Intelligent isotope labeling system with concentration prediction control capability

Abstract

The invention provides an intelligent isotope labeling system with concentration prediction control capability. The intelligent isotope labeling system comprises a data acquisition subsystem, a rate prediction subsystem, a concentration target trajectory calculation subsystem, a control instruction generation subsystem and a periodic updating subsystem. According to the method, the problems that an existing method is inaccurate in speed prediction, uncontrollable in track, split in control strategy and the like are solved, and through structural design and module linkage, the system has good engineering landing performance, adaptability and sustainable optimization capacity.

Description

Technical Field

[0001] This invention belongs to the field of intelligent isotope labeling, and particularly relates to an intelligent isotope labeling system with concentration prediction and control capabilities. Background Technology

[0002] In modern experimental setups for biomarkers, gas sensors, isotope tracing, and environmental simulation, precisely controlling the concentration evolution trajectory of specific components (such as CO2, especially its stable isotopes like 13CO2) is a crucial prerequisite for achieving experimental repeatability and safety. However, most existing methods rely on statically setting the inlet flow rate or a fixed proportion of mixed gases, lacking dynamic response and real-time feedback mechanisms. This makes it difficult to meet the control requirements of rapidly changing or nonlinear release scenarios, especially in non-standard release source environments such as liquid-liquid reactions that generate CO2, where this control bottleneck is even more pronounced. More seriously, existing CO2 control systems often use empirical formulas or lookup tables for concentration prediction, failing to accurately model the dynamic characteristics of the "start-up delay—peak burst—slow decay" process during release, leading to a mismatch between the concentration trajectory and the control behavior.

[0003] Furthermore, existing technologies generally neglect the coupling relationship between gas diffusion, dilution, and exchange in confined spaces, failing to construct a rate-concentration mapping mechanism that reflects the real physical process, thus limiting the ability to determine the engineering accessibility of concentration trajectories. Even if some systems possess closed-loop control capabilities, their control command generation often relies too heavily on classical PID or fuzzy rule bases, lacking a pre-assessment mechanism for the rationality and accessibility of the target trajectory itself, making them prone to ineffective regulation or system oscillations.

[0004] In actual deployment scenarios, due to issues such as equipment aging and environmental disturbances, the system control performance often tends to drift over time. The lack of a unified mechanism to update and optimize parameters further weakens the overall control robustness.

[0005] Therefore, there is an urgent need for a complete methodology that integrates reaction modeling, concentration trajectory design, accessibility assessment, fusion control strategies, and adaptive parameter updates. This methodology should form a reproducible, predictable, and controllable closed-loop process from the acquisition of underlying variables to the execution of high-level control, in order to solve the problem of accurate modeling and execution of concentration control under the dynamic process of CO2 release. Summary of the Invention

[0006] The purpose of this invention is to propose an intelligent isotope labeling system with concentration prediction and control capabilities. It constructs a complete technology chain consisting of five interdependent and interconnected steps, enabling the system to have good engineering feasibility, adaptability, and sustainable optimization capabilities.

[0007] To achieve the above objectives, this invention provides an intelligent isotope labeling system with concentration prediction and control capabilities, the method comprising:

[0008] The data acquisition subsystem is used to acquire and construct a five-dimensional normalized vector of CO2 isotope gas.

[0009] A rate prediction subsystem is used to construct a physically embedded feedforward network to output a CO2 gas release rate prediction sequence.

[0010] The concentration target trajectory calculation subsystem is used to collect the current environmental background concentration, combine it with the gas release rate prediction sequence to generate a CO2 concentration target curve, and generate an reachability indicator;

[0011] The control command generation subsystem is used to output a sequence of isotope-labeled control commands through the controller, based on the results of the data acquisition subsystem, the rate prediction subsystem, and the concentration target trajectory calculation subsystem.

[0012] Furthermore, the system also includes:

[0013] A periodic update subsystem is used to periodically update the system based on the execution results of the control command generation subsystem and to archive the data; each data archive package includes: timestamp, experiment number, control model version, environmental metadata, and anomaly flag;

[0014] The results obtained by the rate prediction subsystem are used to optimize the parameters of the periodic update.

[0015] Furthermore, the CO2 isotope gas is generated through the reaction of liquid 1 and liquid 2;

[0016] In the data acquisition subsystem, the volume input is acquired by an electromagnetic flowmeter, the initial reaction temperature is measured by a thermocouple, the initial reaction pressure is measured by a differential pressure sensor, and the set reaction duration is set through an HMI interface.

