Dual-probe circulation monitoring and dynamic compensation closed-loop control system and method for cell culture

Abstract

The application discloses a kind of double-probe circulation monitoring and dynamic compensation closed-loop control system and method for cell culture, system includes double-probe circulation in-situ monitoring module, it is installed in reactor single opening by sealing plug, internal integration Raman spectrum probe and dielectric spectrum probe, is driven by cavity motor dynamic fan block and corrugated pipe mechanism control two probes along U-shaped guide rod independent lifting, realize the in-situ circulation collection of multiple height, double mode;And dynamic delay compensation closed-loop control module, it runs model-free adaptive predictive control algorithm with dynamic delay compensation, without kinetic model, on-line identification mixed delay and execution delay, multiple-step prediction metabolic state and rolling optimization feed strategy.The application realizes multidimensional in-situ monitoring with single opening, solves dynamic sterile problem through corrugated tube full-sealing structure, eliminates control oscillation by delay on-line identification, with the advantages of low cost, strong compatibility, high control precision, process good universality.

Description

Technical Field

[0001] This invention belongs to the field of animal cell culture process control technology, specifically relating to an online metabolite monitoring and dynamic feeding control system and method for suspension culture of livestock and poultry cell microcarriers used in veterinary vaccine production. Background Technology

[0002] Large-scale cell culture is a key technological step in the production of biological products such as veterinary vaccines and monoclonal antibodies. During the culture process, real-time monitoring of metabolite concentrations (such as glucose, lactic acid, and glutamine) and dynamic adjustment of feeding strategies accordingly are core control measures to maintain high-density cell growth, ensure product expression levels, and maintain batch stability.

[0003] Currently, the mainstream methods for metabolite monitoring in this field include: offline sampling biochemical analysis, online flow injection analysis (such as YSI, Nova Biomedical), and in-situ Raman spectroscopy monitoring, which has been developed in recent years. Offline sampling has a time lag of more than 30 minutes, making it unsuitable for real-time closed-loop control; online analysis equipment is expensive (over 400,000 RMB per set) and requires external piping, posing a risk of contamination; while in-situ Raman spectroscopy can achieve non-contact, non-consumable continuous monitoring, existing solutions have two structural defects: First, the contradiction between the need for multi-point monitoring and the insufficient reactor opening resources. In large-scale reactors (≥50L), due to the non-ideal nature of stirring and mixing, there are gradient differences in metabolite concentrations in different areas, and single-point monitoring cannot characterize the overall state. If a multi-probe distributed arrangement is used to achieve multi-point monitoring, each probe needs to occupy an independent opening. However, conventional reactor top end caps only have 6-8 standard openings, which are already occupied by pH electrodes, DO electrodes, temperature sleeves, sampling tubes, feed tubes, etc., leaving only 1-2 available spare openings. Multi-point probe solutions are not feasible due to physical opening limitations. Secondly, there is the issue of aseptic sealing during dynamic probe movement. To address insufficient openings, some studies have attempted to use a single probe for monitoring at different heights by lifting and lowering it. However, the probe requires dynamic sealing during lifting and lowering. Traditional O-rings or packing seals are prone to wear and tear on the sealing surface after long-term reciprocating motion, creating gaps and leading to the risk of microbial leakage. Current technology has not yet provided a feasible solution that can withstand repeated high-pressure steam sterilization and maintain an absolute seal after tens of thousands of reciprocating movements.

[0004] In addition to the aforementioned issues with the monitoring end, existing feeding control strategies also have significant shortcomings. Constant-rate or batch feeding cannot match the dynamic changes in cell metabolism, easily leading to insufficient nutrition during the logarithmic growth phase or accumulation of metabolic waste during the plateau phase. DO / pH-stat indirect feeding judges metabolic status by the rebound of dissolved oxygen or pH, resulting in low control precision and delayed response. Existing closed-loop control systems generally employ fixed delay compensation models. However, during culture, as cell density and microcarrier concentration increase, the viscosity of the culture medium rises, and the reactor mixing time constant can extend from 30s to over 90s; simultaneously, the Raman spectroscopy acquisition integration time also needs dynamic adjustment due to signal attenuation. Fixed delay models lead to a dynamic time mismatch between feeding actions and the actual metabolic needs of cells, easily causing periodic oscillations of "overfeeding-underfeeding" in large-scale systems, which ultimately undermines culture stability.

[0005] Therefore, developing a system and method that can solve the above problems simultaneously has significant industrial application value. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a closed-loop control system and method for large-scale cell culture with dual-probe cyclic in-situ monitoring and dynamic delay compensation. The aim is to achieve dual-mode, multi-height cyclic in-situ monitoring of Raman and dielectric spectra using a single reactor opening. By identifying dynamically changing system delay parameters online and combining this with a model-free adaptive predictive control algorithm that requires no pre-modeling, a low-cost, highly compatible, and robust closed-loop feeding control system is constructed, thereby solving the aforementioned long-standing technical challenges.

[0007] The solution to the technical problem of this invention is as follows: a closed-loop control system for dual-probe cyclic monitoring and dynamic compensation in cell culture, comprising: a monitoring module for installation at a single opening of a reactor, the monitoring module including at least two probes, each probe configured to collect different culture state parameters; an independent drive mechanism associated with each probe for driving the probe to independently reciprocate vertically within the reactor to perform in-situ measurements at multiple different heights; and a control module communicatively connected to the monitoring module, the control module including: a dynamic delay identification unit for identifying and updating, in each control cycle, dynamic delay parameters characterizing the entire link from feeding action to response of the system based on real-time data collected by the monitoring module; a model-free predictive controller that does not require a pre-established cell metabolic kinetic model, but instead performs multi-step predictions of the culture state for one or more future control cycles based on the real-time collected data and the dynamic delay parameters, and generates feeding control instructions for the current cycle based on the results of the multi-step predictions; and an execution module communicatively connected to the control module for receiving and executing the feeding control instructions to adjust the feeding operation.

