A medical institution safety production dynamic evaluation method and system based on multi-source heterogeneous data fusion

Abstract

The application provides a medical institution safety production dynamic evaluation method and system based on multi-source heterogeneous data fusion, and is applied to the field of data processing. The application first collects multi-source heterogeneous full-dimensional data of equipment operation, environment monitoring, personnel operation, fire safety, logistics energy consumption and business process, realizes standardized conversion through cleaning, normalization and feature extraction, completes data deep coupling through weighted fusion, space-time correlation analysis and deep learning, and forms a fusion data set. Then, according to risk identification, grade determination and trend prediction, three-dimensional correlation nodes are constructed in layers, and a safety feature correlation system is established; combined with the analytic hierarchy process and fuzzy comprehensive evaluation modeling, core features are extracted, and a safety feature vector is generated. Relying on the edge-cloud collaborative engine, real-time updating, early warning and iterative optimization are completed, and through accuracy verification and dynamic correction, the final dynamic evaluation result supporting fine safety management and active risk prevention and control is formed.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method and system for dynamic assessment of safety production in medical institutions based on the fusion of multi-source heterogeneous data. Background Technology

[0002] With the expansion of medical institutions' scale of diagnosis and treatment, the increase in the types of equipment, and the improvement of the intelligence level of hospital areas, safety production management has covered multiple dimensions such as equipment operation, environmental monitoring, personnel operation, fire protection, logistics energy consumption, and business processes. The data sources are wide-ranging, the structure is complex, and the collection frequency varies greatly, exhibiting significant multi-source heterogeneous characteristics. Traditional medical institution safety management relies heavily on manual inspections, paper records, regular inspections, and post-event rectification. The data is scattered and stored in different business systems, making it difficult to achieve unified aggregation and efficient utilization. The data standards between subsystems are not uniform, and the spatiotemporal benchmarks are inconsistent, making it impossible to conduct effective correlation and fusion analysis. This easily leads to problems such as delayed hazard detection, inaccurate risk assessment, and one-sided assessment results.

[0003] While some security monitoring systems currently achieve partial data collection and simple alarms, they generally suffer from the following shortcomings: First, multi-source heterogeneous data lacks a standardized processing mechanism; data collected from different devices and systems has inconsistent formats, units, and timestamps, making direct integration and use difficult. Second, security assessments often rely on single indicators or static thresholds, failing to incorporate scientific methods such as hierarchical analysis and fuzzy evaluation for quantitative modeling, resulting in highly subjective and unstable assessment results. Third, most systems only have passive alarm functions, unable to conduct trend prediction and proactive prevention based on historical data, thus failing to meet the needs of dynamic and refined security management. Fourth, there is a lack of collaborative computing mechanisms between the edge and cloud, leading to high local data processing pressure and low transmission efficiency, hindering real-time updates, rapid early warnings, and closed-loop handling. Fifth, security characteristics and assessment results lack a traceability system; no clear correlation is established between indicators, data, and judgment conclusions, making the models uninterpretable and difficult to verify.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore includes information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] According to one aspect of this application, a method for dynamic assessment of safety production in medical institutions based on multi-source heterogeneous data fusion is provided, comprising: acquiring multi-source heterogeneous safety production data from multiple sources, including structured and unstructured data on equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes; completing the standardization transformation of multi-source heterogeneous data based on data cleaning, normalization processing, and feature extraction; achieving deep data coupling by weighted fusion, spatiotemporal correlation analysis, and deep learning models to generate a dynamic assessment fusion dataset for safety production; layering the dynamic assessment fusion dataset for safety production according to the dimensions of risk identification, level determination, and trend prediction; constructing a three-element association node of indicator-data-assessment result in each layer; integrating multi-dimensional safety information through a safety indicator association algorithm to generate a safety assessment dataset for medical institutions. A safety feature association system for comprehensive production control is established. Based on this system, a model is created using the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation. Core features of operational status, environmental parameters, operating procedures, security levels, and energy consumption indicators are extracted to generate safety feature embedding vectors. These vectors are then imported into an edge computing and cloud-based collaborative computing engine, and iterative optimization is performed using real-time updates and anomaly warning mechanisms. Simultaneously, a risk correction decision chain is generated through visualization of safety indicator deviations, forming a basic model for dynamic assessment of medical institution safety production that integrates multi-source heterogeneous data. Based on this model, accuracy verification and dynamic correction are performed using real-time monitoring data and historical safety data to generate dynamic safety assessment results adapted to the refined safety management and proactive risk prevention of medical institutions.

[0006] Another aspect of this application discloses a dynamic safety assessment system for medical institutions based on multi-source heterogeneous data fusion, comprising: a multi-source data acquisition module for acquiring structured and unstructured multi-source heterogeneous safety production data across all dimensions, including equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes, forming a comprehensive safety data collection set; a data fusion processing module for cleaning, normalizing, and extracting features from the multi-source heterogeneous data to achieve standardized transformation, and achieving deep data coupling through weighted fusion, spatiotemporal correlation analysis, and deep learning to generate a dynamic safety assessment fusion dataset; and a safety feature construction module for layering the fusion dataset according to risk identification, level determination, and trend prediction dimensions, constructing a three-element association node of indicator-data-assessment result, and using safety indicators... The association algorithm integrates multi-dimensional security information to generate a comprehensive security feature association system for control. The evaluation modeling module is used to extract core features of operating status, environmental parameters, operating procedures, security level, and energy consumption indicators based on the security feature association system, combined with the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation modeling, and generate security feature embedding vectors. The edge-cloud collaboration module is used to import the security feature embedding vectors into the edge-cloud collaborative computing engine, and iteratively optimize them with real-time updates and anomaly warning mechanisms. It generates a risk correction decision chain through indicator deviation visualization, forming a dynamic evaluation basic model. The dynamic evaluation output module is used to conduct accuracy verification and dynamic correction based on the evaluation basic model combined with real-time monitoring and historical security data, and output dynamic safety production evaluation results adapted to the refined safety management and proactive risk prevention of medical institutions.

[0007] According to another aspect of this application, an electronic device is provided, on which a computer program is stored, which, when executed by a first processor, implements the above-described dynamic assessment method for medical institution safety production based on multi-source heterogeneous data fusion.

[0008] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for dynamic evaluation of medical institution safety production based on multi-source heterogeneous data fusion.

[0009] The beneficial effects of this application are as follows: By collecting multi-source heterogeneous data from across the entire domain, including equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes, the data undergoes standardization processing through cleaning, normalization, and feature extraction. Then, weighted fusion, spatiotemporal correlation, and deep learning are used to achieve deep data coupling. The system constructs a three-layer structure based on risk identification, level determination, and trend prediction, and establishes a three-element association node for indicators, data, and assessment results, forming a unified safety feature association system. Modeling is completed using the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation, generating standardized safety feature embedding vectors. Real-time calculation, rapid early warning, and closed-loop handling are achieved through an edge-cloud collaborative engine. Through accuracy verification and dynamic correction, the final output is a dynamic assessment result that can be used for safety management, achieving full-domain coverage, dynamic perception, scientific assessment, and proactive prevention and control.

[0010] This application aims to standardize and deeply integrate multi-source heterogeneous data, unify spatiotemporal benchmarks and data formats, solve the problems of data dispersion and inability to conduct linked analysis, and improve data utilization. It constructs a ternary correlation and hierarchical assessment system to make risk assessment traceable and verifiable, improving the objectivity and stability of assessment results. Furthermore, by combining the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation, it quantifies indicator weights and risk levels, better aligning with the complex safety assessment scenarios of medical institutions. Employing an edge-cloud collaborative architecture, it enables rapid local response and global cloud optimization, reducing transmission pressure and improving real-time early warning capabilities. It supports dynamic correction and closed-loop handling, achieving full-process management from hazard discovery and early warning push to rectification feedback, thereby improving safety control efficiency and proactive prevention capabilities.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0012] Figure 1 The flowchart illustrates a method for dynamic assessment of medical institution safety production based on multi-source heterogeneous data fusion, as provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the structure of a dynamic assessment system for medical institution safety production based on multi-source heterogeneous data fusion, provided in an embodiment of this application. Detailed Implementation

[0013] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0014] The following is combined Figure 1 This application describes a method for dynamic assessment of safety production in medical institutions based on multi-source heterogeneous data fusion, according to an exemplary embodiment of the present application: S101 acquires multi-dimensional, heterogeneous, multi-source safety production data from various sources, including structured and unstructured data on equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes in medical institutions.

[0015] In one implementation, the six core data collection dimensions for medical institution safety production are first defined, and the specific collection objects and data scope for each dimension are determined to ensure coverage of all key safety control points. The equipment operation dimension collects the operating status and fault information of medical equipment, special equipment, and power supply equipment. The environmental monitoring dimension collects environmental parameter information such as temperature, humidity, harmful gases, noise, and cleanliness. The personnel operation dimension collects the operational behaviors and compliance information of medical staff, logistics personnel, and security personnel. The fire protection and security dimension collects information on the status of fire protection facilities, security monitoring, intrusion alarms, and passageway occupancy. The logistics energy consumption dimension collects water, electricity, gas, and HVAC energy consumption data and pipeline operation parameters. The business process dimension collects data on treatment processes, material flow, emergency response, and hospital management processes.

[0016] Structured data is collected using a targeted approach with fixed formats and fields, directly accessed through existing information system interfaces. Equipment operation data is collected from the equipment monitoring system, with inputs including equipment number, runtime, voltage, current, fault codes, and start / stop status, and output as a standardized equipment operation data table. Environmental monitoring data is collected from the environmental sensing system, with inputs including sensor number, monitoring location, real-time values, and exceedance flags, and output as a standardized environmental parameter record table. Fire safety and security data is collected from the fire safety and security platform, with inputs including facility location, inspection status, alarm records, and linkage status, and output as a standardized fire safety and security status table. Logistics energy consumption data is collected from the energy management system, with inputs including energy consumption points, usage data, energy consumption thresholds, and anomaly flags, and output as a standardized energy consumption dataset. Business process data is collected from the hospital information system, with inputs including process nodes, processing time, executors, and completion status, and output as standardized business process records.

[0017] For unstructured data, text extraction, video parsing, audio recognition, and image acquisition are used to complete the digital conversion. Personnel operation unstructured data comes from surveillance videos and operation record texts; inputs include on-site operation videos, handwritten records, and voice reports. Through behavior recognition and text extraction, descriptions of personnel operation behaviors and compliance statements are output. Fire safety unstructured data comes from on-site photos and alarm recordings; inputs include on-site facility photos, alarm audio, and inspection records. Through image recognition and audio transcription, descriptions of abnormal situations and location information are output. Business process unstructured data comes from meeting minutes, emergency reports, and notification documents; inputs include paper reports, electronic documents, and audio descriptions of processes. Through document parsing and text extraction, key process nodes and execution status are output.

[0018] The structured and unstructured data from the six dimensions mentioned above are uniformly imported into the data collection intermediate library, and formatted and source-marked according to the medical institution safety production data standards. The collected data undergoes integrity verification, checking for missing key dimensions, key fields, and key time points. The collected data undergoes validity verification, removing obviously abnormal, duplicate, and invalid collection records. The collected data is then aligned temporally and spatially, with unified timestamps and location encoding to ensure data correlation and traceability.

