Safety control method and system for special gas cabinet based on multi-parameter monitoring

Abstract

The application discloses a special gas cabinet safety control method and system based on multi-parameter monitoring, which collects multi-dimensional data such as gas concentration, environmental parameters and equipment state inside and outside the special gas cabinet through a distributed sensing network, constructs a dynamic risk portrait in combination with a process situation after adaptive filtering and time sequence feature extraction. An LSTM-VAE hybrid model, a gas reaction knowledge graph, an improved D-S evidence theory and a pipeline topology propagation model are used to respectively calculate an abnormal score, a gas interaction risk score and a spatial risk score, a comprehensive risk value is generated through dynamic weighted fusion, a risk level is divided, and a hierarchical control strategy is triggered. When a risk anomaly occurs, a leakage source positioning and gas mixing tracing are automatically started, and a diagnosis report is generated. The system realizes closed-loop management of monitoring, early warning, response and tracing, integrates data-driven and knowledge-driven methods, and improves risk identification accuracy and intelligent level of disposal.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial safety, and in particular to a safety control method and system for special gas holders based on multi-parameter monitoring. Background Technology

[0002] In high-end manufacturing fields such as semiconductors, photovoltaics, and display panels, the safety of special gas cabinets, as the core unit for special gas supply, is paramount. Traditional special gas cabinet safety monitoring mainly relies on single gas concentration sensors to trigger threshold alarms, lacking the coordinated perception of multiple parameters such as oxygen concentration, ambient temperature and humidity, and equipment status, making it difficult to identify hidden leaks or early faults under complex operating conditions. Furthermore, existing systems generally employ fixed alarm thresholds and static monitoring strategies, failing to dynamically adjust sensitivity according to process stages, leading to frequent false alarms and missed alarms. When a leak occurs, only a local alarm is provided, lacking the ability to locate the leak source, assess gas mixing risks, and predict propagation paths, making it difficult to support rapid emergency response. In recent years, although some research has introduced machine learning models for anomaly detection, these are mostly isolated models, not integrated with physical knowledge such as gas chemical properties and pipeline topology, resulting in a single dimension of risk assessment and limited intelligence. Therefore, there is an urgent need for a special gas cabinet safety control method that can achieve multi-parameter fusion perception, dynamic risk assessment, hierarchical control, and intelligent traceability to improve the safety control capabilities of special gas cabinets.

[0003] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a special gas holder safety control method and system based on multi-parameter monitoring, which solves the problem that the existing special gas holder safety monitoring system is difficult to achieve early warning and intelligent response due to its reliance on a single parameter, fixed threshold and isolated alarm.

[0004] To achieve the above and other related objectives, the present invention provides the following technical solution: The safety control method for special gas holders based on multi-parameter monitoring includes: S1, real-time monitoring of the target special gas concentration, oxygen concentration, environmental parameters, physical state parameters, and safety equipment status parameters inside and around the special gas holder; S2, preprocessing the collected monitoring data and extracting time-series features; S3, based on preset gas safety properties and the pipeline connection topology of the special gas holder, using an intelligent assessment model to perform fusion analysis of multi-dimensional parameters and output a risk level, wherein the intelligent assessment model includes an anomaly detection model, a gas interaction assessment model, and a pipeline topology risk propagation model; S4, executing corresponding graded safety control strategies according to the risk level, and when the risk level is abnormal, initiating the leak source location and gas mixing source tracing process and generating a diagnostic report.

[0005] Furthermore, S1 uses a distributed multi-source sensor network to collaboratively collect multi-dimensional operational and environmental data of the special gas holder's interior and surrounding areas. The data acquisition process is dynamically controlled by an adaptive monitoring engine based on real-time process conditions. Specifically, this includes: analyzing the current process formulation, extracting the physicochemical properties of the gases involved, constructing a risk weight vector associated with each safety parameter, and generating a dynamic risk profile; based on the risk weight vector and a preset benchmark monitoring strategy, dynamically generating a final monitoring strategy through a parameterized adjustment algorithm, which includes at least a nonlinear mapping of sampling frequency, computational resource priority, and alarm threshold; and, according to the generated final monitoring strategy, controlling the distributed multi-source sensor network to collaboratively collect multi-dimensional operational and environmental data of the special gas holder's interior and surrounding areas, and dynamically enabling or disabling specific composite alarm logic based on the dynamic risk profile.

[0006] Further, S2 includes: preprocessing the original monitoring data using an adaptive filtering algorithm based on process stage perception; extracting multi-scale time-series features from the data characteristics of different safety parameters to form an initial feature set; weighting and filtering the initial feature set based on the risk weight vector in the risk profile to generate a key feature subset; and inputting the key feature subset into a feature fusion algorithm for nonlinear transformation and integration to generate a fused feature set for risk assessment.

[0007] Further, S3 includes: S31, the anomaly detection model adopts a hybrid model based on LSTM network and VAE encoder, performs time-series modeling on the fusion feature set generated in S2, calculates the reconstruction error and prediction bias of the current state respectively, and generates a weighted first anomaly score; S32, the gas interaction assessment model constructs a gas reaction knowledge graph with gas type, concentration range, and environmental parameters as nodes, and based on this graph and real-time data, uses an improved DS evidence theory to fuse multi-source evidence to calculate whether gas mixing causes combustion, explosion, or the generation of toxic byproducts. The second risk score for the object; S33, the pipeline topology risk propagation model abstracts the special gas holder pipeline system into a directed graph G=(V,E), where nodes V represent valves, joints, and sensors, and edges E represent pipeline connections; based on the random walk algorithm, the propagation process of abnormal signals or leakage risks in the directed graph is simulated to generate a third risk score characterizing the spatial risk distribution; S34, the first abnormal score, the second risk score, and the third risk score are weighted and fused to generate a comprehensive risk value, and the corresponding comprehensive risk level is determined according to the preset threshold range to which the comprehensive risk value belongs.