[0008] Preferably, the independent drive mechanism includes: a sealing base fixed to a single opening of the reactor; a guide structure extending from the sealing base into the reactor; at least two sliders slidably disposed on the guide structure, each probe being fixed to one of the sliders; at least two flexible sealing expansion members, one end of each flexible sealing expansion member being sealed to the sealing base and the other end being sealed to one of the sliders, each flexible sealing expansion member forming an independent sealing cavity; and an external drive source that drives the flexible sealing expansion members to extend and retract by controlling the fluid pressure within the independent sealing cavity, thereby causing the sliders and corresponding probes to move along the guide structure.

[0009] Preferably, the external driving source includes: a fluid volume adjustment device having a first chamber and a second chamber with variable volumes; the first chamber is connected to the sealing cavity of the first flexible sealing expansion member via a first fluid pipe; the second chamber is connected to the sealing cavity of the second flexible sealing expansion member via a second fluid pipe; and a drive motor for driving the fluid volume adjustment device such that when the volume of the first chamber decreases, the volume of the second chamber increases, and vice versa, thereby achieving synchronous movement in opposite directions in the vertical direction between the first probe connected to the first flexible sealing expansion member and the second probe connected to the second flexible sealing expansion member.

[0010] Preferably, the at least two probes include a Raman spectroscopy probe and a dielectric spectroscopy probe; the Raman spectroscopy probe is used to collect spectral information of metabolites in the culture medium; the dielectric spectroscopy probe is used to collect dielectric spectral information reflecting cell biomass information.

[0011] Preferably, the dynamic delay identification unit is configured to: apply a small disturbance of a predetermined magnitude at the current feeding rate during stable system operation, collect the response curves of the culture state parameters before and after the disturbance, and calculate the current end-to-end dynamic delay parameters, including reactor mixing delay and actuator response delay, based on the response curves.

[0012] Preferably, the model-free predictive controller employs a model-free adaptive predictive control (MFAPC) algorithm, whose multi-step prediction model is constructed based on online-identified pseudo-partial derivatives, which characterize the local dynamic influence of the feeding operation at the current working point on the culture state parameters.

[0013] Preferably, the execution module includes at least two independently controlled feed pumps, each used to deliver different types of feed solutions; the feed control instructions include instructions to control the at least two feed pumps respectively; the model-free predictive controller is configured to perform rolling optimization based on a multi-objective optimization function, the multi-objective optimization function including at least two of the following objectives: maintaining metabolite concentration within a target range, maximizing metabolic health score, and minimizing feed rate fluctuations.

[0014] Preferably, the control module further includes a metabolic health early warning unit, which has a built-in metabolic health assessment model pre-trained based on historical data. This model is used to calculate a metabolic health score based on the real-time data from the monitoring module and send a feedforward signal to the model-free predictive controller when the score is lower than a preset threshold. The model-free predictive controller then adjusts its control strategy based on the feedforward signal.

[0015] A closed-loop control method for large-scale cell culture based on the system includes the following steps: Step S1: Initialization, setting the target range of metabolites and the control period T; Step S2: In-situ monitoring and data acquisition, within each control period T, the Raman spectroscopy probe and dielectric spectroscopy probe are driven to circulate vertically within the reactor through the dual-probe cyclic in-situ monitoring module, acquiring Raman and dielectric spectroscopy data at at least two different height positions, and obtaining real-time metabolite concentration and live cell density data after processing; Step S3: Dynamic delay online identification, every preset period, a small step disturbance is applied to the current feeding rate, the response curve of metabolite concentration before and after the disturbance is acquired, the current full-link dynamic delay parameters including reactor mixing delay and actuator response delay are calculated online, and the delay compensation model is updated; Step S4: Pseudo Partial derivative online estimation: Based on the input-output data sequence of the most recent L control cycles, the pseudo-partial derivative φ(k) of the system is estimated online. The pseudo-partial derivative represents the local dynamic influence of the feeding operation on the metabolite concentration at the current operating point. Step S5: Multi-step prediction: Based on the pseudo-partial derivative φ(k) and the dynamic delay parameter, a multi-step prediction model is constructed using a model-free adaptive predictive control algorithm to predict the metabolite concentration for the next N control cycles. Step S6: Rolling optimization: With the goal of maintaining the metabolite concentration within the target range, a multi-objective optimization function is solved to calculate the optimal feeding rate increment for the current control cycle. Step S7: Feeding execution: According to the optimal feeding rate increment, the control execution module (320) adjusts the feeding operation. Step S8: Feedback iteration: Return to step S2 and enter the next control cycle to form a closed loop.

[0016] The specific method for obtaining real-time metabolite concentration and live cell density data after processing in step S2 includes: performing wavelet transform denoising and airPLS background subtraction on the acquired Raman spectra in sequence, and then using a pre-trained quantitative model combining partial least squares regression (PLSR) and transfer learning to calculate the concentrations of glucose, lactic acid, glutamine, and ammonia; measuring the permittivity at multiple frequency points on the acquired dielectric spectrum data, and calculating the live cell density (VCD) and average cell diameter through nonlinear fitting of the Cole-Cole model.

[0017] The beneficial effects of this invention are as follows: 1. By independently raising and lowering a dual-probe cyclic in-situ monitoring module (Raman spectroscopy probe + dielectric spectroscopy probe) on a U-shaped guide rod, in-situ cyclic acquisition at four different height points (upper, middle, lower, and bottom reflectors) is achieved using a single reactor opening. The total time for a complete cycle (4 points) is ≤3 minutes. The quantitative accuracy of Raman spectroscopy is ±3% for glucose and ±3% for lactate. The consistency between dielectric spectroscopy live cell density (VCD) and offline counting is R²=0.96.

[0018] 2. Dynamic delay compensation eliminates control oscillations. The dynamic delay identification mechanism detects a shift in mixing delay from 30s to over 90s during culture. After adopting the MFAPC-DDC algorithm (delay step compensation d=1), the feeding rate curve shows a smooth change. Compared with the traditional fixed delay model, the feeding oscillation amplitude is reduced by approximately 90%, the glucose concentration fluctuation is ≤±1.5mmol / L, and the batch-to-batch cell density RSD decreases from 35% to 8%.

