Filter membrane degerming and filtering device and filter membrane degerming and filtering system

Abstract

The invention relates to the technical field of terminal sterilization and filtration in a radiopharmaceutical preparation process, in particular to a filter membrane sterilization and filtration device and system. Comprising the steps that a state sensing module collects liquid medicine radioactive activity data and filter membrane structure strain data and generates real-time data flow; the twinborn modeling module constructs a filter membrane digital twinborn body, and outputs an accumulated radiation damage prediction result through data assimilation; the prediction decision module generates a filtering parameter optimization sequence by using a deep reinforcement learning agent; the control execution module adjusts the flow velocity of the liquid medicine through fuzzy self-adaptive PID control and feed-forward compensation; and the verification feedback module performs an integrity test and transmits data back to the twin modeling module to realize model parameter correction. All the modules form a closed-loop control system through a high-speed data bus, so that the damage of radiation to a filter membrane material is effectively inhibited, and the integrity and the sterile guarantee level of a membrane structure in the filtering process are maintained.

Description

Technical Field

[0001] This invention relates to the field of terminal sterilization filtration technology in radiopharmaceutical preparation processes, and particularly to a membrane sterilization filtration device and system. Background Technology

[0002] Existing terminal sterilization filtration technology for positron-emitting radiopharmaceuticals refers to the process of removing microorganisms using microporous membranes during the final filling stage of positron-emitting radiopharmaceutical production. Positron-emitting radiopharmaceuticals typically include short-half-life radionuclides, whose rapid decay characteristics require the preparation process to be completed within a strictly controlled time window while maintaining drug sterility. Therefore, terminal sterilization filtration efficiently removes bacteria and particulate matter through physical retention principles to meet sterile drug specifications and is compatible with rapid, closed production line operations, thereby ensuring the microbiological safety of the drug before clinical use.

[0003] Existing terminal sterilization filtration technologies for positron-emitting radiopharmaceuticals suffer from the following technical challenges: The filter membranes used for sterilization are typically composed of polymer materials, and the radiation energy continuously released during the decay of positron-emitting nuclides such as fluorine-18 directly affects the molecular structure of the membrane material. This radiation energy can cause the polymer backbone to break or undergo cross-linking reactions, resulting in increased material brittleness and irreversible changes in pore morphology at the microscopic level. Since drug filtration must be completed immediately when the radioactivity is high, the filter membrane continuously absorbs radiation dose during contact, potentially causing latent damage to membrane integrity. For example, in the production of fluorine-18-labeled deoxyglucose injections, radiation damage causes submicron-level defects on or inside the membrane surface when the high-activity drug flows through the filter membrane. These defects may not be immediately detected in conventional bubble point tests but can become pathways for microbial penetration, ultimately affecting the sterility assurance level of the drug. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a filter membrane sterilization filtration device and system, which solves the technical problem that continuous radiation irradiation caused by the decay of radioactive nuclides leads to damage to the microstructure of the filter membrane polymer material, resulting in potential failure of membrane integrity during the sterilization filtration process.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows: In a first aspect, the filter membrane sterilization filtration device provided by the present invention includes a state sensing module, a twin modeling module, a prediction and decision module, a control execution module, and a verification feedback module; The state sensing module collects data on the radioactivity of the drug solution flowing through the sterilization filter membrane and the structural strain data of the sterilization filter membrane, and generates a real-time data stream. The twin modeling module receives the real-time data stream, constructs a digital twin of the sterilization filter membrane, dynamically updates the digital twin by assimilating the structural strain data, and outputs the cumulative radiation damage prediction result of the sterilization filter membrane. The prediction and decision-making module includes a decision agent, which is used to receive the cumulative radiation damage prediction result and the radioactivity data, and generate a filter parameter optimization sequence through the decision agent. The filter parameter optimization sequence includes a flow rate setpoint instruction. The control execution module receives the flow rate setpoint command and the cumulative radiation damage prediction result, and adjusts the operating parameters of the precision metering pump to control the drug filtration flow rate. The verification feedback module applies test conditions to the sterilization filter membrane after filtration is completed, obtains integrity test data, and sends the integrity test data back to the digital twin modeling module to correct the model parameters of the digital twin.

[0006] Furthermore, in the filter membrane sterilization filtration device of the present invention, the state sensing module includes a radioactive sensor, a strain sensor network, and a signal processing unit; The radioactive sensor captures radioactive particles and converts them into photon signals. The photon signals are converted into electrical pulse signals by a photoelectric conversion unit. The electrical pulse signals are filtered to remove background noise pulses and enter a counter. The counter counts and outputs the activity time-series electrical signal within a fixed time window. The strain sensor network is composed of optical fibers written into a grating array. The optical fibers are encapsulated in the sterilization filter membrane in a preset topology. The light emitted by the light source is reflected by the grating array, and the resulting optical signal is received and analyzed by a spectrum demodulator to obtain the wavelength shift. The signal processing unit synchronizes the activity time-series electrical signal with the wavelength drift, and converts the wavelength drift into structural strain data according to a preset calibration coefficient. The signal processing unit adds a timestamp to the activity time-series electrical signal and the structural strain data, and packages them to generate the real-time data stream.

[0007] Furthermore, in the filter membrane sterilization filtration device of the present invention, the twin modeling module performs the following data processing: The system receives the real-time data stream, extracts radioactivity data, and calculates the radiation energy deposition rate of each grid cell of the sterilizing filter membrane based on the geometric and intrinsic material parameters of the membrane using the finite element analysis method. The calculated energy deposition rate is input into the polymer radiation damage kinetic equation, and the time-varying field of the porosity and average pore size of the sterilization filter membrane is solved by numerical integration to output the predicted strain field. The predicted strain field is compared with the measured structural strain field extracted from the real-time data stream. An ensemble Kalman filter algorithm is used, with the measured structural strain field as the observation, to back-optimize and adjust the local damage rate parameter in the polymer radiation damage kinetic equation. Based on the optimized parameter, the updated cumulative radiation damage prediction result is output.

