessed and the historical support vector X'. i The high-dimensional similarity information between samples effectively solves the problem of linear inseparability of building fire protection features in the original low-dimensional space. The radial basis function captures the spatial distribution pattern between samples and transforms it into a decision function. The weighting criteria are used to construct an accurate risk level determination boundary, which significantly improves the model's accuracy in identifying complex fire evolution logic.

[0032] S3. The hyperparameters of the three sub-item support vector machine models are optimized using the particle swarm optimization algorithm, and the models are trained using the sub-item training datasets to obtain three optimal parameterized sub-item support vector machine models. The models are then normalized according to the classification accuracy ratio of each optimal parameterized sub-item support vector machine model on the validation set to obtain the dynamic weight vector of each sub-item. Specifically, step S3 includes: S31. Using the classification accuracy on the validation set as the fitness function and the penalty factor as the fitness function. With kernel function parameters To find the optimal target, the particle swarm size is set to be... Initialize the particle swarm's position and velocity to obtain the initial particle swarm; S32. Combine the parameters corresponding to each particle in the initial particle swarm. The sub-items are substituted into the support vector machine model for training. The classification accuracy is calculated on the validation set as the fitness value, and the individual historical best position of each particle is initialized. and the global historical best position The formula for calculating classification accuracy is: in, Indicates classification accuracy. This represents the number of correctly predicted samples on the validation set. This represents the total number of samples in the validation set; S33. The particle swarm optimization (PSO) algorithm with shrinkage factor and adaptive inertia weight is used to iteratively update the velocity and position of the particle swarm. After each iteration, the fitness is re-evaluated and the individual historical best position and the global historical best position are updated until the convergence condition is met. The corresponding parameter combination is taken as the optimal hyperparameter combination. Step S33 specifically includes the following sub-steps: S331. The preset particle swarm size is N=30. The position vector of each particle p (where p∈{1,2,…,N}) in the optimization space represents a set of candidate hyperparameter combinations for the support vector machine, denoted as ; Wherein, the first dimension of the position vector is coordinate The second dimension coordinate represents the current penalty factor value of the particle. The penalty factor represents the current nuclear parameter value of the particle in real time; throughout the iterative optimization process, the penalty factor... The value of the kernel parameter is strictly limited to a preset search range [0.01, 100]. It is strictly limited to the preset search range [0.001, 10].

[0033] S332. At the current iteration number t, the particle swarm optimization algorithm calculates the velocity vector of the p-th particle at the next moment using a velocity update formula with a shrinkage factor, based on the p-th particle's current individual historical best position and the swarm's current global historical best position. And then update its position vector to obtain This allows for the application of a penalty factor to the particle. Numerical values ​​and kernel parameters Dynamic adjustment and optimization of numerical values.

[0034] S333, Set the penalty factor corresponding to the current position of each particle. With kernel parameters Substitute the numerical values ​​into the corresponding support vector machine models for each item, train the models using the item training datasets, and then measure the classification accuracy of the trained models on the validation set. The fitness value of the particle is then used as the basis for dynamic updates to the historical best position, as follows: ① Individual Historical Best Position Update: Compare the current classification accuracy of the p-th particle with the highest classification accuracy recorded by the particle in all previous iterations; if the current classification accuracy is higher than its historical highest record, then update the individual historical best position of the particle using the current position vector, denoted as . , Indicates the first The penalty factor value corresponding to the highest classification accuracy achieved by an individual particle across all historical iterations. Indicates the first The kernel function parameter value corresponding to the highest classification accuracy achieved by an individual particle in all historical iterations; ② Global Historical Best Position Update: Compare the classification accuracy of all particles in the current iteration with the highest classification accuracy recorded in all previous iterations of the entire particle swarm; if a higher classification accuracy is found, update the global historical best position of the entire particle swarm using the current position vector of that best particle, denoted as . , This represents the penalty factor value corresponding to the particle that achieves the highest classification accuracy across all iterations of the entire particle swarm. This represents the kernel function parameter value corresponding to the particle that achieves the highest global classification accuracy across all iterations of the entire particle swarm.

