Fire safety intelligent management system and method based on big data

Abstract

The invention relates to the technical field of fire safety management, and particularly discloses a fire safety intelligent management system and method based on big data. An existing fixed threshold value or single-source trigger fire alarm is prone to high-frequency false alarm, and early symptoms such as smoldering, dark fire and cable overheating are missed and late to report after the threshold is raised, and the system comprises a data acquisition module, a data preprocessing module, a fire behavior evaluation module, an alarm linkage control module and a monitoring display module. The method comprises the steps of data acquisition and preprocessing, candidate event generation, evidence feature extraction, score calculation, secondary discrimination, alarm linkage and write-back, and monitoring display and evidence storage. According to the method, alarm and linkage are stably kept in an effective interval through secondary discrimination of multi-source consistency and scene disturbance combined constraint and adaptive penalty recommendation closed loop, false alarm, missing alarm and late alarm are reduced, and the success rate of smoke exhaust, spraying and power-off linkage is improved.

Description

Technical Field

[0001] This invention relates to the field of fire safety management technology, and in particular to a fire safety intelligent management system and method based on big data. Background Technology

[0002] Building fire safety management typically relies on multiple devices such as smoke detectors, heat detectors, electrical fire monitoring systems, and video surveillance to output alarm signals. On-duty personnel then organize verification, coordinated response, and post-incident tracing based on alarm information. Currently, fixed thresholds or single-source triggering strategies are used: when kitchen fumes, water vapor, dust, aerosol spray, or welding fumes enter the detection range, high-frequency false alarms are easily triggered; raising the detection threshold to reduce false alarms leads to delayed responses to early signs such as smoldering, hidden fires, and localized cable overheating, resulting in missed alarms. The accumulation of false alarms also causes alarm fatigue, consuming patrol resources with numerous invalid events and delaying the issuance of critical response instructions; while missed alarms directly amplify the risks to personnel evacuation and initial fire suppression. Therefore, there is an urgent need for an intelligent fire safety management solution that can minimize missed alarms, reduce the risk of false alarms to below that of conventional building fire alarm systems, and keep alarms and coordinated response within an effective range under complex scene disturbances. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a fire safety intelligent management system and method based on big data. Through a two-level discrimination and adaptive penalty recommendation closed loop with multi-source consistency and scene disturbance joint constraints, alarms and linkages are kept stable within an effective range. Under the premise of minimizing missed alarms, the risk of false alarms is reduced to a level lower than that of conventional building fire alarm equipment, and the success rate and traceability of smoke exhaust, sprinkler, and power failure linkages are improved.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A big data-based intelligent fire safety management system includes a data acquisition module, a data preprocessing module, a fire assessment module, an alarm linkage control module, and a monitoring and display module connected in sequence. The monitoring and display module is connected to the fire assessment module. The data acquisition module collects smoke concentration, temperature, CO concentration, electrical load, electrical arc flash signals, and video images from the building, and records the sampling timestamps and installation locations. The data preprocessing module performs preprocessing on the multi-source data to form a feature sequence of the same observation. After the fire assessment module outputs candidate fire events based on a spatiotemporal fusion model trained on big data, it introduces a joint approach based on multi-source consistency, scene perturbation, and response cost. The constrained two-level discrimination mechanism aggregates candidate events by room topology and calculates the smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture confidence using a sliding window. It dynamically adjusts the confidence threshold lookup table based on ambient humidity, dust concentration, and personnel density. Under the premise of minimizing the cost of missed detections, it reduces the cost of false alarms to below the current minimum false alarm cost. The alarm linkage control module generates alarm levels based on the two-level discrimination results and triggers audible and visual alarms, evacuation commands, smoke exhaust systems, sprinkler valve groups, and power-off controls. At the same time, it writes back the handling results to update the confidence threshold lookup table. The monitoring and display module displays real-time parameter status, alarm trajectories, and equipment status and stores evidence.

[0006] As one aspect of the system of the present invention, the fire assessment module includes a candidate event generation submodule, an evidence feature extraction submodule, a multi-source consistency score calculation submodule, a scene disturbance score calculation submodule, a false alarm / missed alarm estimation submodule, a joint discrimination cost calculation submodule, a penalty recommendation submodule, and a secondary discrimination output submodule. The candidate event generation submodule is connected to the evidence feature extraction module. The evidence feature extraction submodule is connected to both the multi-source consistency score calculation submodule and the scene disturbance score calculation submodule. The multi-source consistency score calculation submodule and the scene disturbance score calculation submodule are both connected to the false alarm / missed alarm estimation submodule. The multi-source consistency score calculation submodule, the scene disturbance score calculation submodule, the false alarm / missed alarm estimation submodule, and the joint discrimination cost calculation submodule are all connected to the joint discrimination cost calculation submodule. The candidate event generation submodule, the evidence feature extraction submodule, the multi-source consistency score calculation submodule, the scene disturbance score calculation submodule, and the penalty recommendation submodule are all connected to the penalty recommendation submodule. The penalty recommendation submodule is connected to the secondary discrimination output submodule. The secondary discrimination output submodule is connected to the false alarm / missed alarm estimation submodule.

[0007] It should be noted that the fire assessment module uses the candidate event generation submodule as the entry point to standardize and organize the candidate fire events generated by spatiotemporal fusion. The event identifiers, time window boundaries, and spatial attributions from the candidate event generation submodule are then passed to the evidence feature extraction submodule. This submodule extracts evidence features such as smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture confidence within the corresponding observation time window and outputs an evidence component sequence. This sequence is then input to the multi-source consistency score calculation submodule and the scene disturbance score calculation submodule. The multi-source consistency score calculation submodule quantifies the consistency relationship between multiple evidence components to obtain a multi-source consistency score. The scene disturbance score calculation submodule quantifies humidity, dust, ventilation intensity, personnel density, and data quality degradation status to obtain a scene disturbance score. The multi-source consistency score and scene disturbance score, along with the evidence component sequence, are further input to the false alarm / missed alarm estimation submodule to estimate the false alarm probability and missed alarm risk. The outputs of the multi-source consistency score calculation submodule, the scene disturbance score calculation submodule, and the false positive / false negative estimation submodule are all input into the joint discrimination cost calculation submodule to form a weighted joint discrimination cost of false positive probability and false negative risk. Simultaneously, the candidate event generation submodule, the evidence feature extraction submodule, the multi-source consistency score calculation submodule, and the scene disturbance score calculation submodule provide event profile elements and score information to the penalty recommendation submodule in parallel. The penalty recommendation submodule determines the accelerated penalty method for the joint discrimination cost within the out-of-bounds threshold range and the mitigated penalty method within the intermediate effective range based on similar sample retrieval and feedback fusion, and outputs the penalty parameters. After receiving the penalty parameters, the joint discrimination cost calculation submodule performs penalty correction on the joint discrimination cost and inputs the corrected joint discrimination cost into the secondary discrimination output submodule to generate candidate event handling results and alarm level suggestions. The secondary discrimination output submodule then sends the handling results back to the false positive / false negative estimation submodule for rolling calibration of false positive probability and false negative risk.

[0008] As one aspect of the system of this invention, in the secondary discrimination output submodule, the secondary discrimination mechanism specifically involves: acquiring the multi-source consistency score and scene disturbance score for each candidate event, and processing the candidate events with a joint discrimination cost. The joint discrimination cost is a weighted sum of the false alarm probability and the missed alarm risk. The joint discrimination cost applies an acceleration penalty to the false alarm probability and the missed alarm risk within a preset excessive threshold range, and a mitigation penalty to the missed alarm probability and the missed alarm risk within a preset intermediate effective range. The acceleration penalty and the mitigation penalty are determined according to the penalty recommendation model of the penalty recommendation submodule. When the scene disturbance and multi-source evidence are inconsistent, the discrimination threshold and processing strategy are stably constrained within the effective range of "controllable false alarms and controllable missed alarms" through adaptive penalties, thereby simultaneously reducing high-frequency false alarms, early missed alarms, and late alarms, and avoiding alarm level jitter.