[0019] 3. The MFAPC algorithm does not require a pre-established cell metabolic kinetic model. When changing cell lines (e.g., from Vero to MDCK), the system can adapt to the new metabolic characteristics within 1-2 culture cycles through online pseudo-partial derivative estimation, shortening the process transfer and debugging cycle from several weeks to zero. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the dual-probe circulating in-situ monitoring module provided in this embodiment of the invention applied to a reactor.

[0021] Figure 2 for Figure 1 A magnified schematic diagram of the dual-probe cyclic in-situ monitoring module.

[0022] Figure 3 for Figure 2 A schematic diagram of the explosion structure.

[0023] Figure 4 This is a schematic diagram of the overall system architecture provided in an embodiment of the present invention.

[0024] Figure 5The flowchart illustrates a model-free adaptive predictive control algorithm with dynamic delay compensation provided in an embodiment of the present invention.

[0025] Figure 6 This is a connection diagram of the dual-path feeding system provided in an embodiment of the present invention.

[0026] Figure 7 This is a curve showing the effect of the present invention in the culture of Vero cells in 100L microcarriers.

[0027] The diagram labels are as follows: 100-Dual-probe cyclic in-situ monitoring module; 110-Intracavity execution module; 120-Extracavity drive module; 111-Sealing plug; 112-U-shaped guide rod; 113-Slider one; 114-Raman spectroscopy probe; 115-Slider two; 116-Dielectric spectroscopy probe; 117-Signal line one; 118-Signal line two; 119-Bellower; 121-Base; 122-Annular cavity; 123-Fixed cone block; 124-Moving fan block; 125-Cap; 126-Drive motor; 127-First breathing tube; 128-Second breathing tube; 220-Miniature fiber optic Raman spectrometer; 230-Side wall reflector; 300-Dynamic delay compensation closed-loop control module; 310-Controller; 320-Execution module; 321-First peristaltic pump; 322-Second peristaltic pump; 400-Host computer monitoring and early warning module; 500-Bioreactor; A-First air chamber; B-Second air chamber. Detailed Implementation

[0028] The invention will now be further described with reference to the accompanying drawings.

[0029] Example 1: See Figure 4 The present invention provides a closed-loop control system for large-scale cell culture, comprising three core components: a dual-probe cyclic in-situ monitoring module 100, a dynamic delay compensation closed-loop control module 300, and an execution module 320. The dual-probe cyclic in-situ monitoring module 100 is installed at a single opening in the top end cap of the bioreactor 500, and is connected to a Raman spectrometer 220 and a dielectric spectrometer (not shown in the figure) via signal line 117 and signal line 118, respectively. The Raman spectrometer 220 and the dielectric spectrometer transmit the processed data to the controller 310 via a communication bus (such as RS-485 or Modbus TCP / IP). The controller 310 incorporates a dynamic delay identification unit and a model-free predictive controller, and its output is connected to the execution module 320. The execution module 320 includes a first peristaltic pump 321 and a second peristaltic pump 322, and the outlets of the two peristaltic pumps are connected to the feed inlet of the reactor 500 via silicone tubing. The controller 310 also controls the pneumatic solenoid valve through the digital output module, thereby controlling the lifting and lowering action of the dual-probe cyclic in-situ monitoring module 100.

[0030] For the specific structure of the dual-probe cyclic in-situ monitoring module, please refer to [link / reference]. Figure 1 , Figure 2 and Figure 3 The dual-probe circulating in-situ monitoring module 100 consists of two parts: an in-cavity execution module 110 and an external drive module 120. The in-cavity execution module 110 is installed inside the reactor 500. A standard assembly opening of DN25 or DN50 is pre-set on the top end cap of the reactor 500, and a sealing plug 111 is fixed inside the opening. The sealing plug 111 is made of 316L stainless steel, and its outer circumference is machined with threads or clamps to form a detachable sealed connection with the reactor opening. Two vertical through holes are formed in the central area of ​​the sealing plug 111 for signal line 117 and signal line 118 to pass through. The through holes are filled with high-temperature resistant epoxy resin or sintered glass to ensure airtightness.

[0031] A U-shaped guide rod 112 is fixed to the bottom of the sealing plug 111 (located on one side inside the reactor). The U-shaped guide rod 112 consists of two parallel vertical rods and a bottom connecting crossbar, and is made of 316L stainless steel with a mirror-polished surface and a friction coefficient ≤0.1. The upper end of the U-shaped guide rod 112 is fixed to the sealing plug 111 by thread or welding, and the lower end extends to the bottom of the reactor, about 5cm from the bottom of the tank.

[0032] Both slider 113 and slider 115 are hollow cylindrical sleeves with a 0.1-0.3mm gap between their inner diameter and the outer diameter of the U-shaped guide rod 112, and are fitted with sliding bearings made of PTFE (polytetrafluoroethylene) or PEEK (polyetheretherketone). Slider 113 is sealed and fitted onto the first side (left side) of the U-shaped guide rod 112, and slider 115 is sealed and fitted onto the second side (right side). Each slider can slide freely up and down along the rod, with a sliding resistance ≤2N.

[0033] Raman spectroscopy probe 114 is fixed to the side of slider 113. This probe is a commercially available miniature immersion Raman probe (such as the Wasatch Photonics RP785), with an outer diameter of 6 mm and a length of 80 mm. The front end has a sapphire optical window, and the rear end is connected to an optical fiber (signal line 117). The optical fiber passes through a pre-set groove inside slider 113, exits from the corresponding through-hole of sealing plug 111, and connects to the external miniature fiber optic Raman spectrometer 220. In this embodiment, the Raman spectrometer 220 uses a 785 nm excitation wavelength, with a spectral range of 200-2000 cm⁻¹. -1 8cm resolution -1 The integration time is adjustable from 10ms to 60s.