[0008] Furthermore, in the filter membrane sterilization filtration device of the present invention, the decision-making agent in the prediction decision module is a deep reinforcement learning agent; The state vector of the deep reinforcement learning agent consists of the feature values ​​of the cumulative radiation damage prediction result, the historical and predicted activity decay curves, the remaining drug volume, and the cumulative filtration time. The deep reinforcement learning agent completes pre-training in an offline simulation environment with the reward objective of maximizing the probability of maintaining filter membrane integrity. When running online, the deep reinforcement learning agent receives the state vector and outputs preliminary flow rate adjustment suggestions. The prediction decision module also includes a model prediction controller. The model prediction controller receives the preliminary flow rate adjustment suggestion action. Within the current control cycle, with the objective function of minimizing the cumulative damage to the filter membrane in the future finite time domain, it encodes and iteratively optimizes a set of future flow rate sequences using a genetic algorithm, and outputs the future flow rate sequence with the lowest total damage cost as the optimized filter parameter sequence.

[0009] Furthermore, in the filter membrane sterilization filtration device of the present invention, the control execution module includes a fuzzy adaptive PID controller and a feedforward compensator; The fuzzy adaptive PID controller receives the flow rate setpoint command and the actual flow rate value fed back by the electromagnetic flowmeter, calculates the deviation and the rate of change of deviation, dynamically tunes the PID parameters online through the fuzzy rule table, and outputs the basic control quantity to the driver of the precision metering pump. The feedforward compensator receives the cumulative radiation damage prediction result output by the twin modeling module, performs a differential operation on the cumulative radiation damage prediction result to obtain the damage change rate, and multiplies the damage change rate with a preset feedforward gain coefficient to generate a feedforward compensation amount. The basic control quantity and the feedforward compensation quantity are superimposed in the adder, and the superimposed signal output by the adder serves as the final control signal, driving the precision metering pump to control the flow rate of the filtered drug solution.

[0010] Furthermore, in the filter membrane sterilization filtration device of the present invention, the control execution module is also connected to a pressure transmitter, which monitors the pressure difference between the upstream and downstream of the sterilization filter membrane and outputs pressure difference data. The control execution module is equipped with a pressure protection sub-circuit, which receives the differential pressure data and compares the differential pressure data with a preset safety threshold. When the differential pressure data exceeds the preset safety threshold, the pressure protection sub-circuit generates a limiting signal, which is then introduced into the output of the fuzzy adaptive PID controller to activate the control quantity softening logic and limit the rate of change of the control signal output to the precision metering pump.

[0011] Furthermore, in the filter membrane sterilization filtration device of the present invention, when the filtration is completed, the verification feedback module switches the pneumatic valve to connect the clean compressed nitrogen source to the upstream cavity of the sterilization filter membrane, and raises the pressure to the preset test pressure and maintains it through the pressure regulating valve; The verification feedback module records the pressure decay time-series curve through a high sampling rate pressure sensor and synchronously records the gas diffusion flow time-series data through a downstream mass flow meter. The verification feedback module performs low-pass filtering on the pressure decay time series curve and the gas diffusion flow time series data to remove noise. Then, it uses an unsteady diffusion physics model to fit the filtered pressure decay time series curve and the gas diffusion flow time series data to calculate the standardized diffusion flow value, which serves as the core integrity index. The verification feedback module encapsulates the time-series data of the entire filtering task process and the core integrity indicators into a batch feedback dataset, and transmits it to the database of the twin modeling module.

[0012] Furthermore, in the sterilization filtration device of the present invention, after receiving the batch feedback dataset, the twin modeling module extracts the core integrity index in the batch feedback dataset as the actual state observation value of the sterilization filter membrane at the end of the current filtration cycle. The twin modeling module compares the actual state observation value with a series of predicted damage state values ​​corresponding to the time when the digital twin was generated within the current filtering cycle. The twin modeling module uses the maximum likelihood estimation algorithm to inversely adjust the key model coefficients in the polymer radiation damage kinetic equation that affect the long-term prediction accuracy. After the twin modeling module completes the coefficient adjustment, it optimizes the initial state of the digital twin before it enters the next working cycle.

[0013] Furthermore, in the filter membrane sterilization filtration device of the present invention, the state sensing module, the twin modeling module, the prediction and decision module, the control execution module, and the verification feedback module are connected via a high-speed data bus; The real-time data stream output by the state awareness module is input to the twin modeling module; The cumulative radiation damage prediction results output by the twin modeling module are input into the prediction decision module; The optimized sequence of filtering parameters output by the prediction decision module is input to the control execution module; The control signal output by the control execution module drives the precision metering pump; The integrity test data output by the verification feedback module is input into the twin modeling module; The twin modeling module receives the integrity test data and corrects the model parameters to form a closed-loop data stream.

[0014] Secondly, the filter membrane sterilization filtration system provided by the present invention is applied to the filter membrane sterilization filtration device as described above, including a state sensing module, a twin modeling module, a prediction and decision module, a control execution module, and a verification feedback module. The state sensing module is configured to collect data on the radioactivity of the drug solution flowing through the sterilization filter membrane and the structural strain data of the sterilization filter membrane, and generate a real-time data stream. The twin modeling module is configured to receive the real-time data stream, construct a digital twin of the sterilization filter membrane, dynamically update the digital twin by assimilating the structural strain data, and output the cumulative radiation damage prediction result of the sterilization filter membrane. The prediction and decision module is configured to include a decision agent, which receives the cumulative radiation damage prediction result and the radioactivity data, and generates a filter parameter optimization sequence through the decision agent. The filter parameter optimization sequence includes a flow rate setpoint instruction. The control execution module is configured to receive the flow rate setpoint command and the cumulative radiation damage prediction result, and adjust the operating parameters of the precision metering pump to control the drug filtration flow rate. The verification feedback module is configured to apply test conditions to the sterilization filter membrane after filtration is completed, obtain integrity test data, and send the integrity test data back to the twin modeling module to correct the model parameters of the digital twin.