[0009] As one aspect of the system of the present invention, in the penalty recommendation submodule, the process by which the penalty recommendation model determines the accelerated penalty and the mitigated penalty method includes:

[0010] Step 1, Construct Event Profile: Combine the smoke increment, temperature rise rate, CO concentration increment, arc pulse density, load mutation amplitude, video flame texture confidence and multi-source consistency score, scene disturbance score, and fire compartment topology location encoding of candidate events within the same observation window into an event profile vector. This unifies multi-source evidence, disturbance state, and spatial topology into a comparable structured representation, ensuring that subsequent retrieval and recommendation are carried out in the same semantic coordinate system and reducing cross-scene feature mismatch.

[0011] Step 2, Similar Sample Retrieval: In the event database storing historical verification conclusions and linkage results, K-nearest neighbor retrieval is performed on the event profile vector using cost-aware similarity. Cost-aware similarity is composed of the weighted sum of multi-source consistency difference, disturbance intensity difference, topological proximity, and the cost difference of false positives, false negatives, or late reports. Reference events are found in the sample domain where "operating conditions are similar and cost consequences are similar", so that the penalty strategy transfer not only looks at the surface similarity, but is also sensitive to the cost of false positives and false negatives, thereby improving the targeting and usability of the recommendations.

[0012] Step 3, Remove spurious similar samples: The retrieved similar samples are screened out based on perturbation gating rules and sensor credibility calibration rules. Perturbation gating rules are used to reduce the similarity weight of smoke and video features and increase the similarity weight of temperature rise rate, CO concentration increment and arc light features when humidity, dust concentration or ventilation intensity exceeds the preset threshold. This suppresses the "spurious similarity" interference caused by perturbations such as oil fumes, water vapor, dust and strong ventilation, reduces the bias of the penalty strategy caused by mistakenly taking non-fire perturbation samples as references, and improves the stability and anti-perturbation ability of the recommendation.

[0013] Step 4, Penalty Method Determination: Statistically analyze the proportion of false alarm samples and the proportion of missed and late reports in the retained samples, and compare them with the preset upper bound of false alarm probability and upper bound of missed report risk. When the proportion of false alarm samples exceeds the upper bound of false alarm probability, select a penalty method that applies accelerated penalty to the false alarm probability exceeding the limit and a gradual penalty to the intermediate effective interval. When the proportion of missed or late reports exceeds the upper bound of missed report risk, select a penalty method that applies accelerated penalty to the missed report risk exceeding the limit and a gradual penalty to the intermediate effective interval. This quickly pulls the system operating point back from the "sensitive end" or "sluggish end" to the effective interval, while reducing the penalty slope within the effective interval to avoid frequent threshold oscillations, thus achieving a balance between false alarm suppression and missed report prevention.

[0014] Step 5, Parameter Recommendation: Output the family of penalty functions and penalty parameters corresponding to the penalty method. The penalty parameters include the upper and lower bounds of the effective interval, the penalty slope within the interval, the acceleration coefficient for exceeding the boundary, the upper limit of the penalty saturation value, and the parameter update step size. This transforms the penalty strategy from "qualitative selection" into an executable set of quantitative parameters, enabling the secondary discrimination cost to have controllable sensitivity and upper limit protection under different working conditions, while taking into account both rapid convergence and preventing over-correction caused by outliers.

[0015] Step 6, Online Constraint Update: Based on manual verification and write-back, the penalty parameters are rolled over and an upper limit is set on the parameter change rate. The penalty parameters are continuously calibrated using the closed-loop verification results to adapt to seasonal changes, equipment aging, and scene migration. At the same time, the upper limit on the change rate suppresses parameter jitter and drift, ensuring the robustness and interpretability of long-term operation.

[0016] As one aspect of the system of the present invention, in the penalty recommendation submodule, the multi-source consistency score and scene perturbation score in step 1 are obtained by the multi-source consistency score calculation submodule and the scene perturbation score calculation submodule, respectively, specifically as follows:

[0017] In the multi-source consistency score calculation submodule, within the observation window of the candidate event, the net change, rate of rise, rate of change, pulse count, mean, and peak value are extracted from smoke, temperature, C concentration, arc light, electrical load, and video flame texture confidence. Each parameter is then normalized to the zero-to-one range according to a preset scale parameter, resulting in six evidence components. The multi-source consistency score is obtained by pairwise comparison of these six evidence components. Specifically, a consistency weight is calculated for the difference magnitude between any two evidence components, and then the consistency weights of all evidence pairs are calculated according to... The preset coupling weights are weighted, summarized, and normalized to obtain a multi-source consistency score. By normalizing the multi-source evidence with different dimensions and sampling frequencies to a unified scale and using pairwise consistency weights to characterize the degree of "synchronous pointing to the fire", the coordinated response of smoke temperature, CO, arc light, load, and video features in the time window under the real fire is amplified, while the single-source or few-source anomalies caused by disturbances such as oil fumes, water vapor, and dust are suppressed. This provides a stable, comparable, and non-sensitive consistency criterion for subsequent false alarm and missed alarm estimation and secondary discrimination.

[0018] In the scene disturbance score calculation submodule, relative humidity, dust concentration, ventilation intensity, and personnel density are normalized and weighted to obtain the score. At the same time, a data quality degradation term is introduced as a disturbance correction. The data quality degradation term is jointly determined by three states: the proportion of missing samples within the observation window, whether the sensor drift exceeds the drift threshold, and whether the data is saturated or truncated. The weighted sum of relative humidity, dust concentration, ventilation intensity, personnel density, and data quality degradation term, after being constrained by upper and lower limits, outputs a scene disturbance score of zero to one. This unifies the quantification of external environmental disturbances and data reliability degradation into a controllable disturbance intensity index, enabling the system to identify and label states such as "high humidity / high dust / strong ventilation / high occupancy" and "missing, drifting, saturation," which are prone to false alarms or masking early fires. Based on this, the threshold lookup table, similar sample screening, and cost penalty parameters are adaptively adjusted to avoid oversensitivity under high disturbance conditions leading to alarm storms or overconservatism under low contrast conditions causing missed or delayed alarms, thus improving the robustness and stability of long-term operation.

[0019] As one aspect of the system of this invention, the penalty recommendation submodule has a first feedback channel and a second feedback channel. The first feedback channel is used to receive the verification tags and linkage action validity tags made by the fire scene response segment for candidate events within a preset short time limit, and adds response delay, location matching degree, and data integrity to each verification tag to form immediate credibility. The second feedback channel is used to receive the review tags, false alarm cause categories, and penalty parameter correction suggestions from the strategy labeling end for closed-loop events, and adds evidence chain completeness and consistency assessment to each review tag to form authoritative credibility. The penalty recommendation submodule performs hierarchical weighted fusion of the first and second feedback channels according to the immediate credibility and authoritative credibility, and triggers arbitration rules when there is a tag conflict between the first and second feedback channels to output the final tag and conflict measurement. This includes prioritizing coverage with authoritative credibility, using conflict metrics to suppress parameter update step size, and using conflict metrics to enhance the strength of pseudo-similarity removal. By hierarchically modeling short-term verification feedback and closed-loop review feedback and assigning them immediate credibility and authoritative credibility respectively, the penalty recommendation submodule can quickly absorb the real performance of the latest on-site conditions on false alarms and missed alarms to achieve timely adaptation, and can also use the review conclusions of a more complete evidence chain to robustly calibrate long-term strategies. At the same time, when the two types of feedback conflict, prioritizing coverage with authoritative credibility and constraining the parameter update step size with conflict metrics to enhance the strength of pseudo-similarity removal can suppress threshold drift and oscillation caused by noise labels and occasional misjudgments, reduce system bias caused by error migration, thereby improving the stability, traceability and cross-scenario generalization ability of penalty parameter updates, and further reducing alarm fatigue and idle processing resources.

