Special gas cylinder state real-time monitoring and abnormal behavior identification system

Abstract

The application discloses a special gas cylinder state real-time monitoring and abnormal behavior identification system, and belongs to the technical field of gas cylinder safety monitoring. The gas cylinder state data acquisition module continuously and real-timely acquires the pressure, temperature, position and gas concentration of the gas cylinder, and the data acquisition frequency is completed within one second. The gas cylinder digital twin module constructs a virtual three-dimensional digital model of the gas cylinder based on gas thermodynamic physical constraints, realizes double-source fusion of physical constraints and measured data by establishing an equivalent mapping relationship between the virtual model and the real-time state of the gas cylinder, and adaptively corrects the equivalent mapping relationship based on a deviation decomposition attribution mechanism. The deviation decomposition attribution mechanism divides the deviation into three categories: random error, systematic cumulative deviation and abnormal deviation, and adopts differentiated correction strategies for different types. The equivalent mapping dynamic updating unit continuously optimizes the mapping function through an online learning algorithm, so that the model has self-evolution ability. The passive monitoring is changed into active prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gas cylinder safety monitoring, in particular to a special gas cylinder state real-time monitoring and abnormal behavior identification system, and especially to a gas cylinder state monitoring system based on a digital twin technology and a physical constraint driven adaptive mapping correction mechanism. BACKGROUND

[0002] Special gas is widely used in semiconductor manufacturing, medical health, chemical production and other fields, and the safety state of its storage container gas cylinder is directly related to production safety and personnel safety. The traditional gas cylinder monitoring method mainly relies on manual periodic inspection and simple pressure gauge reading. This method has problems such as low monitoring frequency, lagging response, and inability to predict potential risks. With the development of Internet of Things technology, some enterprises have begun to use sensors to collect gas cylinder pressure and temperature data in real time, but these systems still have obvious technical defects.

[0003] The main problem of the prior art is that the monitoring dimension is single, only focusing on the change of a single parameter of pressure or temperature, and using a simple threshold alarm mechanism. When the pressure exceeds the preset upper limit or is lower than the preset lower limit, an alarm is triggered. This passive post-response cannot provide early warning of potential dangers. In addition, there is an inherent physical correlation between the pressure and temperature of the gas, and monitoring a single parameter alone can easily result in false positives or false negatives. For example, the increase in the pressure in the gas cylinder due to the increase in the ambient temperature is a normal physical phenomenon, but if only the pressure change is monitored, it may be misjudged as abnormal.

[0004] More importantly, existing systems lack deep understanding and prediction ability of the gas cylinder state. During long-term use, the physical properties of the gas cylinder will change slowly, such as corrosion of the inner wall and aging of the valve, which are difficult to discover through simple data comparison. Some technical solutions attempt to introduce data-driven machine learning models for anomaly detection, but pure data-driven black box models lack physical interpretability, and the prediction may fail when the gas cylinder operates in new conditions not covered by the training data, and the fixed model parameters cannot adapt to the long-term evolution of the physical properties of the gas cylinder.

[0005] Digital twin technology provides a new approach to gas cylinder monitoring by constructing a virtual model of the gas cylinder in real time with the physical entity. However, existing digital twin applications mainly face two challenges: first, the deviation between the virtual model and the physical entity will accumulate over time, and there is a lack of effective self-correction mechanism; second, the fusion of physical modeling and data-driven is not deep enough, and the physical laws of gas thermodynamics are not fully utilized to constrain and guide the evolution of the model.

[0006] Therefore, there is an urgent need to develop an intelligent gas cylinder monitoring system that can fuse physical constraints and measured data, has adaptive correction capability, and can achieve predictive maintenance. SUMMARY

[0007] The application aims to provide a special gas cylinder state real-time monitoring and abnormal behavior identification system, and solve the technical problems of single monitoring dimension, inability to predict potential risks, and cumulative deviation of digital twin models that cannot be self-corrected in the prior art.

[0008] The application provides a special gas cylinder state real-time monitoring and abnormal behavior identification system, which comprises:

[0009] A cylinder state data acquisition module is configured to continuously and real-timely acquire pressure values, temperature values, position information and gas concentration values in the cylinder, and the data acquisition frequency is within 1 second.

[0010] A cylinder data preprocessing module is in communication connection with the cylinder state data acquisition module, configured to receive real-time data sent by the cylinder state data acquisition module, and perform data cleaning, data transformation and data normalization processing on the real-time data, wherein the data cleaning is configured to filter abnormal data and repeated records.

[0011] A cylinder digital twin module is in communication connection with the cylinder data preprocessing module, configured to receive data processed by the cylinder data preprocessing module, construct a virtual three-dimensional digital model of the cylinder based on gas thermodynamic physical constraints, realize double-source fusion of physical constraints and measured data by establishing an equivalent mapping relationship between the virtual three-dimensional digital model and the real-time state of the cylinder, and perform self-adaptive correction on the equivalent mapping relationship based on a deviation decomposition attribution mechanism.

[0012] An abnormal early warning module is in communication connection with the cylinder digital twin module, configured to receive digital twin data output by the cylinder digital twin module, compare the real-time data of the cylinder with the digital twin data to calculate a deviation degree value, and generate a warning information when the deviation degree value exceeds a preset threshold.

[0013] Preferably, the cylinder state data acquisition module comprises a gas sensor, a temperature sensor, a position tag and a pressure sensor, the gas sensor is configured to measure the gas concentration in the cylinder, the pressure sensor is configured to acquire the pressure value in the cylinder, the temperature sensor is configured to measure the temperature value in the cylinder, and the position tag is configured to record the position information of the cylinder.

[0014] Preferably, the cylinder digital twin module comprises:

[0015] A physical constraint modeling unit is configured to determine the physical correlation among the pressure, temperature and volume in the cylinder based on a gas state equation, select corresponding physical model parameters according to the type of the stored gas, establish a coupling relationship model of pressure and temperature, and introduce a fluid mechanics conservation equation as a supplementary constraint.

[0016] A data feature extraction unit is in communication connection with the cylinder data preprocessing module, configured to receive the data processed by the cylinder data preprocessing module, perform multi-scale time series analysis on the processed data, and extract short-term fluctuation features, medium-term trend features, and long-term drift features.

[0017] A dual-source fusion unit is connected with the physical constraint modeling unit and the data feature extraction unit respectively, configured to receive the theoretical value calculated by the physical constraint modeling unit and the measured feature extracted by the data feature extraction unit, establish the correlation between the theoretical calculation value and the measured data feature, and perform comparison and matching, and record the position, size, and mode of deviation when deviation is detected.

[0018] Preferably, the physical constraint modeling unit is further configured to read the physical parameters of the cylinder, including cylinder volume, gas type, gas molar mass, and gas constant, calculate the theoretical pressure value and the theoretical temperature value at the current time based on the physical parameters and the gas state equation, and call a fluid mechanics solver to calculate the temperature field distribution, pressure field distribution, and velocity field distribution inside the cylinder.

[0019] Preferably, the data feature extraction unit maintains three data buffers with different time spans, a short-term buffer for storing data of the last one minute and extracting fluctuation features, a medium-term buffer for storing data of the last one hour and extracting trend features, and a long-term buffer for storing data of the last one week and extracting drift features, quantifies the fluctuation degree by calculating the standard deviation and the coefficient of variation in the sliding window, calculates the trend slope and the trend strength by linear regression fitting, and judges systematic deviation by comparing the baseline values of different time periods.

[0020] Preferably, the cylinder digital twin module further comprises:

[0021] A deviation decomposition and attribution unit is in communication connection with the dual-source fusion unit, configured to receive the deviation information recorded by the dual-source fusion unit, calculate a comprehensive deviation index, analyze the time evolution law of the deviation, and determine the deviation type, the deviation type including random error, systematic cumulative deviation, and abnormal deviation.

[0022] A hierarchical correction strategy unit is in communication connection with the deviation decomposition and attribution unit, configured to receive the deviation type output by the deviation decomposition and attribution unit, and perform differentiated correction operations based on the deviation type, wherein random error is not corrected, a physical parameter adaptive adjustment mechanism is started for systematic cumulative deviation, and a fast response mode is started for abnormal deviation.

[0023] As preferred, the bias decomposition attribution unit is further used to maintain a bias history database, record bias values and their decomposition results at each time point in the past period of time, calculate an autocorrelation function of the bias to determine randomness, calculate a trend line fitting goodness of the bias to determine systematicness, calculate a mutation detection statistic of the bias to determine abnormality, and label the current bias with an attribution label based on the analysis results, the attribution label including a bias type, a bias severity, and a recommended correction strategy.

[0024] As preferred, the physical parameter adaptive adjustment mechanism in the hierarchical correction strategy unit includes: calculating a parameter sensitivity matrix, the parameter sensitivity matrix being used to describe a bias change amount caused by a unit change of each adjustable parameter, using a steepest descent method or a conjugate gradient method to calculate an optimal parameter adjustment vector, an adjustment amount of the parameter adjustment vector being limited by a safety constraint to ensure that a single adjustment does not make the virtual three-dimensional digital model deviate from a physically reasonable range, and returning the adjusted parameter value to the physical constraint modeling unit to trigger re-computation of the virtual model.

[0025] As preferred, the gas cylinder digital twin module further includes:

[0026] An equivalent mapping dynamic updating unit, in communication connection with the hierarchical correction strategy unit, is configured to receive a result after the hierarchical correction strategy unit performs a correction operation, extract input and output samples of this correction as training data, and update weight coefficients of an equivalent mapping function using an online learning algorithm, the equivalent mapping function being used to calculate a parameter configuration that the virtual three-dimensional digital model should adopt according to measured data features;

[0027] An iterative convergence control unit, in communication connection with the equivalent mapping dynamic updating unit, is configured to monitor a convergence state of a correction process, maintain a state machine including normal operation, correction in progress, convergence determination, convergence success, and convergence failure, and calculate a bias change rate and an iteration number in real time, confirm convergence success when the bias remains stable and is lower than a threshold value in three consecutive sampling periods, and determine convergence failure and trigger an exception handling process when the iteration number exceeds an upper limit or the bias continues to oscillate.

[0028] As preferred, the exception early warning module includes:

[0029] A pressure and temperature exception feature recognition unit is configured to receive gas cylinder real-time data, monitor abnormal behaviors of the gas cylinder in terms of pressure and temperature in real time, and determine abnormal modes or alarm modes of the gas cylinder through numerical fluctuations and fluctuation degrees of the pressure and the temperature;

[0030] A pressure and temperature joint analysis unit, in communication connection with the pressure and temperature exception feature recognition unit, is configured to receive exception features output by the pressure and temperature exception feature recognition unit, comprehensively determine a relationship between pressure and temperature changes, and jointly analyze abnormal behaviors of the gas cylinder.