[0015] The beneficial effects of this invention are: This invention utilizes a state-sensing module to collect real-time data on the radioactivity of the drug solution and the strain data of the filter membrane structure. A twin modeling module constructs and dynamically updates a digital twin based on this data, accurately predicting the cumulative radiation damage to the filter membrane. A prediction and decision-making module uses a deep reinforcement learning agent and a model predictive controller to generate optimized filtration parameters. A control execution module achieves precise flow rate adjustment through fuzzy adaptive PID control and feedforward compensation, combined with a pressure protection mechanism to avoid sudden pressure shocks. A verification feedback module uses integrity test data after filtration to reverse-correct the twin model parameters, forming a closed-loop control system from data sensing, model prediction, intelligent decision-making to execution feedback. This architecture can proactively suppress the cumulative damage of radiation to the filter membrane material, effectively maintaining the structural integrity and sterility of the filter membrane during the filtration of radioactive drugs, while adapting to dynamic changes in different process conditions, thus improving the reliability and stability of the filtration process. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a system architecture diagram of the filter membrane sterilization filtration device of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0019] To better understand the purpose of this invention, the invention will now be described in further detail.

[0020] In a first aspect, the filter membrane sterilization filtration device provided by the present invention includes a state sensing module, a twin modeling module, a prediction and decision module, a control execution module, and a verification feedback module; The state sensing module collects data on the radioactivity of the drug solution flowing through the sterilization filter membrane and the structural strain data of the sterilization filter membrane, and generates a real-time data stream. The twin modeling module receives the real-time data stream, constructs a digital twin of the sterilization filter membrane, dynamically updates the digital twin by assimilating the structural strain data, and outputs the cumulative radiation damage prediction result of the sterilization filter membrane. The prediction and decision-making module includes a decision agent, which is used to receive the cumulative radiation damage prediction result and the radioactivity data, and generate a filter parameter optimization sequence through the decision agent. The filter parameter optimization sequence includes a flow rate setpoint instruction. The control execution module receives the flow rate setpoint command and the cumulative radiation damage prediction result, and adjusts the operating parameters of the precision metering pump to control the drug filtration flow rate. The verification feedback module applies test conditions to the sterilization filter membrane after filtration is completed, obtains integrity test data, and sends the integrity test data back to the digital twin modeling module to correct the model parameters of the digital twin.

[0021] The sterilization filtration device provided by this invention dynamically maintains the integrity of the filter membrane during the filtration of radioactive pharmaceutical solutions through the coordinated operation of multiple modules. During operation, the state sensing module is activated first, responsible for collecting two key physical quantities: first, the radioactivity data of the pharmaceutical solution flowing through the sterilization filter membrane; and second, the structural strain data of the sterilization filter membrane itself under the combined effects of irradiation and fluid pressure. For example, the radioactivity data is collected using a non-contact scintillator sensor, while the structural strain data is obtained through a fiber optic grating sensor network embedded in the filter membrane support structure. The collected raw signals are then conditioned and converted from analog to digital, and integrated into a real-time data stream with time-series markers, providing data input for subsequent modules.

[0022] Real-time data streams are transmitted to the digital twin modeling module, which constructs a digital twin based on the filter membrane's material properties and geometric parameters. The digital twin simulates the deposition process of radiative energy in the filter membrane material using the finite element method and couples it with polymer damage kinetics equations to predict the evolution trend of the microporous structure. The module assimilates the real-time measured structural strain data and uses a data fusion algorithm to dynamically correct the model parameters, thereby outputting a high-confidence prediction of cumulative radiation damage. This step maps the state of the physical world to the virtual model, providing a basis for predictive decision-making.

[0023] The prediction and decision-making module receives cumulative radiation damage predictions and real-time radioactivity data from the twin modeling module. An internal decision-making agent, such as one trained using deep reinforcement learning, performs rolling optimization of filtration parameters with the goal of maintaining membrane integrity. The agent integrates current damage predictions, activity decay trends, and remaining drug volume to output a sequence of optimized filtration parameters, including flow rate setpoint instructions. This sequence aims to balance filtration efficiency with the risk of membrane damage.

[0024] The control execution module receives flow rate setpoint commands from the prediction and decision-making module, while simultaneously subscribing to real-time damage prediction data from the twin modeling module. The fuzzy adaptive PID controller within the module dynamically adjusts its output based on the deviation between the setpoint and the actual flow rate feedback from the electromagnetic flowmeter, while the feedforward compensator generates a compensation amount based on the rate of damage change. The sum of these two components drives a precision metering pump to regulate the drug flow rate. This feedforward-feedback composite control structure effectively suppresses nonlinear disturbances caused by radioactivity decay.

[0025] The verification feedback module is automatically activated after the filtration cycle ends, introducing clean compressed gas upstream of the filter membrane for integrity testing. Pressure sensors and mass flow meters record the pressure decay curve and diffusion flow rate, respectively. The data is filtered and fitted to a physical model to obtain a standardized diffusion flow value, which serves as the core integrity indicator. This indicator, along with the entire process data, is packaged into a batch feedback dataset and sent back to the twin modeling module for model parameter calibration, completing a closed-loop feedback from physical execution to model optimization.