[0020] As one aspect of the system of the present invention, in the evidence feature extraction submodule, the video flame texture confidence is calculated using a time-frequency texture consistency and two-sided suppression method. Specifically, within the observation window of the candidate event, the target region is extracted from the video sequence with the spatial location of the candidate event as a constraint. For the target region, the brightness fluctuation sequence and texture sequence are extracted frame by frame. The texture sequence is composed of at least one of local gradient magnitude, gradient direction histogram, or local binary mode. Short-time Fourier transform is performed on the brightness fluctuation sequence and texture sequence to obtain the time spectrum. Within the preset flame flicker frequency band, the frequency band energy ratio, spectral concentration, and temporal continuity are calculated to obtain the time-frequency feature group. When mapping the time-frequency feature group to confidence, incrementally enhanced suppression is applied to the two non-flame ranges that meet the preset "fluctuation too low" and "fluctuation too high" criteria. When the frequency band energy ratio or spectral concentration is lower than the lower bound threshold, the confidence is reduced by increasing the amount of excess. When the value exceeds the upper threshold, the confidence level is reduced incrementally by the amount of excess. When the time-frequency feature group falls into the middle effective interval between the upper and lower boundaries, a preset slope suppression or no suppression is applied, and enhancement is performed according to the temporal continuity. By extracting brightness fluctuations and texture time sequences under the spatial constraints of candidate events and performing time-frequency analysis, the unique flicker frequency band energy distribution, spectral concentration morphology, and persistence characteristics of flames are stably characterized. Then, a confidence mapping mechanism that incrementally enhances the suppression at both ends of "too low fluctuation" and "too high fluctuation" is adopted to suppress false flames caused by static reflection, fixed light source, and low image noise, as well as false flames caused by high-frequency strong fluctuations such as welding arc, strong reflection, and strobe lights. This ensures that the confidence level maintains high recognition within a reasonable fluctuation range and decays rapidly in abnormal intervals on both sides, reducing false alarms from single video sources and improving the detection stability of early signs of real flames in low contrast, occlusion, or disturbance scenes, thereby improving the robustness and consistency of multi-source fusion discrimination.

[0021] A big data-based intelligent fire safety management method, applying the aforementioned big data-based intelligent fire safety management system, includes the following methods:

[0022] Step M1, Data Acquisition: Collect smoke concentration, temperature, CO concentration, electrical load, electrical arc flash signal and video images in the building, and write the sampling timestamp and installation location;

[0023] Step M2, data preprocessing: Denoising, drift correction, missing data completion and clock alignment are performed on the multi-source data to form a feature sequence of the same observation window;

[0024] Step M3, Candidate Event Generation: The spatiotemporal fusion model trained based on big data outputs candidate fire events from the feature sequences obtained in step M2 and aggregates them by room topology;

[0025] Step M4, Evidence Feature Extraction: Calculate the smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture confidence within a sliding window;

[0026] Step M5, Score Calculation: Calculate the multi-source consistency score and scene disturbance score for candidate events, and estimate the false alarm probability and false negative risk;

[0027] Step M6, Secondary discrimination: Construct a joint discrimination cost by weighted sum of false alarm probability and false negative risk, and apply accelerated penalty to the joint discrimination cost within a preset excessive threshold range and a slow-release penalty within a preset effective intermediate range according to the output penalty method and penalty parameters, and output the secondary discrimination result, which includes fire alarm warning and non-fire alarm;

[0028] Step M7, Alarm Linkage and Write-back: When the secondary judgment result is a fire alarm warning, an alarm level is generated and the audible and visual alarms, evacuation instructions, smoke exhaust system, sprinkler valve group and power failure control are triggered. The alarm linkage executes the action sequence of the partition and level according to the alarm level, and the linkage handling result is written back to update the confidence threshold lookup table.

[0029] Step M8, Monitoring Display and Evidence Storage: Display real-time parameter status, alarm trajectory and device status, and store evidence for candidate event links.

[0030] As one aspect of the method of the present invention, in step M7, the alarm linkage executes a zonal and hierarchical action sequence according to the alarm level, specifically as follows: when the alarm level is low, an audible and visual alarm is triggered only in the fire compartment to which the candidate event belongs, and a verification prompt is generated in the monitoring display module; when the alarm level is medium, in addition to triggering the audible and visual alarm, an evacuation command is issued to the fire compartment and its adjacent compartments, and the smoke exhaust system of the fire compartment is started; when the alarm level is high, the sprinkler valve group of the fire compartment is opened on the basis of the medium-level linkage, and the fire pump or pressure stabilizing pump is started; when the alarm level is the highest level, the non-fire load in the fire compartment is powered off on the basis of the high-level linkage, and the emergency lighting, fire control equipment and fire power supply circuit are kept powered on; wherein, the escalation conditions between each level include the secondary discrimination result continuously meeting the escalation threshold within the preset holding time or the candidate event propagating to the adjacent compartment under the preset diffusion criteria, and each linkage action is configured with a minimum holding time.

[0031] As one aspect of the method of the present invention, in step M7, the linkage commands for the smoke exhaust system, spray valve group, and power failure control are implemented using closed-loop confirmation and failure handling. Specifically, the start-up of the smoke exhaust system is confirmed by fan operation feedback or fan current feedback; the opening of the spray valve group is confirmed by valve position signal and pipeline pressure switch or flow switch signal; and the power failure control is confirmed by the state quantity of the circuit breaker tripping auxiliary contact or the circuit voltage being zero. When any linkage command fails to meet the confirmation criteria within the preset confirmation timeout period, it is re-sent according to the preset number of retries. If it still fails, it is switched to the backup equipment or backup circuit to perform degraded linkage. At the same time, the issuance timestamp, confirmation timestamp, success or failure flag, and failure reason code of the linkage command are recorded and written back along with the manual verification conclusion.

[0032] Compared with existing technologies, the technical advantages of this invention are:

[0033] This invention employs unified preprocessing and spatiotemporal fusion modeling of multi-source data (smoke, temperature, CO, load, arc light, and video images) under timestamp and spatial location constraints. At the candidate fire event level, it introduces a two-level discrimination mechanism centered on multi-source consistency scores and scene perturbation scores. False alarm probability and false negative risk are uniformly measured using a joint discrimination cost. Accelerated penalties are applied to both ends of the false alarm and false negative risk in the out-of-bounds range, while a slow-release suppression is applied in the effective middle range to maintain discrimination stability. Furthermore, it combines historical verification and linkage results with similar sample retrieval and dual-channel processing. The system integrates human feedback with adaptive recommendation of penalty methods and parameters, and continuously updates the confidence threshold lookup table. Under complex operating conditions, it can simultaneously reduce the proportion of invalid alarms and the probability of early fire missed. Through alarm level-based zoning and hierarchical linkage, interlock verification, closed-loop confirmation, and failure retry and degradation handling, it achieves reliable linkage execution and traceable write-back, improving the effectiveness of building fire alarms, the timeliness of handling, the success rate of linkage, and the traceability of post-event audits. It also reduces alarm fatigue and resource idleness, expands the window for personnel evacuation and initial fire fighting, and improves the overall stability and robustness of building fire safety management. Attached Figure Description

[0034] Figure 1 This is a system block diagram of the present invention;

[0035] Figure 2 This is a schematic diagram showing the connection of the sub-modules of the fire assessment module of the present invention;

[0036] Figure 3 The network architecture diagram for implementing the penalty recommendation model in this invention is shown below;

[0037] Figure 4 Example diagram of selecting target emotional interest regions for high-rise fires according to the present invention;

[0038] Figure 5 This is a flowchart of the method of the present invention. Detailed Implementation

[0039] 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.

