Laboratory safety whole-process dynamic management system based on Internet of Things

Abstract

The invention relates to the technical field of Internet of Things and machine learning, and particularly discloses a laboratory safety whole-process dynamic management system based on the Internet of Things. The system comprises an Internet of Things sensing layer, an edge calculation layer, a data fusion and knowledge construction layer, a dynamic risk assessment and decision-making layer and an execution and feedback layer. A dynamic security knowledge graph is constructed by fusing multi-source heterogeneous data, risk assessment is performed by integrating mode recognition and probability graph reasoning, a hierarchical management and control instruction is generated, and finally closed-loop management is formed through an execution layer, so that real-time and comprehensive perception, accurate risk assessment and adaptive dynamic management and control of a laboratory security state are realized.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things and machine learning technology, specifically relating to an IoT-based dynamic management system for the entire process of laboratory safety. Background Technology

[0002] As a crucial component of the next generation of information technology, the Internet of Things (IoT) enables ubiquitous connectivity between things and between things and people through information sensing devices and network links, profoundly transforming the fields of production management and safety monitoring. Laboratory safety management is a key link in ensuring the smooth progress of scientific research activities and preventing safety accidents. Its core lies in the real-time and accurate monitoring and risk assessment of factors such as the experimental environment, equipment status, personnel behavior, and the flow of hazardous materials.

[0003] A static risk assessment model based on a fixed rule base is adopted, relying on regular manual inspections to obtain safety status information. This approach has significant limitations: static rules are difficult to effectively adapt to the complex and ever-changing risk scenarios arising from the dynamic combination of reagents, equipment, and operating procedures during experiments, resulting in insufficient real-time performance and accuracy of risk assessment; periodic manual inspections cannot achieve continuous perception of safety status, resulting in monitoring blind spots and response delays, and failing to provide early warnings for sudden anomalies or risk accumulation; furthermore, heterogeneous data from various IoT sensing devices (such as temperature and humidity, gas concentration, equipment current, and access control records) lacks effective fusion analysis and semantic association, making it difficult for the system to construct a global safety situation awareness from discrete data, and early warning information is often isolated and lacks contextual support.

[0004] The hidden, interconnected, and dynamically evolving nature of laboratory safety risks makes these problems particularly prominent. Fixed rules and outdated information cannot capture the real-time evolution of risks, leading to reactive safety management. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic management system for laboratory safety based on the Internet of Things, in order to solve the problems in the existing technology that static risk assessment models are difficult to adapt to dynamic risk scenarios, manual inspections have monitoring blind spots and response delays, and the lack of effective integration of heterogeneous data leads to insufficient understanding of the overall safety situation.

[0006] This invention provides an IoT-based dynamic management system for the entire process of laboratory safety. The system includes an IoT sensing layer, an edge computing layer, a data fusion and knowledge construction layer, a dynamic risk assessment and decision-making layer, and an execution and feedback layer. Each layer interacts with data and instructions through a preset communication protocol to collaboratively achieve closed-loop management of the entire process of laboratory safety.

[0007] The IoT sensing layer, deployed within the laboratory's physical space, is used to collect real-time, multi-dimensional raw safety status data. This layer consists of several heterogeneous sensing and identification nodes. The sensing nodes include environmental monitoring sensors, equipment status sensors, and hazardous materials status sensors. The environmental monitoring sensors specifically collect data on temperature, humidity, light intensity, specific gas concentrations, and smoke concentrations within the laboratory. The equipment status sensors specifically collect real-time data on the operating current, voltage, power, operating time, and surface temperature of high-power instruments and equipment. The hazardous materials status sensors specifically collect data on the internal temperature, volatile organic compound concentrations, and door opening / closing status of hazardous chemical storage cabinets. The identification nodes include personnel identification units and item identification units. The personnel identification unit uses biometric or RFID technology to obtain real-time identification information and spatiotemporal location of personnel entering the laboratory or operating specific equipment. The item identification unit scans RFID tags or QR codes to obtain real-time unique identification information, access records, and current location information of experimental reagents, consumables, and hazardous chemicals. All sensing and identification nodes are connected to the laboratory's local area network via wired or wireless means.

[0008] The edge computing layer consists of several edge computing gateways deployed at the edge of the laboratory network. Each edge computing gateway is responsible for connecting to the IoT sensing layer nodes within the physical area. The edge computing gateway is used to perform localized preprocessing and preliminary analysis on the received raw sensing data. Preprocessing includes data cleaning, format standardization, and timestamp alignment. Preliminary analysis includes real-time comparison of data based on preset single-index thresholds. When any sensor data exceeds its safety threshold range, the edge computing gateway immediately generates and reports a level-one local alarm event, while continuously uploading the preprocessed standardized data stream to the data fusion and knowledge construction layer.