[0034] The dielectric spectroscopy probe 116 is fixed to the side of slider 115. This probe has a four-electrode or two-electrode structure, with electrodes made of 316L stainless steel or platinum-iridium alloy, an exposed electrode length of 15 mm, and an electrode spacing of 3 mm. Signal line 118 is a double-core shielded wire that passes through the groove in slider 115 and the corresponding through-hole in sealing plug 111, connecting to a dielectric spectroscopy analyzer (such as Aber Instruments' Futura or Hamilton's Incyte). The dielectric spectroscopy analyzer can measure permittivity (ε) and conductivity (σ) in the frequency range of 0.1 MHz to 10 MHz.

[0035] There are two bellows 119, each fitted onto one of the two rods of the U-shaped guide rod 112. Each bellows 119 is a welded bellows made of 316L stainless steel, with a wall thickness of 0.15mm, an outer diameter of 16mm, a free length of 40mm, a maximum elongation of 60mm, a compression of 35mm, and an elastic modulus of approximately 2N / mm. The upper end of each bellows 119 is sealed and fixed to the bottom of the sealing plug 111 by clamps or welding, and the lower end is sealed and fixed to the corresponding slider (113 or 115) by clamps. The interior of each bellows 119 forms an independent sealed cavity. When compressed air (0.1-0.4MPa) is filled into this sealed cavity, the bellows 119 elongates, pushing the slider downward; when air is extracted from the sealed cavity (negative pressure -0.05 to -0.08MPa), the bellows 119 contracts, pulling the slider upward. The sealed cavities of the two bellows 119 are independent of each other and are connected to the external drive module 120 through the first breathing tube 127 and the second breathing tube 128, respectively.

[0036] The external drive module 120 is fixed above the top end cap of the reactor 500. See also Figure 2 and Figure 3 The base 121 is fixedly connected to the support of the top end cap of the reactor by screws. A circular groove is machined in the central area of ​​the base 121. The inner wall of the groove and the cover 125 are sealed together by threads and O-rings to form an annular inner cavity.

[0037] Within the annular cavity, a fixed cone block 123 is fixed to the base 121 by a locating pin. The cross-section of the fixed cone block 123 is fan-shaped, with a central angle of approximately 90°. Two independent airways are machined on the inner side of the fixed cone block 123 (the side facing the center of the annular cavity). The inlets of the two airways face the first air chamber A and the second air chamber B in the annular cavity, respectively, and the outlets of the two airways are connected to the connectors of the first breathing tube 127 and the second breathing tube 128, respectively.

[0038] The movable sector block 124 is a sector-shaped block with a central angle of approximately 180°. Its thickness matches the height of the annular inner cavity, and its outer arc surface slides in a sealing manner with the inner wall of the annular inner cavity. A shaft hole is located at the center of the movable sector block 124, which is fixedly connected to the output shaft of the drive motor 126 via a flat key. The drive motor 126 is a stepper motor or servo motor (such as a 42-stepper motor with a holding torque of 0.4 N·m). Its housing is fixed above the cover 125, and its output shaft passes through the central hole of the cover 125 into the annular inner cavity. A rotary sealing ring and a rolling bearing are embedded in the central hole of the cover 125 to ensure sealing and low friction.

[0039] When the drive motor 126 drives the moving fan block 124 to rotate in the forward direction (e.g., clockwise), the moving fan block 124 gradually compresses the volume of the first air chamber A in the annular inner cavity, forcing the gas in the first air chamber A to enter the sealed cavity of the first bellows 119 through the first air passage of the fixed cone block 123 and the first breathing tube 127. The first bellows 119 extends, pushing the Raman spectroscopy probe 114 downward. At the same time, the rotation of the moving fan block 124 causes the volume of the second air chamber B to gradually expand, generating negative pressure. Gas is drawn out from the sealed cavity of the second bellows 119 through the second breathing tube 128. The second bellows 119 contracts, driving the dielectric spectroscopy probe 116 upward. When the drive motor 126 rotates in the reverse direction (counterclockwise), the two probes move in opposite directions. By controlling the rotation angle of the motor 126, the position of the two probes in the vertical direction can be precisely controlled. In this embodiment, every 1° rotation of the motor corresponds to a probe displacement of 0.1mm, and the positioning accuracy can reach ±0.5mm.

[0040] Position detection can be achieved in two ways: one is to install magnetostrictive displacement sensors or Hall switch arrays on the two rods of the U-shaped guide rod 112 to directly read the slider position; the other is to use the encoder of the drive motor 126 to provide feedback on the rotation angle, and calculate the theoretical position of the probe by combining the proportional relationship of the transmission mechanism. This embodiment prefers the second method to reduce internal leads and lower the risk of contamination.

[0041] Before system startup, reactor 500 and dual-probe circulating in-situ monitoring module 100 are subjected to overall in-situ cleaning (CIP) and high-pressure steam sterilization (SIP). The sterilization conditions are 121°C for 30 minutes. The sealing structure of bellows 119 and sealing plug 111 has been tested and can withstand ≥100 sterilization cycles without leakage.

[0042] After sterilization, culture medium containing 10 g / L microcarriers (Cytodex 1) and 10% serum was injected into the reactor, and Vero cells were inoculated to an initial density of 2 × 10⁶ cells / year. 5 cells / mL. The control module is set to monitor a cycle time of T=5min.

[0043] Within each monitoring cycle, the controller 310 executes according to the following timing sequence: t=0s: the drive motor 126 rotates, moving the Raman spectroscopy probe 114 to a position 20cm below the liquid surface (upper position), and the dielectric spectroscopy probe 116 to a position 50cm below the liquid surface (middle position). After stabilizing for 0.5s, the Raman spectrometer 220 begins integration and acquisition, with an integration time of 30s; simultaneously, the dielectric spectroscopy analyzer begins measurement, acquiring data once at each of the six frequency points of 0.2, 0.5, 1, 2, 5, and 10MHz, taking approximately 2s.

[0044] At t=30s: Drive motor 126 reverses direction, Raman probe 114 descends to 50cm below the liquid surface (middle position), and dielectric spectroscopy probe 116 rises to 20cm below the liquid surface (upper position). After stabilizing for 0.5s, spectral and dielectric spectrum acquisition is performed again.