[0040] Example 1

[0041] like Figure 1 As shown, this invention proposes a big data-based intelligent fire safety management system, comprising a data acquisition module, a data preprocessing module, a fire assessment module, an alarm linkage control module, and a monitoring and display module connected in sequence. The monitoring and display module is connected to the fire assessment module. The data acquisition module collects smoke concentration, temperature, CO concentration, electrical load, electrical arc light signals, and video images from the building, and records the sampling timestamps and installation locations. The data preprocessing module performs preprocessing on the multi-source data to form a feature sequence of the same observation. After the fire assessment module outputs candidate fire events based on a spatiotemporal fusion model trained on big data, it introduces a system based on multi-source consistency, scene disturbance, and response. The two-level discrimination mechanism with joint cost constraints aggregates candidate events by room topology and calculates the smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture confidence using a sliding window. It dynamically adjusts the confidence threshold lookup table based on ambient humidity, dust concentration, and personnel density. Under the premise of minimizing the cost of missed detections, it reduces the cost of false alarms to below the current minimum false alarm cost. The alarm linkage control module generates alarm levels based on the two-level discrimination results and triggers audible and visual alarms, evacuation commands, smoke exhaust systems, sprinkler valve groups, and power-off control. At the same time, it writes back the handling results to update the confidence threshold lookup table. The monitoring and display module displays real-time parameter status, alarm trajectories, and equipment status and stores evidence.

[0042] like Figure 2As shown, the fire assessment module includes a candidate event generation submodule, an evidence feature extraction submodule, a multi-source consistency score calculation submodule, a scene disturbance score calculation submodule, a false positive / false negative estimation submodule, a joint discrimination cost calculation submodule, a penalty recommendation submodule, and a secondary discrimination output submodule. The candidate event generation submodule is connected to the evidence feature extraction module. The evidence feature extraction submodule is connected to both the multi-source consistency score calculation submodule and the scene disturbance score calculation submodule. The multi-source consistency score calculation submodule and the scene disturbance score calculation submodule are both connected to the false positive / false negative estimation submodule. The multi-source consistency score calculation submodule, the scene disturbance score calculation submodule, and the false positive / false negative estimation submodule are all connected to the joint discrimination cost calculation submodule. The candidate event generation submodule, the evidence feature extraction submodule, the multi-source consistency score calculation submodule, and the scene disturbance score calculation submodule are all connected to the penalty recommendation submodule. The penalty recommendation submodule is connected to the secondary discrimination output submodule. The secondary discrimination output submodule is connected to the false positive / false negative estimation submodule.

[0043] The fire assessment module is specifically implemented using an edge computing host (or rack server) + time synchronization device + local storage array + multi-standard communication interface unit. Specifically, the candidate event generation submodule, evidence feature extraction submodule, multi-source consistency score calculation submodule, scene disturbance score calculation submodule, false positive / false negative estimation submodule, joint discrimination cost calculation submodule, penalty recommendation submodule, and secondary discrimination output submodule are all deployed and run in software within the edge computing host. The edge computing host uses an industrial computer platform equipped with a multi-core CPU, GPU inference accelerator card or NPU inference accelerator module, memory, and solid-state drive. The candidate event generation submodule uses GPU or NPU to execute spatiotemporal fusion model inference, and the CPU schedules the task queue. The evidence feature extraction submodule uses CPU to perform sliding window statistics and feature calculation, and on the video side, it calls the GPU video decoding unit or independent video decoding card to complete target area decoding and time-frequency analysis. Analysis: The multi-source consistency score calculation submodule, scene disturbance score calculation submodule, false alarm / missed alarm estimation submodule, and joint discrimination cost calculation submodule are performed by the CPU using matrix operations and threshold lookup table access; the penalty recommendation submodule is performed by the CPU accessing the local event database and performing similarity retrieval and parameter recommendation, with the local event database stored in a local SSD array or RAID disk array; the secondary discrimination output submodule sends the handling results to the alarm linkage control module through a communication interface, and the communication interface unit is equipped with at least an Ethernet interface, an RS485 interface, and a CAN interface to connect to the fire control host, data gateway, or industrial switch; the time synchronization device uses a PTP or NTP time synchronization module to provide a unified time standard for various types of data; and the edge computing host is powered by a dual-power redundant power supply module and UPS, and the hardware watchdog automatically resets critical process anomalies, thereby realizing real-time inference, closed-loop write-back, and stable operation of fire assessment on the building site.

[0044] like Figure 3 As shown, the technical solution proposed in this invention is specifically accomplished by a hierarchical intelligent network architecture composed of a dual-tower similarity retrieval network, a disturbance gating calibration network, and a penalty method and parameter generation network. The dual-tower similarity retrieval network first encodes candidate events and historical events with a unified structure, mapping event profiles containing information such as smoke, temperature rise, CO, electrical arcing, load mutations, video textures, multi-source consistency, and scene disturbances into measurable embedding vectors. Based on this, K-nearest neighbor retrieval and cost-aware reordering are performed, selecting the set of samples from the historical event database that are most relevant to the current event in terms of semantic morphology and consequences. However, considering the complex disturbance factors commonly present in building fire protection scenarios, such as oil fumes, water vapor, dust, strong ventilation, and sensor degradation, this invention further introduces a disturbance gating calibration network to perform secondary processing on the aforementioned similar sample set. This network takes scene perturbation parameters and data quality status as input, generating feature-level dynamic weights and sample-level credibility weights. It reduces the weight of susceptible evidence components and enhances stable and reliable evidence components under the current conditions. It also suppresses the influence of pseudo-similar samples through soft or hard removal mechanisms, thus outputting a calibrated set of credible samples and their corresponding cost statistics, laying a reliable foundation for subsequent penalty decisions. Based on this, the penalty method and parameter generation network, as the core control unit of the entire architecture, performs cost situation aggregation encoding on the calibrated sample set. It introduces the proportion of false positives, the proportion of missed or delayed reports, cost differences, and their degree of exceeding limits, and combines this with the multi-source consistency score and scene perturbation score of the current event to form a cost situation embedding vector reflecting the overall risk status. The network uses an expert gating mechanism to adaptively select or weightedly fuse false positive suppression experts and missed positive suppression experts, generating penalty parameter candidates for different risk aspects. Under parameter constraints and rate of change limits, it outputs the final executable penalty method and parameter set, ensuring that the joint discrimination cost remains smooth and stable within the intermediate effective range and converges quickly when the risk exceeds the limit. Through the coordinated operation of the three-layer network, the present invention can simultaneously suppress high-frequency false alarms and early missed alarms and delayed alarms in complex and ever-changing building fire protection scenarios, avoid frequent fluctuations in alarm thresholds, and improve the reliability, interpretability, and long-term operational stability of alarms and linkage responses.