[0009] The data fusion and knowledge construction layer is the core data processing hub of the system, used to receive and fuse standardized multi-source heterogeneous data streams from all edge computing gateways and construct a dynamically evolving laboratory safety knowledge graph. This layer includes a spatiotemporal alignment module, an entity relationship extraction module, and a graph dynamic update module. The spatiotemporal alignment module, based on a unified Coordinated Universal Time (UTC) time reference and a laboratory three-dimensional spatial coordinate model, aligns all incoming sensory data, personnel identification data, and item identification data to the same spatiotemporal framework, forming basic fact data tuples with unified spatiotemporal labels.

[0010] The entity relationship extraction module parses basic factual data tuples based on a pre-defined entity type and relationship type ontology library. Entity types include personnel, experimental equipment, chemical reagents, experimental operations, and environmental regions. Relationship types include personnel located in regions, personnel operating equipment, equipment using reagents, reagents stored in regions, and equipment generating environmental parameters. This module automatically identifies entity instances involved in the data tuples and creates or updates the relationships between entities in real time based on data semantics.

[0011] The graph dynamic update module maintains a laboratory safety knowledge graph stored in a graph database. Nodes in this graph represent entity instances, and edges represent the relationships between entities. Both nodes and edges have time-series attributes. The graph dynamic update module continuously receives the output from the entity relationship extraction module and updates the node attributes and edge relationship strengths in the knowledge graph in real time in an incremental manner. It also records the precise timestamp of each change, enabling the knowledge graph to dynamically reflect the real-time status of laboratory safety elements and their historical evolution trajectory.

[0012] The dynamic risk assessment and decision-making layer, based on the dynamic knowledge graph output by the data fusion and knowledge construction layer, performs real-time, online security risk assessments and generates tiered control decisions. This layer includes a risk pattern recognition engine, a probabilistic graphical reasoning module, and a decision generation module.

[0013] The risk pattern recognition engine has a built-in risk pattern feature library trained from historical accident cases and expert rules. The engine continuously performs subgraph matching and pattern scanning on the dynamic knowledge graph to identify whether there are topological structures and attribute combinations that match known risk patterns in the feature library. For example, personnel operating high-risk equipment without wearing protective equipment, incompatible chemicals being stored too close together and the ambient temperature exceeding the standard, and equipment running continuously for too long with abnormal current fluctuations.

[0014] For the identified matching patterns, the engine outputs the corresponding risk type identifier and confidence level. The probabilistic graphical reasoning module is used to evaluate potential associated risks not covered by the risk pattern feature library. This module transforms the dynamic knowledge graph into a probabilistic graphical model, where the state of the nodes is a random variable and the edges represent the conditional dependencies between variables. The module injects evidence into the state of some nodes using received real-time observation data and, based on a Bayesian network reasoning algorithm, propagates the probabilistic influence along the topology of the graph to calculate the posterior probability that key risk nodes are in a dangerous state. For example, after observing evidence that a specific reagent has been used and the ventilation equipment current is zero, the probability that the concentration of toxic gas in the experimental area exceeds the standard can be calculated.

[0015] The decision generation module receives risk identifiers and confidence levels from the risk pattern recognition engine and key risk posterior probabilities from the probabilistic graph inference module. Based on a preset multi-level decision matrix, it generates specific control instructions. The decision matrix defines response strategies corresponding to different risk levels and types, including generating early warning information, sending control suggestions, and triggering automatic control instructions.

[0016] The execution and feedback layer receives and executes control commands issued by the dynamic risk assessment and decision-making layer, and feeds back the execution results and changes in environmental status to the system. This layer includes an information presentation unit, an auxiliary control unit, and a feedback acquisition unit. The information presentation unit, through display terminals and mobile applications within the laboratory, provides real-time safety warnings, risk analysis reports, and operational guidance to management and laboratory personnel at different levels. The auxiliary control unit, through an industrial control interface, sends direct control commands to environmental control equipment and safety facilities within the laboratory. These commands include automatically activating the emergency ventilation system, shutting down specific power circuits, and locking hazardous materials storage cabinet doors. The feedback acquisition unit, through the IoT sensing layer, continuously monitors changes in relevant environmental parameters and equipment status after the execution of commands, and inputs this data as new sensing data into the system, thus forming a complete closed loop from sensing, analysis, decision-making to execution and re-sensing.

[0017] Furthermore, the risk pattern feature library in the risk pattern recognition engine has the following update mechanism: The system establishes a case learning module, which automatically collects data from the entire closed-loop processing of each confirmed security incident, including the initial map status, identified risk patterns, decisions made, and final results; The case learning module uses a difference comparison algorithm to compare and analyze successfully handled cases with unsuccessfully handled or newly emerging abnormal cases, extracting new risk association features or correcting the confidence weights of existing patterns; After the extracted or corrected features are reviewed and confirmed by security experts, they are updated to the risk pattern feature library incrementally, thereby achieving adaptive evolution of the system's risk identification capabilities.