[0017] The steps performed by the data acquisition subsystem further include: normalizing the acquired data to generate a five-dimensional normalized vector of the CO2 isotope gas.

[0018] Furthermore, the modeling expression of the physically embedded feedforward network is as follows:

[0019]

[0020] Among them, f NN (·) represents a shallow neural network structure, with the input being the five-dimensional normalized vector and the current time, and the output being the basic generation rate; e -ktIt is an exponential delay function used to simulate the initial lag of CO2 release, where k is an empirical adjustment coefficient based on historical experimental fitting; T th This is the experimentally determined threshold temperature. When the temperature is too high, the reaction is prone to violent shaking and bubble bursts. λ1 is a parameter for adjusting the intensity of the suppression of the release rate in the high-temperature region; max(0,T) r -T th This option indicates that it only takes effect when the temperature exceeds the safety threshold, in order to limit the instability of the equipment caused by sudden rate changes; A sequence for predicting CO2 gas release rates.

[0021] Furthermore, the physical embedded feedforward network is trained using a weighted temporal loss function to enhance the fitting accuracy for the time period t∈[5s,20s], which is the main peak stage of CO2 release.

[0022] The CO2 gas release rate prediction sequence is formatted as a fixed-length vector, predicting once every 0.5 seconds, covering the array [0, t]. r The total length of the output array is 2·t, covering the entire interval. r .

[0023] Furthermore, the target concentration trajectory C target (t), calculated as follows:

[0024]

[0025] Among them, C bg V represents the current environmental background concentration. box Let t be the volume of the box, and t be the time point t. This is a predicted sequence for CO2 gas release rates, where τ is a floating-point vector of length τ, and e -μ(t-τ) The gas exchange factor represents the dilution of CO2 gas over time; the highest exhaust frequency indicates the fastest dilution. (1-e) -η(t-τ This is used to simulate the sluggish process of gas diffusing uniformly from the release port into the chamber; μ is the current environmental exchange rate parameter; η is the natural diffusion coefficient of CO2; The coupling penalty term is controlled by λ3; It is the instantaneous acceleration of concentration, used to locate abrupt change points; The first derivative is mapped to a smooth penalty coefficient restricted to [-1,1] for soft suppression of drastic changes; γ is the sensitivity coefficient of the penalty function, used to reflect the strength of suppression of mutations.

[0026] Furthermore, the concentration target trajectory calculation subsystem is also used to determine trajectory reachability:

[0027] The maximum permissible error is designed, and the error between the target concentration trajectory and the predicted concentration trajectory is calculated point by point. If the maximum permissible error is exceeded at any time, the current trajectory is considered unreachable, and the user or system is prompted to automatically roll back and adjust the response conditions of the data acquisition subsystem to regenerate the prediction.

[0028] Furthermore, the controller output includes the intake valve opening time, the exhaust valve opening time, and the vacuum pump pumping speed.

[0029] Furthermore, in the control command generation subsystem, the actual concentration trajectory and the predicted concentration trajectory are calculated point by point to construct a feedback error term, and a complete fusion control command generation structure is constructed based on the feedback error term:

[0030]

[0031] Among them, u t =[δ in (t),δ out (t),f pump [t] represents the complete control action at the current time point; δ in (t) represents the pulse width modulation (PWM) value of the intake valve, δ out (t) represents the opening duration of the exhaust valve, f pump (t) represents the speed of the vacuum pump; The feedforward control module is structured as a single fully connected neural network. Its input is a 2D vector (gas velocity and target concentration), with a hidden layer size of 32, and its output is a 3D control signal. For the CO2 gas release rate prediction sequence, C target (t) represents the target concentration trajectory; u t-1 The control values ​​executed in the previous time slice are used to introduce equipment inertia; α∈(0,1) represents the control fusion weight, empirically set to 0.6~0.8; λ4 is the coefficient of the error adjustment gain term; ε(t) is the error adjustment gain term, used for fine control compensation;

[0032] The control command generation subsystem refreshes the control signal every 0.5 seconds.

[0033] Furthermore, in the control command generation subsystem, if the current trajectory is unreachable, a conservative fault-tolerant mode is entered:

[0034] The pulse width modulation (PWM) value of the current intake valve and the opening duration of the exhaust valve are limited to the maximum physical limit to prevent the system from responding violently and causing disturbance accumulation; at the same time, the error adjustment gain term will be scaled to the upper limit to enhance stability.

