Intelligent fire-fighting hidden danger identification method based on large-model multi-mode

Abstract

The invention, which relates to the technical field of fire safety, discloses a multi-modal intelligent identification method for fire-fighting hidden troubles based on a large model, comprising the following steps: S1, acquiring multi-source heterogeneous feature data through a sensor network deployed in a fire-fighting water pump room to obtain an initial feature set; s2, performing unified scale conversion on the initial feature set, calculating heterogeneous bridging parameters according to the environmental parameters and the equipment operation data, constructing a fusion feature vector, inputting the fusion feature vector into a preset neural network for feature weight adaptive adjustment, and outputting an adjusted weight vector; according to the fire-fighting hidden danger intelligent identification method based on the large-model multi-mode, intelligent closed loop from hidden danger identification to prevention and control decision making is realized, and the intelligent, dynamic and precise management level of a fire-fighting safety system is comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of fire safety technology, specifically to a method for intelligent identification of fire hazards based on a large-scale multimodal model. Background Technology

[0002] Fire hazard management involves the identification and prevention of potential risks. Its key significance lies in its direct relation to public safety and property protection. In industrial production, urban buildings, and daily facilities, any oversight can trigger large-scale disasters and cause irreversible losses.

[0003] Current methods largely rely on manual inspections or simple rule matching, failing to effectively integrate multi-source information. This leads to risk assessments being detached from real-world dynamic scenarios. For example, in complex environments, hazards often involve multiple attribute changes, but existing solutions struggle to capture these cross-influences, amplifying assessment biases. Especially when external conditions fluctuate, risk judgments can easily deviate from the actual hazard evolution path. The core technical challenge lies in the quantification, transformation, and fusion mechanism of multi-dimensional feature information. Firstly, the highly heterogeneous nature of multi-source features such as hazard type, environmental factors, and time series makes it difficult to unify risk calculations. This directly causes biases in probability estimation. For instance, the ignition probability of an electrical equipment hazard in a high-temperature and high-humidity environment should vary over time. However, the lack of effective bridging between heterogeneous features causes the calculation model to ignore the amplification effect of the environment on hazard types, further amplifying the underestimation of potential losses. In specific business scenarios, such as when flammable materials are stored in a factory warehouse, the combination of rising nighttime temperatures and historical small fires should have increased the risk score, but due to the lack of fusion, it is treated as an isolated event, resulting in distorted risk scoring. Therefore, how to transform these multi-source heterogeneous characteristics into unified quantitative risk indicators, and based on this, accurately calculate the probability and extent of damage caused by each hidden danger to fire, has become a key issue in building a reliable prevention and control system. This issue further extends to decision support based on assessment results. How to automatically generate prevention and control strategies and resource allocation plans according to risk distribution, avoiding the unevenness and delays caused by subjective human intervention, such as resource allocation in high-risk areas needing to match the spatiotemporal dynamics of hidden dangers, but currently there is a lack of such an adaptive mechanism, leading to a disconnect between prevention and control measures and actual threats, which has become a core obstacle to improving management efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent identification method for fire hazards based on a large model and multimodal approach, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification of fire hazards based on large-scale multimodal models, the method comprising the following steps: S1. Obtain multi-source heterogeneous feature data through a sensor network deployed in the fire pump room to obtain an initial feature set; S2. Perform a uniform scale transformation on the initial feature set, calculate heterogeneous bridging parameters based on environmental parameters and equipment operation data, construct a fused feature vector, input the fused feature vector into a preset neural network for adaptive adjustment of feature weights, and output the adjusted weight vector. S3. Calculate the cross-influence value of environmental parameters and equipment vibration based on the adjusted weight vector. If the cross-influence value exceeds the preset threshold, the fused features are weighted by combining the material attenuation coefficient with the vibration peak value, and an enhanced weighted feature representation is generated by comparing historical trends. S4. Construct a risk probability model using enhanced weighted feature representation and environmental corrosion influencing factors as input, and output window-level probability values; connect the continuous window probability values ​​in time order to form a probability distribution curve, and determine whether the time series integration parameter adjustment is triggered based on the overall fluctuation characteristics of the probability distribution curve. If triggered, generate an optimized probability curve, and calculate the final window probability value of the fire pump hazard based on the optimized probability curve. S5. Compare the final window probability value with the preset baseline. If it is higher than the baseline, retrieve similar events from the historical damage record database based on the window corresponding to the probability distribution curve, calculate the event similarity, and perform weighted fusion with the window probability value to obtain the probability estimation result.

[0006] As can be seen from the above technical solution, the present invention has the following beneficial effects: This invention effectively overcomes the problems of difficulty in uniformly quantifying multi-source heterogeneous features, large risk assessment biases, and lagging prevention and control strategies in existing technologies. By deploying a multi-type sensor network within the fire protection system, it collects multi-dimensional data such as equipment vibration, environmental parameters, and operational sequences in real time. A deep neural network model is used to normalize, fuse, and weight the initial features, achieving a unified expression and adaptive correction of heterogeneous features, fundamentally improving the accuracy of hazard feature quantification. Based on this, the system, through dynamic feature correlation analysis and risk probability modeling, can calculate the occurrence rate and loss probability of each hazard type in real time, significantly improving the accuracy and response speed of risk identification. Simultaneously, this invention introduces historical sample calibration and model self-learning mechanisms to automatically correct prediction biases, ensuring stability and reliability during long-term operation. Furthermore, based on the identification results, this invention can automatically generate prevention and control and resource allocation schemes based on the spatiotemporal distribution of risks, and form adaptive prevention and control strategies through optimization and iterative cycles, thereby achieving an intelligent closed loop from hazard identification to prevention and control decision-making, comprehensively improving the intelligent, dynamic, and precise management level of the fire safety system. Attached Figure Description

[0007] Figure 1 This is an internal operation diagram of the intelligent identification method for fire hazards based on large-scale multimodal models in this invention. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] like Figure 1 As shown, this invention provides a technical solution: a method for intelligent identification of fire hazards based on a large-scale multimodal model, comprising: S1, acquiring multi-source heterogeneous feature data through a sensor network deployed in a fire pump room to obtain an initial feature set; S2, performing a unified scale transformation on the initial feature set, calculating heterogeneous bridging parameters based on environmental parameters and equipment operation data, constructing a fusion feature vector, inputting the fusion feature vector into a preset neural network for adaptive adjustment of feature weights, and outputting an adjusted weight vector; S3, calculating the cross-influence value of environmental parameters and equipment vibration based on the adjusted weight vector, if the cross-influence value exceeds a preset threshold, then weighting the fusion features by combining the correlation between material attenuation coefficient and vibration peak value, and generating an enhanced weighted feature representation by comparing historical trends; S4, constructing a risk probability model using the enhanced weighted feature representation and environmental corrosion influencing factors as input, and outputting window-level probability values; connecting continuous window probability values ​​in chronological order to form a probability distribution curve, determining whether time series integration parameter adjustment is triggered based on the overall fluctuation characteristics of the probability distribution curve, if triggered, generating an optimized probability curve, and then adjusting the optimized probability curve based on the time series integration parameter adjustment; S5. Calculate the final window probability value of fire pump hazards; S6. Compare the final window probability value with the preset baseline. If it is higher than the baseline, retrieve similar events from the historical damage record database based on the corresponding window of the probability distribution curve, calculate the event similarity, and weight and fuse it with the window probability value to obtain the probability estimation result; S7. Based on the probability estimation result and equipment damage data, dynamically extract loss features using a self-evolving risk perception model, and achieve multi-dimensional loss assessment through a cost feedback mechanism to generate preliminary loss indicators; S8. Adjust the data weights for the preliminary loss indicators to highlight key loss factors. Adjust the data weights through dynamic weight allocation and correlation matrix construction, while incorporating risk score calculation and historical data calibration to determine the comprehensive risk score; S9. If the comprehensive risk score exceeds the warning threshold, generate a resource allocation plan based on spatiotemporal dynamic distribution, and simultaneously start an optimization iteration loop to incorporate a deviation correction mechanism to obtain a preliminary prevention and control strategy; S10. Analyze the score threshold setting according to the preliminary prevention and control strategy. If the deviation exceeds the preset range, execute an adaptive adjustment loop to integrate historical data verification and scheme compatibility checks to determine the final adaptive prevention and control scheme.

[0010] This method collects heterogeneous data from multiple sources, including ambient temperature, humidity, vibration, current, and voltage, by deploying a multimodal sensor network within a fire pump room. The data is first standardized to unify the scale, and bridging coefficients between different data types are calculated to achieve deep fusion of data from different sources. The fused feature vector undergoes adaptive weight learning through a configured multi-layer neural network structure to output an optimized weight vector, further reflecting the importance of each feature in risk modeling. Subsequently, based on the nonlinear interaction between equipment vibration and environmental parameters, cross-influence values ​​are calculated and dynamically weighted using material property coefficients to generate enhanced features. These enhanced features and corrosion factors are input into a probabilistic model, which, based on time-series sliding window analysis, generates a time probability curve reflecting risk changes. When the curve exhibits significant fluctuations, dynamic parameter integration is triggered, thereby optimizing the window probability output. High-probability outputs will trigger searches of historical databases to match potential damage events and estimate similarity, further improving probability accuracy. Based on this, a multi-dimensional loss index is constructed through a risk perception model and a feedback evaluation mechanism. Key risk weights are dynamically updated based on historical loss data. Combined with the correlation matrix, a comprehensive risk scoring system is formed, which is ultimately used to drive the generation of resource strategies and the real-time adaptation and evolution of prevention and control plans.

[0011] This method enables unified integration and intelligent processing of complex, multi-source data, effectively addressing the issues of information fragmentation and slow response in traditional manual inspections. Adaptive learning via neural networks automatically adjusts feature importance, improving the generalization ability of the identification model and the accuracy of risk assessment. Dynamic monitoring of cross-influence quantities combined with vibration characteristics enhances feature representation capabilities, facilitating the timely identification of potential hazards. A sliding window mechanism in the probabilistic model supports continuous time-series risk modeling, effectively avoiding misjudgments caused by single-point data distortion. The loss assessment system, combining self-evolutionary algorithms and cost feedback, supports risk management from prediction to quantification and decision support. Furthermore, the proposed prevention and control strategy possesses adaptive optimization capabilities, adapting to changing on-site environments and improving overall fire safety prevention and control efficiency.

[0012] The technical process of simultaneously adjusting probability values ​​and calculating event similarity to obtain probability estimation results refers to the system correcting the initial risk probability value by incorporating similarity analysis of historical events after obtaining the initial risk probability value, thereby improving the accuracy and stability of fire hazard identification. The entire logical process includes the calculation steps for probability value adjustment, the implementation steps for event similarity calculation, and the coupling relationship between the two, as detailed below: First, when performing hazard identification, the system calculates a preliminary probability value based on weighted feature representation and a risk probability model. This value reflects the likelihood of a hazard occurring on the current equipment under existing monitoring conditions. Since single probability calculation relies solely on real-time data, it may be affected by factors such as data fluctuations, monitoring errors, or short-term environmental changes. Therefore, historical event information needs to be incorporated for correction. At this stage, the system first establishes a baseline probability, calculated from the average hazard probability of similar equipment in historical sample data, serving as a reference standard for the current judgment.

[0013] Secondly, the system performs event similarity calculation. The system retrieves historical events from the historical database that match the characteristics of the current equipment hazard. Each event record includes data such as a timestamp, event type, equipment operating conditions, and damage severity. The system represents the current hazard characteristics and historical event characteristics as vectors, and obtains the event similarity value by calculating the cosine similarity between the two sets of vectors. This similarity value ranges from 0 to 1; the closer the value is to 1, the more similar the current hazard is to the historical event, and the stronger its reference significance for the occurrence of risk. The system typically selects several historical events with the highest similarity (e.g., the top 5) as a reference set and calculates their average similarity value, which serves as the association parameter for this probability adjustment.

[0014] After completing the similarity calculation, the system adjusts the initial probability value based on the similarity results. The adjustment method is weighted fusion, which linearly weights the initial probability value and the average similarity value according to a fixed ratio. The weight allocation is based on empirical parameters determined during the model training phase, typically with an initial probability weight of 0.7 and a similarity weight of 0.3. After this weighting process, if the similarity value is high, the final probability value will be increased, indicating a higher risk; if the similarity value is low, the probability value will be relatively decreased, indicating a lower risk. In this way, the system maintains real-time data responsiveness while fully incorporating statistical references from historical experience to achieve dynamic balance.

[0015] The logical relationship between event similarity calculation and probability value adjustment is as follows: event similarity acts as a factor for probability correction, directionally adjusting the initial probability value. The system establishes a "similarity influence function" to bind similarity to the probability adjustment range, ensuring that the adjustment range is proportional to the similarity intensity. When the similarity is higher than 0.8, the system considers the current potential hazard to be almost identical to historical high-risk events, automatically increasing the probability by 10% to 20%; when the similarity is between 0.5 and 0.8, the system controls the probability adjustment range within 5%; when the similarity is lower than 0.5, the system does not make any adjustments, maintaining the initial calculation result. This logic ensures that the probability estimation results reflect both real-time feature data and inherit historical event patterns, thus stably outputting reliable risk probability values ​​under different environmental conditions.