[0008] Further, S31 includes: using a time window [ ,..., Using historical fusion features as input, an LSTM network is used to predict the feature value at the current time. And calculate the prediction bias. ,in The current time's actual observation feature vector; the current time's fused features As input, it is mapped to the latent space using a VAE encoder, and then reconstructed by a decoder. And calculate the reconstruction error = The prediction deviation and reconstruction error via normalization function and Processing yields dimensionless standardized anomaly confidence scores; weights are then assigned to the standardized anomaly confidence scores. and The weighted sum is calculated and then input into the final transformation function. Generate the first anomaly score The calculation formula is as follows: in and Let be the weighting coefficients of the anomaly detection sub-model, and satisfy . .

[0009] Further, S32 includes: constructing a gas reaction knowledge graph with gas type, concentration range, and environmental parameters as nodes and chemical reaction relationships as edges; matching real-time monitored multi-gas concentration data and environmental parameters with the gas reaction knowledge graph to generate a set of evidence that is activated in real time, wherein each piece of evidence corresponds to a triggered risk condition in the graph; for each piece of evidence in the evidence set, based on its monitored value, it is assigned a membership function that conforms to its physicochemical characteristics. Dynamically calculate its basic probability allocation for the target risk proposition: ,

[0010] in, As evidence The monitored values; To identify the frame, represent the union of all possible risks; Indicates the first The degree of support or trust in a particular set of threats from each source of evidence; For each piece of evidence Assign a dynamic credibility weight , , For sensor reliability, For operating condition effectiveness, For context consistency, , , These are the normalized weighting coefficients; using The basic probability assignment is modified. , All revised basic probability assignments are fused using DS evidence theory. in, This indicates that after integrating all the evidence, the proposition is... The basic probability distribution; K is the normalization factor; Represents all subpropositions The intersection equals the proposition ; Represented as the first Each source of evidence pairs the proposition The corrected basic probability assignment; This represents the sum of probabilities when all combinations of evidence are disjoint; Represented as all subpropositions The intersection of these sets is an empty set; The fused basic probability allocation Projected into the risk space, a second risk score is calculated to determine if gas mixing will lead to combustion, explosion, or the generation of toxic byproducts. The specific formula is as follows: The The normalized risk value for the interval [0,1].

[0011] Further, S33 includes: modeling the special gas holder piping system as a directed graph G=(V,E), where the node set V represents valves, joints, and sensors, and the edge set E represents the piping connection direction; initializing the risk confidence of each node based on real-time monitoring data, and using a random walk algorithm to simulate the propagation process of abnormal signals or leakage risks in the directed graph to obtain the final risk value of each node. ,in , Given the total number of nodes; calculate the maximum value among the final risk values ​​of all nodes. and arithmetic mean ; Introducing distance decay factor , This represents the typical distance from the node to the gas source; This refers to the distance attenuation rate. The third risk score is finally generated using the following formula. ; in, This is the third risk assessment score; This is the global gain coefficient; This is a risk weighting factor.

[0012] Furthermore, S34 includes: scoring the first anomaly. Second risk score and third risk score Context-aware dynamic weighted fusion is performed to generate a comprehensive risk value. ; ( ) in, , , Is with , , Corresponding weighting coefficients; For the first Individual risk dynamic weights; For the first Risk score The static basic preference coefficient; For the first A sensitivity function for each risk; Normalization factor; This represents the input parameters or state variables related to the safety control of special gas holders; it also includes the comprehensive risk value. With respect to the preset level threshold range , , Compare; if Then it is a security level; if Then it is ;if Then it is ;if Then it is .

[0013] Furthermore, the graded safety control strategy in S4 includes: if the risk level is at the attention level, a local low-frequency warning is activated and monitoring is enhanced; if the risk level is at the warning level, a remote alarm is activated and the exhaust system is activated to maximum power; if the risk level is at the danger level, the emergency shut-off valve is immediately activated to close the gas source and the exhaust gas treatment device is activated. The leak source location and gas mixing tracing process includes: when the comprehensive risk level output by S3 is at the attention level, warning level, or danger level, the leak source location and gas mixing tracing process is activated; wherein, the leak source location is based on the pipeline topology of the special gas holder and the spatiotemporal concentration data collected by multiple sensors, and the candidate nodes for the leak are determined by the time difference inversion method and the graph neural network model; the gas mixing tracing is based on a preset gas interaction knowledge graph, identifies gas convergence paths with reaction risks, and traces upstream pollution sources or cross-contamination points; a structured diagnostic report is generated based on the location results and tracing paths, and the diagnostic report includes at least an anomaly summary, risk propagation path, leak source confidence, cause analysis, and operation and maintenance disposal recommendations.

[0014] In another embodiment of the present invention, a special gas cabinet safety control system based on multi-parameter monitoring is provided. The system includes: a distributed multi-source sensor network for real-time monitoring of the target special gas concentration, oxygen concentration, environmental parameters, physical state parameters, and safety equipment status parameters within the special gas cabinet; a feature extraction module for preprocessing the collected monitoring data and extracting time-series features to generate a fused feature set; and an intelligent assessment and risk fusion module for performing multi-dimensional risk assessment on the fused feature set, which integrates: an anomaly detection sub-model, employing a hybrid neural network structure based on an LSTM network and a VAE encoder, for performing time-series modeling of the fused feature set and outputting a first anomaly score; and a gas interaction assessment sub-model, with a built-in gas reaction knowledge graph, combined with an improved DS evidence theory to assess the risk of multi-gas coexistence. The system integrates several risk assessment mechanisms: a fusion reasoning module to output a second risk score; a pipeline topology risk propagation sub-model to model the special gas holder pipeline system as a directed graph with attributes, simulating the spatial propagation process of leakage risk through a random walk algorithm, and outputting a third risk score; a risk fusion model to weightedly fuse the first, second, and third risk scores to generate a comprehensive risk value, which is then mapped to a comprehensive risk level; a hierarchical control module to execute corresponding hierarchical safety control strategies based on the comprehensive risk level; a leakage tracing and diagnostic report generation module to initiate the leakage source location and gas mixing tracing process and generate a diagnostic report when the comprehensive risk level is at the attention, warning, or danger level; and a central computing and communication unit to coordinate data and control flows between modules, schedule task execution, and provide a human-machine interface.