[0031] The position data anomaly analysis unit is configured to receive real-time position information of the gas cylinder, determine the abnormal state of the position of the gas cylinder based on the real-time position information of the gas cylinder, and determine whether the position of the gas cylinder is safe based on the position abnormality degree value and the early warning.

[0032] The leakage degree analysis unit is configured to receive real-time pressure data of the gas cylinder, determine the leakage of the gas cylinder based on the real-time leakage degree of the pressure of the gas cylinder, and perform early warning and abnormal warning based on the output results of the pressure-temperature anomaly feature recognition unit, the pressure-temperature joint analysis unit, the position data anomaly analysis unit, and the leakage degree analysis unit.

[0033] The present application has the following advantages:

[0034] First, the present application constructs a gas cylinder digital twin module based on physical constraints, uses the gas thermodynamic law as the internal constraint of the model, and ensures that the evolution of the virtual model follows the basic physical law. This physical driving modeling method is different from the pure data-driven black box model. Even when the gas cylinder runs into a new working condition, as long as the gas still follows the basic state equation, the model prediction will not deviate from the physically reasonable range. The physical constraint plays a regularization role, significantly improving the generalization ability and reliability of the model.

[0035] Second, the present application creatively establishes an equivalent mapping relationship between virtual data and measured data, and realizes the fusion of physical constraints and measured data. Physical modeling provides a theoretical basis to ensure the essential correctness of the model, while data-driven captures the complex factors of the actual gas cylinder such as aging and environmental impact to fine-tune the physical model. This complementary fusion mechanism makes the model both physically interpretable and highly accurate.

[0036] Third, the bias decomposition attribution mechanism introduced in the present application can accurately identify the root cause of the deviation between the virtual model and the physical entity. Through statistical analysis, the deviation is decomposed into three categories: random error, systematic cumulative deviation, and abnormal deviation. Different correction strategies are used for different types. Random errors are not corrected to avoid introducing noise, systematic deviations are adjusted gradually with small steps to maintain stability, and abnormal deviations are quickly responded to track mutations. This divide-and-conquer method avoids the mis-correction and under-correction problems of traditional adaptive systems, ensuring the accuracy and efficiency of correction.

[0037] Fourthly, the equivalent mapping dynamic updating mechanism of the present application endows the system with the ability of continuous learning and self-evolution. The system takes each correction process as a learning opportunity, and constantly optimizes the mapping function parameters through online learning algorithm, so that the model can track the slow evolution of the physical characteristics of the gas cylinder. With the extension of the running time, the model accuracy is continuously improved, showing a benign trend of getting more and more accurate, greatly reducing the cost of manual maintenance.

[0038] Fifthly, the multi-level iterative convergence control mechanism of the present application sets three convergence criteria of absolute value, change rate and iteration number, and cooperates with physical constraint hard limit and adaptive step adjustment, effectively avoiding system oscillation and divergence caused by parameter adjustment, ensuring the robustness and stability of the correction process.

[0039] Sixthly, the present application realizes the paradigm shift from passive monitoring to active prediction. Through continuous learning of the gas cylinder operation law by the digital twin model, the system can predict the future state change trend of the gas cylinder, provide early warning before danger occurs, and gain disposal time for the operator, significantly improving the safety of gas cylinder use.

[0040] Seventhly, the system architecture of the present application adopts modular design, and each subsystem is relatively independent and interface standardized, which is convenient for engineering implementation and system upgrading. The calculation complexity is optimized, and the system can run in real time on ordinary industrial servers, with good industrial practicability and generalizability. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 Figure 1 is a schematic diagram of the overall architecture of the special gas cylinder state real-time monitoring and abnormal behavior recognition system of the present application. DETAILED DESCRIPTION

[0042] Please refer to Figure 1 , the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0043] As Figure 1 shown, the special gas cylinder state real-time monitoring and abnormal behavior recognition system provided by the present application includes a gas cylinder state data acquisition module 1, a gas cylinder data preprocessing module 2, a gas cylinder digital twin module 3 and an abnormal warning module 4. These modules are connected with each other through a communication network, and together constitute a complete gas cylinder intelligent monitoring system.

[0044] The gas cylinder state data acquisition module 1 is installed on the gas cylinder body and is responsible for continuously and real-timely acquiring key state parameters of the gas cylinder. In an embodiment of the present application, the parameters acquired by the module include pressure value, temperature value, position information and gas concentration value in the gas cylinder. The data acquisition frequency is set to complete one complete multi-parameter acquisition within 1 second, and the selection of this frequency is based on the time scale characteristics of the state change of the gas cylinder. For special gas cylinders, the pressure and temperature changes during normal use usually occur within minutes to hours, and abnormal events such as leakage may occur within seconds to tens of seconds. The 1-second sampling period can capture early signs of rapid abnormal events, and will not cause data redundancy and communication burden due to too high sampling frequency.

[0045] Preferably, the gas cylinder state data acquisition module 1 uploads the acquired data to the server side by using wireless power carrier communication technology. This communication method uses the gas cylinder gas supply pipeline as the signal carrier, without the need for additional wiring, and is particularly suitable for application scenarios with a large number of gas cylinders and dispersed locations. The data transmission protocol uses a compression encoding mechanism, and each transmission includes a timestamp, a sensor number and the measurement values of each parameter, and the size of a single data packet is usually within 128 bytes, ensuring the real-time nature of the communication.

[0046] The gas cylinder data preprocessing module 2 and the gas cylinder state data acquisition module 1 establish a data transmission channel through communication connection. After receiving the original sensor data, the module performs a series of preprocessing operations, including data cleaning, data transformation and data normalization processing. Data cleaning is the first step of preprocessing, which is used to identify and filter abnormal data and repeated records. In actual industrial environments, sensors may produce abnormal values that deviate significantly from the normal range due to electromagnetic interference, communication errors and other reasons. The present application uses an abnormality detection method based on statistical characteristics to compare the statistical distribution of newly arrived data points with historical data.

[0047] In a preferred embodiment of the present application, the data cleaning unit maintains a sliding window, and the window size is set to the last 60 data points, corresponding to the sampling data of the last 1 minute. For the newly arrived pressure measurement value, the system calculates the mean and standard deviation of all data points in the sliding window. If the deviation of the new data point from the mean exceeds 3 times the standard deviation, it is preliminarily determined as a suspicious point. Considering that the state of the gas cylinder may change rapidly, the system further checks the adjacent data points before and after the suspicious point. If the adjacent data points are within the normal range, and only the point is abnormal, it is determined as measurement noise and is removed. If the point and the subsequent data points all show the same trend of change, it is determined as a real state jump and is retained. This abnormality detection method considering the time continuity can effectively distinguish measurement noise and real abnormal events.

[0048] The data transformation operation normalizes the cleaned data, eliminating the differences in dimension and numerical range between different physical quantities. The typical range of pressure data is 0 to 15 MPa, and the typical range of temperature data is -50 to 80 degrees Celsius, with a large difference in numerical scale. Data transformation maps each parameter value to a unified numerical interval through linear mapping, usually choosing a normalized interval of 0 to 1 or -1 to 1. This transformation facilitates subsequent modules for comprehensive analysis and comparison of multi-dimensional data.

[0049] Data normalization adjusts the distribution of data characteristics to meet the input requirements of subsequent algorithms. The normalization method used in this invention calculates the maximum and minimum values of each parameter based on historical data statistics, which are used as the upper and lower bounds of normalization. For pressure parameters, if the maximum value is 12 MPa and the minimum value is 0.5 MPa, then the current pressure value is subtracted by 0.5 MPa and divided by 11.5 MPa to get the normalized value. This normalization method based on historical extreme values can fully utilize the numerical space and improve the accuracy of data representation.

[0050] The gas cylinder digital twin module 3 is the core innovation of this invention, which receives pre-processed data and constructs a virtual three-dimensional digital model of the gas cylinder. This module contains multiple functional units working together to achieve deep integration of physical constraints and measured data, as well as adaptive correction of the model. This module design fully considers the physical laws of gas thermodynamics, while also being able to flexibly learn the individualized characteristics of actual gas cylinders.

[0051] The anomaly warning module 4 is connected to the gas cylinder digital twin module 3 and receives the virtual model data output by the digital twin module. The core function of this module is to continuously compare the real-time data of the gas cylinder with the digital twin data and calculate the deviation degree value between them. When the deviation degree value exceeds the preset threshold, the system determines that the gas cylinder may have an anomaly and generates a warning information to push to the management personnel. This warning mechanism based on digital twin comparison is different from traditional single-parameter threshold alarm, which can consider the associated changes of multi-dimensional parameters, significantly reducing the false positive rate and false negative rate.

[0052] The hardware components of the gas cylinder state data acquisition module 1 include gas sensors, temperature sensors, location tags, and pressure sensors. The selection and configuration of these sensors directly affect the monitoring accuracy and reliability of the system.

[0053] The pressure sensor is a core component of the gas cylinder monitoring system, which is used to collect the pressure value inside the gas cylinder. In the preferred embodiment of the present application, the pressure sensor adopts a piezoresistive or capacitive principle, with a measurement range covering 0 to 25 MPa and an accuracy level reaching 0.5 level, i.e. the measurement error is not more than 0.5% of the full scale. For a working pressure of 15 MPa, the measurement error is controlled within 0.075 MPa. The pressure sensor is installed at the cylinder valve interface, and the sealing is ensured by threaded connection. The response time of the sensor is less than 100 milliseconds, which can quickly capture the transient changes in pressure. Considering the possible vibration and impact of the gas cylinder in use, the sensor shell is made of stainless steel and is designed to be shock-resistant, ensuring long-term stable operation.

[0054] The temperature sensor is used to measure the temperature value inside the gas cylinder. The temperature of the gas inside the cylinder is affected by many factors such as environmental temperature, gas flow and compression expansion. The present application adopts a platinum resistance temperature sensor with a measurement range of minus 50 degrees Celsius to 80 degrees Celsius and an accuracy of A level, i.e. the error at 0 degrees Celsius is not more than 0.15 degrees Celsius. The temperature sensor is installed at the middle position of the cylinder body, and the metal probe with good heat conductivity is in contact with the cylinder wall to collect the cylinder wall temperature as a representation of the gas temperature inside the cylinder. Although there may be some difference between the bottle wall temperature and the core temperature of the gas, for the state monitoring of the cylinder, the bottle wall temperature can reflect the trend of the thermal state change of the cylinder, and the installation is more simple and reliable.

