Virtual-real fusion method for digital twinning operation of cell drug factory in combination with Internet of Things

Abstract

The invention provides a virtual-real fusion method for digital twinning operation of a cell drug factory in combination with the Internet of Things, which comprises the following steps: constructing geometric models of a culture medium storage tank and a bioreactor through three-dimensional modeling software, mapping and associating historical process parameters of a culture medium supply system and temperature, pH value and dissolved oxygen concentration monitored in real time with the geometric models, and obtaining a virtual-real fusion model of the culture medium storage tank and the bioreactor; constructing a digital twinborn model of the reactor; analyzing historical production data through a machine learning algorithm, extracting a quantitative relationship among key parameters of a cell metabolism rate, a glucose concentration in a culture medium, a pH value and a reactor temperature, establishing a cell growth prediction model, and predicting a growth trend and a metabolism demand of cells in a target time period; a virtual-real mapping mechanism of a culture medium supply system is integrated with an intelligent manufacturing execution system, so that the culture medium supply process and the whole pharmaceutical production process are subjected to collaborative optimization control, and the working procedures of the culture medium production process are automatically dispatched according to a production plan and cell growth requirements.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a virtual-reality fusion method for digital twin operation of a cell drug factory combined with the Internet of Things. Background Art

[0002] In the culture medium supply system of the cell drug factory, the real-time fluctuation of the redox potential of the culture medium storage tank will affect cell metabolism. The dissolved oxygen fluctuation signal collected by the IoT sensor needs to be analyzed to predict and dynamically adjust the culture medium replacement cycle. However, in the actual production process, due to the complex composition of the culture medium, the interaction between the components is difficult to accurately grasp, resulting in unclear correlation between the dissolved oxygen fluctuation signal and cell metabolism. At the same time, there are differences in the growth state and metabolic level of cells in different batches, making it impossible for a fixed culture medium replacement cycle to meet the needs of dynamic adjustment. In addition, fluctuations in environmental factors during the culture process, such as temperature and pH value, will also affect the dissolved oxygen level, further increasing the difficulty of prediction. How to establish an accurate mapping relationship between dissolved oxygen fluctuation signals and cell metabolism in a complex production environment, and based on this, to achieve dynamic optimization of the culture medium replacement cycle, is a technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention provides a virtual-real fusion method for digital twin operation of a cell drug factory combined with the Internet of Things, which mainly includes: A dissolved oxygen sensor is used to monitor the fluctuation of dissolved oxygen concentration in the culture medium in the bioreactor in real time, and the time series data of dissolved oxygen concentration is obtained. The time series data is input into the long short-term memory neural network model for training and learning, and the change trend of dissolved oxygen concentration in a preset time period is predicted. Combined with the preset dissolved oxygen concentration threshold, it is determined whether the culture medium replacement cycle needs to be adjusted; If adjustment is needed, the cell density and cell viability during the cell growth process are collected in real time to generate a cell growth curve. The support vector machine algorithm is used to establish a correlation model between the cell growth curve and the culture medium replacement cycle, and the changes in the cell growth curve under different replacement cycle conditions are obtained to determine the target replacement cycle; The geometric model of the culture medium storage tank and bioreactor is constructed through 3D modeling software. The historical process parameters of the culture medium supply system and the real-time monitored temperature, pH value and dissolved oxygen concentration are mapped and associated with the geometric model to build a digital twin model of the reactor. Acquire the internal environmental parameters of the culture medium storage tank and bioreactor in real time, map and synchronize the internal environmental parameters with the virtual parameters in the digital twin model in real time, build a virtual-real fusion closed-loop control model for the culture medium supply system, and dynamically adjust the flow and concentration of the supply system according to cell growth requirements and metabolic status; Combined with the digital twin model, multi-scale and multi-physics simulation of the fusion culture medium tank and bioreactor during cell growth was carried out to simulate and analyze the impact of fluctuations in the redox potential and dissolved oxygen concentration of the culture medium on cell growth, and to change the ratio and supply of key nutrients in the culture medium; Analyze historical production data through machine learning algorithms, extract the quantitative relationship between key parameters such as cell metabolic rate, glucose concentration and pH value in the culture medium, and reactor temperature, establish a cell growth prediction model, and predict the growth trend and metabolic needs of cells in the target time period; Integrate the virtual-reality mapping mechanism of the culture medium supply system with the intelligent manufacturing execution system to achieve coordinated optimization control of the culture medium supply process and the entire drug production process, and automatically schedule the culture medium production process according to the production plan and cell growth requirements.

[0004] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses a method for integrating virtual and real operations of a digital twin operation of a cell drug factory in combination with the Internet of Things. The dissolved oxygen concentration of the culture medium in the bioreactor is monitored in real time by a dissolved oxygen sensor, and its changing trend is predicted by a long short-term memory neural network. In combination with cell density and vitality data, a correlation model between the cell growth curve and the culture medium replacement cycle is established. The present invention constructs a digital twin model of the culture medium storage tank and the bioreactor, and realizes closed-loop control of virtual and real integration. The influence of culture medium parameters on cell growth is analyzed by multi-scale and multi-physical field simulation, and the cell growth trend and metabolic demand are predicted by a machine learning algorithm, thereby realizing intelligent regulation of the culture medium supply process. The present invention integrates the culture medium supply system with the intelligent manufacturing execution system, realizes collaborative optimization control of the entire process, and improves the efficiency and quality of biopharmaceutical production. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] Figure 1 This is a flow chart of the virtual-reality fusion method of the digital twin operation of a cell drug factory combined with the Internet of Things of the present invention. DETAILED DESCRIPTION

[0006] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0007] like Figure 1 In this embodiment, a virtual-reality fusion method for digital twin operation of a cell drug factory combined with the Internet of Things may specifically include: Step S101, using a dissolved oxygen sensor to monitor the fluctuation of the dissolved oxygen concentration of the culture medium in the bioreactor in real time, obtaining the time series data of the dissolved oxygen concentration, inputting the time series data into the long short-term memory neural network model for training and learning, predicting the change trend of the dissolved oxygen concentration in a preset time period, and combining with the preset dissolved oxygen concentration threshold to determine whether it is necessary to adjust the culture medium replacement cycle.