[0016] Ultimately, the system defines the adjusted probability value as the "probability estimation result." This result not only reflects the potential risks calculated from the current data but also integrates statistical correlations and feature matching results from historical events, forming a comprehensive indicator that has undergone dual verification. The probability estimation result is directly input into the subsequent loss assessment and risk score calculation module to quantify the potential damage level and determine the risk grade.

[0017] S1 includes acquiring multi-source heterogeneous feature data from a building beam and column monitoring system via a sensor network to obtain an initial feature set; for the initial feature set, using principal component analysis (PCA) to extract equipment vibration factors and operating time series information to obtain a refined feature subset, where the PCA algorithm takes heterogeneous features from the initial feature set as input and outputs the dimensionality-reduced refined feature subset; based on the refined feature subset, fusing pre-acquired beam and column stress distribution data from the beam and column monitoring system to determine structural stability indices, where beam and column stress distribution data is acquired in real time by beam and column sensors; if the structural stability indices exceed a preset threshold, supplementary vibration data is acquired from backup sensors in the pump house network to obtain the initial feature set.

[0018] In this embodiment, in the monitoring system, the sensor network continuously collects relevant physical quantities of beams and columns according to a preset sampling time interval. At each sampling moment, the stress data collected by the beam and column sensors, the vibration amplitude and frequency data collected by the vibration sensors installed near the equipment, and the time information characterizing the equipment's operating status together form a multi-source heterogeneous data record. Multiple data records acquired within several consecutive sampling moments are superimposed in chronological order to form an initial feature set covering a fixed time window. Each record includes at least the sampling time, beam / column location identifier, stress value at that location, equipment vibration value associated with that location, and the start and end duration of equipment operation. To avoid the impact of differences in the dimensions and numerical ranges of different physical quantities on subsequent calculations, each numerical field in the initial feature set is normalized using the same standard. The normalization process calculates the average value and fluctuation range of the field within a selected time period, scales and shifts the original value by a determined ratio, so that the same field is distributed around zero in the normalized data and the fluctuation range is limited to a preset interval. All records after normalization constitute the principal component score. The input data set for the principal component analysis algorithm is as follows: When executing the principal component analysis algorithm, the pairwise correlation of each normalized field within a selected time period is first calculated to obtain a correlation matrix that reflects the correlation between the fields. Then, eigenvalue decomposition is performed on the correlation matrix to obtain several independent principal directions. Each principal direction corresponds to an eigenvector and an eigenvalue representing the importance of the principal direction. By arranging all eigenvalues ​​in descending order, their cumulative contribution ratio is calculated. When the cumulative contribution ratio first reaches or exceeds 80% of the total amount of all change information, the corresponding principal directions are determined as the retained principal components. The number of principal components is the dimension of the refined feature subset. After obtaining the principal components, each normalized feature record is linearly projected onto these principal component directions. The projection result is the coordinate of the record in the refined feature subset. The projection results of all records are arranged in chronological order to form the refined feature subset. Each dimension in the refined feature subset corresponds to comprehensive factors such as the overall intensity of equipment vibration, vibration change trend, energy distribution of vibration in different frequency ranges, and stability of the operating time series.When determining structural stability indices, beam-column sensors first collect stress distribution data at sampling intervals that are the same as or integer multiples of the vibration data. Each stress data point includes the beam-column location identifier, sampling time, and stress value at that location. Stress data from multiple beam-column locations within the same time window are organized into a stress distribution sequence according to the beam-column location index. Then, the refined feature subset within that time window is time-aligned with the stress distribution sequence of the corresponding time window, establishing a one-to-one correspondence on a time-window basis. Within each time window, the maximum, minimum, and average stress values ​​at all beam-column locations are first calculated, and then... Within this time window, the average values ​​of dimensions related to vibration intensity and the fluctuation amplitudes of dimensions related to vibration change trends in the refined feature subset are used. By assigning pre-defined weighting coefficients to the maximum stress, average stress, average vibration intensity, and vibration change amplitude, these quantities are weighted and summed. The result is then divided by the weighted sum of the allowable stress reference value and the allowable vibration change amplitude reference value within this window, resulting in a dimensionless structural stability index. A larger index value indicates that the structure is closer to an unstable state. The sampling time intervals involved in the above process are determined during the system design phase based on the rate of change in equipment operation and the rate of change in beam and column stress. To ensure that the changes in beam and column stress and equipment vibration between two adjacent samplings are within a sufficiently characterizable range, the time window length is determined by selecting a period of time no less than ten times the sampling time interval and ensuring that it covers the typical cycle of equipment operation. The number of principal components is determined by the condition that the cumulative contribution ratio of the principal components reaches or exceeds 80% of the total amount of all change information. During the system debugging phase, the weighting coefficients are determined according to the order that the impact of beam and column stress on structural safety is greater than that of vibration amplitude, and the impact of vibration amplitude on structural safety is greater than that of operating time fluctuation. Through multiple rounds of experiments, the deviation between the structural stability index and the actual observed structural response is calculated, and each weight value is gradually corrected until the deviation is controlled within the predetermined range. The preset threshold is determined by historical data and safety specifications. Specifically, during the trial operation phase before the building is put into use, beam and column stress distribution data and equipment vibration data are continuously collected for no less than 30 days. The structural stability index within each time window is calculated, and the structural stability indices corresponding to all time windows are sorted from smallest to largest. The element in the sequence whose position is equal to 90% of the total number of windows is found, and the structural stability index value corresponding to this element is used as the preset threshold. When the structural stability index exceeds this threshold, it indicates that the state is higher than the vast majority of normal operating states.During system operation, the real-time calculated structural stability index is compared with a preset threshold. If the structural stability index within a certain time window is greater than or equal to the preset threshold, it is determined that there is an abnormal risk to the current structural stability. The pump house network immediately sends a start command to the backup sensors. Upon receiving the command, the backup sensors collect supplementary vibration data at a higher sampling frequency than the conventional sensors within the same or immediately following time window. This supplementary vibration data includes vibration amplitude, vibration frequency, and vibration duration. This supplementary vibration data is then merged with the original vibration data in chronological order to reconstruct a new initial feature set. This new initial feature set undergoes normalization and principal component analysis again to update and refine the feature subset and structural stability index. This allows for a more refined characterization of the risk state when structural stability approaches the abnormal boundary by improving the temporal resolution and richness of the vibration data, reducing errors caused by insufficient data resolution or local data loss in the previous structural stability calculation.

[0019] S2 involves collecting ambient temperature data and equipment vibration sequences, performing mean subtraction and variance division on both to unify their scales, resulting in a unified-scale dataset; calculating the correlation coefficient between the temperature and vibration sequences based on the unified-scale dataset, and combining this with equipment operation logs to form a preliminary fusion vector; inputting the preliminary fusion vector into a convolutional neural network, where weights are adaptively adjusted through convolutional and fully connected layers to obtain an adjusted weight vector; and then weighting and fusing the adjusted weight vector with the average stress, maximum stress, and stress fluctuation amplitude collected by the beam-column monitoring system to calculate the structural environmental stability characteristics.

[0020] In this embodiment, ambient temperature and equipment vibration data are synchronously collected in the fire pump room using pre-deployed ambient temperature and vibration sensors at pre-set fixed sampling time intervals. Temperature values ​​acquired over several consecutive sampling periods are arranged chronologically to form a temperature sequence, and vibration amplitude values ​​at corresponding times are arranged chronologically to form a vibration sequence. The sampling time interval is determined during the system design phase based on the dominant frequency of equipment vibration and the rate of change of ambient temperature, ensuring that a time window contains at least several complete vibration cycles and reflects slow temperature changes. The length of the time window is selected as a multiple greater than or equal to the number of dominant vibration cycles and not less than [a certain value]. A time interval of 10 times the sampling time interval was determined to ensure that the data within the same window is both representative and computationally stable. After obtaining the temperature and vibration sequences, to eliminate the differences in the dimensions and numerical ranges of the two physical quantities, mean subtraction and variance division were performed on both for scaling. Specifically, for the temperature sequence within a time window, the arithmetic mean of all temperature samples within the window was first calculated as the temperature mean. Then, the difference between each temperature sample and the temperature mean was calculated. The sum of the squares of these differences was divided by the total number of samples minus 1 to obtain the temperature variance. The square root of the temperature variance was then taken to obtain the temperature standard deviation. Finally, the standard deviation was calculated for each temperature sample. The difference between the mean and the mean temperature is divided by the standard deviation of the temperature to obtain the standardized temperature value at that moment. For vibration sequences within the same time window, the same method is used to calculate the vibration mean, vibration variance, and vibration standard deviation. The difference between each vibration sample value and the mean is then divided by the standard deviation to obtain the standardized vibration value at that moment. The resulting standardized temperature and vibration sequences constitute a unified-scale dataset. Each time point in this dataset contains one standardized temperature value and one standardized vibration value. The length of the unified-scale dataset is equal to the number of sampling points within the time window. The number of sampling points is determined during the system configuration phase by the sampling time interval and... The window length is determined jointly. After obtaining the unified scale dataset, in order to describe the synchronicity between temperature changes and vibration changes, the correlation coefficient between the temperature sequence and the vibration sequence is calculated based on the unified scale dataset. Specifically, within a time window, for each sampling moment, the product of the standardized temperature value and the standardized vibration value at that moment is taken. The sum of the products of all sampling moments is divided by the total number of samples minus 1 to obtain a correlation coefficient in the range of negative 1 to positive 1. The closer the correlation coefficient is to positive 1, the higher the degree of simultaneous occurrence of temperature increase and vibration enhancement. The closer it is to negative 1, the more obvious the inverse relationship between temperature increase and vibration weakening. The closer it is to 0, the weaker the correlation between the two.After the correlation coefficient calculation is completed, in order to incorporate both the environment and equipment operating status into the subsequent identification model, the unified-scale dataset and the equipment operation logs are time-aligned. The equipment operation logs have a defined recording format during the equipment installation and commissioning phase. Each log entry includes at least a timestamp, equipment on / off status, equipment operating condition level, and alarm flags. In the algorithm implementation, the log records are first divided into various time windows based on the timestamps. For each time window, the total duration of the equipment being on, the proportion of time the main operating condition levels occur, and whether alarm flags appear are statistically analyzed. The equipment on / off status and operating condition level are encoded into numerical values ​​according to a preset mapping rule, such as the off status. The operating condition is encoded as 0 for the active state and 1 for the active state. Each operating condition level is mapped sequentially to integers 1, 2, 3, etc., and standardized using the same mean subtraction and variance division methods as the unified-scale dataset. This ensures that the values ​​in the operation logs are within a similar range to the temperature and vibration data. Subsequently, within each time window, the standardized temperature and vibration value sequences from all moments in the unified-scale dataset, along with the operating condition encoding sequence obtained from the operation logs, are concatenated chronologically into a three-channel time series. The correlation coefficient corresponding to this window is added as an additional numerical feature reflecting the overall correlation degree to the description of this time series, forming the preliminary fusion for that time window. The vector consists of a preliminary fusion vector for each time window. The length of the preliminary fusion vector is determined by the number of sampling points and channels within the time window. An additional correlation coefficient is used as an independent component of this vector. The set of preliminary fusion vectors will serve as the input to the convolutional neural network. In the design of the convolutional neural network, the structural parameters of the convolutional layers and fully connected layers are first determined. The number of convolutional kernels, the length of the convolutional kernels in the time direction, and the convolutional stride are determined experimentally during the system training phase. The experimental process is as follows: sample windows that have been manually analyzed and labeled with historical events are selected from historical running data to construct training and validation datasets. Multiple candidate convolutional kernel numbers, lengths, and strides are set respectively. The convolutional kernel length and stride are combined, and the convolutional neural network is trained for several rounds for each combination. The prediction error of the risk-related indicators on the validation dataset is recorded, and the set of parameters with the smallest prediction error is used as the final convolutional layer parameters. The number of neurons in the fully connected layer is determined according to the feature dimension of the output of the convolutional layer and the dimension of the weight vector to be output. The dimension of the weight vector is determined in the system design stage according to the number of structural environment stability-related features to be described. For example, when it is necessary to characterize the importance of the temperature channel, vibration channel and operating status channel respectively, the weight vector contains at least 3 components. If it is necessary to distinguish the importance of different time periods or frequency bands more finely, the number of components can be increased accordingly.During the training phase, the preliminary fusion vectors corresponding to historical sample windows are input into a convolutional neural network. The convolutional layer extracts temperature change patterns, vibration change patterns, and operational state change patterns in the time dimension through sliding convolution operations. The extracted local patterns are combined into higher-level temporal feature representations, which are then fed into a fully connected layer. The fully connected layer maps these features into a weight vector and, based on the error between the weight vector and the target output, gradually adjusts the parameter values ​​in the convolutional and fully connected layers using a pre-selected parameter update rule. The target output can be the structural environment stability level or hazard risk level labeled according to historical events. The parameter update rule is implemented using... A fixed learning rate, a fixed number of training epochs, and a fixed batch size are determined through a small number of trials before training to ensure stable convergence of the training error and prevent significant overfitting. After training, the preliminary fusion vectors corresponding to each time window during the current operation are input into the trained convolutional neural network. The network outputs an adjusted weight vector through convolutional and fully connected layers. Each component of this weight vector corresponds to the importance value of a specific input feature or a certain type of input feature; the larger the value, the greater the impact of that feature or feature category on the stability of the structural environment. To integrate the adjusted weight vector with the average stress and maximum stress values ​​collected by the beam-column monitoring system... The stress fluctuation amplitude and the maximum stress value are weighted and fused. First, the beam-column monitoring system collects stress data at each beam-column location within the time period corresponding to the aforementioned time window. The stress average value for that window is obtained by averaging the stress data at all sampling times within the same time window. The maximum stress value for that window is then identified as the maximum stress value for all sampling times within the same time window. Simultaneously, the difference between the stress value at each sampling time within the window and the stress average value is calculated. The absolute values ​​of these differences are averaged to obtain the stress fluctuation amplitude for that window. The mean subtraction and variance division processes, similar to those used for temperature and vibration, are then applied to the stress average value, maximum stress value, and stress fluctuation amplitude to achieve the desired result. The quantities are also at a uniform scale; subsequently, several components pre-assigned to the structural stress-related features are selected from the adjusted weight vector. These components are then summed after absolute value processing, and each component is divided by this sum to obtain three normalized weight coefficients. These coefficients are multiplied by the average stress, maximum stress, and stress fluctuation amplitude at the uniform scale, respectively. The three products are then summed to obtain a single value, which serves as the structural environmental stability feature, representing the overall stability under the combined effects of environmental temperature changes, equipment vibration changes, and beam-column stress responses within the current time window. The selected weight component numbers establish a fixed correspondence with the stress-related input features during the network design phase.

