Digital twin system for a laboratory gas drying process

Abstract

The present application relates to the technical field of laboratory gas treatment and digital twin monitoring control, in particular to a digital twin system for laboratory gas drying process; the system maps data acquisition end, control execution end and processing end; the processing end includes environment agent data acquisition module, physical state prediction module, out-of-step entropy quantification evaluation module and adaptive safety decision module; by acquiring external environment agent sequence data, inputting physical information neural network prediction model and classical physical simulation model respectively, outputting virtual internal state parameters and confidence based on residual error; combining fluctuation characteristics to solve digital-physical out-of-step entropy and predict failure probability; comparing out-of-step safety threshold and risk threshold to output state maintenance instruction or physical backup intervention instruction, and generating simulation model failure warning signal during intervention, so as to reduce the risk of moisture breakthrough into precision instruments.

Description

Technical Field

[0001] This invention relates to the field of laboratory gas treatment and digital twin monitoring and control technology, specifically a digital twin system for laboratory gas drying processes. Background Technology

[0002] As a crucial link in ensuring the gas supply for precision analytical instruments, the laboratory gas drying process directly affects the stability of downstream instruments and the accuracy of detection results in scenarios such as high-precision detection and trace analysis. Therefore, timely assessment of the adsorption state of the dryer is key to ensuring gas supply safety. Existing monitoring methods have the following technical shortcomings. For example, passive mechanical gas dryers usually do not have electronic sensors installed inside, making it difficult to directly obtain the internal state of porous media, such as water content and saturation. If judgment is made based solely on operating time, manual inspection, or single external temperature and humidity data, it is easy to overlook the risks caused by gas station pressure pulses, high humidity disturbances in the environment, and inaccurate model predictions. Especially when the dryer is close to its limit of operation, it is difficult to identify the failure trend before the moisture exceeds the limit and trigger reliable safety intervention in a timely manner. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a digital twin system for laboratory gas drying processes. Specifically, the technical solution of this invention includes: The system includes a data acquisition end, a control execution end, and a processing end that communicates with both. The processing end includes an environmental agent data acquisition module, a physical state prediction module, a step-out entropy quantification and evaluation module, and an adaptive safety decision module. The environmental proxy data acquisition module is used to acquire the external environmental proxy sequence data of the physical entity through the data acquisition terminal; The physical state prediction module is used to input the external environment proxy sequence data into a preset physical information neural network prediction model and a preset classical physical simulation model, respectively, calculate residual data based on the output results of the two, and output the virtual internal state parameters of the physical entity and the corresponding confidence level based on the residual data. The out-of-step entropy quantification and evaluation module is used to extract the fluctuation characteristics of the external environment proxy sequence data, calculate the digital-physical out-of-step entropy based on the fluctuation characteristics and the confidence level, and predict the failure probability of the physical entity reaching the physical failure critical point based on the digital-physical out-of-step entropy. The adaptive safety decision module is used to obtain a preset out-of-step safety threshold and a preset risk threshold based on historical failure sample statistics or the physical tolerance limit calibration of the intake air moisture content by downstream precision instruments; compare the digital-physical out-of-step entropy with the preset out-of-step safety threshold, compare the failure probability with the preset risk threshold, and output a state maintenance command or physical backup intervention command to the control execution terminal according to the comparison result, and generate a simulation model failure warning signal when outputting the physical backup intervention command.

[0004] Optionally, the process of obtaining the external environment proxy sequence data of the physical entity includes: Obtain the fluid boundary state sequence at the air intake end of the physical entity and use the fluid boundary state sequence as initial proxy data; extract the high-frequency pulse features of the initial proxy data within a preset sliding time window; use the high-frequency pulse features to perform noise reduction filtering on the initial proxy data to obtain the external environment proxy sequence data.

[0005] Optionally, the process by which the physical state prediction module outputs the virtual internal state parameters and corresponding confidence levels of the physical entity includes: The external environment proxy sequence data is input into the physical information neural network prediction model, and a nonlinear dynamic prediction result is output; the external environment proxy sequence data is input into the classical physical simulation model, and a deterministic benchmark prediction result is output; the residual data between the nonlinear dynamic prediction result and the deterministic benchmark prediction result is calculated; the virtual internal state parameters are generated based on the residual data as a compensation term; the variance or standard deviation of the residual data is calculated, and the confidence level of the virtual internal state parameters is determined based on the variance or standard deviation.

[0006] Optionally, the process by which the out-of-step entropy quantization evaluation module calculates the digital-physical out-of-step entropy includes: The fluctuation characteristics of the external environment proxy sequence data are extracted, including the amplitude change rate and the frequency change rate; the confidence level of the virtual internal state parameters is obtained, the confidence level is normalized and inverted to obtain the model cognitive uncertainty; the amplitude change rate and the frequency change rate are weighted and summed according to preset weight coefficients to construct an environmental interference penalty term; the model cognitive uncertainty is multiplied by the environmental interference penalty term to calculate the digital-physical desynchronization entropy.

[0007] Optionally, the process of predicting the failure probability of the physical entity reaching the physical failure critical point based on the digital-physical out-of-sync entropy includes: constructing a Markov state transition matrix; using the digital-physical out-of-sync entropy and the virtual internal state parameters as the current state input to the Markov state transition matrix; calculating the transition probability of the current state transitioning to the physical failure critical point within a preset future time period using the Markov state transition matrix; and marking the transition probability as the failure probability.

[0008] Optionally, the process by which the adaptive safety decision module outputs a state maintenance command or a physical backup intervention command to the control execution terminal based on the comparison result includes: Step 1: Determining whether the digital-physical out-of-synchronization entropy is less than a preset out-of-synchronization safety threshold. If it is less, proceed to Step 2; if it is greater than or equal to, proceed to Step 3. Step 2: Determining whether the failure probability is less than a preset risk threshold. If it is less, outputting a state maintenance command to the control execution terminal; if it is greater than or equal to, outputting a physical backup intervention command to the control execution terminal. Step 3: Directly outputting a physical backup intervention command to the control execution terminal and generating a simulation model failure warning signal.

