Intelligent isotope gas concentration regulation and control method based on concentration prediction

By using neural network models and state-deviation mapping models, the problems of response lag and regulation oscillation in isotope gas concentration regulation in enclosed spaces were solved, achieving highly responsive and stable gas concentration control that can adapt to complex environmental changes and multiple batch condition switching.

CN120853715AInactive Publication Date: 2025-10-28SHENZHEN ZHONGTING TECH CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202510928934.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-07
Publication Date
2025-10-28
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing control systems struggle to achieve high response speed and steady-state control in enclosed spaces. Traditional PID control methods suffer from lag and are prone to concentration overshoot or regulation oscillations. They also lack the ability to model long-term deviations in control strategy behavior and cannot adapt to high-frequency experiments and complex environmental changes.

Method used

An intelligent control method based on concentration prediction is adopted, which combines dynamic modeling of sensor characteristics, optimization algorithms and data super-resolution technology. The isotope gas concentration is predicted through a neural network model, feedforward control commands are generated, and a state-deviation mapping model is constructed to optimize the long-term strategy.

Benefits of technology

It achieves high-precision isotope gas concentration control, improves the system's responsiveness and stability, and constructs a closed-loop control architecture of data perception, control decision-making, deviation feedback and strategy self-learning, breaking through the bottleneck problem of traditional concentration control systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120853715A_ABST
    Figure CN120853715A_ABST
Patent Text Reader

Abstract

The invention provides an intelligent isotope gas concentration regulation and control method based on concentration prediction, which comprises the following steps: collecting multi-dimensional data in a closed box subjected to emptying treatment to construct a standardized control-response data sequence; predicting the isotope gas concentration through a neural network model to obtain the predicted isotope gas concentration and a long-term concentration drift trend estimation item; inputting the current standardized control-response data sequence into Transform to obtain current state representation, and generating a control instruction of valve opening and closing time and sucking pump power on the basis of the current state representation in combination with predicted concentration and target concentration; executing the control instruction, comparing the actual concentration with the predicted concentration, and coding the deviation into a deviation risk code through a multi-layer perceptron in combination with the control instruction; receiving the current state representation and the deviation risk code to establish a state-deviation mapping model, fitting deviation distribution parameters of the control deviation, and accumulating control experience to optimize a long-term strategy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent regulation of isotope gas concentration, and particularly relates to an intelligent regulation method for isotope gas concentration based on concentration prediction. Background Technology

[0002] In numerous applications involving stable isotopes (such as 13CO2), such as precision experiments, agricultural breeding, and medical metabolic research, the demand for precise control of gas concentrations within small spaces is increasingly urgent. Especially in closed systems, due to limited space volume and strong boundary airtightness, the diffusion, accumulation, and dilution processes of gases exhibit strong nonlinearity and hysteresis. Traditional PID-based concentration regulation methods are insufficient to meet the dual requirements of response speed and steady-state control. Existing control systems generally rely on real-time concentration feedback for regulation, i.e., adjusting ventilation or extraction strategies only after concentration changes occur. This "post-feedback" mechanism not only has a delayed response but is also prone to concentration overshoot or regulation oscillations. Furthermore, most systems lack the ability to structurally model long-term deviations in control strategy behavior; even if deviations exist, experience cannot be accumulated or cross-batch strategy optimization cannot be performed. Regarding concentration prediction, although some methods attempt to introduce neural network models for trend judgment, they generally suffer from problems such as opaque model structure, insensitivity to control behavior, and lack of physical consistency in prediction results, failing to support the design of demanding control systems.

[0003] Furthermore, existing control systems rarely consider combining predictive and feedforward control, feedback deviation identification, and strategy experience evolution to construct a closed-loop, self-learning concentration control system. As a result, concentration control systems exhibit slow response to disturbances, lack of strategy generalization ability, and gradual performance degradation over long-term operation, making them ill-suited for high-frequency experiments, complex environmental changes, or multiple batch condition switching scenarios. Summary of the Invention

[0004] The purpose of this invention is to propose an intelligent control method for isotopic gas concentration based on concentration prediction, which combines dynamic modeling of sensor characteristics, optimization algorithms and data super-resolution technology, aiming to solve the problems of insufficient accuracy and poor environmental adaptability encountered by existing technologies in the process of multimodal sensor data alignment.