[0019] S102 completes the standardization transformation of multi-source heterogeneous data based on data cleaning, normalization processing and feature extraction, and achieves deep data coupling by using weighted fusion, spatiotemporal correlation analysis and deep learning models to generate a dynamic assessment fusion dataset for safe production.

[0020] In one implementation, considering the actual engineering needs of multi-source heterogeneous safety production data fusion in medical institutions, and the technical characteristics of data cleaning, normalization, and feature extraction in medical safety scenarios, three core processing steps—standardization transformation, feature extraction, and deep coupling—are introduced to build a multi-source data fusion execution system specifically adapted to the safety production scenarios of medical institutions. This system adopts a layered, progressive architecture, with a data preprocessing layer, a feature extraction layer, and a data fusion layer sequentially set from top to bottom. These three layers are connected in series according to the data flow direction, forming a stable processing link with unidirectional advancement and no feedback jumps.

[0021] As the first layer of the system, the data preprocessing layer directly receives raw heterogeneous data output from the multi-source data acquisition modules. The data processed covers six categories of structured and unstructured information: equipment operation, environmental monitoring, personnel operation, fire protection and security, logistics and energy consumption, and business processes. This layer mainly performs data cleaning and normalization. For missing fields, it uses the same type of mean filling rule; for outlier values, it uses threshold interval removal rules; and for duplicate records, it uses a combined time and location deduplication rule. At the same time, it maps values ​​of different units and ranges to the same value range, so that all types of raw data have a unified calculation basis.

[0022] The feature extraction layer receives standardized data from the data preprocessing layer and extracts effective information strongly related to safe production for different types of data. It extracts operating status, fault identifiers, and load level features from equipment operation data; parameter values, exceedance status, and fluctuation amplitude features from environmental monitoring data; behavior type, compliance, and operation time features from personnel operation data; and facility status, alarm information, and linkage features from fire and security data. Finally, it outputs standardized feature data with a unified format and fixed dimensions, providing effective input for subsequent fusion processing.

[0023] The data fusion layer receives standardized feature data output from the feature extraction layer and sequentially performs weighted fusion, spatiotemporal correlation analysis, and deep learning coupling processing. Weights are assigned to features of different dimensions according to their importance in medical safety management. Multiple types of data from the same time and location are correlated and matched. Then, a deep learning model is used to complete feature denoising and depth fitting, achieving the organic fusion of multi-source heterogeneous data.

[0024] The three layers are connected in a serial data stream manner. The processing result of the previous layer is directly used as the only input of the next layer. No external additional data is introduced, and the core meaning of the data is not changed. Only format conversion, feature extraction and deep coupling are completed to form a complete data processing link from raw data to standardized fused features. It can stably support subsequent security feature association and dynamic evaluation modeling.

[0025] Following the processing requirements of unified data specifications, complete feature representation, and reliable fusion results, standardized format conversion and full-domain joint integration were carried out on six types of data: equipment operation, environmental monitoring, personnel operation, fire protection and security, logistics energy consumption, and business processes. This unified data structure, field meanings, and storage formats, and completed the alignment and preprocessing of multi-source heterogeneous data. For equipment operation data, it was uniformly converted into a standard format including equipment identifiers, status codes, operating parameters, and timestamps. Equipment identifiers are used to uniquely distinguish medical equipment, special equipment, and power supply equipment; status codes are used to mark operating states such as normal, fault, standby, and alarm; operating parameters include key indicators such as voltage, current, temperature, and load; and timestamps are used to achieve time-series alignment of data across the entire domain.

[0026] For environmental monitoring data, a standardized format is adopted, including monitoring points, parameter types, values, units, and alarm tags. Monitoring points clearly specify the locations of indoor spaces, computer rooms, wards, fire exits, etc. Parameter types cover temperature and humidity, harmful gases, noise, and cleanliness. Values ​​and units are uniformly quantified, and alarm tags directly indicate whether safety thresholds have been exceeded. For personnel operation data, a standardized format is adopted, including personnel identity, operation object, behavior content, and compliance status. Personnel identity distinguishes between medical staff, logistics personnel, security personnel, and third-party personnel. Operation objects correspond to physical entities such as equipment, valves, switches, and fire-fighting equipment. Behavior content records specific operational actions, and compliance status indicates whether it conforms to medical safety operation procedures.

[0027] For fire safety data, a standardized format is implemented, including facility number, location, status, and alarm information. Facility numbers correspond to facilities such as fire extinguishers, fire hydrants, sprinklers, smoke detectors, and access control systems; locations are precise down to the area and floor; status reflects whether the facility is intact, missing, faulty, or overdue for inspection; and alarm information records the trigger time, type, and linkage details. For logistics energy consumption data, a standardized format is implemented, including energy consumption type, usage, threshold, and time period information. Energy consumption types are categorized as electricity, water, gas, and HVAC energy consumption; usage is the actual metered value; thresholds are the upper limits for safe energy consumption; and time period information is divided by hour, shift, and date for easy identification of abnormal fluctuations. For business process data, a standardized format is implemented, including process nodes, entities, time, and completion status. Process nodes cover aspects such as diagnosis and treatment, material transfer, hazardous chemical use, medical waste disposal, and emergency response; entities are the executing departments or responsible persons; time records the start and end times of nodes; and completion status is marked as normal, delayed, interrupted, or abnormal. Through the above standardization transformation, data collected from different systems, protocols, and structures are made to have unified fields, unified units, unified time series, and unified location codes, achieving dual alignment of multi-source heterogeneous data at both the logical and physical levels, and providing a stable and reliable unified data foundation for subsequent feature extraction, data fusion, and dynamic evaluation.

[0028] Based on the execution standards of reliable data quality, satisfactory fusion accuracy, and stable evaluation input, data cleaning rules, normalization parameters, and feature extraction strategies are uniformly configured to achieve standardized processing and effective information extraction of multi-source heterogeneous data, ensuring that the data entering subsequent models is accurate, standardized, and usable. Regarding data cleaning rule configuration, three main categories of rules are set: missing value imputation, outlier removal, and deduplication. The missing value imputation rule addresses field gaps in equipment operating parameters and environmental monitoring values ​​by filling them with historical averages of similar indicators or valid values ​​from adjacent times, ensuring data continuity and integrity. The outlier removal rule identifies and removes values ​​exceeding safety thresholds or exhibiting extreme fluctuations, such as equipment current far exceeding the rated range or unreasonable surges in environmental concentration. The deduplication rule uses a combination of timestamps, location numbers, and indicator types to determine the validity of duplicate records for the same time, location, and indicator, retaining only one valid record to avoid redundancy and interference.

[0029] Regarding normalization parameter configuration, unified settings are applied to numerical data, character data, and time-series data. Numerical data employs a minimum-maximum range mapping, uniformly mapping data of different dimensions such as voltage, temperature, energy consumption, and concentration to a fixed range, eliminating differences in magnitude. Character data uses encoded mapping parameters to convert textual descriptions of equipment status, personnel identity, and compliance status into unified codes for easier calculation. Time-series data is normalized according to a fixed time alignment step, ensuring all data are arranged at the same time interval, achieving system-wide time synchronization.

[0030] In terms of feature extraction strategy configuration, core information strongly related to safe production is extracted from different data sources. Operating status and fault characteristics are extracted from equipment operation data to reflect whether the equipment is working properly and whether potential hazards exist. Parameter and exceedance characteristics are extracted from environmental monitoring data to reflect whether the environment meets medical safety requirements. Behavioral and compliance characteristics are extracted from personnel operation data to reflect whether operations are standardized and whether risky actions exist. Facility and alarm characteristics are extracted from fire and security data to reflect whether the security and fire protection system is intact and effective. Consumption and anomaly characteristics are extracted from logistics energy consumption data to reflect whether energy use is reasonable and whether there are leaks or overloads. Node and efficiency characteristics are extracted from business process data to reflect whether processes such as diagnosis, treatment, supplies, and emergency response are smooth and safe. Through the above configuration, raw heterogeneous data can be transformed into clean, standardized, and unified high-quality data, achieving complete extraction of effective information and providing a stable and reliable input foundation for subsequent data fusion, feature association, and dynamic evaluation.

[0031] Structured and unstructured data, after standardization, cleaning, normalization, and effective information extraction, are uniformly input into a multi-source data fusion execution system. The system employs a three-tiered processing flow—weighted fusion, spatiotemporal correlation analysis, and deep learning models—to achieve deep coupling of multi-source heterogeneous data, ultimately outputting standardized fused feature information with fixed dimensions, unified format, and effective features. In the weighted fusion process, the weighted fusion module assigns independent weight coefficients to six dimensions of data based on the control level and risk impact degree of the medical institution's safety production. Weight settings are determined according to industry safety management standards, with fire safety and equipment operation dimensions receiving higher weights than other dimensions such as environmental monitoring and logistics energy consumption. The module multiplies the extracted feature values ​​of each dimension by their corresponding weights and then sums them to complete the weighted superposition of multi-source features. Through weighted fusion, high-safety-priority indicators dominate subsequent evaluations while preserving the independent physical meaning of each dimension's features.

[0032] In spatiotemporal correlation analysis, the module uses globally unified timestamps and spatial location codes as a benchmark to align and match weighted feature data. The module associates and binds multiple types of monitoring data with the same timestamp and location code, constructing multi-dimensional data relationships at the same time and location. For example, it associates and integrates equipment operating parameters, environmental monitoring values, personnel operation records, and fire protection facility status in the same operating room at the same time, forming a spatiotemporal data unit containing comprehensive safety information, ensuring that the data calculated by the subsequent model maintains a consistent correspondence in time and space.

[0033] In deep learning model processing, the model receives all data after weighted fusion and spatiotemporal correlation. The model consists of an input layer, two fully connected layers, a feature mapping layer, and an output layer connected sequentially. The fully connected layers perform global feature fusion and nonlinear fitting on multi-source heterogeneous features, eliminating redundancy and conflicts between data from different sources. The feature mapping layer reduces and enhances high-dimensional mixed features, filters out data noise, and strengthens key features related to security risks. The model ultimately outputs standardized fused features with fixed dimensions, uniform feature length, continuous and stable values, and can be directly used to construct a security feature association system.

[0034] The entire process takes cleaned and normalized structured and unstructured heterogeneous data as input. After weighted allocation, spatiotemporal alignment, deep fitting and noise reduction optimization, it outputs standardized fusion features with unified dimensions, standardized format, spatiotemporal alignment and significant features. It fully realizes the deep coupling of multi-source heterogeneous data, and provides high-quality and reliable data support for subsequent risk identification, level determination and trend prediction.

[0035] Based on the complete processing flow described above—including standardization transformation, data cleaning, normalization, feature extraction, weighted fusion, spatiotemporal correlation analysis, and deep learning coupling—the full-domain integration and deep fusion of multi-source heterogeneous data can be achieved, generating a fused dataset specifically adapted to the dynamic assessment needs of medical institutions' safety production. This dataset serves as the sole data foundation for subsequent safety feature association, risk identification, level determination, and trend prediction, possessing the characteristics of being standardized, unified, stable, and reliable.

