A PINN-based dac performance prediction and control method and system

Abstract

The application discloses a DAC performance prediction and control method and system based on PINN, comprising the following steps: constructing and training an inverse physical information neural network, inputting each sensor data at different times, outputting a CO2 concentration prediction value, and iterating to minimize the total loss function; inputting real-time sensor data of array arrangement on the adsorption bed into the network to obtain an optimal deactivation parameter vector, and diagnosing the deactivation dominant mechanism and predicting the remaining service life of the adsorption bed according to the deactivation parameter vector; if the overall remaining service life of the current adsorption bed is lower than a preset safety threshold, triggering an optimization control operation, and after the adjustment is completed, continuing the performance prediction. The application can independently and in real time perform sensing, performance prediction, dominant mechanism diagnosis, decision-making and execution.

Description

Technical Field

[0001] This invention relates to the field of carbon negative technology in environmental engineering, and in particular to a method and system for predicting and controlling DAC performance based on PINN. Background Technology

[0002] Against the backdrop of global efforts to address climate change and achieve carbon neutrality, Direct Air Capture (DAC) technology, which directly captures carbon dioxide (CO2) from the atmosphere, is considered one of the key technological pathways to achieve net-zero or even negative emissions. Thermochromic / Pressure Swing Adsorption (TSA / PSA) based on solid adsorbents is the mainstream DAC technology. Its core process involves using solid adsorbent materials with special pore structures and surface chemistry (such as amino-functionalized silicon-based materials and metal-organic frameworks (MOFs)) to selectively capture low concentrations of CO2 molecules from the air at ambient temperature and pressure. Once the adsorbent reaches saturation, desorption is achieved by increasing the temperature or decreasing the pressure, releasing a high concentration of CO2 gas for subsequent Carbon Capture and Storage (CCS) or Carbon Capture and Utilization (CCU). The adsorbent is then regenerated and enters the next adsorption cycle.

[0003] In this cyclical process, the adsorbent is the "heart" of the entire DAC system, and its performance stability and lifespan directly determine the system's capture efficiency, energy consumption, and ultimate operating costs. Therefore, effectively monitoring and managing the long-term performance of the adsorbent is one of the core issues driving the large-scale commercial application of DAC technology. Currently, the following technical solutions are commonly used to assess and predict the performance degradation (i.e., "deactivation") of the adsorbent in DAC systems:

[0004] (1) Offline sampling and laboratory characterization analysis: After the DAC equipment has been running for a certain period of time, it is shut down. A small amount of adsorbent sample is taken from different positions of the adsorption bed and sent to the laboratory. The performance is characterized by professional analytical instruments. By comparing the performance parameters of the new and old materials, the degree of aging is quantitatively evaluated.

[0005] This method suffers from severe lag, making online, real-time monitoring impossible. By the time a significant performance degradation is detected, the adsorbent has often been operating inefficiently for hundreds or even thousands of previous cycles, resulting in long-term energy and efficiency losses. Furthermore, this method lacks representativeness and is disconnected from actual operating conditions: gram-level samples taken from a large adsorption bed rarely fully represent the average condition of the entire bed, especially when aging is unevenly distributed along the bed. Additionally, ideal laboratory testing conditions (such as pure gas and constant temperature and humidity) cannot reproduce the complex and dynamically changing real-world operating conditions of the DAC (such as humidity fluctuations and the influence of trace contaminants), leading to discrepancies between the evaluation results and actual performance.

[0006] (2) Empirical model prediction based on macroscopic operating data: This method does not directly analyze the material itself, but rather establishes a macroscopic, empirical correlation model based on the long-term historical operating data of the DAC system through statistical analysis to describe the relationship between the system capture efficiency and the cumulative operating time or number of cycles; based on this type of model, the factory formulates a fixed, time-based maintenance plan.

[0007] This method lacks mechanistic insights and cannot guide optimization: the empirical model is a "black box" that can only describe "performance degradation" but cannot answer "why the degradation occurred" or "which key performance parameter degraded." Therefore, its only guidance is the crude "periodic replacement," which cannot achieve refined operation and maintenance. The method also has poor environmental adaptability: an empirical model built in a specific location (such as a dry inland area) will completely fail once applied to another location (such as a humid coastal area) due to significant differences in environmental factors (such as average annual humidity). The model lacks the ability to adapt to different or changing operating environments.

[0008] (3) Detailed mechanistic model simulation based on first principles: This is a method based on theoretical calculations. Researchers construct detailed mathematical models that include fluid dynamics (CFD), heat and mass transfer, adsorption kinetics, and complex chemical reaction networks, and perform high-precision simulations of the physicochemical processes within the adsorbent bed on a computer. By introducing hypothetical deactivation reactions (such as hydrolysis caused by water vapor, functional group degradation caused by oxidizing gases, etc.) into the model, the long-term evolution trend of adsorbent performance under specific operating conditions can be simulated and predicted. This type of model is mainly used in the theoretical research of new materials and the design stage of adsorption reactors.

[0009] This method is computationally extremely expensive and cannot be applied online: the computational demands of such high-fidelity models are enormous, with a single simulation typically requiring hours or even days, completely lacking the capability for real-time calculation and parameter identification within minute-level DAC runtime cycles. Furthermore, its role is that of a design tool, not an online diagnostic tool.

[0010] In summary, existing technologies suffer from drawbacks such as lag, black-box characteristics, poor environmental adaptability, and high computational costs. There is a need to develop an online, real-time method that incorporates physical mechanisms, simultaneously identifying the decay rates of multiple key performance parameters (such as diffusion coefficient and adsorption capacity) of the adsorbent under real, complex operating conditions. This would enable accurate prediction of the adsorbent's remaining lifetime and in-depth diagnosis of the deactivation mechanism, providing technical support for intelligent, refined maintenance and cost optimization of DAC systems. Summary of the Invention

[0011] To address the shortcomings of existing technologies in simultaneously overcoming the problems of evaluation lag, operational condition disconnect, mechanism ambiguity, and spatial non-uniformity, as well as the inefficiency and high cost of adsorbent management and maintenance, this invention proposes a DAC performance prediction and control method and system based on PINN.

[0012] The specific technical solution is as follows:

[0013] A PINN-based method for DAC performance prediction and control includes the following steps:

[0014] S1: For the modularly arranged adsorption bed in the direct air capture system, acquire real-time sensing data from multiple sensors arrayed within the adsorption bed at different times.