[0005] To achieve the above objectives, this invention provides an intelligent control method for isotope gas concentration based on concentration prediction, the method comprising:

[0006] S1: Collect multidimensional data from the emptied sealed enclosure to construct a standardized control-response data sequence;

[0007] S2: Using a neural network model, the concentration of isotopic gases in a closed environment is predicted to obtain the predicted concentration of isotopic gases.

[0008] S3: Identify the current state representation of the current standardized control-response data sequence, and combine the predicted concentration with the target concentration to generate control commands for valve opening and closing time and pump power;

[0009] S4: Execute the control command, compare the current actual concentration with the predicted concentration, and encode the deviation into a deviation risk code through a multilayer perceptron;

[0010] S5: Receive the current state representation and deviation risk code to establish a state-deviation mapping model, fit the deviation distribution parameters of the control deviation, and optimize the long-term strategy.

[0011] Furthermore, the multidimensional data acquisition includes real-time monitoring of CO2 concentration by infrared sensors, wind speed collected by hot-film anemometers, temperature and humidity by temperature and humidity sensors, valve switch status and pump power records, wherein the sampling frequency is not less than 1Hz and is aligned with a unified timestamp by the controller.

[0012] Furthermore, the structure of the neural network model is as follows:

[0013] The encoder part uses a two-layer Transformer structure to encode timing information, and the decoder part is a linear mapping module that fuses state embeddings and trend terms;

[0014] The prediction expression of the neural network model is as follows:

[0015]

[0016] in, X′ represents the predicted CO2 concentration τ seconds from now. t-L+1:t Θ is a normalized input time window of length L; Attn(·) is an encoder module with physical attention guidance, whose weight adjustment part is controlled by two sets of structural variables; phys Θ is the tensor of physical parameters for gas diffusion; ctrl The parameters for controlling the perturbation estimation are: W, W, and Δ. trend (t) is the long-term concentration drift trend estimate, used to handle the slow shift caused by the decrease in system airtightness or long-term background gas infiltration.

[0017] Furthermore, the loss function of the neural network model is expressed as:

[0018]

[0019] in, C is the loss function for the neural network model. t+τ The actual observed concentration value was collected by an infrared CO2 sensor; For the local first-order difference of the predicted sequence; φ ctrl λ1 and λ2 are the expected concentration change slope estimated based on the amplitude of abrupt changes in the control signal; λ1 and λ2 are the weight hyperparameters of the regularization term.

[0020] in, This is used to ensure that the predicted value is not less than zero and to avoid physical inconsistencies.

[0021] Furthermore, during the model training phase of the neural network model, a sliding window method is used to construct sample sequences from historical data, with each window having a length of L = 60 and a sliding step of 1 second;

[0022] All model training, tuning, and deployment are completed by the host computer, built using the PyTorch platform, and support dynamic online updates of physical prior tensors for later system adaptation to different experimental container structures or air duct configurations.

[0023] Furthermore, step S3 specifically includes:

[0024] The control strength ρ is calculated using a structurally interpretable policy function. t This is used to reflect the current system's demand for control.

[0025]

[0026] Among them, the first item This indicates the necessary control strength based on the prediction error; ∈ is a small constant to prevent division by zero; the second term is MLP. res (h t The dynamic disturbance adjustment term is based on the system state and consists of a two-layer MLP structure. res (h t The output implies the impact of sudden system disturbances, unstable trends, or historical disturbance memories; h t The current state is represented by α; α is the adjustment hyperparameter, which represents the proportion of the influence of the control state on the final control response.

[0027] Based on control intensity ρ t Two specific control output quantities are derived:

[0028]

[0029] Where u0,P0 are the current control state reference values; λ u ,λ P : These are the mapping coefficients from control strength to actual control variables; clip(·): are the upper and lower bound functions to prevent output overload; This is the valve opening and closing time for the next cycle; The operating power of the air pump for the next cycle; umin Minimum valve opening and closing time; u max P represents the maximum valve opening and closing time. min This represents the minimum operating power of the air pump; P max This represents the maximum operating power of the air pump.

[0030] Furthermore, the multilayer perceptron is a three-layer feedforward network. The first layer is ReLU activation, the second layer is linearly mapped to the intermediate state, and the final output is a Tanh code in the range of [-1,1], which expresses the direction and intensity of the deviation risk.

[0031] Furthermore, the state-deviation mapping model employs a dual-head neural network to output deviation distribution parameters.