[0045] It should be noted that the penalty recommendation submodule has a first feedback channel and a second feedback channel. The penalty recommendation submodule performs hierarchical weighted fusion of the first feedback channel and the second feedback channel based on the immediate credibility and the authority credibility, and triggers arbitration rules to output the final label and conflict metric when there is a label conflict between the first feedback channel and the second feedback channel. In the penalty recommendation model, the "first feedback channel / second feedback channel" is implemented through an integrated network structure of dual-branch feedback encoding, credibility estimation, conflict arbitration, and hierarchical fusion. The first feedback channel encoding branch quantifies / embedding inputs such as verification labels, linkage action validity markers, handling delays, location matching degrees, and data integrity. This data is then passed through a lightweight multilayer perceptron (MLP) or gated recurrent unit (GRU) to obtain the first channel feedback embedding vector. The handling delay can be gated with time decay, so that feedback closer to the event occurrence contributes more to the embedding. The second feedback channel encoding branch performs category embedding and numerical concatenation on inputs such as verification labels, false alarm cause categories, penalty parameter correction suggestions, evidence chain integrity, and consistency assessment. This data is then passed through a deeper MLP or self-attention encoding layer to obtain the second channel feedback embedding vector. The "cause category / suggestion type" is encoded using an embedding layer to preserve semantic differences. An instant credibility estimation head is concatenated at the end of the first channel encoding branch to fuse "processing delay, location matching degree, data integrity" and the first feedback channel embedding vector, outputting an instant credibility scalar (Sigmoid output) ranging from 0 to 1. An authoritative credibility estimation head is concatenated at the end of the second channel encoding branch to fuse "evidence chain integrity, consistency assessment" and the second channel embedding vector, outputting an authoritative credibility scalar (Sigmoid output) ranging from 0 to 1. The "label distribution / discrimination logit" predicted by each of the two channels (which can be output by a small classification head connected to each channel) is input to the conflict detection head, and the degree of inconsistency is calculated using absolute difference, cosine distance, or KL divergence, and the conflict metric (0 to 1) is output through MLP. The arbitration logic is implemented using a gated arbitration layer: the gated arbitration layer takes "authoritative credibility, instant credibility, and conflict metric" as input and outputs the fused weights of the two channels; when the conflict metric is higher than a preset value, the gated arbitration layer automatically increases the weight of the second channel and simultaneously outputs an "update suppression coefficient" to limit the subsequent parameter update step size.The hierarchical weighted fusion layer first aggregates multiple feedbacks of the same event within a channel using attention pooling to obtain "first feedback channel aggregated embedding / second feedback channel aggregated embedding". Then, it performs gated weighted fusion between channels: generating fusion weights based on "immediate credibility, authoritative credibility, and conflict metric", and performing a weighted summation on the aggregated embeddings of the two channels to obtain the final label embedding and the final label / confidence output. At the same time, it outputs the conflict metric and update suppression coefficient, which are used to implement constraints such as "suppressing update step size and enhancing pseudo-similarity removal strength" when the penalty parameters are updated online.

[0046] It should be noted that in the evidence feature extraction submodule, the confidence score of video flame texture is calculated using a combination of time-frequency texture consistency and two-sided suppression. For example... Figure 4 As shown, a specific example of image information calculation is presented using a publicly available building fire video sequence. Approximately 6 seconds of continuous frames are extracted from a high-rise building fire video to describe the calculation of the video flame texture confidence. The observation window is set to 6 seconds, comprising 180 frames, with a frame rate of 29.97 fps. The target region of interest (ROI) is constrained by the spatial location of the candidate event, and a 160*160 region is extracted within the frame (ROI coordinates range X from 136 to 296, y from 192 to 352). This region covers a clearly visible flame band on the facade (see...). Figure 4(Red box in the middle) Within the evidence feature extraction submodule of this invention, the video sequence of the target region of interest is processed frame by frame: First, a brightness fluctuation sequence and a texture sequence are constructed. The brightness fluctuation sequence is obtained by extracting the average brightness of each frame of the target region of interest to obtain a one-dimensional frame sequence that changes over time (reflecting the brightness fluctuations caused by flame flickering). The texture sequence is obtained by extracting the texture amount of each frame of the target region of interest and combining them into a one-dimensional sequence. For example, the gradient magnitude mean (which can also be replaced by gradient direction histogram statistics or local binary mode statistics) is used to characterize the texture dynamics caused by flame edge jitter, fine textures, and the mixing of smoke and flame. Each sequence is subjected to a short-time Fourier transform to obtain a spectrum that slides over time. Within a preset flame flicker frequency band (2-8 Hz), three time-frequency features are calculated: frequency band energy proportion, spectral concentration, and temporal continuity. Lower and upper thresholds are set for the frequency band energy proportion and spectral concentration, respectively. If the value is below the lower threshold, it is judged as "too low fluctuation" (more like a static bright object / reflection), and the confidence level is reduced by increasing the threshold. If the value is above the upper threshold, it is judged as "too high fluctuation" (more like a strobe light / strong noise / compression artifact), and the confidence level is reduced by increasing the threshold. If the value falls within the middle effective range, the high confidence level is maintained (or only a slight suppression with a preset slope is applied). Following the above operations, the frequency band energy ratio at the window center time of 3.737s was obtained as 0.249, the spectral concentration as 0.110, the temporal continuity as 0.695, and the confidence score of the flame texture after bilateral suppression and continuity enhancement as 0.816. This indicates that the brightness fluctuation and texture dynamics of the region of interest of this target are neither too low (excluding static bright spots / reflections) nor too high (suppressing strobe / strong noise) within the typical flame flicker frequency band, and have good temporal continuity, thus being assigned a high confidence score.

[0047] Finally, the confidence level is enhanced by time-domain continuity (the better the continuity, the higher the confidence level).

[0048] Example 2

[0049] Unlike Example 1, this example is a specific illustration of a big data-based intelligent fire safety management method proposed in this invention.

[0050] like Figure 5 As shown, the present invention proposes a big data-based intelligent fire safety management method, applying the big data-based intelligent fire safety management system described in Example 1. The method includes:

[0051] Step M1, Data Acquisition: Collect smoke concentration, temperature, CO concentration, electrical load, electrical arc flash signal and video images in the building, and write the sampling timestamp and installation location;

[0052] Step M2, data preprocessing: Denoising, drift correction, missing data completion and clock alignment are performed on the multi-source data to form a feature sequence of the same observation window;

[0053] Step M3, Candidate Event Generation: The spatiotemporal fusion model trained based on big data outputs candidate fire events from the feature sequences obtained in step M2 and aggregates them by room topology;

[0054] Step M4, Evidence Feature Extraction: Calculate the smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture confidence within a sliding window;

[0055] Step M5, Score Calculation: Calculate the multi-source consistency score and scene disturbance score for candidate events, and estimate the false alarm probability and false negative risk;

[0056] Step M6, Secondary discrimination: Construct a joint discrimination cost by weighted sum of false alarm probability and false negative risk, and apply accelerated penalty to the joint discrimination cost within a preset excessive threshold range and a slow-release penalty within a preset intermediate effective range according to the output penalty method and penalty parameters, and output the secondary discrimination result;

[0057] Step M7, Alarm Linkage and Write-back: Based on the secondary discrimination results, generate alarm levels and trigger audible and visual alarms, evacuation commands, smoke exhaust systems, sprinkler valve groups and power failure controls, and write back the linkage response results to update the confidence threshold lookup table;

[0058] Step M8, Monitoring Display and Evidence Storage: Display real-time parameter status, alarm trajectory and device status, and store evidence for candidate event links.