[0015] As described above, the special gas cabinet safety control method and system based on multi-parameter monitoring of the present invention has the following beneficial effects: First, by constructing a distributed multi-source sensor network, comprehensive real-time monitoring of the target special gas concentration, oxygen content, environmental parameters, physical state, and safety equipment status in and around the special gas cabinet is achieved, breaking through the limitations of traditional systems that rely solely on single gas concentration alarms, and significantly improving the comprehensiveness of anomaly detection and early warning capabilities. Second, the introduction of an adaptive filtering and dynamic risk profiling mechanism based on process stage perception enables the monitoring strategy to automatically adjust the sampling frequency, resource priority, and alarm threshold according to the process formula, avoiding false alarms or missed alarms caused by fixed parameters, and improving the system's adaptability and stability under varying operating conditions. Furthermore, through the synergy of a hybrid model of LSTM network and VAE encoder, gas reaction knowledge graph and improved DS evidence theory, and pipeline topology risk propagation model, risk quantification is performed from three dimensions: temporal anomaly, chemical reaction risk, and spatial propagation path, significantly improving the accuracy and interpretability of risk assessment. The dynamic weighted fusion mechanism of comprehensive risk value can automatically adjust the weight of each risk item according to the context, ensuring that the judgment criteria are scientific and reasonable. When the risk level rises, the system executes a tiered control strategy, achieving a step-by-step response from early warning and ventilation to emergency shutdown, ensuring a balance between safety and production. After an anomaly occurs, it automatically initiates the leak source location and gas mixing tracing process, generating a structured diagnostic report, significantly shortening troubleshooting time and improving operational efficiency. The overall solution features closed-loop management, clear logic, and strong implementability, making it suitable for high-risk process scenarios such as semiconductors and photovoltaics, comprehensively enhancing the intrinsic safety level and intelligent control capabilities of special gas cabinets. Attached Figure Description

[0016] Figure 1 The diagram shown is a flowchart of the method of the present invention. Detailed Implementation

[0017] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be noted that, unless otherwise specified, the following embodiments and features described herein can be combined with each other.

[0018] Example 1

[0019] like Figure 1As shown, this invention provides a safety control method for special gas holders based on multi-parameter monitoring, including: S1, real-time monitoring of the target special gas concentration, oxygen concentration, environmental parameters, physical state parameters, and safety equipment status parameters inside and around the special gas holder; S2, preprocessing the collected monitoring data and extracting time-series features; S3, based on preset gas safety properties and the pipeline connection topology of the special gas holder, using an intelligent assessment model to perform fusion analysis of multi-dimensional parameters and output a risk level, wherein the intelligent assessment model includes an anomaly detection model, a gas interaction assessment model, and a pipeline topology risk propagation model; S4, executing corresponding graded safety control strategies according to the risk level, and when the risk level is abnormal, initiating the leak source location and gas mixing source tracing process and generating a diagnostic report.

[0020] This invention deploys a distributed multi-source sensor network within and around a special gas holder to collect real-time multi-dimensional data, including the concentration of the target special gas, oxygen concentration, environmental parameters, physical state, and the status of safety equipment. After preprocessing and extracting temporal features from the collected data, an intelligent assessment model is introduced that integrates anomaly detection, gas reaction risk assessment, and pipeline topology propagation analysis. This model comprehensively considers the safety properties of the gas and the pipeline connection structure, performing deep fusion analysis of multi-source information and outputting a graded risk assessment result. When the risk level reaches a preset threshold, the system automatically executes the corresponding graded control strategy and, under abnormal conditions, initiates a leak source location and gas mixing tracing mechanism. Combining spatiotemporal concentration data and knowledge graph reasoning, a diagnostic report containing causal analysis and treatment recommendations is generated, achieving closed-loop safety control from monitoring and early warning to source tracing.

[0021] This invention significantly improves the accuracy and timeliness of safety risk identification and response in special gas cabinets through multi-parameter collaborative monitoring and intelligent fusion assessment. Compared to traditional single-gas concentration alarm methods, this solution comprehensively considers gas interactions, equipment status, and pipeline topology, effectively reducing false alarm and missed alarm rates. Through dynamic risk profiling and context-aware hierarchical control strategies, it achieves adaptive safety management for process scenarios. After a risk occurs, it proactively initiates leak source location and gas mixing tracing processes, not only quickly curbing the spread of the accident but also providing structured diagnostic reports to assist maintenance personnel in accurately identifying potential hazards. This significantly enhances the intelligent management level and intrinsic safety capabilities of special gas systems in high-risk process scenarios such as semiconductors and photovoltaics.

[0022] S1 uses a distributed multi-source sensor network to collaboratively collect multi-dimensional operational and environmental data of the special gas holder's interior and surrounding areas. The data acquisition process is dynamically controlled by an adaptive monitoring engine based on real-time process conditions. Specifically, this includes: analyzing the current process formulation, extracting the physicochemical properties of the gases involved, constructing a risk weight vector associated with various safety parameters, and generating a dynamic risk profile; based on the risk weight vector and a preset benchmark monitoring strategy, dynamically generating a final monitoring strategy through a parameterized adjustment algorithm, which includes at least a nonlinear mapping of sampling frequency, computational resource priority, and alarm thresholds; and, according to the generated final monitoring strategy, controlling the distributed multi-source sensor network to collaboratively collect multi-dimensional operational and environmental data of the special gas holder's interior and surrounding areas, and dynamically enabling or disabling specific composite alarm logic based on the dynamic risk profile.