[0035] S334. Repeat steps S332 to S333 until the maximum number of iterations T is reached. max=100, or the fitness value of the global historical best position no longer increases for 10 consecutive generations, indicating that the convergence condition has been met; at this point, the iteration is terminated, and the final global historical best position is obtained. The first dimension value is determined as the optimal penalty factor. The second dimension value is determined as the optimal kernel parameter. In this way, the optimal parameterized support vector machine classification model is constructed.

[0036] S34. Using the optimal hyperparameter combination Train the corresponding sub-item support vector machine model completely on the sub-item training dataset to obtain the optimal parameterized sub-item support vector machine model. Input the validation set samples into the optimal parameterized sub-item support vector machine model to calculate the risk level label of each validation set sample. Compare the risk level label with the true label of the validation set sample one by one, and count and record the optimal fitness value of the model on the validation set. , These correspond to ignition, spread, and disaster, respectively. S35. Normalize the optimal fitness value to obtain the dynamic weight vector. : in, For sub-items The corresponding dynamic weight vector, , , These represent the optimal fitness values ​​of the support vector machine models for the three optimal parameterized components: fire ignition, spread, and disaster, on the validation set.

[0037] In one embodiment of the present invention, the number of particles is 30, N=30, and the maximum number of iterations is set to... Punishment factor The range of values ​​is set to nuclear parameters The range of values ​​is set to Initialize particle velocity The speed range is set to , Individual learning factor c1 and social learning factor c2 are set as follows: .

[0038] In one embodiment of the present invention, step S32, the training of each binary classification support vector machine includes: For each binary sub-problem, the parameter combination corresponding to the current particle is used. The optimal classification hyperplane is determined by solving the following dual convex quadratic programming problem to find the hyperparameters: in, The number of samples in the binary classification training subset; , The first The and the first The feature vector of each historical building sample is composed of the values ​​of each secondary risk indicator in Table 1 after quantitative mapping and normalization. For the first Category labels for each historical building sample; , For Lagrange multipliers, representing the decisive weights of the corresponding historical samples when defining risk boundaries; This is the penalty factor corresponding to the current particle, used to set the upper limit of the model's tolerance for misclassified samples; For radial basis kernel functions; The optimal Lagrange multiplier vector is obtained by solving the problem. Afterwards, based on satisfying Support vector set The risk level label is calculated as follows: in, Indicates the sample to be predicted The risk level label is output after calculation by the decision function. Represents a symbolic function. The expression represents the result obtained by solving the dual problem. The optimal Lagrange multiplier corresponding to each training sample This represents the index of the historical building samples in the binary classification training subset. Indicates satisfaction The set of support vectors composed of historical building samples. Indicates the first Category labels for each historical building sample. For classification bias, Represent the radial basis function kernel and calculate the support vectors. With the sample to be predicted Nonlinear similarity in high-dimensional feature spaces Indicates satisfaction The set of free support vectors, Represents the set of free support vectors The sample index in Represents free support vectors Corresponding category tags, The sample to be predicted; Input the validation set samples into the decision function. The outputs of the six binary classification decision functions are integrated using the majority voting method, and the classification accuracy on the validation set is calculated as the current parameter combination. The fitness value is updated accordingly. and .

[0039] By introducing a penalty factor C and a kernel function parameter γ, an effective trade-off between "minimizing empirical risk" and "minimizing structural risk" is achieved in the dual space. This prevents the model from distorting the classification boundary by overly catering to individual extreme and abnormal fire conditions, and ensures that the early warning model has extremely high generalization ability and evaluation stability when facing new target buildings.

[0040] In one embodiment of the present invention, the convergence condition is to satisfy one of the following three conditions: First, the number of iterations reaches the preset maximum number of iterations. ; Second, continuous Global optimal fitness in the next iteration Maximum relative rate of change Continuously below the accuracy threshold : in, Indicates continuity The maximum relative rate of change of the global optimal fitness in the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The global optimal fitness value at the next iteration; Third, the globally optimal particle in parameter space continuous The Euclidean distance of each iteration is less than the length of the diagonal of the search interval. The search area is , .