[0015] S2: Construct and train an inverse physical information neural network that couples a physical model describing the adsorption process with an empirical function describing the deactivation of the adsorbent material. The input is the sensor data, and the output is a predicted CO2 concentration. It iterates to minimize the total loss function, which includes data loss and physical loss, based on the effective diffusion coefficient D. eff Maximum adsorption capacity q max The network prediction error is updated; the D eff and q max Calculated from the deactivation parameter vector;

[0016] S3: Input the real-time sensing data collected by S1 into the inverse physical information neural network to solve for a set of deactivation parameter vectors that quantify multiple deactivation mechanisms. ;

[0017] S4: According to the above The numerical values ​​of each parameter are used to diagnose the dominant mechanism of deactivation and predict the remaining service life of the adsorption bed.

[0018] S5: Determine whether the remaining service life of the current adsorption bed is lower than the preset safety threshold. If not, continue normal operation and predict the performance of the direct air capture system. If so, trigger the optimization control operation, complete the adjustment, and continue normal operation and predict the performance of the direct air capture system.

[0019] Furthermore, the deactivation parameter vector includes: cyclic aging deactivation rate constant k1, humidity-induced deactivation rate constant k2, contaminant poisoning deactivation rate constant k3, and capacity decay rate constant k. q .

[0020] Furthermore, in S2, the inverse physical information neural network includes: a fully connected feedforward neural network structure module, a deactivation function calculation module, a loss calculation module, and an optimizer module;

[0021] The fully connected feedforward neural network structure module includes an input layer, a hidden layer, and an output layer; the input layer receives the normalized variable (t, z), and the output layer outputs the predicted CO2 concentration C. pred (t,z); where z is the position coordinate along the axial direction of the adsorption bed;

[0022] The inverse physical information neural network mathematizes and couples the physical model describing the adsorption process with the empirical function describing material deactivation.

[0023] The physical model is based on the mass conservation relationship of the gas phase in the adsorption bed and the linear or nonlinear driving force relationship of the solid phase adsorption kinetics. It imposes simultaneous constraints on the gas phase CO2 concentration and the adsorption amount of the solid phase adsorbent: the gas phase part follows the conservation relationship between the transport of substances along the axial direction of the adsorption bed and the source terms generated by solid phase adsorption and desorption; the solid phase part follows the kinetic relationship that the actual adsorption amount is equal to the equilibrium adsorption amount corresponding to the current gas phase conditions. The two types of relationships work together to ensure that the output of the inverse physical information neural network simultaneously satisfies the coupling requirements of fixed bed mass conservation and linear driving force adsorption kinetics in both time and space.

[0024] The deactivation function calculation module includes D eff and q max The calculation of the current value of the effective diffusion coefficient D; eff The current value is based on the number of adsorption-desorption cycles completed (N), the relative humidity of the inlet air (H(t)), and the pollutant concentration (C). p (t) in a single cycle duration t cycle The monotonic decay relationship, including the cumulative exposure within the range, is expressed as an initial value. Using k1 as the baseline, k2 as the weight parameter of N, k3 as the weight parameter of H(t), and k3 as the weight parameter of C... p The weighting parameters of (t) are determined; the maximum adsorption capacity q max The current value is determined based on a monotonically decaying relationship related to the cumulative cycle or runtime, starting from the initial value. Based on k q The parameters for N are determined; the {k1, k2, k3, k q} constitute the deactivation parameter vector It is updated synchronously with the network parameters during training.

[0025] Furthermore, in S2, the total loss function is composed of the data loss L. data and physical loss L physics We get the result by weighted summation;

[0026] The data loss L data Used to measure z at each sensor location i Location and sampling time t j Below, the network outputs the predicted concentration values. The overall mean square error relative to the corresponding measured value;

[0027] The physical loss L physics Used to measure the rate of change of the gas and solid phase fields represented by the network output relative to time and space; physical loss L physics The calculation specifically involves: calculating the output C of the fully connected feedforward neural network through automatic differentiation. pred The partial derivatives of the input variables (t, z) are substituted into the system of partial differential equations to obtain the residuals of the system of partial differential equations. The mean square values ​​of these residuals at all locus points in the spatiotemporal domain are then calculated to obtain the physical loss L. physics .

[0028] Further, in step S4, the step of predicting the remaining service life of the adsorbent includes: substituting the deactivation parameter vector into a preset decay function used to describe the effective diffusion coefficient and / or maximum adsorption capacity to obtain a decay curve of the effective diffusion coefficient and / or maximum adsorption capacity changing with the number of future cycles; and calculating the remaining service life based on the decay curve and its preset failure threshold.

[0029] Furthermore, S4 also includes: a multi-objective optimization module aimed at minimizing the total lifecycle capture cost, based on the... Generate the regeneration temperature T that combines positive and negative economic impacts. reg Adjustment strategy;

[0030] The logic of the multi-objective optimization module is as follows: when deactivation acceleration is detected, the positive and negative economic impacts of increasing or decreasing the set temperature step size of the regeneration temperature are evaluated respectively, thereby calculating an optimal regeneration temperature that minimizes the expected value of the total life cycle capture cost, and sending it as a set value to the direct air capture system.

[0031] Furthermore, the estimated total lifecycle capture cost (LCOC) is obtained by dividing the total cost per cycle by the amount of CO2 captured per cycle.

[0032] The total cost of a single cycle includes the variable cost of the single cycle and the allocated cost of capital;

[0033] The CO2 capture amount per cycle is the actual capture amount, obtained by subtracting the theoretical capture amount from the actual capture amount using a capture efficiency factor. The theoretical capture amount is the product of the total mass of adsorbent in the adsorption bed and the effective working capacity. The capture efficiency factor is obtained based on a material balance model at a characteristic time, which is based on the effective length of the adsorption bed. The degree of mass transfer limitation is determined by combining structural and operating parameters, including the overall mass transfer coefficient and the apparent flow rate of the gas. Based on the degree of mass transfer limitation, the reduction ratio of the actual capture amount to the equilibrium capacity is calculated, and this reduction ratio is used as the value of the capture efficiency factor.

[0034] Furthermore, in S5, the optimized control operation includes:

[0035] (1) Execute the preset life extension priority operation mode and control the operating frequency of specific components according to the preset instructions to slow down the aging process of the direct air capture system;

[0036] (2) Adjust the corresponding operating parameters according to the dominant inactivation mechanism obtained from the diagnosis;

[0037] (3) Determine whether the performance of an adsorbent module is below the performance failure threshold. If the performance of the upstream adsorbent module is below the performance failure threshold and the downstream adsorbent module still has a large margin, the operation and maintenance platform generates a maintenance instruction and sends a high-level warning to the user, prompting them to prepare spare parts and arrange a maintenance plan.