[0009] Optionally, the external environment proxy sequence data includes gas supply pressure fluctuation data and temperature and humidity data at the inlet of the passive mechanical gas dryer; the virtual internal state parameters include the virtual saturation of the porous medium; and the physical failure critical point is when the porous medium reaches a fully saturated state.

[0010] Optionally, the physical backup intervention command includes a forced offline regeneration command for the passive mechanical gas dryer; the state maintenance command includes a command to allow the passive mechanical gas dryer to continue operating to a preset limit saturation threshold, wherein the limit saturation threshold is less than the saturation corresponding to the fully saturated state.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. To address the problem in existing passive mechanical gas dryers that lack internal electronic sensors and struggle to directly obtain the water content and saturation of porous media, this system extracts external environmental proxy sequence data through an environmental proxy data acquisition module and synchronously inputs it into a preset physical information neural network prediction model and a classical physical simulation model. The system calculates the residual data between the nonlinear dynamic prediction results and the deterministic benchmark prediction results, using this as a compensation term to generate virtual internal state parameters reflecting the virtual saturation of the porous media, and uses the variance or standard deviation of the residual data to determine the confidence level. This mechanism can dynamically compensate for prediction deviations under complex operating conditions without relying on internal sensors, significantly improving the accuracy and confidence level of internal state assessment. 2. To address the issue that traditional methods often overlook pressure pulses at gas stations and high-humidity disturbances in the environment, the out-of-step entropy quantification and evaluation module of this system can extract the amplitude and frequency change rates of external environmental proxy sequence data within a sliding time window as fluctuation features. The system constructs the extracted fluctuation features into an environmental disturbance penalty term and multiplies it with the model cognitive uncertainty derived from the confidence level of the virtual internal state parameters to calculate the digital-physical out-of-step entropy. This mechanism integrates external high-frequency water vapor impact features with the model's internal cognitive bias, effectively identifying and quantifying the synchronization risk of state deviation between the digital twin system and the physical entity. 3. To address the issue of difficulty in triggering timely safety intervention when the dryer approaches its limit of utilization, this system constructs a Markov state transition matrix, transforming the digital-physical out-of-step entropy and virtual internal state parameters into the probability of reaching the physical failure critical point within a future time period. The adaptive safety decision module compares the out-of-step entropy with the out-of-step safety threshold and the failure probability with the risk threshold in a hierarchical manner, adaptively outputting state maintenance instructions or physical backup intervention instructions, and generating early warning signals. This mechanism can proactively implement intervention and protection before the porous medium reaches the physical failure critical point of complete saturation, thereby ensuring the continuous and safe gas supply to downstream precision instruments in the laboratory. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] like Figure 1 As shown, this embodiment provides a digital twin system for a laboratory gas drying process. The system is mapped to a physical entity, which is a passive mechanical gas dryer with porous media inside. The system includes a data acquisition end, a control execution end, and a processing end that communicates with both. The processing end includes an environmental proxy data acquisition module, a physical state prediction module, a step-out entropy quantification assessment module, and an adaptive safety decision module. The environmental proxy data acquisition module is used to acquire the external environmental proxy sequence data of physical entities through the data acquisition terminal; The physical state prediction module is used to input external environment proxy sequence data into a preset physical information neural network prediction model and a preset classical physical simulation model, respectively. Based on the output results of the two, it calculates residual data and outputs the virtual internal state parameters of the physical entity and the corresponding confidence level based on the residual data. The out-of-step entropy quantification and evaluation module is used to extract the fluctuation characteristics of the external environment proxy sequence data. Based on the fluctuation characteristics and confidence level, it calculates the digital-physical out-of-step entropy and predicts the failure probability of the physical entity reaching the physical failure critical point based on the digital-physical out-of-step entropy. The adaptive safety decision module is used to obtain preset out-of-step safety thresholds and preset risk thresholds based on historical failure sample statistics or the physical tolerance limit calibration of intake air moisture content by downstream precision instruments. It compares the digital-physical out-of-step entropy with the preset out-of-step safety thresholds and the failure probability with the preset risk thresholds. The preset out-of-step safety thresholds and preset risk thresholds are obtained during the system debugging phase based on historical failure sample statistics or the physical tolerance limit calibration of intake air moisture content by downstream precision instruments. Based on the comparison results, it outputs state maintenance commands or physical backup intervention commands to the control execution terminal, and generates simulation model failure warning signals when outputting physical backup intervention commands.

