Bioreactor multi-parameter adaptive control system and method based on machine learning

Abstract

The invention discloses a bioreactor multi-parameter adaptive control system and method based on machine learning, and belongs to the technical field of bioreactor intelligent control, and the bioreactor multi-parameter adaptive control system comprises a bioreactor hepatocyte culture unit, a data acquisition and control unit and a machine learning intelligent decision-making unit. And the data acquisition and control lower computer controls operation variables such as gas supply, sampling and perfusion in the culture unit execution mechanism, collects culture process data returned by the parameter detection module, constructs a data-driven cell state prediction model by utilizing machine learning, and optimizes model parameters and operation variables based on a prediction result to realize closed-loop control. The method can overcome the defect that there is no precise mechanism model for hepatic cell culture, and multi-parameter collaborative optimization can be realized only depending on process data, so that the state of the hepatic cells is sensed in real time, full-cycle optimization of cell viability and functions is realized, and the activity and density of hepatic cell culture are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for bioreactors, and in particular to a multi-parameter adaptive control system and method for bioreactors based on machine learning. Background Technology

[0002] A bioartificial liver is an extracorporeal support system that allows artificially cultured hepatocytes to exchange substances bidirectionally with the patient's blood through a semi-permeable membrane or direct contact. Its function is similar to that of a human liver, making it an effective means of compensating for liver function in patients with severe liver failure. The bioreactor, as the core component of the bioartificial liver, provides sufficient gas and nutrient supply to the hepatocytes and promptly removes metabolic waste, ensuring good hepatocyte activity and enabling them to perform their specific liver functions, thus directly affecting the efficacy of the bioartificial liver. Therefore, controlling the hepatocyte culture process by regulating relevant parameters of the bioreactor to achieve high-density, large-scale expansion and culture of hepatocytes in vitro, obtaining a large number of functionally active hepatocytes, is currently one of the research hotspots in the clinical application of bioartificial livers.

[0003] Hepatocyte culture is characterized by high hysteresis, multiple parameters, and strong coupling. However, most bioreactor control methods rely on empirical exploration and manual adjustment of operating conditions. The differences in bioreactor types and culture methods mean that some of these experiences lack universality, significantly increasing time and economic costs. Alternatively, single-variable PID control or fixed setpoint control may be used, but the independent control of each parameter lacks coordination, easily leading to process control failure. Furthermore, the nonlinear and time-varying nature of hepatocyte culture makes it difficult to construct accurate mathematical models (mechanistic models) and accurately determine the dynamic changes in cell state and culture environment during the culture process.

[0004] Existing technologies have optimized bioreactor control methods to some extent, achieving preliminary applications of machine learning-based bioreactor control. For example, patent CN113168141B discloses a bioprocess control method based on spectrophotometry, which uses orthogonal partial least squares analysis to establish a single model to predict multiple key biological processing variables at multiple growth stages, and generates PID parameters based on the prediction results, achieving a direct correlation between spectral data and process control parameters under a simple system structure. However, this system lacks multi-parameter collaborative optimization and has a single control strategy. Patent CN120065750B discloses a modular CSTR fermentation simulation equipment and an adaptive collaborative control method, which adopts a two-layer cascaded control architecture and dynamically adjusts target parameters based on an expert system. This can improve processing efficiency while maintaining system stability, but the expert system rules rely on experience, have limited learning ability, and have high model complexity. The intelligent bioreactor system disclosed in patent CN120276251A uses multimodal data analysis and cross-attention mechanism to process multimodal data. It decodes fusion features through neural network and performs real-time online regulation based on the state of cell clusters, realizing an end-to-end control architecture and reducing human intervention. However, it does not take into account key indicators such as cell metabolic state and functional expression, nor does it fully consider the dynamic evolution of time-series data during the culture process.