[0032] Furthermore, the objective of the state-deviation mapping model is to fit the probability density estimate of the deviation, calculated as follows:

[0033]

[0034] The dual-head neural network outputs the mean μ(h) representing the current state. t The variance of the current state representation is σ. 2 (h t ), which is the parameter for controlling the deviation distribution; the input in the current state represents h. t The input is encoded into an intermediate hidden layer via a two-layer ReLU activated feedforward neural network, which outputs the mean μ(h) of the current state representation. t The variance of the current state representation is σ. 2 (h t );

[0035] The parameter training uses historical datasets. It is obtained through training using the maximum likelihood loss function.

[0036] Furthermore, S2 also generates a long-term concentration drift trend estimate; this long-term concentration drift trend estimate is used to handle the slow shift caused by decreased system airtightness or long-term background gas infiltration, and to integrate the chronic background gas infiltration effect over a long experimental period; the expected concentration change slope φ estimated based on the abrupt change amplitude of the control signal... ctrl The gas diffusion physics parameter tensor is obtained by adaptive fitting with historical mutation data; it is used to adjust the attention weight path to form structural physics guidance.

[0037] The beneficial technical effects of the present invention are at least as follows:

[0038] This invention proposes an isotope gas concentration prediction and control method based on gas diffusion mechanism. By constructing a control-response time series data structure and combining it with a physical guided attention model, high-precision concentration prediction is achieved. Furthermore, a feedforward control command is generated through a structured strategy function, and a prediction deviation feedback module and an empirical strategy evolution mechanism are introduced to construct a complete intelligent concentration control closed-loop system.

[0039] The system of this invention first extracts key factors affecting concentration changes from control behavior and environmental variables, constructs a standardized input structure, and provides a unified data representation for the model. Then, it incorporates physical priors such as control-driven disturbance trends, physical non-negativity, and system slow variability into concentration prediction, improving the model's accuracy and interpretability. In terms of control strategy generation, it integrates prediction errors and system state regulation characteristics to form a structured control intensity function, achieving a more responsive feedforward control output than traditional PID control. The system also constructs a structured coding module in the prediction deviation feedback stage to identify control effects and construct a quantified deviation representation. Finally, through control deviation density modeling and joint state expression, it proposes a long-term control experience expression mechanism, providing a reference for subsequent strategy optimization and model evolution.

[0040] The entire system is centered on concentration prediction and has constructed a closed-loop control architecture that integrates data perception, control decision-making, deviation feedback, and strategy self-learning, thus overcoming the bottleneck problems of "one-way prediction", "static control" and "no experience accumulation" in existing concentration regulation systems. Attached Figure Description

[0041] 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.

[0042] Figure 1 This is a flowchart of an intelligent control method for isotope gas concentration based on concentration prediction according to the present invention. Detailed Implementation

[0043] 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.

[0044] In one or more embodiments, such as Figure 1 As shown, a method for intelligent control of isotope gas concentration based on concentration prediction is disclosed, including:

[0045] S1: Collect multidimensional data from the emptied sealed enclosure to construct a standardized control-response data sequence.

[0046] Specifically, the goal of this step is to construct a multi-dimensional time-series data input structure for concentration prediction tasks. Since the application scenario of this invention is within a sealed chamber... 13 For isotopic gases like CO2, high-precision concentration control at the ppm level requires data that includes not only the gas concentration itself but also multiple dimensions such as gas flow, ambient temperature and humidity, and system control actions to reflect the true driving factors of concentration changes. To ensure the accuracy of subsequent concentration predictions and the interpretability of the control logic, the initialization phase must ensure the system is in a stable, clean, and known state, and simultaneously collect all relevant parameters to construct a time-series structure acceptable to subsequent models.

[0047] Since it is S1, all inputs come from the system's internal initialization settings and data acquisition devices, including:

[0048] Concentration sensor module for enclosure (infrared non-dispersive CO2 sensor);

[0049] Control actuator recording module (including valve control log and air pump power setting);

[0050] Environmental status acquisition module (including temperature probe, humidity sensor, and hot-film anemometer);

[0051] System configuration table (static parameters such as enclosure volume and flow channel distribution).

[0052] First, close all intake valves and start the vacuum pump at maximum power to reduce the gas concentration in the chamber to below 10 ppm. This process is scheduled by a lower-level controller, and the concentration values ​​(C) at various moments during the evacuation process are collected by an infrared CO2 sensor (e.g., SenseAir S8 model) at a sampling frequency of 1 Hz. t The sensor is positioned in the central area of ​​the chamber to easily reflect the residual gas level in the main mixing zone. The system determines that purging is complete and proceeds to the next stage once the concentration at 20 consecutive time points is below 10 ppm.