[0055] The gas sensor is used to measure the gas concentration inside the cylinder, which is particularly important in the case of mixed gas or possible change of gas composition. For example, some special gases may decompose or react with impurities during storage, resulting in a decrease in gas purity. The gas sensor selects the corresponding detection principle according to the type of gas to be monitored, such as electrochemical sensor, infrared absorption sensor or thermal conductivity sensor. For active gases such as oxygen, electrochemical sensors can achieve high sensitivity detection. For gases such as carbon dioxide with obvious infrared absorption characteristics, infrared sensors have the advantages of good selectivity and strong anti-interference ability. The gas sensor extracts a small amount of gas from the inside of the cylinder through the sampling pipeline for analysis, and the sampling flow is controlled at a few milliliters per minute, which has negligible effect on the total pressure of the cylinder.

[0056] The position tag is used to record the position information of the gas cylinder, and realize the spatial tracking of the gas cylinder. The special gas cylinder usually needs to be transported between different places for use, and accurate grasp of the position of the gas cylinder is very important for inventory management and emergency response. The position tag of the application adopts integrated GPS and Beidou dual-mode positioning, and the positioning accuracy can reach within 5 meters in an open outdoor environment. Considering that the gas cylinder may be used in indoor or underground places, the position tag also integrates the function of Bluetooth beacon. In the indoor environment, the indoor positioning is realized through the triangular positioning or proximity detection of the arranged Bluetooth receiving base station. The position tag reports the position information every 5 minutes, and when the movement is detected, the reporting frequency is increased to once a minute, which not only ensures the timeliness of the position tracking, but also controls the data traffic and power consumption.

[0057] The data collected by the four types of sensors is gathered and preliminarily processed by the data acquisition controller. The data acquisition controller adopts a low-power embedded microcontroller, and a built-in analog-to-digital converter converts the analog signals of the sensors into digital signals. The controller synchronously samples each sensor, ensuring that the multi-dimensional data collected at the same time is time-consistent, which is crucial for analyzing the correlation between parameters. The collected digital data is packaged into data frames, and after adding time stamps and data check codes, it is sent to the server side through the wireless communication module.

[0058] In terms of power supply configuration, the gas cylinder state data acquisition module 1 adopts a replaceable battery power supply mode. Considering that the use cycle of the gas cylinder is usually several months to one year, the battery capacity is selected as 5000 mAh lithium battery. Through optimization of the sensor working mode and the communication protocol, the average power consumption of the system is controlled below 15 mW, and a single battery full charge can support continuous operation for more than 6 months. When the battery power drops to 20%, the system automatically sends a low power alarm to remind the management personnel to replace the battery in time, avoiding monitoring interruption due to power depletion.

[0059] The gas cylinder digital twin module 3 includes a physical constraint modeling unit, a data feature extraction unit and a dual-source fusion unit. These three units constitute the basic framework of digital twin, realizing the complete chain from physical law to measured data to fusion modeling.

[0060] The physical constraint modeling unit is based on the gas state equation to build the physical correlation between pressure, temperature and volume. The gas state equation describes the essential relationship between the macroscopic state parameters of the gas, which is the basic law of thermodynamics. For an ideal gas, the state equation is expressed as the product of pressure and volume equal to the product of the amount of gas substance, gas constant and temperature. Although the actual gas deviates from the ideal gas behavior under high pressure or low temperature conditions, it can still be described by the state equation after introducing the compression factor correction. The physical constraint modeling of the application is based on this basic physical law.

[0061] The physical constraint modeling unit first needs to determine the geometric parameters and gas property parameters of the cylinder. The geometric parameters of the cylinder include the inner diameter, wall thickness and total volume, which are determined during the manufacture of the cylinder and remain unchanged during use as fixed boundary conditions of the model. Taking a standard 40-liter steel cylinder as an example, the inner diameter is 232 mm, the wall thickness is 5.3 mm, and the total volume is 40 liters. The gas property parameters depend on the type of gas stored, and the molar mass, gas constant, specific heat capacity and other parameters of different gases are different. For nitrogen, the molar mass is 28 grams per mole, and the gas constant is 297 joules per kilogram kelvin. For oxygen, the molar mass is 32 grams per mole, and the gas constant is 260 joules per kilogram kelvin. The system reads the corresponding physical parameters from the parameter database according to the gas type input by the user during initial configuration.

[0062] After establishing the basic physical association, the physical constraint modeling unit further introduces the fluid mechanics conservation equation as a supplementary constraint. The mass conservation equation ensures that the total mass of the gas in the cylinder remains constant unless leakage or charging occurs. The momentum conservation equation describes the change in momentum when the gas flows, although the flow of gas in a stationary storage cylinder is very weak, but in the case of charging, discharging or natural convection caused by temperature gradient, the flow effect still needs to be considered. The energy conservation equation describes the heat transfer process, and there is heat exchange between the gas in the cylinder and the environment, and the combined action of heat conduction, convection and radiation of the cylinder wall makes the temperature in the cylinder tend to the ambient temperature.

[0063] These physical constraints together constitute the mathematical representation of the virtual model. At each time step, the physical constraint modeling unit calculates the theoretical state value at the next time according to the current pressure and temperature state, combined with the physical constraint equation. This calculation process uses numerical solution method to discretize the continuous partial differential equation into algebraic equation system for solution. Considering the real-time requirement, the calculation method is optimized in the present application.

[0064] The data feature extraction unit performs multi-scale time series analysis on the real-time collected pressure and temperature data. The purpose of time series analysis is to extract feature information that can represent the change rule of the cylinder state from continuous data stream. Unlike physical modeling, which focuses on theoretical values, data feature extraction focuses on the statistical characteristics and change trend of measured values.

[0065] In one embodiment of the application, the data feature extraction unit maintains three data buffers of different time spans. The short-term buffer stores data of the last minute, corresponding to 60 sampling points, for extracting short-term fluctuation features. The system calculates the standard deviation and coefficient of variation of these 60 data points, the standard deviation reflecting the dispersion of data, and the coefficient of variation, the ratio of standard deviation to mean, reflecting the relative fluctuation amplitude. For a normally operating cylinder, the short-term standard deviation of pressure is usually less than 0.05 MPa, and the coefficient of variation is less than 0.5%. If the short-term standard deviation suddenly increases to more than 0.2 MPa, it may indicate that the cylinder is disturbed by external factors or has an internal abnormality.

[0066] The medium-term buffer stores data of the last hour, corresponding to 3600 sampling points, for extracting medium-term trend features. The system performs linear regression fitting on these data points, calculating the slope of the trend line and the goodness of fit. The sign and magnitude of the trend slope reflect the direction and speed of parameter change. For the temperature parameter, if the ambient temperature is constant and the cylinder is in a stationary state, the temperature trend slope should be close to zero. If the trend slope is positive and persists, it indicates that the temperature is rising continuously, which may be a normal response to the rising ambient temperature or a temperature rebound process after cylinder filling. The goodness of fit measures the degree of agreement between data points and the trend line, a high goodness of fit indicates that the data presents a clear linear trend, and a low goodness of fit indicates that the data fluctuates greatly or presents a nonlinear change.

[0067] The long-term buffer stores data of the last week, due to the large amount of data, the system adopts a downsampling strategy, retaining one data point every 5 minutes, a total of 2016 data points in a week, for extracting long-term drift features. Long-term drift features are identified by comparing baseline values at different times. The system divides the weekly data by day, calculates the median of each day's data as the baseline value for that day, and then compares the changes between the first day's baseline and the last day's baseline. If the pressure baseline slowly decreases from 10 MPa on the first day to 9.7 MPa on the seventh day, a total decrease of 0.3 MPa, this long-term slow pressure decrease may indicate a small leak in the cylinder. This slow leak is difficult to detect from short-term data alone, but long-term statistical features can reveal its existence.

[0068] The data feature extraction unit organizes the extracted multi-scale features into a feature vector for output. The dimension of the feature vector is usually between 20 and 50, containing various statistical quantities such as the mean, standard deviation, coefficient of variation, trend slope, goodness of fit, baseline change, etc. This rich feature representation provides a comprehensive information base for subsequent fusion analysis.

[0069] The dual-source fusion unit receives the theoretically calculated values output by the physical constraint modeling unit and the measured feature vectors output by the data feature extraction unit, establishes a correlation between the two and performs a comparison and matching. The core idea of fusion is to verify whether the prediction of the physical model is consistent with the actual measurement. When the two deviate, the deviation information is recorded for subsequent correction.

[0070] The first step of dual-source fusion is to extract features from the theoretically calculated values, making them comparable to the measured feature vectors. The physical constraint modeling unit calculates the instantaneous values of pressure and temperature at each time. The dual-source fusion unit also calculates statistical features such as mean, standard deviation, and trend slope for these theoretical value sequences, generating a theoretical feature vector. Then the system compares the measured feature vector and the theoretical feature vector dimension by dimension, calculating the deviation of each dimension.

[0071] The deviation is calculated using relative error rather than absolute error. For the pressure mean, if the measured mean is 10.2 MPa and the theoretical mean is 10.0 MPa, the absolute deviation is 0.2 MPa and the relative deviation is 2%. The reason for using relative deviation is that the numerical scales of different physical quantities differ greatly, and relative deviation can uniformly measure the deviation degree of different features. The system sets a deviation weight for each feature dimension, reflecting the importance of the feature to the gas cylinder state judgment. The weight of the pressure mean is set higher, for example 0.3, because pressure is the core indicator of the gas cylinder state. The weight of the short-term fluctuation standard deviation is set lower, for example 0.1, because short-term fluctuations are greatly affected by random factors.

[0072] The dual-source fusion unit calculates the weighted deviation value as a measure of the overall deviation degree. The comprehensive deviation value is obtained by multiplying the relative deviation of each dimension by the corresponding weight and then summing. If the comprehensive deviation value is less than the set normal threshold, for example 5%, it is considered that the measured data and the theoretical model agree well, the physical model can accurately describe the gas cylinder state, and the system continues to operate normally. If the comprehensive deviation value exceeds the normal threshold but is less than the abnormal threshold, for example between 5% and 15%, it is considered that there is a certain deviation, and the correction mechanism needs to be started. If the comprehensive deviation value exceeds the abnormal threshold, for example greater than 15%, it may indicate that a significant abnormal event has occurred in the gas cylinder, and the system immediately triggers an alarm.