[0036] This fusion dataset comprehensively integrates six categories of safety-related data: equipment operation, environmental monitoring, personnel operation, fire safety and security, logistics energy consumption, and business processes. It includes all standardized and valid feature information. Equipment operation features cover equipment status, fault identification, operating parameters, and load changes. Environmental monitoring features cover temperature and humidity, harmful gases, cleanliness, noise, and exceedances. Personnel operation features cover operational behavior, compliance, operation time, and the objects being operated on. Fire safety and security features cover facility status, alarm information, passageway status, and linkage status. Logistics energy consumption features cover water, electricity, gas, and heating consumption and abnormal fluctuations. Business process features cover nodes and efficiency information for diagnosis and treatment, supplies, hazardous chemicals, medical waste, and emergency response.

[0037] The dataset features a unified data format, with all field names, units, numerical ranges, timestamps, and location codes adhering to the medical institution safety production data specifications, achieving complete alignment across systems, devices, and data types. The dataset boasts comprehensive feature coverage, retaining all valid information strongly related to safety production while eliminating redundant, noisy, and invalid data. It achieves strict spatiotemporal alignment, allowing direct association and retrieval of multi-dimensional data from the same time and location. The dataset exhibits low data noise, enhanced by outlier removal, duplicate data filtering, and deep learning denoising, improving data accuracy and reliability. With high stability and consistency, the dataset can be directly input into subsequent models for computation without further preprocessing.

[0038] This fused dataset serves as the standard input data for constructing a safety feature association system. It provides the risk identification module with criteria for judging anomalies and potential hazards, the risk level assessment module with the required indicator values ​​for risk classification, and the trend prediction module with the basis for historical patterns and temporal changes. Stable, high-quality data input ensures the accuracy of subsequent analytic hierarchy process (AHP) and fuzzy comprehensive evaluation modeling, enhances the credibility and practicality of dynamic assessment results, and provides solid data support for achieving refined safety management and proactive risk prevention in medical institutions.

[0039] S103 stratifies the dynamic assessment and fusion dataset for safe production into layers according to the dimensions of risk identification, level determination, and trend prediction. Each layer constructs a three-element association node of indicator-data-assessment result. Through the safety indicator association algorithm, multi-dimensional safety information is integrated to generate a safety feature association system for the whole-domain management of safe production in medical institutions.

[0040] In one implementation, based on the actual needs of comprehensive safety management in medical institutions and combined with the business logic of multi-dimensional data association and safety assessment, three functional modules—risk identification, level determination, and trend prediction—operate collaboratively to jointly determine the hierarchical method and feature association strategy for the fused dataset. The three modules have clearly defined roles and process in parallel, forming a full-process assessment support capability covering hazard discovery, level assessment, and trend prediction.

[0041] The risk identification module is responsible for anomaly retrieval and hazard location within the fused dataset. Using preset safety thresholds and anomaly judgment rules, it systematically screens data related to equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes, accurately identifying items deviating from normal conditions. Based on real-time data, this module directly marks the location, type, and manifestation of anomalies, providing clear targets for subsequent handling. For example, the module can automatically identify hazards such as excessive equipment temperature, excessive ambient gas concentration, personnel not operating according to regulations, offline fire protection facilities, sudden increases or decreases in energy consumption, and delays in process nodes.

[0042] The risk grading module is used to classify and assess identified risk items. Based on medical institution safety management regulations, and considering factors such as the probability of occurrence, scope of impact, degree of harm, and difficulty of recovery, risks are classified into different safety levels. The module automatically completes the grading through indicator comparison and rule matching, ensuring that the assessment results are consistent, objective, and traceable. For example, high-impact events such as blocked fire exits and critical equipment failures are classified as high-risk events, while minor environmental fluctuations and incomplete general records are classified as low-risk events.

[0043] The trend prediction module, based on historical safety data, cyclical characteristics, and temporal patterns, uses trend fitting and extrapolation algorithms to predict the safety situation over a future period. The module extracts fluctuation characteristics, high-incidence patterns, and changing trends from historical operational data to proactively identify potential risks such as increased risk, equipment aging, environmental degradation, and abnormal energy consumption, providing a basis for proactive prevention and control. For example, it predicts equipment heat dissipation risks based on historical summer high-temperature data and predicts peak energy consumption risks based on cyclical changes in energy consumption.

[0044] The three modules read the dynamic safety assessment fusion dataset in parallel, without interfering with each other and sharing the same data source. During processing, the unified fusion data is first divided into three independent dimensions: risk identification, level determination, and trend prediction. Then, through indicator mapping rules, six categories of raw data are mapped to the assessment stage. Data items such as equipment malfunctions, exceeding environmental parameter limits, personnel operational violations, fire-fighting facility failures, sudden energy consumption changes, and process interruptions are uniformly mapped to the risk identification dimension for rapid hazard detection. The severity of risks, the size of the affected area, the potential harm, and the scope of business impact are mapped to the level determination dimension to determine the risk level. Data such as historical time series changes, seasonal patterns, time period characteristics, and long-term drift trends are mapped to the trend prediction dimension to predict future safety conditions.

[0045] A four-stage linkage mechanism of collection, fusion, layering, and association is adopted to carry out integrated collaborative processing of dynamic assessment datasets for safety production. The four stages are connected in sequence and the data flows in a closed loop to ensure that multi-source heterogeneous data completes structured, layered, and associated processing under a unified framework, ultimately forming standardized safety feature associated data.

[0046] The first phase involves data review. This phase reviews and reuses the multi-source data from the entire domain that has already been collected, cleaned, normalized, and anomaly removed. No duplicate collection or additional processing is performed; instead, valid data from equipment operation, environmental monitoring, personnel operations, fire safety, logistics energy consumption, and business processes are directly accessed. The integrity, temporal continuity, spatial coding consistency, and field validity of data from all areas, departments, and monitoring points across the entire hospital are checked item by item to ensure that the data sources for subsequent processing are consistent, the temporal sequence is continuous, and the location is traceable.

[0047] The second stage is the reading of fused data. The system loads the fused dataset after standardization transformation, weighted fusion, spatiotemporal correlation, and deep learning denoising. This dataset has a unified format, complete features, and spatiotemporal alignment. All indicators, values, units, and codes comply with the medical institution safety production data specifications and can be directly used for hierarchical and correlation construction without further preprocessing.

[0048] The third stage involves dimensional layering. Based on the three assessment purposes—risk identification, risk level determination, and trend prediction—the fused dataset is divided into three independent yet interconnected layers. The risk identification layer is used to extract abnormal indicators in real time, locate potential hazards, and determine safety incidents; the risk level determination layer is used to classify identified risks into safety levels and determine the scope and severity of impact; and the trend prediction layer is used to extrapolate future safety trends and predict potential risks based on historical data and time-series characteristics. These three layers process in parallel without interference, collectively forming the data foundation for the comprehensive assessment.

[0049] The fourth stage is the construction of the association. Within each of the above layers, a three-element association node is established, consisting of indicators, data, and assessment results. The indicator layer corresponds to key items in the safety production assessment, used to clarify the assessment object and monitoring content; the data layer corresponds to the real-time monitoring values, status information, and statistical results of this item, used to provide objective evidence; and the result layer contains preliminary judgment conclusions based on this data, used to identify anomalies, levels, and trends. The three-element association nodes maintain a one-to-one correspondence, ensuring that each assessment is traceable, verifiable, and reproducible.

[0050] At the risk identification layer, a ternary correlation node is constructed, using equipment temperature as an indicator, real-time temperature sampling value as data, and over-temperature risk as the judgment result, to quickly identify abnormal equipment operation. At the level determination layer, a ternary correlation node is constructed, using the fire protection facility integrity rate as an indicator, the ratio of the actual number of intact facilities to the required number as data, and low risk as the judgment result, to intuitively reflect fire protection and security capabilities. At the trend prediction layer, a ternary correlation node is constructed, using energy consumption fluctuation as an indicator, hourly energy consumption sampling sequence as data, and potential future overload as the judgment conclusion, to provide early warning of abnormal energy consumption trends.

[0051] After the three-element node is constructed, four core safety elements—equipment operating status, environmental monitoring parameters, fire safety information, and personnel operation records—are connected to the corresponding nodes, achieving data binding and information association at the same location, time, and theme. Through cross-verification and linkage integration of multi-dimensional information, standardized safety feature association data with complete logic, standardized structure, and comprehensive content is formed, providing a stable and reliable input basis for subsequent feature verification, full-domain association, and assessment modeling.

[0052] A full-domain feature verification is conducted on standardized security feature association data. The full-domain feature verification covers all dimensions, all three-element association nodes, and all security elements. The verification content covers three core contents: data consistency, association accuracy, and risk judgment rationality. The reliable and usable feature data is ensured through item-by-item verification and automatic correction.

[0053] To address data discrepancies caused by different sources, acquisition devices, and transmission formats of heterogeneous data, a consistency correction process is immediately initiated. Based on a unified standard across the entire domain, consistency correction performs unified timestamp standardization, unified physical unit conversion, and unified spatial location coding matching on all data. This ensures that data collected from different systems, devices, and time periods are fully aligned in terms of time reference, measurement units, and spatial coordinates, enabling the correlation, comparison, and fusion of multi-source data. When there are discrepancies in time or value between ambient temperature data and equipment temperature data at the same monitoring location, consistency correction is achieved through unified timestamp alignment and binding with the same spatial location coding, ensuring that both sets of data point to the same monitoring object and the same state at the same time.

[0054] To address correlation errors caused by mismatches between indicators and data, logical conflicts between indicators of different dimensions, and incorrect feature associations, a closed-loop calibration mechanism is automatically triggered. Closed-loop calibration re-matches corresponding monitoring data according to indicator definitions, corrects erroneous correlations according to safety feature association rules, and eliminates conflicts between indicators according to business logic, ensuring that the indicators, data, and evaluation results in each ternary correlation node remain consistent and reasonable. When equipment fault indicators do not match or contradict actual operating data, closed-loop calibration re-verifies the equipment status and fault codes to establish a correct correspondence between indicator data.

[0055] To address the issues of inaccurate assessment results caused by ambiguous safety risk status determination, unclear level boundaries, and ambiguous thresholds, a strict tiered assessment process is implemented. Based on the medical institution's safety production management regulations and preset safety thresholds, this process quantifies and grades various risks related to equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes, clearly classifying risks into three levels: low, medium, and high risk, ensuring clear, objective, and consistent assessment results. When risk levels are ambiguous or cannot be directly distinguished, precise tiering is achieved based on pre-set threshold ranges and scoring rules. After consistency correction, closed-loop calibration, and tiered assessment, optimized safety feature information is generated, characterized by clear dimensional hierarchy, accurate ternary correlation, complete feature integration, and consistent data logic. This optimized safety feature information can be directly used for subsequent full-domain correlation execution and the construction of a safety feature correlation system, providing stable, accurate, and reliable input data for the dynamic assessment model.