[0041] In one embodiment of the present invention, ≥10, .

[0042] In one embodiment of the present invention, the speed update formula is as follows: in, Indicates the first The particle in the first During the nth iteration Velocity components in one parameter dimension Indicates the contractility factor. Indicates the first The adaptive inertia weights for the next iteration. Indicates the first The particle in the first During the nth iteration Velocity components in one parameter dimension Indicates the individual's cognitive acceleration coefficient. express Uniformly distributed random numbers within the range. Indicates the first The particle in the first During the nth iteration The individual's historical best position in each parameter dimension. Indicates the first The particle in the first During the nth iteration The current position of each parameter dimension. Indicates the social learning acceleration factor. Represents a random number uniformly distributed within the range [0,1]. Indicates the first In the nth iteration, the global historical best position is at the . Components of each parameter dimension Indicates the first The particle in the first During the nth iteration The updated position of each parameter dimension. This represents the total acceleration constant determined by both the individual cognitive acceleration coefficient c1 and the social learning acceleration coefficient c2. , Indicates the initial inertia weight. Indicates the termination of inertia weight. Indicates the maximum number of iterations. This indicates the current iteration number.

[0043] This invention is based on the classification accuracy of risk weights and support vector machine sub-models. It automatically labels the weights of each dimension through the optimization results of particle swarm optimization algorithm, so that the evaluation system is completely driven by data features, which significantly improves the objectivity and credibility of the evaluation results.

[0044] In one embodiment of the present invention, The value is 0.9. The value is 0.4. It is a random number. .

[0045] S4. Input the real-time collected building feature data into three optimal parameterized sub-item support vector machine models respectively, obtain the decision margin of each sub-item, map the decision margin into continuous risk potential value, construct the real-time risk state point in the three-dimensional risk space with the risk potential value of the three sub-items; use the dynamic weight vector of each sub-item to calculate the distance between the real-time risk state point and the preset disaster target point, and convert the distance into the overall risk index. Specifically, step S4 includes: The real-time collected building feature data is preprocessed and then input into three optimal parameterized component support vector machine models. The decision margins output by each model are then extracted. , This yields the decision margins for the three components; The Sigmoid mapping function is used to determine the decision margins of each component. Convert to range Continuous risk potential value : in, This is the steepness parameter of the mapping curve. This is the offset of the mapping center; The larger the value, the higher the risk level of the corresponding sub-item; Risk potential values ​​of three sub-items Construct real-time risk status points in a three-dimensional risk space for coordinates. The upper limit of the risk potential energy corresponding to when all three sub-items are in an extremely dangerous state is defined as the disaster target point. ; Using dynamic weight vectors Real-time risk status points Disaster target point The weighted Euclidean distance is calculated by weighting the differences of each component's coordinates. and will Normalization converted into overall risk index .

[0046] This invention maps the risks of fire ignition, spread, and disaster into three-dimensional coordinates, and uses weighted Euclidean distance to characterize the degree to which a building approaches the "disaster point." It exhibits an exponential nonlinear amplification effect when dealing with concurrent multidimensional risks, significantly improving the system's early warning sensitivity to extreme coupled risks. By mapping the decision margin to continuous risk potential energy values, the evolution of the overall risk index TRI exhibits excellent mathematical continuity and smoothness, avoiding abrupt jumps in warning levels caused by minor changes in a single indicator, and better aligning with the objective physical laws of fire risk evolution.

[0047] In one embodiment of the present invention, the disaster target point If the value is (100, 100, 100), and this point represents an extreme dangerous state in all three dimensions—ignition, spread, and disaster—then: in, This represents the dynamic weight vector of the fire risk components. This represents the dynamic weight vector of the spread risk components. This represents the dynamic weight vector of the disaster risk components. This represents the potential energy value indicating the risk of fire. This represents the potential energy value indicating the risk of contagion. This represents the potential risk value for a potential disaster.