[0038] Furthermore, in step S5, if the remaining service life of the entire adsorption bed is lower than the emergency replacement threshold, an adsorption bed switching is performed, and the airflow is introduced into the standby adsorption bed to ensure production continuity.

[0039] A PINN-based DAC performance prediction and control system, used to implement the PINN-based DAC performance prediction and control method, includes: a variable frequency blower, an electrically controlled switching valve, a heater, several adsorption beds, a computing unit, and a control execution unit; airflow enters the system pipeline from an air inlet, and the variable frequency blower, electrically controlled switching valve, heater, and one adsorption bed are sequentially and coaxially arranged in the pipeline along the airflow direction, with the airflow finally exiting from an exhaust port; the adsorption beds are designed to be detachable and replaceable, with the remaining adsorption beds serving as backups; the variable frequency blower is used to control the intake airflow, the electrically controlled switching valve is used to control the switching between the working adsorption bed and the backup adsorption bed, and the heater is used to precisely regulate the temperature of the regeneration heat fluid;

[0040] Each adsorption bed includes an environmental sensor and multiple adsorbent modules arranged along the gas flow direction. Each adsorbent module can be independently disassembled and replaced. The environmental sensor is used to detect information at the inlet of the adsorption bed, including: air temperature, relative humidity, and pollutant concentration. Each adsorbent module is equipped with a fiber optic sensor at its outlet to detect information within the corresponding adsorbent module, including: specific gravity of adsorbent particles, apparent gas flow rate, and CO2 concentration.

[0041] The environmental sensor and the fiber optic sensor of the adsorbent module are both connected to the computing unit, which is connected to the control execution unit. After processing the received sensor data, the computing unit sends control commands to the control execution unit. The control execution unit is connected to and controls the variable frequency blower, the electrically controlled switching valve, and the heater, respectively.

[0042] The beneficial effects of this invention are:

[0043] (1) It has achieved a fundamental leap from “offline, delayed, black box” assessment to “online, real-time, mechanism-level” diagnosis.

[0044] This invention combines a reverse physical information neural network (PINN) with a distributed sensor array. This distributed, multi-point sensing scheme enables high-precision spatial resolution identification of adsorbent aging states, providing the necessary data foundation for the PINN model to accurately decouple various aging mechanisms with uneven distribution along the bed (e.g., more severe pollutant poisoning in the inlet module). The PINN architecture can process high-dimensional spatiotemporal data in real time, while the sensor array provides precise boundary conditions that reflect the internal gradient of the bed, thus providing a precise basis for subsequent targeted maintenance of specific modules. Existing technologies that rely on offline characterization through downtime sampling (lagging and unrepresentative) or on building empirical models based on macroscopic historical data (black box and unable to extrapolate) cannot simultaneously overcome four major technical challenges: assessment lag, operational condition disconnect, mechanistic ambiguity, and spatial non-uniformity.

[0045] In summary, this invention enables the online acquisition of a high-fidelity "digital twin" of the adsorbent's health status, synchronized in real time with the physical entity. This "digital twin" not only displays the current performance value but also reveals the underlying physical decay rate, achieving a qualitative leap in diagnostic capabilities.

[0046] (2) It has achieved a deep understanding of the transformation from “single performance characterization” to “multidimensional decoupling of inactivation mechanism”.

[0047] This invention innovatively incorporates multiple decay parameters describing cyclic aging, moisture effects, contaminant poisoning, and active site loss as latent variables to be solved, embedding them into a unified PINN model for simultaneous inversion. Existing adsorbent testing technologies, even through experiments, typically only measure comprehensive indicators such as total adsorption capacity or a single mass transfer coefficient, failing to isolate the contributions of different aging factors and thus unable to answer the fundamental question of "why performance declines," leading to technical limitations where maintenance decisions are based on guesswork.

[0048] This invention can quantitatively decouple and identify the technical effects of the dominant inactivation mechanism at the current stage, which upgrades the subsequent maintenance strategy from a "one-size-fits-all" replacement to a "targeted" precise intervention (for example, if it is affected by moisture, then optimize regeneration; if it is due to loss of active points, then replacement is necessary).

[0049] (3) It has realized the intelligent transformation from "passive manual operation and maintenance" to "predictive, economically optimal autonomous closed-loop control".

[0050] This invention combines accurate diagnostic and predictive results with positive and negative economic impacts to generate a strategy for adjusting regeneration temperature, and incorporates a modular hardware design. Existing technologies, which fragment the monitoring, alarm, decision-making, and execution processes, relying entirely on manual operation and maintenance based on fixed procedures or experience, suffer from technical drawbacks such as slow response, inability to dynamically adapt to changing operating conditions, high maintenance costs, and significant resource waste.

[0051] This invention transforms the entire direct air capture system performance prediction and control system from a passive tool into an intelligent entity capable of autonomously performing "sensing-diagnosis-prediction-decision-execution"; the system can automatically find the optimal balance between energy consumption and material loss, proactively avoid risks, and precisely guide targeted maintenance, ultimately maximizing the economic benefits throughout the system's entire life cycle. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the architecture of the DAC performance prediction and control system based on PINN in an embodiment of the present invention.

[0053] Figure 2 This is a flowchart of the DAC performance prediction and control method based on PINN in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the structure and training loop of the inverse PINN model in an embodiment of the present invention.

[0055] In the diagram, 1 is the air inlet, 2 is the variable frequency blower, 3 is the electrically controlled switching valve, 4 is the heater, 5 is the environmental sensor, 6 is the first adsorbent module, 7 is the second adsorbent module, 8 is the third adsorbent module, 9 is the exhaust port, 10 is the computing unit, and 11 is the control execution unit. Detailed Implementation

[0056] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The objectives and effects of the present invention will become clearer as a result. The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0057] like Figure 1 As shown, a PINN-based DAC performance prediction and control system includes: a variable frequency blower 2, an electrically controlled switching valve 3, a heater 4, several adsorption beds, a computing unit 10, and a control execution unit 11. Airflow enters the system pipeline from the air inlet 1. The variable frequency blower 2, the electrically controlled switching valve 3, the heater 4, and one adsorption bed are sequentially and coaxially arranged in the pipeline along the airflow direction (i.e., the positive z-axis direction). The airflow is finally output from the exhaust port 9. The adsorption beds are designed to be detachable and replaceable, with the remaining adsorption beds serving as backups. The variable frequency blower 2 controls the intake airflow, the electrically controlled switching valve 3 controls the switching between the working and backup adsorption beds, and the heater 4 precisely regulates the temperature of the regeneration heat fluid.