[0053] After purging, the system automatically activates the background air injection device, injecting background gas from an isotope-free CO2 gas source at a constant flow rate (set value, for example, 200 ml / min). Simultaneously, the internal fan module of the chamber is activated to enhance gas uniformity. The wind speed is measured by a hot-film anemometer (e.g., FS7 series), installed near the exhaust port to capture the dominant flow state of the mixed airflow, with a sampling frequency of 5 Hz. The collected wind speed is recorded as V. t The variable is then smoothed (sliding window length is 2 seconds) for subsequent variable normalization.

[0054] Simultaneous collection of ambient temperature T t With relative humidity H t Temperature data was acquired using a DS18B20 single-point digital temperature probe, installed in a dead-end area of ​​the airflow to capture any potential temperature differences. Humidity was measured using an SHT35 digital humidity module, installed adjacent to the temperature probe to ensure compatibility. All sampling frequencies were 1Hz and aligned with a unified timestamp via the controller.

[0055] Control behavior recording is performed simultaneously with environmental data acquisition. Valve opening / closing signal u t Recorded by the lower-level controller log module, using 0 to indicate off and 1 to indicate on, with millisecond-level precision. Air pump power P t The record is the current duty cycle percentage of the PWM output, recorded once per second by the controller, to represent the dynamic pumping capacity of the pump.

[0056] All the above variables are organized into a multidimensional control-response input sequence:

[0057] X t =[C t ,T t H t V t ,u t ,P t ]

[0058] in:

[0059] C t CO2 concentration in the chamber is obtained from an infrared gas sensor, sampled at 1Hz;

[0060] T t Temperature, from DS18B20 probe;

[0061] H t Humidity, from SHT35 module;

[0062] V t Wind speed, measured by FS7 anemometer;

[0063] u t Valve control status (0 or 1), sourced from controller log;

[0064] P t : Pump power, expressed as PWM duty cycle percentage.

[0065] Furthermore, after establishing the time series, the scales of different physical quantities need to be normalized to facilitate subsequent model processing. The mean and standard deviation of all variables were calculated using running data from a continuous 20-minute period during the initialization phase, defined as μ.* ,σ * The normalized input is:

[0066]

[0067] in:

[0068] μ C ,μ T ,μ H ,μ V ,μ P : The mean of each variable within its stable operating range;

[0069] σ C ,σ T ,σ H ,σ V ,σ P : represents the corresponding standard deviation;

[0070] u t Since these are discrete variables, no normalization is performed.

[0071] The normalization process is completed by the host computer preprocessing module, which calculates the mean and standard deviation based on historical data and updates them dynamically. Taking wind speed as an example, assuming the mean value collected within 5 minutes after the venting is completed is 0.83 and the standard deviation is 0.11, then at any given time V t The above formula can be used to convert it into dimensionless input.

[0072] The key to this step is to unify the physical environmental variables and control behaviors into a time-series input format. This design reflects the diffusion lag and non-uniform aggregation behavior of isotopic gases in this scenario, which is superior to the traditional CO2 prediction method that only uses gas concentration as a single variable for modeling.

[0073] The final output includes: the normalized input sequence X′ t This will serve as the input to the concentration prediction model; the initial system state parameter set S0 = {μ C ,μ T ,μ H ,μ V ,μ P} is used as a benchmark for normalization and strategy adjustment.

[0074] S2: Using a neural network model, the concentration of isotopic gases in a closed environment is predicted to obtain the predicted concentration of isotopic gases.

[0075] Specifically, the goal of this step is to utilize the standardized control-response data sequence X′ constructed in the previous stage. tBased on this, a neural network model for predicting isotopic CO2 concentration in closed environments is established. This model must not only possess high-precision prediction capabilities for future concentration trends but also demonstrate interpretability and physical consistency to meet the "feedforward control" requirements in intelligent control scenarios for high-value gas concentrations. This step is not merely a prediction task but also a critical node in the control system, as its output directly impacts the valve and pump action decisions in the next step. Therefore, the structural design must balance prediction accuracy, system response stability, and physical interpretability.

[0076] The input is the structured sequence X′ output from step one. t =[C t ,T t H t V t ,u t ,P t ], and the initial state parameter set S0 = {μ C ,μ T ,μ H ,μ V ,μ P}

[0077] In this step, a hybrid network structure combining attention mechanisms, local physical diffusion field modeling, and predictive stability regularization control is used for model design. Special modeling design is specifically tailored to the following scene characteristics:

[0078] Isotopic gases (such as 13CO2) diffuse slowly in the experimental chamber and exhibit significant spatial accumulation delay;

[0079] The effects of environmental variables (such as wind speed and temperature gradient) on concentration changes are nonlinear and have a strong lag.