[0056] The optimized safety feature information is correlated across the entire domain, establishing a seamless workflow through a safety indicator correlation algorithm. This algorithm is a fully automated multi-source feature correlation and matching algorithm designed for safety production scenarios in medical institutions. With temporal consistency, spatial consistency, and indicator correlation as core constraints, it achieves precise correlation of data across the entire domain through a four-step sequential execution process. The entire process requires no manual intervention and can run stably on both edge and cloud platforms. The algorithm uses standardized spatial location encoding, globally unified timestamps, and standardized indicator type encoding as three unique matching keys, sequentially performing six operations: traversal reading, temporal alignment, spatial grouping, indicator binding, feature aggregation, and logical verification, to complete the automatic matching, correlation, and integration of features across the entire domain.

[0057] After the algorithm starts, it first enters the full-domain traversal reading step. Following a pre-defined sequence of six dimensions—equipment, environment, personnel, fire protection, energy consumption, and processes—it simultaneously reads equipment operating status characteristics, real-time environmental monitoring parameters, personnel operation records, fire safety level assessment results, logistics energy consumption fluctuation data, and overall security situation statistics. During the reading process, the algorithm scans each data point in the risk identification layer, level assessment layer, and trend prediction layer, analyzing each indicator-data-assessment result tripartite correlation node layer by layer. It extracts five types of information from the nodes: indicator name, monitoring value, collection time, location code, and status identifier. All feature data scattered across different layers and nodes are then loaded into a unified correlation calculation unit.

[0058] The second step is to perform time alignment. The algorithm uses the hospital-wide unified time as a benchmark, performing timestamp verification and consolidation on all read data. For multiple data entries with time deviations within the allowable range, they are automatically aligned to the same standard time. For data with discontinuous or missing times, valid values ​​from adjacent times are used to complete the data. For example, if three data entries from the same operating room—equipment temperature, ambient temperature and humidity, and personnel operation records—have collection times differing within the allowable range, the algorithm will align them to the same timestamp to ensure time sequence consistency in subsequent associations.

[0059] The third step involves spatial grouping. The algorithm categorizes data hierarchically according to a four-level structure: hospital area, floor, region, and monitoring point, based on standardized location coding. All data belonging to the same operating room, machine room, fire safety zone, or energy consumption monitoring point are grouped into a single spatial group, allowing data from different dimensions to be aggregated within the same spatial unit. For example, equipment operation data, environmental monitoring data, fire safety facility data, and personnel operation data from the operating room on the east side of the third floor are grouped into the same spatial association group.

[0060] The fourth step involves performing indicator classification and binding. Based on a pre-defined dictionary of medical institution safety production indicators, the algorithm matches and binds multiple data points within the same spatiotemporal group. This binds equipment operation indicators with environmental parameter indicators, personnel operation indicators with fire safety indicators, and energy consumption indicators with business process indicators. For example, the operating status of anesthesia equipment is bound to the operating room's ambient gas concentration indicator, and records of medical staff's operational procedures are bound to the fire escape route status indicator, forming mutually corroborating indicator combinations.

[0061] The fifth step involves performing multi-level feature aggregation. Starting from the lowest-level monitoring points, the algorithm aggregates multiple related features from the same point into point-level features; then, it aggregates features from multiple points within the same region into region-level features; finally, it aggregates all region features into hospital-wide features. The aggregation process preserves the original indicators, data, and evaluation results for each level, forming a hierarchical feature structure that can be traced downwards.

[0062] The sixth step is to perform a global logic verification operation. The algorithm performs business logic checks on the feature combinations that have been associated. It verifies whether there are conflicts between the device status and environmental parameters, whether personnel operations contradict the security status, and whether the energy consumption data matches the process nodes. If a logical conflict occurs, such as a device being marked as stopped but the energy consumption data showing a high load, the algorithm automatically marks it as abnormal and initiates a local re-association; if there is no logical conflict, the association is confirmed to be valid.

[0063] During the association process, all data is strictly matched and associated across the entire domain according to three key values: spatial location code, collection timestamp, and indicator type code. Spatial location code is based on a four-level coding system (campus, floor, department, monitoring point) to group all safety data from the same physical area together. Collection timestamps are aligned chronologically based on the unified synchronized time. Indicator type codes are categorized into six major categories: equipment, environment, personnel, fire protection, energy consumption, and processes, aggregating data of similar indicators. Through the combination and matching of these three key values, precise binding of multi-source data is achieved, forming a structured feature network that is traceable in origin, queryable in detail, and computationally executable.

[0064] This structured feature network comprehensively covers all security dimensions, including equipment operation, environmental monitoring, personnel operation, fire protection, logistics and energy consumption, and business processes. It also runs through the entire assessment process of risk identification, level determination, and trend prediction. It can provide unified feature support for the identification of abnormal hazards, risk level classification, and future situation projection, ultimately forming a security feature association system that is structurally complete, logically coherent, and data reliable.

[0065] S104, based on the security feature association system, combined with the hierarchical analysis method and fuzzy comprehensive evaluation, completes the modeling, extracts the core features of operating status, environmental parameters, operating procedures, security level, and energy consumption indicators, and generates security feature embedding vectors.

[0066] In one implementation, combining the quantitative modeling requirements and multi-dimensional feature mapping logic of dynamic safety assessment in medical institutions, an analytic hierarchy process (AHP) weight configuration rule and a fuzzy comprehensive evaluation and judgment mechanism are introduced to build an overall assessment and modeling framework adapted to medical safety scenarios. This framework uses a safety feature association system as data input, weight calculation and risk assessment as core components, and aims to output standardized feature vectors, fully aligning with the comprehensive control logic of equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes.

[0067] The framework consists of a two-tiered structure: a weight allocation layer and a comprehensive evaluation layer, connected sequentially. Data flows unidirectionally and operates collaboratively between the two layers. The weight allocation layer uses the analytic hierarchy process (AHP) to assign weights to safety indicators, establishing indicator hierarchies, constructing judgment matrices, and ranking and calculating indicator importance. The comprehensive evaluation layer uses fuzzy comprehensive evaluation to determine risk levels, establishing membership functions, performing fuzzy calculations, and outputting quantitative risk scores. Both layers directly interface with the safety feature association system, reading the three-element association node information of indicators, data, and evaluation results in real time, ensuring that the modeling data is consistent, traceable, and verifiable.

[0068] The Analytic Hierarchy Process (AHP) is primarily used to construct a hierarchical structure of safety production indicators. It breaks down complex safety assessment objectives into a multi-level indicator system, determining the degree of influence of each indicator on the final assessment result through pairwise comparisons. This method first sets the overall safety production assessment as the target layer, then sets equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes as the criteria layer, and finally sets the specific monitoring items under each dimension as the indicator layer. Through expert scoring and consistency verification, reasonable weights are assigned to each layer of indicators, highlighting high-priority elements such as fire safety and critical equipment operation. For example, in the criteria layer, fire safety and equipment operation have higher weights than environmental monitoring and logistics energy consumption, making the assessment results more aligned with the key safety management priorities of medical institutions.

[0069] Fuzzy comprehensive evaluation is primarily used to transform difficult-to-quantify qualitative safety descriptions into precise quantitative evaluation values, adapting to the safety assessment needs in medical scenarios where states are ambiguous and boundaries are unclear. This method addresses fuzzy concepts such as whether equipment is functioning properly, whether the environment meets standards, whether operations are standardized, whether security measures are effective, and whether energy consumption is abnormal. It establishes membership functions corresponding to low, medium, and high risks, mapping monitored values ​​to membership degrees corresponding to the safety level, and then obtains a comprehensive score through synthesis calculations. For example, fuzzy states such as slightly exceeding environmental parameters, non-standard personnel operations, and occasional equipment malfunctions are uniformly transformed into calculable quantitative scores, making the evaluation results objective and consistent.

[0070] Combining the quantitative modeling needs and multi-dimensional feature mapping logic for dynamic assessment of safety production in medical institutions, this paper introduces the weight configuration rules of the analytic hierarchy process (AHP) and the fuzzy comprehensive evaluation and judgment mechanism to build an overall assessment modeling framework adapted to medical safety scenarios. This framework uses a safety feature association system as data input, weight calculation and risk assessment as core components, and aims to output standardized feature vectors, fully aligning with the comprehensive control logic of equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes.

[0071] The framework consists of a two-tiered structure: a weight allocation layer and a comprehensive evaluation layer, connected sequentially. Data flows unidirectionally and operates collaboratively between the two layers. The weight allocation layer uses the analytic hierarchy process (AHP) to assign weights to safety indicators, establishing indicator hierarchies, constructing judgment matrices, and ranking and calculating indicator importance. The comprehensive evaluation layer uses fuzzy comprehensive evaluation to determine risk levels, establishing membership functions, performing fuzzy calculations, and outputting quantitative risk scores. Both layers directly interface with the safety feature association system, reading the three-element association node information of indicators, data, and evaluation results in real time, ensuring that the modeling data is consistent, traceable, and verifiable.

[0072] The Analytic Hierarchy Process (AHP) is primarily used to construct a hierarchical structure of safety production indicators. It breaks down complex safety assessment objectives into a multi-level indicator system, determining the degree of influence of each indicator on the final assessment result through pairwise comparisons. This method first sets the overall safety production assessment as the target layer, then sets equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes as the criteria layer, and finally sets the specific monitoring items under each dimension as the indicator layer. Through expert scoring and consistency verification, reasonable weights are assigned to each layer of indicators, highlighting high-priority elements such as fire safety and critical equipment operation. For example, in the criteria layer, fire safety and equipment operation have higher weights than environmental monitoring and logistics energy consumption, making the assessment results more aligned with the key safety management priorities of medical institutions.

[0073] Fuzzy comprehensive evaluation is primarily used to transform difficult-to-quantify qualitative safety descriptions into precise quantitative evaluation values, adapting to the safety assessment needs in medical scenarios where states are ambiguous and boundaries are unclear. This method addresses fuzzy concepts such as whether equipment is functioning properly, whether the environment meets standards, whether operations are standardized, whether security measures are effective, and whether energy consumption is abnormal. It establishes membership functions corresponding to low, medium, and high risks, mapping monitored values ​​to membership degrees corresponding to the safety level, and then obtains a comprehensive score through synthesis calculations. For example, fuzzy states such as slightly exceeding environmental parameters, non-standard personnel operations, and occasional equipment malfunctions are uniformly transformed into calculable quantitative scores, making the evaluation results objective and consistent.

[0074] Based on the operational rules of the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation, each safety indicator is assigned an independent weight and level determination attribute, allowing each indicator to participate in calculations individually or in joint comprehensive evaluations, thus completing the core characteristics of the dynamic evaluation model. In the practical application of AHP, a judgment matrix is ​​first constructed for the safety production evaluation indicator system, comparing the importance of each indicator at the same level pairwise. The comparison process incorporates the actual scenario of safety management in medical institutions, with industry experts objectively scoring indicators based on the actual control priorities of fire protection, equipment, personnel, environment, energy consumption, and processes. The initial weights of each indicator are then obtained through matrix operations. Consistency verification is subsequently conducted to ensure that the weight allocation logic is reasonable and conflict-free, ultimately generating stable and reliable weight coefficients. In this model, the weights of the six dimensions are allocated according to safety importance as follows: fire protection and security dimension weight 0.30, equipment operation dimension weight 0.25, personnel operation dimension weight 0.18, environmental monitoring dimension weight 0.12, business process dimension weight 0.10, and logistics and energy consumption dimension weight 0.05. The sum of all weights is 1, satisfying the normalization requirement. This weighting method does not rely on a single data type, nor is it limited to a single monitoring dimension. It can simultaneously accommodate structured numerical features such as equipment current and environmental concentration, as well as unstructured transformation features such as personnel operation videos and inspection texts, ensuring that data from different sources can participate in the evaluation with reasonable weights.