[0021] S3 includes calculating the mean temperature, standard deviation of temperature, rate of change of temperature, root mean square value of vibration, maximum peak value of vibration, and count of abnormal peak values ​​from real-time collected ambient temperature and equipment vibration sequences, and forming an initial vector in a fixed order; using the initial vector and the fused feature vector obtained in S2 as input, the interaction strength between environmental parameters and equipment vibration is calculated through cross-influence assessment to obtain a cross-influence value; if the cross-influence value exceeds a preset threshold, the weights of the fused feature vector are adjusted, and the components of temperature change rate, peak vibration value, and count of abnormal peak values ​​are amplified by a preset ratio to generate an adjusted weight vector; based on the adjusted weight vector and the correlation parameters between the equipment material attenuation coefficient and the peak vibration value, a weighted operation is performed on the fused feature vector to obtain a weighted feature representation; the weighted feature representation is compared with the historical operating trend, and the weighted feature representation is enhanced according to the trend deviation ratio to obtain an enhanced weighted feature representation.

[0022] In this embodiment, during the operation of the fire pump room, ambient temperature and equipment vibration sequences are synchronously collected at fixed sampling time intervals. The sampling time interval is determined during the system design phase based on the dominant frequency range of equipment vibration and the actual rate of temperature change. This ensures that a time window can cover several complete vibration change processes while also providing sufficient data points for statistical analysis of temperature changes within that window. The length of the time window is determined by selecting a duration that is not less than 10 times the sampling time interval and can cover typical operating conditions of the equipment. Within each time window, the system first calculates the average temperature for all temperature samples within that window. Specifically, this involves summing all temperature samples one by one within the window and then calculating the sum. Dividing the average temperature by the number of temperature samples within the window yields the mean temperature for that window. The standard deviation is then calculated by subtracting the mean temperature from each sample value, obtaining the deviation of that sample value relative to the mean. The squares of each deviation are then summed, and the sum is divided by the number of temperature samples minus 1 to obtain the variance. The square root of the variance is then taken to obtain the standard deviation. The rate of change of temperature is calculated based on the change between two adjacent temperature samples within the time window. For each pair of adjacent sampling times, the difference between the temperature at the later sampling time and the temperature at the earlier sampling time is calculated. This difference is then divided by the time interval between the two sampling times to obtain the first rate of change. The absolute value of all first rates of change within the window is then calculated. The average value is used to obtain the temperature change rate of the window, which is used to characterize the strength of the overall temperature fluctuation within that time period. The processing of the equipment vibration sequence is as follows: First, the root mean square value of vibration is calculated. Within a time window, all vibration amplitudes are squared one by one, then all squared values ​​are added together, and the sum is divided by the number of vibration sampling values. Finally, the square root of the sum is taken to obtain the root mean square value of vibration for that window, which reflects the overall level of vibration energy within that time period. The maximum vibration peak value is determined by the largest value among all vibration sampling values ​​within the window, and this value is directly used as the maximum vibration peak value for that window. Abnormal peak counts are achieved by setting an abnormal peak judgment threshold. This threshold is determined during the system debugging phase by analyzing a large number of vibrations under normal operating conditions. The vibration data is statistically analyzed to determine the abnormal peak value. Specifically, during the initial operation or trial operation of the equipment, vibration data is continuously collected for no less than 30 days. The maximum vibration values ​​within all time windows are arranged in ascending order, and the value corresponding to 95% of the time windows is selected as the abnormal peak value judgment threshold. This ensures that under normal operating conditions, only a small number of time windows will have a maximum vibration value that reaches or exceeds this value. During the formal operation of the system, for each vibration sequence within a time window, all vibration sample values ​​are checked one by one. When a vibration sample value is greater than or equal to the above abnormal peak value judgment threshold, the sample value is regarded as an abnormal peak value. The number of all sample points that meet the condition within the window is accumulated as the abnormal peak value count for that window.After calculating the mean temperature, standard deviation of temperature, rate of temperature change, root mean square value of vibration, maximum peak value of vibration, and count of abnormal peaks, these six quantities are arranged in a fixed order to form an initial vector. This order is determined during the system design phase and remains unchanged throughout the entire operation. The order is: mean temperature, standard deviation of temperature, rate of temperature change, root mean square value of vibration, maximum peak value of vibration, and count of abnormal peaks, ensuring that the meaning of each component is stable and clear in subsequent calculations. After obtaining the initial vector, it is aligned with the fused feature vector obtained in step S2 within the same time window. The initial vector is used as the reference for environmental and vibration systems. To supplement the feature set, the fused feature vector is used as a feature reflecting the results of multi-source fusion and neural network adaptive learning. These features are numerically concatenated in a predetermined order to form an extended input vector for the window. Based on this extended input vector, a cross-influence assessment is performed to calculate the interaction strength between environmental parameters and equipment vibration. In specific implementation, the system first recalculates the correlation between the temperature and vibration sequences processed at a uniform scale within the time window. This is done by multiplying the standardized temperature and vibration values ​​at each sampling time, averaging the products across all sampling times, and obtaining a value reflecting the degree of unidirectional change. The correlation value is calculated; simultaneously, a coupling strength is constructed using the temperature change rate and the root mean square value of vibration. This coupling strength is obtained by directly multiplying the temperature change rate and the root mean square value of vibration, and is used to represent the linkage strength between temperature fluctuations and vibration energy. Then, an anomaly proportion is constructed using the abnormal peak count. Dividing the abnormal peak count by the number of vibration samples within the window yields a proportion between 0 and 1, reflecting the frequency of abnormal vibration peaks within the window. After calculating the above three intermediate quantities, the correlation value, coupling strength, and anomaly proportion are multiplied by three values ​​pre-set during the system debugging phase. The three proportional coefficients are calculated and summed, and the sum is used as the cross-influence value of the time window. The above three proportional coefficients are determined by offline calculation of historical data. Specifically, historical time window data including known hidden danger events and normal operation status are selected, and the relevance value, coupling strength and abnormality ratio of each window are calculated respectively. These quantities are combined into different weighted forms, and the system's ability to distinguish between hidden danger windows and normal windows under different weighted combinations is compared. The classification accuracy and the identification recall rate of hidden danger windows are used as evaluation criteria, and the proportional coefficient set that can take into account both high accuracy and high recall rate is selected as the final proportional coefficient.The preset threshold for cross-influence values ​​is also determined based on historical data. Specifically, in the aforementioned historical data, only time windows confirmed to be in normal operation are counted. The cross-influence value for each normal window is calculated, and these cross-influence values ​​are sorted from smallest to largest. The value corresponding to the element in the sequence whose position equals 95% of the total number of normal windows is selected as the preset threshold. This ensures that states with cross-influence values ​​greater than this threshold only occur at a low proportion in normal samples. Therefore, during the online operation phase, when the cross-influence value of a certain time window is greater than or equal to this preset threshold, it can be considered that the interaction intensity between environmental parameters and equipment vibration within that window is abnormally high, requiring increased attention to related characteristics. Once the cross-influence value... If the threshold is exceeded, the system performs weight adjustment on the fused feature vector obtained in S2. Specifically, the system pre-reserves the positions of components corresponding to the three types of information—temperature change rate, vibration peak value, and abnormal peak count—in the fused feature vector. During system design, it is clearly defined which type of physical quantity each component corresponds to. When an abnormal cross-influence value is detected, these three components are read from the fused feature vector. The component representing the temperature change rate is multiplied by a first amplification factor, the component representing the vibration peak value is multiplied by a second amplification factor, and the component representing the abnormal peak count is multiplied by a third amplification factor. All three amplification factors are set to values ​​greater than 1 to amplify information related to rapid changes and abnormal vibrations in the feature space. These three amplification factors are used in the system... During the debugging phase, multiple sets of experiments were conducted to determine the final proportions. The experimental method involved varying the combination of three proportional coefficients on the same historical sample set, comparing the impact of each combination on the subsequent hazard identification model output, and selecting the combination that significantly improved confidence in hazard samples and resulted in an acceptable increase in false alarms in normal samples as the final proportion. After amplifying these three components, other components in the fused feature vector remained unchanged, and the resulting vector was the adjusted weight vector. Based on the adjusted weight vector, the system combined the equipment material attenuation coefficient with the vibration peak correlation parameter to perform a weighted operation on the fused feature vector. The equipment material attenuation coefficient was determined during the equipment documentation phase based on the equipment material type, design service life, and factory test data. For different materials or different models... Different values ​​are given to the equipment. The larger the value, the more significant the material performance degradation under the same vibration load. After installation, the coefficient is calibrated through a period of operation monitoring and maintenance records to make it closer to the actual use. The vibration peak correlation parameter is determined by analyzing historical fault records and vibration peak statistics. Specifically, the distribution of the maximum vibration peak and abnormal peak counts in the historical data are statistically analyzed within a certain period before the occurrence of various faults. The parameters corresponding to the working conditions in which high vibration peaks and high abnormal counts frequently occur before the fault are set to larger values, and the parameters corresponding to the working conditions in which abnormal peaks rarely occur under normal operation are set to smaller values, thus forming a vibration peak correlation parameter library divided by equipment type and operating environment.In the specific execution of the weighted calculation, the system first reads the corresponding material attenuation coefficient from the equipment file according to the equipment type corresponding to the current time window, and selects the vibration peak correlation parameter that matches the equipment and the current operating condition from the vibration peak correlation parameter library. Then, the vibration-related component in the adjusted weight vector is multiplied by these two parameters, and then multiplied by the corresponding component in the original fused feature vector. This ensures that the vibration-related features in the final weighted feature representation simultaneously reflect the material vulnerability and the correlation strength between historical vibration peaks and faults. For components related to ambient temperature or other operating conditions, only the corresponding weight in the adjusted weight vector is multiplied by the value in the original fused feature vector. Weighting is performed without additional multiplication by the material attenuation coefficient or vibration peak correlation parameter. After this processing, the weighted results of all components are recombined into a new feature vector, which is the weighted feature representation of the current time window. To further enhance the weighted feature representation using historical operating trend information, the system maintains a historical operating trend database. This database is gradually built and updated during the long-term operation of the system, recording the statistical characteristics of each feature component over time under normal and slightly abnormal conditions for different equipment types. For each feature component, the historical operating trend database stores its average value and allowable fluctuation range under normal conditions. After obtaining the weighted feature representation of the current time window, the system... The system performs trend comparison on each component in the vector. Specifically, it reads the average value and allowable fluctuation range of the component under normal conditions from the historical trend database, calculates the absolute value of the difference between the current component value and the historical average value, and then divides this difference by the allowable fluctuation range to obtain the trend deviation ratio of the component. The larger the trend deviation ratio, the greater the deviation of the current state from the historical normal state. Then, it averages the trend deviation ratios of all components in the weighted feature representation to obtain an overall trend deviation ratio, which reflects the strength of the deviation of the overall feature state from the historical normal trend. During enhancement processing, when the overall trend deviation ratio is less than 1, the current weighted feature representation is considered to be... The system maintains the overall trend of the current state being similar to the historical normal trend. Each component in the weighted feature representation remains unchanged and is directly output as the enhanced weighted feature representation. When the overall trend deviation ratio is greater than or equal to 1, the current state is considered to deviate significantly from the historical normal trend. The system then multiplies each component of the weighted feature representation by the overall trend deviation ratio, thus amplifying the window with a large deviation from the historical trend in the feature space. This generates an enhanced weighted feature representation. In subsequent steps, when using this enhanced weighted feature representation for risk probability modeling, it can better highlight the impact of rapid temperature changes, strong vibrations, and their coupling effect with equipment material degradation on the formation of fire hazards.