[0058] Each adsorption bed includes an environmental sensor 5 and multiple adsorbent modules arranged along the gas flow direction. The inlet of the adsorption bed is defined as z=0, and the outlet as z=L. In this embodiment, three adsorbent modules are arranged: a first adsorbent module 6, a second adsorbent module 7, and a third adsorbent module 8. Each adsorbent module can be independently disassembled and replaced. Each adsorbent module has a built-in fiber optic sensor at its outlet to detect information such as the specific gravity of the adsorbent particles, the apparent flow rate of the gas, and the CO2 concentration within the corresponding adsorbent module. The environmental sensor 5 is used to detect the air temperature, relative humidity, and pollutant concentration at the inlet of the adsorption bed.

[0059] Both the environmental sensor 5 and the fiber optic sensor of the adsorbent module are connected to the computing unit 10, which in turn is connected to the control execution unit 11. The control execution unit 11 can receive commands via a standard industrial bus (such as Modbus TCP / IP). The control execution unit 11 is connected to and controls the variable frequency blower 2, the electrically controlled switching valve 3, and the heater 4, respectively. In a typical adsorption-desorption cycle, the system collects data detected by the two types of sensors at a fixed sampling frequency (e.g., 1Hz), and appends the current cycle number N and a timestamp. The collected data is transmitted to the computing unit 10 for further processing, and the computing unit 10 sends control commands to the control execution unit 11 accordingly to execute the control of the corresponding components.

[0060] To provide high-fidelity input data, this system abandons a single outlet analyzer and deploys a multi-dimensional sensor array to achieve high spatial resolution CO2 concentration sensing. Distributed fiber optic sensors are installed at multiple key locations along the z-axis of the adsorption bed (e.g., at z=0.3L, z=0.6L, and z=L for the three adsorbent modules). Specifically, fiber optic probes based on tunable diode laser absorption spectroscopy (TDLAS) technology are used. This technology features high selectivity and millisecond-level response speed, enabling the system to capture the dynamic propagation process of the CO2 adsorption front within the bed in real time, which is crucial for improving the system's identification accuracy.

[0061] Using the aforementioned PINN-based DAC performance prediction and control system, this embodiment also proposes a PINN-based DAC performance prediction and control method, such as... Figure 2 As shown, the method includes the following steps:

[0062] S1: In a typical adsorption-desorption regeneration cycle, the system acquires data from environmental sensor 5 and fiber optic sensor in real time at a fixed sampling frequency.

[0063] S2: Construct and train a Physics-Informed Neural Network (PINN) capable of retrieving internal inactivation parameters, such as... Figure 3 As shown, the inverse PINN includes: a fully connected feedforward neural network (FNN) structure module, a deactivation function calculation module, a loss calculation module, and an optimizer module.

[0064] The FNN structure module includes an input layer, multiple hidden layers, and an output layer; the input layer receives the normalized variables (t, z), and the output layer outputs the predicted CO2 concentration C. pred(t,z); where t is time [s] and z is the position coordinate [m] along the axial direction of the adsorption bed.

[0065] The reverse PINN approach mathematically couples the physical model describing the adsorption process with the empirical function describing material deactivation. The physical model, based on the gas-phase mass conservation relationship of the adsorption bed and the linear or nonlinear driving force relationship of solid-phase adsorption kinetics, imposes simultaneous constraints on the gas-phase CO2 concentration and the adsorption capacity of the solid-phase adsorbent. Taking the linear driving force relationship as an example, the physical model uses a linear driving force (LDF) model (a type of adsorption equilibrium model) to describe the fixed-bed mass balance expression. The partial differential equations (PDEs) formed by the two (LDF model and fixed-bed mass balance expression) are as follows:

[0066] (1)

[0067] In the formula, C is the molar concentration of CO2 in the gas phase [mol•m]. -3 Hereinafter referred to as gas phase concentration, is a function of time t and position z; u is the apparent flow rate of the gas in the adsorption bed [m•s]. -1 ], ρ b The bulk density of the adsorbent particles [kg•m] -3 ], q represents the bed porosity of the adsorption bed [dimensionless]; q represents the average molar adsorption capacity of the solid-phase adsorbent [mol•kg]. -1 ] is a function of time t and position z. k LDF The mass transfer coefficient of the total package [s] -1 Its value is related to the effective diffusion coefficient D. eff Positive correlation, for example , q represents the particle radius of the adsorbent. eq To determine the equilibrium molar adsorption capacity [mol•kg] -1 The adsorption capacity q represents the amount of gas at which adsorption reaches equilibrium at the current gas phase concentration C. It is typically described by the Langmuir or Toth isotherm model, and its expression includes the maximum adsorption capacity q. max .

[0068] The deactivation function calculation module includes the effective diffusion coefficient D. eff The decay function and maximum adsorption capacity q maxThe decay function. Experimental results show that after the adsorbent decays, it exhibits a decrease in two apparent parameters: the effective diffusion coefficient (porous media blockage) and the maximum adsorption capacity. Therefore, in this embodiment, the effective diffusion coefficient and the maximum adsorption capacity are changed from constants to functions that include the unknown decay parameters to be solved, which helps to quantitatively decouple and identify the dominant deactivation mechanism at the current stage.

[0069] Effective diffusion coefficient D eff The decay function is as follows:

[0070] (2)

[0071] In the formula, D eff,0 The initial effective diffusion coefficient of the new material [m] 2 ·s -1 [ ], as a known baseline value; N is the number of adsorption-desorption cycles completed so far [dimensionless], H(t) is the real-time measured relative humidity of the inlet air [%RH], C p (t) represents the real-time measured concentration of inlet air pollutants [ppm or mol·m]. -3 ], t cycle is the total time of a single cycle [s]. k1 is the cycle aging deactivation rate constant [dimensionless], an implicit variable to be solved, quantifying the performance degradation caused purely by thermodynamic cycling; k2 is the humidity-affected deactivation rate constant [(%RH·s)]. -1 ] represents the implicit variable to be solved, quantifying the negative impact of water vapor accumulation on mass transfer performance; k3 is the pollutant poisoning inactivation rate constant [(ppm·s)]. -1 or (mol•m -3 ·s) -3 ] represents the implicit variable to be solved, quantifying the toxic effect of the cumulative effect of a specific pollutant on mass transfer performance.