[0045] t=60s: Drive motor 126 reverses again, Raman probe 114 descends to 80cm below the liquid surface (lower position, about 10cm from the bottom of the tank), and dielectric spectroscopy probe 116 descends to 80cm below the liquid surface (lower position). Data collection begins.

[0046] At t=90s: the drive motor 126 lowers the Raman probe 114 further to the corresponding height of the bottom reflector 230 (approximately 2cm from the bottom of the tank) to collect the reflection spectrum. The dielectric spectrum probe 116 remains at the lower position.

[0047] t=120s: Complete one full cycle. Raman spectral data (4 spectra in total) and dielectric spectral data (4 depths × 6 frequencies = 24 data points) from four locations have all been acquired and transmitted to the controller 310.

[0048] After receiving the raw Raman spectrum, the controller 310 first performs preprocessing: denoising is achieved using a three-layer wavelet transform based on the sym5 wavelet basis, with a soft thresholding method used; subsequently, the airPLS (adaptive iterative weighted penalized least squares) algorithm is used to subtract the fluorescence background, with 10 iterations and a lambda parameter of 10. 5 The processed spectral range is 200-2000 cm⁻¹. -1 .

[0049] Quantitative analysis was performed using a partial least squares regression (PLSR) model. The PLSR model was pre-trained using 217 batches of publicly available metabolomics data and 58 batches of experimental data from our university. The input was pre-processed Raman spectra (200-2000 cm⁻¹). -1The dataset contains 1801 wavenumber points, and the outputs are the concentrations of glucose, lactate, glutamine, and ammonia. To adapt to different cell lines and culture media, transfer learning was introduced during training: first, the base model was trained on a general dataset, and then fine-tuned on a small batch of data (5 batches) from the target cell line. During fine-tuning, only the weights of the output layer were adjusted, while the first 5 hidden layers were fixed. The quantitative accuracy verification results were: glucose ±3% (concentration range 1-25 mmol / L), lactate ±3% (0-30 mmol / L), glutamine ±5% (0-8 mmol / L), and ammonia detection limit 5 μmol / L.

[0050] Dielectric spectrum data were fitted using the Cole-Cole model. The measured permittivity ε(f) is related to the frequency f as follows: ε(f) = ε_∞ +(ε_0-ε_∞) / [1 +(j2πfτ)^(1-α)] (1), where ε_0 is the low-frequency limiting permittivity, ε_∞ is the high-frequency limiting permittivity, τ is the relaxation time, and α is the distribution parameter. ε_0 was obtained by nonlinear least squares fitting (Levenberg-Marquardt algorithm). The live cell density VCD is linearly related to ε_0: VCD = k·(ε_0-ε_b), where ε_b is the background permittivity without cells, and k is the calibration coefficient (calibrated by offline sampling and counting, typical value 0.8×10). 6 Cells / mL·pF / cm). Cell diameter d is calculated from τ: d ∝ (τ)^(1 / 2). In this embodiment, the VCD measurement range is 1×10⁻⁶. 5 -1×10 7 cells / mL, with an average absolute deviation of ≤8% compared to offline counts.

[0051] Controller 310 performs dynamic delay identification every two control cycles (i.e., every 10 minutes). The specific steps are as follows:

[0052] 1. Based on the current feeding rate u(k), apply a step increase of 5% of the current rate. In this embodiment, the current glucose feeding rate is assumed to be 3.2 mL / min, then the rate after the step increase becomes 3.36 mL / min.

[0053] 2. After the step is applied, the system continuously monitors the change in glucose concentration at a sampling interval of 0.5 min for 10 minutes. Since the complete monitoring cycle of Raman spectroscopy is 5 min, the system still acquires data according to the normal cycle during the first 0-5 min after the step, but the processed data is immediately used for delayed identification; during the second 5-10 min, the system acquires another data point.

[0054] 3. Apply Kalman filtering to the glucose concentration-time series. The state equation is set as x(k+1) = x(k) + w(k), and the observation equation is z(k) = x(k) + v(k), where w(k) and v(k) are the process noise and measurement noise, respectively, with initial covariance values ​​set to Q = 0.01 and R = 0.5. A smooth response curve is obtained after filtering.

[0055] 4. Define the time point corresponding to the peak value of the first derivative of the response curve as the hybrid delay τ. m (That is, the time when the probe first detects a significant concentration change after the feed is introduced, due to stirring and diffusion). The time point when the response curve reaches 63.2% of the new steady-state value is compared with τ. m The difference is defined as the time constant T. s Execution delay τ e Defined as the time from when the feed pump receives the instruction to when it actually dispenses the feed liquid (in this embodiment, the peristaltic pump τ). e ≈0.5s, negligible). For a 50L reactor, the stirring speed is 80rpm, τ m Typically, this ranges from 30 to 90 seconds and increases with increasing cell density.

[0056] 5. Calculate the total delay steps d = round((τ) m +τ e ) / T), where T=5min. If τ m =70s, then the total delay is approximately 70.5s, divided by 300s gives 0.235, which is rounded down to 0. However, in actual biological processes, due to mixing and metabolic responses, the true effective delay may be close to one cycle. This embodiment adopts a conservative strategy, with d set to 1. The delay compensation model is D(z) = z -1 .

[0057] Model-Free Adaptive Predictive Control (MFAPC) Algorithm: The MFAPC algorithm embedded in the controller 310 does not require a pre-established cell metabolic kinetic model, but instead learns online based on real-time input and output data. The algorithm parameters are set as follows: control period T = 5 min, prediction time domain N = 4, control time domain M = 2, penalty factor λ = 0.01, weight matrix Q = diag(1, 0.5), R = 0.1.