[0008] The dissolved oxygen concentration value collected by the dissolved oxygen sensor of the bioreactor is obtained, and a dissolved oxygen concentration time series data set is obtained according to the sampling time mark; the dissolved oxygen concentration time series data set is standardized by a normalization method, and a standardized dissolved oxygen concentration time series data set is obtained by removing outliers and supplementing missing values; the mean, variance, and peak statistical features of the dissolved oxygen concentration are extracted according to the standardized dissolved oxygen concentration time series data set, and the statistical features are input into a long short-term memory neural network for training to obtain a dissolved oxygen concentration prediction model; the dissolved oxygen concentration in a fixed time period in the future is predicted by the dissolved oxygen concentration prediction model, and if the predicted dissolved oxygen concentration is lower than a preset concentration threshold and the decline rate exceeds a preset rate threshold, it is marked as a warning time point for culture medium replacement.

[0009] Specifically, the dissolved oxygen data of the culture medium in the bioreactor is collected by the dissolved oxygen sensor, and the dissolved oxygen concentration value is obtained according to the fixed sampling time interval, and the dissolved oxygen concentration time series data set is formed according to the sampling time mark. The dissolved oxygen concentration time series data set is cleaned, the outliers are removed and the missing values ​​are supplemented. The dissolved oxygen concentration value is standardized by the normalization method to obtain the standardized dissolved oxygen concentration time series data set. The statistical features such as the mean, variance, and peak value of the dissolved oxygen concentration are extracted from the standardized dissolved oxygen concentration time series data set, and the training sample matrix is ​​constructed. The training is input into the long short-term memory neural network for training to obtain the dissolved oxygen concentration prediction model. The dissolved oxygen concentration in the future fixed time period is predicted by the dissolved oxygen concentration prediction model to generate a predicted dissolved oxygen concentration time series data set. The dissolved oxygen concentration decrease rate is calculated according to the predicted dissolved oxygen concentration time series data set. When the decrease rate exceeds the preset rate threshold and the predicted dissolved oxygen concentration is lower than the preset concentration threshold, it is marked as the warning time point for culture medium replacement. The historical culture medium replacement time record is extracted from the dissolved oxygen monitoring database, and the average replacement cycle is calculated in combination with the warning time point to generate the next culture medium replacement time. In the process of cell culture in bioreactor, dissolved oxygen concentration is a key parameter affecting the cell growth state, and the change of dissolved oxygen concentration reflects the intensity of cell metabolic activity. When the dissolved oxygen concentration in the culture medium continues to decrease, it indicates that the cell metabolism is vigorous and the oxygen consumption rate increases. It also indicates that the nutrients in the culture medium are gradually consumed. In practical applications, the dissolved oxygen sensor collects dissolved oxygen concentration values ​​every 10 minutes, continuously monitors for 48 hours, and obtains dissolved oxygen concentration data at 288 time points. The normal range of dissolved oxygen concentration fluctuates between 35% and 70%. During the data cleaning process, outliers below 20% or above 85% are eliminated, and the missing data are supplemented by linear interpolation to ensure the continuity of the data. During the standardization process, the mean of dissolved oxygen concentration in the historical data is selected as the benchmark value, with a variance of 15%, and the collected dissolved oxygen concentration data is converted into a standard normal distribution. The extracted statistical features include the arithmetic mean, geometric mean, standard deviation, kurtosis and skewness of dissolved oxygen concentration every 4 hours. These features can fully reflect the change law of dissolved oxygen concentration. The long short-term memory neural network model adopts a three-layer network structure. The input layer contains 12 neurons, corresponding to the dissolved oxygen concentration data of the previous 12 time points, the hidden layer contains 8 neurons, and the output layer predicts the dissolved oxygen concentration values ​​at the next 4 time points. The model training uses 120 sets of historical data, of which 80% are used for training and 20% are used for verification. The prediction results show that when the cell growth enters the logarithmic phase, the dissolved oxygen concentration shows a significant downward trend, with an average decrease of 2% to 3% per hour. The warning threshold is set at 35% dissolved oxygen concentration, and the decline rate threshold is 3% per hour. When the predicted dissolved oxygen concentration is lower than 35% in the next 6 hours and the decline rate exceeds 3% per hour, the medium replacement warning is triggered.By analyzing historical culture data, it was found that the average culture medium replacement cycle during cell culture is 72 hours, but the actual replacement time will fluctuate with factors such as cell density and metabolic activity. During the period of vigorous cell growth, the replacement cycle is shortened to 54 hours, while during the growth plateau period, the replacement cycle can be extended to 96 hours. Combining the warning time point with the historical replacement records, it is calculated that the next replacement time is generally between 12 hours and 24 hours after the warning time point.

[0010] Step S102, if adjustment is needed, the cell density and cell viability during the cell growth process are collected in real time to generate a cell growth curve, and a support vector machine algorithm is used to establish a correlation model between the cell growth curve and the culture medium replacement cycle to obtain the change of the cell growth curve under different replacement cycle conditions and determine the target replacement cycle.

[0011] The cell density value and fluorescence intensity signal in the culture medium are collected by a cell density detector, and the original cell growth data set is obtained according to the fixed time interval records; outlier detection is performed according to the original cell growth data set, and the missing data is supplemented by the cubic spline interpolation method to obtain the processed cell growth data set; the cell density change rate and the cell viability change rate are calculated for the processed cell growth data set, and the cell growth feature vector is constructed in combination with the cell morphological parameters; the historical culture medium replacement time points are extracted from the culture medium replacement record database, and the corresponding relationship between the cell growth feature vector and the replacement time point is established, and the relationship is input into the support vector machine algorithm to obtain the culture medium replacement cycle model, and the cell density integral value and the cell viability average value at different replacement time intervals are predicted according to the culture medium replacement cycle model; it is determined whether the cell density integral value and the cell viability average value are within the preset threshold interval, and if they are within the preset threshold interval, the time point with the largest cell density integral value and the highest cell viability average value is selected as the target culture medium replacement cycle.