[0023] "Enhanced weighted feature representation" is an important intermediate result of this invention in the multi-source feature fusion and risk correlation enhancement stages. It is a high-dimensional feature expression generated after dynamic weight adjustment, historical trend comparison, and risk correlation enhancement, based on the initial weighted feature representation. This technical feature is used to more accurately reflect the potential coupling relationship between equipment operating status and environmental changes in complex fire monitoring scenarios, thereby providing more reliable input data for subsequent risk probability calculation and hazard identification. Its specific connotation and formation process are as follows: First, the foundation of enhanced weighted feature representation is "weighted feature representation." This weighted feature representation originates from the preliminary weighted calculation results of multi-source heterogeneous data, such as equipment vibration parameters, ambient temperature parameters, and beam-column stress parameters, within the feature fusion model. This result is a comprehensive vector formed by assigning adaptive weights to different types of features, used to describe the overall state of the operating environment of the fire pump room. For example, the combined effects of changes in equipment vibration intensity, temperature rise rate, and structural stress fluctuations, through preliminary weighting, form a multi-dimensional feature set that numerically reflects the contribution of different factors to potential risks.

[0024] Secondly, in generating the enhanced weighted feature representation, the system first performs cross-influence analysis to calculate the interaction strength between environmental parameters and equipment vibration. When the cross-influence value exceeds a set threshold (e.g., 0.6), the system identifies a significant coupling relationship between the two and automatically performs weight adjustment. The logic of weight adjustment is as follows: if environmental changes and equipment vibration have a positive reinforcing relationship (i.e., increased temperature leads to increased vibration), the weight ratio of vibration parameters is increased; if they have a suppressive relationship (i.e., increased temperature leads to decreased vibration), the weight of environmental parameters is increased. After adjustment, the new feature weight vector redistributes the importance of each parameter, thus making the fusion result more closely reflect the actual operating state.

[0025] Next, the system introduces equipment material attenuation characteristics and vibration peak correlation parameters to correct for the time-cumulative effect. The equipment material attenuation parameter reflects the trend of declining mechanical performance under long-term operation; this parameter is determined by the equipment's service life, operating time, and historical vibration spectrum change rate. The vibration peak correlation parameter identifies the impact of instantaneous impact vibrations on system stability, using the ratio of the average maximum vibration amplitude over a past period as an indicator. The system then uses these two parameters as correction coefficients and fuses them again with the aforementioned weighted feature vector, ensuring that the output features not only reflect the current state but also include response characteristics to historical attenuation and instantaneous anomalies.

[0026] Finally, to ensure the model's temporal continuity and trend consistency, the system introduces a historical trend comparison process after fusion. The system extracts weighted feature records of the same type of equipment under similar operating conditions from the historical database and calculates the consistency between the current weighted features and historical records over time. If the consistency is higher than 0.8, the system considers the current state to conform to historical safe operating patterns and makes no adjustments; if the consistency is lower than 0.8, the system determines that there is a potential anomaly and amplifies the anomaly signal by increasing the weight ratio of risk-related features, thereby strengthening the risk perception capability. The feature result output after this process is the "enhanced weighted feature representation".

[0027] In summary, the enhanced weighted feature representation is not a simple feature superposition, but a dynamic representation achieved through a triple enhancement mechanism: (1) adaptive weight enhancement, adjusting feature weights through cross-influence analysis; (2) historical memory enhancement, correcting time factors by introducing material attenuation and vibration peak correlation; and (3) trend comparison enhancement, strengthening potential risk signals through historical consistency comparison. This representation vector combines real-time performance and stability, and can comprehensively reflect the immediate state of equipment operation, historical degradation characteristics, and potential hidden danger trends, providing a highly reliable data foundation for subsequent risk probability calculation, event matching, and prevention and control decisions.

[0028] Therefore, the "enhanced weighted feature representation" is the core intermediate result of the present invention in realizing dynamic risk association enhancement in the feature fusion stage. It is a high-dimensional fusion expression of multi-source heterogeneous data after weight optimization, time correction and trend verification. Its technical essence lies in enabling the model to more accurately identify hidden danger patterns and potential failure trends in the fire protection system through precise control of parameter weights and adaptive alignment of historical features.

[0029] S4 includes collecting equipment vibration data and environmental corrosion data through sensors, aligning them by timestamps, and generating a time-series dataset; based on the obtained enhanced weighted feature representation, calculating the root mean square value, the maximum peak value of vibration, and the count of abnormal peaks exceeding three standard deviations from the equipment vibration sequence, multiplying the three indicators by preset weights and summing them to obtain the equipment vibration correlation strength; using the equipment vibration correlation strength and environmental corrosion rate as inputs to construct a risk probability model, outputting a window-level preliminary probability value, and forming a preliminary probability distribution curve in chronological order; if the preliminary probability value exceeds a preset threshold, adjusting the time-series integration parameters, including increasing the analysis window length from 10 seconds to 20 seconds, shortening the sliding step from 1 second to 0.5 seconds, and adjusting the vibration smoothing length from 5 points to 7 points; recalculating the probability sequence based on the adjusted time-series integration parameters and generating an optimized probability curve; performing probability averaging and trend integration on the optimized probability curve at fixed intervals to obtain the final window probability value of the hidden danger.

[0030] In this embodiment, at the fire pump room site, installed sensors continuously collect equipment vibration data and environmental corrosion data. Equipment vibration data is recorded at fixed sampling intervals by vibration sensors positioned near the pump and its foundation. During the system design phase, this sampling interval is set to be no more than one-tenth of the equipment vibration's main cycle to ensure a detailed depiction of vibration changes. Environmental corrosion data is periodically collected by corrosion monitoring sensors located at typical locations, including indicators such as the concentration of corrosive media in the environment, the corrosion depth on the metal surface, or the corrosion potential. The environmental corrosion data sampling interval is set as an integer multiple of the vibration data sampling interval based on the corrosion change rate to ensure that different transmission rates are maintained. The data collected by the sensors are temporally correlated. The system uses the same time base across all acquisition devices, adding a timestamp accurate to the second or higher to each vibration and corrosion data point. At the data receiving end, the data is aligned according to the timestamps, and all vibration and corrosion data within the same time period are grouped into consecutive, non-overlapping or partially overlapping time windows. The length of each time window is set to a continuous 10 seconds by default. All vibration sampling values ​​sorted by time within a time window are combined into a vibration time series, and corrosion monitoring values ​​collected within that time window are combined into a corrosion time series. If no new corrosion sampling value is found at a certain time point, the most recent acquisition value before that point is used. The corrosion values ​​are filled to form a time-series dataset aligned to timestamps. Within each time window, based on the enhanced weighted feature representation obtained in step S3, the system calculates the root mean square (RMS) value of vibration, the maximum peak value of vibration, and the count of outliers exceeding three standard deviations from the device vibration time series within that window. The calculation process for the RMS value is as follows: first, square each vibration sample value within the window; sum all the squared values; then divide the sum by the total number of vibration sample values ​​within the window to obtain an intermediate result of the average square value; finally, take the square root of this intermediate result to obtain the RMS value of vibration for that window. This value is used to characterize the vibration within that time period. The overall level of vibration energy; the calculation of the maximum peak value of vibration is to iterate through all vibration sample values ​​within the window, and take the sample value with the largest value as the maximum peak value of vibration in that window, which is used to reflect the strongest vibration intensity that occurred in that time period; the calculation of the number of abnormal peak values ​​exceeding three standard deviations is as follows: first, calculate the average value of vibration sample values ​​within the window, by summing all vibration sample values ​​within the window and dividing by the number of vibration sample values ​​to obtain the vibration mean, then subtract the vibration mean from each vibration sample value to obtain the deviation, square each deviation, sum all the squared deviation values ​​and divide by the number of vibration sample values ​​minus 1 to obtain the vibration variance, and take the square root of the vibration variance to obtain the vibration standard deviation;After calculating the vibration standard deviation, the system sets the threshold for abnormal peak values ​​within the window to three times the vibration mean plus the vibration standard deviation. During the process of iterating through all vibration samples within the window, the system counts the number of samples with values ​​greater than or equal to this threshold. This count represents the abnormal peak count exceeding three times the standard deviation, reflecting the frequency of strong abnormal vibrations. After obtaining the vibration root mean square value, the maximum vibration peak value, and the abnormal peak count, the system multiplies each of these three indicators by their corresponding preset weights according to preset rules and sums them to obtain the equipment vibration correlation intensity. The three weights are determined based on historical data during the system's offline training phase, specifically selecting values ​​including normal operation, minor abnormalities, and confirmed abnormalities. Historical time window samples, including potential hazard events, are used to calculate three indicator values ​​for each sample. Whether a sample belongs to a hazard window or a risk window is used as the labeling result. By iterating through multiple different weight combinations, the weighted sum of the three indicators is calculated for each weight combination, and its distinguishing ability from the labeled results is observed. The hazard window identification rate and the false alarm rate of the normal window are used as evaluation indicators. A weight combination where the hazard window identification rate is not lower than a specified percentage and the false alarm rate of the normal window is not higher than a specified percentage in historical data is selected as the final preset weight. In this way, the three weights remain fixed after the system is put into operation and do not automatically adjust with changes in online data. After obtaining the equipment vibration correlation intensity, this intensity value is... The environmental corrosion rate, along with other parameters, is used as input to construct a risk probability model. The environmental corrosion rate is obtained within each time window by comparing the current corrosion index with the changes in corrosion indices over several previous time windows. Specifically, the average change in corrosion indices over a fixed time period is calculated and divided by the period length to obtain the corrosion rate for that period. This corrosion rate is then mapped to the input value corresponding to the current time window. The risk probability model is implemented using a multi-layered model with a fixed structure. Before system deployment, this model is trained offline using a large number of historical time window samples. Each sample includes the corresponding equipment vibration correlation intensity, environmental corrosion rate, and an annotation by experts indicating whether there are fire hazards during that time period. The training process involves gradually adjusting the model's internal parameters to make the model's predicted probability of historical samples as close as possible to the label results. After training, the model parameters are fixed. During system operation, the equipment vibration correlation intensity and environmental corrosion rate of each time window are input into the risk probability model. The model outputs a value between 0 and 1, representing the probability of a fire hazard existing in that time window. This value is the window-level preliminary probability value. Connecting the preliminary probability values ​​of each time window in chronological order forms a preliminary probability distribution curve. The horizontal axis of the curve represents the time window order, and the vertical axis represents the preliminary probability value of the corresponding time window, which is used to intuitively represent the change of hazard risk over time.To avoid oversensitivity to extremely short-duration transient anomalies, the system sets a preset threshold for the initial probability value. This threshold is determined from historical normal samples during the offline phase. Specifically, all time windows confirmed to be in normal operation are selected from historical data. The initial probability value for each window is calculated, and these probability values ​​are sorted from smallest to largest. The probability value corresponding to the element whose sorting position equals 95% of the total number of normal windows is taken as the preset threshold. This ensures that under normal operating conditions, only no more than 5% of the time windows will have an initial probability value exceeding this threshold. When the system is running online, if the initial probability value of a certain time window is greater than or equal to the preset threshold, it is considered that the current initial probability distribution curve may not yet portray the risk near that window smoothly and stably enough. More refined analysis is needed by adjusting the time series integration parameters. At this point, the system triggers parameter adjustment, increasing the analysis window length used to calculate the probability series from the default 10 seconds to 20 seconds. That is, instead of 10 seconds per analysis window, it is now 20 seconds per analysis window, to cover a longer time range of vibration and corrosion data within each window, thereby improving the stability of the probability estimation. Simultaneously, the sliding step size is increased from 1 second. The time interval is shortened to 0.5 seconds. This means that when generating continuous analysis windows, the starting time of each window movement is shifted backward by only 0.5 seconds, increasing the overlap between adjacent windows, improving the temporal resolution of the probability sequence, and better capturing the details of risk changes. For vibration data smoothing, the smoothing length is adjusted from 5 points to 7 points. This means that when calculating vibration features used in the risk probability model, instead of using the local average of 5 consecutive sampling points as the smoothing result, the local average of 7 consecutive sampling points or a similar smoothing operation is used, reducing the sensitivity of vibration features to noise at individual sampling points. After adjusting the above time series integration parameters, the system, based on the new analysis window length, sliding step size, and vibration smoothing length, re-divides the time windows of vibration and corrosion data collected within the same time period, recalculates the features, and re-inputs the updated equipment vibration correlation intensity and environmental corrosion rate into the risk probability model. This results in a new window probability value sequence, which is then connected in chronological order to form an optimized probability curve. This optimized probability curve has smoother and more continuous variation characteristics compared to the initial probability distribution curve, while retaining the main trend of risk fluctuations.To obtain the final window probability value of potential hazards for subsequent decision-making from the optimized probability curve, the system performs probability averaging and trend integration on the optimized probability curve according to a fixed period. The fixed period is determined during the system design phase based on the typical working cycle of the equipment. For example, for continuously operating equipment, the fixed period is set to 60 seconds, and for intermittently operating equipment, the fixed period is set to a certain integer segment of the entire operating cycle. In actual operation, the system divides the time axis into several continuous intervals with the fixed period length. Within each fixed period, the arithmetic mean of all optimized probability values ​​within that period is first calculated as the periodic average probability, used to reflect the overall risk level within that period. Then, trend integration is performed, specifically by traversing the sequence of optimized probability values ​​sequentially from the beginning to the end of the period, judging the probability difference between two adjacent time points, and when the probability difference between the two adjacent time points is reached, the trend integration is performed. When the probability value at an intermediate time point is greater than the probability value at the previous time point, the difference is considered as an increase and accumulated until the entire cycle is traversed, obtaining the sum of all increases within that cycle. This sum represents the degree to which risk accumulates and increases over time. After completing the probability averaging and trend integration, the system combines the cycle average probability and the trend integral value using a fixed weighting method. Specifically, the cycle average probability is multiplied by an average weighting coefficient, and the trend integral value is scaled proportionally to a range close to the cycle average probability and then multiplied by a trend weighting coefficient. The two are then added together to obtain the final window probability value of the hidden danger within that cycle. The average weighting coefficient and the trend weighting coefficient are also determined through repeated trials using historical data during the offline phase, ensuring that the final window probability value reflects both the overall risk level within that cycle and the amplifying effect of the continuously rising trend on risk.

