An optimized control system and method for escherichia coli tyrosine production fermentation

Abstract

The application provides an optimization control system and method for Escherichia coli tyrosine fermentation, and relates to the technical field of bio-chemical industry and microbial fermentation engineering. The optimization control system and method for Escherichia coli tyrosine fermentation comprises the following steps: based on the mixed raw material of carbon source and composite nitrogen source feeding, raw material component data is obtained in real time by using an online spectrum analysis technology, initial environmental parameters are associated by using a multivariate fusion analysis method, a raw material component characteristic code and an optimal starting environmental parameter set are generated, and the adaptability and stability of the fermentation starting stage are improved; a dynamic prediction model based on time series data is used to calculate the pH variation trend in real time, a parameter adaptive control algorithm is used, and the dissolved oxygen control and energy consumption output are optimized in the fermentation tank, so that the metabolic efficiency of Escherichia coli is improved; and finally, the tyrosine yield reaches 61.7 g / L.

Description

Technical Field

[0001] This invention relates to the fields of biochemical engineering and microbial fermentation engineering technology, and in particular to an optimized control system and method for tyrosine production fermentation by Escherichia coli. Background Technology

[0002] Bio-fermentation engineering is an engineering technology field that uses microbial metabolic activities as its core to transform raw materials such as carbon and nitrogen sources into high-value-added products with specific economic value through biotransformation technology. In the process of synthesizing tyrosine using *Escherichia coli*, as the fermentation system is scaled up to industrial scale, it faces technical limitations such as limited mass transfer, uneven dissolved oxygen distribution, and the accumulation of metabolic byproducts (such as acetic acid). Existing fermentation control technologies often rely on offline detection in the raw material assessment stage, resulting in a lag in the setting of start-up parameters and feed-in strategies, leading to delayed cell growth in the early stages of fermentation or metabolic imbalance in the later stages. Process control often uses simple fixed parameters or threshold alarm mechanisms, which lag in responding to dynamic changes in key parameters such as pH and local dissolved oxygen in large tanks, causing deviations in the metabolic pathways of *E. coli* and a decrease in tyrosine yield.

[0003] Meanwhile, traditional process optimization typically focuses on a single objective (such as maximizing bacterial biomass), neglecting the synergistic relationship between the synthesis rate of the target product (tyrosine), byproduct inhibition, and the energy consumption of the large-scale system. This results in low energy efficiency during scale-up production and difficulty in breaking through the final yield bottleneck. Dissolved oxygen control in giant fermenters generally employs open-loop regulation, lacking refined coordination between vigorous stirring and high-flow-rate aeration, leading to extremely high aeration energy consumption. Therefore, there is an urgent need for an optimized control system for high-density fermentation of *E. coli* and high-yield tyrosine production to solve the challenges of mass transfer and metabolic control in scale-up production. Summary of the Invention

[0004] This invention provides an optimized control system and method for Escherichia coli tyrosine production fermentation. Addressing the problem of traditional fermentation assessment relying on offline detection, which leads to a disconnect between start-up parameter settings and feedstock characteristics, this invention utilizes online feed spectral analysis and multivariate fusion (S1) to acquire feedstock component characteristic codes in real time and accurately match optimal start-up parameters, thus eliminating control lag and ensuring growth stability in the early stages of fermentation. For the issue of lag in the dynamic response of key parameters such as pH and dissolved oxygen, this invention introduces a dynamic prediction model based on time-series data and adaptive PID control (S2). By predicting pH change trends and implementing feedforward adjustment, drastic fluctuations in the fermentation microenvironment are avoided, thereby ensuring the continuity of the E. coli metabolic pathway. Addressing the problem of byproduct accumulation and yield bottlenecks caused by traditional processes focusing on a single objective, this invention constructs a metabolite-operating parameter prediction model and employs a Pareto front-based multi-objective optimization algorithm (S3). By comprehensively balancing the tyrosine synthesis rate and the inhibition of byproducts such as acetic acid, the globally optimal combination of operating parameters is found under multivariate dynamic constraints, effectively overcoming metabolic overflow. To address the problem of excessive energy consumption due to difficulties in coordinated control in large fermenters, this invention employs a deep reinforcement learning network to optimize DO control (S4). By globally searching for the optimal combination of stirring and aeration while meeting dissolved oxygen biological requirements, it significantly reduces the operating energy consumption of the giant fermentation system. Compared to the limited fermentation yield of traditional processes, this invention achieves a technological breakthrough by reaching a final tyrosine yield of 61.7 g / L.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an optimized control system and method for tyrosine production fermentation by *Escherichia coli*, comprising the following steps:

[0006] S1: Based on the mixed raw materials of carbon source and composite nitrogen source in the feed, online spectral analysis technology is used to obtain raw material composition data in real time. The initial environmental parameters are correlated through multivariate fusion analysis method to generate raw material component feature codes and optimal start-up environmental parameter set;

[0007] S2: Based on the feature codes of the raw material components and real-time fermentation time data, a dynamic prediction model based on time data is used to calculate the pH change trend, and combined with the parameter adaptive control algorithm, dynamic evaluation of the main fermentation state and pH stabilization control instructions are generated.

[0008] S3: Based on the feature codes of the raw material components and the dynamic evaluation of the main fermentation state, and by integrating the detection data of key products and by-products, a metabolite-operating parameter prediction model is constructed. The comprehensive optimization target of fermentation quality and the set value of operating parameters are generated by solving the Pareto frontier-based multi-objective optimization algorithm.

[0009] S4: Based on the comprehensive optimization target of fermentation quality and the set values ​​of operating parameters, a deep reinforcement learning control strategy model is adopted to optimize the energy consumption output of the fermenter under the premise of meeting the dissolved oxygen control requirements, and generate low-energy DO optimization control commands.

[0010] Furthermore, following step S4, step S5, which involves full-process visual monitoring and anomaly-assisted decision-making, is also included: based on the parameter set, instructions, and set values ​​generated in the aforementioned steps, a synchronously updated high-fidelity digital twin is constructed, the process status is displayed through three-dimensional visualization and mixed reality technology, and a full-process visual monitoring map and intelligent auxiliary decision-making suggestions are generated based on anomaly diagnosis algorithms. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the main steps of the present invention;

[0012] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0013] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0014] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0015] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0016] Figure 6 This is a detailed schematic diagram of S5 of the present invention;

[0017] Figure 7 This is a system block diagram of the present invention;

[0018] Figure 8 shows the tyrosine yield curve for Example 1;

[0019] Figure 9 shows the tyrosine yield curve for Example 2;

[0020] Figure 10 shows the tyrosine yield curve for Example 3;

[0021] Figure 11 shows the tyrosine yield curve for Example 4;

[0022] Figure 12 shows the tyrosine production curve for Example 5. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Phase S1:

[0025] S101: Based on the fed-batch production of a mixture of carbon source and composite nitrogen source (yeast powder and inorganic salts), an online near-infrared spectral scanner was used to collect spectral reflectance data in the wavelength range of 1100-2500 nm. The scanning frequency was set to 10 times per second, and the average value was taken after 60 seconds of continuous scanning, recording a total of 150 sets of raw spectral data points.

[0026] S102: Based on the raw material's original spectral data, a pre-established quantitative analysis model is invoked. This model is trained using PLS. Reflectance data at a specific wavelength is input and substituted into the model formula to calculate the carbon source concentration in the material as 45.5%, the total nitrogen source concentration as 12.8%, and the precursors and salts as 2.5%, generating a quantitative table of raw material component proportions.

[0027] S103: Based on the raw material composition ratio quantification table and the initial fermenter parameters of 34 ℃ temperature, 7.0 pH, and 25% dissolved oxygen, a multidimensional input matrix was constructed using principal component analysis. Carbon sources and total nitrogen with absolute loading values ​​greater than 0.4 were extracted as key feature combinations to generate a raw material-environment correlation feature vector.

[0028] S104: Based on the feature vector input adaptive fuzzy logic controller, activate the rules in the fuzzy rule base for high-density fermentation of E. coli: if the carbon source is medium-high and the nitrogen source is medium, output the optimal start-up temperature setting value of 34.5 ℃, pH setting value of 6.95, and dissolved oxygen target value of 28% (about 2.0 mg / L), and generate the optimal start-up environment parameter set.

[0029] S105: Integrate features and quantification tables to form key identifiers that characterize the unique properties of raw materials and generate feature codes for raw material components.

[0030] Phase S2:

[0031] S201: Based on the raw material component feature codes and real-time acquired fermentation time-series data of the fermenter (temperature 34.5 ℃, pH 6.95, dissolved oxygen 2.1 mg / L, E. coli OD... 600 (Data), build and run an LSTM dynamic prediction model.