[0072] Maximum adsorption capacity q max The decay function is as follows:

[0073] (3)

[0074] In the formula, q max,0 The initial maximum molar adsorption capacity of the new material [mol·kg] -1 ], as a known baseline value; k q Let be the capacity decay rate constant [dimensionless], and let be the implicit variable to be solved, quantifying the permanent decrease in adsorption capacity caused by the loss of active sites.

[0075] The set of latent variables to be solved yields the deactivation parameter vector. This vector, as part of the model, is updated by inverse solving using external observation data.

[0076] In the loss calculation module, the total loss function includes data loss and physical loss. The training objective of PINN is to minimize the total loss function, which is expressed as follows:

[0077] (4)

[0078] In the formula, w data w is the weighting factor [dimensionless] for the data loss term. physics The two weighting factors are dimensionless weighting factors for the physical loss term. These factors balance the importance of data fitting and physical constraints and are adjustable hyperparameters, specifically adjusted through the optimizer module (such as the Adam optimizer). L represents the total loss. data L represents the data loss, used to measure the mean square error between the FNN predictions and the measurements from all spatial sensing points. physics This is a physical loss.

[0079] Specifically, data loss L data The expression is as follows:

[0080] (5)

[0081] In the formula, K is the total number of sensors in the sensor array [dimensionless], and M is the number of time-series data points collected by each sensor [dimensionless]. For FNN at time point t j Position z i Predicted concentration at [mol•m] -3 ], For the sensor at time point t j Position z i Concentration measurement at [mol•m] -3 ].

[0082] Physical loss L physics The calculation specifically involves: calculating the FNN output C through automatic differentiation. pred Partial derivatives with respect to the input variable (t, z) (e.g.) , Substitute these values ​​into equation (1) to obtain the PDE residuals. Calculate the mean square value of these residuals at all locus points in the spatiotemporal domain to obtain the physical loss L. physics Crucially, when calculating this PDE residual, D eff and q max It is based on the vector of deactivation parameters to be solved. The expressions (2) and (3) are used for calculation.

[0083] During the training of the inverse PINN, in each evaluation cycle (i.e., the time to perform a complete DAC performance prediction, such as every 24 hours), computing unit 10 performs the following operations:

[0084] (a) Extract the data collected in the most recent evaluation periods as the training set of sensor measured data.

[0085] (b) Initialize the network weights of the inverse PINN (including w) data and w physics ) and the deactivation parameter vector to be solved .

[0086] (c) Input the training set data into the inverse PINN, and use optimizers such as Adam to minimize the total loss function L through the backpropagation algorithm; during the optimization process, the weights and inactivation parameter vectors of the inverse PINN are... It will be updated simultaneously. Iterate and optimize until the total loss function converges, complete the network weight update, and obtain the trained inverse PINN.

[0087] S3: Using the real-time sensing data collected in S1 as training data, the inverse physical information neural network is trained online, and the inactive parameter vector is updated through an iterative optimization algorithm. This continues until the total loss function converges. At this point, the inverse PINN not only learns the adsorption process in the current state, but also obtains a set of quantified inactivation parameter vectors that best explain the observed breakthrough curve behavior. This enables online parameter recognition.

[0088] S4: Based on the inactivation parameter vector, simultaneously perform inactivation-dominant mechanism diagnosis, Remaining Useful Life (RUL) prediction [unit: number of cycles or hours], and regeneration strategy adjustment, as detailed below:

[0089] (4.1) Diagnosis of the dominant inactivation mechanism: comparison The magnitude of each component can be used to diagnose the dominant mechanism of inactivation. For example, if the value of k3 is much larger than that of other components, it indicates that pollutants in the current air are the primary cause of performance degradation.

[0090] (4.2) RUL Prediction: Substituting expressions (2) and (3), we can extrapolate D. eff and q max The decay curve over the next N cycles. Combined with the preset D. eff and q maxThe system can accurately predict the RUL (Recovery Limit) of each adsorbent module by identifying the performance failure threshold (i.e., the minimum value). The average RUL of each adsorbent module is then used to obtain the overall RUL of the current adsorption bed.

[0091] Furthermore, due to the multi-dimensional sensor array and the spatially resolvable reverse PINN set in this embodiment, the parameters of each adsorbent module can be identified individually, thereby obtaining an aging (i.e., decay curve) distribution map along the z-axis, providing a basis for targeted replacement of the adsorbent module.

[0092] (4.3) Construct a multi-objective optimization module with the goal of minimizing the total lifecycle capture cost (LCOC), used to optimize the real-time identification... Determine how to dynamically adjust the regeneration temperature T controlled by heater 4. reg The logic of the multi-objective optimization module is as follows: when accelerated deactivation is detected, it evaluates the positive and negative economic impacts of increasing or decreasing the regeneration temperature by a set temperature step, thereby calculating an optimal regeneration temperature that minimizes the expected LCOC value, and sends this as a setpoint to the proportional heating valve. As an example, the multi-objective optimization module evaluates and determines that increasing or decreasing the regeneration temperature T... reg The positive economic impact of increasing the temperature by 5°C is that it may delay deactivation and save on future material replacement costs; the negative economic impact is that it may increase steam or electricity consumption in the current cycle.

[0093] In the multi-objective optimization module, based on different operating parameters (T) reg ,v air Predicted LCOC, v air The apparent flow rate of the gas is expressed by the following formula:

[0094] (6)

[0095] In the formula, C cycle The variable cost per cycle consists of two parts: ① Energy cost C energy ① Based on the energy consumption of heater 4 and variable frequency blower 2; ② Adsorbent degradation cost C degradation : Allocate the total adsorbent replacement cost to the predicted overall RUL of the adsorption bed. C cap_amortized To amortize capital costs (capital costs amortized over cycles), specifically, the total system capital expenditure (CAPEX) is amortized over each independent cycle within the design life cycle (i.e., the initial RUL of the material), this cost is a fixed value for the system, relative to real-time T. reg v air Irrelevant; total system capital expenditure includes all initial investments such as system equipment procurement and engineering construction. The CO2 capture amount per cycle, based on real-time identification. The calculated maximum adsorption capacity q max It is obtained through real-time correction.