[0031] S5 includes obtaining the probability value of the potential hazard, determining whether it is higher than the preset threshold baseline; if it is higher than the baseline, obtaining matching events from historical damage records, performing a matching degree evaluation by comparing the matching events with the vectors of the current potential hazard features, the matching degree evaluation being based on the comparison of event time and type attributes; calculating event similarity using cosine similarity based on the matching degree evaluation, the cosine similarity being obtained by dividing the inner product of event vectors by the product of moduli; adjusting the probability value by weighted fusion of the similarity index and the original probability value, the weighted fusion being multiplied by the similarity index using preset weights and then added to the original probability value to obtain the probability estimation result.

[0032] In this implementation, the probability value of the potential hazard is first read from the current time window and compared with a pre-set threshold baseline. The threshold baseline is determined using historical data before system deployment. Specifically, a sample of window probability values ​​from a long period of normal operation is selected, sorted from smallest to largest, and the value at the 95th percentile of all samples is taken as the baseline. This ensures that under normal operating conditions, no more than 5% of window probability values ​​will exceed this baseline. Once determined, the threshold baseline remains fixed and can be re-determined by re-analyzing new historical data when the system version is upgraded or the operating environment changes significantly. During online operation, when the system detects that the probability value of the potential hazard in the current window is greater than or equal to the threshold baseline, it determines that the window has a high risk and requires comparison and correction using historical damage records. Subsequently, the system retrieves several records from the historical damage record database that are most similar to the current potential hazard characteristics. The historical damage record database is continuously accumulated during long-term system operation, and each record includes at least the start time of the damage, duration, damage type, involved equipment number, the probability value of the potential hazard within the corresponding time window, and the current... When constructing matching events, the system first marks all records in the record library with a damage level not lower than a preset level as valid candidate records. This preset level is determined by safety management personnel during system debugging based on actual needs. For example, all records that lead to equipment shutdown for maintenance or more serious consequences can be selected as valid records. Then, the system constructs a current hazard feature vector based on the hazard characteristics of the current time window. This vector consists of several numerical components in a fixed order. The components include the hazard probability value of the current window, the root mean square value of vibration within the current window, the maximum peak value of vibration, the count of abnormal peak values, the environmental corrosion rate, and several components reflecting key risk characteristics from the enhanced weighted features generated in the previous steps. The specific components included are fixed during the system design phase and do not change during the entire operation. For each valid record in the historical damage record library, the system also extracts the corresponding values ​​in the same field order as the current hazard feature vector to form the event vector of that record, so that all event vectors and the current hazard feature vector are completely consistent in the number of dimensions and the meaning of each dimension.After obtaining the current hazard feature vector and all candidate event vectors, the system first performs a preliminary screening based on the event's time and type attributes. The time attribute includes the span between the damage occurrence time and the current time, the time period within a day where the damage occurred, and whether it falls on a weekday or non-weekday within a week. The type attribute includes the damage type code and the equipment type code. In implementation, the system uniformly encodes these attributes into numerical form; for example, different damage types are mapped to different integer numbers, different equipment types are mapped to different integer numbers, and the occurrence time is divided into several time periods by hour, with each time period assigned a different code. The system first identifies the current hazard and then calculates the difference between the time attribute code corresponding to the current hazard and the time attribute code of the candidate record. Records with differences within the allowable range are assigned a higher time matching score, while those with differences exceeding the range are assigned a lower time matching score. During type attribute comparison, when a candidate record and the current hazard share the same damage type code and equipment type code, the type matching score is higher; when only one code is the same or both codes are different, the type matching score is lower. By setting weight coefficients for the time matching score and the type matching score, the two are weighted and summed to obtain the matching score for each candidate record. During the system debugging phase, historical data analysis is used to determine, for example, a weight combination of time and type attributes that aligns with safety management experience can be selected through multiple trials. The system sorts all candidate records from highest to lowest matching score, selecting the top-ranked records as matching events. The number of selected records is determined during system design, for example, it can be fixed at 10 or 20 to achieve a balance between computational complexity and comprehensive matching. After obtaining the set of matching events, the system calculates event similarity based on the numerical relationship between these event vectors and the current hazard feature vector. The method uses cosine similarity, specifically: for each matching event, the system multiplies the event vector with the current hidden danger feature vector dimension by dimension and accumulates them to obtain an accumulated value representing the degree of consistency between the two directions. At the same time, the magnitude of the current hidden danger feature vector and the magnitude of the event vector are calculated separately. The magnitude is calculated by summing the squares of all components in the vector and taking the square root. Then, the accumulated value of the consistency between directions is divided by the product of the two magnitudes to obtain a similarity index between 0 and 1. When the magnitudes of the components in each dimension of the two vectors change in a similar trend, the similarity index is close to 1. When the trends of change in each dimension are unrelated, the similarity index is close to 0.For each matching event, a similarity index is calculated as described above. The system can select the event with the highest similarity as the representative event, or it can calculate a comprehensive similarity index by weighting the similarity of multiple matching events according to their matching scores. In one feasible solution, to balance computational efficiency and stability, the system selects the event with the highest similarity and uses its similarity index as the event similarity of the current window. This selection rule is determined during the system design phase and will not be changed during operation. After obtaining the event similarity, the system weights and fuses the event similarity with the original hazard probability value to incorporate historical actual damage conditions for correction while considering the current model output. Specifically, the system pre-sets a fusion weight coefficient, which is selected between 0 and 1 and determined through multiple experiments on historical samples during the system debugging phase. The experimental method involves calculating the consistency between the fused probability value and the historical actual damage occurrence under different fusion weights. The degree of consistency can be obtained by comparing windows with probability values ​​above a certain threshold with windows where actual damage events have occurred. A fusion weight that balances detection rate and false alarm rate is selected as the final value. During online operation, the system multiplies the event similarity by the aforementioned fusion weight coefficient to obtain a correction amount. This correction amount is then added to the original hazard probability value to obtain the probability estimation result. To ensure the probability estimation result is within a reasonable range, if the sum is greater than 1, it is truncated to 1; if the sum is less than 0, it is truncated to 0, thus keeping the final output probability estimation result between 0 and 1. If the current hazard probability value is below the threshold baseline, the system considers the current window risk level to be within the normal or slightly high range. To reduce unnecessary computational overhead, it no longer searches for matching events in the historical damage record database, nor performs event similarity calculation and weighted fusion steps, but directly outputs the original hazard probability value as the probability estimation result.

[0033] S6 includes extracting three data points from equipment damage records: direct damage costs, downtime losses, and component replacement costs. These three data points are then added together to obtain a base loss value. The degree of quantified loss is calculated by multiplying the base loss value by a probability estimation result. The level is determined based on the quantified loss degree and a loss degree grading standard, which is divided into low, medium, and high ranges based on historical loss samples. The average maintenance cost corresponding to the level is extracted from maintenance cost estimation data as the calculation basis. This maintenance cost estimation data includes average labor costs, material costs, and repair time. The average of the quantified loss degree and the average maintenance cost is used to determine the initial loss index base. The equipment model, service life, and load rate are then used to determine if the level exceeds the upper limit threshold. If the initial loss index base exceeds the threshold, the maintenance cost estimate is adjusted upwards by a preset ratio and re-averaged with the quantified loss degree to obtain a revised initial loss index base.

[0034] In this implementation, the system first reads data from the established equipment damage records one by one. Each equipment damage record contains three cost fields: direct damage cost, downtime loss, and component replacement cost. The direct damage cost field records the repair or scrap value of the damaged equipment body and auxiliary components in this failure. The downtime loss field records the amount of production capacity or service capacity loss caused by the downtime of the equipment. The component replacement cost field records the purchase amount of the vulnerable parts, critical parts, and auxiliary parts actually replaced in this repair. When the system is running, for the equipment corresponding to the potential hazard that needs to be assessed, the system retrieves the historical damage events corresponding to that equipment from the equipment damage records based on its equipment number and time information. In each retrieved damage event, the system first reads the three cost values: direct damage cost, downtime loss, and component replacement cost one by one. Take out the three cost values ​​and add them together to get the basic loss value of the damage event. Then, organize the basic loss values ​​of all relevant historical damage events by time or time window to get a set of basic loss value samples. In the current hazard assessment, in order to reflect the possible economic consequences under the probability of the hazard, the system selects a set of historical damage event samples that are closest to the current equipment type and the most recent operating conditions. The basic loss values ​​in the sample are averaged or a basic loss value representing the current equipment's historical loss level is calculated according to preset statistical rules. Then, the basic loss value is multiplied by the probability estimation result obtained in the previous step S5 to get the quantified loss degree. The quantified loss degree is represented by a specific monetary value to indicate the expected economic loss under the current probability. The probability estimation result ranges from 0 to 1. Therefore, the quantified loss degree value is never greater than the basic loss value.After obtaining the quantified loss level, to facilitate subsequent classification and management, the system needs to divide continuous loss values ​​into three levels: low loss, medium loss, and high loss. To achieve this, the system uses historical loss samples to establish a loss level classification standard before deployment. Specifically, all damage events within a sufficiently long time period are extracted from the equipment damage records. The quantified loss level of each event is recalculated using the method described above. All quantified loss level values ​​are sorted from smallest to largest. Based on the sorting result, the entire value range is divided into three consecutive, non-overlapping intervals. The upper limit of the low-loss interval is determined by the quantified loss value located at approximately 33% of the total sample size after sorting. The upper limit of the medium loss interval is determined by the quantified loss value located at approximately 66% of the total sample size after sorting. The lower limit of the low loss interval is zero, the lower limit of the medium loss interval is the upper limit of the low loss interval, the lower limit of the high loss interval is the upper limit of the medium loss interval, and the upper limit of the high loss interval is the largest quantified loss value in the historical samples or a value with a certain safety margin. These three intervals form the classification standard for low, medium, and high loss levels. During system operation, after calculating the quantified loss level corresponding to the current hidden danger, the system compares this value with the upper and lower limits of the above three intervals. If the quantified loss level value is greater than or equal to the low loss interval, the system will determine the level. If the quantified loss value is below the lower limit of the medium loss interval, it is classified as a low loss level. If the quantified loss value is greater than or equal to the lower limit of the medium loss interval but less than the lower limit of the high loss interval, it is classified as a medium loss level. Otherwise, it is classified as a high loss level, and this level is used as an index for subsequent maintenance cost estimation. After determining the level, the system extracts the average maintenance cost corresponding to that level from the maintenance cost estimation data as the basis for calculation. The maintenance cost estimation data is compiled from historical maintenance records and budget data. Each sample contains three fields: actual labor costs, material costs, and repair time in the current maintenance. The labor cost field represents the total labor input in the maintenance. The "Expense Amount" and "Material Cost" fields represent the cost of replacing parts and consumable materials, while the "Repair Time" field represents the time from equipment downtime to normal operation. When constructing the average maintenance cost, the system groups these historical samples according to the loss level. For each group, the average of the sum of labor costs and material costs is calculated as the average direct maintenance cost for that level. The repair time is then multiplied by the standard labor cost to convert it into time-based indirect costs. This indirect cost is added to the average direct maintenance cost to obtain the average maintenance cost corresponding to that level. The average maintenance cost is calculated and stored before the system is officially put into operation. During operation, it is retrieved by looking up a table according to the level.In the current hazard assessment, after the system determines the quantified loss level and corresponding grade, it reads the average maintenance cost corresponding to that grade from the maintenance cost estimation data. The quantified loss level value is added to the average maintenance cost value, and then divided by 2 to obtain a median value, which serves as the basis for the preliminary loss indicator. This base value numerically integrates the expected loss based on probability and the typical maintenance cost level. Subsequently, the system determines whether this preliminary loss indicator base exceeds the upper limit threshold of that grade based on the equipment model, service life, and current load rate. The upper limit threshold is set separately for different equipment models, service lives, and load rate combinations before system deployment. Specifically: First, the historical loss samples are divided into multiple subsets according to the equipment model. Within each subset of equipment models, the service life is further divided into several time intervals, such as 0 to 5 years, 5 to 10 years, and over 10 years. Then, within each time interval, the load rate is divided into different load level intervals, such as load rate less than 50%, between 50% and 80%, and greater than 80%. For each combination interval, the historical samples... The preliminary loss index base values ​​of loss events belonging to this combination are sorted from smallest to largest. The value corresponding to approximately 90% of the sample size of this combination is selected as the upper limit threshold for this combination under this loss level. This indicates that under this equipment model, this service life, and this load rate, the preliminary loss index base value will generally not exceed this value. When the system is running, for the equipment corresponding to the current hidden danger, the system first determines its equipment model and manufacturing date based on the equipment file information, then calculates the service life based on the current date, and calculates the current load rate based on real-time monitoring data. The service life and load rate are respectively assigned to the pre-divided service life interval and load interval to obtain a specific equipment combination interval. Then, based on the combination interval and the current loss level, the corresponding upper limit threshold is read from the threshold table. The currently calculated preliminary loss index base value is compared with the upper limit threshold value. When the preliminary loss index base value is less than or equal to the upper limit threshold value, the system considers that the loss assessment has not exceeded the normal upper limit under this equipment state, keeps the base value unchanged, and outputs it as the preliminary loss index of the hidden danger.When the initial loss indicator base value exceeds the upper threshold, it indicates that the economic impact of the potential hazard may be underestimated given the current equipment model, service life, and load rate combination. To conservatively adjust the maintenance cost, the system adjusts the maintenance cost estimate upwards according to a preset ratio. Specifically, it reads a preset upward adjustment ratio greater than 1 from the parameter configuration based on the current equipment combination range and loss level. This preset upward adjustment ratio is determined through repeated trials using historical samples before system deployment, ensuring that the coverage of the adjusted average maintenance cost in the severe damage event samples meets predetermined requirements. During runtime, the system multiplies the original average maintenance cost by this upward adjustment ratio to obtain the adjusted maintenance cost estimate. Then, it adds the adjusted maintenance cost estimate to the original quantified loss level and divides it by 2 to recalculate and obtain the corrected initial loss indicator base value, which is then used as the final initial loss indicator output.