[0032] S202: Input current and previous historical data, the model outputs the pH prediction sequence for the next 30 minutes, and the calculated short-term dynamic prediction value is 6.90.

[0033] S203: Based on the current actual pH measurement of 6.95, the set value of 7.0, and the predicted value of 6.90, calculate the current pH deviation Δe. c =6.95-7.0=-0.05, and simultaneously calculate the predicted pH deviation Δe. p=6.90-7.0=-0.10. A dynamic evaluation index for pH deviation is generated.

[0034] S204: Based on evaluation indicators, the PID control algorithm is adjusted in real time using an online optimization method to generate an adaptive PID control parameter set.

[0035] S205: Execute the algorithm to calculate the adjustment amount:

[0036] (Equation 1)

[0037] Where u(t) is the precise adjustment amount of the system output, in L / min;

[0038] K p This is the proportionality coefficient;

[0039] e(t) represents the current pH deviation of the system, denoted as Δec;

[0040] T i The integral time constant is expressed in minutes. This is the cumulative integral term of the deviation;

[0041] T d The differential time constant is expressed in minutes. The differential term is the rate of change of the deviation;

[0042] t is a time parameter, in minutes.

[0043] It outputs the precise alkali replenishment amount (e.g., 1.5 L / min) and cooling water adjustment amount for the ammonia water feed system, and generates pH stabilization control commands.

[0044] S206: Generates a dynamic snapshot of the main fermentation state, comprehensively judging that the current E. coli metabolism is in a highly active state and the environment is stable.

[0045] Phase S3:

[0046] S301: Combine key product data obtained from periodic HPLC detection (current tyrosine concentration 35.2 g / L, byproduct acetic acid 1.8 g / L) to generate a target metabolite-state association dataset.

[0047] S302: The random forest algorithm is used to train and update the predictive correlation model between metabolite concentration and feed-in and dissolved oxygen operation parameters.

[0048] S303: Define the set of multi-objective optimization functions

[0049] Max(0.5×tyrosine synthesis rate - 0.3×acetic acid production rate - 0.2×energy consumption) (Equation 2)

[0050] Where: Max represents finding the maximum value of the overall function formula within the parentheses;

[0051] Tyrosine synthesis rate, in g / (L·h);

[0052] Acetic acid formation rate, in g / (L·h);

[0053] Energy consumption is normalized (converted to a dimensionless value of 0 to 1) for easier calculation on the same order of magnitude;

[0054] 0.5, 0.3, and 0.2 represent the weight ratios of each parameter in the optimization function.

[0055] S304: The non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy is used for online solution, with constraints set as temperature 32-36 ℃, pH 6.8-7.2, and dissolved oxygen 15-30%. The optimal combination of operating parameters is obtained.

[0056] S305: Extract set values ​​(such as temperature 34.0 ℃, pH 6.90, flow rate adjustment), generate comprehensive optimization targets and operating parameter set values ​​for fermentation quality, and suppress by-product overflow.

[0057] Phase S4:

[0058] S401: Extracts real-time DO value (1.5 mg / L), rotation speed (80 rpm), and aeration (40 m). 3 The state observation vector is generated from the cumulative energy consumption data ( / min) and the cumulative energy consumption data.

[0059] S402: Build a reinforcement learning environment with action space for adjusting stirring speed (-5 to +5 rpm / level) and valve opening.

[0060] S403: Load a pre-trained proximal policy optimization network model that can make decisions within the reinforcement learning state-action-reward framework and generate a deep reinforcement learning control policy network instance.

[0061] S404: Based on the DO control state observation vector and the deep reinforcement learning control strategy network instance, input the current state, obtain the optimal action output by the network (e.g., stirring speed +2 rpm, aeration maintained), and generate recommended energy consumption optimization actions.

[0062] S405: Based on the recommended energy consumption optimization actions, considering equipment safety constraints (such as the maximum speed not exceeding 120 rpm), after verifying the safety upper limit, generate the final executable stirrer speed control command and air intake valve opening control command, and generate the low energy consumption DO optimization control command.

[0063] S5 stage:

[0064] S501: Integrate raw material component feature codes, dynamic evaluation of primary fermentation status, comprehensive optimization targets and operating parameter settings for fermentation quality, low-energy DO optimization control instructions and their feedback data after execution, construct the core input dataset of a high-fidelity digital twin, and generate the core input dataset of the digital twin.