[0059] It should be noted that in step M7, the alarm linkage executes a tiered action sequence according to the alarm level. The escalation conditions between each level include the secondary discrimination result continuously meeting the escalation threshold within a preset holding time or the candidate event propagating to adjacent partitions under a preset diffusion criterion, and each linkage action is configured with a minimum holding time. Preferably, the preset diffusion criterion can include one or more of the following: multi-partition consistency triggering criterion, temporal sequence correlation criterion, ventilation and airflow coupling criterion, topological adjacency and structural correlation judgment, video space extension criterion, and joint risk accumulation criterion. The multi-zone consistency trigger criterion is that within a preset time window, at least two of the smoke concentration, temperature rise rate, or CO concentration increment independently collected in adjacent fire compartments simultaneously exceed the corresponding baseline threshold within the compartment, and their changing trends are consistent with the original compartment's candidate event. In this case, the event is determined to have spread to adjacent compartments. The temporal correlation criterion is that when the abnormal occurrence time of relevant monitoring indicators in adjacent fire compartments lags behind the abnormal time in the original compartment by no more than a preset propagation delay threshold, and the abnormal amplitude shows an increasing trend, it is determined to meet the temporal characteristics of fire propagation along the spatial direction. The ventilation and airflow coupling criterion is that if the compartment to which the candidate event belongs and the adjacent compartment are within the same ventilation or smoke exhaust path, and the ventilation direction, wind speed, or pressure difference conditions meet the criteria for smoke potentially flowing along that path, and simultaneously, the adjacent compartments monitor a synchronous rise in smoke or CO concentration. If the candidate event is directly adjacent to the adjacent partition in the fire compartment topology, or connected through structural units such as connecting corridors, stairwells, and pipe shafts, and abnormal temperature or smoke is detected continuously within the structural unit, it is determined to be structurally associated diffusion. If the candidate event is directly adjacent to the adjacent partition in the fire compartment topology, or connected through structural units such as connecting corridors, stairwells, and pipe shafts, and abnormal temperature or smoke is detected continuously within the structural unit, it is determined to be visual diffusion. If the high confidence area of ​​flame texture or smoke texture moves from the edge of the original partition's field of view to the corresponding field of view area of ​​the adjacent partition in consecutive frames, and this movement continues to exceed a preset frame number threshold, it is determined to be visual diffusion. If the joint risk accumulation criterion is that when the secondary discrimination result of the candidate event in the original partition is continuously in the high-risk range, and the joint discrimination cost in the adjacent partition increases monotonically within a preset accumulation time and exceeds the diffusion trigger threshold, it is determined that the overall fire risk has spread.

[0060] Specifically, the spatial affiliation (room / fire compartment number) and topological location code of the candidate events in the secondary discrimination results are received, and the following are found in the fire compartment topology table: the belonging compartment Z0, the adjacent compartment set Z1, and the related compartment set Z2 of the same smoke exhaust compartment / same smoke-proof stairwell; at the same time, it is determined whether to extend the linkage domain from Z0 to Z1 or Z2 according to the preset "diffusion criteria" (such as continuous smoke temperature rise rate, continuous CO increment, and repeated occurrence of the event in adjacent compartments).

[0061] Map the alarm levels output by the secondary discrimination to preset linkage action templates (the action list table is indexed by "level × region"):

[0062] Low level: Only Z0 local alarm + prompt for verification;

[0063] Medium level: Z0+Z1 evacuation prompts and Z0 smoke extraction is activated;

[0064] High level: Start sprinkler / fire pumps based on the medium level;

[0065] Highest level: Based on the higher level, implement Z0 non-fire load power cut-off while retaining fire power supply / emergency lighting.

[0066] The action template also includes fixed action priority, minimum hold time, number of retries, and degradation alternative strategies.

[0067] The action sequence is arranged in the order of "first personnel prompting, then smoke control, then fire extinguishing and water supply, and finally power cut-off and isolation" to form an executable queue. Each action includes: target device ID, target status (on / off / mode), effective zone, execution timeout, confirmation conditions and failure handling.

[0068] Perform interlock checks before distribution, including at least:

[0069] Power outage interlock: Confirm that the fire control equipment, fire pumps, and emergency lighting circuits are on the "uninterrupted power supply whitelist";

[0070] Smoke exhaust interlock: Confirm that the corresponding fire damper / smoke exhaust damper is in a state that allows opening, to avoid conflict with the air supply / smoke exhaust logic;

[0071] Spray interlock: Confirm that the valve group / pump group is in an available state and meets the pressure / water source conditions;

[0072] If the interlock is not satisfied, the "downgrade replacement" will be executed according to the template (such as changing to expand the evacuation prompt, changing to start the backup fan / backup circuit, etc.).

[0073] Commands are sent to the corresponding devices via the fire control host interface or building control / equipment bus interface (Ethernet / RS485 / CAN, etc.):

[0074] Audible and visual alarm: Press Z0 or Z0+Z1 to select the circuit trigger;

[0075] Evacuation instructions: Issued via zoned broadcast or voice loop;

[0076] Smoke exhaust system: start the fan, open the valve, and close the relevant fire damper according to the zone;

[0077] Sprinkler valve assembly / pump assembly: Open valves and start pumps according to zones;

[0078] Power outage control: Press Z0 to select the non-fire load circuit to trip.

[0079] At the same time, the timestamp of each instruction being issued is recorded.

[0080] Perform closed-loop verification according to the verification conditions of the action template: for example, fan operation / current feedback, valve position in place, pressure / flow switch, circuit breaker tripping contact or circuit voltage is zero;

[0081] If a timeout is confirmed, a retry will be performed according to the template (limited number of times). If it still fails, a backup device / backup circuit will be switched and a degraded reason code will be marked.

[0082] Set minimum holding time and release conditions to avoid frequent start-stop; if the secondary judgment continues to exceed the limit or the diffusion area expands within the holding window, an upgrade is triggered and the action list is recalculated; if the release conditions are met (risk decreases in multiple consecutive windows and verification confirms no fire), the system is restored in the safe order of "first release power off / sprinkler, then release smoke exhaust, and finally release sound and light," and the restoration confirmation is retained.

[0083] The "action list, execution status, confirmation timestamp, failure reason code, degradation path, and manual verification conclusion" are written back to update the confidence threshold lookup table and form event link evidence to support subsequent review and parameter rolling calibration.

[0084] It should be noted that in step M7, the linkage commands for the smoke exhaust system, spray valve group, and power failure control are implemented using closed-loop confirmation and failure handling.

[0085] Specifically, an action unit (target device ID, target status, confirmation conditions, confirmation timeout Tack, retry count N, and degradation strategy) is generated for each linkage action; the feedback points (DI / AI / communication registers) of the corresponding device are read through the fire control host / building control gateway; feedback is polled / subscribed within the Tack; if the confirmation conditions are met, success is determined and the confirmation timestamp is recorded; if no confirmation is made after timeout, the device is retried N times; if it still fails, the device / circuit switching or degradation linkage is executed; a failure reason code is generated at the same time; the issuance timestamp, confirmation timestamp, success / failure flag, failure reason code, degradation path, and final status are recorded and written back for subsequent threshold table updates and auditing.

[0086] More specifically, the smoke extraction system is treated as follows:

[0087] The exhaust fan's "operation feedback" DI (RUN) or the inverter's "operation status word"; the fan current / power AI (I / P > minimum operating threshold), or the duct differential pressure / speed switch, or the corresponding exhaust valve's "position feedback" DI (OPEN_FB); RUN = 1 and (I > threshold or OPEN_FB = 1) is considered successful; optionally, "fan failure" DI can be added as a negative condition. Failure handling strategies include: timeout tack, such as 10–30s (configurable according to the project); retry, first resend the start command 1–2 times, if the valve is not in place, first issue "valve open" and then issue "fan start"; standby switching, switch the standby smoke exhaust fan in the same zone (master-standby switching), or switch the standby frequency converter channel, if a branch valve fails, start the upstream main valve / bypass valve and expand the smoke exhaust zone range (degraded linkage); safety interlock, if "fire damper closed / interlock conflict" is detected, first execute the interlock release sequence (such as valve open, confirmation, fan start).