[0015] This embodiment provides an overall operation mechanism for a digital twin system of a laboratory gas drying process. Specifically, the gas drying process in a high-precision analytical laboratory is taken as a unified application scenario: the centralized gas station in the plant supplies gas to the precision analytical instruments in the laboratory through pipelines with a length greater than a preset length threshold. A passive mechanical gas dryer is installed in the gas path to adsorb moisture and impurities. Since the dryer does not have electronic sensors inside, the system acquires external proxy data through data acquisition terminals set on the air inlet side and the environment side, and establishes a digital twin corresponding to the dryer through the processing terminal. Then, the control execution terminal performs operations such as continuing operation, alarm, shutdown switching, or offline regeneration. Specifically, the environmental proxy data acquisition module does not collect the actual moisture content inside the dryer, but rather the external boundary conditions that reflect changes in its adsorption load, such as inlet pressure, inlet temperature, inlet humidity, and their time-series changes; the physical state prediction module simultaneously calls two types of models: one is a physical information neural network prediction model, used to characterize the dynamic changes under nonlinear conditions such as high humidity pulses, adsorption heat accumulation, and local capillary condensation; specifically, the physical information neural network prediction model includes a long short-term memory network or a multilayer perceptron network to extract the temporal features of the data, and its training process loss function includes physical constraint terms that integrate the moisture diffusion dynamics and thermodynamic conservation equations of porous media; Another type is the classical physical simulation model, which provides a deterministic benchmark based on the laws of adsorption thermodynamics and kinetics. Specifically, the classical physical simulation model uses a linear driving force model combined with a preset adsorption isotherm to calculate the mass transfer rate and heat exchange relationship. The residual data between the outputs of the two models reflects the degree of consistency between the digital models. When the residual is stable, it means that the confidence level of the digital twin's description of the internal state of the physical entity meets the preset confidence threshold condition. When the residual continues to expand, it means that the external operating conditions may have deviated from the classical assumptions, and the confidence level of the virtual internal state parameters has decreased. To facilitate engineering implementation, the state data in the processing end can be tagged and managed according to normal adsorption state, disturbed adsorption state, and critical adsorption state. For example, if the internal state trend output by the physical information neural network prediction model is basically consistent with the trend output by the classical physical simulation model within a certain time period, the processing end marks this state as a normal adsorption state; when the two are in the same direction but the deviation increases, it is marked as a perturbed adsorption state; when the two output trends are opposite and the deviation exceeds the preset consistency tolerance threshold and the confidence decreases, it is marked as a critical adsorption state. This labeling process does not change the model output itself, but is only used by the subsequent safety decision module to quickly identify operational risks. Furthermore, the aforementioned normal adsorption state, perturbed adsorption state, and critical adsorption state are operational labels for model consistency and operating conditions, used to describe the confidence performance of the digital twin within the current window. They are not the same hierarchical concept as the safe adsorption state, accelerated adsorption state, critical warning state, and physical failure threshold used in the subsequent Markov state transition matrix. The former is used for operational label management, while the latter is used for future failure probability evolution calculation. The processing end can first use the aforementioned operational labels to quickly screen the current window, and then combine the virtual internal state parameters and digital-physical out-of-sync entropy mapping to the Markov state, thereby avoiding semantic conflicts caused by the same state definition in a multi-level logical architecture. In case of anomalies, if a short-term communication interruption occurs at the data acquisition end, the processing end can maintain the model state within the most recent valid time window and reduce the confidence level; if the data loss continues for more than the preset sampling period, the system will not continue to rely on the extrapolation results, but will increase the risk of loss of synchronization and issue a conservative processing instruction to the control execution end; if the control execution end does not provide feedback on the execution result, the processing end can repeatedly send the instruction and generate a communication anomaly record on the device side to avoid the failure of a single transmission failure to implement the safety instruction. A passive mechanical gas dryer is installed at the front end of a gas chromatograph-mass spectrometer in an analytical laboratory. During normal operation, the system allows the dryer to operate in a range close to its limit to maintain the equipment's preset continuous operating cycle. When the gas station in the plant experiences an inlet pressure pulse due to the start-up or shutdown of the production line, and the outdoor humidity is continuously high, causing the rate of change of the inlet moisture content boundary to exceed the preset humidity change rate threshold, the processing end monitors that the classical physical simulation model still outputs a smooth adsorption process, while the physical information neural network prediction model predicts that the internal state is rapidly approaching the preset physical failure critical zone. At this point, the residuals of the two types of models expand, the out-of-step entropy quantization evaluation module outputs a higher digital-physical out-of-step entropy, and the adaptive safety decision module no longer simply maintains the current operating state, but instead outputs physical backup intervention instructions and generates simulation model failure warning signals. The purpose of this mechanism is to connect external boundary disturbances, model consistency and safety decisions into a closed loop in the dryer scenario without internal sensors, so that the digital twin system can not only estimate the internal state, but also actively trigger conservative protection when the model confidence decreases, reducing the risk of moisture breaking through into precision instruments. In this embodiment, the process of obtaining the external environment proxy sequence data of the physical entity includes: Obtain the fluid boundary state sequence at the air intake end of the physical entity and use the fluid boundary state sequence as the initial proxy data; extract the high-frequency pulse features of the initial proxy data within a preset sliding time window; use the high-frequency pulse features to perform noise reduction and filtering on the initial proxy data to obtain the external environment proxy sequence data.

[0016] This embodiment provides a mechanism for acquiring external environmental proxy sequence data. Specifically, in the above-mentioned laboratory gas drying process, if all pressure, temperature, and humidity readings collected at the air inlet are directly input into the subsequent model, sensor spikes, instantaneous valve opening and closing disturbances, and real water vapor pulses are easily mixed, causing the digital twin to misjudge the measurement noise as changes in physical adsorption load. Therefore, this embodiment introduces high-frequency pulse feature extraction and noise reduction filtering within a sliding time window before the initial proxy data enters the prediction stage. Specifically, the fluid boundary state sequence may include the inlet pressure sequence, inlet temperature sequence, inlet relative humidity sequence, and sampling timestamps; a sliding time window is used to observe local changes in data over continuous time; for example, continuous sampling points can be divided into continuous sliding windows. , and ,like If only a single isolated peak appears in internal pressure and humidity, and no change in the same direction is observed in adjacent windows, then the peak is more likely to be caused by electromagnetic interference or transient sampling error. like and If a continuous increase in humidity is accompanied by pressure fluctuations, this change is more likely to reflect the actual water vapor pulse in the gas source. Based on this, the processing end retains fluctuations that have physical significance for the adsorption load and suppresses noise that is less related to the evolution of water content inside the dryer. In abnormal processing conditions, if the sliding time window is too short, the system may misjudge the actual short-term water vapor impact as noise; if the window is too long, it may weaken the rapid changes in the initial stage of the high-intensity water vapor pulse in the gas source. Therefore, the window length can be preset according to the gas path volume, sampling frequency, and instrument protection level, and allows maintenance personnel to perform calibration during the system debugging phase; if the data of a certain type of sensor at the acquisition end is obviously out of bounds or the timestamp is out of order, the processing end can temporarily remove that channel, using only the remaining effective channels to maintain the external environment proxy sequence, and simultaneously reducing the subsequent confidence level; in the above-mentioned analytical laboratory, the aging of the gas station pipeline caused the intake pressure to have a momentary spike when the valve group switched; the system found that the spike was not accompanied by the same direction of humidity and temperature change within the sliding window, so it was filtered as a low physical correlation disturbance; The combination of persistent high humidity weather conditions and sudden changes in factory gas consumption resulted in increased humidity and stronger pressure fluctuations over multiple consecutive windows. The system retains this change as a proxy sequence of external environmental data, which is then used by the subsequent prediction module to identify the actual increase in the dryer's adsorption load. The purpose of this step is to organize the original fluid boundary data at the inlet end into a proxy sequence that better reflects the dryer's adsorption load, reduce the interference of high-frequency redundant noise on the digital twin state assessment, and retain high-frequency water vapor impact information that has physical guidance significance for safety decisions. In this embodiment, the process by which the physical state prediction module outputs the virtual internal state parameters of a physical entity and the corresponding confidence level includes: External environment proxy sequence data is input into a physical information neural network prediction model, which outputs nonlinear dynamic prediction results; external environment proxy sequence data is input into a classical physical simulation model, which outputs deterministic benchmark prediction results; residual data between the nonlinear dynamic prediction results and the deterministic benchmark prediction results are calculated; virtual internal state parameters are generated based on the residual data as compensation terms; the variance or standard deviation of the residual data is calculated, and the confidence level of the virtual internal state parameters is determined based on the variance or standard deviation.