[0035] S7 includes extracting fixed key loss factors from preliminary loss indicators, including five data points: direct maintenance, downtime loss, material replacement, labor costs, and potential damage; calculating the proportion of each of these five data points in the total loss, with the proportion representing the degree of influence of the factor; determining initial weights according to a preset allocation rule based on the degree of influence of the factors from largest to smallest, assigning the highest weight to the factor with the highest influence, and decreasing the weights of the remaining factors in order, forming an initial weight vector; calculating the correlation coefficients between each pair of the aforementioned five data points using historical loss samples, constructing a correlation matrix using the absolute values ​​of the correlation coefficients, normalizing each row of the correlation matrix, and multiplying it by the initial weight vector to obtain adjusted weights; calculating a risk score based on the adjusted weights using a weighted summation method for the aforementioned five data points, with the risk score being the weighted result of the standardized value of each factor and the corresponding adjusted weight; comparing the risk score with the reference score in historical records, performing a fixed proportion of numerical correction based on the average deviation generated by past samples, to obtain a calibrated score; assigning corresponding interval weights according to the risk interval where the calibrated score is located, and merging the calibrated score and interval weights according to a preset proportion to obtain a comprehensive risk score.

[0036] In this implementation, after completing step S6, a preliminary loss index basis has been calculated for the current potential hazard window. Simultaneously, detailed cost breakdown information constituting this loss index is retained in the equipment damage and maintenance records. To highlight key loss factors that contribute more significantly to the risk, the system first extracts five fixed key loss factors from historical loss data and current assessment data related to the current potential hazard: direct repair costs, downtime loss costs, material replacement costs, labor costs, and potential damage costs. Direct repair costs represent the expenditure on repairing and replacing the equipment body and its auxiliary structures in this or similar historical events. Downtime loss costs represent the production or service revenue generated per unit time due to equipment downtime. The loss is multiplied by the downtime. Material replacement costs represent the amount spent on replacing various spare parts and consumable materials during the maintenance process. Labor costs represent the wages and related labor expenses of personnel involved in the maintenance, testing, and reset operations throughout the process. Potential damage costs represent the indirect losses that may be caused by this hazard under the most unfavorable and reasonable circumstances. The determination of potential damage costs is based on an estimation range provided by the safety management department according to historical major damage samples, fire drill assessment results, and applicable industry standards. Then, the median or larger value that matches the current equipment type and site conditions is selected within this range as the basis for calculation. In the loss assessment of the current hazard, the system reads the above five costs item by item under a unified time period and unified statistical caliber. The five expenses are then summed to obtain a total loss amount. Using this total loss amount as the denominator and each expense as the numerator, the proportion of each expense in the total loss is calculated, resulting in five percentage values. Each percentage value represents the degree of influence of the corresponding loss factor on the overall loss; a higher percentage value indicates a greater contribution of the factor to the total loss. These percentages are considered the factor's degree of influence. After obtaining the degree of influence of the five factors, the system sorts them in descending order and determines their initial weights according to a preset allocation rule. This preset allocation rule is determined before system deployment through historical data fitting and expert experience. For example, in one feasible setting, the initial weight of the factor with the highest degree of influence is limited to a certain percentage of the total weight. Above a certain percentage, the initial weight of the least influential factor is limited to a certain percentage of the total weight, while the middle three factors are distributed proportionally within this range. To ensure that the sum of the weights equals 100 percentage points, the system sets a fixed weight template for each ranking position during the design phase. For example, the factor with the highest influence corresponds to a higher base weight percentage, and the second, third, fourth, and fifth factors correspond to several base weight percentages that gradually decrease. During runtime, the system only matches the specific factors to the corresponding positions based on the current ranking result, without changing the base weight of each ranking position. In this way, an initial weight vector containing five values ​​can be formed, and each value in the vector corresponds one-to-one with a loss factor.To reflect the degree of correlation between different loss factors in weight allocation, the system uses historical loss samples to calculate the correlation coefficients between each of the aforementioned five cost data points. These historical loss samples are collected from all damage events that have occurred since the system began operation. Each sample record contains the specific values ​​of these five costs. During the offline analysis phase, the system constructs a set of sample pairs for each different cost type, treating the values ​​of these two costs in all samples as pairs of observations. By calculating the deviation of each cost in each pair of samples from its respective mean and statistically analyzing the joint changes in these deviations, a value reflecting the degree of mutual or inverse change between the two costs is obtained; this value is the correlation coefficient. After completing all pairs of analysis... After calculating the correlation coefficients of the two combinations, the system takes the absolute values ​​of these correlation coefficients and fills them into a five-row, five-column numerical table according to a fixed order of cost types. Each row and column in the table corresponds to a loss factor. The value at the intersection of a row and a column represents the absolute magnitude of the correlation between the factor corresponding to that row and the factor corresponding to that column. This numerical table constitutes the correlation matrix. To use the correlation matrix to correct the initial weights, the system performs row normalization on each row of the correlation matrix. This involves summing the five values ​​in a row, dividing each value in that row by the sum, and obtaining five new percentage values. The sum of these percentages equals one hundred percentage points, representing the correlation distribution of the factor corresponding to that row with other factors. After normalizing all rows, the system multiplies this normalized association matrix with the initial weight vector. Specifically, for each loss factor, the adjusted weight is equal to the five weight values ​​in the initial weight vector multiplied by the five normalized values ​​in the column of the association matrix containing that factor. These five products are then summed. In other words, each new weight comprehensively considers the association effects of other factors on that factor. This operation yields an adjusted weight vector containing five values, which also corresponds one-to-one with the five loss factors, and the sum remains 100 percentage points. After obtaining the adjusted weights, to uniformly map the magnitude of the loss factors and their weights to the risk metric, the system... The risk score is calculated by weighted summation of the aforementioned five loss data. Before the calculation, in order to avoid unreasonable influence of different cost levels on the weighted result, the system first standardizes the five costs. That is, for each cost type, the average value and the fluctuation range within the normal range are calculated in the historical samples. The current cost value is subtracted from the average value of that type and scaled by the standard fluctuation range of that type, so that each cost after standardization is approximately within a uniform range. After standardization, the system multiplies each standardized cost value by the corresponding adjusted weight, and then adds the five products to obtain the risk score of the current hidden danger. This risk score is a single value used to comprehensively measure the risk contribution of the five loss factors in the current state.To improve the fit between risk scores and actual loss outcomes, the system compares risk scores with reference scores from historical records. These reference scores are derived offline by substituting the five costs at the time into the same calculation process based on actual losses in historical loss samples. These risk scores are then compared with the actual loss levels and economic consequences of these events, selecting the scores that best represent the typical risk levels for each category as references. During offline analysis, the system calculates the deviation between risk scores and actual loss outcomes. The deviation is calculated by subtracting the baseline score (mapped from the actual loss amount) from the risk score. The absolute values ​​of the deviations are then averaged across all historical samples to obtain an average deviation value. The average deviation value is used to determine the correction magnitude required for calibration. During online operation, the system determines a fixed percentage correction coefficient based on the average deviation obtained from historical statistics. For example, the average deviation is multiplied by a scaling factor less than one as the correction magnitude. When the current risk score is significantly higher than the historical average level of the reference score interval, the system subtracts the correction magnitude from the risk score. When the current risk score is significantly lower than the historical average level of the reference score interval, the system adds the correction magnitude to the risk score. The resulting value is the calibrated score. To avoid overcalibration due to repeated directional corrections, the above-mentioned increase / decrease directions and thresholds are fixed after repeated trials with historical samples and are not automatically changed during operation. After obtaining the calibrated integral, the system performs final risk mapping based on the risk interval in which the integral falls. Before system deployment, the risk interval is divided into several continuous segments according to the numerical range of the calibrated integral, for example, corresponding to low, medium, and high risk categories. The upper and lower limits of each segment are determined by the percentile position of the integral range of that type of event in historical data. Simultaneously, a corresponding interval weight is assigned to each risk interval. The interval weight reflects the overall importance of the risk level in the subsequent comprehensive assessment. The interval weight value is between zero and one, set by safety management personnel in conjunction with management strategies and fault tolerance requirements. When the calibrated integral of a hazard falls into a certain risk interval, the system reads the corresponding value from the configuration based on the interval number. The system assigns interval weights; finally, it merges the calibrated integral with the interval weights according to a preset ratio. This preset ratio is determined offline through experiments with multiple weight combinations. Typically, a larger proportion is allocated to the calibrated integral to highlight the quantitative impact of actual loss factors, while a smaller but non-zero proportion is allocated to the interval weights to reflect the management's subjective emphasis on different risk levels. In implementation, the system multiplies the calibrated integral by an integral fusion coefficient, multiplies the interval weights by an interval fusion coefficient, and then adds the two together to obtain a final value, which is the comprehensive risk score. If the sum exceeds a predetermined maximum value, it is truncated to the maximum value; if it is below zero, it is truncated to zero, ensuring that the comprehensive risk score is always output within the effective range.

[0037] S8 includes obtaining a comprehensive risk score and comparing it with a preset warning threshold. When the comprehensive risk score exceeds the warning threshold, it extracts three types of distribution features from the spatiotemporal dynamic distribution data: event time distribution, spatial distribution, and event density, and normalizes these features. Based on the normalized spatial distribution and event density features, it calculates the event proportion of each monitoring area, generates a regional resource allocation ratio according to the event proportion, and determines the inspection frequency and response priority for each time period based on the time distribution features, thereby generating a resource allocation plan. Based on the resource allocation plan, it initiates an optimization iteration loop, calculates the deviation value based on the difference between the actual response time and the planned response time, and if the deviation value is higher than a preset range, it adjusts the regional resource allocation amount by a fixed ratio and updates the response priority, repeating the iteration until the deviation converges. The updated resource allocation plan is used as the initial prevention and control strategy.