[0058] In each control cycle k, the algorithm executes the following sub-steps: Step 1: Online estimation of pseudo-partial derivatives. Using the input-output data sequence {u(k-4), u(k-3), u(k-2), u(k-1), u(k)} and {y(k-4),..., y(k)} of the most recent L=5 cycles, the pseudo-partial derivative φ(k) of the current system is estimated using an improved projection algorithm: φ(k) = φ(k-1) +η·Δu(k-1)·[Δy(k) -φ(k-1)·Δu(k-1)] / (μ +‖Δu(k-1)‖ 2 (2), where Δy(k)=y(k)-y(k-1), Δu(k)=u(k)-u(k-1), η=0.6 is the step size factor, and μ=1 is the small amount to prevent zeroing. φ(k) is a scalar or vector (a matrix when there are multiple inputs and multiple outputs). In this embodiment, a single input (glucose feeding rate) and a single output (glucose concentration) control are used. φ(k) is a scalar, and its physical meaning is the local dynamic gain of the change in feeding rate on the change in glucose concentration at the current working point.

[0059] Step 2: Dynamic delay compensation, the identified delay step number d (d=1 in this embodiment) is introduced into the prediction model. Construct the delayed input increment sequence: Δu_d(k) = u(kd) u(kd-1) (3).

[0060] Step 3: Multi-step prediction, using the following model to predict glucose concentrations for the next 4 cycles:

[0061] ŷ(k+1|k) = y(k) +φ(k)·Δu(k+1-d); ŷ(k+2|k) = y(k) +φ(k)·Δu(k+1-d)+φ(k+1)·Δu(k+2-d); ŷ(k+3|k) = y(k) +φ(k)·Δu(k+1-d) +φ(k+1)·Δu(k+2-d)+φ(k+2)·Δu(k+3-d); ŷ(k+4|k) = y(k) +Σ_{i=0}^{3} φ(k+i)·Δu(k+1+id)(4). Since φ(k+1) and φ(k+2) are unknown, they are approximated by the current estimated value φ(k), that is, it is assumed that the pseudo-partial derivative remains unchanged in the prediction time domain.

[0062] Step 4: Rolling optimization, solve the following quadratic programming problem to obtain the optimal control increment sequence Δu(k), Δu(k+1), Δu(k+2), min J = Σ_{j=1}^{4} [Q·(ŷ(k+j)- y_ref) 2 ] +R·Σ_{i=0}^{2} [Δu(k+i)] 2(5), where y_ref is the target glucose concentration setpoint (7.5 mmol / L in this embodiment), Q=1, R=0.1. Constraints: 0.5 ≤ u(k)+Δu(k) ≤ 10 mL / min, |Δu(k)| ≤ 1 mL / min. Optimization is performed using gradient descent, iterating 20 times. The first control increment Δu(k) is executed, and the rest are discarded.

[0063] Step 5: Execution and Feedback. Controller 310 converts the calculated new feeding rate u(k+1) = u(k) + Δu(k) into a 4-20mA analog signal and outputs it to the driver of the first peristaltic pump 321. The peristaltic pump 321 (model: LongerPumpBT100-2J) has a flow accuracy of ±1% and a response time ≤0.2s. Meanwhile, the control strategy of the second peristaltic pump 322 is similar to that of the first pump, but its control target is lactate concentration (target value ≤15mmol / L). When the lactate concentration exceeds 12mmol / L, the algorithm automatically reduces the glucose feeding rate and considers switching part of the feeding to galactose.

[0064] For detailed configuration of dual-path feeding, please refer to [link / reference]. Figure 6 The inlet of the first peristaltic pump 321 extends through a silicone tube into a basic feed bottle (containing 10×DMEM concentrate, including glutamine, amino acids, and vitamins, but no glucose), and its outlet is connected to the feed port of reactor 500 through a silicone tube. The inlet of the second peristaltic pump 322 extends into a glucose concentrate feed bottle (containing 500 g / L glucose), and its outlet is also connected to the same feed port of reactor 500 (connected via a Y-type tee). The two pumps are independently controlled, allowing for decoupled regulation of total nutrient and carbon source.

[0065] When the lactate concentration exceeds 15 mmol / L, the controller 310 switches the feed solution of the second peristaltic pump 322 to galactose concentrate (200 g / L galactose), continuing for 2-3 cycles until the lactate concentration drops below 12 mmol / L, then switches back to glucose. The switching is achieved through a three-way solenoid valve, controlled by the PLC's digital output module.

[0066] Figure 7The graphs show the changes in glucose concentration, lactate concentration, and glucose feed rate over time for a typical batch of this embodiment. As can be seen from the graph, the glucose concentration, initially 25 mmol / L (from the culture medium), entered the target range of 5-10 mmol / L after cell consumption within approximately 12 hours, and was then stably controlled between 6-8 mmol / L, with fluctuations ≤ ±1.5 mmol / L. The lactate concentration slowly increased throughout the culture period (120 hours), but never exceeded 16 mmol / L, eventually stabilizing at 13-15 mmol / L. The feed rate exhibited a "low-high-low" trend, highly matching the cell growth cycle (lag phase-log phase-plateau phase). The maximum feed rate occurred at 48-72 hours (late logarithmic growth phase), reaching 8.2 mL / min. The peak viable cell density (monitored in real-time by dielectric spectroscopy) occurred at 96 hours, reaching 3.2 × 10⁻⁶. 6 cells / mL, viability ≥90%. Offline sampling and counting verification showed that the consistency between real-time dielectric spectrometry VCD and offline counting was R 2 =0.96.

[0067] Compared to traditional constant-rate feeding (fixed at 3 mL / min), cell density increased by 45% and lactate peak decreased by 38%. Compared to manual feeding based on offline sampling, the batch-to-batch relative standard deviation (RSD) of cell density decreased from 35% to 8%. Hardware cost statistics: Miniature Raman spectrometer (domestic) 80,000 RMB, dielectric spectrometer (domestic) 35,000 RMB, probe and lifting mechanism (including bellows, motor, and machined parts) 20,000 RMB, PLC and peristaltic pump 20,000 RMB, total cost approximately 155,000 RMB, a 65% reduction compared to the imported YSI+Nova combination (approximately 450,000 RMB).