[0017] This embodiment provides a mechanism for generating virtual internal state parameters and corresponding confidence levels. Specifically, after filtering external proxy data, the system has the following technical limitations: the silica gel and molecular sieve inside the dryer are in a closed tube, and their local water absorption, adsorption heat, and saturation front advancement cannot be directly measured; if only the classical adsorption model is relied upon, the model may be too smooth in the near-saturation stage; if only the data-driven model is relied upon, there may be a lack of physical constraints when the operating conditions exceed the training distribution. Therefore, this embodiment outputs two types of models in parallel and uses residual data to form compensation; specifically, after receiving the external environment proxy sequence, the physical information neural network prediction model outputs prediction results that reflect nonlinear dynamic changes, such as predicting that the internal adsorption state may undergo accelerated changes when humidity pulses are superimposed on temperature fluctuations; after receiving the same set of proxy sequences, the classical physical simulation model outputs deterministic benchmark prediction results according to the preset adsorption isotherm, mass transfer rate and heat exchange relationship. The processing end compares the two types of results to obtain residual data, and uses the residual as a compensation term to participate in the generation of virtual internal state parameters. If the fluctuation amplitude of the residual is less than or equal to the preset stability threshold, it indicates that although there are differences between the nonlinear model and the classical model, the differences have interpretable working condition compensation significance. If the fluctuation amplitude of the residual is greater than the preset stability threshold, it indicates that the consistency measurement of the internal state of the digital twin is weakened, and the confidence of the virtual internal state parameters should be reduced. For example, in a certain window, the processing end obtains the nonlinear prediction result and the baseline prediction result, and the difference between the two forms a compensation term; when the direction of change of the compensation term is consistent in consecutive windows and the fluctuation range is less than or equal to the preset tolerance threshold, the system regards the compensation term as a reasonable correction to the classic model; when the compensation term frequently jumps in the opposite direction in adjacent windows, the system identifies the phenomenon as a deterioration of model consistency and attaches a low confidence level when outputting virtual internal state parameters. In case of anomalies, if the physical information neural network prediction model outputs an abnormal null value or exceeds the physical upper limit, the processing end can temporarily use the classical physical simulation model as the benchmark and downgrade the confidence level. If the classical physical simulation model diverges due to abnormal boundary conditions, the processing end can limit its output to within the physical allowable range and increase the residual abnormal label. If neither of the two models can generate effective results, the system will not continue to provide suggestions for maintaining the normal state, but will hand over the running segment to the safety decision module for high-risk processing. In the aforementioned laboratory, after a period of operation, the dryer has entered a high utilization stage; external proxy data shows that the inlet humidity has continuous pulses. The classical physical simulation model still outputs the baseline result according to the relatively stable adsorption front, while the physical information neural network prediction model identifies that the high humidity pulse may cause local capillary condensation and outputs a nonlinear trend that approaches the critical region faster; the processing end uses the difference between the two as a compensation term to generate virtual internal state parameters that are closer to the current complex working conditions, and simultaneously calculates the confidence level according to the degree of residual stability. The purpose of this mechanism is to estimate the internal state that cannot be directly measured by combining classical physical benchmarks with nonlinear compensation, and to use residual fluctuations to reflect the confidence level of the model, so as to provide state input and uncertainty measurement basis for subsequent out-of-step entropy assessment. In this embodiment, the process of the out-of-step entropy quantization evaluation module calculating the digital-physical out-of-step entropy includes: The fluctuation characteristics of the external environment proxy sequence data are extracted, including the amplitude change rate and the frequency change rate. The confidence of the virtual internal state parameters is obtained, and the confidence is normalized and inverted to obtain the model cognitive uncertainty. The amplitude change rate and the frequency change rate are weighted and summed according to the preset weight coefficients to construct the environmental disturbance penalty term. The model cognitive uncertainty and the environmental disturbance penalty term are multiplied to calculate the digital-physical desynchronization entropy.