[0012] Specifically, the cell density value in the culture medium is collected in real time by a cell density detector, and the fluorescence intensity signal is collected synchronously to calculate the cell viability value, and the original cell growth data set is recorded at fixed time intervals. The original cell growth data set is detected for outliers, and the cell density values ​​and cell viability values ​​beyond the physiological range are eliminated. The missing data are supplemented by the cubic spline interpolation method to obtain the processed cell growth data set. According to the processed cell growth data set, characteristic parameters such as the cell density change rate, cell viability change rate, and density-viability ratio are calculated, and the cell growth characteristic vector is constructed in combination with the cell morphological parameters at the corresponding time points. The historical culture medium replacement time points are extracted from the culture medium replacement record database, and the corresponding relationship between the cell growth characteristic vector and the replacement time point is established, which is input into the support vector machine algorithm to train the culture medium replacement cycle model. The culture medium replacement cycle model is used to predict the cell density curve at different replacement time intervals, and the cell density integral value and cell viability average value corresponding to each time interval are calculated. According to the preset cell density threshold interval and cell viability threshold interval, the qualified culture medium replacement cycle values ​​are screened out, and the time point with the largest cell density integral value and the highest cell viability average value is selected as the target culture medium replacement cycle. During cell culture, cell density and cell viability are important indicators reflecting the cell growth status. The number of cells in the culture medium is monitored in real time by a cell density detector, and data is collected every 30 minutes. The measurement range of cell density is between 1×10^5 and 1×10^7 cells / ml. At the same time, fluorescent dyes are used to mark living cells, and cell viability is calculated by detecting fluorescence intensity. Fluorescence intensity is positively correlated with cell metabolic activity. In practical applications, due to failure of detection equipment or external interference, abnormal values ​​often appear in the original data, such as cell density suddenly drops to 0 or surges to 1×10^8 cells / ml. These data that are obviously beyond the physiological range of cells need to be eliminated. For the time points where data are missing, the cubic spline interpolation method is used to supplement. This method can maintain the continuity and smoothness of the data curve and avoid drastic fluctuations. Cell growth characteristic parameters contain multiple dimensions. The cell density change rate reflects the proliferation rate of the cell population. Under normal circumstances, the density change rate in the logarithmic growth period is between 0.15 and 0.25. The cell viability change rate indicates the changing trend of cell metabolic activity. A decrease in viability of more than 20% usually indicates an adverse change in the culture environment. The density-viability ratio comprehensively reflects the metabolic level of a single cell. A value between 2 and 5 indicates that the cell is growing well. Historical culture data show that different culture medium replacement cycles can lead to significant differences in cell growth curves. If the replacement cycle is too short, the cells fail to fully utilize the nutrients in the culture medium, resulting in a waste of resources; if the replacement cycle is too long, cell growth is inhibited by the accumulation of metabolic waste, resulting in a decrease in cell viability.By analyzing historical data, it was found that replacing the culture medium when the cell density reached 5×10^6 cells / ml and the cell viability remained above 85% could achieve a higher cell proliferation efficiency. The support vector machine algorithm established a mapping relationship between cell growth characteristics and the optimal replacement time point by learning the rules in historical culture data. In the model verification stage, simulation predictions were performed on different replacement cycles, with multiple candidate replacement time points set within the range of 48 hours to 96 hours at intervals of 24 hours. By calculating the cell density integral value and average viability value corresponding to each time point, the total cell yield and quality indicators were comprehensively evaluated. The selected target replacement cycle should ensure that the cell density integral value reaches the expected level while maintaining a high average cell viability.

[0013] Step S103, constructing the geometric model of the culture medium storage tank and the bioreactor through 3D modeling software, mapping and associating the historical process parameters of the culture medium supply system and the real-time monitored temperature, pH value and dissolved oxygen concentration with the geometric model, and constructing a digital twin model of the reactor.

[0014] A three-dimensional structural model of the storage tank and the bioreactor is constructed using three-dimensional modeling software, wherein the three-dimensional structural model includes geometric data of a feed port, a discharge port, a stirring device, and a sensor installation position; parameter monitoring points are established according to the surface of the three-dimensional structural model, culture medium supply system data is obtained from a process parameter database, and real-time monitoring data of the sensor installation position is collected; the fluid force field, temperature field, and concentration field are divided according to the parameter monitoring points, and the distribution values ​​of the fluid force field, temperature field, and concentration field in space are calculated using a finite element algorithm; a deep neural network algorithm is used to train the coupling relationship between the fluid force field, temperature field, and concentration field, and a reactor digital mapping structure is established according to the coupling relationship to obtain a digital twin model.

[0015] Specifically, according to the physical size parameters of the culture medium storage tank and the bioreactor, a three-dimensional structural model of the storage tank and the reactor is constructed using three-dimensional modeling software, including the geometric data of the feed port, the discharge port, the stirring device, and the sensor installation position, to generate the first structural model. The flow, pressure, and stirring rate values ​​of the culture medium supply system are extracted from the process parameter database, the values ​​are normalized, and the real-time monitoring data of the temperature sensor, pH sensor, and dissolved oxygen sensor in the bioreactor are collected to generate a process parameter data set. According to the numerical distribution in the process parameter data set, parameter monitoring points are established on the surface of the first structural model, and the distribution law of the parameters in space is calculated using the interpolation algorithm to generate the second parameter structural model. In the second parameter structural model, the fluid force field, temperature field, and concentration field are divided, and a group of field quantity calculation equations are established. The distribution law of the field quantity in space is calculated by the finite element algorithm to generate the third physical field model. The deep neural network algorithm is used to train the coupling relationship between the fluid force field, temperature field, and concentration field, and the characteristic vector including the field quantity value, field quantity gradient, and inter-field force is input, and the evolution law under the interaction of field quantities is output. According to the law of field quantity coupling evolution, the reactor digital mapping structure is established on the basis of the third physical field model to realize the dynamic association of process parameters, monitoring data, and physical field distribution, and form a digital twin model. The digital twin modeling of bioreactors involves the coupling analysis of multiple physical fields, among which the reactor structure directly affects the flow characteristics and mass transfer efficiency of the internal culture medium. Taking the stirred bioreactor as an example, the reactor is 300 mm high and 200 mm in diameter. It is equipped with a four-blade propeller stirring device. The diameter of the stirring paddle is 1 / 3 of the inner diameter of the container, and the height from the bottom is 1 / 4 of the height of the container. The feed inlet is located at the top center, the discharge port is located at the bottom side wall, and the dissolved oxygen sensor, pH sensor and temperature sensor are evenly arranged along the reactor height. The acquisition frequency of process parameters is set according to the parameter change characteristics. The temperature and pH value are collected every 10 minutes, the dissolved oxygen concentration is collected every 5 minutes, and the stirring speed and flow rate are collected in real time. The original data was normalized, and the temperature range of 35 to 38 degrees Celsius was mapped to the interval of 0 to 1, the pH value of 6.8 to 7.2 was mapped to the interval of 0 to 1, and the dissolved oxygen concentration of 30% to 70% was mapped to the interval of 0 to 1. A grid of monitoring points was arranged on the surface of the reactor structure. The grid density changed with the parameter gradient, and the grid was encrypted in the area where the parameter changed dramatically. Taking the distribution of dissolved oxygen as an example, the spacing between monitoring points was set to 10 mm in the stirring area and near the gas-liquid interface, and the spacing in other areas was set to 20 mm. The radial basis function interpolation algorithm was used to calculate the parameter distribution between grid points to achieve continuous distribution of parameters in space. The division of the physical field is based on fluid mechanics and heat and mass transfer theory. The fluid force field includes pressure field and velocity field. The pressure field reflects the pressure distribution of the culture medium under gravity and stirring, and the velocity field describes the flow characteristics of the culture medium. The temperature field is affected by the heating device and the ambient temperature, and a temperature gradient is formed near the reactor wall.The dissolved oxygen concentration field is affected by both bubble distribution and cell oxygen consumption. Deep neural networks are used to establish coupling relationships between physical fields. Input features include local field values, field spatial gradients, and inter-field forces. Taking the coupling of temperature field and dissolved oxygen field as an example, an increase in temperature will reduce the dissolved oxygen content of the culture medium. For every 1 degree Celsius increase, the dissolved oxygen concentration decreases by about 2%. At the same time, the density difference caused by temperature changes will affect the natural convection of the culture medium, thereby changing the mass transfer efficiency of dissolved oxygen. The digital twin model reflects the internal state of the reactor by updating the physical field distribution in real time. When the stirring speed increases from 200 rpm to 400 rpm, the model can predict the dynamic process of the dissolved oxygen concentration increasing from 45% to 65% within 15 minutes, while showing slight fluctuations in the temperature field and the stability of the pH value. This multi-physical field coupled digital mapping can accurately reflect the dynamic changes in the culture environment.