[0073] When a deviation is detected, the dual-source fusion unit records detailed deviation feature information, including the time when the deviation occurs, the most significant feature dimension of the deviation, the size and sign of the deviation, the duration of the deviation, etc. These information are packaged into a deviation feature vector and passed to the subsequent deviation decomposition attribution unit to provide a basis for accurate correction.

[0074] The physical constraint modeling unit needs to read the physical parameters of the gas cylinder and call the fluid mechanics solver when calculating the theoretical values. This embodiment describes the technical details of physical modeling in depth.

[0075] The acquisition of physical parameters is divided into static parameters and dynamic parameters. Static parameters are inherent properties of the cylinder and gas, which are configured at system initialization and remain unchanged. These parameters are stored in the system's configuration database, including cylinder volume, gas type, gas molar mass, and gas constant. Taking a 40-liter steel cylinder storing oxygen as an example, the static parameters read by the system are volume 40 liters, gas type oxygen, molar mass 32 grams per mole, and gas constant 260 joules per kilogram kelvin. These parameters are valid throughout the entire service life of the cylinder.

[0076] Dynamic parameters are parameters that may change with the use of the cylinder, mainly including the heat transfer coefficient and the compression factor correction term. The heat transfer coefficient describes the ability of heat exchange between the cylinder and the environment, which is affected by many factors such as bottle wall material, surface coating, environmental wind speed, etc. The heat transfer coefficient of a new cylinder can be estimated according to the empirical formula, but it will gradually change with the oxidation and dust accumulation on the surface of the cylinder. The compression factor correction term is used to correct the deviation of the actual gas from the ideal gas behavior, which changes with pressure and temperature, and the deviation is more significant under high pressure or low temperature conditions. These dynamic parameters use standard values at the beginning of system operation, and are dynamically updated by the equivalent mapping dynamic updating unit according to actual data in the subsequent running process.

[0077] The physical constraint modeling unit calculates the theoretical pressure value and the theoretical temperature value at the next time based on the current pressure and temperature state and the above-mentioned physical parameters. The calculation process is divided into two sub-steps.

[0078] The first sub-step calculates the theoretical state of the gas in the closed cylinder. Assuming that the current pressure in the cylinder is 10 megapascals, the temperature is 20 degrees Celsius, i.e. 293 kelvin, the cylinder volume is 40 liters, and the gas stored is oxygen. According to the gas state equation, the mass of oxygen in the cylinder can be calculated. Since the cylinder is closed, the mass of the gas remains unchanged at adjacent times. If the ambient temperature remains unchanged and there is no heat exchange in the cylinder at the next time, the temperature in the cylinder should remain at 293 kelvin, and the corresponding pressure should also remain at 10 megapascals. This is an ideal isothermal and isochoric process.

[0079] However, in reality, the cylinder exchanges heat with the environment, and the second sub-step calculates the temperature change caused by heat exchange. According to Fourier's law of heat conduction, the heat flow between the cylinder and the environment is proportional to the temperature difference and the heat transfer coefficient. Assuming the ambient temperature is 25 degrees Celsius, i.e. 298 Kelvin, and the cylinder temperature is 293 Kelvin, the temperature difference is 5 Kelvin. The surface area of the cylinder is about 0.5 square meters, and the heat transfer coefficient is taken as a typical value of 10 watts per square meter Kelvin, so the heat flow is 25 watts. These heat is transferred into the cylinder in one second, causing the gas inside the cylinder to absorb 25 joules of heat. According to the specific heat capacity of the gas, the magnitude of the temperature rise of the gas inside the cylinder caused by these heat can be calculated. The specific heat capacity of oxygen under constant volume conditions is about 650 joules per kilogram Kelvin, and the mass of oxygen inside the cylinder is about 0.5 kilograms, so absorbing 25 joules of heat causes the temperature to rise by about 0.08 Kelvin. Therefore, the theoretical value of the temperature inside the cylinder at the next moment is 293.08 Kelvin.

[0080] The temperature change will cause a corresponding change in pressure. According to Charles' law, the pressure of a gas under constant volume conditions is proportional to the temperature. The temperature rises from 293 Kelvin to 293.08 Kelvin, with a temperature rise ratio of 0.027%, and the pressure also increases by 0.027%, from 10 megapascals to 10.0027 megapascals. This is the theoretical state calculated by the physical constraint model at the next moment.

[0081] For more complex scenarios, the physical constraint modeling unit calls a fluid mechanics solver for numerical calculation. When there is a temperature gradient or gas flow inside the cylinder, a simple lumped parameter model is no longer applicable, and a system of partial differential equations describing fluid motion and heat transfer needs to be solved. The present invention uses computational fluid dynamics methods to divide the cylinder interior into a grid, and solves the mass conservation, momentum conservation and energy conservation equations at the grid nodes.

[0082] Considering that the cylinder is usually axisymmetric, the present invention simplifies the three-dimensional problem to a two-dimensional problem using the axisymmetric assumption, significantly reducing the computational load. The cylinder interior is divided into a number of annular grid elements, and the physical quantities within each element are assumed to be uniformly distributed. The density of the grid affects the calculation accuracy and calculation time, and the present invention uses a non-uniform grid, with denser grids near the bottle wall to capture the boundary layer effect and sparser grids in the central region to improve efficiency. A typical grid configuration has 20 radial nodes, 50 axial nodes, and a total of 1000 grid elements.

[0083] The fluid mechanics solver uses the finite volume method for discretization, converting the partial differential equations into a system of algebraic equations. The convection term uses a second-order upwind scheme to ensure numerical stability, and the diffusion term uses a central difference scheme to ensure accuracy. The implicit scheme is used for time advancement, allowing the use of larger time steps without losing stability. The pressure-velocity coupling uses the SIMPLE algorithm, which iteratively solves the velocity field and pressure field through the pressure correction equation.

[0084] In actual operation, the fluid dynamics solver does not perform full three-dimensional flow calculation at every time step, but is selectively activated as needed. When the cylinder is in a stable state, the pressure and temperature change gently, and the simplified lumped parameter model is sufficient to be accurate, at this time the solver is in a dormant state. When a large fluctuation in pressure or temperature is detected, or the cylinder is filled or discharged, the solver is activated to perform detailed calculations. This adaptive calculation strategy balances the needs of accuracy and efficiency.

[0085] The output of the physical constraint modeling unit is the theoretical pressure value, the theoretical temperature value, and the temperature field distribution, pressure field distribution, and velocity field distribution inside the cylinder at the current time. These data are extracted and compared with the measured data to support the operation of the entire digital twin system.

[0086] The data feature extraction unit maintains three data buffers with different time spans, which are used to extract short-term fluctuation features, medium-term trend features, and long-term drift features. This embodiment describes the data structure and feature calculation method of the buffer in detail.

[0087] The short-term buffer uses a circular queue data structure, and the queue length is fixed at 60 elements, corresponding to the sampling data of the last minute. When new data arrives, the tail element is enqueued, and if the queue is full, the head element is dequeued, always keeping the latest 60 data points in the queue. This circular queue realizes the sliding effect of a fixed window, without the need to frequently allocate and release memory, and has high running efficiency.

[0088] Each element in the short-term buffer stores multi-dimensional data at a sampling time, including timestamp, pressure value, temperature value, gas concentration value, etc. Taking the pressure parameter as an example, the system calculates the statistical features of the 60 pressure values in the buffer. The mean value is obtained by dividing the sum of all pressure values by 60, reflecting the average level of the pressure in the last minute. The standard deviation is obtained by calculating the average of the square sum of the deviation of each pressure value from the mean value, and then taking the square root, reflecting the fluctuation amplitude of the pressure. In a preferred embodiment of the present invention, if the mean value of the 60 pressure values is 10.15 MPa and the standard deviation is 0.03 MPa, the coefficient of variation is 0.03 divided by 10.15, which is approximately 0.003, i.e. 0.3%. This indicates that the pressure is relatively stable in the last minute, with very small fluctuation amplitude.

[0089] Another important indicator of short-term fluctuation characteristics is the peak-to-peak value, which is the difference between the maximum and minimum values. If the maximum pressure value in 60 sampling points is 10.21 MPa and the minimum value is 10.09 MPa, the peak-to-peak value is 0.12 MPa. The peak-to-peak value reflects the range of parameter changes in the short term and is an intuitive indicator of whether there is rapid fluctuation. For a stable running cylinder, the pressure peak-to-peak value is usually less than 0.1 MPa. If the peak-to-peak value suddenly increases to more than 0.5 MPa, it may indicate that the cylinder is impacted or improperly operated.

[0090] The medium-term buffer stores a large amount of data, so a segmented compression storage strategy is adopted. The original sampling frequency is once per second, and there are 3600 data points in one hour, which occupies a lot of memory if stored in full. The present application compresses data in 1 minute segments, retaining the maximum, minimum and average values of each segment. This way, one hour of data is compressed into 60 segments, each with 3 values, for a total of 180 data points, with a compression ratio of 20:1. This compression retains the main characteristics of the data while significantly reducing storage requirements.

[0091] The medium-term trend characteristics are calculated by linear regression fitting of the compressed data. The system takes the average values of the 60 segments as the dependent variable and time as the independent variable, and fits a straight line using the least squares method. The slope of the fitted line reflects the rate of change of the parameter over time and is a key indicator of the direction and strength of the trend. Taking temperature data as an example, if the temperature rises from 18 degrees Celsius to 22 degrees Celsius in one hour, the total temperature rise is 4 degrees Celsius, and the slope of the fitted line is about 0.067 degrees Celsius per minute. This positive slope indicates an upward trend in temperature.

[0092] The goodness of fit is measured by calculating the coefficient of determination, which is equal to the regression sum of squares divided by the total sum of squares, with a value range of 0 to 1. The closer the coefficient of determination is to 1, the closer the data points are to the fitted line, and the more obvious the trend. If the goodness of fit of the temperature data is 0.95, it means that the temperature change is highly consistent with the linear trend, which may correspond to a scenario of stable rising ambient temperature or continuous heat absorption of the cylinder. If the goodness of fit is only 0.3, it means that the temperature fluctuates greatly and does not show a clear linear trend, which may correspond to a scenario of frequent changes in ambient temperature or intermittent use of the cylinder.

[0093] The long-term buffer covers a week of time span, and if the original sampling frequency is used, 600,000 data points need to be stored, which is obviously not practical. The present application uses a larger compression ratio, taking one representative point every 5 minutes, 288 points per day, and 2016 points per week. The selection of representative points uses the median rather than the average, which is not sensitive to outliers and can better reflect the central tendency of the data.