[0035] The beneficial technical effects of the present invention are at least as follows:

[0036] This invention first collects initial reaction variables using high-precision instruments and employs a standardized mechanism to form a unified model input, ensuring data source quality and dimensional compatibility. Second, it constructs a rate prediction model incorporating physical knowledge, introducing exponential delay and temperature suppression terms to significantly improve the fitting accuracy and robustness of typical release curves. Subsequently, in the concentration trajectory design stage, it introduces for the first time a method combining a double exponential integral model and a second-order penalty mechanism to systematically model the process characteristics of "release-diffusion-dilution" and perform physical reachability verification, effectively avoiding uncontrollable trajectory input to the control module. Regarding control command generation, it constructs a fusion controller combining predicted rate and error feedback, introducing a dynamic inertia maintenance mechanism and fault-tolerant protection logic to ensure control accuracy and equipment safety. Finally, it designs a periodic data archiving and incremental parameter update mechanism to achieve long-term adaptive optimization of the model without affecting system real-time performance, enhancing system robustness and maintainability.

[0037] The overall solution not only overcomes the problems of inaccurate rate prediction, uncontrollable trajectory, and fragmented control strategies in existing methods, but also enables the system to have good engineering feasibility, adaptability, and sustainable optimization capabilities through structural design and modular linkage. Attached Figure Description

[0038] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0039] Figure 1 This is a framework diagram of the intelligent isotope labeling system with concentration prediction and control capabilities according to the present invention.

[0040] Figure 2 This is a flowchart of the intelligent isotope labeling system according to an embodiment of the present invention. Detailed Implementation

[0041] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0042] In one or more embodiments, such as Figure 1 As shown, a smart isotope labeling system with concentration prediction and control capabilities is disclosed, comprising:

[0043] The data acquisition subsystem 101 is used to acquire and construct a five-dimensional normalized vector of CO2 isotope gas.

[0044] Specifically, the goal of the data acquisition subsystem 101 is to collect key variables related to the isotopic CO2 generation process from the actual reaction system, and to standardize and structure these variables as input for subsequent modeling steps. The system design must ensure that the collected variables are verifiable, sampling stable, and engineering compatible, ultimately forming a structured and dimensionally consistent input vector.

[0045] In the system design, CO2 isotope gases (such as...) 13 CO2 is released through the reaction of liquid 1 (e.g., carbonate solution) and liquid 2 (e.g., dilute sulfuric acid). Therefore, it is necessary to first collect the actual volumes of both liquids added. Liquid 1 and liquid 2 are injected by two independent solenoid valves. Each injection port is equipped with a high-precision miniature electromagnetic flowmeter (model SFM3400, sampling frequency 100Hz, accuracy 0.1mL) to measure the inflow volume in real time. The system integrates these volumes to obtain the total injected volume, denoted as V1 and V2. The system has an internal timestamp synchronization mechanism to ensure that the flow-volume integration is completed before the reaction begins.

[0046] To ensure consistent reaction conditions, it is also necessary to collect the initial temperature T of the reaction chamber. r With initial pressure P r Temperature is collected in real time by a K-type thermocouple embedded in the reaction chamber wall and uploaded via an AD module at a sampling frequency of 1Hz. Initial pressure is read by a miniature differential pressure sensor (model MPX5010) at the top of the chamber, in the form of an analog voltage signal, which is then processed into a floating-point pressure value by an A / D conversion module. To avoid signal noise interference, the system performs a moving average of the pressure and temperature over the first 5 seconds during the initialization phase, using this average as the valid initial input.

[0047] In addition, the system is equipped with a programmable logic controller (PLC), which allows the planned response time t to be set via a human-machine interface (HMI). r This refers to the duration from the completion of liquid injection to the closed reaction stage, typically set between 15 and 60 seconds, depending on the required gas volume and reaction intensity. This variable is directly written into the control flow register for use by the predictive model.

[0048] Organize the above 5 variables into the original input vector:

[0049] X r =[V1,V2,T r ,P r ,t r ]

[0050] Where: V1: Volume input of liquid 1, collected by an electromagnetic flow meter; V2: Volume input of liquid 2, collected by an electromagnetic flow meter; T rThe initial reaction temperature, measured by a thermocouple; P r The initial pressure of the reaction is measured by a differential pressure sensor; t r The duration of the set response is configured through the HMI interface.