[0018] Furthermore, the structure learning and parameter learning processes of the probabilistic graphical model in the probabilistic graphical inference module are as follows: Structure learning is based on historical safety data, employs a constraint-based method, and utilizes conditional independence tests between variables in the data to initially determine the network topology of dependencies between nodes; parameter learning, based on a fixed network topology, uses maximum likelihood estimation to learn the conditional probability table of node state transitions from historical data. The system sets a model optimization cycle, periodically using newly accumulated closed-loop feedback data to re-estimate and optimize the conditional probability table, enabling the probabilistic graphical model to reflect the latest evolution of laboratory safety status.

[0019] Furthermore, the multi-level decision matrix in the decision generation module is constructed based on the two-dimensional assessment results of the likelihood and severity of the risk. The likelihood dimension is determined by the risk pattern confidence and posterior probability; the severity dimension is graded and assigned values ​​according to the potential for personal injury, property damage, and environmental damage associated with the risk type. The decision matrix divides likelihood and severity into 5 levels and defines 3-level response strategies corresponding to different combinations of the two: Level 1 is low risk, generating only records and prompts; Level 2 is medium risk, generating warnings and pushing disposal suggestions to the mobile terminals of relevant personnel; Level 3 is high risk, immediately generating warnings and triggering the auxiliary control unit to execute automatic intervention, while simultaneously notifying safety management personnel.

[0020] Furthermore, the system operates within a unified time window management framework, which defines three parallel analysis time scales: a real-time stream processing scale, a short-cycle rolling analysis scale, and a long-cycle trend analysis scale. The real-time stream processing scale processes continuous data streams from the IoT sensing layer with a second-level latency, driving risk pattern recognition and immediate alerts. The short-cycle rolling analysis scale, measured in minutes, aggregates and analyzes data from past time windows, driving probabilistic graphical reasoning and situation assessment. The long-cycle trend analysis scale, measured in days or weeks, analyzes the statistical regularities and pattern evolution trends of security events, used to optimize the risk pattern feature library and probabilistic graphical model parameters. The analysis results from these three scales are fused in the decision generation module to ensure that decisions are both real-time and consider historical patterns and trends.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a data fusion and knowledge building layer, fusing and semantically associating heterogeneous multi-source data from the environment, equipment, personnel, and objects within a unified spatiotemporal framework to form a dynamically evolving safety knowledge graph. This design fundamentally changes the isolated and unconnected state of traditional systems, enabling the system to construct a continuous and three-dimensional understanding of the overall safety elements of the laboratory and their interrelationships from discrete sensor readings. This provides a structured data foundation rich in contextual information for subsequent risk analysis, significantly improving the system's depth of understanding of complex risk scenarios.

[0022] 2. This invention proposes a dual-engine architecture for dynamic risk assessment and decision-making, combining a risk pattern recognition engine based on historical pattern matching with a reasoning module based on a probabilistic graphical model. The pattern recognition engine can quickly respond to known, typical risk scenarios, while the probabilistic graphical reasoning module can handle unknown, implicit associated risks, quantifying the probability of potential dangers through probability propagation. This combination enables the system not only to efficiently identify explicit risks but also to deduce and warn of implicit and derivative risks, greatly enhancing the comprehensiveness, foresight, and accuracy of risk assessment, achieving a leap from static rule-based judgment to dynamic scenario assessment.

[0023] 3. This invention achieves a complete closed-loop management system from perception, analysis, decision-making to execution and re-perception, and endows the system with self-evolution capabilities through a case learning module and model optimization mechanism. The automatic control capability of the execution layer ensures immediate intervention in high-risk situations, while the feedback mechanism reintegrates the intervention effect into the analysis cycle.

[0024] 4. This invention can learn from each case and continuously optimize its risk model library and inference model parameters, so that its risk assessment and decision-making capabilities can continuously adapt to new changes and challenges in the laboratory safety environment as the system runs for longer, realizing a fundamental transformation of safety management from passive response to proactive adaptation and continuous optimization. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the IoT-based dynamic management system for laboratory safety throughout the entire process proposed in this invention. Figure 2 This is a schematic diagram illustrating the core principle framework of constructing a dynamic security knowledge graph through data fusion and knowledge construction layers in this invention. Figure 3 This is a logical framework diagram of the dual-engine risk assessment of dynamic risk assessment and decision-making layer in this invention; Figure 4 This is a schematic diagram of the complete closed-loop management process of the system in this invention, from perception, analysis, decision-making to execution and re-perception; Figure 5 This is a schematic diagram of the time window management framework of the system based on multi-timescale parallel analysis in this invention. Detailed Implementation

[0026] Example 1: The overall architecture of the IoT-based dynamic management system for laboratory safety is shown in the attached figure. Figures 1 to 5As shown, the system comprises five main parts: an IoT sensing layer, an edge computing layer, a data fusion and knowledge construction layer, a dynamic risk assessment and decision-making layer, and an execution and feedback layer. These layers interact efficiently and reliably through pre-defined communication protocols, collectively forming a complete security management system with real-time sensing, intelligent analysis, dynamic decision-making, and closed-loop execution capabilities. The specific implementation methods of each component will be described in detail below with reference to the accompanying drawings.