[0088] The handling of sprinkler valve groups is as follows: The deluge valve / electric valve is "open to position" (DI) (OPEN_FB), or the valve position analog value reaches the target opening degree; the status of the pipeline pressure switch, flow switch, and end-point test device, or the pressure AI reaches the set value range; the fire pump / pressure stabilizing pump operation feedback RUN=1, and the pump outlet pressure rises; valve OPEN_FB=1 and (pressure / flow satisfied) and (pump RUN=1 when needed) are considered successful. Failure handling strategies include: timeout T_ack, e.g., 10–60s (valve action and hydraulic establishment take time); retry, including re-issuing the valve opening command; if the pressure does not rise, start the fire pump / switch the pressure stabilizing pump to the fire pump; if the pump fails to start, switch to the standby pump and alarm "pump failure"; degraded linkage: if the valve group fails to reach position, expand to adjacent valve groups on the same riser or zone linkage (according to the template), and forcibly upgrade the evacuation level and smoke extraction strategy; if hydraulic anomalies occur, maintain the alarm and prompt manual intervention (e.g., prompt for manual valve opening / local startup of pump room equipment).

[0089] More specifically, power outage control measures include:

[0090] The primary confirmation is the circuit breaker tripping auxiliary contact DI (TRIP_FB / OPEN_FB) or the smart circuit breaker status word; the secondary confirmation is that the circuit voltage AI≈0 or the power of the end-user load drops below the threshold; the whitelist interlock is that the fire power supply, fire control equipment, and emergency lighting circuits are on the "uninterrupted power supply whitelist", and the tripping command of the whitelist circuits is prohibited or automatically reverted; the comprehensive judgment is that the non-fire circuit OPEN_FB=1 and the voltage≈0 is considered successful, and the whitelist circuits maintain power supply as a constraint condition. Failure handling strategies include: timeout tack, e.g., 5–15 seconds; retry, resending the trip command 1–2 times, and executing the "closing and tripping self-test sequence" if necessary (limited and optional); backup switching / isolation, if a branch circuit breaker fails to operate, switch to the upper-level feeder trip (only for non-fire feeders) or switch the backup isolating switch, while marking the circuit as a "failed to operate circuit", raising the alarm level and prompting manual on-site isolation; safety assurance, when power failure occurs, other linkages are not blocked (smoke exhaust, evacuation, and sprinkler systems continue to operate), and the "power failure not achieved" reason code is forcibly recorded for review.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0092] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fire safety intelligent management system based on big data, characterized in that, The system comprises a data acquisition module, a data preprocessing module, a fire assessment module, an alarm linkage control module, and a monitoring and display module, all connected in sequence. The monitoring and display module is connected to the fire assessment module. The data acquisition module collects smoke concentration, temperature, CO concentration, electrical load, electrical arc flash signals, and video images from the building, and records the sampling timestamps and installation locations. The data preprocessing module preprocesses the multi-source data to form a feature sequence of the same observation. After the fire assessment module outputs candidate fire events based on a spatiotemporal fusion model trained on big data, it introduces a two-level discrimination mechanism based on joint constraints of multi-source consistency, scene perturbation, and response cost. Candidate events are aggregated by room topology, and the confidence levels of smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture are calculated using a sliding window. The confidence threshold lookup table is dynamically adjusted based on ambient humidity, dust concentration, and personnel density. Under the premise of minimizing the cost of missed alarms, the cost of false alarms is reduced to below the current minimum false alarm cost. The alarm linkage control module generates alarm levels based on the secondary discrimination results and triggers audible and visual alarms, evacuation commands, smoke exhaust systems, sprinkler valve groups, and power-off control. At the same time, the handling results are written back to update the confidence threshold lookup table. The monitoring and display module displays real-time parameter status, alarm trajectories, and equipment status and stores evidence.

2. The fire safety intelligent management system based on big data according to claim 1, characterized in that, The fire assessment module includes a candidate event generation submodule, an evidence feature extraction submodule, a multi-source consistency score calculation submodule, a scene disturbance score calculation submodule, a false positive / false negative estimation submodule, a joint discrimination cost calculation submodule, a penalty recommendation submodule, and a secondary discrimination output submodule. The candidate event generation submodule is connected to the evidence feature extraction module. The evidence feature extraction submodule is connected to both the multi-source consistency score calculation submodule and the scene disturbance score calculation submodule. The multi-source consistency score calculation submodule and the scene disturbance score calculation submodule are both connected to the false positive / false negative estimation submodule. The multi-source consistency score calculation submodule, the scene disturbance score calculation submodule, and the false positive / false negative estimation submodule are all connected to the joint discrimination cost calculation submodule. The candidate event generation submodule, the evidence feature extraction submodule, the multi-source consistency score calculation submodule, and the scene disturbance score calculation submodule are all connected to the penalty recommendation submodule. The penalty recommendation submodule is connected to the secondary discrimination output submodule. The secondary discrimination output submodule is connected to the false positive / false negative estimation submodule.

3. The fire safety intelligent management system based on big data according to claim 2, characterized in that, In the secondary discrimination output submodule, the secondary discrimination mechanism is as follows: obtain the multi-source consistency score and scene perturbation score for each candidate event, and process the candidate events with a joint discrimination cost. The joint discrimination cost is the weighted sum of the false positive probability and the false negative risk. The joint discrimination cost applies accelerated penalties to the false positive probability and the false negative risk within a preset excessive threshold range, and applies mitigation penalties within a preset intermediate effective range. The methods of accelerated penalties and mitigation penalties are determined according to the penalty recommendation model of the penalty recommendation submodule.

4. The fire safety intelligent management system based on big data according to claim 3, characterized in that, In the penalty recommendation submodule, the process by which the penalty recommendation model determines the methods of accelerating and mitigating penalties includes: Step 1, Construct Event Profile: Combine the smoke increment, temperature rise rate, CO concentration increment, arc pulse density, load mutation amplitude, video flame texture confidence and multi-source consistency score, scene disturbance score, and fire compartment topology location code of the candidate event within the same observation window into an event profile vector. Step 2, Similar Sample Retrieval: In the event database that stores historical verification conclusions and linkage results, K-nearest neighbor retrieval is performed on the event profile vector with cost-aware similarity. The cost-aware similarity is composed of the weighted sum of multi-source consistency difference, disturbance intensity difference, topological proximity, and false alarm cost difference, false negative or late reporting cost difference. Step 3, Remove spurious similar samples: The retrieved similar samples are screened out based on perturbation gating rules and sensor credibility calibration rules. Perturbation gating rules are used to reduce the similarity weight of smoke and video features and increase the similarity weight of temperature rise rate, CO concentration increment and arc light features when humidity, dust concentration or ventilation intensity exceeds a preset threshold. Step 4, Penalty Method Determination: Statistically analyze the proportion of false alarm samples and the proportion of missed and late reports in the retained samples, and compare them with the preset upper bound of false alarm probability and upper bound of missed report risk. When the proportion of false alarm samples exceeds the upper bound of false alarm probability, select a penalty method that applies accelerated penalty to the amount of false alarm probability exceeding the limit and a mitigating penalty to the intermediate effective interval. When the proportion of missed or late reports exceeds the upper bound of missed report risk, select a penalty method that applies accelerated penalty to the amount of missed report risk exceeding the limit and a mitigating penalty to the intermediate effective interval. Step 5, Parameter Recommendation: Output the penalty function family and penalty parameters corresponding to the penalty method. The penalty parameters include the upper and lower bounds of the intermediate effective interval, the penalty slope within the interval, the acceleration coefficient for exceeding the boundary, the upper limit of the penalty saturation value, and the parameter update step size. Step 6, Online Constraint Update: Based on manual verification and write-back, perform rolling updates on the penalty parameters and set an upper limit on the parameter change rate.