[0016] Step S104, acquire the internal environmental parameters of the culture medium storage tank and bioreactor in real time, map and synchronize the internal environmental parameters with the virtual parameters in the digital twin model in real time, build a virtual-reality fusion closed-loop control model of the culture medium supply system, and dynamically adjust the flow and concentration of the supply system according to cell growth requirements and metabolic status.

[0017] The internal environmental parameters of the culture medium storage tank and the bioreactor are collected by temperature sensors, pH sensors, and dissolved oxygen sensors, and the sliding window method is used to reduce the noise of the environmental parameters to obtain an environmental parameter data set; the glucose concentration and lactate concentration in the cell culture process are collected according to the environmental parameter data set, and the cell metabolism state data set is obtained by combining the cell density data; the sensor collection position is calibrated in the digital twin structure using the cell metabolism state data set, and the parameter space distribution is calculated by the Kriging interpolation algorithm to obtain a virtual parameter distribution data set; the root mean square error value is calculated based on the virtual parameter distribution data set and the environmental parameter data set, and the deep reinforcement learning algorithm is used to train the culture medium supply controller to obtain the culture medium flow compensation value and concentration compensation value.

[0018] Specifically, the internal environmental parameters of the culture medium tank and bioreactor are collected through temperature sensors, pH sensors, dissolved oxygen sensors, liquid level sensors, and pressure sensors. The sliding window method is used to reduce noise and process outliers on the collected data to generate an environmental parameter data set. The glucose concentration, lactate concentration, and amino acid concentration during the cell culture process are collected, and the metabolite concentration change trend is calculated. Combined with the cell density data and cell viability data, a cell metabolic state data set is generated. The sensor acquisition position is calibrated in the digital twin structure, and the corresponding relationship between the actual environmental parameters and the virtual environmental parameters is established. The parameter space distribution is calculated using the Kriging interpolation algorithm to generate a virtual parameter distribution data set. The root mean square error between the virtual parameter distribution data set and the environmental parameter data set is calculated, and an error compensation function is established to correct the virtual parameters and generate a corrected virtual parameter distribution data set. The deep reinforcement learning algorithm is used to train the culture medium supply controller, and the cell metabolic state data set and the corrected virtual parameter distribution data set are input, and the culture medium flow compensation value and concentration compensation value are output. The speed of the culture medium supply pump is adjusted according to the compensation value, and the mixing ratio of nutrients in the culture medium preparation unit is adjusted synchronously. The actual supply parameters are collected and the virtual parameter distribution data set is continuously updated. The virtual-real fusion control of the culture medium supply system involves real-time monitoring and regulation of multiple environmental parameters. In practical applications, the measurement range of the temperature sensor is 30 to 40 degrees Celsius with an accuracy of 0.1 degrees Celsius; the measurement range of the pH sensor is 6.0 to 8.0 with an accuracy of 0.01; the measurement range of the dissolved oxygen sensor is 0 to 100% with an accuracy of 0.1%. A 60-second sliding time window is used to process the raw data to remove abnormal values ​​beyond the normal range. Cell metabolic state monitoring focuses on glucose consumption rate and metabolite accumulation. The glucose concentration gradually decreases from the initial 4.5 g / L, and a supplement signal is triggered when it drops to 1.0 g / L. Lactic acid, as the main metabolite, gradually increases in concentration from 0, and inhibits cell growth when it exceeds 2.0 g / L. Amino acid monitoring includes changes in the concentrations of 20 amino acids such as glutamine and alanine. 48 sensor collection positions are set in the digital twin structure to form a spatial monitoring grid in the reactor. The Kriging interpolation algorithm predicts the parameter distribution of unknown points based on the measured values ​​of known points, taking into account spatial correlation and anisotropic characteristics. Taking the dissolved oxygen distribution as an example, the spatial correlation distance in the stirring area is 20 mm, while the correlation distance in the static area is 50 mm. There is an error between the virtual parameters and the actual parameters. The root mean square error of the temperature field is usually within 0.5 degrees Celsius, the error of the pH field is within 0.05 units, and the error of the dissolved oxygen field is within 2%. By establishing an error compensation function, the virtual parameters are dynamically corrected according to the statistical laws of historical data. The culture medium supply controller is trained using a deep reinforcement learning algorithm, and the reward function is set to include multiple indicators such as cell density growth rate and metabolite concentration change rate.The controller outputs a control instruction every 5 minutes. The adjustment range of the culture medium flow is 50 to 200 ml / hour, and the concentration adjustment range is 0.5 to 2 times the basic concentration. The culture medium preparation unit contains multiple peristaltic pumps and mixers to adjust the ratio of each nutrient according to the metabolic needs of the cells. The glucose storage concentration is 200 g / L, and the amino acid storage concentration is 50 times the basic culture medium concentration. The on-demand replenishment of materials is achieved by precisely controlling the valve opening. The real-time feedback of the supply parameters is synchronized with the update frequency of the virtual parameters to ensure that the digital twin model can accurately reflect the dynamic changes of the actual culture environment. During the culture process, the cell density doubles every 24 hours, and the metabolite concentration fluctuates periodically. The culture medium supply controller maintains a steady-state environment for cell growth through predictive regulation.