[0005] A comprehensive analysis of existing technologies reveals that, despite some progress in data monitoring, data processing, and control strategies, the following common technical problems still exist:

[0006] (1) Data utilization is insufficient, with most data sources being used from a single source. There is a lack of deep integration and collaborative utilization of multimodal data, and the dynamic evolution of time series data is not fully considered.

[0007] (2) The control strategy is inflexible, mostly based on fixed rules or simple models, lacks predictive control, has poor adaptive ability, insufficient multi-parameter collaborative optimization ability, and cannot learn and optimize autonomously during operation;

[0008] (3) Lack of biological understanding, failure to consider the complex effects of cell metabolic activities, lack of quantitative assessment of cell functional status, and poor correlation between control targets and cell physiological status. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a machine learning-based multi-parameter adaptive control system and method for bioreactors, enabling full-cycle optimization of cell activity and function, and effectively improving the activity and density of hepatocyte culture.

[0010] The technical solution of the present invention is as follows:

[0011] A machine learning-based multi-parameter adaptive control system for a bioreactor, the system comprising:

[0012] The bioreactor hepatocyte culture unit includes a bioreactor, a parameter detection module, and an actuator.

[0013] Data acquisition and control unit, including lower-level machine and human-machine interface;

[0014] The machine learning intelligent decision-making unit includes a cell state prediction model and a multi-objective optimization decision-making module.

[0015] Furthermore, the bioreactor can be of different types, including stirred bioreactors, hollow fiber bioreactors, corrugated bioreactors, and other bioreactors suitable for hepatocyte culture.

[0016] Furthermore, the parameter detection module can monitor important culture parameters such as extracellular environmental parameters and biochemical parameters of hepatocytes online through real-time detection methods such as sensors, including pH value, dissolved oxygen (DO) concentration, temperature, glucose concentration, lactate concentration, etc.

[0017] Furthermore, the actuator mainly includes: a gas supply module, a sampling module, a perfusion module, a stirring module, a heating module, etc., wherein:

[0018] The gas supply module includes nitrogen valves, oxygen valves, air valves, and carbon dioxide valves for adjusting pH and DO concentration, as well as a gas mass flow controller for controlling and returning gas flow values.

[0019] The sampling module can sample cells and culture medium through sampling pumps or syringes to perform offline detection of specific functional parameters and growth status of hepatocytes, such as urea, albumin, cell concentration, and morphological characteristics of hepatocytes (cell aggregation degree).

[0020] The perfusion module uses one or more peristaltic pumps with adjustable speed and flow direction to realize the liquid feeding, liquid discharging, liquid replenishment, and liquid replacement operations of the bioreactor;

[0021] The stirring module is used to stir the suspension culture operation of the bioreactor, and can be stirred by magnetic force or motor.

[0022] The heating module can provide a constant temperature for the bioreactor through heating jackets, temperature controllers, and other means.

[0023] Furthermore, the lower-level machine can be a programmable logic controller (PLC) or a computer to collect data from the parameter detection module, transmit it to the machine learning intelligent decision-making unit for processing, and use the returned optimized operating conditions to control the actuator, and send the actual detection values ​​and operating variable values ​​to the human-machine interface for display.

[0024] Furthermore, the human-computer interaction interface can initially set the state of the cultured target cells, operating conditions and constraints, send the set values ​​to the lower-level machine, and visualize various detection parameters and operating variables.

[0025] Furthermore, the data preprocessing module receives real-time collected time-series data and performs cleaning, alignment, standardization, and feature extraction based on data characteristics.

[0026] Furthermore, the cell state prediction module is used to build sequence models that capture spatiotemporal dependencies, such as one of the following: recurrent neural network (RNN), long short-term memory network (LSTM), Transformer model, gated recurrent unit (GRU).

[0027] Furthermore, the multi-objective optimization decision module employs an optimization algorithm that is one of the following: model predictive control (MPC) framework, Bayesian optimization, reinforcement learning, etc. The goal is to optimize the predicted cell state value. The optimal operating parameter setpoint is calculated using the optimization algorithm, and the bioreactor is controlled according to the setpoint.