[0048] Understandably, when the risk increases in any dimension, the state point approaches the disaster point. This decreases, thus driving the TRI value to rise. The larger the value, the higher the overall fire risk of the building.

[0049] S5. Compare the overall risk index with the classification threshold to obtain a fire risk warning signal.

[0050] Understandable, when At that time, a Level 1 red alert was triggered, indicating that the building's condition was extremely close to the disaster target point. A building-wide coordinated fire emergency response was initiated, and a real-time disaster-causing factor analysis report was sent to the local command center. When a Level II orange alert is triggered, indicating that the building's risk deviation exceeds the safety benchmark, a terminal instruction is sent to the fire safety manager, requiring them to complete a targeted on-site verification within a predetermined timeframe; when When a Level 3 yellow alert is triggered, it indicates that some building indicators deviate from safety benchmarks, automatically generating a list of potential hazards requiring rectification and prompting the strengthening of management of weak indicators; when When the building is deemed to be within a safe threshold, routine IoT data monitoring is maintained.

[0051] This invention decomposes building fire risk into three sub-dimensions: ignition risk, spread risk, and disaster risk. It constructs support vector machine classification models for each sub-dimension and automatically calibrates dynamic weight vectors based on the accuracy of each model's validation set. Then, it generates an overall risk index through weighted distance mapping in a three-dimensional risk space. This enables feature-driven assessment under multi-source data fusion, accurately depicts the risk throughout the entire fire life cycle, and achieves scientific dynamic hierarchical early warning.

[0052] In one embodiment of the present invention, the intelligent assessment and early warning method for building fire risk further includes: Sample labels are generated for the risk levels of warning signals confirmed through on-site verification, and the confirmed risk levels are used as the actual labels. Write the sample record at the corresponding monitoring time to obtain labeled samples with real labels; For the three optimal parameterized component support vector machine models, establish a capacity of [missing information]. sliding window Continuously compare the support vector machine models of each component in the sliding window. Prediction results within Calculate the inconsistency rate between the fire risk level prediction label and the prediction. : in, For indicator functions, For sliding windows The sample index within, Let represent the fire risk level prediction label calculated by the optimal parameterized support vector machine model of the uth sub-item on the nth validation set sample, u∈{I,S,C}, which correspond to ignition, spread, and disaster, respectively; like Continuously exceeding the preset drift threshold If this occurs, incremental retraining of the corresponding sub-item support vector machine model will be triggered; For the component support vector machine models that trigger retraining, construct an incremental training set. : in, This refers to the set of support vectors obtained from training the optimal parameterized component support vector machine model. For sliding windows The set of all labeled samples whose internal prediction results are inconsistent with the true labels; To from the sliding window A subset of labeled samples whose internal prediction results are consistent with the true labels is randomly selected to maintain the class distribution balance of the training set. The optimal combination of hyperparameters for the optimal parameterized component support vector machine model that triggers incremental retraining. Using prior information, initialize the particle swarm in its neighborhood, and then use it in the incremental training set. The particle swarm optimization algorithm is re-executed to obtain the updated optimal hyperparameter combination. and in After completing the full training, the updated optimal parameterized component support vector machine model is obtained; The initial particle swarm is: in, This represents the penalty factor value for the i-th particle at the initial time of incremental optimization. This represents the kernel parameter value of the i-th particle at the initial moment of incremental optimization. The term represents the neighborhood random perturbation of the penalty factor. The term represents the neighborhood random perturbation of the kernel parameters. To trigger the optimal penalty factor corresponding to the original model before incremental retraining, The optimal kernel parameters of the original model before triggering incremental retraining are defined as follows: i is the particle number index, and i∈{1,2,…,N}, where N is the particle swarm size, and t=0 represents the initial iteration number of the particle swarm algorithm. The updated optimal parameterized sub-item support vector machine model is replaced by the optimal parameterized sub-item support vector machine model that was triggered for incremental retraining using a hot deployment method. Simultaneously, the dynamic weight vector of each sub-item is recalculated using the optimal fitness value of each updated optimal parameterized sub-item support vector machine model on the validation set. The updated three optimal parameterized sub-item support vector machine models re-enter the sliding window prediction inconsistency rate monitoring.