[0096] The calculation is based on classical adsorption theory, combined with the influence of dynamic operating parameters, and its expression is as follows:

[0097] (7)

[0098] In the formula, M bed q represents the total mass of the adsorbent in the adsorption bed, and is a fixed parameter of the system. work Effective working capacity refers to the amount of CO2 that a unit mass of adsorbent can reversibly exchange in one complete adsorption-desorption cycle. Numerically, it is equal to the difference between the equilibrium adsorption capacity of the adsorbent in the adsorption phase and the equilibrium adsorption capacity in the desorption phase. η capture This is the capture efficiency factor, used to quantify the reduction in actual capture amount relative to the theoretical equilibrium capacity due to mass transfer limitations under dynamic flow conditions.

[0099] This invention uses a material balance model based on characteristic time to calculate this capture efficiency factor, the expression of which is as follows:

[0100] (8)

[0101] In the formula, Let be the effective length of the adsorption bed, which is a geometric constant.

[0102] S5: Determine whether the overall RUL of the current adsorption bed is lower than the preset safety threshold. If not, continue normal operation and collect sensor data to perform DAC performance prediction (i.e., repeat S1, S3-S5); if yes, trigger the optimization control operation, complete the adjustment, and continue normal operation. In this embodiment, the preset safety threshold is 2000 cycles.

[0103] The triggered optimization control operations include:

[0104] (1) Execute the preset "life extension priority" operation mode, for example, automatically send an instruction to the variable frequency blower 2 to reduce its operating frequency by 10% in order to slow down the aging process.

[0105] (2) Adjust the operating parameters (including regeneration strategy adjustment and adsorption strategy adjustment) based on the deactivation dominant mechanism obtained from the diagnosis.

[0106] For example, if the dominant mechanism is high air humidity (relatively high k2), the corresponding adsorption strategy would be adjusted to: reduce wind speed v. airThat is, the air volume of the variable frequency blower 2 is adjusted and reduced to reduce the total water and vapor load; the regeneration strategy is adjusted to: increase the regeneration temperature T. reg That is, increase the heating temperature of heater 4 to achieve more thorough moisture desorption.

[0107] If the dominant mechanism is pollutant poisoning (relatively high K3), the corresponding adsorption strategy is adjusted to: reduce wind speed v. air This means adjusting and reducing the airflow of the variable frequency blower 2 to reduce the total intake of pollutants. The specific regeneration strategy adjustment depends on the type of pollutant. If the pollutant is one that can be decomposed or desorbed at high temperatures (such as certain volatile organic compounds), the regeneration temperature T should be increased. reg If the pollutant forms a permanent chemical bond and cannot be removed by high temperature (such as NOx causing loss of active sites), then the targeted replacement module instruction is triggered.

[0108] If the dominant mechanism is cyclic aging (with a relatively high k1), the corresponding adsorption strategy is adjusted as follows: no direct adjustment is made, or a smoother operating curve is adopted. The regeneration strategy is adjusted as follows: no adjustment is made.

[0109] (3) Based on the aging distribution map along the z-axis obtained in S4, determine whether the performance (equivalent to RUL) of any adsorbent module is below the performance failure threshold; when it is determined that the performance of the upstream adsorbent module is below the performance failure threshold, while the downstream adsorbent module still has a large margin, the operation and maintenance platform will generate maintenance instructions, such as "It is recommended to replace only adsorbent module No. 1", and send a high-level warning to the user to remind them to prepare spare parts and arrange a maintenance plan; this targeted replacement of modules can reduce material waste and maintenance costs.

[0110] Furthermore, when the overall RUL of the adsorption bed falls below the emergency replacement threshold, the electrically controlled switching valve 3 is automatically switched to direct the airflow into the backup adsorption bed, ensuring production continuity. In this embodiment, the emergency replacement threshold is set to 800 cycles.

[0111] The present invention will now be described in detail through specific embodiments.

[0112] Current operating status: The DAC equipment has been running stably for approximately 4000 adsorption-desorption cycles. The adsorption bed adopts a three-module series design (i.e., the first adsorbent module 6, the second adsorbent module 7, and the third adsorbent module 8, hereinafter referred to as M1, M2, and M3), and a complete sensing and control system has been deployed. The plant's intelligent computing core automatically performs a complete adsorbent health status assessment and operating strategy optimization every 24 hours according to a preset program.

[0113] Over the past week, the average relative humidity (H) has remained high due to weather conditions; simultaneously, temporary abnormal emissions from nearby industrial areas have led to a decrease in NOx concentration (C) in the air. pA small peak appeared. However, the operator did not pay much attention to it. The adsorbent abnormality was identified and monitored by the PINN-based DAC performance prediction and control system of this invention. The corresponding process is as follows.

[0114] Phase 1: Online parameter identification and inactivation parameter vector calculation.

[0115] Data Acquisition and Input: The computing unit 10 acquires CO2 penetration curve data recorded by fiber optic sensors deployed at the outlets (z1=0.33L, z2=0.67L, z3=L) of adsorbent modules M1, M2, and M3 over the past 24 cycles (approximately 24 hours, i.e., one evaluation cycle), as well as concurrent environmental data recorded by environmental sensor 5 deployed at the inlet, including: average relative humidity H=85%, NOx concentration C. p A peak value of 0.5 ppm appeared, lasting for 2 hours. The system recorded the current total number of cycles N=4320.

[0116] Reverse PINN calculation: The calculation unit 10 processes the spatiotemporal data collected above. Environmental data The training process of the inverse PINN is initiated by taking the number of iterations N as input. During training, the goal is to find a set of inactive parameter vectors. This allows the solution to the partial differential equations determined by these components to best fit the multi-point penetration curves actually measured over the past 24 hours. After approximately 15 minutes of calculation, the inverse PINN converges, outputting the current optimal set of deactivation parameter vectors; and, due to its spatial resolution characteristics, each adsorbent module corresponds to a deactivation parameter vector. In this embodiment, the specific deactivation parameter vectors for the three modules are as follows:

[0117] Module M1 (entry segment): k1=0.8e-5, k2=1.5e-6, k3=3.2e-7, k q =1.5e-5.

[0118] Module M2 (intermediate segment): k1=0.8e-5, k2=1.5e-6, k3=0.5e-7, k q =0.5e-5.

[0119] Module M3 (Exit Section): k1=0.8e-5, k2=1.5e-6, k3=0.1e-7, k q =0.4e-5.