[0028] Furthermore, the online learning module utilizes new data to continuously update the cell state prediction module and the multi-objective optimization decision module, thereby adaptively optimizing the model parameters.

[0029] A machine learning-based method for multi-parameter control of bioreactors includes the following steps:

[0030] (1) Input the target cell status, operating conditions and constraints into the human-computer interaction interface, and send the command to the lower computer to execute each module of the bioreactor hepatocyte culture unit;

[0031] (2) The data acquisition module acquires the time-series data of the multi-source sensors in the bioreactor in real time, and the data is collected by the lower-level machine and sent to the machine learning intelligent decision-making module;

[0032] (3) Send the time series data to the data preprocessing module for cleaning, alignment, standardization and feature extraction to obtain feature vectors;

[0033] (4) Input the feature vector into the trained cell state prediction model to obtain the predicted cell state value at future time.

[0034] (5) With the goal of optimizing the predicted value of cell state, the multi-objective optimization decision module uses the optimization algorithm to solve the optimal combination of operation parameter settings based on the predicted value and the set target.

[0035] (6) The online learning module uses new data to continuously update the cell state prediction module and the multi-objective optimization decision-making module;

[0036] (7) Return the set value to the lower-level machine, so that the lower-level machine can output instructions to adjust the actuator and realize closed-loop control.

[0037] Preferably, the machine learning-based bioreactor multi-parameter control method further incorporates multimodal data fusion and multi-task learning mechanisms to improve the accuracy of cell state prediction. Specifically:

[0038] (1) The system optionally configures a multimodal data fusion module in the machine learning intelligent decision-making unit to further integrate the real-time acquired time-series data. The data not only includes process environmental parameters and biochemical parameters detected by the parameter detection module, but also integrates image analysis unit and offline detection data through the sampling module, fusing morphological features reflecting the state of cell populations (such as cell aggregation degree) and molecular marker features reflecting cell-specific functions (such as urea and albumin). The system unifies these heterogeneous data into high-information-density feature vectors through feature-level fusion or model-level fusion architecture to enhance the accuracy of state prediction.

[0039] (2) The prediction model of the cell state prediction module is preferably a recurrent neural network (such as LSTM, GRU) or Transformer architecture with attention mechanism, so as to learn the temporal dynamic characteristics and spatial distribution characteristics of the monitoring data at the same time, and more accurately capture the evolution of cell state with culture time and space.

[0040] (3) The cell state prediction model adopts a multi-task learning framework, and its output includes cell state classification signals and key indicator regression prediction values, thus providing a more comprehensive and reliable basis for the multi-objective optimization decision module.

[0041] Beneficial effects:

[0042] Compared with existing technologies, this invention achieves fully automated closed-loop control from "monitoring-prediction-decision-execution", breaking the limitations of traditional single-variable control; it overcomes the shortcomings of the lack of precise mechanistic models in hepatocyte culture, and can achieve multi-parameter collaborative optimization based solely on process data, with strong generalization ability; it can dynamically adjust according to the real-time metabolic feedback of cells, and has strong anti-interference and adaptability to the time-varying characteristics of the process, thereby sensing the state of hepatocytes in real time, realizing the full-cycle optimization of cell activity and function, and effectively improving the activity and density of hepatocyte culture. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0044] Figure 2 This is a schematic diagram of the hardware architecture of the present invention;

[0045] Figure 3This is a detailed flowchart of the control method of the present invention;

[0046] Figure 4 This is a schematic diagram of the system expansion architecture;

[0047] Figure 5 The flowchart is an extension of the control method of the present invention. Detailed Implementation

[0048] To facilitate understanding of the present invention, a more comprehensive description will be given below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0049] Example:

[0050] like Figure 1 As shown, a machine learning-based multi-parameter adaptive control system for a bioreactor includes:

[0051] The bioreactor hepatocyte culture unit includes a bioreactor, a parameter detection module, and an actuator.