[0018] This embodiment provides a quantification mechanism for digital-physical desynchronization entropy. Specifically, relying solely on model residuals for state assessment based on existing virtual internal state parameters and confidence levels still has limitations: under some operating conditions, the model output remains consistent in the short term, but the external boundary conditions show a drastic fluctuation trend; or the external environmental disturbance is small, and the increase in residuals is mainly attributed to single sampling anomalies. Therefore, this embodiment simultaneously introduces external fluctuation characteristics and model cognitive uncertainty to characterize the risk intensity of decoupling between the digital twin and the physical entity. Specifically, the amplitude change rate reflects the drastic change in the magnitude of external proxy data within a short period of time. For example, a sudden rise in the humidity boundary will increase the adsorption load of porous media, and a sudden temperature change will affect the adsorption equilibrium and water vapor condensation conditions. The frequency change rate reflects the density of pulse occurrences. For example, continuous high humidity pulses will cause porous media to be impacted again before heat release and moisture diffusion are completed, thereby increasing the possibility of classical model assumptions failing. The confidence level is normalized and inverted to obtain the model cognitive uncertainty, which physically means that the system's evaluation bias of virtual internal state parameters has increased. The environmental disturbance penalty term is used to combine the amplitude and frequency of external disturbances, and together with the model cognitive uncertainty, it forms the digital-physical desynchronization entropy. To facilitate engineering implementation, confidence levels can be mapped to three levels: high, medium, and low, and external disturbances can be mapped to three levels: stable, increasing, and severe. When the confidence level is high and the external disturbance is stable, the out-of-step entropy is at a low level; when the confidence level is low and the external disturbance is severe, the out-of-step entropy is at a high level. When only one of the two is abnormal, the system places it at an intermediate risk level and continues to observe changes in adjacent windows. This hierarchical implementation can coexist with continuous numerical implementation for alarm interface display and to assist manual judgment. Furthermore, to avoid unclear boundaries during normalization and inversion in engineering implementation, the processing end can pre-save the effective upper and lower limits of the confidence level; if the original confidence level value is denoted as... The lower limit obtained from history or calibration is denoted as The upper limit is recorded as Then follow Solve for the cognitive uncertainty I of the model and restrict I to between 0 and 1; when and When the values ​​are the same or the initialization is not yet complete, the system does not perform meaningless division, but directly uses a conservative intermediate uncertainty as a temporary input. This method not only ensures the monotonic relationship that the lower the confidence level corresponds to the higher the uncertainty, but also avoids the incomparability of out-of-step entropy caused by the inconsistency of confidence scale between different batches of equipment. Furthermore, the extraction of amplitude change rate and frequency change rate can be accomplished using local statistical rules within the same sliding time window; for amplitude change rate, the processing end extracts the change amplitude between adjacent sampling points or the mean of adjacent windows; for frequency change rate, the processing end extracts the change in the number of occurrences of events exceeding the preset pulse discrimination threshold per unit time; wherein, the preset pulse discrimination threshold is a floating filter threshold set based on the historical statistical mean and standard deviation of the external environmental proxy sequence data within the sliding time window; when performing weighted summation, the weights are preset according to the laboratory's sensitivity to moisture breakthrough risk; For example, for high-precision trace analysis instruments, the weight of the humidity pulse correlation channel can be increased, while for general protection scenarios, the pressure, temperature and humidity channels can be kept relatively balanced. If a channel temporarily fails, the system can renormalize the weights according to the remaining effective channels and simultaneously lower the confidence level of the final out-of-step entropy. In abnormal handling situations, if the rate of change of amplitude within a certain window experiences a non-physical abrupt change due to sensor range switching, the system can combine the trends of adjacent windows to suppress it, avoiding amplification of out-of-step entropy at a single point of abnormality; if the confidence level is missing or has not been initialized, the processing end can use a conservative medium-low confidence level as a temporary input until the model completes a stable output; if the external environment proxy sequence is contradictory across multiple channels, for example, if the humidity display shows a sharp increase while the temperature and pressure show no response at all, the system can mark it as a data consistency anomaly and increase the weight of the safety side; In the aforementioned laboratory, during the initial stage of high humidity disturbance, the amplitude of the intake humidity pulse did not remain at its highest level, but the frequency of pulse occurrence exceeded the preset frequency threshold; the confidence level of the physical state prediction module also began to decrease; the out-of-step entropy quantification assessment module calculated the high-frequency high humidity impact together with the increase in model uncertainty, instead of just waiting for the virtual saturation to reach the limit value before alarming, thus avoiding the mapping out-of-step problem caused by the inability to represent in real time; the purpose of this step is to combine the objective intensity of environmental disturbance with the internal uncertainty of the model into a comparable risk quantity, so that the system can identify the state of decreased model consistency, rather than just identifying whether the dryer is close to saturation; In this embodiment, the process of predicting the failure probability of a physical entity reaching the physical failure threshold based on digital-physical out-of-step entropy includes: constructing a Markov state transition matrix; using the digital-physical out-of-step entropy and virtual internal state parameters as the current state input to the Markov state transition matrix; calculating the transition probability of the current state transitioning to the physical failure threshold within a preset future time period using the Markov state transition matrix; and marking the transition probability as the failure probability.