[0094] The long-term drift feature is calculated by baseline comparison method. The system divides one week data by day, and calculates the total median of 288 data points per day as the baseline value of that day. Then the trend of baseline value of each day is calculated. If the pressure baseline of the first day is 10.00 MPa, the second day is 9.98 MPa, and the third day is 9.96 MPa, showing a continuous downward trend, and the baseline decreases to 9.76 MPa by the seventh day, a total of 0.24 MPa is decreased in seven days. This long-term slow decline is likely to correspond to a small amount of leakage. A single day pressure drop of 0.03 to 0.04 MPa is within the range of short-term fluctuations and is difficult to detect, but the cumulative effect of seven days is significant, and the long-term drift feature successfully captures this anomaly.

[0095] The features extracted from the three buffer zones respectively reflect different time scale characteristics of the state of the gas cylinder. The short-term feature corresponds to transient disturbance, the medium-term feature corresponds to gradual trend, and the long-term feature corresponds to slow evolution. The comprehensive analysis of such multi-scale features greatly improves the comprehensiveness and accuracy of anomaly identification, avoiding the one-sidedness that may be caused by single time scale analysis.

[0096] The gas cylinder digital twin module 3 also includes a deviation decomposition attribution unit and a hierarchical correction strategy unit. These two units realize accurate attribution and targeted correction of virtual and real deviations.

[0097] The deviation decomposition attribution unit receives the deviation information recorded by the dual-source fusion unit, first calculates the comprehensive deviation index. The comprehensive deviation index is a scalar value that compresses the deviation of multiple dimensions into a single index for easy judgment. The calculation method is to weight and average the relative deviations of each dimension. In an embodiment of the present invention, the feature dimensions include pressure mean, temperature mean, pressure standard deviation, temperature standard deviation, pressure trend slope, temperature trend slope, etc. a total of 20 dimensions. The weight of each dimension is determined according to its importance to the judgment of the state of the gas cylinder, and the sum of all weights is normalized to 1. The weight of the pressure mean is set to 0.25, the weight of the temperature mean is set to 0.2, the weight of the standard deviation type feature is 0.08, and the weight of the trend slope type feature is 0.1.

[0098] Suppose the relative deviation of the pressure mean at the current time is 3%, the relative deviation of the temperature mean is 2%, the relative deviation of the pressure standard deviation is 10%, and the deviation of the remaining dimensions is less than 5%, then the comprehensive deviation index is calculated as 0.25 times 3% plus 0.2 times 2% plus 0.08 times 10% plus other items, which is approximately equal to 2.9%. This comprehensive deviation value falls within the normal range, and the system determines that the current deviation belongs to acceptable small fluctuations and does not need to be corrected.

[0099] The core of bias attribution is to statistically analyze the time evolution of the bias. The bias decomposition attribution unit maintains a bias history database, which records the integrated bias value and the bias value of each dimension at each time point in the past period. The length of the history window is set to the last 1 hour, corresponding to 3600 time points. The system performs various statistical tests on this time series to determine whether the bias is a random error, a systematic cumulative bias, or an abnormal bias.

[0100] The determination of random error is based on autocorrelation test. The system calculates the autocorrelation function of the bias time series, and if there is no significant correlation between the bias values at different times, i.e., the autocorrelation coefficient is close to zero, it is determined to be a random error. Mathematically, a random error sequence should exhibit white noise characteristics, with each time point's bias value being independent and identically distributed. For a bias sequence of 3600 time points, the system calculates the autocorrelation coefficients with a lag of 1 to 10 steps. If the absolute values of these autocorrelation coefficients are all less than 0.1, the randomness test is passed. At this time, the bias is attributed to random factors such as sensor measurement noise and numerical calculation rounding error, and there is no need to correct the model parameters.

[0101] The determination of systematic cumulative bias is based on trend test. The system performs linear regression fitting on the bias time series, calculates the slope of the trend line and the goodness of fit. If the slope is significantly not zero and the goodness of fit is high, it indicates that the bias shows a monotonic increasing or decreasing trend, and it is determined to be a systematic bias. The specific criteria are set as the absolute value of the trend slope greater than 0.5% per hour and the goodness of fit greater than 0.6. For example, the bias gradually increases from 1% an hour ago to 3% now, with a cumulative increase of 2 percentage points and a goodness of fit of 0.8, which is obviously a systematic cumulative bias. This bias is usually caused by slight inaccuracies in physical model parameters, such as estimation bias of heat transfer coefficient, correction term bias of compression factor, etc. As the running time increases, the cumulative effect of small parameter bias leads to the gradual deviation of model prediction from actual measurement.

[0102] The determination of abnormal bias is based on mutation test. The system uses a sliding window comparison method to calculate the difference between the current bias and the average bias of a certain period of time. If the difference exceeds the set threshold, it is determined to be a mutation. The threshold is set to 3 times the standard deviation of the previous bias, which is based on the 3σ criterion in statistics, and the probability of normal fluctuation exceeding 3 times the standard deviation is extremely small. Assuming that the bias fluctuates around 2% for the past 10 minutes, with a standard deviation of 0.3%, and the current bias suddenly jumps to 5%, with a jump amplitude of 3%, which is much higher than the 0.9% of 3 times the standard deviation, triggering mutation detection. This mutation may correspond to an abnormal event such as gas cylinder leakage, abnormal opening of valve, sudden change of environmental temperature, etc.

[0103] After the deviation type determination is completed, the deviation decomposition attribution unit labels the current deviation with an attribution tag. The tag contains three elements: deviation type, deviation severity, and recommended correction strategy. The deviation type is one of random error, systematic cumulative deviation, or abnormal deviation. The deviation severity is divided into three levels: mild, moderate, and severe. According to the size of the comprehensive deviation index, less than 5% is mild, 5% to 15% is moderate, and greater than 15% is severe. The recommended correction strategy is mapped according to the deviation type. Random error is recommended to be uncorrected, systematic deviation is recommended to be parameter fine-tuned, and abnormal deviation is recommended to be quickly responded.

[0104] The hierarchical correction strategy unit receives the deviation information with the attribution tag and performs the corresponding correction operation according to the tag indication. For random error, the unit does not perform any model parameter adjustment, but reduces the influence of random fluctuations by increasing the data sampling frequency or lengthening the smoothing window length of feature extraction. For example, the pressure sampling frequency is temporarily increased from 1 per second to 10 per second, and the average value of 10 measurements is taken as the representative value of the second, which is equivalent to 10 times square root noise reduction on measurement noise.

[0105] For systematic cumulative deviation, the hierarchical correction strategy unit starts the physical parameter adaptive adjustment mechanism. The core idea of this mechanism is to find the model parameter that needs to be adjusted the most and make a small correction in the direction of reducing the deviation. First, the system performs sensitivity analysis to evaluate the influence of each adjustable parameter on the model output. The adjustable parameters usually include the heat transfer coefficient, the compression factor correction term, the convective heat transfer coefficient, etc. Sensitivity analysis is achieved by the perturbation method, which increases a certain parameter by 1%, re-runs the model calculation, and observes the change in the output. The larger the change, the more sensitive the parameter is to the output.

[0106] In an embodiment of the present application, it is assumed that the current deviation is mainly reflected in the temperature feature, and the theoretical value of the temperature is continuously about 1 degree Celsius lower than the measured value. The system performs sensitivity analysis on the heat transfer coefficient and finds that increasing the heat transfer coefficient by 10% will increase the calculated temperature by about 0.8 degrees Celsius, with a sensitivity coefficient of 0.08 degrees Celsius per percentage point. In contrast, the compression factor correction term has little effect on the temperature. Therefore, the system determines that the heat transfer coefficient is the parameter that needs to be adjusted the most.

[0107] The parameter adjustment adopts the gradient descent idea, and the adjustment direction is the direction of reducing the deviation. The adjustment step is adaptively determined according to the size of the deviation. The current temperature deviation is 1 degree Celsius, and the sensitivity coefficient is 0.08 degrees Celsius per percentage point. In theory, it needs to adjust the heat transfer coefficient by 12.5 percentage points to completely eliminate the deviation. However, considering the nonlinearity of the model and the coupling effect between parameters, a single large adjustment may cause system instability. The present application adopts a conservative strategy, and the single adjustment amplitude does not exceed 30% of the theoretical requirement, that is, the heat transfer coefficient is adjusted by about 4 percentage points this time. If the initial value of the heat transfer coefficient is 10 watts per square meter kelvin, the adjusted value is 10.4 watts per square meter kelvin.

[0108] After parameter adjustment, the system returns the new parameter value to the physical constraint modeling unit, triggering the model to recalculate. The new theoretical value is obtained by recalculation, and the deviation is calculated by comparing with the measured value again. If the deviation is significantly reduced, it means that the adjustment direction and amplitude are reasonable, and the system enters the next round of operation. If the deviation does not decrease or even increases, it means that the adjustment direction is wrong or the amplitude is too large, and the system reverts to the parameter value before adjustment and reduces the step to try again. This exploratory parameter adjustment combined with feedback verification ensures that each adjustment is moving in the right direction.

[0109] For abnormal deviation, the hierarchical correction strategy unit starts the fast response mode. Abnormal deviation usually corresponds to a sudden event, and the model needs to quickly track the mutation of the physical entity. At this time, the degree of relaxation of the physical constraint is relaxed, and the model parameters are allowed to be adjusted by a large amplitude. At the same time, the system directly learns the mutation characteristics of the measured data, and quickly updates the mapping relationship through incremental learning mechanism. Specifically, the system extracts the feature comparison before and after the mutation, and uses this comparison as a sample for reinforcement learning, greatly increases the learning weight of this sample, and makes the model quickly adapt to the new state mode.

[0110] The triggering of the fast response mode is accompanied by the generation of warning information. The system sends an abnormal deviation alarm to the abnormal warning module 4, and pushes a notification to the gas cylinder application terminal, informing the management personnel that the gas cylinder may have an abnormal event, and suggesting on-site verification. This human-machine cooperation mechanism retains the channel for manual intervention while automatically correcting, fully utilizing the experience and flexibility of human judgment.

[0111] The physical parameter adaptive adjustment mechanism in the hierarchical correction strategy unit includes parameter sensitivity matrix calculation, optimal parameter adjustment vector solution and safety constraint limitation. This embodiment details the mathematical principles and implementation details of this mechanism.