[0023] In step S1, this invention introduces an adaptive monitoring mechanism based on real-time process conditions, collecting multi-dimensional operational and environmental data of the special gas holder's interior and surrounding areas through a distributed multi-source sensor network. This monitoring process is dynamically controlled by an adaptive monitoring engine: first, it analyzes the currently executed process formulation, identifies the types of special gases involved and their physicochemical properties (such as toxicity, flammability, and diffusion rate), and, combined with equipment status and environmental conditions, constructs a risk weight vector reflecting the relative importance of each safety parameter, forming a dynamic risk profile; then, this risk profile is used as input to apply a preset baseline monitoring strategy—that is, a set of standardized monitoring configurations defaulted to by the system (including basic sampling frequency, computational resource allocation priority, and initial alarm threshold); through a parameterized adjustment algorithm, the above parameters are nonlinearly mapped and weighted to generate a final monitoring strategy adapted to the current operating conditions; finally, based on this final strategy, the collaborative acquisition behavior of the sensor network is controlled, and specific composite alarm logic is enabled or disabled in real time according to the dynamic risk profile, achieving precise matching between monitoring intensity and risk level.

[0024] S2 includes: preprocessing the original monitoring data using an adaptive filtering algorithm based on process stage perception; extracting multi-scale time series features from the data characteristics of different safety parameters to form an initial feature set; weighting and filtering the initial feature set based on the risk weight vector in the risk profile to generate a key feature subset; and inputting the key feature subset into a feature fusion algorithm for nonlinear transformation and integration to generate a fused feature set for risk assessment.

[0025] First, an adaptive filtering algorithm based on process stage awareness is employed. The filtering parameters are dynamically adjusted according to the current stage of the process formulation (e.g., purging, deposition, exhaust), effectively removing noise interference under different operating conditions while retaining the true and valid signal components. Then, for different types of safety parameters such as target gas concentration, oxygen content, temperature, and pressure, multi-scale time-series features (e.g., mean, variance, trend term, wavelet coefficients, frequency domain energy, etc.) are extracted based on their dynamic time-series characteristics (e.g., abrupt changes, periodicity, drift) to construct an initial feature set containing multi-dimensional dynamic information. Further, the risk weight vector contained in the dynamic risk profile is used to weight and score each feature item in the initial feature set, and key features highly correlated with the current risk are selected to form a key feature subset. Finally, this subset is input into a nonlinear feature fusion algorithm (e.g., kernel method, deep autoencoder, or attention mechanism) for high-dimensional spatial mapping and information integration, generating a compact and discriminative fusion feature set, providing high-quality input for subsequent risk level assessment.

[0026] S3 includes: S31, the anomaly detection model adopts a hybrid model based on LSTM network and VAE encoder, performs time-series modeling on the fusion feature set generated in S2, calculates the reconstruction error and prediction deviation of the current state respectively, and generates a first anomaly score by weighting; S32, the gas interaction assessment model constructs a gas reaction knowledge graph with gas type, concentration range, and environmental parameters as nodes, and based on this graph and real-time data, uses an improved DS evidence theory to fuse multi-source evidence to calculate a second risk score for gas mixing causing combustion, explosion or generation of toxic byproducts; S33, the pipeline topology risk propagation model abstracts the special gas cabinet pipeline system into a directed graph G=(V,E), where nodes V represent valves, joints, and sensors, and edges E represent pipeline connections; based on a random walk algorithm, it simulates the propagation process of abnormal signals or leakage risks in the directed graph to generate a third risk score characterizing the spatial risk distribution; S34, the first anomaly score, the second risk score, and the third risk score are weighted and fused to generate a comprehensive risk value, and the corresponding comprehensive risk level is determined according to the preset threshold range to which the comprehensive risk value belongs.

[0027] In step S3, this invention constructs an intelligent assessment model consisting of three parts: anomaly detection, gas reaction assessment, and pipeline topology propagation, enabling multi-dimensional fusion analysis of the safety status of special gas holders. First, in S31, a hybrid model based on an LSTM network and a variational autoencoder (VAE) is used to perform temporal modeling on the fusion feature set generated in S2: the LSTM is used to predict the feature value at the current moment and calculate the prediction bias, while the VAE is used to reconstruct the input and calculate the reconstruction error. After normalization and weighted fusion, a first anomaly score reflecting the degree of dynamic anomaly in the system is generated. In S32, a gas reaction knowledge graph is constructed with gas type, concentration range, and environmental parameters as nodes. Real-time monitoring data is matched with the graph to activate potential reaction paths. An improved DS evidence theory, incorporating dynamic weights such as sensor credibility and operating context, is used to fuse multi-source evidence, quantify the risk of gas mixing leading to combustion, explosion, or the generation of toxic byproducts, and output a second risk score. In step S33, the special gas holder piping system is modeled as a directed graph G=(V,E), where nodes V represent critical components and edges E represent connections. A random walk algorithm is used to simulate the propagation of leakage risk along the pipeline, generating a third risk score that characterizes the spatial risk distribution. Finally, in step S34, the three scores are dynamically weighted and fused using context awareness to generate a comprehensive risk value. This value is then mapped to the corresponding comprehensive risk level based on a preset threshold range, achieving unified quantification and hierarchical identification of multi-source heterogeneous risks.

[0028] S31 includes: using a time window [ ,..., Using historical fusion features as input, an LSTM network is used to predict the feature value at the current time. And calculate the prediction bias. ,in The current time's actual observation feature vector; the current time's fused features As input, it is mapped to the latent space using a VAE encoder, and then reconstructed by a decoder. And calculate the reconstruction error. = The prediction deviation and reconstruction error via normalization function and Processing yields dimensionless standardized anomaly confidence scores; weights are then assigned to the standardized anomaly confidence scores. and The weighted sum is calculated and then input into the final transformation function. Generate the first anomaly score The calculation formula is as follows: ,in and Let be the weighting coefficients of the anomaly detection sub-model, and satisfy . .

