Fire safety intelligent evaluation system and closed-loop management and control method

Abstract

The invention discloses a fire safety intelligent evaluation system and a closed-loop management and control method, and the method comprises the steps: analyzing national standard and local standard texts through a natural language processing technology, disassembling terms into a three-stage quantifiable index system, and constructing a dynamic index library comprising at least 200 subdivisions; a dynamic rule base is generated based on a BERT-NER model, a standard parameter threshold value is automatically extracted, and version difference comparison and real-time updating are supported; deploying a fire-fighting special BERT model to analyze an unstructured document, extracting violation events and converting the violation events into structured evaluation factors, aggregating multi-source sensor data and video streams through an edge computing gateway, and adding space-time tags to realize millisecond-level real-time monitoring; generating a comprehensive score in combination with the static weight and the dynamic weight, constructing a risk prediction model based on an Attention-LSTM network, inputting historical hidden danger data, environment variables and enterprise operation data, and outputting future risk probability distribution; according to the invention, compliance management is efficient, assessment accuracy is improved, risk response is real-time, and a management and control process is digital.

Description

Technical Field

[0001] The present invention relates to the field of fire safety management and digitization, and in particular to a fire safety intelligent assessment system and a closed-loop control method. Background Art

[0002] In modern urban management, fire protection has always been an important part of urban safety management. Fire safety is not only related to the overall safety level of the city, but also to social harmony. Once an accident occurs, it may cause huge losses to people's lives and property. Current fire safety assessment technology has the following core flaws:

[0003] Rigid standard adaptation: The existing system relies on manually configured assessment rules and cannot dynamically parse the provisions of national standards (GB) and local standards (DB). For example, when the "Code for Fire Protection Design of Buildings" is updated, indicator thresholds must be manually adjusted, resulting in inefficient compliance management.

[0004] Single evaluation dimension: Most systems focus solely on the status of fire protection facilities, such as sensor data, and lack quantitative analysis of unstructured data such as management system integrity, safety system integrity, building fire protection, and fire zoning compliance, resulting in one-sided evaluation results.

[0005] Risk response lags: Static assessment models based on historical data, such as multivariate linear regression, cannot capture real-time environmental variables such as seasonal risks and peak traffic flow. Warning lead times are less than three days, and corrective recommendations lack specificity.

[0006] Broken control processes: Hazard rectification relies on manual tracking and lacks a digital closed loop of "assessment-warning-rectification-acceptance." This results in a rectification completion rate of only 60%-70%, and data cannot be traced back, making it difficult to meet regulatory audit requirements.

[0007] The current research results on fire safety intelligent assessment systems and closed-loop control methods have been disclosed in Chinese invention patent application documents.

[0008] For example, Chinese invention patent application publication number CN11191942 A discloses a method for assessing and warning fire safety risks. An urban fire risk assessment model is established, comprising five levels: primary, secondary, tertiary, quaternary, and quinary indicators. Wireless transmission devices can be used to upload information on various indicators in real time, allowing for simultaneous observation of data over a large area. Areas with high fire risk values ​​can be captured in real time, and areas with potential safety hazards and equipment failures can be transmitted to relevant units for maintenance. The historical risk assessment model provides management decision-makers with information on the performance of government departments and the fire safety quality of social units. This allows them to identify weak links in urban fire safety, further reduce fire risks, and propose forward-looking fire risk management strategies for assessing cities.

[0009] For another example, Chinese invention patent application publication number CN 116957321 A discloses a fire accident destructiveness assessment method and system. The method includes: obtaining a target fire area and dividing it into multiple target unit areas according to area type; assigning levels to the multiple target unit areas according to preset fire risk levels to obtain fire risk levels for the multiple target unit areas; obtaining fire risk indexes for the multiple target unit areas based on the levels; conducting field surveys of the multiple target unit areas to obtain their fire protection forces, and inputting these forces into a protection assessment model to obtain fire protection indexes for the multiple target unit areas; obtaining multiple fire control risk thresholds, and performing fire control on the target fire areas. This method solves the problem of low assessment accuracy in existing fire accident destructiveness assessment methods due to imperfect assessment indicator settings, and improves the accuracy of fire accident destructiveness assessment, thereby enabling timely and correct measures to control risk areas.

[0010] The technical solution in the aforementioned invention patent application is deficient in that existing technologies fail to achieve the closed-loop technical solution of "intelligent adaptation of standards - multi-source data fusion - dynamic risk assessment - closed-loop control," necessitating a more efficient, accurate, and compliant solution. Against this backdrop, the invention patent application proposes an intelligent fire safety assessment system and closed-loop control method, addressing the shortcomings of traditional approaches. Summary of the Invention

[0011] In order to overcome the above technical problems, the present invention provides a fire safety intelligent assessment system and a closed-loop control method.