[0053] The real label y is obtained asynchronously through the following channels. true It also achieves closed-loop backfilling: first, through manual verification status collected by the fire supervision and inspection APP; second, through fire verification signals automatically generated by IoT sensor trigger logic; and third, through association with historical fire accident statistics records.

[0054] To illustrate this invention with a specific example, building A is selected as the evaluation object. All 21 secondary indicator status values ​​of the building are collected. Based on the four-level quantitative standard (1: Level I risk, 2: Level II risk, 3: Level III risk, 4: Safety), the original observation vector X of building A is obtained and normalized, where the theoretical extreme value is taken as the constant. , The normalization results are shown in Table 2: Table 2 Normalization results of secondary indicator state values Extract the corresponding normalized features according to the sub-item dimension, and construct the real-time input vectors for the three sub-items: Fire risk component input vector ( — (5 items in total) Contagion risk component input vector ( — (Total 11 items) Disaster risk component input vector ( — (5 items in total) For the three sub-dimensions of fire risk, spread risk, and disaster risk, a one-to-one strategy is adopted to decompose the four-class problem into six independent binary sub-problems. The corresponding binary support vector machines are constructed using radial basis kernel functions, resulting in three sub-sub-support vector machine models.

[0055] Using the classification accuracy on the validation set as the fitness function, and the penalty factor as... With kernel function parameters To determine the optimization target, the particle swarm optimization algorithm was used to optimize the hyperparameters of the three sub-models. The particle swarm size was set to... acceleration coefficient Corresponding contraction factor Maximum number of iterations After the particle swarm optimization algorithm converges, the optimal combination of hyperparameters for each component is used. The three optimal parameterized component support vector machine models were obtained by fully training on the corresponding component training datasets.

[0056] The optimal fitness values ​​(classification accuracy) of each sub-model on the validation set are as follows: Fire Model F I =91.5%, propagation model F S =86.1%, disaster model F C =91.4%.

[0057] Normalize using the following formula to obtain the dynamic weight vector. : Fire ignition weight: ; Spread weight: ; Disaster weighting: ; The real-time input vector is fed into three optimally parameterized component support vector machine models, and the decision margins of each model's output are extracted. And discrete risk level classification results: Fire sub-model: (Level II risk); Spreading sub-model: (Level II risk); Disaster-causing sub-model: (Level III risk); The Sigmoid mapping function is used to convert the margins of each decision item into intervals. Continuous risk potential value , , , ; This constitutes the real-time coordinates of building A in the three-dimensional risk space. ; Define an extremely dangerous state as a negative ideal solution. Calculate the current state point. Disaster site Weighted Euclidean geometric distance between them: Building A has a comprehensive risk index (TRI) of 81.00, which meets the judgment range of 80≤TRI<90, and automatically triggers a level 2 (orange) warning signal.