[0029] In step S31, a hybrid model combining an LSTM network and a VAE encoder is used to detect anomalies in the preprocessed fused feature data. First, the system uses historical fused feature data from a period prior to the current moment as input. The LSTM network predicts the feature value that should appear at the current moment, and this predicted value is compared with the actual observed value to calculate the prediction bias. Simultaneously, the actual feature value at the current moment is input into the VAE encoder, and data reconstruction is completed through compression and restoration processes. The reconstruction result is compared with the original data to obtain the reconstruction error. These two errors reflect anomalies in the system's time trend and data structure, respectively. To eliminate the incomparability caused by differences in dimensions and distributions, the system normalizes these two errors: for the prediction bias, linear scaling is performed using the minimum and maximum values ​​obtained from statistical analysis of historical normal operation data, ensuring it falls between 0 and 1; for the reconstruction error, a function with an S-shaped response is used for non-linear normalization. This function outputs smoothly when the error is small, and rapidly increases the output after the error increases to a certain threshold. After normalization, the system weights and sums the two normalized anomaly confidence scores according to pre-set weight ratios. The weight values ​​are configured based on actual operating conditions such as gas type and process stage to ensure a reasonable balance between different types of risks. Finally, the weighted result is input into a nonlinear transformation function. This function can be in the form of a hyperbolic tangent function (tanh) or a standard sigmoid function. These functions all have bounded outputs and smooth responses, capable of converting intermediate values ​​into the final first anomaly score. The score ranges from 0 to 1, with higher values ​​indicating a more severe system anomaly.

[0030] S32 includes: constructing a gas reaction knowledge graph with gas type, concentration range, and environmental parameters as nodes and chemical reaction relationships as edges; matching real-time monitored multi-gas concentration data and environmental parameters with the gas reaction knowledge graph to generate a set of evidence that is activated in real time, wherein each piece of evidence corresponds to a triggered risk condition in the graph; for each piece of evidence in the evidence set, based on its monitored value, it is assigned a membership function that conforms to its physicochemical characteristics. Dynamically calculate its basic probability allocation (BPA) for the target risk proposition: , ;in, As evidence The monitored values; To identify the frame, represent the union of all possible risks; Indicates the first The degree of support or trust in each source of evidence for this specific set of threats; for each piece of evidence Assign a dynamic credibility weight , , For sensor reliability, For operating condition effectiveness, For context consistency, , , These are the normalized weighting coefficients; using The basic probability assignment is modified. , All revised basic probability assignments are fused using DS evidence theory: in, This indicates that after integrating all the evidence, the proposition is... The basic probability distribution; K is the normalization factor; Represents all subpropositions The intersection equals the proposition ; Represented as the first Each source of evidence pairs the proposition The corrected basic probability assignment; This represents the sum of probabilities when all combinations of evidence are disjoint; Represented as all subpropositions The intersection of the two sets is an empty set; the basic probabilities after fusion are assigned. Projected into the risk space, a second risk score is calculated to determine if gas mixing will lead to combustion, explosion, or the generation of toxic byproducts. The specific formula is as follows: The The normalized risk value for the interval [0,1].

[0031] In step S32, a gas reaction knowledge graph is constructed to assess the risks of combustion, explosion, or the generation of toxic byproducts that may occur when multiple gases coexist. This knowledge graph uses gas types, concentration ranges, and environmental parameters (such as temperature and pressure) as nodes, and known chemical reaction relationships as connecting edges, forming a structured risk reasoning network. The system inputs real-time monitored multi-gas concentration data and current environmental parameters into this graph, automatically matching and activating potentially triggered risk conditions to generate a set of activated evidence. For each activated piece of evidence, the system calculates its support for the target risk proposition—the basic probability assignment—based on its monitored value using a specific membership function. This membership function can be selected in different forms depending on the gas characteristics; for example, an S-shaped ascending function is used for flammable gases, and a Gaussian function is used for toxic gases, ensuring that the risk support responds reasonably to changes in concentration. Simultaneously, the system assigns a dynamic credibility weight to each piece of evidence. This weight comprehensively considers the reliability of the sensor, the effectiveness of the current operating conditions, and the consistency of the data context, and ensures overall credibility balance through normalization processing. Subsequently, the system employs an improved DS evidence theory to fuse all the corrected basic probability assignments. The fusion process effectively handles conflicts and redundancies among multi-source information, ultimately outputting a second risk score that comprehensively reflects the risk of gas mixing.

[0032] S33 includes: modeling the special gas holder piping system as a directed graph G=(V,E), where the node set V represents valves, joints, and sensors, and the edge set E represents the piping connection direction; initializing the risk confidence of each node based on real-time monitoring data, and using a random walk algorithm to simulate the propagation process of abnormal signals or leakage risks in the directed graph to obtain the final risk value of each node. ,in , Given the total number of nodes; calculate the maximum value among the final risk values ​​of all nodes. and arithmetic mean Introducing a distance attenuation factor , This represents the typical distance from the node to the gas source; The distance attenuation rate is used as the final risk score, which is generated using the following formula. ; in, This is the third risk assessment score; This is the global gain coefficient; This is a risk weighting factor.