[0051] Since these five variables come from different physical dimensions (such as volume, temperature, pressure, and time), directly using them as input to the neural network can easily cause training skew and gradient instability. Therefore, this step uses a min-max normalization method for unified processing. The normalization parameters are based on the most recent 100 sets of experimental data from the system's historical records, maintained by a local SQLite database. The maximum and minimum values ​​of each variable are extracted in vector form. The normalization operation is as follows:

[0052]

[0053] Where: min(X) r The vector () represents the minimum values ​​of each variable in historical data; max(X) r ) represents the vector of maximum values ​​for each variable in historical data; all maximum and minimum values ​​are pre-stored and do not change with the current experiment.

[0054] To ensure input stability, the normalization operation is performed as a matrix operation in the host computer and accelerated in real time by the FPGA module. All floating-point calculations are fixed to four decimal places.

[0055] The final output is A normalized five-dimensional vector serves as the direct input to the gas generation prediction model in subsequent steps. This vector is temporarily stored in a buffer after generation and is associated with an experiment number to support subsequent control and archiving.

[0056] Rate prediction subsystem 102 is used to construct a physically embedded feedforward network to output a CO2 gas release rate prediction sequence.

[0057] Specifically, the goal of the rate prediction subsystem 102 is based on the normalized input vector obtained in the previous step. Construct a rate prediction model specifically for stable isotope gas release scenarios. This model not only needs to predict the release rate of CO2 gas during the reaction process, but also needs to consider the three-stage dynamic characteristics of "start-up delay - reaction peak - hysteresis decay" that are common in experimental scenarios. At the same time, it needs to meet the high requirements of industrial control systems for model simplicity, deployability and real-time performance.

[0058] First, the input variable is the five-dimensional normalized vector output from step one. in:

[0059] V1 and V2 are the injection volumes of carbonate and acid solution, respectively, which determine the initial carbon source and acid strength of the reaction.

[0060] T r The initial temperature of the reaction chamber significantly affects the peak reaction rate.

[0061] P r The initial chamber pressure determines the gas release rate driven by the reaction pressure difference;

[0062] t r The target reaction time set for the experiment is consistent with the sampling window size.

[0063] Furthermore, to enhance the representation of the dynamic characteristics of the reaction, this step designs a physically embedded feedforward network structure, which incorporates two special control terms abstracted from the experimental physical process into the basic network structure: the initiation delay factor e. -kt A high-temperature rate penalty term is added to enhance the model's ability to capture reaction change trends at different temperatures and initial stages.

[0064] The complete modeling expression is as follows:

[0065]

[0066] Where: f NN (·) represents a shallow neural network structure with input as... The output is the basic generation rate based on the current time t; e -kt It is an exponential delay function that simulates the initial lag of CO2 release. k is an empirical adjustment coefficient, based on historical experimental fitting, and is typically taken as 0.05 to 0.1; T th This is the experimentally determined threshold temperature. When the temperature is too high, the reaction is prone to violent shaking and bubble bursts. λ1 is a parameter for adjusting the intensity of the suppression of the release rate in the high-temperature region; max(0,T) r -T th The option indicates that it only takes effect when the temperature exceeds a safe threshold, in order to limit the rate of change that could cause equipment instability.

[0067] Understandably, this structure has two major innovations compared to traditional single-layer neural networks:

[0068] By explicitly multiplying by the exponential decay term, the physical consistency of the model with the three-stage response of "start-up delay-peak-gradual decline" is enhanced;

[0069] By incorporating a suppression term for abnormal behavior at high temperatures in the experiment, the model prediction results exhibit greater engineering stability, avoiding nonlinear surges caused by temperature fluctuations.

[0070] The training data came from a high-frequency mass flow meter (Sensirion SFM3300) on the CO2 outlet pipe of the system, which recorded the real gas flow rate over 60 seconds at a frequency of 20Hz to form the target label.

[0071] Furthermore, the neural network structure is as follows:

[0072] Input layer: accepts A total of 6 dimensions;

[0073] Hidden layer: Dense(32), activation function ReLU;

[0074] Output layer: Dense(1), linear output;

[0075] The network training employs a weighted temporal loss function, focusing on enhancing the fitting accuracy for the time interval t∈[5s,20s], which is the peak period of CO2 release and is crucial for valve control. The loss function is as follows:

[0076]

[0077] in: This represents the actual rate value collected. The model predicts the value; α(t) i ) is a time-weighted function, which takes a value of 1 in t∈[5,20] seconds and decreases to 0.3 in other time periods to highlight the fitting requirements of the main peak segment; N is the total number of sampling points in the prediction window.