[0075] In practical applications of fuzzy comprehensive evaluation, corresponding membership functions are established for various indicators related to equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes. These membership functions map continuously changing monitoring values ​​to membership levels of low, medium, and high risk, transforming vaguely defined and difficult-to-determine safety states into clear and calculable quantitative results. For example, environmental monitoring indicators are assigned threshold ranges: values ​​≤ 80% of the safety upper limit are low risk, 80%–100% are medium risk, and > 100% are high risk; equipment operation indicators are assigned: parameters within ±5% of the rated range are low risk, ±5%–±10% are medium risk, and > ±10% are high risk. Through membership mapping, each indicator can independently determine its risk level, without relying on other indicators, retaining independent risk identification capabilities. For example, exceeding environmental concentration standards can be directly determined as a corresponding risk level without waiting for confirmation from other data.

[0076] This dynamic evaluation model supports parallel computation of multi-dimensional indicators. Within the same computation cycle, it can simultaneously process equipment operation characteristics, environmental parameter characteristics, personnel operation characteristics, fire safety and security characteristics, logistics energy consumption characteristics, and business process characteristics. Data from these six dimensions is input, calculated, and processed synchronously without interference or overlap, preventing data conflicts, logical confusion, or computational delays. The indicators employ isolated computational logic, each completing its own weighting and membership mapping before being uniformly aggregated, ensuring a stable and efficient overall evaluation.

[0077] In terms of specific weighting, the model allocates weights according to the safety risk level and accident severity of medical institutions. Equipment operation indicators have a weight of 0.25, higher than environmental monitoring indicators (0.12), directly related to treatment safety and equipment failure risk. Fire safety indicators have a weight of 0.30, higher than logistics energy consumption indicators (0.05), prioritizing personal safety and fire prevention. Meanwhile, indicators related to treatment safety and personnel health, such as ambient temperature and gas concentration, are set as key items that can independently trigger early warnings and risk level determinations. When their values ​​exceed safety thresholds, a corresponding risk level can be directly generated. For example, if the ambient concentration reaches 105% of the rated value, it is directly determined as high risk, without relying on the calculation results of other indicators, further improving the timeliness and accuracy of risk response.

[0078] By integrating multi-source heterogeneous security feature fusion logic, the entire process maintains the tripartite correlation structure of indicators, data, and assessment results without splitting, losing, or altering their correspondence. Stable calculations and feature extraction are performed simultaneously across six major security scenarios: equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes. This ensures traceable data sources, reproducible calculation processes, and verifiable assessment results. Five core features—operating status, environmental parameters, operating procedures, security level, and energy consumption indicators—are uniformly extracted from the security feature-correlated data. Each feature directly corresponds to the indicator item and measured data in the tripartite correlation node, without adding irrelevant dimensions or omitting key information.

[0079] Operational status characteristics are extracted from equipment operation data to reflect whether medical equipment, special equipment, and power supply equipment are in normal operation, standby, fault, or alarm status, providing a basis for identifying potential equipment hazards. Environmental parameter characteristics are extracted from environmental monitoring data to reflect whether indoor temperature, humidity, harmful gas concentration, noise, and cleanliness meet the safety standards of medical institutions, determining whether the environment is compliant and controllable. Operational procedure characteristics are extracted from personnel operation records to reflect whether the operational behaviors of medical, logistics, and security personnel comply with safety procedures, marking compliance, violations, and misoperations. Security level characteristics are extracted from fire protection and security data to reflect the integrity rate, alarm status, and protection reliability of fire protection facilities, access control, smoke detectors, sprinklers, and other systems, assessing the area's security level. Energy consumption index characteristics are extracted from logistics energy consumption data to reflect whether there are sudden increases, leaks, overloads, or other abnormalities in electricity, water, gas, and heating consumption, identifying energy risks.

[0080] The entire feature extraction process strictly preserves the ternary association structure. Each extracted core feature is bound to the original indicator name, real-time monitoring data, and preliminary evaluation results, ensuring direct traceability to the corresponding collection point, monitoring time, and original record. This gives the model complete interpretability and facilitates verification and validation. The model can stably perform feature extraction under six major security scenarios and is compatible with various data formats such as structured numerical values, text states, and time-series curves, without encountering format incompatibility, computational interruptions, or feature loss.

[0081] After feature extraction, the five core features are arranged in a pre-defined fixed-dimensional order. Min-maximum normalization is performed on all numerical features, mapping them uniformly to the 0-1 range to eliminate differences in the magnitude of different indicators. Subsequently, vectors are concatenated in the order of operating status, environmental parameters, operating procedures, security level, and energy consumption indicators to generate a security feature embedding vector with reasonable weight configuration, unified evaluation logic, and complete feature dimensions.

[0082] This vector has a fixed-dimensional structure, containing normalized values ​​and corresponding weight coefficients of all core features. It fully preserves the indicator weights assigned by the analytic hierarchy process and the risk level results obtained from fuzzy comprehensive evaluation. The vector format is standardized, and the values ​​are continuous and stable, allowing direct input into edge computing and cloud-based collaborative computing engines. It supports real-time updates, iterative optimization, and visualization. As a standard output, this vector is directly used for building the subsequent dynamic assessment model, providing a unified, reliable, and stable feature input for real-time risk warnings, indicator deviation corrections, and the generation of risk management decision chains.

[0083] S105 embeds safety features into vectors and imports them into the edge computing and cloud collaborative computing engine. It uses real-time updates and anomaly warning mechanisms for iterative optimization. At the same time, it generates a risk correction decision chain through the visualization of safety indicator deviations, forming a basic model for dynamic assessment of medical institution safety production that integrates multi-source heterogeneous data.

[0084] In one implementation, four types of collaborative constraints—dynamic assessment of medical institution safety production, edge-cloud collaboration, real-time indicator updates, and risk anomaly early warning—are adapted, analyzed, and transformed into scenarios to generate four core features, which constitute the basic information for constructing the dynamic assessment model. This process transforms business constraints into quantifiable, configurable, and verifiable model parameters, clarifies the computing architecture, data format, computing power allocation, and early warning logic, and forms the input basis for a model that can be directly implemented in engineering.

[0085] The system embeds security features into a fixed 5-dimensional structure, a numerical range of 0-1, and a floating-point format, and performs full matching and adaptation with the edge-cloud collaborative engine to ensure that the vectors can be imported and calculated correctly. The system maps the 5-dimensional feature vector (running status, environmental parameters, operating procedures, security level, energy consumption indicators) into a 32-bit floating-point array format that the engine can recognize, fully aligning with the input interfaces of the edge and cloud, achieving compatibility and integration. For example, the vector format is uniformly set to [0.22, 0.15, 0.33, 0.88, 0.41], and the parsing results at both ends are completely consistent, with no calculation deviations.

[0086] Based on the hardware capabilities of the medical institution's on-site edge gateway (CPU clock speed of 1.2GHz, memory of 2GB, sampling frequency of 1Hz), local real-time computing tasks are allocated to specifically handle low-latency indicators such as device status and environmental parameters. The edge node is responsible for preliminary calculation and anomaly detection of the collected data once per second, and only uploads the detection results, reducing the amount of data transmitted from the original 5KB / s to 128Byte / s, reducing the cloud transmission pressure by more than 97%.

[0087] The cloud server is responsible for global data fusion, weight calculation, trend inference, and model iteration, undertaking high-computing tasks. Based on nearly 30 days of historical data from the entire hospital, the cloud performs cross-regional weighted fusion and long-term trend prediction. For example, the cloud uses global weights of 0.30 for fire safety, 0.25 for equipment operation, 0.18 for personnel operation, 0.12 for environmental monitoring, and 0.05 for energy consumption to perform a comprehensive risk score on 128 monitoring points throughout the hospital and output a safety situation change curve for the next 24 hours.

[0088] Establish rules for triggering and pushing abnormal indicators to achieve a risk response within ≤1 second. Preset trigger thresholds: environmental parameters exceeding limits by >100%, equipment parameters deviating from rated parameters by ±10%, and fire protection facilities being offline for ≥3 seconds are considered abnormal. When the concentration of combustible gas in the operating room reaches the 105% safety limit, the edge node immediately triggers a local warning and simultaneously pushes it to the cloud, achieving rapid response through edge alarms, cloud verification, and global notification.

[0089] The system transforms five categories of dynamic on-site factors—equipment operation, environmental monitoring, personnel operation, fire safety, and logistical energy consumption—using logical rules and defining threshold boundaries. This generates five key parameters: safety indicator operation parameters, risk level judgment parameters, trend prediction constraint parameters, abnormal state triggering parameters, and data fusion stability parameters. These parameters collectively constitute the on-site safety state constraints that the model must adhere to, ensuring that the dynamic assessment calculation process closely matches the actual operating scenario of medical institutions, with reasonable values ​​and logical consistency. The safety indicator operation parameters define the operating range and numerical change rules of various safety indicators under normal conditions, providing a standard reference benchmark for the model. The normal operating range for equipment temperature is set to 18℃~45℃, with temperatures exceeding this range considered abnormal; the normal range for ambient humidity is set to 40%RH~60%RH; the reasonable fluctuation range for electricity consumption is set to ≤±10%; and the safe upper limit for gas concentration is ≤50ppm. All indicators are implemented according to the safety production standards of medical institutions, ensuring a unified assessment benchmark.

[0090] Risk level determination parameters are used to set the threshold ranges for indicators corresponding to low, medium, and high risk levels, enabling automatic risk level determination. A fire protection facility integrity rate ≥95% is low risk, 80%–95% is medium risk, and <80% is high risk; environmental parameters ≤80% of the safety upper limit are low risk, 80%–100% are medium risk, and >100% are high risk; equipment operation deviations within ±5% of rated values ​​are low risk, ±5%–±10% are medium risk, and >±10% are high risk; a single alarm signal trigger is a prompt, while continuous triggering for ≥3 seconds indicates high risk. Trend prediction constraint parameters are used to limit the time range, fluctuation upper limit, and reliability conditions for trend extrapolation, preventing prediction results from exceeding reasonable limits. The model limits the trend prediction time range to the next 24 hours; the energy consumption fluctuation upper limit does not exceed 1.2 times the historical average; and the risk change probability reliability threshold is set to ≥85% before outputting prediction results. For example, the energy consumption trend for the next hour and the probability of risk changes for the next two hours are calculated within the above constraints to ensure that the predictions are effective, usable, and reliable. Abnormal state triggering parameters are used to set triggering conditions for abnormal situations such as sudden changes in indicators, continuous exceedances, and data loss, enabling real-time response to anomalies. An anomaly is immediately triggered if the equipment current change is greater than ±15%; an anomaly is triggered if environmental parameters continuously exceed limits for ≥3 minutes; data loss is determined and an anomaly is triggered if monitoring data is not updated for 5 consecutive seconds; and an anomaly alert is triggered if there are ≥2 records of personnel operating in violation of regulations. Data fusion stability parameters are used to specify the alignment rules for multi-source data, time synchronization accuracy, and data validity period to ensure the reliability of the fusion results. Data from different sources at the same location must be within ±1 second timestamps to be fused; spatial location coding accuracy must be down to the specific monitoring point; the validity period for a single data entry is ≤5 minutes, and data exceeding the time limit automatically becomes invalid and is not included in the calculation.