[0033] In step S33, the piping system of the special gas holder is first abstracted into a directed graph structure to simulate the propagation of leakage risk along the physical connection paths. Nodes in the graph represent key components such as valves, joints, and sensors, while edges represent the connection directions of the pipelines, forming a topological network reflecting the actual fluid pathways. The system initializes the risk confidence of each node based on real-time monitored data such as gas concentration and pressure changes, serving as the starting point for risk propagation. Subsequently, a random walk algorithm is used to simulate the diffusion process of abnormal signals or leakage risks in this directed graph. The algorithm calculates the probability of risk transferring from one node to an adjacent node based on pipeline connections and distances between nodes, and iteratively updates the risk value of each node. During propagation, a distance attenuation factor is introduced, meaning the risk gradually weakens as the propagation distance increases, simulating the physical characteristics of real gas diffusion. Finally, the system integrates the highest risk value, average risk level, and typical distance from the gas source for all nodes to generate a third risk score, which characterizes the spatial distribution breadth and concentration of leakage risk.

[0034] S34 includes: scoring the first anomaly. Second risk score and third risk score Context-aware dynamic weighted fusion is performed to generate a comprehensive risk value. ; ( )in, , , Is with , , Corresponding weighting coefficients; For the first Individual risk dynamic weights; For the first Risk score The static basic preference coefficient; For the first A sensitivity function for each risk; Normalization factor; This represents the input parameters or state variables related to the safety control of special gas holders; it also includes the comprehensive risk value. Compared with the preset level threshold range , , Compare; if Then it is a security level; if Then it is ;if Then it is ;if Then it is .

[0035] In step S34, the first anomaly score, second risk score, and third risk score are first merged to generate a comprehensive risk value for final risk level determination. The fusion process employs a context-aware dynamic weighting mechanism, meaning the weights of each score are not fixed but dynamically adjusted based on the current operating status, process stage, and environmental conditions of the special gas holder. The system pre-sets a basic preference coefficient for each risk score, reflecting its default importance in the overall safety assessment; simultaneously, a sensitivity function is introduced to perceive the response intensity of a certain risk type under the current operating conditions. For example, in the multi-gas mixing stage, the sensitivity to gas interaction risk automatically increases. The system also receives other input parameters, such as equipment aging level and historical alarm frequency, as context variables in the weight calculation. Finally, the dynamic weights of each score are jointly determined by the basic preference coefficient and the sensitivity function, and are normalized to ensure a sum of 1. The comprehensive risk value obtained after weighted fusion is mapped to a preset level range, corresponding to four risk levels: safety, caution, warning, and danger, achieving a unified risk criterion output.

[0036] The graded safety control strategy in S4 includes: if the risk level is at the attention level, a local low-frequency warning is activated and monitoring is enhanced; if the risk level is at the warning level, a remote alarm is activated and the exhaust system is activated to maximum power; if the risk level is at the danger level, the emergency shut-off valve is immediately activated to close the gas source and the exhaust gas treatment device is activated. The leak source location and gas mixing tracing process includes: when the comprehensive risk level output by S3 is at the attention level, warning level, or danger level, the leak source location and gas mixing tracing process is activated; wherein, the leak source location is based on the pipeline topology of the special gas holder and the spatiotemporal concentration data collected by multiple sensors, and the candidate nodes for the leak are determined by the time difference inversion method and the graph neural network model; the gas mixing tracing is based on a preset gas interaction knowledge graph, identifies gas convergence paths with reaction risks, and traces upstream pollution sources or cross-contamination points; a structured diagnostic report is generated based on the location results and tracing path, and the diagnostic report includes at least an anomaly summary, risk propagation path, leak source confidence, cause analysis, and operation and maintenance disposal recommendations.

[0037] In step S4, the corresponding graded safety control strategy is executed based on the comprehensive risk level output from S3. When the risk level is at the attention level, the system activates a local low-frequency audible and visual warning and automatically increases the monitoring frequency of key parameters to enhance data acquisition. When the warning level is reached, the system triggers a remote alarm to notify the operator and activates the exhaust system to run at maximum power to prevent gas accumulation. When the risk level is reached, the system immediately drives the emergency shut-off valve to close the gas source and simultaneously activates the exhaust gas treatment device to prevent the leakage of harmful gases. In addition, when the comprehensive risk level is at the attention level or higher, the system automatically initiates the leak source location and gas mixing source tracing process. Leak source location is based on the pipeline topology of the special gas holder and the spatiotemporal concentration data of multiple sensors. The time difference inversion method combined with a graph neural network model is used to infer the node location where the leak is most likely to occur. Gas mixing source tracing is based on a preset gas reaction knowledge graph to identify gas convergence paths with reaction risks and traces upstream possible pollution sources or cross-contamination points. Finally, the system generates a structured diagnostic report, which includes an anomaly summary, risk propagation path, confidence ranking of leakage sources, possible cause analysis, and operation and maintenance handling suggestions.

[0038] Example 2

[0039] A special gas cabinet safety control system based on multi-parameter monitoring is provided for executing the method described in Embodiment 1. The system includes: a distributed multi-source sensor network for real-time monitoring of the target special gas concentration, oxygen concentration, environmental parameters, physical state parameters, and safety equipment status parameters within the special gas cabinet; a feature extraction module for preprocessing the collected monitoring data and extracting temporal features to generate a fused feature set; and an intelligent assessment and risk fusion module for performing multi-dimensional risk assessment on the fused feature set, which integrates: an anomaly detection sub-model, employing a hybrid neural network structure based on LSTM and VAE encoders, for temporal modeling of the fused feature set and outputting a first anomaly score; and a gas interaction assessment sub-model, incorporating a built-in gas reaction knowledge graph and combining improved DS evidence theory to assess the risk of multi-gas coexistence. The system performs fusion reasoning to output a second risk score; a pipeline topology risk propagation sub-model models the special gas holder pipeline system as a directed graph with attributes, simulates the spatial propagation process of leakage risk through a random walk algorithm, and outputs a third risk score; a risk fusion model is used to weight and fuse the first anomaly score, the second risk score, and the third risk score to generate a comprehensive risk value, which is then mapped to a comprehensive risk level; a hierarchical control module is used to execute corresponding hierarchical safety control strategies based on the comprehensive risk level; a leakage tracing and diagnostic report generation module initiates the leakage source location and gas mixing tracing process and generates a diagnostic report when the comprehensive risk level is at the attention level, warning level, or danger level; and a central computing and communication unit is used to coordinate the data flow and control flow between modules, schedule task execution, and provide a human-machine interface.