[0019] This embodiment provides a mechanism for predicting failure probability based on digital-physical out-of-sync entropy. Specifically, the out-of-sync entropy at a single moment can only indicate the synchronization risk between the current digital twin and the physical entity. However, whether the dryer will reach the failure threshold in the future depends on the current virtual internal state and the evolution trend of the subsequent state. Therefore, this embodiment constructs a Markov state transition matrix to transform the current state into the future failure probability. Specifically, the processing end can divide the dryer state into several discrete states, such as safe adsorption state, accelerated adsorption state, critical warning state, and physical failure critical point; digital-physical out-of-step entropy is used to reflect the uncertainty of the current state judgment, virtual internal state parameters are used to reflect the adsorption progress of porous media; Markov state transition matrix is ​​used to record the probability of transitioning from one state to another under a given operating condition. For example, if the system is set to four states , , , ,in Representing the physical failure critical point, the matrix contains... point to The element represents the transition probability from the accelerated adsorption state to the critical warning state, which is determined by... point to The elements represent the transition probability of entering the failure critical point; this matrix can be pre-constructed based on historical operation records, offline simulation results and safety calibration data, and can be updated during the maintenance cycle; furthermore, in order to avoid the digital-physical out-of-step entropy and virtual internal state parameters being too abstract in implementation as the current state input, the processing end can first perform joint segmentation on the two and then map them to discrete states; Specifically, the adsorption progress level can be determined first based on the range of virtual saturation, and then the synchronization risk level can be determined based on the range of digital-physical out-of-step entropy. The combination of the adsorption progress level and the synchronization risk level is mapped to the current state number in the Markov matrix. For example, when the virtual saturation is less than the first preset saturation threshold and the out-of-step entropy is less than the first preset out-of-step entropy threshold, it is mapped to S0. When the virtual saturation is greater than or equal to the first preset saturation threshold and less than the second preset saturation threshold and the out-of-step entropy is rising, it is mapped to S1 or S2. When the virtual saturation is close to complete saturation or the out-of-step entropy is continuously high, it can be directly mapped to the most conservative critical warning input state. In this way, the current state input is obtained by a traceable range judgment. Furthermore, the aforementioned safe adsorption state, accelerated adsorption state, critical warning state, and physical failure critical point specifically refer to the discrete prediction states in the Markov state transition matrix, used for probabilistic evolution within a finite future time period; they do not replace the normal adsorption state, perturbed adsorption state, and critical adsorption state operation labels used above to describe model consistency; the former is used for probability calculation, and the latter is used for operation monitoring and display, and the two can be associated one by one but should not be used interchangeably. For example, when the running label indicates a perturbed adsorption state, combining virtual saturation and digital-physical out-of-step entropy, the current Markov input state may fall into either... It is also possible that due to the high risk of synchronization, it will directly fall into Furthermore, the preset future time period is not an infinite extrapolation, but a finite prediction window that matches the sampling frequency, the dryer regeneration cycle, and the laboratory's allowable response time. For scenarios where the sampling frequency is higher than the preset frequency threshold and the protection level corresponds to the preset high safety standard, the future time period can be set to tens of minutes to several hours; for scenarios where the gas supply status change rate is lower than the preset change rate threshold, it can be appropriately extended, but should still be limited to the effective range of the model that has been verified; the processing end obtains the cumulative probability of transitioning to the physical failure critical point through matrix multiplication or equivalent iteration within this time period, and outputs it as the failure probability; this mechanism ensures that the failure probability corresponds to the probability calculation value of entering the critical point within the preset short future, rather than an unbounded extrapolation that is out of the constraints of the operating condition; In abnormal handling situations, if the current state input falls between two state boundaries, the system can use a more conservative high-risk state as the matrix input to avoid classification bias near the critical boundary; if historical data is insufficient to support a certain state transition, the system can use a conservative transition relationship given by physical simulation and expert rules; if the future time period exceeds the effective prediction range of the model, the processing end will not continue to expand the prediction, but will indicate that the current result is only applicable to the preset time range, and will hand over the safety decision to the short-term risk result to trigger. In the aforementioned laboratory, the system assesses that the current virtual internal state of the dryer is in a high utilization range, and the digital-physical out-of-sync entropy is increasing. The processing end inputs this state into the Markov state transition matrix to calculate whether it is possible to transition from the critical warning state to the physical failure critical point within the next few hours. If the transition result shows that there is a probability of entering the critical point that exceeds the tolerance within a preset future time period, the system marks the transition probability as the failure probability for the adaptive safety decision module to call. The purpose of this mechanism is to extend the current state assessment to the failure risk prediction within a limited time range, so that the system can perform early intervention before the physical breakthrough occurs. In this embodiment, the process by which the adaptive safety decision module outputs a state maintenance command or a physical backup intervention command to the control execution terminal based on the comparison result includes: Step 1: Determine whether the digital-physical out-of-synchronization entropy is less than a preset out-of-synchronization safety threshold. If it is less, proceed to Step 2; if it is greater than or equal to, proceed to Step 3. Step 2: Determine whether the failure probability is less than a preset risk threshold. If it is less, output a state maintenance command to the control execution terminal; if it is greater than or equal to, output a physical backup intervention command to the control execution terminal. Step 3: Directly output a physical backup intervention command to the control execution terminal and generate a simulation model failure warning signal.

[0020] This embodiment provides an adaptive safety decision-making mechanism. Specifically, after the aforementioned state prediction and failure probability estimation are completed, if the decision to continue operation is based solely on the virtual internal state parameters, it may still incorrectly pursue the maximum utilization rate when the model confidence decreases. If the decision to shut down is based solely on the out-of-step entropy, it may over-interrupt the experiment after a short-term disturbance. Therefore, this embodiment compares the digital-physical out-of-step entropy and failure probability in a hierarchical manner, forming a safety decision-making process that first judges the model consistency and then judges the future failure risk. Specifically, the first-level judgment is used to confirm whether the digital twin is still within the range where the confidence level meets the requirements; when the digital-physical desynchronization entropy is less than the preset desynchronization safety threshold, it indicates that the external disturbance and model uncertainty have not yet jointly formed a serious desynchronization risk, and the system enters the second-level judgment; the second-level judgment is used to confirm whether the probability of entering the physical failure critical point within the preset time period in the future is within the tolerance, provided that the model consistency meets the requirements. If the failure probability is less than the preset risk threshold, the control execution terminal receives the status maintenance instruction; if the failure probability is greater than or equal to the preset risk threshold, the system outputs the physical backup intervention instruction; if the digital-physical out-of-synchronization entropy in the first-level judgment is greater than or equal to the out-of-synchronization safety threshold, the system directly outputs the physical backup intervention instruction and generates a simulation model failure warning signal to remind maintenance personnel that the current working condition not only includes the physical failure risk of the dryer approaching saturation, but also indicates that the confidence boundary of the digital twin model has been triggered. Furthermore, to make the execution results of the two-level judgment clearer, the status maintenance command can be set to a conservative release result that allows continued operation but maintains the original monitoring cycle or increases the monitoring frequency; the physical backup intervention command can be set to a protection result that requires switching, shutdown, regeneration, or isolation of the current drying branch. When physical backup intervention is triggered in step two, the main reason is that the probability of future failure exceeds the tolerance. At this time, the system can simultaneously provide a high failure risk warning on the host computer interface. When physical backup intervention is triggered in step three, in addition to performing protection actions, a simulation model failure warning signal is also generated to clearly indicate that the model confidence boundary has been breached. Through the above distinction, intervention due to high future failure risk of equipment and intervention due to abnormal model consistency can be distinguished at the interface and record levels. Furthermore, the processing end can synchronously generate early warning information and retain event records when any physical backup intervention command is output; among them, the simulation model failure early warning signal generated by step three has a specific meaning, clearly indicating that the corresponding model confidence boundary has been breached; while when step two needs to provide synchronous prompts, it can use the same early warning signal channel or its risk prompt subtype for reporting, but the reason for recording should be marked as failure probability exceeding the threshold triggering intervention, so as to distinguish it from the out-of-step entropy exceeding the threshold triggering model failure early warning in step three; Furthermore, after outputting the state maintenance instruction, the processing end does not stop risk tracking, but enters the next sampling cycle to continue rolling updates of digital-physical out-of-sync entropy and failure probability; if multiple consecutive windows are in a low out-of-sync, low failure probability state, the maintenance strategy is maintained; if any window meets the conditions of step two or step three, the previous maintenance instruction is immediately overwritten, and physical fallback intervention is initiated; this continuous rolling mechanism ensures that the system will not ignore the sudden high humidity impact of the next window because the previous window is judged to be safe. In abnormal handling situations, for boundary conditions equal to the threshold, the system handles them in the high-risk direction. That is, when the out-of-step entropy equals the out-of-step safety threshold, physical backup intervention is triggered, and when the failure probability equals the risk threshold, physical backup intervention is also triggered. If the threshold configuration is missing, the processing end can call the factory safety configuration. If the maintenance personnel modify the threshold beyond the allowable range, the system refuses to take effect and retains the original configuration. If the control execution end is executing an uninterruptible instrument protection process, the system can first output an audible and visual alarm and a shutdown suggestion from the host computer, and then perform a switching or offline operation according to the gas path safety strategy. In the aforementioned analytical laboratory, the system allows the dryer to operate close to its limit saturation level on a daily basis. When the digital-physical out-of-step entropy is detected to be within a safe range and the probability of failure to reach full saturation in the next few hours is low, the control actuator maintains the gas supply and prompts continued monitoring. When high humidity disturbance causes the out-of-step entropy to reach the threshold, even if the virtual internal state parameters have not yet shown full saturation, the system directly triggers a physical backup intervention command and generates a simulation model failure warning signal on the monitoring interface, prompting maintenance personnel not to continue to rely on the current prediction curve to schedule delayed regeneration. The purpose of this mechanism is to manage model consistency metrics and physical failure risks separately, ensuring experimental continuity when model confidence is high and future risks are within tolerance, and prioritizing the protection of the safety of precision instruments when the model loses synchronization or future risks exceed tolerance. In this embodiment, the external environment proxy sequence data includes the gas supply pressure fluctuation data and temperature and humidity data at the inlet of the passive mechanical gas dryer; the virtual internal state parameters include the virtual saturation of the porous medium; and the physical failure threshold is when the porous medium reaches complete saturation.