[0058] The corresponding sub-categorization results show that both the fire risk and spread risk are at Level II, while the disaster risk is at Level III. Among these, the personnel responsible for firefighting and evacuation are rated Level I (highest risk), representing the primary risk factor for the current spread risk. The system sends terminal instructions to the fire safety manager, requiring them to complete on-site verification and rectification of the above indicators within a predetermined timeframe. The verified and confirmed actual risk level will serve as the actual label. The sample records at the corresponding monitoring time are written and incorporated into the sliding window drift monitoring of each sub-item support vector machine model, triggering the subsequent incremental retraining process.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional risk evolution assessment and early warning method for buildings throughout their entire fire lifecycle, characterized in that, include: S1. Construct a hierarchical risk indicator system based on three sub-dimensions: fire risk, spread risk, and disaster risk. Collect historical building fire data and classify, organize, and preprocess it according to the hierarchical risk indicator system to obtain sub-item training datasets for different sub-dimensions. S2. For different sub-dimensions, a one-to-one multi-classification strategy and radial basis kernel function are used to construct the corresponding support vector machine classification model, resulting in three sub-sub-dimension support vector machine models. S3. The hyperparameters of the three sub-item support vector machine models are optimized using the particle swarm optimization algorithm, and the models are trained using the sub-item training datasets to obtain three optimal parameterized sub-item support vector machine models. The models are then normalized according to the classification accuracy ratio of each optimal parameterized sub-item support vector machine model on the validation set to obtain the dynamic weight vector of each sub-item. S4. Input the real-time collected building feature data into three optimal parameterized sub-item support vector machine models respectively, obtain the decision margin of each sub-item, map the decision margin into continuous risk potential value, and construct the real-time risk state point in the three-dimensional risk space with the risk potential value of the three sub-items. The distance between the real-time risk status point and the preset disaster target point is calculated using the dynamic weight vector of each sub-item, and the distance is converted into an overall risk index; S5. Compare the overall risk index with the classification threshold to obtain a fire risk warning signal.

2. The three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle as described in claim 1, characterized in that: The hierarchical risk indicator system comprises three primary indicators: fire risk, spread risk, and disaster risk. The fire risk primary indicator includes five secondary indicators: inherent building risk, seasonal risk, holiday risk, combustible gas concentration, and electrical spark alarm frequency. The spread risk primary indicator includes eleven secondary indicators: fire compartments, automatic fire extinguishing systems, smoke control systems, fire hydrant systems, automatic alarm systems, personnel responsible for firefighting and evacuation, employee fire safety training, fire drills, fire prevention inspections, facility inspections, and regular maintenance of fire-fighting equipment and fire safety signs. The disaster risk primary indicator includes five secondary indicators: fire control room staffing, evacuation routes, safety exits, emergency lighting, and evacuation signs. Each secondary indicator corresponds to one of four risk levels: safe, Level III risk, Level II risk, and Level I risk.

3. The three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle as described in claim 1, characterized in that: Step S2 specifically includes: For each sub-dimension, the number of risk level categories is determined to be four. A one-to-one strategy is used to decompose the four-class problem into 6 independent binary sub-problems, resulting in a set of 6 binary sub-problems corresponding to that sub-dimension. For each binary sub-problem, the corresponding subsets of samples of the two classes are extracted from the sub-item training dataset, and a binary support vector machine is constructed using the radial basis kernel function as the kernel function, resulting in 6 binary support vector machines in the sub-item dimension; The samples to be classified are sequentially input into 6 binary support vector machines for classification voting. The number of votes for each risk level is counted by majority voting. The level with the most votes is taken as the risk level classification result for that sub-item dimension, thus obtaining the sub-item support vector machine model for the sub-item dimension.

4. The three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle as described in claim 3, characterized in that: Step S3 specifically includes: S31. Using the classification accuracy on the validation set as the fitness function and the penalty factor as the fitness function. With kernel function parameters To find the optimal target, the particle swarm size is set to be... Initialize the particle swarm's position and velocity to obtain the initial particle swarm; S32. Combine the parameters corresponding to each particle in the initial particle swarm. The sub-items are substituted into the support vector machine model for training. The classification accuracy is calculated on the validation set as the fitness value, and the individual historical best position of each particle is initialized. and the global historical best position ; S33. The particle swarm optimization algorithm with shrinkage factor and adaptive inertia weight is used to iteratively update the velocity and position of the particle swarm. After each iteration, the fitness is re-evaluated and the individual historical best position and the global historical best position are updated until the convergence condition is met. The corresponding parameter combination is taken as the optimal hyperparameter combination. S34. Train the corresponding sub-item support vector machine model on the sub-item training dataset with the optimal hyperparameter combination to obtain the optimal parameterized sub-item support vector machine model. Input the validation set samples into the optimal parameterized sub-item support vector machine model, calculate the risk level label of each validation set sample, and compare the risk level label with the true label of the validation set sample one by one. Statistically record the optimal fitness value of the model on the validation set. , These correspond to ignition, spread, and disaster, respectively. S35. Normalize the optimal fitness value to obtain the dynamic weight vector.