[0120] Phase Two: Simultaneously conduct diagnostic analysis, predict remaining service life, and adjust regeneration strategies.

[0121] (1) Deactivation mechanism diagnosis: By comparing each parameter in the deactivation parameter vector obtained in stage one with the historical baseline, it was found that the k2 (humidity effect) values ​​of all modules increased significantly, and the k3 (pollutant effect) and k of module M1 increased significantly. q The capacity decay value is an order of magnitude higher than that of the downstream M2 and M3 modules.

[0122] The final diagnosis was: the mass transfer performance of the system was accelerated due to high humidity. At the same time, the inlet module M1 was significantly affected by pollutant (NOx) poisoning, which has caused irreversible loss of active sites.

[0123] (2) RUL Prediction: Prior to this assessment, the system predicted the overall RUL of the current adsorption bed based on historical data to be 3500 cycles remaining. Now, the system will predict the latest identified and deteriorated RUL. Substitute the vector into the decay function, recalculate the extrapolation, and output the new prediction results. Based on the prediction results, correct the overall RUL of the current adsorption bed to 1800 cycles. The RUL of each adsorbent module are: M1=500 cycles, M2=2500 cycles, and M3=2800 cycles.

[0124] (3) Regeneration strategy generation: The multi-objective optimization module with the goal of minimizing LCOC is started and the regeneration strategy is adjusted to: the regeneration temperature for the next week is slightly adjusted from 100°C to 105°C. Although this will increase energy consumption by about 3%, it is expected to partially restore the mass transfer performance caused by water vapor, thereby extending the overall RUL of the adsorption bed by about 200 cycles.

[0125] Phase Three: Autonomous Decision-Making and Closed-Loop Control (Automatic Response Process). The specific execution sequence is as follows:

[0126] (1) It is determined that the overall RUL (1800) of the adsorption bed is lower than the set safety threshold (2000), and the RUL (500) of module M1 is lower than the emergency replacement threshold (800).

[0127] (2) Implement short-term optimization control:

[0128] Regeneration strategy execution: The calculation unit 10 sends the adjusted temperature setpoint of the regeneration strategy to the heater 4 through the control execution unit 11.

[0129] Adsorption strategy adjustment: In order to mitigate further impact on the M1 module, the system automatically enters the "life extension priority" mode and sends a command to the variable frequency blower 2 to reduce the airflow rate during the adsorption stage by 5%.

[0130] (3) Send operator alarm:

[0131] The system sent an alert to the management engineer: the adsorbent is experiencing accelerated systemic aging, and the overall RUL has dropped to 1800 cycles. Module M1 has severely degraded, with an RUL of only 500. The system has automatically initiated optimized regeneration and de-loading life extension procedures.

[0132] Phase Four: Guided Targeted Maintenance.

[0133] Maintenance work order generation: Since the RUL of module M1 has triggered the emergency threshold, the system automatically creates a new maintenance work order in the factory's Computerized Maintenance Management System (CMMS). The work order clearly instructs: "It is recommended to perform maintenance on the adsorption bed within the next 500 cycles. Maintenance content: Replace the first adsorbent module 6 (M1)."

[0134] Manual execution: Upon receiving this work order, the factory's maintenance engineer will, within the predetermined time, only remove the aging M1 module and replace it with a new adsorbent module, without touching the still-functioning M2 and M3 modules.

[0135] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A DAC performance prediction and control method based on PINN, characterized in that, Includes the following steps: S1: For the modularly arranged adsorption bed in the direct air capture system, acquire real-time sensing data from multiple sensors arrayed within the adsorption bed at different times. S2: Construct and train an inverse physical information neural network that couples a physical model describing the adsorption process with an empirical function describing the deactivation of the adsorbent material. Its input is the sensor data, and its output is a predicted CO2 concentration. The network iterates to minimize the total loss function, which includes data loss and physical loss, based on the effective diffusion coefficient D. eff Maximum adsorption capacity q max The network prediction error is updated; the D eff and q max Calculated from the deactivation parameter vector; When training the inverse physical information neural network, perform the following operations in each evaluation cycle: (a) Extract the data collected in the most recent evaluation periods as the training set for sensor data; (b) Initialize the network weights and the deactivation parameter vector to be solved for the inverse physical information neural network; (c) Input the training set data into the inverse physical information neural network, use the optimizer, and minimize the total loss function through the backpropagation algorithm; during the optimization process, the weights and inactivation parameter vectors of the inverse physical information neural network will be updated simultaneously; iterate the optimization until the total loss function converges, complete the network weight update, and obtain the trained inverse physical information neural network; S3: Input the real-time sensing data collected by S1 into the inverse physical information neural network to solve for a set of deactivation parameter vectors that quantify multiple deactivation mechanisms. ; S4: According to the above The numerical values ​​of each parameter are used to diagnose the dominant mechanism of deactivation and predict the remaining service life of the adsorption bed. S5: Determine whether the remaining service life of the current adsorption bed is lower than the preset safety threshold. If not, continue normal operation and predict the performance of the direct air capture system. If so, trigger the optimization control operation, complete the adjustment, and continue normal operation and predict the performance of the direct air capture system.

2. The DAC performance prediction and control method based on PINN according to claim 1, characterized in that, The deactivation parameter vector includes: cyclic aging deactivation rate constant k1, humidity-induced deactivation rate constant k2, contaminant poisoning deactivation rate constant k3, and capacity decay rate constant k. q .