[0078] The final output is a gas release rate prediction sequence. The format is a fixed-length vector, predicted every 0.5 seconds, covering [0, t]. r The total length of the output array is 2·t, covering the entire interval. r This sequence will be used in the next step for concentration trajectory prediction and environmental control strategy generation.

[0079] The concentration target trajectory calculation subsystem 103 is used to collect the current environmental background concentration, combine it with the gas release rate prediction sequence to generate a CO2 concentration target curve, and generate an reachability indicator.

[0080] Specifically, this step aims to predict the gas release rate based on the gas release rate sequence generated by the rate prediction subsystem 102. Combined with the current background concentration C of the system bg Design a target CO2 concentration trajectory C that satisfies the equipment's response characteristics, has physical accessibility, and ensures environmental stability. target(t). Simultaneously, a systematic reachability determination mechanism is introduced to assess whether the trajectory is engineering-feasible under the current reaction settings. This step serves as a bridge between "rate prediction" and "closed-loop control logic," ensuring the control plan possesses both physical rationality and control robustness.

[0081] Further collection of current environmental background concentration C bg (Measured in real time by a high-precision CO2 sensor inside the chamber); Chamber volume V box (System preset constant, unit L); Current environmental exchange rate parameter μ (air exchange ratio set by exhaust / replenishment valve); CO2 natural diffusion coefficient η (empirical constant, reflecting the gas uniformity inside the chamber).

[0082] The system first predicts the gas release rate sequence. The CO2 concentration accumulation curve is obtained by integration. Considering the physical mechanisms of hysteresis-exchange-diffusion in a closed chamber, this invention introduces a correction term to enhance the engineering reliability of the trajectory prediction:

[0083]

[0084] Where: e -μ(t-τ) The gas exchange factor represents the dilution of CO2 over time; the higher the exhaust frequency, the faster the dilution. (1-e -η(t-τ) The simulation simulates the delayed process of gas diffusion from the release port into the chamber; the two exponential terms together construct a non-transient mapping function of "rate → concentration" under real-world conditions.

[0085] The integral is approximated in the program by a 0.5-second sliding window convolution, which is highly efficient and keeps the error within 5 ppm.

[0086] Next, to ensure that the predicted curve can be transformed into the control target, the actual set trajectory C needs to be constructed. target (t), while maintaining the feasibility of the control target, aims to closely approximate natural concentration changes. To this end, two innovative treatment mechanisms are introduced:

[0087] Nonlinear weighted difference smoothing term: reduces local fluctuations;

[0088] Punishment-type jump control function: Explicitly punishes second-order concentration jumps to prevent frequent system regulation.

[0089] The formula for generating the target trajectory is as follows:

[0090]

[0091] Wherein: the first term is the basic concentration prediction value; the second term is the coupling penalty term, the intensity of which is controlled by λ3 (empirical value is 5-10 ppm); It is the instantaneous acceleration of concentration, used to locate the point of abrupt change; The first derivative is mapped to a smooth penalty coefficient restricted to [-1,1] to softly suppress drastic changes; γ is the sensitivity coefficient of the penalty function, reflecting the strength of suppression of mutations.

[0092] The innovation of this structural design lies in using an asymmetric second derivative to couple the first-order change slope, which strongly suppresses the sudden increase in the region while retaining the predicted trend in the gradual change region. This adapts to the inertial response characteristics of the box and the physical reality of gas flow velocity, avoiding the technical dead end of "a beautiful trajectory but uncontrollable".

[0093] Finally, trajectory reachability is determined. The system sets a maximum permissible error ∈ max (e.g., 5ppm), and calculate |C point by point. target (t)-C pred (t)|。 If the threshold is exceeded at any time t, the system will determine that the trajectory is unreachable and prompt the user or the system to automatically roll back and adjust the reaction conditions (such as increasing or decreasing V1 or V2) to regenerate the prediction.

[0094] The final output is: Target concentration trajectory C target (t), in the format of a fixed-length floating-point sequence, with a unit time of 0.5 seconds; reachability flag R feasible A Boolean value indicating whether the current trajectory is controllable.

[0095] The control command generation subsystem 104 is used to output a sequence of isotope-labeled control commands through the controller based on the results of the data acquisition subsystem, the rate prediction subsystem, and the concentration target trajectory calculation subsystem.