[0040] This invention provides a special gas cabinet safety control system for implementing the aforementioned method, comprising multiple functional modules working collaboratively. A distributed multi-source sensor network is responsible for real-time acquisition of multi-dimensional data such as the target special gas concentration, oxygen concentration, environmental parameters, physical state, and safety equipment status inside and around the special gas cabinet. A feature extraction module filters, denoises, and extracts temporal features from the received raw data to generate a fused feature set for risk assessment. The intelligent assessment and risk fusion module is the core processing unit, internally integrating an anomaly detection sub-model, a gas interaction assessment sub-model, a pipeline topology risk propagation sub-model, and a risk fusion model, respectively calculating and synthesizing the scores for three types of risks. A hierarchical control module outputs control commands based on the comprehensive risk level, driving the alarm device, ventilation system, or emergency shut-off valve to operate. A leak tracing and diagnostic report generation module initiates the location and tracing process when a risk anomaly occurs and outputs a structured report. A central computing and communication unit is responsible for coordinating data transmission, task scheduling, and human-machine interaction between modules to ensure the overall efficient operation of the system.

[0041] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. All equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this invention should still be covered by the claims of this invention.

Claims

1. A safety control method for special gas holders based on multi-parameter monitoring, characterized in that, include: S1, real-time monitoring of target special gas concentration, oxygen concentration, environmental parameters, physical state parameters and safety equipment status parameters in and around the special gas cabinet; S2, preprocesses the collected monitoring data and extracts time-series features to generate a fusion feature set for risk assessment; S3, based on the preset gas safety properties and the pipeline connection topology of the special gas cabinet, uses an intelligent assessment model to perform fusion analysis on multi-dimensional parameters and outputs the risk level. The intelligent assessment model includes an anomaly detection model, a gas interaction assessment model, and a pipeline topology risk propagation model. S4. Based on the risk level, execute the corresponding graded safety control strategy. When the risk level is abnormal, initiate the leak source location and gas mixing source tracing process and generate a diagnostic report. S3 includes: S31, the anomaly detection model adopts a hybrid model based on LSTM network and VAE encoder, performs time series modeling on the fusion feature set generated in S2, calculates the reconstruction error and prediction deviation of the current state respectively, and generates the first anomaly score by weighting. S32, the gas interaction assessment model constructs a gas reaction knowledge graph with gas type, concentration range, and environmental parameters as nodes, and based on this graph and real-time data, it uses an improved DS evidence theory to fuse multi-source evidence and calculates a second risk score for gas mixing to cause combustion, explosion, or the generation of toxic byproducts. S33, the pipeline topology risk propagation model abstracts the special gas holder pipeline system into a directed graph G=(V,E), where nodes V represent valves, joints, and sensors, and edges E represent pipeline connections; based on the random walk algorithm, it simulates the propagation process of abnormal signals or leakage risks in the directed graph and generates a third risk score that characterizes the spatial risk distribution; S34, the first anomaly score, the second risk score and the third risk score are weighted and fused to generate a comprehensive risk value, and the corresponding comprehensive risk level is determined according to the preset threshold range to which the comprehensive risk value belongs.

2. The special gas holder safety control method based on multi-parameter monitoring according to claim 1, characterized in that, The S1 uses a distributed multi-source sensor network to collaboratively collect multi-dimensional operational and environmental data from the interior and surrounding areas of the special gas holder; the data acquisition process is dynamically controlled by an adaptive monitoring engine based on real-time process conditions, specifically including: The process formula currently in operation is analyzed, the physicochemical properties of the gases involved are extracted, and a risk weight vector associated with each safety parameter is constructed to generate a dynamic risk profile. Based on the risk weight vector and the preset benchmark monitoring strategy, the final monitoring strategy is dynamically generated through a parameterized adjustment algorithm. The parameterized adjustment algorithm includes at least a nonlinear mapping of sampling frequency, computing resource priority, and alarm threshold. Based on the generated ultimate monitoring strategy, the distributed multi-source sensor network is coordinated to collect multi-dimensional operation and environmental data of the special gas cabinet and its surrounding related areas, and specific composite alarm logic is dynamically enabled or disabled according to the dynamic risk profile.

3. The special gas holder safety control method based on multi-parameter monitoring according to claim 2, characterized in that, S2 includes: An adaptive filtering algorithm based on process stage perception is used to preprocess the raw monitoring data; Based on the data characteristics of different security parameters, multi-scale temporal features are extracted from their time series to form an initial feature set; Based on the risk weight vector in the risk profile, the initial feature set is weighted and filtered to generate a key feature subset; the key feature subset is input into the feature fusion algorithm for nonlinear transformation and integration to generate a fused feature set for risk assessment.

4. The special gas holder safety control method based on multi-parameter monitoring according to claim 1, characterized in that, S31 includes: With time window [ ,..., Using historical fusion features as input, an LSTM network is used to predict the feature value at the current time. And calculate the prediction bias. ,in This represents the actual observed feature vector at the current moment; Based on the fusion characteristics at the current moment As input, it is mapped to the latent space using a VAE encoder, and then reconstructed by a decoder. And calculate the reconstruction error = ; The prediction bias and reconstruction error via normalization function and The process yields dimensionless standardized anomaly confidence scores. Assign weights to the standardized anomaly confidence scores. and The weighted sum is calculated and then input into the final transformation function. Generate the first anomaly score The calculation formula is as follows: in and Let be the weighting coefficients of the anomaly detection sub-model, and satisfy . .