[0012] A fire safety intelligent assessment system, comprising:

[0013] The national and local standards adaptive engine module uses natural language processing technology to parse the texts of national and local standards, breaking down the standard clauses into a three-level quantifiable indicator system and building a dynamic indicator library containing at least 200 subdivisions;

[0014] Generate a dynamic rule base based on the BERT-NER model, automatically extract standard parameter thresholds, and support version difference comparison and real-time updates;

[0015] The multimodal data fusion layer module deploys a fire-related BERT model to parse unstructured documents, extract violation events, and convert them into structured assessment factors. It aggregates multi-source sensor data and video streams through an edge computing gateway, and adds spatiotemporal tags to achieve millisecond-level real-time monitoring.

[0016] The intelligent assessment and diagnosis module uses a dual-weight dynamic assessment algorithm, combining static and dynamic weights to generate a comprehensive score. It builds a risk prediction model based on the Attention-LSTM network, inputting historical hidden danger data, environmental variables, and enterprise operation data to output the risk probability distribution for the next 30 / 60 / 90 days.

[0017] The closed-loop control module locates hidden dangers through BIM models and automatically issues work orders, links rectification standards with acceptance processes, and integrates blockchain evidence storage technology to achieve data traceability for the entire process of "assessment-warning-rectification-acceptance";

[0018] The data output by the national standard and landmark adaptive engine module and the multimodal data fusion layer module are respectively input into the intelligent evaluation and diagnosis module for processing. The data processed by the intelligent evaluation and diagnosis module are input into the closed-loop control module for processing and displayed through the visualization platform.

[0019] Furthermore, the dynamic rule base includes:

[0020] Using NLP technology to parse national and local standard texts, the BERT-NER model was used to construct a knowledge graph containing over 100,000 entities. The graph is a mapping relationship between standard parameters, and entity types include clauses, parameters, thresholds, equipment, and locations.

[0021] The improved Word Mover's Distance algorithm is used to calculate the semantic distance between the monitoring data and the standard terms, with a matching accuracy of ≥95%.

[0022] Furthermore, the semantic distance between the monitoring data and the standard clause is calculated by the following formula:

[0023]

[0024] Among them, d1 is the sensor data vector, d2 is the standard term vector, T is the transportation matrix, and c(i,j) is the Euclidean distance between word vectors.

[0025] Furthermore, the dual-weight dynamic evaluation algorithm specifically includes:

[0026] Static weight, based on the mandatory level of the standard, the weight of the strong clause × 1.5;

[0027] Dynamic weights are verified through Monte Carlo simulation, combined with seasonal correction factors, the company's historical risk level and regional differences to ensure that the standard deviation of the assessment results is ≤5%.

[0028] Furthermore, the risk prediction model includes:

[0029] The input layer integrates three types of time series data with a time window of 180 days;

[0030] The feature enhancement layer dynamically focuses on key influencing factors through the attention mechanism. In winter, it automatically assigns a 30% weight to "electrical detection" data, improving the ability to capture long sequence dependencies.

[0031] Output layer: outputs the risk probability distribution for the next 30 / 60 / 90 days and generates a risk heat map;

[0032] Risk list and early warning generation: combining real-time monitoring data with forecast results to automatically generate a structured risk list;

[0033] The multi-dimensional early warning mechanism includes graded early warning and trend early warning:

[0034] Graded warnings: high-risk alerts are sent via SMS, app pop-ups, and sound and light alarms, while medium and low-risk alerts are marked on the management platform.

[0035] Trend warning: when the predicted risk level rises for three consecutive periods, the warning level will be automatically upgraded.

[0036] Furthermore, the risk prediction model input data includes:

[0037] Historical hidden danger data, environmental variables, and enterprise operation data.

[0038] Furthermore, the closed-loop control module includes:

[0039] The work order system module automatically associates BIM model coordinates with sensor positioning data, and the accuracy of hidden danger location marking is ≤0.5 meters;

[0040] It supports mobile photo upload and electronic signature acceptance, shortening the rectification cycle to less than 36 hours.

[0041] The present invention further provides a fire safety closed-loop control method, which is applied to a fire safety intelligent assessment system as described in any one of the above claims, comprising:

[0042] Step 1: Dynamically analyze national and local standards using natural language processing technology, build a three-level evaluation indicator system, generate a dynamic rule base based on the BERT-NER model, automatically extract standard parameter thresholds, and support version difference comparison and real-time updates;

[0043] Step 2: Fusion multimodal data, deploying a fire-specific BERT model to parse unstructured documents, extracting violation text, sensors, and videos, and converting them into structured assessment factors. Multi-source sensor data and video streams are aggregated through an edge computing gateway, and spatiotemporal tags are added to achieve millisecond-level real-time monitoring.