[0027] First, the IoT sensing layer is deployed within the physical space of the laboratory and serves as the data source for the entire system. This layer consists of a large number of heterogeneous sensing and identification nodes. All nodes are connected to the laboratory's local network via wired Ethernet or wireless LAN and follow a unified time synchronization protocol to ensure the consistency of the time reference of the collected data. Environmental monitoring sensors are used to acquire key environmental parameters within the laboratory in real time, including but not limited to temperature, humidity, light intensity, specific gas concentrations (such as carbon monoxide, hydrogen sulfide, ammonia, etc.), and smoke concentration. These sensors employ high-precision electrochemical or optical detection principles, with a sampling frequency of no less than once per second, and the output signal is uploaded in digital form after analog-to-digital conversion.

[0028] Equipment status sensors are integrated inside or outside high-power experimental instruments (such as high-temperature furnaces, centrifuges, lasers, etc.) to collect real-time operating current, voltage, power, cumulative operating time, and shell surface temperature. These sensors typically employ Hall effect current transformers, thermocouples, or infrared temperature measurement modules, with their range and accuracy customized according to equipment specifications. Data update cycles range from 500 milliseconds to 2 seconds. Hazardous materials status sensors are deployed inside hazardous chemical storage cabinets to monitor cabinet temperature, volatile organic compound (VOC) concentration, and cabinet door opening / closing status. VOC sensors utilize metal-oxide-semiconductor (MOS) technology with a response time of less than 10 seconds; cabinet door status is detected via reed switches or Hall position sensors, and status changes trigger immediate event reporting.

[0029] Personnel identification units are deployed at laboratory entrances, critical equipment operating areas, and high-risk area boundaries. They employ biometric identification (such as fingerprints, facial recognition, or iris scanning) or radio frequency identification (RFID) technology to acquire real-time identification information of personnel entering specific areas or operating designated equipment. Upon successful identification, the unit immediately generates an identity event record containing a unique personnel identifier, an identification timestamp, and spatial coordinates (obtained through multi-base station triangulation or UWB ultra-wideband positioning technology).

[0030] The item identification unit uses fixed or handheld reading and writing devices to scan RFID tags or QR codes on reagent bottles, consumable packaging, or hazardous material containers to obtain their unique identifier, name, specifications, expiration date, access history, and current location information. All identification events are accurately timestamped and bound to spatial coordinates, forming complete spatiotemporal trajectory data.

[0031] All the aforementioned raw sensing data and recognition events are transmitted to the edge computing layer via the laboratory's local area network. The edge computing layer consists of several edge computing gateways, each responsible for managing all sensing nodes within a physical sub-area (such as a laboratory room or functional area). The edge computing gateways utilize embedded industrial computer platforms, equipped with real-time operating systems, and possess local data processing and preliminary analysis capabilities. Their core functions include data cleaning, format standardization, and timestamp alignment.

[0032] During data cleaning, outliers caused by sensor malfunctions or communication interference (such as values ​​exceeding physically reasonable ranges or consecutive repetitive values) are removed. Smoothing is achieved using a sliding window mean filter or Kalman filter algorithm. Format standardization converts raw data from different manufacturers and using different protocols into a system-defined JSON Schema format, with fields including data source identifier, measurement value, unit, accuracy level, and acquisition timestamp (UTC format). Timestamp alignment utilizes Network Time Protocol (NTP) or Precision Time Protocol (PTP) to synchronize all data streams at the microsecond level, ensuring strict time consistency of data from different sensors at the same time.

[0033] After preprocessing, the edge computing gateway performs preliminary analysis: for each type of sensor data, an independent safety threshold range is set (e.g., a temperature limit of 35 degrees Celsius and a VOCs concentration limit of 50 ppm). When any data point exceeds its corresponding threshold, the gateway immediately generates a Level 1 local alarm event. The event content includes the alarm type, the value exceeding the limit, the location of occurrence, the timestamp, and the associated device / area identifier, and is immediately reported to the data fusion and knowledge construction layer through a priority queue. Simultaneously, regardless of whether an alarm is triggered, all standardized data streams are continuously uploaded to the central data processing layer at fixed intervals to ensure the completeness and timeliness of the global analysis.

[0034] The data fusion and knowledge construction layer, as the core data hub of the system, has the following structure: Figure 2As shown, the system consists of three collaborative modules: a spatiotemporal alignment module, an entity relationship extraction module, and a dynamic graph update module. This layer receives standardized data streams from all edge computing gateways and builds and maintains a dynamically evolving laboratory safety knowledge graph based on these streams. The spatiotemporal alignment module first establishes a unified spatiotemporal reference system: the time dimension uses Coordinated Universal Time (UTC) as the global benchmark, while the spatial dimension is based on a three-dimensional coordinate system constructed from the laboratory building information model (BIM), assigning unique spatial coordinates (x, y, z) to each sensor, device, and personnel activity area. All incoming data tuples (such as "temperature sensor A measured 28.5℃ at time T" and "person Zhang San is located at coordinates (3.2, 4.1, 1.5) at time T+0.2 seconds") are mapped to this unified framework, forming basic fact data tuples with precise spatiotemporal labels.