[0080] Control variable (valve state u) t With the power P of the air pump t This will cause strong disturbances, and it is necessary to model the potential disturbance trends in advance;

[0081] To maintain the system physical consistency of the predicted values ​​(e.g., non-negative concentration, gradual variation, and controllability), a custom regularization term is introduced into the model for soft constraints.

[0082] The model structure employs a "physically coupled perceptual attention prediction network," with an encoder-decoder architecture at its core. The encoder uses a two-layer Transformer structure to encode temporal information, while the decoder is a linear mapping module that fuses state embeddings and trend terms. The model's prediction expression is as follows:

[0083]

[0084] in:

[0085] The predicted CO2 concentration 13 seconds from now;

[0086] X′ t-L+1:t : is a normalized input time window of length L;

[0087] Attn(·): This is an encoder module with physical attention guidance, whose weight adjustment part is controlled by two sets of structured variables;

[0088] Θ phys : This refers to the gas diffusion physical parameter tensor (e.g., local airflow disturbance estimate, hysteresis coefficient), derived from the box structure modeling;

[0089] Θ ctrl : To control disturbance estimation parameters (such as current u) t 、P t The degree of induced CO2 fluctuations is estimated based on the gradient after abrupt changes within a historical window.

[0090] W: is the linear decoder weight vector;

[0091] Δ trend (t): This is the long-term concentration drift trend estimate, used to handle slow shifts caused by decreased system airtightness or long-term background gas infiltration.

[0092] Furthermore, to further enhance the model's prediction stability and physical interpretability in complex scenarios, a custom loss function is introduced, which consists of three parts: the main task prediction error term, the trend perturbation suppression term, and the physical non-negativity constraint term.

[0093]

[0094] in:

[0095] C t+τ The actual observed concentration values ​​were collected by an infrared CO2 sensor.

[0096] The local first-order difference of the predicted sequence (representing the predicted rate of change);

[0097] φ ctrl : The slope of the expected concentration change estimated based on the amplitude of the abrupt change in the control signal;

[0098] λ1, λ2: These are the hyperparameters for the regularization term weights;

[0099] The last item ensures that the predicted value is not less than zero, avoiding physical inconsistencies.

[0100] in, Main task error; Align with the disturbance trend; Non-negative positive regularization term

[0101] Furthermore, during the model training phase, a sliding window approach is used to construct sample sequences from historical data. Each window has a length of L = 60, and the sliding step is 1 second. All model training, tuning, and deployment are completed by the host computer, built using the PyTorch platform, and support dynamic online updates of the physical prior tensor Θ. phys It is used for later system adaptation to different experimental container structures or air duct configurations.

[0102] Understandably, the innovation of this structure lies in:

[0103] A disturbance trend regularization term φ driven by control variables was introduced. ctrl This item is not used in the usual way, but is obtained by adaptive fitting with historical mutation data, emphasizing direct modeling of "control action → concentration perturbation";

[0104] Add a physical nonnegative regularization term to the concentration prediction (to penalize negative prediction values ​​twice);

[0105] The gas diffusion dynamic parameter Θ is introduced into the structure in a tensor manner. phys This involves adjusting the attention weight path to create structural physical guidance.

[0106] Using the concentration drift trend estimate Δ trend (t) is used to fit the chronic background gas permeation effect over a long experimental period. Such terms are rarely seen in general prediction tasks, but are of great practical significance in the scenario of this invention.

[0107] The final output includes: Provides predicted values ​​for the generation of subsequent control actions; Δ trend (t): Provides a basis for predicting medium- and long-term concentration trends for the strategic planning layer.

[0108] S3: Identify the current state representation of the current standardized control-response data sequence, and combine the predicted concentration with the target concentration to generate control commands for valve opening and closing time and pump power.

[0109] Specifically, this step aims to build upon the concentration prediction results from the previous stage. and the current system state representation h tWhen abnormal fluctuations in gas concentration are predicted in the future, a feedforward control strategy is proactively generated to drive valves and pumps to intervene in advance, achieving precise control of 13CO2 gas concentration at the ppm level. Compared to traditional closed-loop systems that wait for the concentration to exceed the limit before triggering control, the feedforward strategy in this step has a higher response speed and control accuracy, making it particularly suitable for the special needs of isotope experiments with stringent concentration requirements and delayed diffusion reactions.