[0019] Step S105, combining the digital twin model to perform multi-scale multi-physics field simulation of the fusion culture medium storage tank and bioreactor during cell growth, simulate and analyze the impact of fluctuations in the redox potential and dissolved oxygen concentration of the culture medium on cell growth, and change the ratio and supply of key nutrients in the culture medium.

[0020] The geometric structure data of the storage tank and the reactor in the digital twin structure are obtained, and the Navier-Stokes equations describing the fluid motion and the convection-diffusion equations describing the mass transfer process are established according to the geometric structure data; a multi-scale grid structure is established by using an adaptive grid division method according to the Navier-Stokes equations and the convection-diffusion equations, and multi-physical field distribution data are obtained through the multi-scale grid structure; a cell-scale calculation module is constructed according to the multi-physical field distribution data by using a discrete element algorithm, and the redox potential distribution and dissolved oxygen concentration gradient data in the microenvironment are obtained through the cell-scale calculation module; a cell metabolic flux balance equation group is established according to the dissolved oxygen concentration gradient data, and the glucose, amino acid, fatty acid consumption rate and metabolite generation rate data are calculated through the cell metabolic flux balance equation group.

[0021] Specifically, the geometric structure data of the tank and reactor are extracted from the digital twin structure. According to the spatial distribution law of the potential field, flow field, temperature field, and concentration field, the Navier-Stokes equations describing fluid motion are established at the macroscopic scale, and the convection-diffusion equations describing the mass transfer process are established at the mesoscopic scale. The adaptive meshing method is used to encrypt the mesh in the area where the fluid moves violently and sparse the mesh in the area where the fluid is stationary. A multi-scale mesh structure is established, and the finite element method is used to solve the multi-physics field control equations. The discrete element algorithm is used to construct the cell-scale calculation module, divide the cell membrane, cytoplasm, and nucleus into three calculation domains, and calculate the redox potential distribution and dissolved oxygen concentration gradient in the microenvironment around the cell. According to the cell-scale calculation results, a group of cell metabolic flux balance equations is established to calculate the consumption rate of glucose, amino acids, fatty acids and other substances and the rate of metabolite generation. A metabolic flux predictor is established using a deep neural network algorithm, and characteristic vectors such as cell density, metabolite concentration, and environmental parameters are input to predict cell growth curves under different culture conditions. According to the prediction results of metabolic flux, the matching relationship between the consumption rate and the replenishment rate of nutrients in the culture medium is calculated, a new culture medium formula is generated, and the formula parameters are continuously optimized through the feedback regulation mechanism. The multi-scale multi-physics simulation of the cell culture process involves cross-scale calculations from the molecular level to the reactor scale. At the macro level, the fluid movement in the bioreactor is mainly driven by the stirring paddle, forming a complex three-dimensional flow field structure. When the stirring speed is 100 rpm, the flow velocity in the center area of ​​the reactor reaches 0.5 m / s, while the flow velocity in the edge area drops below 0.1 m / s. This flow velocity difference directly affects the mass transfer efficiency of dissolved oxygen. When meshing, a finer mesh is used for areas with larger flow velocity gradients, and the mesh size is dynamically adjusted between 0.5 mm and 5 mm. The mesh density near the stirring paddle is the highest, containing 1000 mesh cells per unit volume, while the mesh density in the static area is reduced to 100 cells / cubic centimeter. This adaptive mesh can accurately capture the detailed characteristics of the flow field. Cell-scale calculations focus on the microenvironment around individual cells. The cell membrane is about 10 nanometers thick and constitutes the main barrier to material transport. The cytoplasm has a volume of about 2000 cubic microns and is the main site of metabolic reactions. In the microenvironment, the redox potential drops from 150 millivolts outside the cell to -50 millivolts inside the cell, forming a transmembrane potential difference, which affects the metabolic activity of the cell. Metabolic flux analysis shows that a single cell consumes 1 picomoles of glucose per hour, produces 2 picomoles of lactic acid and 0.5 picomoles of carbon dioxide. The consumption patterns of amino acids show significant differences, with glutamine having the highest consumption rate of 0.8 picomoles / hour, while tryptophan has a consumption rate of only 0.05 picomoles / hour. This difference reflects the complexity of the cellular metabolic network. The deep neural network adopts a three-layer structure, and the input layer contains 48 nodes, corresponding to cell density, concentrations of 12 key metabolites, and various environmental parameters.The hidden layer processes these features through 1024 neurons, and the output layer predicts the cell growth curve within the next 24 hours. The training data set contains 1000 sets of historical culture data, and the validation set accounts for 20%. The culture medium formula optimization is based on the dynamic balance of metabolic needs. When the cell density reaches 2×10^6 cells / mL, the glucose consumption rate increases to 3 times the initial value, while the glutamine consumption rate increases to 4 times. According to this trend of change, the concentration of glucose in the culture medium is increased from 4.5 g / L to 6.0 g / L, and glutamine is increased from 2.0 mmol / L to 3.5 mmol / L, while the content of other essential amino acids is appropriately increased. Through this dynamic regulation, the specific growth rate of the cells is maintained between 0.028 and 0.032 / hour, and metabolic homeostasis is well maintained.

[0022] Step S106, analyze historical production data through machine learning algorithms, extract the quantitative relationship between key parameters such as cell metabolic rate, glucose concentration and pH value in the culture medium, and reactor temperature, establish a cell growth prediction model, and predict the growth trend and metabolic needs of cells in the target time period.

[0023] Acquire culture medium parameter data from a historical database, process the culture medium parameter data by a maximum and minimum value normalization method to obtain a standardized culture parameter data set; based on the standardized culture parameter data set, use a wavelet transform denoising algorithm and a preset decomposition depth to perform signal reconstruction on the standardized culture parameter data set to obtain a smoothed data set; for the smoothed data set, calculate eigenvalues ​​and eigenvectors by a principal component analysis method, and if the cumulative contribution rate of the eigenvalues ​​reaches a preset threshold, obtain a reduced dimension feature data set; use the reduced dimension feature data set to train a recursive neural network model, and if the prediction error of the validation set is less than a preset threshold, obtain a cell growth prediction model, and calculate the metabolite generation rate by the cell growth prediction model.