[0044] Step 3: A dual-weight dynamic assessment algorithm is used to calculate the comprehensive risk score. A risk prediction model is constructed based on the Attention-LSTM network. Historical hidden danger data, environmental variables, and enterprise operation data are input. The model outputs the risk probability distribution for the next 30 / 60 / 90 days and generates a graded warning.

[0045] Step 4: Use the BIM model to locate hidden dangers and automatically issue rectification work orders, and achieve closed-loop management of the entire process through blockchain evidence storage.

[0046] Furthermore, the dual-weight dynamic evaluation algorithm in step 3 includes:

[0047] Step 31: static weight setting: according to the mandatory level of the standard, the weight of the strong indicator is increased by 50%;

[0048] Step 32: Dynamic weight adjustment. Based on seasonal risk adjustment, the weight of winter electrical fires is increased by 30% and the weight of historical enterprise data is adjusted by 20%.

[0049] Step 33: Aggregate the indicator scores of the three dimensions of management, construction, and facilities to generate independent sub-item scores;

[0050] Step 34: Based on the final weights and actual scores of each evaluation indicator, the following weighted summation formula is used to calculate the system comprehensive score:

[0051]

[0052] Among them, Score i is the score of the i-th indicator, W 最终i is the corresponding dynamic weight, and β is the scene correction coefficient.

[0053] Furthermore, the closed-loop management in step 4 includes:

[0054] Step 41: Mark the hidden danger locations using the BIM model and sensor positioning data, and associate them with specific national standard clauses;

[0055] Step 42: After rectification, the IoT data is re-checked and the AI ​​image is reviewed, automatically triggering the acceptance process.

[0056] Compared with the existing technologies, the fire safety intelligent assessment system and closed-loop control method of the present invention have the following advantages:

[0057] 1. The fire safety intelligent assessment system and closed-loop control method described in the present invention improves compliance management efficiency, with a standard update response time of ≤1 hour, and automatically generates compliance comparison reports to ensure that the assessment rules deviate from the latest national and local standards by ≤1%;

[0058] 2. The fire safety intelligent assessment system and closed-loop control method described in this invention improves assessment accuracy, with 100% coverage of indicators in the three dimensions of management, construction, and facilities. The quantitative reliability of subjective indicators, such as system implementation, is increased by 30%, and the comprehensive assessment accuracy rate is ≥ 92%;

[0059] 3. The fire safety intelligent assessment system and closed-loop control method described in this invention provide real-time risk response, automatic early warning of high-risk events, and multi-terminal push notifications within 30 seconds. The prediction model identifies risk trends 45 days in advance, reducing the missed detection rate of major hidden dangers to less than 5%.

[0060] 4. The fire safety intelligent assessment system and closed-loop control method described in this invention digitize the control process, improve closed-loop management efficiency by 50%, automatically dispatch rectification work orders with 100% accuracy, and support data backtracking within seconds by regulatory authorities through blockchain evidence storage.

[0061] 5. The fire safety intelligent assessment system and closed-loop control method described in the present invention provide intelligent decision support and an intelligent suggestion engine solution adoption rate exceeding 85%, helping companies reduce fire safety management costs by 40% and promoting the industry's transformation from "post-event disposal" to "pre-event prevention." BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a schematic diagram of the fire safety intelligent assessment system architecture of the present invention;

[0063] Figure 2 This is a schematic diagram of the fire safety intelligent assessment algorithm architecture of the present invention;

[0064] Figure 3 This is a flow chart of the fire safety closed-loop control method described in the present invention. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0066] like Figure 1 As shown, the present invention provides a fire safety intelligent assessment system, comprising:

[0067] The national and local standards adaptive engine module uses natural language processing technology to parse the texts of national and local standards, breaking down the clauses into a three-level quantifiable indicator system and building a dynamic indicator library containing at least 200 subdivisions;

[0068] Generate a dynamic rule base based on the BERT-NER model, automatically extract standard parameter thresholds, and support version difference comparison and real-time updates;

[0069] The multimodal data fusion layer module deploys a fire-related BERT model to parse unstructured documents, extract violation events, and convert them into structured assessment factors. It aggregates multi-source sensor data and video streams through an edge computing gateway, and adds spatiotemporal tags to achieve millisecond-level real-time monitoring.

[0070] The intelligent assessment and diagnosis module uses a dual-weight dynamic assessment algorithm, combining static and dynamic weights to generate a comprehensive score. It builds a risk prediction model based on the Attention-LSTM network, inputting historical hidden danger data, environmental variables, and enterprise operation data to output the risk probability distribution for the next 30 / 60 / 90 days.