[0110] Input variables come from the previous stage:

[0111] The predicted concentration value output from step two;

[0112] h t The Transformer structure outputs a representation of the system's current state (e.g., a 64-dimensional vector).

[0113] r resp The concentration change (ppm / %) caused by unit control is derived from experimental calibration.

[0114] C target Target concentration setting (e.g., 1000 ppm) is a system constant.

[0115] Furthermore, to avoid instability and computational burden during policy generation, the scheme employs a structured interpretable policy function instead of a deep black-box policy network. This invention defines the control strength ρ. t This is used to reflect the current system's demand for control.

[0116]

[0117] The first item represents the necessary control strength based on the prediction error (if the predicted value is greater than the target value, a "vacuuming" operation is required);

[0118] ∈ is a small constant (e.g., 10). -3 ), to prevent division by zero;

[0119] The second term is a dynamic disturbance adjustment term based on the system state, consisting of a two-layer MLP structure. res (h t The output implies the impact of sudden system disturbances, unstable trends, or historical disturbance memories.

[0120] α is the adjustment hyperparameter (e.g., 0.2–0.5), representing the proportion of influence of the control state drive on the final control response.

[0121] This expression integrates two parts: "prediction error-driven control intent" and "state characteristic-aware adjustment," forming the core estimation logic of the entire control output strategy. Unlike traditional control strategies based on PID or single error feedback, this structural combination enhances the model's robustness to nonlinear factors (such as temperature and humidity fluctuations or wind speed changes).

[0122] Next, based on ρ t Two specific control output quantities are derived:

[0123]

[0124] in:

[0125] u0, P0: The current control state reference values ​​(control parameters used in the previous cycle);

[0126] λ u ,λ P These are the mapping coefficients from control intensity to actual control variables (obtained through calibration);

[0127] clip(·): This is a bounding function to prevent output overload, such as u∈[0,8] seconds, P∈[0,100]%.

[0128] Final output The signals are converted into control signals via an interface and applied to the solenoid valve (to control ventilation) and the PWM pump control module (to control the pumping rate), respectively.

[0129] The final output is: Valve opening / closing time (seconds) for the next cycle; The operating power (%) of the air pump in the next cycle.

[0130] S4: Execute the control command, compare the current actual concentration with the predicted concentration, and encode the deviation into a deviation risk code through a multilayer perceptron.

[0131] Specifically, this step aims to perform feedforward control actions on the system. Subsequently, the actual data collected by the infrared sensor 13 CO2 concentration Compared with the previous prediction model output value By comparing the results, we can determine whether the current control measures have successfully achieved the target concentration control, and construct a system deviation identification and trend risk characterization model.

[0132] The goal of this step is not to directly output control commands, but to structure and encode the detected prediction bias as input to the next policy optimization module (step five). This structured bias information should have the following characteristics: ① real-time updates, ② quantifiability, ③ ability to distinguish the source of disturbances, and ④ usability as a criterion for model updates or policy reconstruction.

[0133] enter: The valve opening and closing time control quantity generated in the previous step; The pump power control value generated in the previous step; The predicted future concentration values ​​generated in step two; The actual concentration value is obtained by a non-dispersive infrared (NDIR) sensor (e.g., real-time access via an RS-485 interface); h t The system state representation obtained by Transformer encoding is used to understand potential fluctuations under current environmental conditions.

[0134] Furthermore, the specific procedures are as follows:

[0135] Error construction:

[0136] First, this invention calculates the prediction error e. t+τ :

[0137]

[0138] This error represents the deviation between the predicted and actual concentrations, and can be positive (prediction underestimated) or negative (prediction overestimated).

[0139] Construction of biased attribution model:

[0140] Since prediction bias can be caused by various factors, such as inaccurate control behavior, system disturbances (temperature and humidity), and modeling bias, this invention designs a prediction method based on the system state representation h. t and control actions The bias representation model outputs a coded quantity ∈ t :

[0141]

[0142] in:

[0143] ∈ t It encodes the deviation risk and can be used for subsequent strategy accumulation optimization, control weight adjustment, model retraining triggering, etc.

[0144] MLP δ It is a three-layer feedforward network. The first layer is ReLU activation, the second layer is linear mapping to the intermediate state, and the final output is a Tanh code in the range of [-1,1], which expresses the direction and intensity of the deviation risk.