[0096] Specifically, the task of the control command generation subsystem 104 is to obtain the target concentration trajectory C target (t) and trajectory reachability flag R feasible Based on this, combined with the real-time collected current concentration C in the chamber real (t), generates control instructions executable at the device layer. The system drives the electronic intake valve, exhaust valve, and vacuum pump to work in tandem, ensuring that the actual concentration matches the set trajectory. To enhance control accuracy and stability, the system employs a control strategy that integrates feedforward regulation and error feedback control mechanisms. Furthermore, considering the equipment's inertial characteristics and control stability requirements, a structural vibration damping and inertia holding module is incorporated into the control commands.

[0097] Furthermore, to achieve an action update every 0.5 seconds, the system controller output is designed as the following triplet:

[0098] Intake valve opening time δ in(t) (unit: PWM pulse width control value);

[0099] Exhaust valve opening time δ out (t);

[0100] Vacuum pump pumping speed f pump (t)(unit is dimensionless scaling factor).

[0101] The system first constructs a feedback error term based on the target and actual concentrations:

[0102] ε(t)=C target (t)-C real (t)

[0103] Where: C target (t) represents the concentration value that the system aims to achieve at the current time point; C real ε(t) represents the actual monitored concentration; ε(t) is used as the dynamic error quantity for subsequent feedback adjustment.

[0104] Next, we will construct a complete fusion control instruction generation structure:

[0105]

[0106] Where: u t =[δ in (t),δ out (t),f pump [(t)] represents the complete control action at the current time point; The feedforward control module is structured as a single fully connected neural network. Its input is a 2D vector (gas velocity and target concentration), with a hidden layer size of 32, and its output is a 3D control signal. t-1 The control values ​​executed in the previous time slice are used to introduce equipment inertia (to prevent drastic changes in action); α∈(0,1) represents the control fusion weight, empirically set to 0.6~0.8; ε(t) is the error adjustment gain term (empirically set to 1~5) used for fine control compensation. λ4 is the coefficient of the error adjustment gain term.

[0107] In practical execution, the controller outputs u t Encoded into three digital signals, each controlling:

[0108] The pulse width modulation (PWM) value of the intake valve (achieved via an electromagnetic driver); the opening duration of the exhaust valve; and the vacuum pump speed (controlled by the drive module via PWM-voltage conversion).

[0109] The system refreshes the aforementioned control signals every 0.5 seconds. Specifically, when R... feasible When δ = 0, it indicates that the target trajectory is unattainable, and the system enters a conservative fault-tolerant mode: at this time, δ is limited.in (t) and δ out (t) will not exceed the maximum physical limit (e.g., 0.2 seconds) to prevent the system from responding violently and causing disturbance accumulation; at the same time, the feedback term ε(t) will be scaled to the upper limit (e.g., within ±10ppm) to enhance stability.

[0110] The output is: a sequence of control commands. Each iteration generates new control variables; control status labels: normal execution / fault-tolerant degradation (based on R). feasible Decide).

[0111] The periodic update subsystem 105 is used to periodically update the system according to the execution results of the subsystem generated by the control instructions, and to archive the data; each package of the data archive includes: timestamp, experiment number, control model version, environmental metadata, and anomaly marker.

[0112] Specifically, this step, as an extension module in the system's operational cycle, aims to achieve data archiving and periodic updates of model parameters during the concentration control process, addressing prediction inaccuracies caused by long-term changes in control targets, equipment aging, and environmental drift. To determine whether the system's self-update mechanism is triggered, an objective function for dynamic evaluation of system performance is first constructed.

[0113]

[0114] The first item is the concentration control error assessment, which reflects the stability of the system's deviation from the set target; the second item is the penalty for changes in control actions, which is used to measure the severity of control fluctuations; λ6 is a fixed weight item used to balance the two parts of the index; both items use historical data within the execution period T.

[0115] The system has a built-in threshold judgment module: if there are K consecutive periods (e.g., K=3) in... If it does, the archiving and incremental parameter optimization logic will be triggered; otherwise, the update will be delayed.

[0116] At this point, the system initiates the following two specific operations:

[0117] 1. Data archiving: The system will... The data packets are compressed and stored in a binary structured format. Each packet includes: a timestamp; an experiment number; a control model version; environmental metadata (temperature, humidity, equipment status); and anomaly markers (if any).

[0118] Furthermore, data is written to a local database or synchronized to a remote log management server. Data packets are named in the format {ExperimentID}_{TimeStamp}.bin for traceability.