[0035] The entity relation extraction module performs semantic parsing on basic fact data tuples based on a pre-defined ontology library. The ontology library explicitly defines five core entities: personnel, experimental equipment, chemical reagents, experimental operations, and environmental regions; and six core relations: The system identifies the location of personnel, the equipment they operate, the reagents used by the equipment, the storage location of the reagents, the environmental parameters generated by the equipment, and the reagents involved in the operation. For example, when the system receives two data points: "Personnel Zhang San swipes his card at time T to turn on centrifuge B" and "Centrifuge B starts at time T with a current of 15 amps," the module automatically identifies the entities "Zhang San" and "centrifuge B" and creates the relationship "Zhang San operates centrifuge B." Simultaneously, by combining the configuration information of centrifuge B (such as its commonly used reagent being ethanol), the potential relationship "Centrifuge B uses ethanol" can be inferred.

[0036] All entity instances are assigned a unique Global Identifier (GUID), whose attributes include static attributes (such as device model, reagent CAS number) and dynamic attributes (such as current status, last operation time). Relationship edges also have temporal attributes, recording the start time, duration, and confidence level of the relationship.

[0037] The knowledge graph dynamic update module uses a graph database (such as Neo4j or JanusGraph) as its underlying storage to maintain the real-time state of the aforementioned knowledge graph. Whenever the entity-relation extraction module outputs a new entity or relation, this module performs an incremental update operation: if the entity already exists, its dynamic attributes are updated; if the relation already exists, its valid time interval is extended or its confidence level is adjusted; if it is a completely new entity or relation, a new node or edge is created. All changes are recorded with precise timestamps and support backtracking to any historical snapshot of the knowledge graph by time slice. Therefore, the knowledge graph not only reflects the current state of the laboratory's safety elements but also fully records their evolutionary trajectory, providing rich contextual information for subsequent risk analysis.

[0038] The dynamic risk assessment and decision-making layer conducts real-time risk assessments based on the aforementioned dynamic knowledge graph. Its logical architecture is shown in the attached figure. Figure 3 As shown, a dual-engine driven mode is adopted: a risk pattern recognition engine and a probabilistic graphical inference module. The risk pattern recognition engine has a built-in risk pattern feature library, which is trained from historical accident reports, expert experience rules, and simulation exercise data.

[0039] Each risk pattern is stored as a subgraph template, containing specific combinations of node types, edge relationship topology, and attribute constraints. For example, a typical high-risk pattern is defined as: "The personnel node is not associated with the 'wearing protective equipment' attribute, and has an 'operation' relationship with the 'high-risk equipment' node, and the 'toxic gas concentration' attribute value of the area where the equipment is located is greater than the threshold."

[0040] The engine performs a full scan of the current knowledge graph at a fixed frequency, executes a subgraph isomorphic matching algorithm, and once a matching instance is found, it outputs a risk type identifier (such as "unprotected operation of high-risk equipment"), a matching confidence score (calculated based on attribute matching degree and historical occurrence frequency), and a list of involved entities.

[0041] For potential risks not covered by the feature library, the probabilistic graphical reasoning module performs supplementary evaluation. This module transforms the dynamic knowledge graph into a Bayesian network: each key entity in the graph (such as "ventilation system status", "reagent evaporation rate", "regional gas concentration") is modeled as a random variable node, whose possible states include "normal", "abnormal", "fault", etc.; the relationships between entities are transformed into conditional dependency edges, representing the probabilistic influence of one node's state on the state of another node.

[0042] The system injects evidence into some observable nodes through real-time observation data (such as sensor readings and equipment switching signals), and then performs Bayesian inference to calculate the posterior probability of unobservable but critical risk nodes (such as the "probability of poisoning of experiment personnel"). This inference process uses approximate inference algorithms (such as belief propagation or Markov chain Monte Carlo sampling), and can achieve a response time in seconds with limited computing resources.

[0043] ; The above equation is the standard form of Bayes' theorem, where Indicates the status of the risk node to be assessed. This represents the set of all observed evidence. The system uses this formula to calculate the conditional probability that a risk node is in a dangerous state given the current observed evidence. This posterior probability is then directly input into the decision generation module as a quantitative indicator.