[0112] The parameter sensitivity matrix describes the mapping between model parameters and output characteristics. Assuming the model has n adjustable parameters and m output characteristics, the sensitivity matrix is an m by n matrix. The element in the i-th row and j-th column represents the change of the i-th characteristic caused by a unit change of the j-th parameter. The sensitivity matrix is calculated using numerical differentiation method. A small perturbation is applied to each parameter in turn, and the response of each output characteristic is observed.

[0113] In one embodiment of the application, the adjustable parameters include three: heat transfer coefficient, compression factor correction term, and convective heat transfer coefficient; the output characteristics include two: pressure mean value and temperature mean value; and the sensitivity matrix is 2 by 3. The calculation process is as follows: first, keep all parameters unchanged, run the model to get the baseline output, pressure mean value 10.0 megaPascal, temperature mean value 20 degrees Celsius. Then increase the heat transfer coefficient from 10 watts per square meter Kelvin to 10.1, an increase of 1%, and keep other parameters unchanged, run the model again to get the pressure mean value 10.0 megaPascal, temperature mean value 20.05 degrees Celsius. It can be seen that an increase of 1% in the heat transfer coefficient increases the temperature mean value by 0.05 degrees Celsius, and has almost no effect on the pressure mean value. Therefore, the element in the 2nd row and 1st column of the sensitivity matrix is 0.05 degrees Celsius per hundredth, and the element in the 1st row and 1st column is 0. Similarly, other elements are calculated.

[0114] The complete sensitivity matrix may show that the heat transfer coefficient mainly affects the temperature characteristic, the compression factor correction term mainly affects the pressure characteristic, and the convective heat transfer coefficient has a moderate effect on both. This matrix reveals the causal relationship between parameters and characteristics, providing a quantitative basis for parameter adjustment.

[0115] The solution of the optimal parameter adjustment vector is an optimization problem. The optimization goal is to minimize the deviation between the adjusted model output and the measured value. Mathematically, given the current deviation vector, solve the parameter adjustment vector so that the product of the sensitivity matrix and the adjustment vector is equal to the negative deviation vector. This is a linear equation system solving problem.

[0116] When the number of parameters is equal to the number of characteristics, the sensitivity matrix is a square matrix, and if the square matrix is invertible, the equation system has a unique solution, which can be directly inverted. When the number of parameters is greater than the number of characteristics, the equation system is underdetermined, and there are infinitely many solutions. At this time, the least norm solution is adopted, that is, among all the solutions that meet the conditions, the one with the smallest adjustment amplitude is selected to maintain the stability of the system. When the number of parameters is less than the number of characteristics, the equation system is overdetermined, and there may be no exact solution. At this time, the least squares solution is adopted, which minimizes the sum of the squares of the residuals.

[0117] The solving method adopted by the present application is flexibly selected according to specific conditions. For a simple scene, an explicit solving method is directly calculated. For a complex scene, an iterative optimization method such as the steepest descent method or the conjugate gradient method is used to gradually approach the optimal solution. The iterative formula of the steepest descent method is: the next parameter is equal to the current parameter minus the step size multiplied by the deviation gradient. The deviation gradient, that is, the transpose of the sensitivity matrix multiplied by the deviation vector, points to the direction in which the deviation increases the fastest, and the negative direction is the direction in which the deviation decreases the fastest. The step size is a key parameter for controlling the adjustment amplitude of each step, and a too large step size may cause oscillation, and a too small step size may slow down the convergence. The present application adopts an adaptive step size strategy, and the initial step size is set to a small value such as 0.1, and is dynamically adjusted according to the change of the deviation at each step. If the deviation continues to decrease, the step size is gradually increased to speed up the convergence. If the deviation rebounds, the step size is immediately halved to ensure stability.

[0118] The safety constraint limit ensures that the parameter adjustment does not deviate the model from the physically reasonable range. Each adjustable parameter has a corresponding value range of its physical meaning. The typical range of the heat transfer coefficient is 5 to 50 watts per square meter Kelvin, and less than 5 indicates that the bottle wall is almost adiabatic, and more than 50 exceeds the general convective heat transfer capacity. The range of the compression factor correction term is usually 0.8 to 1.2, and equal to 1 corresponds to an ideal gas, and deviates from 1 indicates the actual gas effect. The system sets the upper and lower limits of each parameter when it is configured, and checks whether the adjusted parameter value is within the limited range in real time during the adjustment process.

[0119] In an embodiment of the present application, the initial value of the heat transfer coefficient is 10, the lower limit is set to 5, and the upper limit is set to 20. If the adjustment amount calculated by the optimization algorithm is 15, the heat transfer coefficient becomes 25, which exceeds the upper limit, and the system limits the adjustment amount to 10, so that the heat transfer coefficient just reaches the upper limit 20. This truncation processing ensures that the parameter will not enter an unreasonable region.

[0120] In addition, in order to avoid instability caused by a single large adjustment, the system also sets an upper limit for the single adjustment amplitude. It is stipulated that the single adjustment amount does not exceed 30% of the current parameter value. For example, if the heat transfer coefficient is currently 10, the single adjustment amount does not exceed 3, and the adjusted parameter range is 7 to 13. If the optimization algorithm suggests adjusting 5, the system will limit it to 3, and the adjustment will be completed in two steps. Although this conservative strategy of step-by-step adjustment prolongs the convergence time, it greatly improves the robustness and avoids system oscillation caused by aggressive adjustment.

[0121] After the parameter adjustment is completed, the new parameter value is stored in the system configuration database for subsequent modeling. At the same time, the system records the detailed information of this adjustment, including the adjustment time, the parameter value before adjustment, the adjustment amount, the parameter value after adjustment, the adjustment reason, the adjustment effect, etc., to form the parameter evolution history track. These historical records are used to analyze the long-term trend of the physical characteristics of the gas cylinder, and provide data support for equipment maintenance and life assessment.

[0122] The cylinder digital twin module 3 further comprises an equivalent mapping dynamic updating unit and an iterative convergence control unit. The two units work together to realize continuous learning and stable optimization of the model.

[0123] The core function of the equivalent mapping dynamic updating unit is to maintain a mapping function from the measured data feature space to the virtual model parameter space. The role of this mapping function is to automatically recommend the optimal parameter configuration for the virtual model to be adopted given the current measured features of the cylinder. The establishment of the mapping function is based on the learning of historical operation data.

[0124] The input of the mapping function is the feature vector of the measured data, including pressure mean, temperature mean, pressure fluctuation, temperature trend, and other multi-dimensional features. The output is the adjustable parameter vector of the virtual model, including heat transfer coefficient, compression factor correction term, and other parameters. The mapping function can adopt various mathematical forms, such as linear mapping, polynomial mapping, or neural network mapping. The present application preferably adopts the neural network form because it can express complex nonlinear relationships.

[0125] The neural network mapping function adopts a multi-layer perceptron structure, including an input layer, a hidden layer, and an output layer. The number of input layer neurons is equal to the dimension of the feature vector, such as 20. The number of output layer neurons is equal to the dimension of the parameter vector, such as 3. The number of hidden layer neurons is set to between 10 and 50, with 20 hidden neurons selected in this embodiment. Full connection is used between layers, the hyperbolic tangent function is selected as the activation function of the hidden layer to introduce nonlinearity, and the linear function is selected as the activation function of the output layer to output continuous parameter values.

[0126] When the mapping function is initialized, the neural network weights are initialized with small random values, at which time the mapping relationship is not accurate. The system gradually optimizes the weights through continuous learning in actual operation. Each time the hierarchical correction strategy unit completes a parameter adjustment, the equivalent mapping dynamic updating unit obtains a new training sample. The input of this sample is the measured feature vector before correction, and the target output is the optimized parameter vector after correction. This sample reflects that under this feature condition, using this parameter configuration can make the model achieve high accuracy.

[0127] The unit adds the new sample to the training sample library and starts the incremental learning process. The difference between incremental learning and offline training is that incremental learning only uses new samples to update the network weights slightly, rather than retraining all historical samples from scratch. The advantage of incremental learning is that the computational load is small and it can be performed online in real time. The disadvantage is that catastrophic forgetting may occur, i.e., new knowledge covers old knowledge. To avoid catastrophic forgetting, the present application uses the experience replay technique, which not only uses new samples but also randomly selects a number of old samples from the historical sample library for training each time incremental learning is performed.

[0128] In a preferred embodiment of the present application, 10 historical samples and 1 new sample are extracted to form a mini-batch for training each incremental learning, and the training is iterated for 3 to 5 rounds. The weight update adopts the back propagation algorithm, and the learning rate is set to 0.01, which is a small learning rate to ensure that each update does not change the weight much and maintain the stability of the model. After several incremental learning, the mapping function gradually masters the law between different characteristic working conditions and the optimal parameter configuration.

[0129] When the measured data of a new time arrives, the equivalent mapping dynamic updating unit inputs the data characteristics into the mapping function, calculates the output by forward propagation, and obtains the recommended parameter configuration. This parameter configuration is based on the experience of historical learning, and should be able to give a parameter close to the historical optimal solution for similar characteristic working conditions, avoiding the repeated trial and error process. The physical constraint modeling unit runs the virtual model with the recommended parameters, which can obtain a high-precision prediction with a high probability, reducing the workload of subsequent correction.

[0130] The learning of the mapping function is a continuous process, and the generalization ability of the mapping function is continuously enhanced as the number of samples increases. In the early stage of fewer samples, the mapping function mainly relies on interpolation, giving accurate recommendations for the working conditions covered by the training samples, and relatively conservative recommendations for the working conditions not covered. After the sample is rich, the mapping function has a certain extrapolation ability, and even if it encounters a new working condition that has never appeared in the training set, it can also make reasonable reasoning based on adjacent samples to give an acceptable parameter configuration.

[0131] The function of the iterative convergence control unit is to monitor the convergence state of the entire correction cycle to prevent infinite loop and oscillation divergence. The unit maintains a state machine, which includes five states: normal operation, correction in progress, convergence judgment, convergence success, and convergence failure. The system is initially in the normal operation state, and when the deviation decomposition attribution unit determines that correction is needed, the state machine enters the correction in progress state.

[0132] In the correction in progress state, the iterative convergence control unit monitors two key indicators in real time: the deviation change rate and the iteration number. The deviation change rate is defined as the ratio of the current deviation to the last deviation minus 1, reflecting the change amplitude and direction of the deviation. If the current deviation is 3% and the last deviation is 4%, the deviation change rate is negative 25%, indicating that the deviation has decreased by 25%. Continuous reduction of the deviation is a good signal of convergence.