[0068] Example 2: A preferred scheme incorporating metabolic health early warning and feedforward control. Building upon Example 1, this example further introduces a metabolic health early warning unit and a feedforward control strategy based on this early warning. Construction of the metabolic health scoring model: A fuzzy comprehensive evaluation model based on metabolomics data is pre-stored within the controller 310. The training data for this model comes from the NCBI GEO database (GSE series, 217 batches in total) and 58 batches of Vero, ST, and MDCK cell culture data accumulated in our laboratory. Each batch of data includes the concentrations of six metabolites (glucose, lactate, glutamine, ammonia, glutamate, and glutathione) at eight time points, along with the corresponding final cell density and viability labels.

[0069] Model building steps:

[0070] 1. Principal Component Analysis (PCA) Dimensionality Reduction: After standardizing the 6-dimensional metabolite concentration data, the first 3 principal components (cumulative variance contribution rate ≥ 85%) are extracted.

[0071] 2. Fuzzification: The membership function of each principal component adopts a triangular membership function, which is divided into three levels: "low", "medium" and "high".

[0072] 3. Rule Base: Based on 58 batches of experimental data, 15 fuzzy rules were extracted through cluster analysis. For example: IF glucose is "low" AND lactate is "high" AND ammonia is "high" THEN health score is "low".

[0073] 4. Defuzzification: The centroid method is used to calculate the output, resulting in a continuous health score H from 0 to 100.

[0074] 5. Threshold setting: Statistical analysis was conducted on batches that eventually experienced metabolic collapse (cell viability ≤70%) in historical data. H<40 was defined as a red alert (high risk), H between 40-60 as a yellow alert (sub-health), and H>60 as normal.

[0075] Model Validation: On 10 independent test sets, the model's early warning time for metabolic collapse was 12-18 hours, with a false alarm rate (H<40 but no actual collapse) of 13.5% and a false negative rate (H≥40 but actual collapse) of 5%. Feedforward Control Strategy: During normal operation, the controller 310 calculates the current metabolic health score H(k) after each monitoring cycle (5 minutes). A yellow warning is triggered when H(k) is below 60 twice consecutively; a red warning is triggered when it is below 40 twice consecutively. The warning signal is directly fed into the rolling optimization stage of MFAPC, dynamically adjusting the optimization target. Under a yellow warning: the metabolite set range is tightened. For example, the glucose target range is adjusted from 5-10 mmol / L to 6-8 mmol / L, and the lactate target upper limit is adjusted from 20 mmol / L to 15 mmol / L. At the same time, the weight Q1 (metabolite deviation weight) in the optimization objective function is increased by 50%, and R (feed fluctuation weight) is reduced by 30%, making the controller more proactive in maintaining metabolic homeostasis. Under a red alert: In addition to the adjustments mentioned above, "active intervention" is implemented: the feed solution in the second peristaltic pump 322 is forcibly switched to galactose for 4 cycles (20 min), while the total feed rate is reduced by 30% to alleviate metabolic stress. If H does not recover after 3 consecutive red alerts, the system will issue an alarm to "suggest termination of culture".

[0076] This feedforward control strategy was tested in three batches of Vero cell cultures. In one batch, due to serum batch quality issues, cells began exhibiting metabolic abnormalities at 72 hours. The system in Example 2 issued a yellow alert at 60 hours (i.e., when cell viability was still 85%) and automatically adjusted the control parameters. Ultimately, the highest cell density for this batch still reached 2.8 × 10⁻⁶. 6The cell viability was maintained above 80%, while the control batch (same serum batch) that did not use the warning module experienced metabolic collapse after 72 hours, with cell viability dropping below 60%, forcing premature termination.

[0077] In a preferred embodiment, the controller 310 also dynamically adjusts the monitoring period and Raman integration time based on the real-time live cell density (VCD). The specific rule is as follows: when VCD < 0.5 × 10⁻⁶... 6 When the cell density is low (0.5-2.0 × 10⁻¹¹ cells / mL), cell metabolism is slow. Therefore, the monitoring period is extended to 10 min, and the Raman integration time is shortened to 10 s (signal-to-noise ratio is sufficient) to reduce the potential impact of the laser on cells and save energy. When the VCD is in the range of 0.5-2.0 × 10⁻¹¹ cells / mL... 6 When the number of cells / mL is between 2.0 and 10, a standard cycle of 5 minutes and an integration time of 30 seconds are used. When VCD > 2.0 × 10 6 When the cell / mL ratio is high, cell metabolism is vigorous, the system viscosity increases, and the mixing delay becomes larger. Therefore, the monitoring cycle is shortened to 3 min, the integration time is extended to 60 s (to improve the spectral signal-to-noise ratio), and the frequency of dynamic delay identification is increased from once every 2 cycles to once per cycle.

[0078] In a 100L scale test, compared with a fixed 5-minute cycle, the control precision (standard deviation of glucose concentration) at the peak of cell density decreased from 0.9 mmol / L to 0.5 mmol / L, and the total Raman laser irradiation time was reduced by about 40%, thus reducing potential phototoxicity.

[0079] It should be noted that the above embodiments and accompanying drawings are merely illustrative examples of the core principles and key structures of the system and method of the present invention. The accompanying drawings are simplified schematic diagrams, intended to clearly illustrate the structural, process, or data flow relationships related to the innovative points of the technical solution, and are not intended to limit the complete form of the actual product. This specification focuses on the innovative technical means necessary to achieve the purpose of the invention and solve the technical problem; auxiliary or common-sense details such as pipe insulation, cable routing, conventional filter circuits, standard component selection, and mounting brackets, which can be implemented by those skilled in the art without creative effort, are not described in detail, but should be understood as naturally included in the specific implementation of the present invention and fall within the protection and implementation scope of this technical solution. Any non-creative adjustments or equivalent substitutions to the dimensions, materials, and control parameters of various components based on the technical principles of the present invention do not depart from the spirit and protection scope of the present invention.