[0026] The modules interact via a high-speed data bus, forming a collaborative system encompassing state awareness, virtual modeling, intelligent decision-making, precise control, and verification feedback. This architecture enables the device to proactively predict and suppress cumulative radiation damage to the filter membrane, thereby maintaining the integrity of the membrane structure under the process conditions of high-speed filtration of radioactive pharmaceutical solutions.

[0027] The radioactive sensor in the state-aware module employs a combination of a thallium-doped sodium iodide scintillator and a photomultiplier tube. When a radioactive particle passes through the scintillator, fluorescence is generated, and the photomultiplier tube converts the fluorescence signal into an electrical pulse signal. This electrical pulse signal is initially amplified by a preamplifier and then sent to a discriminator. The discriminator filters out low-amplitude noise pulses generated by ambient radiation by setting a voltage threshold. The valid pulses after discrimination are accumulated by a counter, forming a time-series distributed activity signal. The strain sensor network uses a grating array etched onto a single-mode fiber using a femtosecond laser. The reflection wavelength of each grating corresponds to the strain state at a specific location on the filter membrane. Light emitted from a broadband source is incident on the fiber grating array via a circulator. The reflection spectrum is peak-detected by a demodulator. By comparing the drift between the initial calibration wavelength and the real-time detection wavelength, and combining this with a temperature compensation algorithm, the structural strain value is calculated. The signal processing unit uses an FPGA chip to achieve multi-channel data synchronization, aligning the activity signal with the strain data in time. Based on the sensor calibration curve, the wavelength drift is converted into micro-strain units, and finally, a millisecond-level timestamp is added before encapsulation into a standardized data packet.

[0028] Upon receiving the real-time data stream, the twin modeling module first extracts the radioactivity data from the data packet. Combining this with the decay gamma-ray energy spectrum characteristics of fluorine-18, it uses the Monte Carlo method to simulate the particle transport process within the filter membrane material. Based on the 3D CAD model of the filter membrane, a finite element mesh is generated. For each mesh element, the mass decay coefficient is calculated based on the material density and elemental composition. The radiation energy deposition distribution field is then solved using the real-time activity data. The energy deposition rate is input as a source term into the polymer damage kinetics equation, which includes the rate constants for chain breakage and crosslinking reactions. Numerical integration is performed using the fourth-order Runge-Kutta method to calculate the spatiotemporal evolution data of porosity and Young's modulus. The assimilation process between the predicted and measured strain fields employs an ensemble Kalman filter algorithm. The predicted strain values ​​of the finite element mesh nodes form a state vector, and the strain values ​​at the sensor measurement points serve as the observation vector. By minimizing the covariance matrix between the prediction and observation, the reaction rate parameters in the damage kinetics equation are optimized inversely.

[0029] The deep reinforcement learning agent employs a dual-deep Q-network architecture. Its state space includes features such as porosity variance and maximum strain gradient extracted from cumulative radiation damage prediction results, while the activity decay curve uses an exponentially fitted decay constant as input features. The agent is pre-trained in a PyBullet simulation environment, which constructs a multiphysics coupled model of the filter membrane. The reward function is designed as a negative weighted sum of damage increment and filtration progress. During online deployment, the agent collects its state vector every 5 seconds and selects actions for output using an ε-greedy strategy. The action space is discretized into 10 flow rate adjustment levels. The model predictive controller uses a nonlinear predictive control framework, solving an optimization problem with damage constraints in each control cycle. The genetic algorithm population size is set to 50, and the optimal flow rate sequence is generated through tournament selection and simulated binary crossover.

[0030] The fuzzy adaptive PID controller uses the deviation between the flow rate setpoint and the actual flow rate, along with its derivative, as input variables. It fuzzifies these variables into linguistic variables such as "negatively large," "zero," and "positively large" through a membership function. The fuzzy rule base includes 49 rules, such as "if the deviation is positively large and the rate of change is negatively small, then output positively large." Defuzzification employs the centroid method to calculate precise PID parameter tuning values. The feedforward compensator performs first-order difference calculations on the cumulative radiation damage prediction results. The damage change rate is mapped to a compensation quantity through a gain scheduling module; this mapping relationship was established through prior calibration experiments. The adder is implemented using analog arithmetic circuitry. The basic control quantity and the feedforward compensation quantity are combined into a PWM signal in the driver to control the metering pump motor.

[0031] The pressure transmitter uses the diffused silicon principle and is installed in the inlet and outlet pipes of the filter membrane, with a measurement range covering 0-500 kPa. The pressure protection sub-loop is implemented in the PLC. When the detected differential pressure data exceeds the safety threshold, a soft-start algorithm is triggered within 1 millisecond, limiting the slope of the control signal change through an exponential decay function. The softening logic is specifically implemented by processing the PID output through a first-order inertial element, with the time constant dynamically adjusted according to the rated pressure bearing capacity of the filter membrane to avoid water hammer effects caused by sudden changes in flow velocity.

[0032] During integrity testing, a three-position five-way solenoid valve was used as the pneumatic valve. After switching, nitrogen gas was stabilized at the preset test pressure via a pressure reducing valve. The pressure sensor adopted the ceramic capacitive principle, with a sampling frequency of 100Hz. The recorded pressure decay curve was smoothed using Savitzky-Golay filtering. The diffusion flow was calculated based on the Hagen-Poiseuille equation, fitting the slope of the pressure-time curve using the least squares method, and then converting it to the volumetric flow rate under standard conditions according to the ideal gas law. The calculation of the core integrity index included temperature and pressure compensation for the diffusion flow value, and the final result was compared with the threshold specified in the ISO 13408 standard.