[0052] Data acquisition and control unit, including lower-level machine and human-machine interface;

[0053] The machine learning intelligent decision-making unit includes a data preprocessing module, a cell state prediction module, a multi-objective optimization decision-making module, and an online learning module.

[0054] The hardware structure of the hepatocyte culture unit in the bioreactor is as follows: Figure 2 As shown, the bioreactor uses a double-arm cell culture flask. A gas supply module consisting of a gas cylinder, solenoid valve, and flow meter supplies four streams of mixed gas, which mixes with the culture medium flowing out from the inlet pump via an oxygenator, flows into one arm, and then flows out through the outlet pump. It then passes through a parameter detection module comprised of a detection tank and flows back to the storage bottle, forming a perfusion loop. The detection tank includes at least the following sensors: pH sensor, DO sensor, temperature sensor, glucose-lactate sensor, or online biochemical analyzer. The bioreactor is stirred by a magnetic stirrer and heated in a constant temperature chamber. Samples can be taken out via a sampling pump for image analysis and offline detection. Alkali solution can be introduced via a replenishment pump to adjust the pH. All these components send operational data to a PLC, which displays the data on the human-machine interface and controls the system.

[0055] The detailed flowchart of the control method of the present invention is shown below. Figure 3 As shown, it includes the following steps:

[0056] (1) Input the target cell status, operating conditions and constraints into the human-computer interaction interface, and send the command to the lower computer to execute each module of the bioreactor hepatocyte culture unit;

[0057] (2) The data acquisition module acquires the time-series data of the multi-source sensors in the bioreactor in real time, and the data is collected by the lower-level machine and sent to the machine learning intelligent decision-making module;

[0058] (3) Send the time series data to the data preprocessing module for cleaning, alignment, standardization and feature extraction to obtain feature vectors;

[0059] (4) Input the feature vector into the trained cell state prediction model to obtain the predicted cell state value at future time.

[0060] (5) With the goal of optimizing the predicted value of cell state, the multi-objective optimization decision module uses the optimization algorithm to solve the optimal combination of operation parameter settings based on the predicted value and the set target.

[0061] (6) The online learning module uses new data to continuously update the cell state prediction module and the multi-objective optimization decision-making module;

[0062] (7) Return the set value to the lower-level machine, so that the lower-level machine can output instructions to adjust the actuator and realize closed-loop control.

[0063] Specifically, data preprocessing includes: data cleaning, standardization, and feature extraction.

[0064] Data cleaning mainly includes outlier detection; this embodiment uses the Isolation Forest algorithm. Let the sensor data sequence be... ,in This represents the parameters of a d-dimensional process.

[0065] Outliers are detected by constructing a randomized split tree. For each tree:

[0066] Randomly select a feature and split value The segmentation criteria are: ;

[0067] The abnormal values ​​are then categorized as follows:

[0068] ;

[0069] in, For the sample Path length for tree species; The expected path length in all trees; Let be the average path length of a binary search tree for n samples. ; For harmonic numbers, , is Euler's constant.

[0070] Data standardization uses Z-score standardization. Let the standardization matrix be... Then the j-th component of the i-th sample after standardization is:

[0071] ;

[0072] in, Let j be the average value of the j-th variable. Let j be the standard deviation of the j-th variable. .

[0073] For sliding window size , construct features:

[0074] Moving average: ;

[0075] Sliding trend: ;

[0076] Difference characteristics: .

[0077] The cell state prediction module employs the GRU algorithm to reduce model complexity and accelerate training. The GRU unit is represented as follows:

[0078] Reset Door: ;

[0079] Update Gate: ;

[0080] Candidate hidden state: ;

[0081] Final hidden state: ;

[0082] in, Enter information for the current moment; This is the hidden state from the previous moment; Represented as function, ; It is the hyperbolic tangent function. ; This indicates element-wise multiplication.