3. The DAC performance prediction and control method based on PINN according to claim 2, characterized in that, In S2, the inverse physical information neural network includes: a fully connected feedforward neural network structure module, a deactivation function calculation module, a loss calculation module, and an optimizer module; The fully connected feedforward neural network structure module includes an input layer, a hidden layer, and an output layer; the input layer receives the normalized variable (t, z), and the output layer outputs the predicted CO2 concentration C. pred (t,z); where z is the position coordinate along the axial direction of the adsorption bed; The inverse physical information neural network mathematizes and couples the physical model describing the adsorption process with the empirical function describing material deactivation. The physical model is based on the mass conservation relationship of the gas phase in the adsorption bed and the linear or nonlinear driving force relationship of the solid phase adsorption kinetics. It imposes simultaneous constraints on the gas phase CO2 concentration and the adsorption amount of the solid phase adsorbent: the gas phase part follows the conservation relationship between the transport of substances along the axial direction of the adsorption bed and the source terms generated by solid phase adsorption and desorption; the solid phase part follows the kinetic relationship that the actual adsorption amount is equal to the equilibrium adsorption amount corresponding to the current gas phase conditions. The two types of relationships work together to ensure that the output of the inverse physical information neural network simultaneously satisfies the coupling requirements of fixed bed mass conservation and linear driving force adsorption kinetics in both time and space. The deactivation function calculation module includes D eff and q max The calculation of the current value of the effective diffusion coefficient D; eff The current value is based on the number of adsorption-desorption cycles completed (N), the relative humidity of the inlet air (H(t)), and the pollutant concentration (C). p (t) in a single cycle duration t cycle The monotonic decay relationship, including the cumulative exposure within the range, is expressed with an initial value D. eff,0 Using k1 as the baseline, k2 as the weight parameter of N, k3 as the weight parameter of H(t), and k3 as the weight parameter of C... p The weighting parameters of (t) are determined; the maximum adsorption capacity q max The current value is determined based on a monotonically decaying relationship related to the cumulative loop or runtime, starting from the initial value q. max,0 Based on k q The parameters for N are determined; the {k1, k2, k3, k q } constitute the deactivation parameter vector It is updated synchronously with the network parameters during training.

4. The DAC performance prediction and control method based on PINN according to claim 3, characterized in that, In S2, the total loss function is composed of the data loss L data and physical loss L physics We get the result by weighted summation; The data loss L data Used to measure z at each sensor location i Location and sampling time t j Below, the network outputs the predicted concentration values. The overall mean square error relative to the corresponding measured value; The physical loss L physics Used to measure the rate of change of the gas and solid phase fields represented by the network output relative to time and space; physical loss L physics The calculation specifically involves: calculating the output C of the fully connected feedforward neural network through automatic differentiation. pred The partial derivatives of the input variables (t, z) are substituted into the system of partial differential equations to obtain the residuals of the system of partial differential equations. The mean square values ​​of these residuals at all locus points in the spatiotemporal domain are then calculated to obtain the physical loss L. physics .

5. The DAC performance prediction and control method based on PINN according to claim 1, characterized in that, In step S4, the step of predicting the remaining service life of the adsorbent includes: substituting the deactivation parameter vector into a preset decay function used to describe the effective diffusion coefficient and / or maximum adsorption capacity to obtain a decay curve of the effective diffusion coefficient and / or maximum adsorption capacity as a function of future cycles; and calculating the remaining service life based on the decay curve and its preset failure threshold.

6. The DAC performance prediction and control method based on PINN according to claim 1, characterized in that, S4 further includes: a multi-objective optimization module aimed at minimizing the total lifecycle capture cost, based on the... Generate the regeneration temperature T that combines positive and negative economic impacts. reg Adjustment strategy; The logic of the multi-objective optimization module is as follows: when deactivation acceleration is detected, the positive and negative economic impacts of increasing or decreasing the set temperature step size of the regeneration temperature are evaluated respectively, thereby calculating an optimal regeneration temperature that minimizes the expected value of the total life cycle capture cost, and sending it as a set value to the direct air capture system.

7. The DAC performance prediction and control method based on PINN according to claim 6, characterized in that, The estimated total lifecycle capture cost (LCOC) is obtained by dividing the total cost per cycle by the amount of CO2 captured per cycle. The total cost of a single cycle includes the variable cost of the single cycle and the allocated cost of capital; The CO2 capture amount in a single cycle is the actual capture amount, which is obtained by subtracting the theoretical capture amount from the capture efficiency factor. The theoretical capture amount is the product of the total mass of the adsorbent in the adsorption bed and the effective working capacity. The capture efficiency factor is obtained based on a material balance model for a characteristic time, which is based on the effective length of the adsorption bed. The degree of mass transfer limitation is determined by combining structural and operating parameters, including the overall mass transfer coefficient and the apparent flow rate of the gas. Based on the degree of mass transfer limitation, the reduction ratio of the actual capture amount to the equilibrium capacity is calculated, and this reduction ratio is used as the value of the capture efficiency factor.

8. The DAC performance prediction and control method based on PINN according to claim 1, characterized in that, In S5, the optimized control operation includes: (1) Execute the preset life extension priority operation mode and control the operating frequency of specific components according to the preset instructions to slow down the aging process of the direct air capture system; (2) Adjust the corresponding operating parameters according to the dominant inactivation mechanism obtained from the diagnosis; (3) Determine whether the performance of an adsorbent module is below the performance failure threshold. If the performance of the upstream adsorbent module is below the performance failure threshold and the downstream adsorbent module still has a large margin, the operation and maintenance platform generates a maintenance instruction and sends a high-level warning to the user, prompting them to prepare spare parts and arrange a maintenance plan.

9. The DAC performance prediction and control method based on PINN according to claim 1, characterized in that, In step S5, if the remaining service life of the adsorption bed as a whole is lower than the emergency replacement threshold, an adsorption bed switching is performed, and the airflow is introduced into the standby adsorption bed to ensure production continuity.

10. A PINN-based DAC performance prediction and control system, used to implement the PINN-based DAC performance prediction and control method according to any one of claims 1-9, characterized in that, include: Variable frequency blower, electrically controlled switching valve, heater, several adsorption beds, computing unit, and control execution unit; Airflow enters the system pipeline through the air inlet. A variable frequency blower, an electrically controlled switching valve, a heater, and an adsorption bed are arranged coaxially in the pipeline along the airflow direction. The airflow is finally output from the exhaust port. The adsorption bed is designed to be detachable and replaceable, with other adsorption beds serving as backups. The variable frequency blower is used to control the airflow rate, the electrically controlled switching valve is used to control the switching between the working adsorption bed and the backup adsorption bed, and the heater is used to precisely regulate the temperature of the regeneration heat fluid. Each adsorption bed includes an environmental sensor and multiple adsorbent modules arranged along the gas flow direction. Each adsorbent module can be independently disassembled and replaced. The environmental sensor is used to detect information at the inlet of the adsorption bed, including: air temperature, relative humidity, and pollutant concentration. Each adsorbent module is equipped with a fiber optic sensor at its outlet to detect information within the corresponding adsorbent module, including: specific gravity of adsorbent particles, apparent gas flow rate, and CO2 concentration. The environmental sensor and the fiber optic sensor of the adsorbent module are both connected to the computing unit, which is connected to the control execution unit. After processing the received sensor data, the computing unit sends control commands to the control execution unit. The control execution unit is connected to and controls the variable frequency blower, the electrically controlled switching valve, and the heater, respectively.

Images