5. The fire safety intelligent management system based on big data according to claim 4, characterized in that, In the penalty recommendation submodule, the multi-source consistency score and scene perturbation score in step 1 are obtained through the multi-source consistency score calculation submodule and the scene perturbation score calculation submodule, respectively, as follows: In the multi-source consistency score calculation submodule, within the observation window of the candidate event, the net change, rate of rise, rate of change, pulse count, mean, and peak value are extracted from smoke, temperature, C concentration, arc light, electrical load, and video flame texture confidence, respectively. Each parameter is normalized to the range of zero to one according to the preset scale parameter to obtain six evidence components. The multi-source consistency score is obtained by comparing the six evidence components pairwise. Specifically, the consistency weight is calculated for the difference magnitude of any two types of evidence components, and then the consistency weights of all evidence pairs are weighted and normalized according to the preset coupling weight to obtain the multi-source consistency score. In the scene disturbance score calculation submodule, relative humidity, dust concentration, ventilation intensity, and personnel density are normalized and weighted to obtain the score. At the same time, a data quality degradation term is introduced as a disturbance correction. The data quality degradation term is determined by three states: the proportion of missing samples within the observation window, whether the sensor drift exceeds the drift threshold, and whether the data is saturated or truncated. The weighted sum of relative humidity, dust concentration, ventilation intensity, personnel density, and data quality degradation term is constrained by upper and lower limits and outputs a scene disturbance score of zero to one.

6. The fire safety intelligent management system based on big data according to claim 5, characterized in that, The penalty recommendation submodule has a first feedback channel and a second feedback channel. The first feedback channel is used to receive the verification tags and linkage action validity tags made by the fire scene response section for candidate events within a preset short time limit, and adds response delay, location matching degree and data integrity to each verification tag to form immediate credibility. The second feedback channel is used to receive the review tags, false alarm cause categories and penalty parameter correction suggestions from the strategy labeling end for closed-loop events, and adds evidence chain completeness and consistency assessment to each review tag to form authoritative credibility. The penalty recommendation submodule performs hierarchical weighted fusion of the first and second feedback channels based on immediate credibility and authority credibility. When there is a label conflict between the first and second feedback channels, it triggers arbitration rules to output the final label and conflict metric. The arbitration rules include prioritizing authority credibility, suppressing parameter update step size with conflict metric, and increasing the intensity of pseudo-similar removal with conflict metric.

7. The fire safety intelligent management system based on big data according to claim 2, characterized in that, In the evidence feature extraction submodule, the confidence of video flame texture is calculated using time-frequency texture consistency and two-sided suppression. Specifically, within the observation window of the candidate event, the target region is extracted from the video sequence with the spatial location of the candidate event as a constraint. For the target region, the brightness fluctuation sequence and texture sequence are extracted frame by frame. The texture sequence consists of at least one of local gradient magnitude, gradient direction histogram, or local binary mode. Short-time Fourier transform is performed on the brightness fluctuation sequence and texture sequence to obtain the time spectrum. Within the preset flame flicker frequency band, the frequency band energy ratio, spectral concentration, and temporal continuity are calculated to obtain the time-frequency feature group. When mapping time-frequency feature groups to confidence levels, incrementally enhanced suppression is applied to the two non-flame ranges that meet the preset criteria of "too low fluctuation" and "too high fluctuation". When the frequency band energy ratio or spectral concentration is lower than the lower threshold, the confidence level is decreased by increasing the amount of overshoot. When the frequency band energy ratio or spectral concentration is higher than the upper threshold, the confidence level is decreased by increasing the amount of overshoot. When the time-frequency feature group falls into the middle effective interval between the upper and lower boundaries, a preset slope suppression or no suppression is applied, and enhancement is performed according to the continuity of the time domain.

8. A fire safety intelligent management method based on big data, employing the fire safety intelligent management system based on big data as described in any one of claims 1-7, characterized in that, The methods include: Step M1, Data Acquisition: Collect smoke concentration, temperature, CO concentration, electrical load, electrical arc flash signal and video images in the building, and write the sampling timestamp and installation location; Step M2, data preprocessing: Denoising, drift correction, missing data completion and clock alignment are performed on the multi-source data to form a feature sequence of the same observation window; Step M3, Candidate Event Generation: The spatiotemporal fusion model trained based on big data outputs candidate fire events from the feature sequences obtained in step M2 and aggregates them by room topology; Step M4, Evidence Feature Extraction: Calculate the smoke temperature rise rate, CO concentration increment, arc pulse density, and video flame texture confidence within a sliding window; Step M5, Score Calculation: Calculate the multi-source consistency score and scene disturbance score for candidate events, and estimate the false alarm probability and false negative risk; Step M6, Secondary discrimination: Construct a joint discrimination cost by weighted sum of false alarm probability and false negative risk, and apply accelerated penalty to the joint discrimination cost within a preset excessive threshold range and a slow-release penalty within a preset effective intermediate range according to the output penalty method and penalty parameters, and output the secondary discrimination result, which includes fire alarm warning and non-fire alarm; Step M7, Alarm Linkage and Write-back: When the secondary judgment result is a fire alarm warning, an alarm level is generated and the audible and visual alarms, evacuation instructions, smoke exhaust system, sprinkler valve group and power failure control are triggered. The action sequence of the partition and level is executed according to the alarm level, and the linkage handling result is written back to update the confidence threshold lookup table. Step M8, Monitoring Display and Evidence Storage: Display real-time parameter status, alarm trajectory and device status, and store evidence for candidate event links.

9. The intelligent fire safety management method based on big data according to claim 8, characterized in that, In step M7, the alarm linkage executes a zonal and hierarchical action sequence according to the alarm level. Specifically: when the alarm level is low, an audible and visual alarm is triggered only in the fire compartment to which the candidate event belongs, and a verification prompt is generated on the monitoring display module; when the alarm level is medium, in addition to triggering the audible and visual alarm, evacuation instructions are issued to the fire compartment and its adjacent compartments, and the smoke exhaust system of the fire compartment is activated; when the alarm level is high, the sprinkler valve group of the fire compartment is opened on the basis of the medium-level linkage, and the fire pump or pressure stabilizing pump is started; when the alarm level is the highest, the non-fire loads in the fire compartment are cut off on the basis of the high-level linkage, while keeping the emergency lighting, fire control equipment and fire power circuits uninterrupted; wherein, the escalation conditions between each level include the secondary discrimination result continuously meeting the escalation threshold within the preset holding time or the candidate event propagating to the adjacent compartment under the preset diffusion criteria, and each linkage action is configured with a minimum holding time.

10. A fire safety intelligent management method based on big data according to claim 8, characterized in that, In step M7, the linkage commands for the smoke exhaust system, sprinkler valve group, and power failure control are implemented using closed-loop confirmation and failure handling. Specifically, the start-up of the smoke exhaust system is confirmed by fan operation feedback or fan current feedback; the opening of the sprinkler valve group is confirmed by valve position signal and pipeline pressure switch or flow switch signal; and the power failure control is confirmed by the state quantity of the circuit breaker tripping auxiliary contact or the circuit voltage being zero. If any linkage command fails to meet the confirmation criteria within the preset confirmation timeout period, it is re-sent according to the preset number of retries. If it still fails, it switches to the backup equipment or backup circuit to perform degraded linkage. At the same time, the timestamp of the linkage command issuance, the confirmation timestamp, the success or failure flag, and the failure reason code are recorded and written back along with the manual verification conclusion.

Images