[0083] Let the input sequence be , The output is the cell state (cell viability, cell density), denoted as The state prediction output at time t is obtained by connecting a fully connected layer after the GRU unit:

[0084] ,

[0085] Multi-objective optimization decision module selection uses Model Predictive Control (MPC) method. Define the system state vector:

[0086] ,

[0087] in, For cell activity, pH value Dissolved oxygen concentration, This refers to the glucose concentration. This represents the lactic acid concentration.

[0088] Define the control input vector:

[0089] ,

[0090] in, For gas flow rate, The rate of alkali addition. For the culture medium perfusion rate, This represents the stirring rate.

[0091] Use a pre-trained GRU model as the predictor:

[0092] ,

[0093] Where m is the historical window size and N is the prediction time domain.

[0094] The objective function for MPC optimization is:

[0095] ,

[0096] Where Q, R, S are weight matrices. For reference only. To control the time domain, , To predict the time domain, For the input vector, Let be the predicted value at step i.

[0097] The constraints are:

[0098] ,

[0099] ,

[0100] For key biological parameters, soft constraints based on historical data are introduced:

[0101] ,

[0102] in, , For the penalty function, The minimum value of the input vector. The maximum value of the input vector. The minimum input vector difference, The maximum input vector difference, To predict cell viability, For maximum cell viability, This represents the minimum lactic acid concentration. To predict lactic acid concentration.

[0103] Therefore, the overall optimization problem is:

[0104] ,

[0105] The solution is obtained using the Sequential Quadratic Programming (SQP) method.

[0106] The online learning module is based on newly collected data. Update the GRU model:

[0107] ,

[0108] in, For loss function, For learning rate, Let this be the state vector of the new system. To predict the system vector, The original weights.

[0109] In addition, the weight matrix is ​​dynamically adjusted based on control performance:

[0110] ,

[0111] in, Let be the tracking error for the i-th state. For adaptive coefficients, This is the weight matrix.

[0112] As a high-level embodiment of the present invention, its system extension architecture diagram is as follows: Figure 4 As shown. The core prediction model of the machine learning intelligent decision-making unit is constructed as follows:

[0113] Three types of data: process parameters Image features Molecular marker characteristics Each part is preprocessed, and the preprocessing method is as described above.

[0114] The three feature vectors are concatenated into a single multimodal feature vector:

[0115] , Let p be the eigenvector. Let i be the feature vector. Let m be the eigenvectors.

[0116] The multimodal feature sequences are then input into the GRU layer. Let the output of the GRU layer be... An attention mechanism is introduced to calculate the weights at each time step:

[0117] ,

[0118] ,

[0119] ,

[0120] Among them, v, and These are learnable parameters. Let be the tracking error at time step t. It is a context vector. In hidden state, Let be the normalized attention weights at time step t. For time steps.

[0121] Use context vectors and the hidden state of the last time step Multi-task learning mainly consists of two tasks:

[0122] Regression task predicts cell activity :

[0123] ,

[0124] Probability distribution of cell states (adherence, proliferation, functional maintenance, stress) predicted by a classification task:

[0125] ,

[0126] The content of other methods remains unchanged, and the control method extension process is as follows: Figure 5 As shown.

[0127] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, this should not be construed as a limitation of the present invention. Those skilled in the art should understand that any modifications and improvements to the present invention without departing from the concept of the present invention are within the scope of protection of the present invention.

Claims

1. A multi-parameter adaptive control system for a bioreactor based on machine learning, characterized in that, The system includes: The bioreactor hepatocyte culture unit includes a bioreactor, a parameter detection module, and an actuator. The data acquisition and control unit, including the lower-level machine and the human-machine interface, is used to acquire sensor data and control the actuators; The machine learning intelligent decision-making unit includes a data preprocessing module, a cell state prediction module, a multi-objective optimization decision-making module, and an online learning module. The prediction model is a recurrent neural network architecture, and the decision-making module adopts a model prediction control framework. Closed-loop control mechanism: The system predicts cell state through real-time time-series data, dynamically optimizes operating parameters, and feeds them back to the actuator.