5. The three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle as described in claim 4, characterized in that: In step S32, the training of each binary classification support vector machine includes: For each binary sub-problem, the parameter combination corresponding to the current particle is used. The optimal classification hyperplane is determined by solving the following dual convex quadratic programming problem to find the hyperparameters: ； ； in, The number of samples in the binary classification training subset; , The first The and the first Feature vectors of historical building samples; For the first Category labels for each historical building sample; , For Lagrange multipliers; The penalty factor corresponding to the current particle; It is a radial basis kernel function; The optimal Lagrange multiplier vector is obtained by solving the problem. Afterwards, based on satisfying Support vector set The risk level label is calculated as follows: ； ； in, Indicates the sample to be predicted The risk level label is output after calculation by the decision function. Represents a symbolic function. The expression represents the result obtained by solving the dual problem. The optimal Lagrange multiplier corresponding to each training sample This represents the index of the historical building samples in the binary classification training subset. Indicates satisfaction The support vector set composed of historical building samples, Indicates the first Category labels for each historical building sample. For classification bias, Represent the radial basis function kernel and calculate the support vectors. With the sample to be predicted Nonlinear similarity in high-dimensional feature spaces Indicates satisfaction The set of free support vectors, Represents the set of free support vectors The sample index in Represents free support vectors Corresponding category tags, The sample to be predicted; The validation set samples are input into the decision functions. The outputs of the six binary classification decision functions are integrated using a majority voting method, and the classification accuracy of the validation set is calculated as the current parameter combination. The fitness value is then updated accordingly. and .

6. The three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle as described in claim 5, characterized in that: Step S33 specifically includes: S331. The preset particle swarm size is N=30. The position vector of each particle p∈{1,2,…,N} in the optimization space represents a set of candidate hyperparameter combinations for the support vector machine, denoted as... Wherein, the first dimension of the position vector is the coordinate. The real-time representation of the penalty factor value corresponding to the particle, and the second dimension of the position vector. The current nuclear parameter value of the particle is displayed in real time. S332. At the current iteration number t, the particle swarm optimization algorithm calculates the velocity vector of the p-th particle at the next moment using a velocity update formula with a shrinkage factor, based on the p-th particle's current individual historical best position and the swarm's current global historical best position. And update its position vector to obtain ; S333, Set the penalty factor corresponding to the current position of each particle. With kernel parameters The numerical values ​​are substituted into the corresponding support vector machine model, the model is trained using the training dataset of each item, and the classification accuracy of the trained model on the validation set is used as the fitness value of the particle; the historical best position is dynamically updated based on the fitness value. S334. Repeat steps S332 to S333 until the final globally historical best position is obtained. The first dimension value is determined as the optimal penalty factor. The second dimension value is determined as the optimal kernel parameter. In this way, the optimal parameterized support vector machine classification model is constructed.

7. A three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle, as described in claim 6, characterized in that: The convergence condition is to satisfy one of the following three conditions: First, the number of iterations reaches the preset maximum number of iterations. ; Second, continuous Global optimal fitness in the next iteration Maximum relative rate of change Continuously below the accuracy threshold : ； in, Indicates continuity The maximum relative rate of change of the global optimal fitness in the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The globally optimal fitness value at the next iteration. Indicates the first The global optimal fitness value at the next iteration; Third, the globally optimal particle in parameter space continuous The Euclidean distance of each iteration is less than the length of the diagonal of the search interval. The search area is , .