5. The special gas holder safety control method based on multi-parameter monitoring according to claim 1, characterized in that, S32 includes: Construct a gas reaction knowledge graph with gas type, concentration range, and environmental parameters as nodes and chemical reaction relationships as edges; match real-time monitored multi-gas concentration data and environmental parameters with the gas reaction knowledge graph to generate a set of evidence that is activated in real time, where each piece of evidence corresponds to a risk condition that is triggered in the graph; For each piece of evidence in the evidence set, based on its monitoring value, a preset membership function that conforms to its physicochemical characteristics is used. Dynamically calculate its basic probability allocation for the target risk proposition: ， in, As evidence The monitored values; To identify the frame, represent the union of all possible risks; Indicates the first The degree of support or trust in a particular set of threats from each source of evidence; For each piece of evidence Assign a dynamic credibility weight , , For sensor reliability, For operating condition effectiveness, For context consistency, , , These are the normalized weighting coefficients; use The basic probability assignment is modified. , ; All modified basic probability assignments are fused using DS evidence theory: in, This indicates that after integrating all the evidence, the proposition is... The basic probability distribution; K is the normalization factor; Represents all subpropositions The intersection equals the proposition ; Represented as the first A source of evidence for a subproposition The corrected basic probability assignment; This represents the sum of probabilities when all combinations of evidence are disjoint; Represented as all subpropositions The intersection of these sets is an empty set; The fused basic probability allocation Projected into the risk space, a second risk score is calculated to determine if gas mixing will lead to combustion, explosion, or the generation of toxic byproducts. The specific formula is as follows: The The normalized risk value for the interval [0,1].

6. The special gas holder safety control method based on multi-parameter monitoring according to claim 1, characterized in that, S33 includes: The special gas holder piping system is modeled as a directed graph G=(V,E), where the node set V represents valves, joints, and sensors, and the edge set E represents the piping connection direction. The risk confidence of each node is initialized based on real-time monitoring data, and a random walk algorithm is used to simulate the propagation process of abnormal signals or leakage risks in the directed graph to obtain the final risk value of each node. ,in , The total number of nodes; Calculate the maximum value of the final risk value among all nodes. and arithmetic mean ; Introducing distance decay factor , This represents the typical distance from the node to the gas source; This refers to the distance attenuation rate. The third risk score is finally generated using the following formula. ; in, This is the third risk assessment score; This is the global gain coefficient; This is a risk weighting factor.

7. The special gas holder safety control method based on multi-parameter monitoring according to claim 1, characterized in that, S34 includes: The first abnormal score Second risk score and third risk score Context-aware dynamic weighted fusion is performed to generate a comprehensive risk value. ; （ ） in, , , Is with , , The corresponding weighting coefficients; For the first Individual risk dynamic weights; For the first Risk score The static basic preference coefficient; For the first A sensitivity function for each risk; Normalization factor; These represent input parameters or state variables related to the safety control of special gas holders. Comprehensive risk value With respect to the preset level threshold range , , Compare; if Then it is a security level; if Then it is ;if Then it is ;if Then it is .

8. The special gas holder safety control method based on multi-parameter monitoring according to claim 7, characterized in that, The hierarchical security control strategy in S4 includes: If the risk level is at the alert level, a local low-frequency alert will be activated and monitoring will be enhanced. If the risk level is warning level, activate the remote alarm and activate the exhaust system to maximum power; If the risk level is dangerous, the emergency shut-off valve will be activated immediately to shut off the gas supply, and the exhaust gas treatment device will be activated in conjunction with it. The process for locating the leak source and tracing the gas mixture includes: When the overall risk level output by S3 is at the attention level, warning level, or danger level, the leak source location and gas mixing source tracing process is initiated. The leak source location is based on the pipeline topology of the special gas holder and spatiotemporal concentration data collected by multiple sensors, using time difference inversion and graph neural network models to determine candidate nodes where leaks occur. The gas mixing source tracing is based on a pre-set gas interaction knowledge graph to identify gas convergence paths with reaction risks and trace their upstream pollution sources or cross-contamination points. A structured diagnostic report is generated based on the location results and tracing paths. The diagnostic report includes at least an anomaly summary, risk propagation path, leak source confidence level, causal analysis, and operation and maintenance recommendations.

9. A system for implementing the method according to any one of claims 1-8, characterized in that, The system includes: A distributed multi-source sensor network is used to monitor in real time the concentration of target special gas, oxygen concentration, environmental parameters, physical state parameters, and safety equipment status parameters inside the special gas cabinet. The feature extraction module preprocesses the collected monitoring data and extracts time-series features to generate a fused feature set. The intelligent assessment and risk fusion module is used to perform multi-dimensional risk assessment on the fused feature set. It integrates the following: The anomaly detection sub-model adopts a hybrid neural network structure based on LSTM network and VAE encoder to perform temporal modeling of the fused feature set and output the first anomaly score; The gas interaction assessment sub-model incorporates a gas reaction knowledge graph and combines an improved DS evidence theory to perform fusion reasoning on the risk of multi-gas coexistence, outputting a second risk score. The pipeline topology risk propagation sub-model models the special gas holder pipeline system as a directed graph with attributes, simulates the spatial propagation process of leakage risk through a random walk algorithm, and outputs a third risk score. The risk fusion model is used to weight and fuse the first anomaly score, the second risk score, and the third risk score to generate a comprehensive risk value and map it to a comprehensive risk level. The hierarchical control module is used to execute corresponding hierarchical security control strategies based on the comprehensive risk level. The leak tracing and diagnostic report generation module initiates the leak source location and gas mixing tracing process and generates a diagnostic report when the comprehensive risk level is at the attention level, warning level, or danger level. The central computing and communication unit is used to coordinate the data and control flows between modules, schedule task execution, and provide a human-machine interface.

Images