[0119] 2. Parameter update (optional):

[0120] The system calls the archived data from the first N rounds and uses an incremental learning mechanism to fine-tune the parameters θ of the original Transformer concentration prediction model:

[0121]

[0122] in, η is the concentration prediction value of the old model for the i-th round of historical data; η is the learning rate (adaptively set by the system); θ new The new parameter set after fine-tuning only updates the weights and does not change the structure; all fine-tuning is completed through the local host computer to prevent external intervention.

[0123] This parameter update process uses an "optional extension" mechanism, which is activated when the error index continuously exceeds the limit. It is not automatically updated under normal circumstances to ensure stable system operation. If necessary, the system can issue a "manual review required" prompt to enhance security.

[0124] The final output is: the updated concentration prediction model parameters θ. new Archived data file record path and index number; control performance change trend log (for long-term trend analysis or predictive model stability study).

[0125] like Figure 2 As shown, this system is deployed in a closed isotope labeling experimental environment, including a data acquisition subsystem, a rate prediction subsystem, a concentration target trajectory calculation subsystem, a control command generation subsystem, and execution terminals (such as gas valves, pressure control modules, pure CO2 generators, incubators, etc.). The specific process is as follows:

[0126] First, zero-point calibration is performed on the gas flow sensor (used for purging adjustment, flow input, etc.), concentration detector, and temperature and humidity sensor to ensure data acquisition accuracy. The physical parameters of the experimental environment (such as the volume of the experimental chamber, the space parameters of the incubator, etc.) and the theoretical gas production curve of the pure CO2 generator (derived based on the principle of chemical reaction) are preloaded. At the same time, the control parameters of the incubator (such as the initially set flow rate, pressure threshold, etc.) are initialized.

[0127] Secondly, multi-dimensional data is collected at a set frequency (e.g., 1Hz), and the collected multi-dimensional parameters are normalized to construct a five-dimensional normalized vector (covering key parameter dimensions related to CO2 isotope gas regulation, such as flow rate, pressure, temperature, humidity, time, or device operating parameters, etc. The specific dimensions can be determined and optimized through experiments based on the actual scenario) as the basis for system input.

[0128] Furthermore, a physically embedded feedforward neural network (FFN) is used to integrate the physical laws of chemical reactions with data-driven correction, and combined with the environmental background concentration and release rate to generate a CO2 concentration target curve that "meets experimental requirements and is physically achievable".

[0129] Finally, based on the model predictive control (MPC) framework, the control sequence is output by integrating data from multiple subsystems.

[0130] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0131] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0132] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0133] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0134] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0135] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0136] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart isotope labeling system with concentration prediction and control capabilities, characterized in that, The system includes: The data acquisition subsystem is used to acquire and construct a five-dimensional normalized vector of CO2 isotope gas. A rate prediction subsystem is used to construct a physically embedded feedforward network to output a CO2 gas release rate prediction sequence. The concentration target trajectory calculation subsystem is used to collect the current environmental background concentration, combine it with the gas release rate prediction sequence to generate a CO2 concentration target curve, and generate an reachability indicator; The control command generation subsystem is used to output a sequence of isotope-labeled control commands through the controller, based on the results of the data acquisition subsystem, the rate prediction subsystem, and the concentration target trajectory calculation subsystem.

2. The intelligent isotope labeling system with concentration prediction and control capability according to claim 1, characterized in that, The system also includes: A periodic update subsystem is used to periodically update the system based on the execution results of the control command generation subsystem and to archive the data; each data archive package includes: timestamp, experiment number, control model version, environmental metadata, and anomaly flag; The results obtained by the rate prediction subsystem are used to optimize the parameters of the periodic update.

3. The intelligent isotope labeling system with concentration prediction and control capability according to claim 1, characterized in that, The five-dimensional normalized vector includes the volume input of liquid 1, the volume input of liquid 2, the initial reaction temperature, the initial reaction pressure, and the set reaction duration. The CO2 isotope gas is generated through the reaction of liquid 1 and liquid 2. In the data acquisition subsystem, the volume input is acquired by an electromagnetic flowmeter, the initial reaction temperature is measured by a thermocouple, the initial reaction pressure is measured by a differential pressure sensor, and the set reaction duration is set through an HMI interface. The steps performed by the data acquisition subsystem further include: normalizing the acquired data to generate a five-dimensional normalized vector of the CO2 isotope gas.