[0033] During model calibration, the digital twin compiles a time series of predicted damage state values ​​recorded every 30 seconds within the current filtering cycle and performs correlation analysis with the core integrity indicators provided by the validation feedback module. Maximum likelihood estimation employs the EM algorithm, iteratively optimizing the activation energy parameter and frequency factor in the damage kinetic equation in two steps. After coefficient adjustment, cross-validation is used to evaluate the model improvement effect, and the leave-one-out method is used to calculate the root mean square value of the prediction error, ensuring improved prediction accuracy in subsequent cycles.

[0034] The high-speed data bus employs a time-sensitive networking protocol, and each module achieves microsecond-level clock synchronization via the IEEE 802.1AS protocol. The state-aware module pushes data to the twin modeling module using a publish-subscribe model, with data packets including a sequence number verification mechanism. The specific path of the closed-loop data flow is as follows: integrity test data from the verification feedback module is uploaded to the database via the OPC UA protocol; the data acquisition service of the twin modeling module periodically pulls the latest batch of data; the model calibration task is encapsulated as a Docker container and executed on the edge server; and the calibrated model parameters are updated to the real-time prediction system via remote procedure calls.

[0035] Secondly, the filter membrane sterilization filtration system provided by the present invention is applied to the filter membrane sterilization filtration device as described above, including a state sensing module, a twin modeling module, a prediction and decision module, a control execution module, and a verification feedback module. The state sensing module is configured to collect data on the radioactivity of the drug solution flowing through the sterilization filter membrane and the structural strain data of the sterilization filter membrane, and generate a real-time data stream. The twin modeling module is configured to receive the real-time data stream, construct a digital twin of the sterilization filter membrane, dynamically update the digital twin by assimilating the structural strain data, and output the cumulative radiation damage prediction result of the sterilization filter membrane. The prediction and decision module is configured to include a decision agent, which receives the cumulative radiation damage prediction result and the radioactivity data, and generates a filter parameter optimization sequence through the decision agent. The filter parameter optimization sequence includes a flow rate setpoint instruction. The control execution module is configured to receive the flow rate setpoint command and the cumulative radiation damage prediction result, and adjust the operating parameters of the precision metering pump to control the drug filtration flow rate. The verification feedback module is configured to apply test conditions to the sterilization filter membrane after filtration is completed, obtain integrity test data, and send the integrity test data back to the twin modeling module to correct the model parameters of the digital twin.

[0036] In the context of radiopharmaceutical production, this device addresses the radiation damage caused by high-activity drug solutions flowing through sterile filter membranes during the preparation of fluorine-18 labeled drugs. It employs multi-module collaborative control to dynamically maintain the integrity of the filter membrane. The collaborative mechanism of each module is explained below with reference to specific implementation methods.

[0037] The state-aware module's radioactive sensor employs a combination of a sodium iodide scintillator and a photomultiplier tube. When fluorine-18 in the solution decays and releases gamma rays, these rays strike the scintillator, generating fluorescent pulses. The photomultiplier tube converts the optical signal into electrical pulses, which are then filtered to remove environmental noise. A counter counts the number of pulses per unit time to generate an activity time-series signal. The strain sensor network monitors filter membrane deformation using a Bragg grating array etched onto an optical fiber by a femtosecond laser. The reflection wavelength shift of each grating is linearly related to the local strain. The signal processing unit uses a field-programmable gate array (FPGA) to synchronously process multiple signals, packaging activity and strain data into timestamped data packets, which are then transmitted via Ethernet to the twin modeling module.

[0038] After receiving the real-time data stream, the twin modeling module first analyzes the radioactivity data and, combined with the gamma-ray energy spectrum characteristics of fluorine-18, uses the Monte Carlo method to simulate the energy deposition process of particles in the filter membrane material. Based on the finite element mesh generated from the three-dimensional model of the filter membrane, the absorbed dose rate of each mesh element is calculated. The dose rate data is input into the polymer damage kinetic equation, which includes the kinetic parameters of chain breakage and crosslinking reactions. The evolution of filter membrane porosity is solved through numerical integration. The module assimilates the measured strain data using an ensemble Kalman filter algorithm, using the strain values ​​at the sensor measurement points as observations, and back-optimizes the reaction rate constant in the damage equation, outputting updated cumulative radiation damage prediction results.

[0039] The deep reinforcement learning agent in the prediction and decision-making module adopts a dual-deep Q-network architecture. The network input layer receives state variables such as damage prediction feature values, activity decay curve fitting parameters, and remaining drug volume. The agent undergoes millions of training iterations in a simulation environment to learn a policy function that maximizes filter membrane integrity. During online operation, the agent collects system status every five seconds and outputs flow rate adjustment suggestions. The model predictive controller uses the agent's suggestions as the initial solution and employs a genetic algorithm to continuously optimize the flow rate sequence for the next ten minutes. The objective function comprehensively considers damage accumulation and filtration efficiency, ultimately outputting the optimal flow rate command sequence.

[0040] The fuzzy adaptive PID controller in the control execution module fuzzifies the deviation and rate of change between the set flow rate and the actual flow rate into seven linguistic variables, and calculates the PID parameter tuning values ​​using a weighted average method. The feedforward compensator performs differential calculations on the damage prediction results and generates a compensation quantity based on a preset gain coefficient table. The basic control quantity and the compensation quantity are superimposed by an analog adder and converted into a pulse width modulation signal to drive the stepper motor of the metering pump. The pressure protection sub-loop monitors the filter membrane pressure difference in real time. When the pressure difference exceeds a set threshold, it automatically switches to a first-order inertial element to limit the rate of change of the control signal and prevent pressure shocks from damaging the filter membrane structure.

[0041] The verification feedback module initiates the integrity test procedure after filtration. A three-position five-way solenoid valve switches the gas path, allowing clean nitrogen to enter the upstream chamber of the filter membrane. A pressure sensor records the pressure decay curve at a sampling frequency of 100 Hz, while a mass flow meter simultaneously monitors the downstream gas flow rate. The collected data is digitally filtered, and a non-steady-state diffusion model is used to calculate the standardized diffusion flow value. The time-series data and integrity indicators of the filtration process are packaged into a dataset and uploaded to the database system via the OPC UA protocol.