[0071] The closed-loop control module locates hidden dangers through BIM models and automatically issues work orders, links rectification standards with acceptance processes, and integrates blockchain evidence storage technology to achieve data traceability for the entire process of "assessment-warning-rectification-acceptance";

[0072] The data output by the national standard and landmark adaptive engine module and the multimodal data fusion layer module are respectively input into the intelligent evaluation and diagnosis module for processing. The data processed by the intelligent evaluation and diagnosis module are input into the closed-loop control module for processing and displayed through the visualization platform.

[0073] The BERT-Based Named Entity Recognition model is a natural language processing technology that combines the BERT pre-trained language model with the named entity recognition (NER) task. BERT, known as the Bidirectional Encoder Representation Model in Chinese, is pre-trained using the Transformer architecture and can understand contextual semantics, particularly excelling at handling ambiguous and complex Chinese sentences.

[0074] NER Chinese Name Entity Recognition: identifying and classifying specific entities (such as names of people, places, time, legal provisions, etc.) from text;

[0075] Attention-LSTM is an attention-based long short-term memory network. It is a deep learning model that combines the time series modeling capabilities of LSTM with the dynamic weighting of the attention mechanism.

[0076] BIM, Building Information Modeling, is an intelligent method that integrates data from the entire life cycle of a building through three-dimensional digital technology.

[0077] Furthermore, the dynamic rule base includes:

[0078] Using NLP technology to parse national and local standard texts, the BERT-NER model was used to construct a knowledge graph containing over 100,000 entities. The graph is a mapping relationship between standard parameters, and entity types include clauses, parameters, thresholds, equipment, and locations.

[0079] The improved Word Mover's Distance algorithm is used to calculate the semantic distance between the monitoring data and the standard terms, with a matching accuracy of ≥95%.

[0080] Specifically, the dynamic rule base generation uses NLP technology to parse national and local standard texts, build a knowledge graph containing more than 100,000 entities, automatically extract parameters such as "fire door fire resistance limit ≥ 1.5h" and generate evaluation rules, support standard version difference comparison and automatic update reminders, and the response time is ≤ 1 hour;

[0081] Through the knowledge graph, sensor data (such as "smoke alarm 3 times / month") is mapped to GB 50116 (monthly alarm ≤ 2 times), generating an over-standard alarm and linking it to specific clauses, enabling one-click traceability of "data-standard-hazard";

[0082] like Figure 2 As shown, a three-layer architecture is used to achieve intelligent mapping between standard clauses and monitoring data;

[0083] By using NLP technology to parse national and local standard texts, we can build a knowledge graph containing more than 100,000 entities. The main entity types include:

[0084] Clause entity: such as "Article 6.5.1 of GB 50016-2014";

[0085] Parameter entities: such as "fire resistance limit of fire door" and "width of evacuation passage";

[0086] Threshold entity: such as "≥1.5h" and "≥1.1m";

[0087] Equipment entities: such as "fire door" and "fire hydrant";

[0088] Location entity: such as "B1-floor power distribution room" and "2nd floor of Building 3";

[0089] Inter-entity relationships include:

[0090] Clause-parameter relationship: such as "Clause 6.5.1 - Fire resistance limit of fire doors";

[0091] Parameter-threshold relationship: such as "fire door fire resistance limit - ≥1.5h";

[0092] Equipment-parameter relationship: such as "fire door-fire resistance limit".

[0093] The multimodal data fusion layer module includes unstructured data processing and real-time access to the Internet of Things. For unstructured data processing, it deploys a fire-specific BERT model to parse inspection reports, drill records, and other documents, extracting violations such as "failure to set up fire warning signs" and "expired plans," and converting them into structured evaluation factors with an accuracy rate of ≥95%. It also supports OCR recognition of inspection forms and automatically extracts key information, such as inspection time and abnormal items.

[0094] Real-time IoT access aggregates 20+ types of sensor data, including smoke detection, water pressure, electrical fire monitoring, and smart camera video streams, through edge computing gateways; additional spatiotemporal tags are added with millisecond accuracy to enable real-time monitoring of facility status, such as "2023-10-01 14:00, the temperature in the B1 floor distribution room is 85°C."

[0095] Multimodal data preprocessing;

[0096] Sensor data cleaning:

[0097]

[0098] Text data entity extraction:

[0099]

[0100] Furthermore, an improved Word Mover's Distance (WMD) algorithm is used, and the semantic distance between the monitoring data and the standard terms is calculated by the following formula:

[0101]

[0102] Among them, d1 is the sensor data vector, d2 is the standard term vector, T is the transportation matrix, and c(i,j) is the Euclidean distance between word vectors.

[0103] The WordMover's Distance (WMD) algorithm is a text similarity calculation method based on natural language processing (NLP) and word embedding, specifically used to measure the semantic difference between two texts.