8. The three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle as described in claim 7, characterized in that: The speed update formula is: ； ； ； in, Indicates the first The particle in the first During the nth iteration Velocity components in one parameter dimension Indicates the contractility factor. Indicates the first The adaptive inertia weights for the next iteration. Indicates the first The particle in the first During the nth iteration Velocity components in one parameter dimension Indicates the individual's cognitive acceleration coefficient. express Uniformly distributed random numbers within the range. Indicates the first The particle in the first During the nth iteration The individual's historical best position in each parameter dimension. Indicates the first The particle in the first During the nth iteration The current position of each parameter dimension. Indicates the social learning acceleration factor. express Uniformly distributed random numbers within the range. Indicates the first In the nth iteration, the global historical best position is at the . Components of each parameter dimension Indicates the first The particle in the first During the nth iteration The updated position of each parameter dimension. This represents the total acceleration constant determined by the individual cognitive acceleration coefficient c1 and the social learning acceleration coefficient c2, satisfying the following condition: and >

4.

9. A three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle, as described in claim 1, characterized in that: Step S4 specifically includes: The real-time collected building feature data is preprocessed and then input into three optimal parameterized component support vector machine models. The decision margins output by each model are then extracted. , This yields the decision margins for the three components; The Sigmoid mapping function is used to determine the decision margins of each component. Convert to range Continuous risk potential value : ； in, This is the steepness parameter of the mapping curve. This is the offset of the mapping center; The larger the value, the higher the risk level of the corresponding sub-item; Risk potential values ​​of three sub-items Construct real-time risk status points in a three-dimensional risk space for coordinates. The upper limit of the risk potential energy corresponding to when all three sub-items are in an extremely dangerous state is defined as the disaster target point. ; Using dynamic weight vectors Real-time risk status points With disaster target point The weighted Euclidean distance is calculated by weighting the differences of each component's coordinates. and will Normalization converted into overall risk index .

10. A three-dimensional risk evolution assessment and early warning method for buildings oriented towards the entire fire life cycle, as described in claim 1, characterized in that: Intelligent assessment and early warning methods for building fire risk also include: Sample labels are generated for the risk levels of warning signals confirmed through on-site verification, and the confirmed risk levels are used as the actual labels. Write the sample record at the corresponding monitoring time to obtain labeled samples with real labels; For the three optimal parameterized component support vector machine models, establish a capacity of [missing information]. sliding window Continuously compare the support vector machine models of each component in the sliding window. Prediction results within Calculate the inconsistency rate between the fire risk level prediction label and the prediction. : ； in, For indicator functions, For sliding windows The sample index within, Let represent the fire risk level prediction label calculated by the optimal parameterized support vector machine model of the uth sub-item on the nth validation set sample, u∈{I,S,C}, which correspond to ignition, spread, and disaster, respectively. like Continuously exceeding the preset drift threshold If this occurs, incremental retraining of the corresponding sub-item support vector machine model will be triggered; For the component support vector machine models that trigger retraining, construct an incremental training set. : ； in, This refers to the set of support vectors obtained from training the optimal parameterized component support vector machine model. For sliding windows The set of all labeled samples whose internal prediction results are inconsistent with the true labels; To use the sliding window A subset of labeled samples whose internal prediction results are consistent with the true labels is randomly selected to maintain the class distribution balance of the training set. The optimal combination of hyperparameters for the optimal parameterized component support vector machine model that triggers incremental retraining. Using prior information, initialize the particle swarm in its neighborhood, and then use it in the incremental training set. The particle swarm optimization algorithm is re-executed to obtain the updated optimal hyperparameter combination. and in After completing the full training, the updated optimal parameterized component support vector machine model is obtained; The updated optimal parameterized sub-item support vector machine model is replaced by the optimal parameterized sub-item support vector machine model that was triggered for incremental retraining in a hot deployment manner. Simultaneously, the dynamic weight vector of each sub-item is recalculated using the optimal fitness value of each optimal parameterized sub-item support vector machine model on the validation set after the update. The three updated optimal parameterized sub-item support vector machine models re-enter the sliding window prediction inconsistency rate monitoring.

Images