Claims

1. A closed-loop control system for dual-probe cyclic monitoring and dynamic compensation in cell culture, characterized in that, include: A monitoring module, for installation at a single opening of the reactor, the monitoring module comprising: At least two probes, each configured to acquire different culture state parameters; An independent drive mechanism, associated with each probe, is used to drive the probe to reciprocate independently in the vertical direction within the reactor to perform in-situ measurements at multiple different height positions; A control module, communicatively connected to the monitoring module, includes: The dynamic delay identification unit is used to identify and update the dynamic delay parameters that characterize the entire link between the feeding action and the response of the system in each control cycle, based on the data collected in real time by the monitoring module. The model-free predictive controller does not require a pre-established cell metabolic kinetic model. Instead, it makes multi-step predictions of the culture status for one or more future control cycles based on the real-time acquired data and the dynamic delay parameters, and generates the feeding control command for the current cycle based on the results of the multi-step predictions. The execution module is communicatively connected to the control module and is used to receive and execute the material replenishment control command to adjust the material replenishment operation.

2. The system according to claim 1, characterized in that, The independent drive mechanism includes: a sealing base fixed to a single opening of the reactor; a guide structure extending from the sealing base into the reactor; at least two sliders slidably disposed on the guide structure, each probe being fixed to one of the sliders; at least two flexible sealing expansion members, one end of each flexible sealing expansion member being sealed to the sealing base and the other end being sealed to one of the sliders, each flexible sealing expansion member forming an independent sealing cavity; and an external drive source that drives the flexible sealing expansion members to extend and retract by controlling the fluid pressure within the independent sealing cavity, thereby causing the sliders and corresponding probes to move along the guide structure.

3. The system according to claim 2, characterized in that, The external drive source includes: a fluid volume adjustment device having a first chamber and a second chamber with variable volumes; the first chamber is connected to the sealing cavity of the first flexible sealing expansion member via a first fluid conduit; the second chamber is connected to the sealing cavity of the second flexible sealing expansion member via a second fluid conduit; and a drive motor for driving the fluid volume adjustment device such that when the volume of the first chamber decreases, the volume of the second chamber increases, and vice versa, thereby achieving synchronous movement in opposite directions in the vertical direction between the first probe connected to the first flexible sealing expansion member and the second probe connected to the second flexible sealing expansion member.

4. The system according to claim 1, characterized in that, The at least two probes include a Raman spectroscopy probe and a dielectric spectroscopy probe; the Raman spectroscopy probe is used to acquire spectral information of metabolites in the culture medium; the dielectric spectroscopy probe is used to acquire dielectric spectral information reflecting cell biomass.

5. The system according to claim 1, characterized in that, The dynamic delay identification unit is configured to: apply a small disturbance of a predetermined magnitude to the current feeding rate during stable system operation, collect the response curves of the culture state parameters before and after the disturbance, and calculate the current end-to-end dynamic delay parameters, including reactor mixing delay and actuator response delay, based on the response curves.

6. The system according to claim 1, characterized in that, The model-free predictive controller employs the model-free adaptive predictive control (MFAPC) algorithm. Its multi-step prediction model is constructed based on online-identified pseudo-partial derivatives, which characterize the local dynamic influence of feeding operations on culture state parameters at the current working point.

7. The system according to claim 1, characterized in that, The execution module includes at least two independently controlled feed pumps, each used to deliver different types of feed solutions; the feed control instructions include instructions to control the at least two feed pumps respectively; the model-free predictive controller is configured to perform rolling optimization based on a multi-objective optimization function, which includes at least two of the following objectives: maintaining metabolite concentration within a target range, maximizing metabolic health score, and minimizing feed rate fluctuations.

8. The system according to any one of claims 1 to 7, characterized in that, The control module also includes a metabolic health early warning unit, which has a built-in metabolic health assessment model pre-trained based on historical data. This model is used to calculate a metabolic health score based on the real-time data from the monitoring module and send a feedforward signal to the model-free predictive controller when the score is lower than a preset threshold. The model-free predictive controller then adjusts its control strategy based on the feedforward signal.

9. A closed-loop control method for large-scale cell culture based on the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: Step S1: Initialization, setting the target range for metabolites and the control period T; Step S2: In-situ monitoring and data acquisition. In each control cycle T, the Raman spectroscopy probe (114) and dielectric spectroscopy probe (116) are driven to move vertically in the reactor through the dual-probe cyclic in-situ monitoring module (100) to collect Raman spectroscopy and dielectric spectroscopy data at at least two different height positions. After processing, real-time metabolite concentration and live cell density data are obtained. Step S3: Dynamic delay online identification. Every preset period, a small step disturbance is applied to the current feeding rate. The response curves of metabolite concentration before and after the disturbance are collected. The current dynamic delay parameters of the entire link, including reactor mixing delay and actuator response delay, are calculated online, and the delay compensation model is updated. Step S4: Online estimation of pseudo-partial derivatives. Based on the input-output data sequence of the most recent L control cycles, the pseudo-partial derivatives φ(k) of the system are estimated online. The pseudo-partial derivatives characterize the local dynamic influence of the feeding operation on the metabolite concentration at the current operating point. Step S5: Multi-step prediction. Based on the pseudo-partial derivative φ(k) and the dynamic delay parameter, a multi-step prediction model is constructed using a model-free adaptive predictive control algorithm to predict the metabolite concentration for the next N control cycles. Step S6: Rolling optimization, with the goal of maintaining metabolite concentration within the target range, solve the multi-objective optimization function and calculate the optimal feed rate increment for the current control cycle; Step S7: Perform material replenishment. Based on the optimal material replenishment rate increment, control the execution module (320) to adjust the material replenishment operation. Step S8: Feedback iteration, return to step S2, enter the next control cycle, and form a closed loop.

10. The method according to claim 9, characterized in that, The specific method for obtaining real-time metabolite concentration and live cell density data after processing in step S2 includes: performing wavelet transform denoising and airPLS background subtraction on the acquired Raman spectra in sequence, and then using a pre-trained quantitative model combining partial least squares regression (PLSR) and transfer learning to calculate the concentrations of glucose, lactic acid, glutamine, and ammonia; measuring the permittivity at multiple frequency points on the acquired dielectric spectrum data, and calculating the live cell density (VCD) and average cell diameter through nonlinear fitting of the Cole-Cole model.

Images