[0104] Application in fire protection standard analysis, assuming:

[0105] Standard clause: "The width of the evacuation passage shall not be less than 1.2 meters";

[0106] Test report: "The minimum effective width of the aisle is 1.1 meters";

[0107] The calculation process of WMD is as follows:

[0108] Segment the sentence and map it into word vector space (e.g., "evacuation channel" ≈ "walkway", "no less than" ≈ "minimum effective");

[0109] Calculate the semantic distance between word pairs (e.g., the numerical difference between "1.2 meters" and "1.1 meters");

[0110] Find the optimal "word pair matching solution" to minimize the total transportation cost and finally output the similarity score.

[0111] Real-time inference engine;

[0112] Real-time reasoning is implemented based on the Drools rule engine. Examples of core rules are as follows:

[0113]

[0114]

[0115] Algorithm execution process;

[0116] Data access, obtaining real-time sensor data, such as "B1 floor fire door opening time > 30 minutes";

[0117] Entity mapping: mapping data to knowledge graph entities, such as “fire door” → “GB 50016 Article 6.5.1”;

[0118] Threshold comparison: compare parameter values ​​with standard thresholds, such as "on time > 30 minutes" vs. "normal ≤ 5 minutes";

[0119] Rule reasoning, triggering corresponding rules to generate alarms, such as "fire door not closed properly";

[0120] Corrective suggestions, which associate corrective measures in the knowledge graph, such as "check if the door closer is damaged";

[0121] Furthermore, the dual-weight dynamic evaluation algorithm specifically includes:

[0122] Static weight, based on the mandatory level of the standard, the weight of the strong clause × 1.5;

[0123] Dynamic weights are verified through Monte Carlo simulation, combined with seasonal correction factors, the company's historical risk level and regional differences to ensure that the standard deviation of the assessment results is ≤5%.

[0124] Specifically, the core function of the dual-weight dynamic assessment algorithm is to solve the problem of fixed indicator weights in traditional assessments and their inability to respond to real-time risk changes through the dual mechanism of "static weight benchmarking and dynamic weight adaptation";

[0125] The dual-weighted dynamic assessment algorithm first sets "static weights" based on the mandatory levels of national standards (such as strong clauses) to ensure the rigid constraints of the core requirements of fire regulations (for example, the weight of strong indicators such as "evacuation route width" is increased by 50%);

[0126] At the same time, a "dynamic correction coefficient" is introduced to integrate dynamic factors such as seasonal risk characteristics (for example, the weight of the electrical fire indicator is automatically increased by 30% in winter), the company's historical risk level (the weight of the indicator for high-risk units fluctuates by ±20%), and regional environmental differences (the weight of the "pipeline anti-freeze" indicator in northern regions is increased by 5%). The rationality of the weight combination is verified through Monte Carlo simulation (ensuring that the standard deviation of the assessment results is ≤5%).

[0127] Finally, a comprehensive score and sub-item scores are generated through a weighted formula, so that the assessment model can not only adhere to the bottom line of the standard, but also dynamically capture real-time risk changes. The accuracy is improved by 25% compared with the traditional fixed-weight assessment method, which significantly enhances the scientificity and adaptability of the assessment in complex scenarios, and provides accurate data support for risk classification, hidden danger location and rectification suggestions.

[0128] A three-tier risk system is divided based on the comprehensive score, as shown in the following table;

[0129] Risk Level Trigger Conditions Trigger content High risk <70 points When a red alert is triggered, the system automatically generates a "High-risk Rectification Notice" Medium risk 70-89 points Yellow alert, it is recommended to complete targeted rectification within 15 days Low risk ≥90 points Green label, included in routine monitoring

[0130] Intelligent suggestion engine, based on a million-level case library, generates customized solutions through association rule algorithms;

[0131] Monte Carlo simulation is a numerical calculation method based on random sampling and probability statistics. It approximates the solution of complex problems through a large number of random experiments. In fire safety assessment, it can be used to verify the rationality of dynamic weights or predict risk probability distribution.

[0132] Furthermore, the risk prediction model includes:

[0133] The input layer integrates three types of time series data with a time window of 180 days;

[0134] The feature enhancement layer dynamically focuses on key influencing factors through the attention mechanism. For example, in winter, it automatically assigns a 30% weight to "electrical detection" data, improving the ability to capture long sequence dependencies.

[0135] The output layer outputs the risk probability distribution for the next 30 / 60 / 90 days, such as "high risk level probability: 25% on the 30th day, 18% on the 60th day, and 12% on the 90th day." It also generates a risk heat map with color gradients by floor and region.