[0042] The data acquisition service of the digital twin modeling module periodically pulls the latest batch of data and compares and analyzes the integrity test results with the predicted values. The expectation-maximization algorithm is used to optimize the key parameters of the damage kinetic equation, and the updated model parameters are deployed to the real-time prediction system via remote procedure call. This closed-loop learning mechanism enables the digital twin to gradually adapt to the process differences between different batches, continuously improving prediction accuracy.

[0043] The modules achieve microsecond-level synchronization via a time-sensitive network, and data is transmitted using a publish-subscribe model. Real-time data from the state-aware module drives the twin model updates, predictions guide intelligent decision-making, control commands adjust the execution mechanism, and verification data further optimizes the model, forming a complete closed loop of perception, decision-making, execution, and optimization. This architecture enables the device to actively suppress radiation damage during the filtration of radioactive pharmaceutical solutions, providing pharmaceutical companies with a reliable sterilization filtration solution.

[0044] The finite element method (FEM) is a numerical computation technique that solves partial differential equations by discretizing a continuum into a finite number of small elements. In this invention, the FEM is used to calculate the radiation energy deposition rate of each grid element in a sterilizing filter membrane. The specific data processing path is as follows: a finite element mesh model is constructed based on the geometric and intrinsic material parameters of the sterilizing filter membrane; radioactivity data is used as input, and the energy deposition rate of each grid element is calculated by solving the radiation transfer equation; the energy deposition rate distribution field is output, providing source terms for the subsequent polymer damage kinetics equation.

[0045] Ensemble Kalman filtering is a data assimilation method that optimizes state estimation by combining model predictions and observational data. In this invention, ensemble Kalman filtering is used to assimilate measured structural strain data with the predicted strain field. The specific data processing path is as follows: the predicted strain field output by the digital twin is used as a priori estimate, and the measured structural strain field collected by the state-aware module is used as the observation value; by minimizing the covariance matrix between prediction and observation, the local damage rate parameter in the polymer radiation damage kinetic equation is optimized inversely; the updated cumulative radiation damage prediction result is output to improve model accuracy.

[0046] The Monte Carlo method is a statistical simulation technique based on random sampling, used to solve probabilistic problems. In this invention, the Monte Carlo method is used to simulate the transport process of radioactive particles in filter membrane materials. The specific data processing path is as follows: random particle trajectories are generated based on the decay gamma-ray energy spectrum characteristics of fluorine-18 nuclide; the energy deposition distribution is calculated by simulating the interaction between particles and the filter membrane material; and the energy deposition rate data is output as input for finite element analysis for radiation damage prediction.

[0047] Numerical integration methods are mathematical tools for approximating integral values ​​and used to solve differential equations. In this invention, numerical integration methods are used to solve the polymer radiation damage kinetic equation. The specific data processing path is as follows: the energy deposition rate obtained from finite element analysis is used as input and coupled with the damage kinetic equation; numerical integration methods such as the fourth-order Runge-Kutta method are used to integrate the equation over time; the time-varying fields of porosity and average pore size of the sterilization filter membrane are output to describe the microstructure evolution.

[0048] Genetic algorithms are optimization algorithms that simulate the natural evolutionary process, searching for optimal solutions through selection, crossover, and mutation operations. In this invention, a genetic algorithm is used to optimize the filter parameter sequence. The specific data processing path is as follows: the initial flow rate adjustment suggestions output by the deep reinforcement learning agent are used as the initial population; the flow rate sequence is represented as chromosomes through encoding, and iterative optimization is performed using tournament selection and simulated binary crossover; with minimizing the cumulative damage to the filter membrane as the objective function, the future flow rate sequence with the lowest total damage cost is output as the optimized filter parameter sequence.

[0049] Maximum likelihood estimation (MLE) is a parameter estimation method that determines model parameters by maximizing the likelihood function of observed data. In this invention, MLE is used to calibrate the model parameters of a digital twin. The specific data processing path is as follows: extract the core integrity index provided by the verification feedback module as the actual state observation value; compare it with the predicted damage state value generated by the digital twin to construct the likelihood function; and adjust the key model coefficients in the polymer radiation damage kinetic equation through an optimization algorithm to make the predicted value best match the observed value.

[0050] The Expectation-Maximization (EM) algorithm is an iterative algorithm used for maximum likelihood estimation in the presence of latent variables. In this invention, the Expectation-Maximization algorithm is used as an implementation method for maximum likelihood estimation in model calibration. The specific data processing path is as follows: during model calibration, it is iteratively performed in two steps: the E-step calculates the expected likelihood, and the M-step maximizes the likelihood function; by repeatedly adjusting the activation energy parameter and frequency factor in the damage kinetic equation, the model prediction error is minimized, thereby improving prediction accuracy.

[0051] The polymer radiation damage kinetic equation is a mathematical model describing the evolution of the microstructure of polymer materials under radiation. The equation is constructed based on the principles of radiation chemical kinetics, including chain breakage reaction rate equations and crosslinking reaction rate equations. Temperature effects are correlated through the Arrhenius equation, and radiation dose rate is introduced as a driving variable. In this invention, the polymer radiation damage kinetic equation receives the radiation energy deposition rate calculated by the finite element method as input. The spatiotemporal evolution of material porosity and average pore size is solved through numerical integration, and the output predicted strain field is used for state updates of the digital twin.