[0091] The above five parameters together constitute complete on-site safety status constraint information, which constrains the model in all dimensions from five aspects: normal value of indicators, risk classification, prediction range, abnormal triggering, and data validity, ensuring that the dynamic evaluation model calculation process is stable, the results are accurate, and it fully conforms to the real safe production and operation logic of medical institutions.

[0092] With the objective function of multi-source heterogeneous data fusion evaluation as the core, and combined with on-site safety status constraint information, we carry out full-dimensional parameter calibration and model accuracy evaluation, and generate four key operating parameters: real-time update frequency parameter, anomaly warning threshold parameter, visualization display configuration parameter, and closed-loop response linkage parameter. Together, they form dynamic evaluation rule optimization information, providing standardized rule basis for stable model operation, accurate early warning, and efficient response.

[0093] The real-time update frequency parameter is used to set the refresh period of data in each dimension and the interval for recalculating indicators, taking into account system load and operational stability while ensuring real-time performance. It is configured differently according to the importance and change rate of data: the device operation status is updated at a high frequency of 1 second per time to ensure that faults and abnormalities can be immediately captured; the environmental monitoring indicators are updated at 5 seconds per time, taking into account both sensitivity and system overhead; the personnel operation and fire and security status are updated at 3 seconds per time; the logistics energy consumption data is updated at 30 seconds per time; the global security risk assessment is updated across the board at 1 minute per time. Through the hierarchical update frequency, it not only ensures the real-time response of highly sensitive indicators but also avoids resource waste caused by high-frequency calculations, enabling the system to run stably in the long term.

[0094] The abnormal warning threshold parameter is accurately calibrated based on historical safety data, equipment rated ranges, and medical institution safety management specifications, reducing the probability of false alarms and missed alarms through threshold optimization. Trigger thresholds are set separately for different indicators: when the device operation parameters deviate from the rated value by more than ±10%, a warning is triggered; when the environmental parameters exceed the safety upper limit by 100%, an immediate warning is issued; when there are 2 or more consecutive violations in personnel operations, a reminder is triggered; when the fire protection facilities are offline for more than 3 seconds or the alarm signal continues to trigger for more than 3 seconds, it is determined as a high-risk warning; when the energy consumption fluctuates by more than ±15%, it is determined as abnormal. By calibrating the sensitivity in different sections and grades, the warning results are made to fit the actual situation on site, improving the reliability of the assessment.

[0095] The visualization display configuration parameter is used to define the display form, refresh rule, and presentation layout of the safety assessment results, risk distribution situation, and indicator deviation. The system supports multi-format visual output: a single indicator is displayed using real-time values; dimension comparison is displayed using bar charts; the全院 risk area distribution is displayed using heat maps; the time-series change is displayed using trend curves. The display interface is automatically refreshed every 30 seconds, and abnormal indicators are highlighted and flashed in red for reminder, medium risks are marked in yellow, and low risks are shown in green, achieving intuitive visualization and quick identification. The visualization configuration covers data dashboards, risk distribution maps, indicator deviation curves, and disposal process tracking diagrams, meeting management and monitoring requirements.

[0096] The closed-loop disposal linkage parameter is used to define the automatic push objects, process flow paths, and task assignment rules after a warning is triggered, establishing a closed-loop mechanism of "warning - determination - push - disposal - feedback". High-risk events (such as fire alarms, environmental concentration exceeding the standard, key equipment shutdown) are automatically pushed to the safety management person in charge and the on-duty leader, and the audible and visual alarm and emergency disposal process are started simultaneously; medium-risk events (such as the local environment approaching the threshold, minor equipment abnormalities, non-standard operations) generate to-do reminder tasks and are pushed to the person in charge of the department; low-risk events are recorded and automatically tracked, and reminders are upgraded after continuous accumulation. All disposal tasks are feedback within a limited time and automatically leave a record, forming a traceable and assessable closed-loop management chain.

[0097] It should be noted that the term "全院" in the original text seems to be a specific term in Chinese, and it might be more appropriate to use a more specific and accurate English expression if there is a more precise one in the relevant context. Here it is tentatively translated as "the whole hospital".The above four parameters work together to form a complete operating rule system for the dynamic evaluation model. They regulate the model's behavior from four aspects: data updates, early warning judgments, visualization, and closed-loop handling. This ensures a stable and reliable evaluation process, accurate and effective early warning outputs, and efficient closed-loop risk handling, fully supporting the implementation of dynamic evaluation of medical institution safety production.

[0098] This model integrates three core elements: basic information for model construction, safety status constraint information, and evaluation rule optimization information. It employs a comprehensive collaborative design and logical binding across the entire process, including safety feature vector import, edge-cloud collaborative computing, real-time data updates, risk anomaly early warning, and indicator visualization. This results in a structurally complete, stable, and timely dynamic assessment model for medical institution safety. The model integrates data access, distributed computing, real-time early warning, visualized control, and closed-loop decision-making, fully aligning with the logic of multi-source heterogeneous data fusion and the safety assessment process. It can directly serve as the operational platform for subsequent accuracy verification and dynamic correction.

[0099] The security feature embedding vector adopts a unified format and standard interface, and is synchronously input to the edge computing node and the cloud server to achieve consistent input from both ends. The edge side is responsible for real-time calculation and on-site judgment of local point data, focusing on processing high real-time data such as equipment operation, environmental parameters, and personnel operations; the cloud side is responsible for global fusion and summary verification of data from multiple areas across the hospital, unifying data standards and spatiotemporal benchmarks. The two ends achieve real-time data synchronization through a dedicated intranet line, ensuring that the local judgment results at the edge end are completely consistent with the global features at the cloud end, without data deviation or timing misalignment. For example, the overheating feature vector of equipment in the same operating room is obtained simultaneously by the edge end and the cloud end [0.75, 0.68, 0.52, 0.83, 0.41], with completely unified calculation benchmarks.

[0100] The edge and cloud adopt a collaborative computing architecture with a division of labor. The edge handles low-latency, high-frequency, and localized computing tasks, performing threshold comparisons, rapid anomaly detection, and on-site early warnings. The cloud handles high-complexity, global, and long-cycle computing tasks, completing analytic hierarchy process (AHP) weight calculations, fuzzy comprehensive evaluation membership degree calculations, historical data trend fitting, and iterative optimization of model parameters. The edge transmits local judgment results back to the cloud in real time, while the cloud simultaneously sends global optimization parameters and comprehensive scores back to the edge, forming a two-way feedback mechanism for calculation results. For example, the edge can determine whether the equipment temperature exceeds the threshold within one second, while the cloud completes comprehensive risk scoring and trend prediction based on data from 128 locations across the hospital.

[0101] Abnormal states are triggered in real-time by the edge device based on preset thresholds, achieving millisecond-level response. Simultaneously, the original abnormal data is uploaded to the cloud. The cloud verifies, confirms, and classifies the abnormal event based on safety status constraints and risk level parameters, and synchronizes the classification results to the visualization platform. The system displays the deviation range of indicators, abnormal location, risk level, and impact range in real-time through a visual interface. Based on the abnormality type, it automatically generates a risk correction decision chain, clearly defining the handling steps, responsible positions, and completion deadlines, achieving a closed-loop push from warning to handling. For example, if the gas concentration in the operating room exceeds the standard, the edge device immediately alarms. After the cloud verifies it as high-risk, it pushes a decision chain of "shutting down equipment—opening windows for ventilation—evacuating personnel—re-inspection and passing."

[0102] The model uses a unified output format at both the edge and cloud levels to generate standardized intermediate assessment results. These results include quantified risk scores, risk levels (low / medium / high), real-time early warning information, model correction parameters, and handling suggestions. All output fields, coding rules, and data structures are completely consistent. These intermediate results can be directly used by the subsequent accuracy verification module for comparison with real-world conditions. They can also be used by the dynamic correction module to adaptively adjust weights, thresholds, and membership parameters. For example, the unified output format is: Risk score 0.83, High risk, Early warning information "Environmental concentration exceeds standard," Correction parameter "Threshold adjusted to 50ppm," and Handling suggestion "Immediately investigate the leak point." This dynamic assessment model is based on an edge-cloud distributed collaborative architecture, uses safety feature embedding vectors as unified computational input, employs real-time updates and anomaly early warnings as its core operating mechanism, and uses indicator deviation visualization and risk correction decision chains as standard output formats. It fully adapts to the actual needs of medical institutions for multi-source heterogeneous data fusion, comprehensive safety monitoring, and dynamic risk assessment. It is engineering-feasible and field-applicable, and can be directly used to generate subsequent dynamic safety production assessment results.

[0103] S106, based on the basic model for dynamic assessment of safety production in medical institutions, combines real-time monitoring data and historical safety data to verify accuracy and dynamically correct, generating dynamic assessment results for safety production that are suitable for the refined safety management and proactive risk prevention and control of medical institutions.

[0104] In one implementation, a unified feature extraction process is performed on the basic model for dynamic assessment of safety production in medical institutions, real-time monitoring data, historical safety data, and safety control constraints. This results in four basic verification features: model accuracy verification features, real-time data adaptation features, historical trend fitting features, and risk level determination features. These features provide standardized and quantifiable basis for subsequent model accuracy verification, dynamic correction, and assessment result output. The model accuracy verification features are used to verify the degree of consistency between the output results of the dynamic assessment basic model and the actual safety status on site. By comparing the model's calculated score with the on-site verification status item by item, the model's accuracy rate, deviation rate, and confidence level are output, achieving an objective judgment of model reliability. This feature uses the risk score and risk level output by the model as input items and the actual safety status obtained from on-site manual verification, equipment measurement, and inspection records as reference items. The accuracy result is obtained through difference calculation and consistency comparison. For example, if the model determines that an operating room area is low-risk, and on-site verification shows that the equipment is operating normally, environmental parameters meet standards, personnel operations are compliant, and the actual situation is also normal, then the model's accuracy is deemed acceptable. However, if the model determines it to be low-risk, but on-site equipment exhibits abnormal overheating, then the model is considered to have a bias and correction is required. Through continuous accuracy verification, the model weights and thresholds are continuously optimized to improve the accuracy of the assessment.

[0105] Real-time data adaptation features are used to convert, normalize, and align the format of the latest collected real-time monitoring data, transforming it into a standard format that the model can directly read and calculate. This ensures that the latest field data can be immediately integrated into the model and used in calculations. This feature uniformly processes real-time uploaded data such as equipment current, voltage, temperature, ambient temperature and humidity, gas concentration, personnel operating status, fire protection facility status, and energy consumption. It maps data with different units, ranges, and acquisition frequencies to the numerical range and encoding system specified by the model, ensuring consistency with the safety feature embedding vector format. For example, raw data such as a real-time uploaded equipment current of 35A and an ambient concentration of 45ppm are converted into standard values ​​in the 0-1 range after minimum-maximum normalization before being entered into the model for real-time calculations and risk assessment.