[0024] Specifically, the values ​​of glucose concentration, pH value, temperature, dissolved oxygen concentration, and cell density in the culture medium were extracted from the historical database, and the missing values ​​and outliers were eliminated. The data were normalized using the maximum and minimum value standardization method to generate a standardized culture parameter data set. The wavelet transform denoising algorithm was applied to the standardized culture parameter data set, the db4 wavelet basis function was selected, and the four-layer decomposition depth was set. After removing the high-frequency noise, the signal was reconstructed to obtain the culture parameter data set after smoothing. According to the culture parameter data set after smoothing, the Pearson correlation coefficient between the parameters was calculated, the parameter association matrix was constructed, the principal component analysis method was used to extract the eigenvalues ​​and eigenvectors, and the principal components whose cumulative contribution rate reached the preset threshold were selected to generate a reduced dimension feature data set. The recursive neural network algorithm was used to construct a cell growth prediction model, the number of neurons in the input layer was set to be the same as the dimension of the reduced dimension feature data set, the hidden layer used a long short-term memory unit structure, and the training samples were divided in chronological order. According to the preset training parameters, the recursive neural network was iteratively trained, the root mean square error was used as the loss function, and the network weights were updated by the back propagation algorithm until the prediction error on the validation set was less than the preset threshold. Using the trained cell growth prediction model, the current culture parameters are input, and the cell density change trend in the future time period is predicted by the autoregressive method, and the metabolite generation rate and nutrient consumption rate at each time point are calculated. The parameter monitoring data during the cell culture process reflects the dynamic changes in cell growth status and metabolic activity. In the original monitoring data, the measurement range of glucose concentration is 0 to 10 g / L, the pH range is 6.5 to 7.8, the temperature range is 35 to 39 degrees Celsius, and the dissolved oxygen concentration range is 20% to 95%. These parameters are difficult to compare directly under different dimensions and numerical ranges. Through the maximum and minimum value normalization processing, all parameters are mapped to the range of 0 to 1 to make the data comparable. During the data preprocessing process, it was found that sensor failure and external interference can cause mutations and anomalies in the measurement data. For example, the temperature sensor drifts after the sterilization process, and the reading suddenly increases by 2 degrees Celsius; the pH electrode jumps during calibration, and the pH value deviates from 0.5 units instantly. Wavelet transform was used to smooth these anomalies. The reason for choosing db4 wavelet basis function is that it has good local feature extraction ability. After four-layer decomposition, it can effectively remove high-frequency noise while retaining the main features of the signal. There is a significant correlation between the parameters. By calculating the Pearson correlation coefficient, it was found that the correlation coefficient between glucose consumption rate and cell density reached 0.85, indicating that the two are strongly positively correlated. The change in pH value is also closely related to the accumulation of metabolites, with a correlation coefficient of -0.72, reflecting that the acidic substances produced during cell metabolism will reduce the pH value of the culture medium. Principal component analysis shows that the cumulative contribution rate of the first three principal components reached 87%, which means that these three principal components can be used to represent the main information of the original data. The recursive neural network adopts a long short-term memory unit structure, which has the advantage of processing long sequence data.The input layer contains 12 neurons, corresponding to the values ​​of 4 culture parameters at 3 consecutive time points. The hidden layer is set with 64 LSTM units, which can capture the temporal characteristics of parameter changes. During the training process, a 24-hour sliding time window is used to predict parameter changes in the next 12 hours. When the root mean square error on the validation set drops below 0.15, the model training is considered to have reached convergence. The prediction results show that the model can accurately predict the different stages of cell growth. During the vigorous growth period, the cell density doubles every 24 hours, and the glucose consumption rate reaches 0.5 g / L / hour; after entering the plateau period, the growth rate decreases and glucose consumption is maintained at 0.2 g / L / hour. Based on these prediction results, the nutritional needs of cells can be calculated in advance. For example, when the cell density reaches 5×10^6 cells / ml, 0.3 mmol of glutamine is consumed per hour and 0.5 mmol of lactate is produced.

[0025] Step S107, integrating the virtual-reality mapping mechanism of the culture medium supply system with the intelligent manufacturing execution system, so as to coordinate and optimize the control of the culture medium supply process and the whole process of drug production, and automatically schedule the culture medium production process according to the production plan and cell growth requirements.

[0026] Collect production plan scheduling data and cell culture progress data, and generate a production resource status data set based on the scheduling data and progress data; obtain production equipment maintenance records and fault records, and use the equipment reliability evaluation index system to calculate the equipment status evaluation data set; based on the production resource status data set and the equipment status evaluation data set, use the integer programming algorithm to optimize and calculate to obtain the initial scheduling plan for culture medium production; for the initial scheduling plan for culture medium production, use the deep reinforcement learning algorithm to construct a culture medium production scheduler, and process the production plan priority and equipment status characteristics through the scheduler input layer to obtain a scheduling instruction sequence.