4. The intelligent isotope labeling system with concentration prediction and control capability according to claim 1, characterized in that, The physical embedded feedforward network is modeled as follows: Among them, f NN (·) represents a shallow neural network structure, with the input being the five-dimensional normalized vector and the current time, and the output being the basic generation rate; e -kt It is an exponential delay function used to simulate the initial lag of CO2 release, where k is an empirical adjustment coefficient based on historical experimental fitting; T th This is the experimentally determined threshold temperature. When the temperature is too high, the reaction is prone to violent shaking and bubble bursts. λ1 is a parameter for adjusting the intensity of the suppression of the release rate in the high-temperature region; max(0,T) r -T th This option indicates that it only takes effect when the temperature exceeds the safety threshold, in order to limit the instability of the equipment caused by sudden rate changes; A sequence for predicting CO2 gas release rates.

5. The intelligent isotope labeling system with concentration prediction and control capability according to claim 4, characterized in that, The physically embedded feedforward network is trained using a weighted temporal loss function to enhance the training of time t∈ The fitting accuracy for the [5s, 20s] time period, where the time period is the main peak stage of CO2 release; The CO2 gas release rate prediction sequence is formatted as a fixed-length vector, predicting once every 0.5 seconds, covering the array [0, t]. r The total length of the output array is 2·t, covering the entire interval. r .

6. The intelligent isotope labeling system with concentration prediction and control capability according to claim 1, characterized in that, The target concentration trajectory C target (t), calculated as follows: Among them, C bg V represents the current environmental background concentration. box Let t be the volume of the box, and t be the time point t. This is a predicted sequence for CO2 gas release rates, where τ is a floating-point vector of length τ, and e -μ(t-τ) The gas exchange factor represents the dilution of CO2 gas over time; the highest exhaust frequency indicates the fastest dilution. (1-e) -η(t-τ) This is used to simulate the sluggish process of gas diffusing uniformly from the release port into the chamber; μ is the current environmental exchange rate parameter; η is the CO2 natural diffusion coefficient; The coupling penalty term is controlled by λ3; It is the instantaneous acceleration of concentration, used to locate abrupt change points; The first derivative is mapped to a smooth penalty coefficient restricted to [-1,1] for soft suppression of drastic changes; γ is the sensitivity coefficient of the penalty function, used to reflect the strength of suppression of mutations.

7. The intelligent isotope labeling system with concentration prediction and control capability according to claim 6, characterized in that, The concentration target trajectory calculation subsystem is also used to determine trajectory reachability. The maximum permissible error is designed, and the error between the target concentration trajectory and the predicted concentration trajectory is calculated point by point. If the maximum permissible error is exceeded at any time, the current trajectory is considered unreachable, and the user or system is prompted to automatically roll back and adjust the response conditions of the data acquisition subsystem to regenerate the prediction.

8. The intelligent isotope labeling system with concentration prediction and control capability according to claim 1, characterized in that, The controller outputs include the inlet valve opening time, the exhaust valve opening time, and the vacuum pump pumping speed.

9. A smart isotope labeling system with concentration prediction and control capability according to any one of claims 8, characterized in that, In the control command generation subsystem, the actual concentration trajectory and the predicted concentration trajectory are calculated point by point to construct a feedback error term. Based on the feedback error term, a complete fusion control command generation structure is constructed. Among them, u t =[δ in (t),δ out (t),f pump [t] represents the complete control action at the current time point; δ in (t) represents the pulse width modulation (PWM) value of the intake valve, δ out (t) represents the opening duration of the exhaust valve, f pump (t) represents the speed of the vacuum pump; The feedforward control module is structured as a single fully connected neural network. Its input is a 2D vector (gas velocity and target concentration), with a hidden layer size of 32, and its output is a 3D control signal. For the CO2 gas release rate prediction sequence, C target (t) represents the target concentration trajectory; u t-1 The control values ​​executed in the previous time slice are used to introduce equipment inertia; α∈(0,1) represents the control fusion weight, empirically set to 0.6~0.8; λ4 is the coefficient of the error adjustment gain term; ε(t) is the error adjustment gain term, used for fine control compensation; The control command generation subsystem refreshes the control signal every 0.5 seconds.

10. A smart isotope labeling system with concentration prediction and control capability according to any one of claims 7 or 9, characterized in that, In the control command generation subsystem, if the current trajectory is unreachable, a conservative fault-tolerant mode is entered: The pulse width modulation (PWM) value of the current intake valve and the opening duration of the exhaust valve are limited to the maximum physical limit to prevent the system from responding violently and causing disturbance accumulation; at the same time, the error adjustment gain term will be scaled to the upper limit to enhance stability.

Images