[0021] This embodiment provides a physical mapping mechanism for passive mechanical gas dryers. Specifically, in the above-mentioned laboratory scenario, the physical entity is a passive mechanical gas dryer with porous media such as silica gel, molecular sieve, or a combination of both. This device is usually composed of a stainless steel tube, end connectors, filter element support structure, and adsorption packing. It does not rely on electronic control components to complete the drying process and therefore cannot directly output the internal moisture content or local saturation. Specifically, the inlet gas supply pressure fluctuation data reflects the gas flow boundary and gas station pipeline disturbance; temperature data affects water vapor partial pressure, adsorption balance and adsorption heat release; humidity data directly reflects the water load entering the dryer; the virtual saturation of the porous medium is used to represent the degree of adsorption capacity occupancy estimated by the digital twin; the physical failure threshold is set when the porous medium reaches a state of complete saturation because at this point the effective adsorption capacity of the porous medium for moisture is basically exhausted, and moisture may penetrate the dryer and enter downstream precision instruments. In case of abnormal situations, if the passive mechanical gas dryer is replaced with silica gel or molecular sieve of different specifications, the processing end should be reloaded with the corresponding adsorption capacity, regeneration conditions and model parameters; if the dryers are used in parallel or series, the system can establish multiple virtual saturations and use the unit closest to the critical state as the basis for safety decision-making; if the temperature and humidity sensor at the inlet end deviates after maintenance, the system can correct the deviation at the acquisition end by calibrating the gas or standard environmental data, otherwise the confidence level of the virtual saturation output needs to be reduced. In the aforementioned analytical laboratory, the dryer is filled with a combination of molecular sieves and color-changing silica gel. Under normal gas supply, the virtual saturation rises slowly. When the external environment brings a high-humidity gas source accompanied by gas station pressure fluctuations, the virtual saturation rises more rapidly. The system sets the fully saturated state as the physical failure critical point and uses it as the termination risk state for Markov state transition and safety decision-making. The purpose of this mechanism is to explicitly map the abstract digital twin state parameters to the porous media adsorption process of the passive mechanical dryer, so that the virtual internal state output by the system has maintainable, calibrable, and executable engineering meaning. In this embodiment, the physical backup intervention command includes a forced offline regeneration command for the passive mechanical gas dryer; the state maintenance command includes a command to allow the passive mechanical gas dryer to continue operating to a preset limit saturation threshold, wherein the limit saturation threshold is less than the saturation corresponding to the fully saturated state.

[0022] This embodiment provides an execution mechanism for state maintenance and forced offline regeneration. Specifically, in the above-mentioned passive mechanical gas dryer application, the final output of the system is not a simple prompt message, but an instruction that can be converted into a maintenance action by the control execution end. The state maintenance instruction is used to allow the equipment to continue operating to a preset limit saturation threshold when the model confidence meets the requirements and the failure probability is within the tolerance. The physical backup intervention instruction is used to forcibly terminate the current usage cycle in advance and arrange offline regeneration when the risk of loss of synchronization or failure exceeds the tolerance. Specifically, the limiting saturation threshold is less than the saturation corresponding to the fully saturated state. Its engineering significance lies in retaining a necessary safety margin. For example, the laboratory can use the range that is close to but not fully saturated as the maximum utilization range to reduce the risk of unplanned gas supply interruptions caused by frequent disassembly. However, when the digital-physical desynchronization entropy increases or the failure probability increases, the system no longer waits for the limiting threshold, but instead outputs a forced offline regeneration command. Forced offline regeneration may include shutting down the gas path where the dryer is located, switching to a backup dryer, prompting the operator to remove the adsorption cylinder, sending silica gel or molecular sieve into the oven for regeneration at the specified temperature and time, and resetting the initial adsorption state of the corresponding digital twin after regeneration. In case of abnormal handling, if a backup dryer is available on site, the control actuator can switch to the backup branch first and issue an offline regeneration work order; if there is no backup branch, the system can link the downstream instruments to enter a protective shutdown or low-risk standby state; if the operator confirms that offline regeneration will not be performed for the time being, the system should record the manual confirmation information and continue to alarm, and will not regard the confirmation as model consistency recovery; if regeneration is completed but the state is not reset, the processing end will continue to maintain the original high-risk state until maintenance confirmation and necessary calibration data are received. In the aforementioned analytical laboratory, the dryer is normally allowed to operate up to the preset limit saturation threshold to fully utilize the adsorption capacity of silica gel and molecular sieves. When the step loss entropy reaches the preset risk threshold range during a sudden increase in water vapor load, the system sends a forced offline regeneration command to the control execution terminal: first, it switches to the standby drying cylinder, then prompts maintenance personnel to remove the original drying cylinder and place it in the oven for regeneration, while marking the current prediction model as being in a failure warning state on the host computer interface; if the environment returns to stability and the regenerated drying cylinder is brought back online, the system reinitializes its virtual saturation and enters the normal monitoring process. The purpose of this mechanism is to translate the risk identification results of the digital twin system into specific physical protection actions, extend the continuous operation time of the system under the tolerance risk, and prioritize blocking the moisture breakthrough link under the high risk of loss of synchronization or high failure.