[0133] The system sets a convergence criterion: the deviation remains stable and below a threshold value in three consecutive sampling periods. The definition of stability is that the absolute value of the rate of change of the deviation is less than 5%, that is, the deviation neither significantly increases nor significantly decreases. The definition of below the threshold value is that the comprehensive deviation index is less than 3%. When the two conditions are met at the same time and last for three periods, the unit determines that the correction process converges, and the state machine enters the convergence judgment state for final confirmation. After confirmation, it enters the convergence success state, the virtual model locks the current parameter configuration, and the system returns to the normal operation state.

[0134] The number of iterations is a safety mechanism to prevent infinite loops. The system sets an upper limit for the maximum number of iterations, which is set to 20 in this embodiment. If the correction process exceeds 20 iterations and still does not converge, the unit determines that the convergence fails, and the state machine enters the convergence failure state. At this time, the system reverts to the parameter configuration before correction, sends an alarm to the abnormality warning module, and marks that the gas cylinder may have complex abnormalities that are difficult to model, suggesting manual intervention for troubleshooting.

[0135] Oscillation detection is another important function of convergence control. If the deviation repeatedly increases and decreases in consecutive rounds and cannot be stabilized, it indicates that parameter adjustment may be trapped in oscillation. The detection method of the present invention is: calculate the standard deviation of the deviation of the last 5 rounds, if the standard deviation is greater than 30% of the mean value and the deviation does not show a monotonic trend, it is determined to be oscillation. Once oscillation is detected, the system immediately reduces the parameter adjustment step to half of the original, and increases the damping term. The role of the damping term is to apply a reverse resistance in the adjustment direction to suppress overshoot and oscillation. After the step size is reduced and the damping is increased, the system can usually escape from oscillation and enter stable convergence.

[0136] The design of the iteration convergence control unit is inspired by the PID control idea in automatic control theory, which makes comprehensive judgments and adjustments through three feedbacks of the current value, rate of change and cumulative value of the deviation, ensures that the correction process always advances in the direction of convergence, avoids divergence and oscillation, and finally reaches a stable state of high precision.

[0137] The abnormality warning module 4 includes a pressure temperature abnormality feature recognition unit, a pressure temperature joint analysis unit, a position data abnormality analysis unit, and a leakage degree analysis unit. These four units monitor the gas cylinder from different angles and make comprehensive judgments on whether there is an abnormality.

[0138] The pressure temperature abnormality feature recognition unit receives real-time pressure and temperature data of the gas cylinder, and judges the abnormality mode by analyzing the numerical fluctuations and fluctuation degrees of pressure and temperature. This unit focuses on the change characteristics of a single parameter itself, without considering the correlation between parameters.

[0139] For pressure parameter, the unit first determines whether the pressure value is out of the safe range. The working pressure range of the gas cylinder is determined at the design stage. For example, the working pressure range of a common nitrogen cylinder is 10 to 15 megapascals. If it exceeds 15 megapascals, it is an overpressure hazard, and if it is below 5 megapascals, it is an underpressure state. The unit checks the current pressure value in real time, and if it exceeds the range, it immediately determines that it is abnormal and triggers an alarm.

[0140] In addition to absolute value judgment, the unit also analyzes the fluctuation characteristics of the pressure. The standard deviation of the pressure in the last 10 seconds is calculated. If the standard deviation exceeds 0.3 megapascals, it indicates that the pressure is fluctuating sharply, which may correspond to frequent opening and closing of the valve, pipe vibration or other abnormal disturbances. The unit also monitors the pressure mutation, which is defined as the pressure difference between the adjacent two samples exceeding 0.5 megapascals. Under normal circumstances, the pressure change within 1 second should be very small, and the mutation usually corresponds to abnormal events such as pipe rupture, valve failure, etc.

[0141] For temperature parameter, the unit determines whether the temperature is out of the safe range. The temperature safety range of the gas cylinder is usually from minus 10 degrees Celsius to 60 degrees Celsius. If it exceeds 60 degrees Celsius, the performance of the gas cylinder material will decrease and the pressure of the gas will increase significantly. If it is below minus 10 degrees Celsius, part of the gas may liquefy. The unit also analyzes the drift of the temperature, calculates the linear trend of the temperature in the last 1 hour, and if the temperature continues to rise and the trend slope exceeds 0.1 degrees Celsius per minute, the cumulative temperature rise exceeds 6 degrees Celsius, which may indicate that the environment is overheating or the gas cylinder is exposed to the sun, etc.

[0142] The pressure temperature abnormal feature recognition unit judges the abnormality based on statistical threshold, and marks the abnormal features of pressure and temperature respectively and outputs them to the subsequent unit. The advantage of this single parameter judgment is simple and direct, and the disadvantage is that it does not consider the physical correlation between parameters, which may misjudge normal related changes.

[0143] The pressure temperature joint analysis unit receives the single parameter abnormal feature and further analyzes the relationship between the pressure and temperature changes. Based on the physical law of the gas state equation, the unit determines whether the joint change of pressure and temperature is reasonable.

[0144] According to the Charley law, under constant volume conditions, the pressure of the gas is proportional to the temperature. If the temperature rises by 10%, the pressure should also rise by about 10%. The pressure temperature joint analysis unit calculates the ratio of the pressure change rate to the temperature change rate to test whether it is close to 1. In one embodiment of the present invention, if the temperature rises from 20 degrees Celsius, i.e. 293 Kelvin, to 30 degrees Celsius, i.e. 303 Kelvin, in 1 hour, the temperature rises by 10 Kelvin, and the change rate is 3.4%. During the same period, the pressure rises from 10 megapascals to 10.34 megapascals, and the pressure rises by 0.34 megapascals, with a change rate of 3.4%. The change rates of the two are equal, which conforms to the physical law, and is determined to be a normal temperature pressure coupling response. Even if the pressure rise of 0.34 megapascals alone may trigger abnormal recognition, after joint analysis it is determined to be normal.

[0145] If the changes in pressure and temperature are not coordinated, there can be an anomaly. Suppose the temperature does not change, but the pressure drops by 1 MPa, which clearly violates the equation of state, and can correspond to a leak or a pressure gauge failure. Suppose the temperature rises by 20 degrees Celsius, but the pressure rises by only 0.1 MPa, much less than the theoretical value of 0.7 MPa, also indicating an anomaly, which can be a gas leak resulting in a mass decrease offsetting the pressure rise due to temperature rise.

[0146] The pressure-temperature joint analysis unit also considers the timing relationship of the changes. Under normal circumstances, changes in ambient temperature first cause changes in the bottle wall temperature, and then conduct to the gas inside to cause changes in pressure, with a certain time delay. The delay time depends on the thermal time constant of the gas cylinder, which is usually several minutes to several tens of minutes for steel gas cylinders. The unit identifies the delay relationship between the two by calculating the cross-correlation function of the pressure and temperature time series. If the pressure change lags behind the temperature change by 5 minutes, which is consistent with the mechanism of heat conduction, it is determined to be normal. If the two change synchronously or even the pressure change precedes the temperature change, which violates the physical causality, it indicates a measurement error or an abnormal event.

[0147] The position data anomaly analysis unit receives real-time location information of the gas cylinder and monitors the abnormal state of the location of the gas cylinder. The abnormal location of the gas cylinder mainly reflects two aspects of unauthorized movement and entry into a dangerous area.

[0148] Unauthorized movement detection is achieved by comparing the current location with the historical location. The system records the last known legal location of the gas cylinder, such as a gas storage area in a certain workshop. If the gas cylinder moves more than a certain distance, such as 50 meters, and there is no corresponding record of the gas cylinder leaving the warehouse, the unit determines that the gas cylinder has been moved without authorization, which can correspond to the theft or misoperation of the gas cylinder. The unit immediately issues a warning, marks the abnormal location of the gas cylinder, and reminds the management personnel to check.

[0149] Dangerous area detection is based on electronic fence technology. The system pre-sets areas where the gas cylinder is prohibited to enter, such as open flame operation area, high temperature area, and densely populated area, etc. The boundaries of these areas are represented by a polygon of GPS coordinates and stored in the system. The unit determines in real time whether the current location of the gas cylinder falls within the forbidden area polygon, and immediately alarms once it detects that the gas cylinder has crossed the boundary. In one embodiment of the present application, the welding workshop of the chemical experiment building is set as a forbidden area, with a coordinate range of XXX to XXX east and XXX to XXX north. When a certain oxygen cylinder GPS coordinate shows that it has entered this range, the unit triggers a high-priority alarm, because oxygen is a fuel and there is a serious safety hazard in the open flame operation area.

[0150] The position data anomaly analysis unit also analyzes the movement pattern of the gas cylinder. The normal gas cylinder movement speed will not be too fast, usually not more than 2 meters per second of walking speed. If it is detected that the gas cylinder moves at a speed of more than 10 meters per second, it may correspond to vehicle transportation, and the unit determines that it is in a transportation state. At this time, the system checks whether there is a corresponding transportation order, and if there is, it is marked as legal transportation, and if there is not, it is determined to be abnormal. The gas cylinder may be subjected to large vibrations and temperature fluctuations during transportation, so the system adjusts the abnormality judgment threshold of pressure and temperature accordingly to avoid false positives in the transportation state.

[0151] The leakage degree analysis unit is specifically designed for in-depth analysis of gas leakage, a high-risk abnormality. Leakage is the most common and dangerous abnormal event for gas cylinders, and timely and accurate detection of leakage is crucial for safety.

[0152] The basic principle of leakage detection is to monitor the abnormal drop of the pressure of the gas cylinder. The closed gas cylinder should maintain constant pressure under constant temperature conditions, and if the pressure continues to slowly drop, it indicates that there is a leak. However, in actual situations, environmental temperature fluctuations will cause corresponding changes in pressure, and the influence of temperature factors needs to be excluded. The leakage degree analysis unit first performs temperature compensation, calculates the pressure change that should be caused by temperature changes, and deducts the temperature effect from the actual measured pressure change to obtain the temperature-compensated pressure change.

[0153] In an embodiment of the present application, it is assumed that the temperature of the gas cylinder decreases from 25 degrees Celsius to 22 degrees Celsius in one hour, a decrease of 3 degrees Celsius, corresponding to a decrease from 298 Kelvin to 295 Kelvin. According to the state equation, the pressure should decrease by about 1%. If the initial pressure is 10 MPa, the temperature effect causes the pressure to decrease to 9.9 MPa. However, the actual measured pressure is 9.7 MPa, which is 0.2 MPa lower than the 9.9 MPa predicted by the temperature effect. This additional 0.2 MPa pressure drop cannot be explained by temperature changes and is determined to be caused by leakage.