[0044] The decision generation module integrates the outputs of the risk pattern recognition engine (risk type, confidence level) and the probabilistic graphical reasoning module (posterior probability of key risks) to generate specific control instructions based on a pre-set multi-level decision matrix. This decision matrix uses two dimensions: risk probability and consequence severity, each divided into five levels (1 to 5). The probability level is calculated by weighting the confidence level and the posterior probability; the consequence severity level is predefined according to the risk type. For example, an event involving a highly toxic chemical leak is assigned a level 5 (catastrophic), while a common equipment overheating event is assigned a level 2 (minor).

[0045] The matrix intersection defines a three-tiered response strategy: Level 1 only generates system logs and operation prompts; Level 2 generates visual early warning information and pushes handling suggestions to relevant personnel via mobile application; Level 3 immediately triggers automatic control commands and notifies safety management personnel to intervene. For example, when the system identifies a risk pattern where "incompatible chemicals (such as strong oxidants and reducing agents) are stored in adjacent cabinets, and the ambient temperature exceeds 30 degrees Celsius," and the confidence level is above 90%, the system determines it to be a Level 3 high-risk scenario, immediately locks the relevant storage cabinets, activates local exhaust ventilation, and sends an emergency alarm to the laboratory director.

[0046] The execution and feedback layer is responsible for implementing the aforementioned decision-making instructions and feeding back the execution results to the system, forming a closed loop. The information presentation unit displays differentiated information to different user roles through LCD screens deployed on the laboratory walls, desktop terminals, and mobile applications: laboratory personnel can see real-time risk warnings and safety guidance related to their current operation; managers can view a global safety situation heatmap, a list of risk events, and historical trend analysis. The auxiliary control unit connects to the actuators in the laboratory, including emergency ventilation fans, power relays, and electromagnetic locks, through standard industrial control interfaces (such as Modbus TCP and OPCUA). When a Level 3 control instruction is received, this unit completes instruction parsing and execution within 100 milliseconds and returns an execution confirmation status. The feedback acquisition unit is not independent hardware but reuses the sensor network of the IoT sensing layer to continuously monitor changes in environmental parameters after execution (such as the rate of VOCs concentration decrease after ventilation is started and the temperature decay curve of equipment surface after power failure), and uses this feedback data as new sensing input, re-entering the system analysis process. This closed-loop mechanism is attached. Figure 4 As shown, this ensures that every intervention can be validated, learned, and optimized by the system.

[0047] Furthermore, the system operates within a multi-timescale parallel analysis framework, as shown in the attached diagram. Figure 5As shown, the framework defines three analytical scales: a real-time streaming scale for driving rapid matching of edge alarms and risk patterns; a short-cycle rolling analysis scale (5-minute window, 1-minute sliding step) for aggregating historical data to support probabilistic graphical reasoning and situation assessment; and a long-cycle trend analysis scale (7-day window) for uncovering periodic patterns in security events, equipment aging trends, and personnel behavior patterns. The analysis results from these three scales are weighted and fused in the decision generation module. For example, if long-cycle analysis reveals a significant increase in the failure rate of a device every Friday afternoon, the assessment weight of its related risks during that period will automatically increase, thereby enhancing the foresight and adaptability of the decision.

[0048] Furthermore, the system possesses self-evolution capabilities. The case learning module automatically captures the entire process data of each security incident, from occurrence, identification, decision-making to completion of handling, including the initial map state, triggered risk patterns, executed instructions, and final result state. This module employs a difference comparison algorithm to compare the features of successful handling cases with failed or novel anomaly cases, extracting new risk association rules (such as "if device A experiences a current fluctuation standard deviation > 2 amps after > 4 hours of continuous operation at > 80% humidity, it indicates an impending failure"), or correcting the confidence weights of existing patterns. All newly extracted features must be reviewed and confirmed online by security experts before being incrementally updated to the risk pattern feature library. Simultaneously, the conditional probability table of the probabilistic graphical inference module is also optimized weekly: using closed-loop feedback data accumulated over the past 7 days, the maximum likelihood estimation method is used to recalculate the state transition probabilities of each node, ensuring that the model parameters continuously approximate the statistical laws of the real physical world.

[0049] In summary, this embodiment, through the deep integration of IoT sensing, edge computing, knowledge graphs, probabilistic reasoning, and automatic control technologies, constructs a laboratory safety management system capable of real-time sensing, dynamic evaluation, intelligent decision-making, and closed-loop execution. This system not only solves the problems of data silos, delayed response, and rigid models found in traditional solutions, but also achieves continuous evolution of safety management capabilities through a self-learning mechanism, providing modern scientific research laboratories with comprehensive, end-to-end, and all-element dynamic safety assurance.

[0050] Example 2: Based on Example 1 above, this example further refines the implementation mechanism of the entity relationship extraction module in the data fusion and knowledge construction layer, and introduces a dynamic relationship confidence calculation method based on attention mechanism to improve the accuracy and robustness of knowledge graph construction.