[0052] A deep reinforcement learning agent is an intelligent system that uses deep neural networks to implement reinforcement learning decisions. The agent is constructed based on a dual-deep Q-network architecture, including a feature extraction network and a policy output network. Training stability is improved through an experience replay mechanism and a target network for policy updates. In this invention, the deep reinforcement learning agent receives state vectors such as feature values ​​of the cumulative radiation damage prediction results, activity decay curve parameters, and remaining drug volume. It outputs suggested actions for flow rate adjustment and achieves adaptive optimization of filtering parameters through offline pre-training and online fine-tuning.

[0053] The Model Predictive Controller (MMC) is an advanced control method based on model prediction and rolling optimization. The controller is constructed with three core components: a predictive model, an objective function, and an optimization algorithm. The predictive model uses state-space equations to describe the system dynamics, and the objective function comprehensively considers tracking performance and constraints. In this invention, the MMC uses the proposed actions of a deep reinforcement learning agent as the initial solution and performs rolling optimization on the flow velocity sequence within a finite time domain using a genetic algorithm, outputting the optimized sequence of filtering parameters with the lowest total damage cost.

[0054] The fuzzy adaptive PID controller is a self-tuning controller that combines fuzzy inference and PID adjustment. The controller architecture includes a fuzzification interface, a fuzzy inference engine, a knowledge base, and a defuzzification interface. It converts precise quantities into fuzzy quantities through membership functions and adjusts PID parameters online according to fuzzy rules. In this invention, the fuzzy adaptive PID controller receives the flow rate setpoint command and the actual flow rate value fed back from the electromagnetic flowmeter. It dynamically tunes the proportional, integral, and derivative coefficients through a fuzzy rule table and outputs the basic control quantity to drive the precision metering pump.

[0055] Low-pass filtering is a signal processing technique that allows low-frequency signals to pass through while suppressing high-frequency noise. Filters are constructed based on difference equations or transfer function models, and Butterworth or Chebyshev filters are commonly used to achieve flat passband characteristics. In this invention, low-pass filtering is applied to the processing of pressure decay time-series curves and gas diffusion flow time-series data in the verification feedback module. A digital filter implementation is used to effectively remove high-frequency measurement noise, providing clean data for subsequent physical model fitting.

[0056] The unsteady-state diffusion physics model is a physical equation describing the unsteady mass transfer process of gas through a porous medium. The model is constructed based on Fick's second law, combined with Darcy's law to describe fluid motion, and uses partial differential equations to characterize the dynamic relationship between pressure gradient and diffusion flow rate. In this invention, the unsteady-state diffusion physics model receives filtered pressure decay curves and diffusion flow data, and calculates standardized diffusion flow values ​​through parameter fitting, serving as the core indicator for assessing filter membrane integrity.

Claims

1. A membrane sterilization filtration device, characterized in that, It includes a state awareness module, a twin modeling module, a prediction and decision-making module, a control execution module, and a verification and feedback module; The state sensing module collects data on the radioactivity of the drug solution flowing through the sterilization filter membrane and the structural strain data of the sterilization filter membrane, and generates a real-time data stream. The twin modeling module receives the real-time data stream, constructs a digital twin of the sterilization filter membrane, dynamically updates the digital twin by assimilating the structural strain data, and outputs the cumulative radiation damage prediction result of the sterilization filter membrane. The prediction and decision-making module includes a decision agent, which is used to receive the cumulative radiation damage prediction result and the radioactivity data, and generate a filter parameter optimization sequence through the decision agent. The filter parameter optimization sequence includes a flow rate setpoint instruction. The control execution module receives the flow rate setpoint command and the cumulative radiation damage prediction result, and adjusts the operating parameters of the precision metering pump to control the drug filtration flow rate. The verification feedback module applies test conditions to the sterilization filter membrane after filtration is completed, obtains integrity test data, and sends the integrity test data back to the digital twin modeling module to correct the model parameters of the digital twin.

2. The membrane sterilization filtration device according to claim 1, characterized in that, The state perception module includes a radioactive sensor, a strain sensor network, and a signal processing unit. The radioactive sensor captures radioactive particles and converts them into photon signals. The photon signals are converted into electrical pulse signals by a photoelectric conversion unit. The electrical pulse signals are filtered to remove background noise pulses and enter a counter. The counter counts and outputs the activity time-series electrical signal within a fixed time window. The strain sensor network is composed of optical fibers written into a grating array. The optical fibers are encapsulated in the sterilization filter membrane in a preset topology. The light emitted by the light source is reflected by the grating array, and the resulting optical signal is received and analyzed by a spectrum demodulator to obtain the wavelength shift. The signal processing unit synchronizes the activity time-series electrical signal with the wavelength drift, and converts the wavelength drift into structural strain data according to a preset calibration coefficient. The signal processing unit adds a timestamp to the activity time-series electrical signal and the structural strain data, and packages them to generate the real-time data stream.

3. The membrane sterilization filtration device according to claim 1, characterized in that, The twin modeling module performs the following data processing: The system receives the real-time data stream, extracts radioactivity data, and calculates the radiation energy deposition rate of each grid cell of the sterilizing filter membrane based on the geometric and intrinsic material parameters of the membrane using the finite element analysis method. The calculated energy deposition rate is input into the polymer radiation damage kinetic equation, and the time-varying field of the porosity and average pore size of the sterilization filter membrane is solved by numerical integration to output the predicted strain field. The predicted strain field is compared with the measured structural strain field extracted from the real-time data stream. An ensemble Kalman filter algorithm is used, with the measured structural strain field as the observation, to back-optimize and adjust the local damage rate parameter in the polymer radiation damage kinetic equation. Based on the optimized parameter, the updated cumulative radiation damage prediction result is output.