[0106] Historical trend fitting features, based on long-term accumulated historical safety data, extract trend patterns such as safety status change patterns, indicator fluctuation cycles, abnormal high-incidence periods, and seasonal impact characteristics through time series analysis, periodic statistics, and pattern mining. These features provide a reliable historical reference for the model's trend prediction module. This feature performs a full-domain fitting of historical equipment operation, environmental monitoring, energy consumption changes, and hazard occurrence records to extract stable patterns. For example, by fitting historical data from the past three years, it extracts the trend of medical equipment overheating and insufficient heat dissipation during the summer high-temperature period; it extracts the fluctuation characteristics of generally low energy consumption in the hospital at night and peak energy consumption during the day; and it extracts the period characteristics of higher rates of fire hazards and non-standard personnel operations before and after holidays. This allows the trend prediction module to combine historical patterns to more accurately predict future safety situations.

[0107] The risk level determination feature is based on the safety production control constraints of medical institutions, preset threshold ranges, and indicator mapping rules. It maps objective data such as the monitored values, fluctuation ranges, and durations of abnormalities of various indicators to three levels: low risk, medium risk, and high risk, forming a standardized and reproducible basis for level determination. This feature strictly adheres to the correspondence between indicator values ​​and risk levels, without relying on subjective judgment. Equipment operating parameters within ±5% of the rated range are considered low risk, ±5%–±10% are medium risk, and exceeding ±10% are high risk; environmental monitoring parameters ≤ 80% of the safety upper limit are low risk, 80%–100% are medium risk, and >100% are high risk; fire protection facility integrity rate ≥95% is low risk, 80%–95% is medium risk, and <80% is high risk. Through unified mapping rules, the risk level determination is ensured to be fair, standardized, and traceable.

[0108] Feature extraction processing was performed on three core mechanisms: the weight correction logic of the Analytic Hierarchy Process (AHP), the optimization rules of fuzzy comprehensive evaluation, and the calibration method of multi-source data fusion. This resulted in three optimized correction features: evaluation index correction features, anomaly early warning calibration features, and trend prediction optimization features. These provide quantifiable, executable, and reproducible parameter basis for model adaptive iteration, improved early warning reliability, and optimized trend prediction accuracy. The evaluation index correction feature carries weight parameter data that adaptively adjusts with data deviation. It automatically adjusts the weight coefficients of each safety index in the AHP based on the magnitude and frequency of deviation between real-time data and model output, achieving dynamic self-optimization of the evaluation model. This feature takes real-time monitoring data, historical averages, and model judgment results as inputs and completes adaptive correction through deviation calculation, stability judgment, and weight gain / attenuation rules. When a certain type of index data shows continuous deviation, frequent fluctuations, or deviates from the historical benchmark by more than a set threshold, the system automatically increases the weight of that type of index, enhancing its proportion and sensitivity in the comprehensive evaluation; when the index is stable for a long period without anomalies or deviations, the weight is appropriately reduced to avoid oversensitivity. When environmental monitoring data deviates from the model prediction three times consecutively within one hour, the system automatically increases the weight of the environmental monitoring dimension from 0.12 to 0.15, and simultaneously adjusts the weights of other indicators according to the normalization rules, so that the assessment model can respond more sensitively to environmental changes and improve the success rate of hazard identification.

[0109] The anomaly warning calibration feature is used to eliminate monitoring noise and instantaneous abnormal fluctuations in data through rules such as threshold determination, duration judgment, and fluctuation pattern recognition, avoiding false alarms, misassessments, and mishandling caused by interference signals. This calibration feature takes raw collected data as input and achieves noise filtering through four rules: instantaneous jump identification, duration determination, interval threshold verification, and repetition frequency statistics. Only when an indicator exceeds the standard for a continuously longer period and multiple consecutive samples exceed the threshold is it determined to be a true anomaly; instantaneous changes, single jumps, and spikes caused by momentary interference are all judged as noise and directly filtered, without triggering warnings or participating in evaluation calculations. If an environmental concentration value jumps instantaneously from 30ppm to 80ppm due to electromagnetic interference but recovers immediately within 1 second, the system judges it as noise and filters it; if the environmental concentration remains above 55ppm for 3 minutes (exceeding the safety limit of 50ppm), it is judged as a true anomaly and an warning is activated, effectively reducing the false alarm rate.

[0110] Trend prediction optimization features are used to fit historical time-series data through deep learning models, uncovering the evolutionary patterns of the safety situation over time and improving the accuracy of predicting future trends in equipment, environment, energy consumption, and risks. This feature takes multi-source historical data as input, including equipment operation, environmental parameters, energy consumption fluctuations, and risk occurrence records for the past 7 and 30 days. It models the evolutionary relationships of the situation through time-series fitting, period extraction, trend smoothing, and outlier removal. After learning the patterns of indicator changes, peak values, seasonal patterns, and time-period patterns, the model can output the risk probability, indicator trends, and overload potential for the next few hours, providing a basis for proactive prevention and control. Based on hourly energy consumption data and equipment load data from the past seven days, the deep learning model learns the peak electricity consumption patterns and equipment temperature rise patterns, outputting the operating room equipment load trend and energy consumption change curve for the next three hours, predicting potential overload risks in advance, and achieving preventative management.

[0111] With the core objective of refined safety management and proactive risk prevention in medical institutions, the system combines four fundamental features—model accuracy verification, real-time data adaptation, historical trend fitting, and risk level determination—with three optimization features—assessment indicator correction, anomaly warning calibration, and trend prediction optimization—into a joint iterative calculation and full-dimensional constraint verification. The calculation process simultaneously matches real-time monitoring of safety status with long-term historical control patterns, ensuring that the output results not only closely reflect actual on-site conditions but also conform to the historical operational patterns and management standards of medical institutions' safe production. During verification, the system performs item-by-item constraint verification on the weight parameters of the analytic hierarchy process (AHP), anomaly warning threshold parameters, trend prediction interval parameters, and risk level boundary parameters, ensuring that all parameters are within the pre-set reasonable range and within the scope of medical institution safety management regulations. After each round of model calculation, the model's output risk score and risk level are immediately compared with the actual on-site conditions, inspection records, and equipment measurement data. Deviations are identified, and the parameter correction process is automatically initiated, gradually improving the accuracy and consistency of the assessment results. When the risk level predicted by the model is inconsistent with the real-time monitoring data and on-site verification results, the system automatically starts an iterative correction mechanism, dynamically adjusting the indicator weights, warning thresholds, membership intervals and fusion coefficients according to the magnitude of the deviation, and repeating the calculation until the model output is completely consistent with the actual safety status, ensuring that the assessment results are true, reliable and unbiased.

[0112] After multiple rounds of iterative optimization and constraint verification, the system generates target result coefficients for dynamic safety production assessment. These coefficients quantify the assessment scores, risk levels, and handling priorities of each safety dimension and region. Higher coefficient values ​​indicate higher risk and priority for the corresponding region or indicator. The target result coefficients are calculated by weighted summation of six categories of indicators: equipment operation, environmental monitoring, personnel operation, fire safety, logistics energy consumption, and business processes. The weights of each category are consistent with the previous analytic hierarchy process (AHP) allocation results. For example, the fire safety dimension has the highest weight, and its target result coefficient increases most significantly when an anomaly occurs, indicating a prominent fire risk in the region that must be prioritized for handling and rectification.

[0113] Based on the target result coefficients of dynamic safety production assessment, the system generates complete, standardized, and directly applicable dynamic safety production assessment results for medical institutions. The assessment results include quantitative safety scores for the entire hospital, each department, each area, and each monitoring point, intuitively reflecting the overall and local safety levels and facilitating rapid identification of weaknesses by managers. Simultaneously, the system classifies the entire hospital and key indicators into three levels—low risk, medium risk, and high risk—using green, yellow, and red respectively, making the risk levels clearly visible and easy to understand. For key hidden dangers such as indicators exceeding limits, data anomalies, and status deviations, the system automatically generates real-time early warning information, clearly specifying the location of the anomaly, the type of indicator, the magnitude of the exceedance, the duration, and the scope of impact, achieving precise hazard location.

[0114] The assessment results synchronously record key operational parameters of the model after iterative correction, including analytic hierarchy process (AHP) weights, early warning thresholds, risk assessment boundaries, and data fusion coefficients, making the entire assessment process traceable, reproducible, and verifiable. The system automatically matches and generates corresponding response measures, control requirements, and rectification suggestions based on risk type and risk level, providing standardized guidance for on-site handling, safety inspections, hazard rectification, and accountability. These assessment results fully align with the daily safety management processes of medical institutions and can be directly applied to safety inspections, hazard identification, risk warnings, emergency response, performance evaluation, and long-term management.

[0115] In one implementation, such as Figure 2 As shown, this application also provides a dynamic assessment system for medical institution safety production based on multi-source heterogeneous data fusion, including: The multi-source data acquisition module 201 is used to acquire structured and unstructured multi-source heterogeneous safety production data from all dimensions, including equipment operation, environmental monitoring, personnel operation, fire protection, logistics energy consumption, and business processes of medical institutions, forming a comprehensive safety data collection set. The data fusion processing module 202 is used to clean, normalize and extract features from multi-source heterogeneous data to achieve standardized transformation. Through weighted fusion, spatiotemporal correlation analysis and deep learning, it completes deep data coupling and generates a dynamic safety production assessment fusion dataset. The security feature construction module 203 is used to layer the fused dataset according to the dimensions of risk identification, level determination and trend prediction, construct a three-element association node of indicator-data-evaluation result, and integrate multi-dimensional security information through the security indicator association algorithm to generate a full-domain control security feature association system; The evaluation modeling module 204 is used to model security features based on the security feature association system combined with the analytic hierarchy process and fuzzy comprehensive evaluation, extract the core features of operating status, environmental parameters, operating procedures, security level, and energy consumption indicators, and generate security feature embedding vectors. The edge-cloud collaboration module 205 is used to embed security features into vectors and import them into the edge-cloud collaborative computing engine for real-time updates and iterative optimization of the anomaly warning mechanism. It generates a risk correction decision chain through indicator deviation visualization to form a dynamic assessment basic model. The dynamic evaluation output module 206 is used to conduct accuracy verification and dynamic correction based on the evaluation basic model combined with real-time monitoring and historical safety data, and output dynamic evaluation results of safety production that are adapted to the refined safety management and proactive risk prevention of medical institutions.

[0116] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the dynamic assessment method, electronic device, electronic device, and readable storage medium for medical institution safety production based on multi-source heterogeneous data fusion are basically similar to the above-described embodiments of the dynamic assessment method for medical institution safety production based on multi-source heterogeneous data fusion, so the description is relatively simple. Relevant parts can be referred to in the descriptions of the above-described embodiments of the dynamic assessment method for medical institution safety production based on multi-source heterogeneous data fusion.