[0038] In this implementation, after completing step S7, a comprehensive risk score is obtained. This comprehensive risk score represents the degree of risk of the current hidden danger under a comprehensive assessment of multiple dimensions, including economic loss, changes in equipment status, and comparison with historical trends. The system first compares this comprehensive risk score with a pre-set warning threshold. The warning threshold is determined before system deployment based on the statistical analysis of comprehensive risk scores from long-term historical samples. Specifically, all comprehensive risk scores in the historical samples are sorted from smallest to largest, and the value corresponding to approximately 85% of the total number of samples is selected as the warning threshold. This ensures that under normal operating conditions, only a small number of high-risk samples will exceed this threshold. When calculated during online operation... When the overall risk score is greater than or equal to the warning threshold, it indicates that the current hidden danger has a high risk level, requiring dynamic optimization of fire-fighting resources. At this point, the system extracts three types of distribution features from continuously collected spatiotemporal dynamic distribution data: event time distribution, event spatial distribution, and event density distribution. The event time distribution represents the frequency of fire-related events occurring in different time periods within a fixed analysis period. For example, dividing a day into several time periods and counting the number of events in each time period to form a time frequency vector. The event spatial distribution represents the cumulative number of events in different spatial areas within the monitoring area. For example, dividing the monitoring area into several sub-areas according to fixed rules and counting past analysis... The location information of each event within a period is mapped to a corresponding sub-region, and the number of times the event occurs within each sub-region is counted. The event density distribution represents the concentration of events per unit area or per unit time. In implementation, the system divides the number of events in a specific region within a past analysis period by the area of ​​that region to obtain the spatial density, or divides it by the duration of the corresponding statistical period to obtain the temporal density, thereby reflecting the degree of clustering of events in the spatial and temporal dimensions. The above three types of distribution features have different numerical ranges and variation amplitudes after extraction. To facilitate subsequent calculations, the system performs normalization processing on the three types of distribution features respectively. The normalization method is to calculate the maximum and minimum values ​​of each type of distribution vector, subtract the minimum value from each value in the vector, and divide by... The difference between the maximum and minimum values ​​is used to ensure that all values ​​fall within the range of 0 to 1. The closer the normalized value is to 1, the more concentrated the event is in that time period or spatial area; the closer it is to 0, the lower the concentration. After normalization, the system calculates the event percentage of each monitoring area based on the normalized spatial distribution and event density characteristics. The specific method is as follows: multiply the normalized spatial distribution value of a sub-area by the normalized event density value, and sum the products of all sub-areas as the normalization coefficient. Then, divide the product value of each sub-area by the normalization coefficient, so that the sum of the event percentages of all sub-areas equals 100 percentage points. The resulting event percentage represents the proportion of total resources that the area should receive when allocating resources.The system then generates a resource allocation ratio for each region based on the event proportion. This ratio numerically reflects the proportion of inspection manpower, monitoring intensity, or emergency response resources that the system should allocate to a specific sub-region. In implementation, the system multiplies the total available resources by the event proportion for that region to obtain the resource allocation for that region in the current period. After the regional resource allocation is determined, the system determines the inspection frequency and response priority for each time period based on the normalized time distribution characteristics. The method is as follows: the normalized time distribution values ​​are sorted from largest to smallest, and time periods with higher event concentration are assigned higher inspection frequencies and response priorities. For example, if a day is divided into several consecutive time periods, the period with the largest normalized value is set as the highest priority. The system assigns a high inspection frequency to the next highest normalized value segment, and so on. Time priorities are mapped to specific response gradients, such as high, medium, and low priorities. To ensure continuity and stability, the system design phase specifies the minimum inspection frequency and minimum personnel configuration for each priority level, which remain unchanged during operation. The aforementioned regional resource allocation ratio, time-period inspection frequency, and response priority together constitute the current resource allocation scheme. After obtaining the initial resource allocation scheme, the system enters an optimization iteration loop to improve the consistency between the scheme and actual operation. In each iteration, the system calculates the actual response time based on real-time monitoring records and event response records. The system calculates the time from the event trigger to the arrival of emergency resources at the scene or the completion of initial processing. Simultaneously, it retrieves the planned response time corresponding to the region and time period of this type of event from the resource allocation plan. This planned response time is preset to a fixed value based on response priority during the system design phase; for example, the planned response time for high-priority events is shorter, while the planned response times for medium-priority and low-priority events increase sequentially. During the iteration process, the system calculates the absolute value of the difference between the actual response time and the planned response time, using this difference as the deviation value. When the deviation value is less than or equal to a preset range, it indicates that the current resource allocation plan can meet the response requirements, and the iteration stops. When the deviation value is greater than the preset range, it indicates that the current regional resource allocation or response is inadequate. If the priority allocation is insufficient, parameter correction is required. At this time, the system adjusts the regional resource allocation according to a fixed ratio. Specifically, the ratio of the deviation value to the preset range is used as the adjustment coefficient. This adjustment coefficient is multiplied by the resource adjustment ratio factor. The resource adjustment ratio factor is determined based on historical response samples before system deployment and is used to control the magnitude of resource adjustment in each iteration. The system adds or subtracts this correction amount from the current regional resource allocation to adjust the allocation in a direction that is more suitable for the current event distribution. At the same time, the response priority of all time periods involving the resources in this region is reassessed. Time periods with larger deviation values ​​are prioritized one level higher or the inspection frequency is increased so that more resources are concentrated in the critical time periods.After each resource adjustment and priority update, the system re-simulates the expected response time based on the updated configuration scheme. If the deviation between the expected and actual response times still exceeds a preset range, the system continues with the next iteration until the deviation converges to within the preset range or the maximum allowed number of iterations is reached. Once the convergence condition is met, the system uses the latest resource configuration scheme obtained from the current iteration as the initial prevention and control strategy.

[0039] S9 includes reading three risk thresholds (high, medium, and low) from the initial prevention and control strategy and comparing them with preset allowable ranges. If any deviation exceeds the preset range, it is determined that the threshold deviation exceeds the limit. When the threshold deviation exceeds the limit, data such as risk score, response time, and resource configuration from historical operation records are obtained. Historical reference thresholds for each risk level are calculated, and the difference between the current threshold and the historical reference threshold is analyzed to obtain the verification result. Based on the verification result, a compatibility check is performed on the current strategy. The resource configuration ratio, response priority, and triggering conditions of the current strategy are compared with the historical average strategy to calculate the similarity ratio and determine whether the compatibility threshold is reached to determine the compatibility status of the scheme. When the compatibility status of the scheme does not meet the requirements, the risk threshold is adjusted upward or downward according to the verification result at a preset step size, and the compatibility check is re-executed. The process is iterated until the similarity ratio reaches the compatibility requirement or the set iteration limit is reached to obtain the data required for the final scheme determination. The risk threshold, response priority, and resource configuration parameters are updated according to the data required for the final scheme determination, and the final adaptive prevention and control scheme is generated after integration with the initial prevention and control strategy.

[0040] In this implementation, a preliminary prevention and control strategy is generated after completing step S8. This strategy includes three specific values: a high-risk threshold, a medium-risk threshold, and a low-risk threshold, determined according to risk levels. It also includes corresponding resource allocation ratios, response priorities, and triggering conditions. To avoid over- or under-response due to improper risk threshold settings in the actual operating environment, the system first reads the high-risk, medium-risk, and low-risk thresholds from the preliminary prevention and control strategy. Each of these three thresholds is then compared with a preset allowable range. This preset allowable range is determined before system deployment based on historical data statistics and security management specifications. Specifically, the system statistically analyzes typical threshold ranges for each risk level in long-term historical samples. For example, it uses percentile values ​​from the risk scores in historical data, within the low-risk, medium-risk, and high-risk intervals, as the upper and lower limits of the allowable range, ensuring that these thresholds fall within this range under normal conditions. During online operation, if the system detects any risk threshold exceeding the upper limit or falling below the lower limit of the allowable range, the state is considered an excessive threshold deviation, requiring further adaptive adjustments to ensure the reliability and continuity of the final prevention and control plan. When the threshold deviation exceeds the limit, the system retrieves data related to risk assessment and response mechanisms from historical operation records. This data includes the risk score of the historical window, the corresponding actual response time, and the resource configuration parameters at that time. Historical operation record data is accumulated over a long period after system deployment and stored in units of windows. The system categorizes and organizes this data, dividing all samples into three groups according to risk level: high-risk, medium-risk, and low-risk samples. Within each group, the risk scores are sorted from smallest to largest, and the risk scores at the top 50% of each group are selected as the historical reference threshold for that level. At the same time, the risk values ​​between 40% and 60% are recorded as the stable reference interval for that level. Subsequently, the system performs difference analysis between the current high, medium, and low risk thresholds and the corresponding historical reference thresholds, calculates the absolute value of the difference, and compares it with the offset range of the historical stable reference interval. When the difference exceeds the offset range of the reference interval, the system records the difference as the main offset item in the verification results. The verification results include the offset direction (higher or lower than the historical reference threshold), offset magnitude, and whether the offset persists for each threshold.Based on the above verification results, the system performs a compatibility check on the current strategy. The goal of the compatibility check is to determine whether the current strategy is consistent with the system's historical average strategy, avoiding excessive strategy changes due to local data fluctuations that could affect the overall stability of the system. To this end, the system constructs a historical average strategy from historical data. This strategy includes parameters such as the average resource allocation ratio, average response priority, and average trigger condition threshold under the same risk level in history. These parameters were calculated and stored during the deployment phase. During the compatibility check, the system takes the resource allocation ratio vector, response priority distribution, and trigger condition threshold of the current strategy respectively, and concatenates them into a multi-dimensional vector in a predetermined order. The strategy vector is constructed by concatenating historical average strategies in the same order to form a reference strategy vector. The similarity ratio between the two strategies is then calculated. This similarity ratio is obtained by taking the absolute value of the difference between each dimension of the vector, summing all the differences, and dividing by a preset normalization constant. The system then subtracts this normalized difference from 1 to obtain the final similarity ratio, ensuring it fluctuates between 0 and 1. The closer the similarity ratio is to 1, the more consistent the two strategies are. The system compares this similarity ratio with a compatibility threshold, which is determined through multiple rounds of testing before deployment to ensure the continuity and controllability of strategy updates. When the similarity ratio falls below the compatibility threshold, the current strategy state is deemed unsuitable for compatibility and requires further adjustments. When the compatibility status of the solution does not meet the requirements, the system adjusts the risk threshold upwards or downwards according to the verification results at a preset step size. The preset step size is fixed in the system design phase and is used to control the stability of the adjustment range in each iteration. It is generally changed linearly between one percent and three percent of the current threshold. The system checks the offset direction of each risk threshold. When the offset direction is higher than the historical reference value, the system fine-tunes the threshold downwards by the preset step size. When the offset direction is lower than the historical reference value, it fine-tunes it upwards by the same step size. After adjusting the risk threshold, the system reconstructs the strategy vector using the new threshold and then performs a new similarity ratio calculation with the historical average strategy. If the new similarity ratio is still lower than the compatibility threshold, the next round of threshold adjustment and compatibility check is performed according to the same rules. To avoid infinite iteration, the system sets a maximum limit on the number of iterations. In the deployment phase, it is generally selected as 5 to 10 times based on experiments. When the number of iterations reaches the limit and the compatibility requirements are still not met, the result of the iteration with the highest similarity ratio is used as the basis for determining the final solution. The system obtains the data required for determining the final solution through this cyclical iteration, including the updated risk threshold group, trend direction information, resource adjustment suggestions, and response priority change suggestions.After obtaining the data required to determine the final solution, the system writes the updated risk thresholds, response priorities, and resource configuration parameters into the current strategy template one by one, integrates them with the preliminary prevention and control strategy, overwrites duplicate fields with updated data, and merges fields without conflicts in sequence. This ensures that the final strategy retains the timeliness of the preliminary prevention and control strategy under the current hidden danger conditions, while also incorporating the stability and adaptability ensured by historical data verification and compatibility checks. The strategy after integration is the final adaptive prevention and control plan, which can be directly used for actual on-site scheduling, inspection frequency planning, and response priority guidance.

[0041] Although both "equipment vibration sequence" and "equipment vibration factors" in this solution involve the analysis and utilization of the vibration characteristics of fire-fighting equipment, they differ in their technical meaning, data format, and role in the entire intelligent identification process, and thus possess a clear hierarchical relationship and logical connection. Specific details are as follows: First, a "device vibration sequence" refers to a sequence of raw vibration signal data continuously collected by sensors during equipment operation. Essentially, it is a set of dynamic measurements that change over time, typically recorded by accelerometers or displacement sensors at a fixed sampling frequency. The equipment vibration sequence is presented as one-dimensional time-series data, with each data point corresponding to the vibration intensity or displacement amplitude of the equipment at a specific moment. This sequence fully reflects the vibration variation patterns of the equipment during operation and can intuitively characterize the dynamic features of its operating state. In its implementation, the system achieves real-time monitoring of equipment operational stability and mechanical condition through the acquisition and preprocessing of the vibration sequence.