[0136] Risk list and early warning generation: Combine real-time monitoring data with forecast results to automatically generate a structured risk list, including:

[0137] Risk description, such as "The cable temperature in the B1-floor distribution room remains above 65°C (threshold 55°C), and the probability of fire in the next 7 days is 82%";

[0138] Risk level, marked in red / yellow / green;

[0139] Impact scope: locate the affected area based on the BIM model (e.g., affecting three fire zones);

[0140] The multi-dimensional early warning mechanism includes graded early warning and trend early warning:

[0141] Graded warnings: high-risk alerts are sent via SMS, app pop-ups, and sound and light alarms, while medium and low-risk alerts are marked on the management platform.

[0142] Trend warning: when the predicted risk level rises for three consecutive periods, the warning level will be automatically upgraded, such as medium risk to high risk.

[0143] Furthermore, the risk prediction model input data includes:

[0144] Historical hidden danger data, environmental variables, and enterprise operation data.

[0145] Specifically, the 20-dimensional features of historical hidden danger data include the number of fires, equipment failure frequency, etc.

[0146] 5-dimensional features of environmental variables, including temperature, humidity, number of holidays, etc.;

[0147] The eight-dimensional characteristics of enterprise operation data include peak traffic flow, equipment startup rate, training duration, etc.

[0148] Furthermore, the closed-loop control module includes:

[0149] The work order system module automatically associates BIM model coordinates with sensor positioning data, and the accuracy of hidden danger location marking is ≤0.5 meters;

[0150] Support mobile terminal photo upload and electronic signature acceptance, shortening the rectification cycle to within 36 hours;

[0151] Specifically, hidden danger location and work order distribution, based on BIM models and sensor positioning data, automatically mark the hidden danger location, such as "the door closer of the east fire door on the second floor of Building 3 is damaged", and push it to the responsible person through the work order system, attaching rectification standards, citing specific national standard clauses and acceptance procedures;

[0152] Tracking of rectification progress supports mobile photo upload and electronic signature confirmation. The system generates a Gantt chart to display progress. Failure to complete the rectification within the deadline triggers a secondary warning and sends a copy to management. The average hidden danger handling cycle has been shortened from 7 days to 36 hours.

[0153] Automatic acceptance and data archiving, after rectification, the Internet of Things data will be re-inspected, such as the fire door status sensor feedback closing signal, AI image review (to confirm that the channel is unobstructed). After the acceptance is passed, the data will be stored in the blockchain evidence system to form an electronic file of the entire process of "assessment-early warning-rectification-acceptance".

[0154] like Figure 3 As shown, the present invention also provides a fire safety closed-loop control method, which is applied to a fire safety intelligent assessment system as described in any one of the above claims, comprising:

[0155] Step 1: Dynamically analyze national and local standards using natural language processing technology, build a three-level evaluation indicator system, generate a dynamic rule base based on the BERT-NER model, automatically extract standard parameter thresholds, and support version difference comparison and real-time updates;

[0156] Step 2: Fusion multimodal data, deploying a fire-specific BERT model to parse unstructured documents, extracting violation text, sensors, and videos, and converting them into structured assessment factors. Multi-source sensor data and video streams are aggregated through an edge computing gateway, and spatiotemporal tags are added to achieve millisecond-level real-time monitoring.

[0157] Step 3: A dual-weight dynamic assessment algorithm is used to calculate the comprehensive risk score. A risk prediction model is constructed based on the Attention-LSTM network. Historical hidden danger data, environmental variables, and enterprise operation data are input. The model outputs the risk probability distribution for the next 30 / 60 / 90 days and generates a graded warning.

[0158] Step 4: Use the BIM model to locate hidden dangers and automatically issue rectification work orders, and achieve closed-loop management of the entire process through blockchain evidence storage.

[0159] Furthermore, the dual-weight dynamic evaluation algorithm in step 3 includes:

[0160] Step 31: static weight setting: according to the mandatory level of the standard, the weight of the strong indicator is increased by 50%;

[0161] Step 32: Dynamic weight adjustment. Based on seasonal risk adjustment, the weight of winter electrical fires is increased by 30% and the weight of historical enterprise data is adjusted by 20%.

[0162] Step 33: Aggregate the indicator scores of the three dimensions of management, construction, and facilities to generate independent sub-item scores;

[0163] Step 34: Based on the final weights and actual scores of each evaluation indicator, the following weighted summation formula is used to calculate the system comprehensive score:

[0164]

[0165] Among them, Score i is the score of the i-th indicator, W 最终i is the corresponding dynamic weight, and β is the scene correction coefficient.

[0166] Specifically, the static weight is based on the mandatory level of the standard, with the weight of the mandatory clause multiplied by 1.5. For example, the mandatory clause "water pressure ≥ 0.15 MPa" in the Technical Specifications for Fire Water Supply and Fire Hydrant Systems has a default weight of 15%;

[0167] Dynamic weighting, with seasonal corrections introduced, includes a +30% weighting for electrical fires in summer and a ±20% weighting for historically high-risk units in the enterprise risk rating. Monte Carlo simulation (100,000 iterations) verifies the rationality of the weightings to ensure the standard deviation of the assessment results is ≤5%.