[0051] After receiving the spatiotemporally aligned basic fact data tuples, the entity relationship extraction module does not simply perform hard matching based on preset rules, but instead adopts a soft matching strategy. Specifically, for each potential entity pair (such as "personnel-equipment"), the system first calculates an initial association score based on their spatiotemporal co-occurrence (i.e., both appear in the same spatial region within the same time window). Subsequently, the system calls a predefined relationship triggering rule base. Each rule contains a set of conditional predicates (such as "equipment type = autoclave", "personnel role = student", "operation duration > 10 minutes"). When all conditions are met, the rule is activated and contributes an increment to the basic confidence level for its corresponding relationship type (such as "violation operation").

[0052] However, the reliability of different rules varies across different contexts. Therefore, this embodiment introduces an attention weighting mechanism. The system maintains a rule-context attention matrix, where rows represent rules and columns represent contextual features (such as time period, season, and laboratory usage density). Whenever a rule is activated, the system queries this matrix based on the current contextual features to obtain the corresponding attention weight. And weight the base confidence level: ; in For the first The base confidence level of the activation rule, For its attention weight, The sigmoid function maps the linearly weighted result to the interval between 0 and 1. This attention weight matrix is ​​dynamically updated through an online learning mechanism: after each manual review and confirmation or correction of a relationship, the system uses a backpropagation algorithm to adjust the relevant weights, so that rules that can more accurately predict the authenticity of relationships in similar contexts receive higher weights.

[0053] Furthermore, the graph dynamic update module employs a batch merging strategy to improve performance when handling high-frequency update scenarios. When a large number of attribute update requests for the same entity are received within a short period (e.g., within 1 second), the module does not perform database writes one by one. Instead, it first caches these updates in memory and merges conflicting attributes in chronological order (e.g., taking the latest valid value for temperature and appending the operation record to the end of the list), and then commits them to the graph database in a single transaction. This strategy significantly reduces the database I / O load, ensuring that the graph maintains sub-second update latency even when high-concurrency data influx occurs.

[0054] In the dynamic risk assessment and decision-making layer, this embodiment optimizes the structure learning process of the probabilistic graphical inference module. Traditional constraint-based methods rely on rigorous conditional independence tests, but in real-world laboratory environments, data noise and missing data can lead to erroneous topological inferences. To address this, the system introduces a hybrid structure learning strategy: First, based on prior knowledge provided by experts (e.g., "the state of the ventilation system directly affects the regional gas concentration"), a minimum necessary dependency skeleton is constructed; then, based on this skeleton, local structural expansion is performed using historical data, allowing only new edges that statistically significantly improve the model fit. This strategy ensures both the physical interpretability of the model and retains the ability to discover new correlations from the data.

[0055] At the execution and feedback layer, this embodiment enhances the fault tolerance capability of the auxiliary control unit. When the system issues an automatic control command (such as power off), if no execution confirmation feedback is received within a preset timeout period (such as 2 seconds), or if the feedback status is inconsistent with expectations (such as the command to power off but the current is still higher than the threshold), the system will immediately escalate the event level and trigger a backup control path (such as switching to another relay control), while simultaneously generating a device fault diagnosis task and pushing it to the maintenance personnel's terminal. This mechanism ensures that the system can still maintain basic safety intervention capabilities even in the event of partial failure of the actuator.

[0056] Through the aforementioned enhancement mechanisms, this embodiment significantly improves the accuracy of knowledge graph construction, the robustness of the reasoning model, and the reliability of the execution process while maintaining the core system architecture, further strengthening the system's security control performance in complex, highly dynamic laboratory environments.

Claims

1. A dynamic management system for laboratory safety throughout the entire process based on the Internet of Things, characterized in that: include: The IoT sensing layer is deployed within the laboratory's physical space to collect multi-dimensional raw data on safety status in real time. The edge computing layer consists of several edge computing gateways deployed at the edge of the laboratory network, used to perform localized preprocessing and preliminary analysis of the raw sensing data received from the IoT sensing layer. The data fusion and knowledge building layer is used to receive and fuse standardized multi-source heterogeneous data streams from all edge computing gateways and build a dynamically evolving laboratory safety knowledge graph. The dynamic risk assessment and decision-making layer is used to conduct real-time security risk assessment and generate hierarchical control decisions based on the dynamic knowledge graph output by the data fusion and knowledge construction layer. The execution and feedback layer is used to receive and execute the control instructions issued by the dynamic risk assessment and decision-making layer, and to feed back the execution results and changes in environmental status to the system.

2. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 1, characterized in that, The IoT sensing layer consists of sensing nodes and identification nodes; The sensing nodes include environmental monitoring sensors, equipment status sensors, and hazardous materials status sensors; the environmental monitoring sensors are used to collect data on temperature, humidity, light intensity, specific gas concentration, and smoke concentration in the laboratory. The device status sensor is used to collect real-time operating current, voltage, power, running time and surface temperature data of high-power instruments and equipment; The hazardous materials status sensor is used to collect data on the internal temperature, volatile organic compound concentration, and cabinet door opening / closing status inside the hazardous chemical storage cabinet. The identification node includes a personnel identification unit and an item identification unit; The personnel identification unit is used to acquire, in real time, the identification information and spatiotemporal location of personnel entering the laboratory or operating specific equipment through biometric identification or radio frequency identification technology. The item identification unit is used to obtain the unique identification information, access records and current location information of experimental reagents, consumables and hazardous chemicals in real time by scanning radio frequency identification tags or QR codes.

3. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 1, characterized in that, The preprocessing includes data cleaning, format standardization, and timestamp alignment. The preliminary analysis includes real-time comparison of data based on preset single-index thresholds. When any sensor data exceeds its safety threshold range, the edge computing gateway immediately generates and reports a level-one local alarm event, while continuously uploading the preprocessed standardized data stream.

4. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 1, characterized in that, The data fusion and knowledge construction layer includes a spatiotemporal alignment module, an entity relationship extraction module, and a graph dynamic update module. The spatiotemporal alignment module is used to align all incoming perception data, personnel identity data, and item identification data into the same spatiotemporal framework based on a unified Coordinated Universal Time (UTC) time reference and a laboratory three-dimensional spatial coordinate model, forming a basic fact data tuple with a unified spatiotemporal label. The entity relationship extraction module is used to parse the basic fact data tuples based on a preset entity type and relationship type ontology library, automatically identify the entity instances involved in the data tuples, and create or update the association relationships between entities in real time according to the data semantics. The graph dynamic update module is used to maintain a laboratory safety knowledge graph stored in a graph database, and continuously receive the output of the entity relationship extraction module to update the relationship strength of node attributes and edges in the knowledge graph in an incremental manner in real time, and record the precise timestamp of each change.

5. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 4, characterized in that, The dynamic risk assessment and decision-making layer includes a risk pattern recognition engine, a probabilistic graphical reasoning module, and a decision generation module. The risk pattern recognition engine has a built-in risk pattern feature library trained from historical accident cases and expert rules. It is used to continuously perform subgraph matching and pattern scanning on the dynamic knowledge graph, identify whether there are topological structures and attribute combinations that match known risk patterns in the feature library, and output the corresponding risk type identifier and confidence level. The probabilistic graphical reasoning module is used to evaluate potential associated risks not covered by the risk pattern feature library. This module transforms the dynamic knowledge graph into a probabilistic graphical model, where the state of the nodes in the graph is a random variable and the edges represent the conditional dependencies between the variables. The probabilistic graph reasoning module injects evidence into the state of some nodes through the received real-time observation data, and based on the Bayesian network reasoning algorithm, propagates the probabilistic influence along the topology of the graph to calculate the posterior probability that the key risk nodes are in a dangerous state. The decision generation module is used to comprehensively receive the risk identifier and confidence level output by the risk pattern recognition engine, as well as the key risk posterior probability output by the probabilistic graph reasoning module, and generate specific control instructions based on the preset multi-level decision matrix.

6. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 5, characterized in that, The execution and feedback layer includes an information presentation unit, an auxiliary control unit, and a feedback acquisition unit; The information presentation unit is used to release safety warning information, risk analysis reports and operational guidance suggestions of different levels to managers and experimental personnel in real time through display terminals and mobile applications in the laboratory. The auxiliary control unit is used to send direct control commands to the environmental control equipment and safety facilities in the laboratory through an industrial control interface; The feedback acquisition unit is used to continuously monitor the changes in relevant environmental parameters and device status after the execution of instructions through the IoT sensing layer, and input these data as new sensing data into the system.

7. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 6, characterized in that, The entity relationship extraction module includes preset entity types such as personnel, experimental equipment, chemical reagents, experimental operations, and environmental areas; and preset relationship types such as personnel located in an area, personnel operating equipment, equipment using reagents, reagents stored in an area, and equipment generating environmental parameters.

8. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 7, characterized in that, The structure learning and parameter learning processes of the probabilistic graphical model in the probabilistic graphical reasoning module are as follows: Structure learning is based on historical security data, adopts the constraint-based method, and uses the conditional independence test between variables in the data to initially determine the network topology of the dependency relationship between nodes; Parameter learning, on the basis of the fixed network topology, adopts the maximum likelihood estimation method to learn the conditional probability table of node state transitions from historical data.

9. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 8, characterized in that, The system sets a model optimization cycle, periodically using newly accumulated closed-loop feedback data to re-estimate and optimize the conditional probability table.

10. The IoT-based dynamic management system for laboratory safety throughout the entire process according to claim 5, characterized in that, The multi-level decision matrix in the decision generation module is constructed based on the two-dimensional assessment results of the probability and severity of the risk. The probability dimension is determined by the risk pattern confidence and posterior probability. The severity dimension is graded and assigned values ​​according to the potential degree of personal injury, property loss and environmental damage associated with the risk type. The decision matrix divides the probability and severity into 5 levels and defines 3-level response strategies corresponding to different combinations of the two.

Images