[0154] The unit further estimates the leakage rate. The pressure drop of 0.2 MPa corresponds to a decrease in the mass of the gas in the cylinder of about 2%. Taking a 40-liter gas cylinder initially filled with 10 MPa of nitrogen gas as an example, the total mass of the gas is about 500 grams, and a decrease of 2% is 10 grams. The leakage occurs within 1 hour, and the average leakage rate is 10 grams per hour or 0.17 grams per minute. This is a trace amount of slow leakage that will not cause significant danger in the short term, but long-term accumulation will lead to gas depletion and severe pressure deficiency.

[0155] The unit determines the severity of the leakage according to the leakage rate. The classification criteria are set as follows: a leakage rate of less than 5 grams per hour is considered to be a mild leakage, a leakage rate of 5 to 50 grams per hour is considered to be a moderate leakage, and a leakage rate of more than 50 grams per hour is considered to be a severe leakage. A mild leakage issues a prompt-level warning, suggesting that the seal be checked during the next maintenance. A moderate leakage issues a warning-level warning, requiring it to be handled within 24 hours. A severe leakage issues an emergency-level warning, requiring the gas cylinder to be immediately deactivated and sent for repair.

[0156] For severe leakage, the unit also attempts to locate the leakage point. Although remote monitoring is difficult to directly locate the physical location of the leakage point, but can infer the type of leakage by data characteristics. If the pressure drops while the temperature drops locally, it may be the cooling effect of gas expansion near the leakage point, suggesting that the leakage point is near the temperature sensor. If the gas concentration sensor detects an abnormal increase in gas concentration at the cylinder valve, it suggests that the leakage may occur at the cylinder valve connection. These inferred information is pushed to the maintenance personnel together with the early warning, helping them to quickly locate and repair the leakage.

[0157] The abnormal information output by the four units of the abnormal early warning module 4 is summarized and comprehensively judged. Only when multiple units determine abnormality at the same time and the abnormal characteristics confirm each other, the system finally confirms the abnormality and issues an early warning, avoiding false alarms caused by single dimension misjudgment. At the same time, the system grades the severity of the abnormality, and only records the log for the light abnormality without disturbing the user, sends a notification to remind the user to pay attention for the moderate abnormality, triggers multiple sound, light and electricity alarms and automatically contacts the emergency contact person for the severe abnormality, ensuring that the abnormality is responded and disposed in time.

[0158] Through the above detailed embodiment description, the technical scheme of the special gas cylinder state real-time monitoring and abnormal behavior recognition system of the present application has been fully disclosed. Based on the above-mentioned embodiment teaching, those skilled in the art can realize the technical scheme of the present application without creative labor, and achieve the expected technical effect. The above-mentioned embodiments are only preferred embodiments of the present application, and are not used to limit the protection scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements of the technical scheme of the present application made by those skilled in the art should fall within the protection scope determined by the claims of the present application.

Claims

1. A real-time monitoring and abnormal behavior identification system for special gas cylinders, characterized in that, include: The gas cylinder status data acquisition module is used to continuously and in real time acquire pressure, temperature, location information and gas concentration values ​​in the gas cylinder, with a data acquisition frequency of less than 1 second; The gas cylinder data preprocessing module is communicatively connected to the gas cylinder status data acquisition module. It is used to receive real-time data sent by the gas cylinder status data acquisition module and perform data cleaning, data transformation and data normalization on the real-time data. The data cleaning is used to filter abnormal data and duplicate records. The gas cylinder digital twin module is communicatively connected to the gas cylinder data preprocessing module. It is used to receive the data processed by the gas cylinder data preprocessing module, construct a virtual three-dimensional digital model of the gas cylinder based on the physical constraints of gas thermodynamics, realize the dual-source fusion of physical constraints and measured data by establishing an equivalent mapping relationship between the virtual three-dimensional digital model and the real-time state of the gas cylinder, and adaptively correct the equivalent mapping relationship based on the deviation decomposition and attribution mechanism. An anomaly warning module is communicatively connected to the gas cylinder digital twin module. It is used to receive digital twin data output by the gas cylinder digital twin module, compare the real-time data of the gas cylinder with the digital twin data to calculate the degree of deviation, and generate warning information when the degree of deviation exceeds a preset threshold. The gas cylinder digital twin module includes: The physical constraint modeling unit is used to determine the physical relationship between pressure, temperature and volume inside the gas cylinder based on the gas state equation. It selects the corresponding physical model parameters according to the type of stored gas, establishes a coupling relationship model between pressure and temperature, and introduces the fluid dynamics conservation equation as a supplementary constraint. The data feature extraction unit is communicatively connected to the gas cylinder data preprocessing module. It is used to receive the data processed by the gas cylinder data preprocessing module, perform multi-scale time series analysis on the processed data, and extract short-term fluctuation features, medium-term trend features, and long-term drift features. The dual-source fusion unit is connected to the physical constraint modeling unit and the data feature extraction unit respectively. It is used to receive the theoretical values ​​calculated by the physical constraint modeling unit and the measured features extracted by the data feature extraction unit, establish the correlation between the theoretical calculated values ​​and the measured data features and perform comparison and matching. When a deviation is detected, the position, size and pattern of the deviation are recorded. The gas cylinder digital twin module also includes: The deviation decomposition and attribution unit is communicatively connected to the dual-source fusion unit. It is used to receive deviation information recorded by the dual-source fusion unit, calculate a comprehensive deviation index, analyze the time evolution of the deviation, and determine the deviation type. The deviation type includes random error, systematic cumulative deviation, and abnormal deviation. The hierarchical correction strategy unit is communicatively connected to the deviation decomposition and attribution unit. It is used to receive the deviation type output by the deviation decomposition and attribution unit and perform differentiated correction operations based on the deviation type. Specifically, random errors are not corrected, a physical parameter adaptive adjustment mechanism is activated for systematic cumulative deviations, and a fast response mode is activated for abnormal deviations.

2. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The gas cylinder status data acquisition module includes a gas sensor, a temperature sensor, a location tag, and a pressure sensor. The gas sensor is used to measure the gas concentration inside the gas cylinder, the pressure sensor is used to collect the pressure value inside the gas cylinder, the temperature sensor is used to measure the temperature value inside the gas cylinder, and the location tag is used to record the location information of the gas cylinder.

3. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The physical constraint modeling unit is further used to read the physical parameters of the gas cylinder, including the gas cylinder volume, gas type, gas molar mass and gas constant. Based on the physical parameters and the gas state equation, it calculates the theoretical pressure and theoretical temperature values ​​at the current moment, and calls the fluid dynamics solver to calculate the temperature field distribution, pressure field distribution and velocity field distribution inside the gas cylinder.

4. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The data feature extraction unit maintains three data buffers with different time spans: a short-term buffer to store data from the most recent minute and extract fluctuation features, a medium-term buffer to store data from the most recent hour and extract trend features, and a long-term buffer to store data from the most recent week and extract drift features. The degree of fluctuation is quantified by calculating the standard deviation and coefficient of variation within the sliding window, and the trend slope and trend strength are calculated by linear regression fitting. The systemic shift is judged by comparing the baseline values ​​of different time periods.

5. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The deviation decomposition and attribution unit is further used to maintain a historical deviation database, record the deviation value and its decomposition results at each moment in the past period, calculate the autocorrelation function of the deviation to determine randomness, calculate the trend line fit goodness of the deviation to determine systematicity, calculate the mutation detection statistic of the deviation to determine abnormality, and assign an attribution label to the current deviation based on the analysis results. The attribution label includes deviation type, deviation severity, and suggested correction strategy.

6. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The physical parameter adaptive adjustment mechanism in the hierarchical correction strategy unit includes: calculating a parameter sensitivity matrix, which describes the amount of deviation change caused by a unit change in each adjustable parameter; calculating the optimal parameter adjustment vector using the steepest descent method or the conjugate gradient method; the adjustment amount of the parameter adjustment vector is limited by safety constraints to ensure that a single adjustment does not cause the virtual three-dimensional digital model to deviate from the physical reasonable range; and sending the adjusted parameter values ​​back to the physical constraint modeling unit to trigger the recalculation of the virtual model.

7. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The gas cylinder digital twin module also includes: The equivalent mapping dynamic update unit is communicatively connected to the hierarchical correction strategy unit. It is used to receive the result after the hierarchical correction strategy unit performs the correction operation, extract the input and output samples of this correction as training data, and update the weight coefficients of the equivalent mapping function using an online learning algorithm. The equivalent mapping function is used to calculate the parameter configuration that the virtual three-dimensional digital model should adopt based on the characteristics of the measured data. The iterative convergence control unit is communicatively connected to the equivalent mapping dynamic update unit. It is used to monitor the convergence status of the correction process, maintain a state machine including normal operation, correction in progress, convergence determination, convergence success, and convergence failure, calculate the deviation change rate and the number of iterations in real time, confirm successful convergence when the deviation remains stable and below the threshold for three consecutive sampling periods, and determine convergence failure and trigger the abnormal handling process when the number of iterations exceeds the upper limit or the deviation continues to oscillate.

8. The special gas cylinder status real-time monitoring and abnormal behavior identification system according to claim 1, characterized in that, The anomaly warning module includes: The pressure and temperature anomaly feature identification unit is used to receive real-time data from the gas cylinder, monitor the abnormal behavior of the gas cylinder in terms of pressure and temperature in real time, and determine the abnormal mode or alarm mode of the gas cylinder by the fluctuation of pressure and temperature values ​​and the degree of fluctuation. The pressure and temperature joint analysis unit is communicatively connected to the pressure and temperature anomaly feature identification unit. It is used to receive the anomaly features output by the pressure and temperature anomaly feature identification unit, make a comprehensive judgment on the relationship between pressure and temperature changes, and perform joint analysis on the abnormal behavior of the gas cylinder. The location data anomaly analysis unit is used to receive real-time location information of gas cylinders, determine the abnormal location status of gas cylinders based on their real-time location, and determine whether the location of gas cylinders is safe based on the degree of location anomaly value and warning. The leakage degree analysis unit is used to receive real-time cylinder pressure data and to judge and warn of cylinder leakage and abnormality based on the real-time leakage degree of cylinder pressure. The abnormality warning module is based on the output results of the pressure and temperature abnormal feature identification unit, the pressure and temperature joint analysis unit, the location data abnormality analysis unit and the leakage degree analysis unit. When the threshold is exceeded, it is defined as abnormal and a warning reason is generated and pushed.

Images