[0168] Item scores: Aggregate the scores of the three dimensions of management, construction, and facilities to generate independent item scores, such as: Score 管理 =∑ J∈管理指标 (Score j ×W 最终j ), the proportion of each sub-item score intuitively shows the weak links of the system. For example, if the Score facility is <60 points, the fire protection facilities will be the key rectification direction.

[0169] The indicators for the three dimensions of management, construction, and facilities are shown in the following table, where management refers to the management system, construction refers to building fire protection, and facilities refers to fire protection facilities;

[0170]

[0171]

[0172] Furthermore, the closed-loop management in step 4 includes:

[0173] Step 41: Mark the hidden danger locations using the BIM model and sensor positioning data, and associate them with specific national standard clauses;

[0174] Step 42: After rectification, the IoT data is re-checked and the AI ​​image is reviewed, automatically triggering the acceptance process.

[0175] The present invention is applicable to the quantitative assessment, risk prediction and full-process closed-loop control of fire safety in special scenarios such as high-rise civil buildings, high-risk fire units, and crowded places; the system of the present invention realizes full-chain intelligent management from risk identification to hidden danger rectification through dynamic adaptation of national standards / local standards, multimodal data fusion, dual-weighted intelligent evaluation and blockchain traceability technology.

[0176] The present invention features intelligent standard adaptation and the world's first dynamic disassembly technology for national and landmark standards. It uses NLP and knowledge graphs to automatically generate and update evaluation rules, increasing efficiency by 80% compared to traditional manual configuration.

[0177] The dual-weighted dynamic assessment of the present invention integrates a dual-weight model that combines standard mandatory levels with real-time risk characteristics, breaking through the limitations of fixed weights and improving assessment accuracy by 25% compared to existing methods.

[0178] The multimodal data fusion of this invention realizes the cross-modal data fusion of BERT text parsing, IoT perception, and AI visual recognition for the first time, increasing the utilization rate of unstructured data from 30% to 90%;

[0179] The invention implements closed-loop control throughout the entire process, constructing a digital closed-loop of "assessment-warning-rectification-acceptance". Through blockchain evidence storage and BIM visualization, it achieves full-process traceability of risk data, increasing the rectification completion rate to 98%.

[0180] The predictive risk management of this invention, based on the Attention-LSTM risk prediction model, realizes probabilistic early warning for the next 90 days, extending the early warning lead time by 45 days compared with traditional threshold alarms.

[0181] The intelligent decision support and risk recommendation engine of the present invention can automatically match rectification plans based on a million-level case library, with a plan adoption rate exceeding 85%, reducing the risk of human decision-making errors.

[0182] The present invention combines visual traceability management, blockchain evidence storage and BIM visualization, supports risk data backtracking at any time point, and meets the full-process audit requirements of regulatory authorities.

[0183] The present invention is not limited to the above-mentioned embodiments. Any modification, improvement, or substitution that can be conceived by those skilled in the art without departing from the essential content of the present invention shall fall within the protection scope of the present invention.

Claims

1. A fire safety intelligent assessment system, characterized in that: include: The national and local standards adaptive engine module uses natural language processing technology to parse the texts of national and local standards, breaking down the standard clauses into a three-level quantifiable indicator system and building a dynamic indicator library containing at least 200 subdivisions; Generate a dynamic rule base based on the BERT-NER model, automatically extract standard parameter thresholds, and support version difference comparison and real-time updates; The multimodal data fusion layer module deploys a fire-related BERT model to parse unstructured documents, extract violation events, and convert them into structured assessment factors. It aggregates multi-source sensor data and video streams through an edge computing gateway, and adds spatiotemporal tags to achieve millisecond-level real-time monitoring. The intelligent assessment and diagnosis module uses a dual-weight dynamic assessment algorithm, combining static and dynamic weights to generate a comprehensive score. It builds a risk prediction model based on the Attention-LSTM network, inputting historical hidden danger data, environmental variables, and enterprise operation data to output the risk probability distribution for the next 30 / 60 / 90 days. The closed-loop control module uses BIM models to locate hidden dangers and automatically assign work orders, linking rectification standards with acceptance processes. It also integrates blockchain evidence storage technology to achieve data traceability for the entire process of "assessment-warning-rectification-acceptance"; The data output by the national standard and landmark adaptive engine module and the multimodal data fusion layer module are respectively input into the intelligent evaluation and diagnosis module for processing. The data processed by the intelligent evaluation and diagnosis module are input into the closed-loop control module for processing and displayed through the visualization platform.