4. The membrane sterilization filtration device according to claim 3, characterized in that, The decision-making agent in the prediction and decision-making module is a deep reinforcement learning agent. The state vector of the deep reinforcement learning agent consists of the feature values ​​of the cumulative radiation damage prediction result, the historical and predicted activity decay curves, the remaining drug volume, and the cumulative filtration time. The deep reinforcement learning agent completes pre-training in an offline simulation environment with the reward objective of maximizing the probability of maintaining filter membrane integrity. When running online, the deep reinforcement learning agent receives the state vector and outputs preliminary flow rate adjustment suggestions. The prediction decision module also includes a model prediction controller. The model prediction controller receives the preliminary flow rate adjustment suggestion action. Within the current control cycle, with the objective function of minimizing the cumulative damage to the filter membrane in the future finite time domain, it encodes and iteratively optimizes a set of future flow rate sequences using a genetic algorithm, and outputs the future flow rate sequence with the lowest total damage cost as the optimized filter parameter sequence.

5. The membrane sterilization filtration device according to claim 1, characterized in that, The control execution module includes a fuzzy adaptive PID controller and a feedforward compensator; The fuzzy adaptive PID controller receives the flow rate setpoint command and the actual flow rate value fed back by the electromagnetic flowmeter, calculates the deviation and the rate of change of deviation, dynamically tunes the PID parameters online through the fuzzy rule table, and outputs the basic control quantity to the driver of the precision metering pump. The feedforward compensator receives the cumulative radiation damage prediction result output by the twin modeling module, performs a differential operation on the cumulative radiation damage prediction result to obtain the damage change rate, and multiplies the damage change rate with a preset feedforward gain coefficient to generate a feedforward compensation amount. The basic control quantity and the feedforward compensation quantity are superimposed in the adder, and the superimposed signal output by the adder serves as the final control signal, driving the precision metering pump to control the flow rate of the filtered drug solution.

6. The membrane sterilization filtration device according to claim 5, characterized in that, The control execution module is also connected to a pressure transmitter, which monitors the pressure difference between the upstream and downstream of the sterilization filter membrane and outputs pressure difference data. The control execution module is equipped with a pressure protection sub-circuit, which receives the differential pressure data and compares the differential pressure data with a preset safety threshold. When the differential pressure data exceeds the preset safety threshold, the pressure protection sub-circuit generates a limiting signal, which is then introduced into the output of the fuzzy adaptive PID controller to activate the control quantity softening logic and limit the rate of change of the control signal output to the precision metering pump.

7. The membrane sterilization filtration device according to claim 1, characterized in that, When the filtration is finished, the verification feedback module switches the pneumatic valve to connect the clean compressed nitrogen source to the upstream cavity of the sterilization filter membrane, and raises the pressure to the preset test pressure and maintains it through the pressure regulating valve. The verification feedback module records the pressure decay time-series curve through a high sampling rate pressure sensor and synchronously records the gas diffusion flow time-series data through a downstream mass flow meter. The verification feedback module performs low-pass filtering on the pressure decay time series curve and the gas diffusion flow time series data to remove noise. Then, it uses an unsteady diffusion physics model to fit the filtered pressure decay time series curve and the gas diffusion flow time series data to calculate the standardized diffusion flow value, which serves as the core integrity index. The verification feedback module encapsulates the time-series data of the entire filtering task process and the core integrity indicators into a batch feedback dataset, and transmits it to the database of the twin modeling module.

8. The membrane sterilization filtration device according to claim 7, characterized in that, After receiving the batch feedback dataset, the twin modeling module extracts the core integrity index from the batch feedback dataset as the actual state observation value of the sterilization filter membrane at the end of this filtration cycle. The twin modeling module compares the actual state observation value with a series of predicted damage state values ​​corresponding to the time when the digital twin was generated within the current filtering cycle. The twin modeling module uses the maximum likelihood estimation algorithm to inversely adjust the key model coefficients in the polymer radiation damage kinetic equation that affect the long-term prediction accuracy. After the twin modeling module completes the coefficient adjustment, it optimizes the initial state of the digital twin before it enters the next working cycle.

9. The membrane sterilization filtration device according to claim 8, characterized in that, The state perception module, the twin modeling module, the prediction and decision-making module, the control execution module, and the verification feedback module are connected via a high-speed data bus. The real-time data stream output by the state awareness module is input to the twin modeling module; The cumulative radiation damage prediction results output by the twin modeling module are input into the prediction decision module; The optimized sequence of filtering parameters output by the prediction decision module is input to the control execution module; The control signal output by the control execution module drives the precision metering pump; The integrity test data output by the verification feedback module is input into the twin modeling module; The twin modeling module receives the integrity test data and corrects the model parameters to form a closed-loop data stream.

10. A membrane sterilization filtration system, applied to the membrane sterilization filtration device as described in any one of claims 1 to 9, characterized in that, It includes a state awareness module, a twin modeling module, a prediction and decision-making module, a control execution module, and a verification and feedback module; The state sensing module is configured to collect data on the radioactivity of the drug solution flowing through the sterilization filter membrane and the structural strain data of the sterilization filter membrane, and generate a real-time data stream. The twin modeling module is configured to receive the real-time data stream, construct a digital twin of the sterilization filter membrane, dynamically update the digital twin by assimilating the structural strain data, and output the cumulative radiation damage prediction result of the sterilization filter membrane. The prediction and decision module is configured to include a decision agent, which receives the cumulative radiation damage prediction result and the radioactivity data, and generates a filter parameter optimization sequence through the decision agent. The filter parameter optimization sequence includes a flow rate setpoint instruction. The control execution module is configured to receive the flow rate setpoint command and the cumulative radiation damage prediction result, and adjust the operating parameters of the precision metering pump to control the drug filtration flow rate. The verification feedback module is configured to apply test conditions to the sterilization filter membrane after filtration is completed, obtain integrity test data, and send the integrity test data back to the twin modeling module to correct the model parameters of the digital twin.

Images