[0027] Specifically, the production schedule data and cell culture progress data are collected, combined with the liquid level information of the culture medium tank and the status information of the preparation equipment, and the relational database is used to record the occupation status of production equipment, raw materials, and personnel resources to generate a production resource status data set. The maintenance records, fault records, and repair records of the production equipment are collected, and an equipment reliability evaluation index system is established. The availability score of the equipment idle time period is calculated to generate an equipment status evaluation data set. According to the culture medium consumption prediction results in the virtual-real mapping database, combined with the production resource status data set and the equipment status evaluation data set, the integer programming algorithm is used to optimize the culture medium production batch arrangement and generate an initial scheduling plan. The deep reinforcement learning algorithm is used to construct the culture medium production scheduler. The input layer contains features such as production plan priority, equipment status, raw material inventory, and personnel scheduling, and the output layer generates a scheduling instruction sequence. The deep reinforcement learning algorithm is trained based on historical scheduling data, and the reward function is set to include indicators such as production plan completion rate, resource utilization, and scheduling stability. The scheduling strategy is optimized through the policy gradient method. According to the scheduling instruction sequence, production resources are allocated, production work orders are generated, and production process parameters are monitored. When equipment failure, raw material shortage, and quality abnormality are detected, the scheduling plan is triggered to be optimized online. Culture medium production scheduling involves the collaborative management of multiple resource dimensions. In the production resource status monitoring, the liquid level of the culture medium tank is collected every 5 minutes through the liquid level sensor, and the replenishment signal is triggered when the liquid level is lower than 30%. The status of the preparation equipment includes indicators such as operating time, maintenance interval, and fault frequency. For example, if the continuous operating time of the agitator exceeds 72 hours, preventive maintenance is required, and the cleaning equipment needs to be disinfected every 8 hours of operation. The equipment reliability assessment adopts a multi-layer indicator system, including equipment integrity rate, fault interval time, repair and recovery time, etc. Taking the sterilizer as an example, the monthly integrity rate is required to reach more than 95%, the average fault interval time is not less than 720 hours, and the single fault repair time is not more than 4 hours. In the evaluation of the equipment idle time period, the equipment status, maintenance plan, staffing and other factors are comprehensively considered to calculate the equipment availability score for each time period. The integer programming algorithm considers multiple constraints when optimizing the culture medium production batch. The time window constraint of the production equipment requires that the production time of each batch is not less than 4 hours, and the equipment cleaning and disinfection time is not less than 2 hours. The raw material inventory constraint ensures that the raw material supply is sufficient during the production process. For example, the inventory of glucose mother solution must meet the production needs for at least 3 days. Personnel scheduling constraints ensure that there are enough operators in each time period. The deep reinforcement learning algorithm handles complex scheduling scenarios. The input features contain 48 dimensions, covering information such as production plans for the next 7 days, equipment status predictions, and raw material inventory warnings. The action space of the scheduling instruction includes decision variables such as equipment allocation, batch scheduling, and personnel deployment. In the reward function design, the weight of the production plan completion rate is 0.4, the weight of resource utilization is 0.3, and the weight of scheduling stability is 0.3. In actual applications, production abnormalities require timely responses.When an abnormal increase in the temperature of the agitator bearing is detected, the algorithm will predict that the equipment may fail, adjust the production plan in advance, and transfer the originally planned tasks to the backup equipment. When the raw material supplier delays delivery, resulting in a decrease in inventory levels, the scheduling algorithm will re-plan the production batch size to prioritize the production of high-priority products. Quality monitoring of the culture medium production process runs through the entire process. Online monitoring of key parameters such as pH value, conductivity, and osmotic pressure, triggering an alarm when the osmotic pressure deviates from the target value by ±10%. The production work order contains information such as the raw material batch number, process parameters, quality inspection results, etc., to achieve traceability of the entire production process. Through multi-dimensional collaborative management, the culture medium production scheduling not only meets the production plan requirements, but also ensures the stability of the production process and the consistency of product quality.

[0028] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A virtual-real fusion method for digital twin operation of a cell drug factory combined with the Internet of Things, characterized in that: The method comprises: A dissolved oxygen sensor is used to monitor the fluctuation of dissolved oxygen concentration in the culture medium in the bioreactor in real time, and the time series data of dissolved oxygen concentration is obtained. The time series data is input into the long short-term memory neural network model for training and learning, and the change trend of dissolved oxygen concentration in a preset time period is predicted. Combined with the preset dissolved oxygen concentration threshold, it is determined whether the culture medium replacement cycle needs to be adjusted; If adjustment is needed, the cell density and cell viability during the cell growth process are collected in real time to generate a cell growth curve. The support vector machine algorithm is used to establish a correlation model between the cell growth curve and the culture medium replacement cycle, and the changes in the cell growth curve under different replacement cycle conditions are obtained to determine the target replacement cycle; The geometric model of the culture medium storage tank and bioreactor is constructed through 3D modeling software. The historical process parameters of the culture medium supply system and the real-time monitored temperature, pH value and dissolved oxygen concentration are mapped and associated with the geometric model to build a digital twin model of the reactor. Acquire the internal environmental parameters of the culture medium storage tank and bioreactor in real time, map and synchronize the internal environmental parameters with the virtual parameters in the digital twin model in real time, build a virtual-real fusion closed-loop control model for the culture medium supply system, and dynamically adjust the flow and concentration of the supply system according to cell growth requirements and metabolic status; Combined with the digital twin model, multi-scale and multi-physics simulation of the fusion culture medium tank and bioreactor during cell growth was carried out to simulate and analyze the impact of fluctuations in the redox potential and dissolved oxygen concentration of the culture medium on cell growth, and to change the ratio and supply of key nutrients in the culture medium; Analyze historical production data through machine learning algorithms, extract the quantitative relationship between key parameters such as cell metabolic rate, glucose concentration and pH value in the culture medium, and reactor temperature, establish a cell growth prediction model, and predict the growth trend and metabolic needs of cells in the target time period; Integrate the virtual-reality mapping mechanism of the culture medium supply system with the intelligent manufacturing execution system to achieve coordinated optimization control of the culture medium supply process and the entire drug production process, and automatically schedule the culture medium production process according to the production plan and cell growth requirements.

2. The method according to claim 1, characterized in that The method uses a dissolved oxygen sensor to monitor the fluctuation of the dissolved oxygen concentration of the culture medium in the bioreactor in real time, obtains time series data of the dissolved oxygen concentration, inputs the time series data into a long short-term memory neural network model for training and learning, predicts the change trend of the dissolved oxygen concentration in a preset time period, and determines whether the culture medium replacement cycle needs to be adjusted in combination with a preset dissolved oxygen concentration threshold, including: Obtain the dissolved oxygen concentration value collected by the dissolved oxygen sensor of the bioreactor, and obtain the dissolved oxygen concentration time series data set according to the sampling time mark; The dissolved oxygen concentration time series data set is standardized by using a normalization method, and a standardized dissolved oxygen concentration time series data set is obtained by removing outliers and supplementing missing values; Extracting the mean, variance and peak statistical features of dissolved oxygen concentration according to the standardized dissolved oxygen concentration time series data set, and inputting them into the long short-term memory neural network training to obtain a dissolved oxygen concentration prediction model; The dissolved oxygen concentration prediction model is used to predict the dissolved oxygen concentration in a fixed time period in the future. If the predicted dissolved oxygen concentration is lower than a preset concentration threshold and the decreasing rate exceeds a preset rate threshold, it is marked as a warning time point for culture medium replacement.