Claims

1. A dynamic assessment method for safety production in medical institutions based on multi-source heterogeneous data fusion, characterized in that, include: Acquire multi-dimensional, heterogeneous, multi-source safety production data from various sources, including structured and unstructured data on equipment operation, environmental monitoring, personnel operation, fire safety and security, logistics energy consumption, and business processes in medical institutions. Based on data cleaning, normalization and feature extraction, the multi-source heterogeneous data is standardized and transformed. Weighted fusion, spatiotemporal correlation analysis and deep learning models are used to achieve deep data coupling and generate a dynamic assessment fusion dataset for safe production. The dynamic assessment and fusion dataset for safe production is layered according to the dimensions of risk identification, level determination, and trend prediction. Each layer constructs a three-element association node of indicator-data-assessment result. Multi-dimensional safety information is integrated through a safety indicator association algorithm to generate a safety feature association system for the whole-domain management of safe production in medical institutions. Based on the security feature association system, the hierarchical analysis method and fuzzy comprehensive evaluation are combined to complete the modeling, extract the core features of operating status, environmental parameters, operating procedures, security level and energy consumption indicators, and generate security feature embedding vectors; Safety features are embedded into vectors and imported into edge computing and cloud collaborative computing engines. Real-time updates and anomaly warning mechanisms are used for iterative optimization. At the same time, a risk correction decision chain is generated by visualizing safety indicator deviations, forming a basic model for dynamic assessment of medical institution safety production that integrates multi-source heterogeneous data. Based on the basic model of dynamic safety assessment of medical institutions, combined with real-time monitoring data and historical safety data, the accuracy is verified and dynamically corrected to generate dynamic safety assessment results that are suitable for the refined safety management and proactive risk prevention of medical institutions.

2. The method as described in claim 1, characterized in that, Based on data cleaning, normalization, and feature extraction, multi-source heterogeneous data is standardized and transformed. Weighted fusion, spatiotemporal correlation analysis, and deep learning models are used to achieve deep data coupling, generating a dynamic safety production assessment fusion dataset, including: Combining the engineering needs of multi-source heterogeneous safety production data fusion in medical institutions, and the technical characteristics of data cleaning, normalization processing, and feature extraction, this paper introduces three core processing steps: standardization transformation, feature extraction, and deep coupling, to build a multi-source data fusion execution system adapted to the safety production scenarios of medical institutions. In accordance with the processing requirements of unified data standards, complete feature representation, and reliable fusion results, the data of equipment operation, environmental monitoring, personnel operation, fire protection and security, logistics energy consumption, and business process are standardized, transformed, and jointly integrated to complete the alignment and preprocessing of multi-source heterogeneous data; Based on the execution standards of reliable data quality, compliant fusion accuracy, and stable evaluation input, configure data cleaning rules, normalization parameters, and feature extraction strategies to complete the standardized processing of heterogeneous data and the extraction of effective information. The system integrates structured and unstructured heterogeneous data inputs, and achieves deep data coupling through weighted fusion, spatiotemporal correlation analysis, and deep learning models, outputting standardized fused feature information. Based on the above process, the standardization and deep coupling of multi-source heterogeneous data are completed to generate a fusion dataset adapted to the dynamic assessment of safety production in medical institutions.

3. The method as described in claim 1, characterized in that, The dynamic assessment and fusion dataset for safe production is layered according to the dimensions of risk identification, level determination, and trend prediction. Each layer constructs a three-element association node for indicators, data, and assessment results. Multi-dimensional safety information is integrated through a safety indicator association algorithm to generate a safety feature association system for comprehensive safety management in medical institutions, including: Based on the overall management and control needs of dynamic assessment of safety production in medical institutions and the multi-dimensional data association rules, the risk identification, level determination and trend prediction modules work together to determine the data layering and feature association strategies, and determine the assessment processing form and safety management and control dimensions through dimension decomposition and indicator mapping. A four-stage linkage mechanism of collection-fusion-layering-association is adopted to process the fused dataset in an integrated manner. First, the dimensional layering is completed according to risk identification, level determination and trend prediction. Then, a three-element association node of indicator-data-assessment result is constructed to associate core safety elements such as equipment operation, environmental monitoring, fire protection and security and personnel operation, and generate standardized safety feature association data. Perform full-domain feature verification on standardized security feature association data, initiate consistency correction processing for deviations in multi-source heterogeneous data, trigger a closed-loop calibration mechanism for indicator association errors, execute a graded judgment process for security risk status, and generate optimized security feature information that includes dimensional layering, ternary association, and feature integration. The optimized safety feature information is executed in a comprehensive manner, and the entire process is linked through the safety indicator association algorithm. The system synchronously reads equipment status, environmental parameters, operating procedures, security level, energy consumption data and global security situation information, forming a complete safety feature association system to support the dynamic assessment of safety production in medical institutions.

4. The method as described in claim 1, characterized in that, Based on a security feature association system combined with the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation, modeling was completed. Core features of operational status, environmental parameters, operating procedures, security level, and energy consumption indicators were extracted, generating security feature embedding vectors, including: Combining the quantitative modeling requirements of dynamic assessment of safety production in medical institutions with the logic of multi-dimensional feature mapping, the weight configuration rules of the analytic hierarchy process and the fuzzy comprehensive evaluation and judgment mechanism are introduced to carry out assessment modeling design for the safety feature association system. Based on the assessment requirements of risk identification, level determination, and trend prediction, hierarchical analysis and fuzzy comprehensive evaluation are performed on the safety feature-related data to make it fully match the calculation format of the dynamic assessment model. Based on the operational principles of the Analytic Hierarchy Process (AHP) and fuzzy comprehensive evaluation, the independent weight assignment and level determination attributes of safety indicators are retained, so that they do not depend on a single data type and monitoring dimension, thus completing the construction of the core characteristics of the dynamic evaluation model. By combining multi-source heterogeneous security feature fusion logic, the three-element correlation structure of indicator-data-evaluation result is maintained to ensure that it can be stably calculated in equipment operation, environmental monitoring, personnel operation, fire protection, logistics energy consumption, and business process scenarios. The model adaptation of the feature extraction process is completed, and a security feature embedding vector with reasonable weight configuration, unified evaluation logic, and complete feature dimensions is generated.

5. The method as described in claim 4, characterized in that, Safety features are embedded into vectors and imported into an edge computing and cloud-based collaborative computing engine. Iterative optimization is performed using real-time updates and anomaly warning mechanisms. Simultaneously, a risk correction decision chain is generated through visualization of safety indicator deviations, forming a fundamental model for dynamic assessment of medical institution safety production that integrates multi-source heterogeneous data. This model includes: The collaborative constraints in the dynamic assessment of safety production in medical institutions, edge-cloud collaboration, real-time indicator updates, and risk anomaly warning are adapted, analyzed, and transformed into scenarios to generate safety feature vector adaptation features, edge computing power allocation features, cloud fusion computing features, and real-time risk warning features, forming the basic information for building a dynamic assessment model. The system transforms dynamic changes in equipment operation, environmental monitoring, personnel operation, fire protection and security, and logistics energy consumption into logical rules and defines threshold boundaries to generate safety indicator operation parameters, risk level judgment parameters, trend prediction constraint parameters, abnormal state triggering parameters, and data fusion stability parameters, thus forming on-site safety status constraint information. For the objective function of multi-source heterogeneous data fusion evaluation, parameter calibration and model accuracy evaluation are carried out in combination with security status constraints. Real-time update frequency parameters, anomaly warning threshold parameters, visualization configuration parameters, and closed-loop handling linkage parameters are generated to form dynamic evaluation rule optimization information. By integrating basic information on model construction, safety status constraints, and evaluation rule optimization, and by conducting a full-process collaborative design for safety feature vector import, edge-cloud collaborative computing, real-time updates, anomaly warnings, and visualization, a dynamic evaluation basic model adapted to the safety production scenarios of medical institutions is generated.

6. The method as described in claim 5, characterized in that, Based on the dynamic assessment model for safety production in medical institutions, and combined with real-time monitoring data and historical safety data, accuracy is verified and dynamically corrected to generate dynamic assessment results for safety production that are suitable for the refined safety management and proactive risk prevention and control of medical institutions. These results include: Feature extraction and processing are performed on the basic model for dynamic assessment of safety production in medical institutions, real-time monitoring data, historical safety data and safety control constraints to generate model accuracy verification features, real-time data adaptation features, historical trend fitting features and risk level determination features. Feature extraction processing is performed on the weight correction logic of the analytic hierarchy process, the optimization rules of fuzzy comprehensive evaluation, and the calibration method of multi-source data fusion to generate evaluation index correction features, anomaly early warning calibration features, and trend prediction optimization features. Among them, the evaluation index correction features include weight parameter data that is adaptively adjusted with data deviation; the anomaly early warning calibration features include monitoring noise and abnormal fluctuation data that are removed based on threshold judgment; and the trend prediction optimization features include the time-series evolution relationship data of security situation obtained by fitting a deep learning model. Based on the core needs of refined management and proactive risk prevention and control in medical institutions' safety production, this study combines model accuracy verification features, real-time data adaptation features, historical trend fitting features, risk level determination features, as well as assessment indicator correction features, anomaly warning calibration features, and trend prediction optimization features for iterative calculation and constraint verification. During parameter optimization, real-time safety status and historical control patterns are matched simultaneously to generate dynamic safety production assessment target result coefficients. These coefficients are used to characterize the target assessment scores and risk handling priorities of each safety dimension, forming a dynamic safety production assessment result that includes assessment scores, risk levels, warning information, correction parameters, and handling suggestions.

7. A dynamic assessment system for safety production in medical institutions based on multi-source heterogeneous data fusion, characterized in that, The system includes: The multi-source data acquisition module is used to acquire structured and unstructured multi-source heterogeneous safety production data from all dimensions, including equipment operation, environmental monitoring, personnel operation, fire protection, logistics energy consumption, and business processes in medical institutions, forming a comprehensive safety data collection set. The data fusion processing module is used to clean, normalize, and extract features from multi-source heterogeneous data to achieve standardized transformation. Through weighted fusion, spatiotemporal correlation analysis, and deep learning, it achieves deep data coupling and generates a dynamic safety production assessment fusion dataset. The security feature construction module is used to layer the fused dataset according to the dimensions of risk identification, level determination, and trend prediction, construct a three-element association node of indicator-data-evaluation result, and integrate multi-dimensional security information through a security indicator association algorithm to generate a security feature association system for full-domain control. The evaluation modeling module is used to extract core features of operating status, environmental parameters, operating procedures, security level, and energy consumption indicators based on the security feature association system combined with the analytic hierarchy process and fuzzy comprehensive evaluation model, and generate security feature embedding vectors. The edge-cloud collaboration module is used to embed security features into vectors and import them into the edge-cloud collaborative computing engine for real-time updates and iterative optimization of the anomaly warning mechanism. It generates a risk correction decision chain through indicator deviation visualization, forming a dynamic assessment basic model. The dynamic evaluation output module is used to conduct accuracy verification and dynamic correction based on the evaluation base model combined with real-time monitoring and historical safety data, and output dynamic safety production evaluation results that are adapted to the refined safety management and proactive risk prevention of medical institutions.

8. An electronic device, characterized in that, include: First processor; The processor also includes a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the dynamic evaluation method for safe production in medical institutions based on multi-source heterogeneous data fusion as described in any one of claims 1 to 6 by executing the executable instructions.

Images