2. The system according to claim 1, characterized in that: The parameter detection module monitors pH value, dissolved oxygen (DO), temperature, glucose concentration and lactic acid concentration in real time; The actuator includes: a gas supply module (nitrogen / oxygen / air / carbon dioxide valve), a sampling module (sampling pump), a perfusion module (peristaltic pump), a stirring module (magnetic stirrer), and a heating module (temperature controller).

3. The system according to claim 2, characterized in that: The gas supply module regulates the flow rate of the mixed gas through a mass flow controller; the sampling module acquires data on cell morphology characteristics and molecular markers (urea, albumin).

4. The method according to claim 1, characterized in that: The data preprocessing module receives real-time time-series data and performs cleaning, alignment, standardization, and feature extraction based on data characteristics. It employs isolated forest outlier detection and Z-score standardization to extract moving average, moving trend, and difference features.

5. The system according to claim 1, characterized in that: The cell state prediction module adopts a gated recurrent unit (GRU) or Transformer architecture; the input data is multimodal time-series data, including real-time sensor data and offline detection features.

6. The system according to claim 1, characterized in that: The multi-objective optimization decision module uses cell activity and density as optimization objectives to solve for the optimal combination of operating parameters; the optimization objective function is: ， Where Q, R, S are weight matrices. For reference only. To control the time domain, , To predict the time domain, For the input vector, Let be the predicted value at step i; The constraints are: ， ， For key biological parameters, soft constraints based on historical data are introduced: ， in, , For the penalty function, The minimum value of the input vector. The maximum value of the input vector. The minimum input vector difference, The maximum input vector difference, To predict cell viability, For maximum cell viability, This represents the minimum lactic acid concentration. To predict lactic acid concentration.

7. The system according to claim 1, characterized in that: The system includes an online learning module that updates the weights of the prediction model based on new data. ， in, For loss function, For learning rate, Let this be the state vector of the new system. To predict the system vector, The original weights; The weight matrix is ​​adaptively adjusted based on the tracking error: ， in, Let be the tracking error for the i-th state. For adaptive coefficients, This is the weight matrix.

8. The system according to claim 1, characterized in that: The machine learning intelligent decision-making unit integrates a multimodal data fusion module, which converts the preprocessed process parameter feature vectors... Image feature vectors and molecular marker feature vectors Concatenate into feature vectors The multimodal feature sequences are then input into the cell state prediction module.

9. The method according to claim 1, characterized in that: The cell state prediction module introduces a multi-task learning framework, whose output includes a classification judgment of cell state and a prediction of continuous values ​​for at least one key cell indicator. An attention mechanism is used to calculate the weights for each time step. ， ， ， Among them, v, and These are learnable parameters. The tracking error at time step t, In hidden state, Let be the normalized attention weights at time step t. For time steps; Use context vectors and the hidden state of the last time step Multi-task learning mainly consists of two tasks: Regression task predicts cell activity : ， Probability distribution of cell states (adherence, proliferation, functional maintenance, stress) predicted by a classification task: .

10. A multi-parameter control method for bioreactors based on machine learning, characterized in that, The method includes: (1) Input the target cell status, operating conditions and constraints into the human-computer interaction interface, and send the command to the lower computer to execute each module of the bioreactor hepatocyte culture unit; (2) The data acquisition module acquires the time-series data of the multi-source sensors in the bioreactor in real time, and the data is collected by the lower-level machine and sent to the machine learning intelligent decision-making module; (3) Perform online preprocessing and feature extraction on the time series data to obtain feature vectors; (4) Input the feature vector into the trained cell state prediction model to obtain the predicted cell state value at future time. (5) With the goal of optimizing the predicted value of cell state, the multi-objective optimization decision module uses the optimization algorithm to solve the optimal combination of operation parameter settings based on the predicted value and the set target. (6) Return the set value to the lower-level machine, so that the lower-level machine can output instructions to adjust the actuator and realize closed-loop control.

Images