[0042] Secondly, "equipment vibration factors" are a set of stability characteristic parameters obtained from "equipment vibration sequences" through feature extraction and analysis. They are an abstract representation of statistics or characteristic values ​​obtained after signal analysis of the vibration sequence. Equipment vibration factors are no longer in time series form, but rather key features extracted through mathematical analysis, such as average vibration amplitude, peak vibration value, spectral energy distribution, dominant frequency, frequency domain energy concentration, and vibration stability index. These factors are typically calculated using algorithms such as principal component analysis, Fourier transform, or wavelet decomposition to describe the overall trend and abnormal characteristics of equipment vibration.

[0043] The relationship between the two can be understood as a correspondence between "data source" and "feature representation": the equipment vibration sequence is the raw signal data, a direct observation of the equipment's operational vibration; the equipment vibration factors are the key variables obtained after feature extraction, dimensionality reduction, and modeling based on this sequence, representing a structured representation of the raw signal. In the method, the system first collects equipment vibration sequence data (as in steps S1 and S3), then uses statistical analysis and feature extraction methods to transform the continuous time series into a set of vibration factors for calculation and analysis in subsequent multi-feature fusion, risk assessment, and probabilistic modeling stages.

[0044] In summary, "equipment vibration sequence" serves as the data input and dynamic monitoring source, while "equipment vibration factors" serve as the feature extraction and model calculation source, representing the features extracted from the sequence data. The relationship between the two is one of dependence and transformation; the vibration sequence is the foundation of the vibration factors, while the vibration factors are the higher-level semantic representation of the vibration sequence. This distinction aims to preserve the temporal dynamic characteristics of equipment vibration while enabling model fusion and weight analysis through abstract feature parameters, thereby improving the stability, accuracy, and applicability of the intelligent fire hazard identification method.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of fire hazards based on large-scale multimodal models, characterized in that, include: S1. Obtain multi-source heterogeneous feature data through a sensor network deployed in the fire pump room to obtain an initial feature set; S2. Perform a uniform scale transformation on the initial feature set, calculate heterogeneous bridging parameters based on environmental parameters and equipment operation data, construct a fused feature vector, input the fused feature vector into a preset neural network for adaptive adjustment of feature weights, and output the adjusted weight vector. S3. Calculate the cross-influence value of environmental parameters and equipment vibration based on the adjusted weight vector. If the cross-influence value exceeds the preset threshold, the fused features are weighted by combining the material attenuation coefficient with the vibration peak value, and an enhanced weighted feature representation is generated by comparing historical trends. S4. Construct a risk probability model using enhanced weighted feature representation and environmental corrosion influencing factors as input, and output window-level probability values; connect the continuous window probability values ​​in time order to form a probability distribution curve, and determine whether the time series integration parameter adjustment is triggered based on the overall fluctuation characteristics of the probability distribution curve. If triggered, generate an optimized probability curve, and calculate the final window probability value of the fire pump hazard based on the optimized probability curve. S5. Compare the final window probability value with the preset baseline. If it is higher than the baseline, retrieve similar events from the historical damage record database based on the window corresponding to the probability distribution curve, calculate the event similarity, and perform weighted fusion with the window probability value to obtain the probability estimation result.

2. The intelligent identification method for fire hazards based on large-scale multimodal models according to claim 1, characterized in that: Specifically, S1 is: A multi-source heterogeneous feature data is acquired from a building beam and column monitoring system using a sensor network to obtain an initial feature set. For the initial feature set, the principal component analysis algorithm is used to extract equipment vibration factors and operating time series information to obtain a refined feature subset. The input of the principal component analysis algorithm is the heterogeneous features in the initial feature set, and the output is the dimensionality-reduced refined feature subset. Based on the refined feature subset, the pre-acquired beam and column stress distribution data is integrated from the beam and column monitoring system to determine the structural stability index, wherein the beam and column stress distribution data is acquired in real time by beam and column sensors; If the structural stability index exceeds the preset threshold, supplementary vibration data is obtained from the backup sensors in the pump house network to obtain the initial feature set.

3. The intelligent identification method for fire hazards based on large-scale multimodal models according to claim 1, characterized in that: Specifically, S2 is: Ambient temperature data and equipment vibration sequences were collected, and the two were scaled by subtracting the mean and dividing the variance to obtain a unified scale dataset. The correlation coefficients between temperature and vibration sequences were calculated based on a unified-scale dataset, and a preliminary fusion vector was formed by combining the equipment operation logs. The initial fused vector is input into the convolutional neural network, and the weights are adaptively adjusted through the convolutional layer and the fully connected layer to obtain the adjusted weight vector. The structural environmental stability characteristics are calculated by weighting and fusing the adjusted weight vector with the average stress, maximum stress, and stress fluctuation amplitude collected by the beam-column monitoring system.

4. The intelligent identification method for fire hazards based on large-scale multimodal models according to claim 1, characterized in that: Specifically, S3 is: The mean temperature, standard deviation of temperature, rate of change of temperature, root mean square value of vibration, maximum peak value of vibration, and count of abnormal peak values ​​are calculated by real-time collected ambient temperature sequence and equipment vibration sequence, and the results are used to form an initial vector in a fixed order. Using the initial vector and the fused feature vector obtained in S2 as input, the interaction strength between environmental parameters and equipment vibration is calculated through cross-influence assessment to obtain the cross-influence value; If the cross-influence value exceeds the preset threshold, the weights of the fused feature vector are adjusted, and the components of temperature change rate, vibration peak value and abnormal peak value are amplified by a preset ratio to generate an adjusted weight vector. Based on the adjusted weight vector and the correlation parameters between the equipment material attenuation coefficient and the vibration peak value, a weighted operation is performed on the fused feature vector to obtain a weighted feature representation. By comparing the weighted feature representation with the historical operating trend, and enhancing the weighted feature representation according to the trend deviation ratio, an enhanced weighted feature representation is obtained.

5. The intelligent identification method for fire hazards based on large-scale multimodal models according to claim 1, characterized in that: Specifically, S4 is: Data on equipment vibration and environmental corrosion are collected by sensors, aligned by timestamps, and a time-series dataset is generated. Based on the obtained enhanced weighted feature representation, the root mean square value, the maximum peak value of vibration, and the count of abnormal peak values ​​exceeding three times the standard deviation are calculated from the equipment vibration sequence. The three indicators are multiplied by preset weights and summed to obtain the equipment vibration correlation intensity. The risk probability model is constructed by taking the equipment vibration correlation intensity and the environmental corrosion rate as inputs, outputting a window-level preliminary probability value, and forming a preliminary probability distribution curve in chronological order. If the initial probability value exceeds the preset threshold, the time series integration parameters are adjusted. The adjustment includes: increasing the analysis window length from 10 seconds to 20 seconds, shortening the sliding step from 1 second to 0.5 seconds, and adjusting the vibration smoothing length from 5 points to 7 points. The probability sequence is recalculated based on the adjusted time series integration parameters, and an optimized probability curve is generated. By performing probability averaging and trend integration on the optimized probability curve at a fixed period, the final window probability value of the hidden danger is obtained.

6. The intelligent identification method for fire hazards based on large-scale multimodal models according to claim 1, characterized in that: Specifically, S5 is: Obtain the probability value of potential hazards and determine whether it is higher than the preset threshold baseline; If the value is above the baseline, matching events are obtained from historical damage records. A matching degree assessment is performed by comparing the matching events with the vector of the current hazard characteristics. The matching degree assessment is based on the comparison of event time and type attributes. Based on the matching degree evaluation, cosine similarity is used to calculate the event similarity. The cosine similarity index is obtained by dividing the inner product of the event vectors by the product of their moduli. The probability value is adjusted by weighted fusion of the similarity index and the original probability value. The weighted fusion is achieved by multiplying the similarity index by a preset weight and then adding it to the original probability value to obtain the probability estimation result.

7. The intelligent identification method for fire hazards based on large-model multimodal analysis according to claim 1, characterized in that, It also includes S6, which uses a self-evolving risk perception model to dynamically extract loss features based on probability estimation results and equipment damage data, and achieves multi-dimensional loss assessment through a cost feedback mechanism to generate preliminary loss indicators, specifically including: Extract three data items from the equipment damage record: direct damage cost, downtime loss, and parts replacement cost. Add these three items together to obtain the basic loss value. The degree of loss is quantified by multiplying the base loss value by the probability estimation result; The level is determined based on the quantitative loss level combined with the loss level grading standard, which is based on the low, medium and high ranges of historical loss samples; The average maintenance cost corresponding to this level is extracted from the maintenance cost estimation data as the basis for calculation. The maintenance cost estimation data includes average labor cost, material cost and repair time. The initial loss index is determined by the average of the quantitative loss level and the average maintenance cost, and it is determined whether the equipment exceeds the upper limit threshold of the level based on the equipment model, service life and load rate. If the initial loss index exceeds the threshold, the estimated maintenance cost will be adjusted upward by a preset ratio and then averaged again with the quantified loss level to obtain the revised initial loss index.

8. The intelligent identification method for fire hazards based on large-model multimodal analysis according to claim 7, characterized in that, This also includes S7, which adjusts data weights for preliminary loss indicators to highlight key loss factors. This adjustment is achieved through dynamic weight allocation and correlation matrix construction, while incorporating risk score calculation and historical data calibration to determine the overall risk score. Specifically, this includes: Fixed key loss factors are extracted from the preliminary loss indicators. These factors include five data items: direct maintenance, downtime loss, material replacement, labor costs, and potential damage. The proportion of each of these five data items in the total loss is calculated, and this proportion represents the degree of influence of the factor. The initial weights are determined according to the factors’ influence from largest to smallest according to a preset allocation rule. The factor with the highest influence is given the highest weight, and the remaining factors are given decreasing weights in order, forming an initial weight vector. The correlation coefficients between the five data items are calculated using historical loss samples, and an association matrix is ​​constructed using the absolute values ​​of the correlation coefficients. After normalizing each row of the association matrix, it is multiplied by the initial weight vector to obtain the adjusted weights. Based on the adjusted weights, the risk score is calculated by weighted summation of the aforementioned five data items. The risk score is the weighted result of the standardized value of each factor and its corresponding adjusted weight. The risk score is compared with the reference score in the historical record, and a fixed percentage of numerical correction is performed based on the average deviation generated by past samples to obtain the calibrated score. Based on the risk interval where the calibrated integral is located, the corresponding interval weight is assigned, and the calibrated integral and the interval weight are merged according to a preset ratio to obtain the comprehensive risk score.

9. The intelligent identification method for fire hazards based on large-scale multimodal models according to claim 8, characterized in that, This also includes S8. If the overall risk score exceeds the warning threshold, a resource allocation plan is generated based on the spatiotemporal dynamic distribution. At the same time, an optimization iteration loop is initiated to incorporate a deviation correction mechanism to obtain a preliminary prevention and control strategy, specifically including: A comprehensive risk score is obtained and compared with a preset warning threshold. When the comprehensive risk score exceeds the warning threshold, three types of distribution features—event temporal distribution, spatial distribution, and event density—are extracted from the spatiotemporal dynamic distribution data, and the distribution features are normalized. Based on the normalized spatial distribution and event density characteristics, the event proportion of each monitoring area is calculated, the regional resource allocation ratio is generated according to the event proportion, and the inspection frequency and response priority of each time period are determined in combination with the time distribution characteristics, thereby generating a resource allocation plan. Based on the resource allocation plan, an optimization iteration loop is started. The deviation value is calculated based on the difference between the actual response time and the planned response time. If the deviation value is higher than the preset range, the regional resource allocation is adjusted by a fixed ratio and the response priority is updated. The iteration is repeated until the deviation converges. The updated resource allocation scheme was used as the initial prevention and control strategy.

10. The intelligent identification method for fire hazards based on large-model multimodal analysis according to claim 9, characterized in that, It also includes S9, which involves setting a threshold for the score based on the initial prevention and control strategy analysis. If the deviation exceeds the preset range, an adaptive adjustment cycle is executed to integrate historical data verification and scheme compatibility checks to determine the final adaptive prevention and control scheme. Specifically, this includes: The three risk thresholds of high, medium and low are read from the initial prevention and control strategy and compared with the preset allowable range. If any one of them deviates from the preset range, it is judged that the threshold deviation exceeds the limit. When the threshold deviation exceeds the limit, data such as risk score, response time and resource configuration in historical operation records are obtained, historical reference thresholds for each risk level are calculated, and the current threshold is compared with the historical reference threshold to obtain the verification results. Based on the verification results, a compatibility check is performed on the current strategy. The resource allocation ratio, response priority and triggering conditions of the current strategy are compared with the historical average strategy. The similarity ratio is calculated and it is determined whether the compatibility threshold is reached in order to determine the compatibility status of the solution. When the compatibility status of the solution does not meet the requirements, the risk threshold is adjusted up or down according to the verification results and the compatibility check is re-executed. The process is iterated until the similarity ratio reaches the compatibility requirement or the set iteration limit is reached, and the data required for the final solution determination is obtained. Based on the final plan, the required data update risk thresholds, response priorities, and resource configuration parameters are determined and integrated with the initial prevention and control strategy to generate the final adaptive prevention and control plan.

Images