3. The method according to claim 1, characterized in that If adjustment is required, the cell density and cell viability during the cell growth process are collected in real time to generate a cell growth curve, and a support vector machine algorithm is used to establish a correlation model between the cell growth curve and the culture medium replacement cycle, to obtain the change of the cell growth curve under different replacement cycle conditions, and to determine the target replacement cycle, including: A cell density detector is used to collect cell density values ​​and fluorescence intensity signals in the culture medium, and the original cell growth data set is obtained according to the fixed time interval recording; Performing outlier detection on the original cell growth data set, supplementing missing data using a cubic spline interpolation method, and obtaining a processed cell growth data set; Calculating the cell density change rate and the cell viability change rate for the processed cell growth data set, and constructing a cell growth feature vector in combination with cell morphology parameters; Extracting historical culture medium replacement time points from a culture medium replacement record database, establishing a correspondence between the cell growth feature vector and the replacement time point, inputting the result into a support vector machine algorithm to obtain a culture medium replacement cycle model, and predicting the cell density integral value and the average value of cell viability at different replacement time intervals according to the culture medium replacement cycle model; It is determined whether the cell density integral value and the cell viability average value are within a preset threshold range. If they are within the preset threshold range, the time point with the maximum cell density integral value and the highest cell viability average value is selected as the target culture medium replacement cycle.

4. The method according to claim 1, characterized in that The geometric model of the culture medium storage tank and the bioreactor is constructed by using the three-dimensional modeling software, and the historical process parameters of the culture medium supply system and the real-time monitored temperature, pH value and dissolved oxygen concentration are mapped and associated with the geometric model to construct a digital twin model of the reactor, including: A three-dimensional structural model of the storage tank and the bioreactor is constructed using three-dimensional modeling software, wherein the three-dimensional structural model includes geometric data of the feed port, the discharge port, the stirring device, and the sensor installation position; Establish parameter monitoring points based on the surface of the three-dimensional structure model, obtain culture medium supply system data from the process parameter database, and collect real-time monitoring data of the sensor installation position; Dividing the fluid force field, temperature field and concentration field according to the parameter monitoring point, and calculating the distribution values ​​of the fluid force field, temperature field and concentration field in space by using a finite element algorithm; A deep neural network algorithm is used to train the coupling relationship between the fluid force field, temperature field and concentration field, and a reactor digital mapping structure is established according to the coupling relationship to obtain a digital twin model.

5. The method according to claim 1, characterized in that The real-time acquisition of the internal environmental parameters of the culture medium storage tank and the bioreactor, the real-time mapping and synchronization of the internal environmental parameters with the virtual parameters in the digital twin model, the construction of a virtual-real fusion closed-loop control model of the culture medium supply system, and the dynamic adjustment of the flow and concentration of the supply system according to the cell growth requirements and metabolic state, including: The internal environmental parameters of the culture medium storage tank and the bioreactor are collected by using a temperature sensor, a pH sensor, and a dissolved oxygen sensor, and the environmental parameters are subjected to noise reduction processing by using a sliding window method to obtain an environmental parameter data set; According to the environmental parameter data set, glucose concentration and lactate concentration in the cell culture process are collected, and combined with the cell density data to obtain a cell metabolic state data set; The cell metabolic state dataset is used to calibrate the sensor acquisition position in the digital twin structure, and the parameter space distribution is calculated by the Kriging interpolation algorithm to obtain a virtual parameter distribution dataset; The root mean square error value is calculated based on the virtual parameter distribution data set and the environmental parameter data set, and the culture medium supply controller is trained using a deep reinforcement learning algorithm to obtain the culture medium flow compensation value and the concentration compensation value.

6. The method according to claim 1, characterized in that The digital twin model is combined to perform multi-scale multi-physics simulation of the fusion culture medium storage tank and bioreactor during cell growth, simulate and analyze the impact of fluctuations in the redox potential and dissolved oxygen concentration of the culture medium on cell growth, and change the ratio and supply of key nutrients in the culture medium, including: Acquire geometric structure data of the storage tank and the reactor in the digital twin structure, and establish the Navier-Stokes equations describing fluid motion and the convection-diffusion equations describing the mass transfer process based on the geometric structure data; A multi-scale grid structure is established by using an adaptive grid division method according to the Navier-Stokes equations and the convection-diffusion equations, and multiple physical field distribution data are obtained through the multi-scale grid structure; A discrete element algorithm is used to construct a cell-scale calculation module according to the multi-physical field distribution data, and the redox potential distribution and dissolved oxygen concentration gradient data in the microenvironment are obtained through the cell-scale calculation module; A cellular metabolic flux balance equation group is established according to the dissolved oxygen concentration gradient data, and the glucose, amino acid and fatty acid consumption rate and metabolite generation rate data are calculated by the cellular metabolic flux balance equation group.

7. The method according to claim 1, characterized in that The method analyzes historical production data through a machine learning algorithm, extracts the quantitative relationship between key parameters such as cell metabolic rate, glucose concentration in the culture medium, pH value, and reactor temperature, establishes a cell growth prediction model, and predicts the growth trend and metabolic demand of cells within a target time period, including: Acquire culture medium parameter data from a historical database, and process the culture medium parameter data using a maximum and minimum value standardization method to obtain a standardized culture parameter data set; According to the standardized culture parameter data set, a wavelet transform denoising algorithm and a preset decomposition depth are used to perform signal reconstruction on the standardized culture parameter data set to obtain a smoothed data set; For the smoothed data set, eigenvalues ​​and eigenvectors are calculated by principal component analysis, and if the cumulative contribution rate of the eigenvalues ​​reaches a preset threshold, a reduced-dimensional feature data set is obtained; The recursive neural network model is trained using the dimension reduction feature data set. If the prediction error of the validation set is less than a preset threshold, a cell growth prediction model is obtained, and the metabolite generation rate is calculated using the cell growth prediction model.

8. The method according to claim 1, characterized in that The virtual-real mapping mechanism of the culture medium supply system is integrated with the intelligent manufacturing execution system to achieve coordinated optimization control of the culture medium supply process and the entire drug production process, and automatically schedule the culture medium production process according to the production plan and cell growth requirements, including: Collecting production plan scheduling data and cell culture progress data, and generating a production resource status data set according to the scheduling data and progress data; Obtain the maintenance records and fault records of production equipment, and use the equipment reliability evaluation index system to calculate the equipment status evaluation data set; According to the production resource status data set and the equipment status evaluation data set, an integer programming algorithm is used to optimize and calculate to obtain an initial scheduling plan for culture medium production; According to the initial scheduling plan for culture medium production, a deep reinforcement learning algorithm is used to construct a culture medium production scheduler. The production plan priority and equipment status characteristics are processed through the scheduler input layer to obtain a scheduling instruction sequence.

Images