2. The fire safety intelligent assessment system according to claim 1, characterized in that: The dynamic rule base includes: Using NLP technology to parse national and local standard texts, the BERT-NER model was used to construct a knowledge graph containing over 100,000 entities. The graph is a mapping relationship between standard parameters, and entity types include clauses, parameters, thresholds, equipment, and locations. The improved Word Mover's Distance algorithm is used to calculate the semantic distance between the monitoring data and the standard terms, with a matching accuracy of ≥95%.

3. The fire safety intelligent assessment system according to claim 2, characterized in that: The semantic distance between the monitoring data and the standard clauses is calculated by the following formula: Among them, d1 is the sensor data vector, d2 is the standard term vector, T is the transportation matrix, and c(i,j) is the Euclidean distance between word vectors.

4. The fire safety intelligent assessment system according to claim 1, characterized in that: The dual-weight dynamic evaluation algorithm specifically includes: Static weight, based on the mandatory level of the standard, the weight of the strong clause × 1.5; Dynamic weights are verified through Monte Carlo simulation, combined with seasonal correction factors, the company's historical risk level and regional differences to ensure that the standard deviation of the assessment results is ≤5%.

5. The fire safety intelligent assessment system according to claim 1, characterized in that: The risk prediction model includes: The input layer integrates three types of time series data with a time window of 180 days; The feature enhancement layer dynamically focuses on key influencing factors through the attention mechanism. In winter, it automatically assigns a 30% weight to "electrical detection" data, improving the ability to capture long sequence dependencies. Output layer: outputs the risk probability distribution for the next 30 / 60 / 90 days and generates a risk heat map; Risk list and early warning generation: combining real-time monitoring data with forecast results to automatically generate a structured risk list; The multi-dimensional early warning mechanism includes graded early warning and trend early warning: Graded warnings: high-risk alerts are sent via SMS, app pop-ups, and sound and light alarms, while medium and low-risk alerts are marked on the management platform. Trend warning: when the predicted risk level rises for three consecutive periods, the warning level will be automatically upgraded.

6. The fire safety intelligent assessment system according to claim 1, characterized in that: The risk prediction model input data includes: Historical hidden danger data, environmental variables, and enterprise operation data.

7. The fire safety intelligent assessment system according to claim 1, characterized in that: The closed-loop control module includes: The work order system module automatically associates BIM model coordinates with sensor positioning data, and the accuracy of hidden danger location marking is ≤0.5 meters; It supports mobile photo upload and electronic signature acceptance, shortening the rectification cycle to less than 36 hours.

8. A fire safety closed-loop control method, characterized in that: A fire safety intelligent assessment system according to any one of claims 1 to 7, comprising: Step 1: Dynamically analyze national and local standards using natural language processing technology, build a three-level evaluation indicator system, generate a dynamic rule base based on the BERT-NER model, automatically extract standard parameter thresholds, and support version difference comparison and real-time updates; Step 2: Fusion multimodal data, deploying a fire-specific BERT model to parse unstructured documents, extracting violation text, sensors, and videos, and converting them into structured assessment factors. Multi-source sensor data and video streams are aggregated through an edge computing gateway, and spatiotemporal tags are added to achieve millisecond-level real-time monitoring. Step 3: A dual-weight dynamic assessment algorithm is used to calculate the comprehensive risk score. A risk prediction model is constructed based on the Attention-LSTM network. Historical hidden danger data, environmental variables, and enterprise operation data are input. The model outputs the risk probability distribution for the next 30 / 60 / 90 days and generates a graded warning. Step 4: Use the BIM model to locate hidden dangers and automatically issue rectification work orders, and achieve closed-loop management of the entire process through blockchain evidence storage.

9. The fire safety closed-loop control method according to claim 8, characterized in that: The dual-weight dynamic evaluation algorithm in step 3 includes: Step 31: static weight setting: according to the mandatory level of the standard, the weight of the strong indicator is increased by 50%; Step 32: Dynamic weight adjustment. Based on seasonal risk adjustment, the weight of winter electrical fires is increased by 30% and the weight of historical enterprise data is adjusted by 20%. Step 33: Aggregate the indicator scores of the three dimensions of management, construction, and facilities to generate independent sub-item scores; Step 34: Based on the final weights and actual scores of each evaluation indicator, the following weighted summation formula is used to calculate the system comprehensive score: Among them, Score i is the score of the i-th indicator, W 最终i is the corresponding dynamic weight, and β is the scene correction coefficient.

10. The fire safety closed-loop control method according to claim 8, characterized in that: The closed-loop management in step 4 includes: Step 41: Mark the hidden danger locations using the BIM model and sensor positioning data, and associate them with specific national standard clauses; Step 42: After rectification, the IoT data is re-checked and the AI ​​image is reviewed, automatically triggering the acceptance process.

Images