[0145] Furthermore, all input quantities are preprocessed by normalization (e.g., concentration ppm → [0,1], power % → [0,1]); model training is based on supervised regression using historical control execution records and concentration acquisition trajectories.

[0146] It is important to note that this step does not directly adjust the control action; it only outputs a deviation structure quantity ∈ for subsequent strategy optimization. t Example: If ∈ t = -0.8 indicates a negative prediction bias (i.e., the prediction is too high and the actual concentration is low). In this case, the valve opening time can be appropriately extended or the pumping rate reduced in the next cycle. Conversely, if ∈ t = +0.7 indicates that the system may be under-predicting, the actual concentration has increased, and the subsequent strategy should strengthen the gas extraction response.

[0147] Output: ∈ t The concentration control deviation risk code in the current state of the system can be used for the control strategy optimization module in the next cycle.

[0148] S5: Receive the current state representation and deviation risk code to establish a state-deviation mapping model, fit the deviation distribution parameters of the control deviation, and optimize the long-term strategy.

[0149] Specifically, the key to this step is designing a long-term deviation behavior modeling module, which continuously receives control states h t Corresponding deviation ∈ t Based on the data pairs, a state-response mapping model is established, thereby forming a control experience base with the following three capabilities:

[0150] Determine whether there were any systematic biases in past control measures under these conditions.

[0151] To provide a reference for future model structure adjustments or prediction confidence assessments;

[0152] Provide data support for the division of applicable regions or the setting of experimental groups for strategy application.

[0153] Furthermore, this invention formalizes the modeling task as a density learning problem, the objective of which is to fit the probability density estimation model p(∈) of the bias. t |h t ):

[0154]

[0155] in:

[0156] The model employs a dual-head neural network: it outputs μ(h) respectively. t ) and σ 2 (h t );

[0157] Enter h t The input is encoded into an intermediate hidden layer through a two-layer ReLU activated feedforward neural network (e.g., 128-dimensional), outputting two scalars (mean and variance).

[0158] This structure represents h in the current state. t Below, the deviation generated by the control behavior ∈ t Distribution, possessing uncertainty modeling capabilities;

[0159] Furthermore, parameter training utilizes historical datasets. The model is trained using the maximum likelihood loss function; it can be deployed and saved as a JSON structure for use by downstream strategy modules or operations and maintenance departments.

[0160] For example, the sample obtained after 10 sets of control-feedback experiments is:

[0161]

[0162] This invention feeds these data into a bias modeling network, and learns the target p(∈) by maximizing the likelihood of a normal distribution. t |h t The structure of this model is as follows. This model does not participate in real-time control, but only serves as an empirical structural expression of the "control deviation distribution under state".

[0163] The key innovation of this mechanism lies in:

[0164] This invention is the first to decouple "control deviation modeling" from the immediate feedback mechanism, storing and expressing it separately in a structured manner; it uses "state-deviation" joint modeling to form a cross-batch control experience expression channel, allowing for long-term system evolution and model migration; the model is in the form of a density function, rather than a single regression output, which has stronger generalization and risk identification capabilities.

[0165] Output M bias A callable control deviation response modeler that supports interface-based input state h t Output control deviation distribution parameter μ(h) t ),σ(h t (This can serve as an auxiliary reference for future experiments or strategy generation modules.)

[0166] 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.

[0167] 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.

[0168] 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.

[0169] 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.

[0170] 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.

[0171] 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.

[0172] 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 method for intelligent control of isotope gas concentration based on concentration prediction, characterized in that, The method includes the following steps: S1: Collect multidimensional data from the emptied sealed enclosure to construct a standardized control-response data sequence; S2: Using a neural network model, the concentration of isotopic gases in a closed environment is predicted to obtain the predicted concentration of isotopic gases. S3: Identify the current state representation of the current standardized control-response data sequence, and combine the predicted concentration with the target concentration to generate control commands for valve opening and closing time and pump power; S4: Execute the control command, compare the current actual concentration with the predicted concentration, and encode the deviation into a deviation risk code through a multilayer perceptron; S5: Receive the current state representation and deviation risk code to establish a state-deviation mapping model, fit the deviation distribution parameters of the control deviation, and optimize the long-term strategy.

2. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 1, characterized in that, The multidimensional data includes CO2 concentration monitored in real time by infrared sensors, wind speed collected by hot-film anemometers, temperature and humidity from temperature and humidity sensors, valve on / off status, and pump power records. The sampling frequency is no less than 1Hz, and the data is aligned with a unified timestamp by the controller.

3. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 1, characterized in that, The structure of the neural network model is as follows: The encoder part uses a two-layer Transformer structure to encode timing information, and the decoder part is a linear mapping module that fuses state embeddings and trend terms; The prediction expression of the neural network model is as follows: in, X′ represents the predicted CO2 concentration τ seconds from now. t-L+1:t Θ is a normalized input time window of length L; Attn(·) is an encoder module with physical attention guidance, whose weight adjustment part is controlled by two sets of structural variables; phys Θ is the tensor of physical parameters for gas diffusion; ctrl The parameters for controlling the perturbation estimation are: W, W, and Δ. trend (t) is the long-term concentration drift trend estimate, used to handle the slow shift caused by the decrease in system airtightness or long-term background gas infiltration.

4. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 3, characterized in that, The loss function of the neural network model is expressed as: in, C is the loss function for the neural network model. t+τ The actual observed concentration value was collected by an infrared CO2 sensor; For the local first-order difference of the predicted sequence; φ ctrl λ1 and λ2 are the expected concentration change slope estimated based on the amplitude of abrupt changes in the control signal; λ1 and λ2 are the weight hyperparameters of the regularization term. in, This is used to ensure that the predicted value is not less than zero and to avoid physical inconsistencies.

5. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 1, characterized in that, The neural network model uses a sliding window method to construct sample sequences from historical data during the model training phase. Each window has a length of L = 60 and a sliding step size of 1 second. All model training, tuning, and deployment are completed by the host computer, built using the PyTorch platform, and support dynamic online updates of physical prior tensors for later system adaptation to different experimental container structures or air duct configurations.

6. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 1, characterized in that, The steps in S3 specifically include: The control strength ρ is calculated using a structurally interpretable policy function. t This is used to reflect the current system's demand for control. Among them, the first item This indicates the necessary control strength based on the prediction error; ∈ is a small constant to prevent division by zero; the second term is MLP. res (h t The dynamic disturbance adjustment term is based on the system state and consists of a two-layer MLP structure. res (h t The output implies the impact of sudden system disturbances, unstable trends, or historical disturbance memories; h t The current state is represented by α; α is the adjustment hyperparameter, which represents the proportion of the influence of the control state on the final control response. Based on control intensity ρ t Two specific control output quantities are derived: Where u0,P0 are the current control state reference values; λ u ,λ P : These are the mapping coefficients from control strength to actual control variables; clip(·): are the upper and lower bound functions to prevent output overload; This is the valve opening and closing time for the next cycle; The operating power of the air pump for the next cycle; u min Minimum valve opening and closing time; u max P represents the maximum valve opening and closing time. min This represents the minimum operating power of the air pump; P max This represents the maximum operating power of the air pump.

7. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 1, characterized in that, The multilayer perceptron is a three-layer feedforward network. The first layer is ReLU activation, the second layer is linear mapping to intermediate states, and the final output is a Tanh code in the range of [-1,1], which expresses the direction and intensity of the bias risk.

8. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 1, characterized in that, The state-deviation mapping model uses a dual-head neural network to output deviation distribution parameters.

9. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 8, characterized in that, The objective of the state-deviation mapping model is to fit the probability density estimate of the deviation, calculated as follows: The dual-head neural network outputs the mean μ(h) representing the current state. t The variance of the current state representation is σ. 2 (h t ), which is the parameter for controlling the deviation distribution; the input in the current state represents h. t The input is encoded into an intermediate hidden layer via a two-layer ReLU activated feedforward neural network, which outputs the mean μ(h) of the current state representation. t The variance of the current state representation is σ. 2 (h t ); The parameter training uses historical datasets. It is obtained through training using the maximum likelihood loss function.

10. The intelligent control method for isotope gas concentration based on concentration prediction according to claim 3, characterized in that, S2 also generates a long-term concentration drift trend estimate; this long-term concentration drift trend estimate is used to handle the slow shift caused by decreased system airtightness or long-term background gas infiltration, and to take into account the chronic background gas infiltration effect over a long experimental period; the expected concentration change slope φ estimated based on the abrupt change amplitude of the control signal... ctrl The gas diffusion physics parameter tensor is obtained by adaptive fitting with historical mutation data; it is used to adjust the attention weight path to form structural physics guidance.

Citation Information

Cited By

  • Method and system for predicting steady-state concentration of carbon dioxide in photosynthesis leaf chamber

    CN121709068A

  • Marine dissolved gas sensor data denoising and time-delay correction real-time processing method

    CN121880941A