[0023] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A digital twin system for a laboratory gas drying process, characterized in that, The system is mapped to a physical entity, which is a passive mechanical gas dryer with porous media inside. The system includes a data acquisition end, a control execution end, and a processing end that communicates with both. The processing end includes an environmental proxy data acquisition module, a physical state prediction module, a step-out entropy quantification evaluation module, and an adaptive safety decision module. The environmental proxy data acquisition module is used to acquire the external environmental proxy sequence data of the physical entity through the data acquisition terminal; The physical state prediction module is used to input the external environment proxy sequence data into a preset physical information neural network prediction model and a preset classical physical simulation model, respectively, calculate residual data based on the output results of the two, and output the virtual internal state parameters of the physical entity and the corresponding confidence level based on the residual data. The out-of-step entropy quantification and evaluation module is used to extract the fluctuation characteristics of the external environment proxy sequence data, calculate the digital-physical out-of-step entropy based on the fluctuation characteristics and the confidence level, and predict the failure probability of the physical entity reaching the physical failure critical point based on the digital-physical out-of-step entropy. The adaptive safety decision module is used to obtain a preset out-of-step safety threshold and a preset risk threshold based on historical failure sample statistics or the physical tolerance limit calibration of the intake air moisture content by downstream precision instruments; compare the digital-physical out-of-step entropy with the preset out-of-step safety threshold, compare the failure probability with the preset risk threshold, and output a state maintenance command or physical backup intervention command to the control execution terminal according to the comparison result, and generate a simulation model failure warning signal when outputting the physical backup intervention command.

2. The digital twin system for the laboratory gas drying process according to claim 1, characterized in that, The process of obtaining the external environment proxy sequence data of the physical entity includes: Obtain the fluid boundary state sequence at the air intake end of the physical entity and use the fluid boundary state sequence as initial proxy data; extract the high-frequency pulse features of the initial proxy data within a preset sliding time window; use the high-frequency pulse features to perform noise reduction filtering on the initial proxy data to obtain the external environment proxy sequence data.

3. The digital twin system for the laboratory gas drying process according to claim 2, characterized in that, The process by which the physical state prediction module outputs the virtual internal state parameters and corresponding confidence levels of the physical entity includes: The external environment proxy sequence data is input into the physical information neural network prediction model, and a nonlinear dynamic prediction result is output; the external environment proxy sequence data is input into the classical physical simulation model, and a deterministic benchmark prediction result is output; the residual data between the nonlinear dynamic prediction result and the deterministic benchmark prediction result is calculated; the virtual internal state parameters are generated based on the residual data as a compensation term; the variance or standard deviation of the residual data is calculated, and the confidence level of the virtual internal state parameters is determined based on the variance or standard deviation.

4. The digital twin system for the laboratory gas drying process according to claim 3, characterized in that, The process by which the out-of-step entropy quantization and evaluation module calculates the digital-physical out-of-step entropy includes: The fluctuation characteristics of the external environment proxy sequence data are extracted, including the amplitude change rate and the frequency change rate; the confidence level of the virtual internal state parameters is obtained, the confidence level is normalized and inverted to obtain the model cognitive uncertainty; the amplitude change rate and the frequency change rate are weighted and summed according to preset weight coefficients to construct an environmental interference penalty term; the model cognitive uncertainty is multiplied by the environmental interference penalty term to calculate the digital-physical desynchronization entropy.

5. The digital twin system for the laboratory gas drying process according to claim 4, characterized in that, The process of predicting the failure probability of a physical entity reaching a physical failure threshold based on the digital-physical out-of-step entropy includes: constructing a Markov state transition matrix; using the digital-physical out-of-step entropy and the virtual internal state parameters as the current state input to the Markov state transition matrix; calculating the transition probability of the current state transitioning to a physical failure threshold within a preset future time period using the Markov state transition matrix; and marking the transition probability as the failure probability.

6. The digital twin system for the laboratory gas drying process according to claim 5, characterized in that, The process by which the adaptive safety decision module outputs a state maintenance command or a physical backup intervention command to the control execution terminal based on the comparison result includes: Step 1: Determining whether the digital-physical out-of-synchronization entropy is less than a preset out-of-synchronization safety threshold. If it is less, proceed to Step 2; if it is greater than or equal to, proceed to Step 3. Step 2: Determining whether the failure probability is less than a preset risk threshold. If it is less, outputting a state maintenance command to the control execution terminal; if it is greater than or equal to, outputting a physical backup intervention command to the control execution terminal. Step 3: Directly outputting a physical backup intervention command to the control execution terminal and generating a simulation model failure warning signal.

7. The digital twin system for the laboratory gas drying process according to claim 6, characterized in that, The external environment proxy sequence data includes the gas supply pressure fluctuation data and temperature and humidity data at the inlet of the passive mechanical gas dryer; the virtual internal state parameters include the virtual saturation of the porous medium; the physical failure critical point is when the porous medium reaches a fully saturated state.

8. The digital twin system for the laboratory gas drying process according to claim 7, characterized in that, The physical backup intervention command includes a forced offline regeneration command for the passive mechanical gas dryer; the state maintenance command includes a command to allow the passive mechanical gas dryer to continue operating to a preset limit saturation threshold, wherein the limit saturation threshold is less than the saturation corresponding to the fully saturated state.

Images