[0065] S502: Based on the core input dataset of the digital twin, the physical fermenter model, the improved E. coli fermentation kinetics model, and the computational fluid dynamics model, update the internal state of the digital twin. The E. coli growth rate is updated using the Monod equation.

[0066] (Equation 3)

[0067] Where: μ is the current specific growth rate of E. coli, in h. -1 ;

[0068] μ max Maximum specific growth rate, in h -1 ;

[0069] S represents the current carbon source substrate concentration, in g / L;

[0070] K s This is the saturation constant, expressed in g / L.

[0071] The rate of tyrosine product formation was updated using the Luedeking-Piret equation:

[0072] (Equation 4)

[0073] in: The product (tyrosine) formation rate, expressed in g / (L·h);

[0074] α is the production constant coupled with the growth of Escherichia coli, in g / g (grams of product / grams of bacteria).

[0075] The value represents the growth rate of Escherichia coli cells, expressed in g / (L·h).

[0076] β is the uncoupled generation constant, in units of h. -1 ;

[0077] X represents the current bacterial cell concentration, expressed in OD0.05. 600 express.

[0078] After the update is complete, an updated digital twin instance will be generated.

[0079] S503: The extended Kalman filter algorithm is used to fuse and assimilate the data collected by the real-time sensor with the predicted value of the updated digital twin instance to ensure that the twin state is synchronized with the physical entity and generate a synchronized digital twin state.

[0080] S504: The system uses a 3D engine to render the flow field and dissolved oxygen distribution inside the tank. For the dissolved oxygen gradient at the bottom and top of the tank, the system intuitively displays dark blue (low oxygen zone) and bright yellow (oxygen-rich zone), as well as the production indicator curve.

[0081] S505: When the system detects abnormal deviations in key parameters (such as a continuous abnormal decrease in dissolved oxygen), it uses a case reasoning algorithm based on the historical case database to match similar historical cases (such as poor mixing around the stirring baffle) and their handling solutions, generating similar historical case matching results.

[0082] S506: Based on the matching results of similar historical cases and the current abnormal data, the conditional probability of different potential fault causes (such as a mixed defect probability of 60% and a replenishment fluctuation probability of 30%) is calculated using a Bayesian network algorithm to generate a fault probability diagnosis report.

[0083] S507: Based on the fault probability diagnosis report and combined with the expert rule base, propose targeted operational intervention suggestions (such as immediately increasing the flow rate of the deep aeration branch and fine-tuning the flow pump frequency) and generate intelligent auxiliary decision-making suggestions.

[0084] Example 1: Standard High-Yield Control Mode

[0085] This embodiment focuses on high-density fermentation of standard batches of E. coli, with the goal of maximizing and stabilizing tyrosine production.

[0086] S101: Based on the mixed raw materials fed in, reflectance data is collected in the wavelength range using an online near-infrared spectroscopy scanner. For example, the reflectance is measured to be 0.38 at 1550 nm and 0.45 at 1700 nm.

[0087] S102: Call the pre-trained PLS model; the carbon source concentration calculation formula is as follows:

[0088] C = 0.15 × R 1550 +0.08×R 1700 +2.5 (Equation 5)

[0089] Where: C is the calculated carbon source concentration parameter (value before normalization), which is finally converted to a mass percentage concentration, in % %.

[0090] R 1550 The spectral reflectance was measured at a wavelength of 1550 nm.

[0091] R1700 The spectral reflectance was measured at a wavelength of 1700 nm.

[0092] 0.15 and 0.08 are weighting coefficients extracted based on historical fermentation data using the PLS algorithm;

[0093] 2.5 represents the constant terms extracted by the algorithm.

[0094] Substitute R 1550 =0.38, R 1700 After setting the value to 0.45, the carbon source concentration parameter is 0.15×0.38+0.08×0.45+2.5=2.593 (value before normalization). Similarly, the total nitrogen concentration parameter is calculated as follows:

[0095] C nitrogen =0.09×R 1550 +0.11×R 1700 +1.2 (Equation 6)

[0096] Where: C nitrogen The calculated total nitrogen concentration parameter (value before normalization) is finally converted to a mass percentage concentration, in % (%).

[0097] R 1550 The spectral reflectance was measured at a wavelength of 1550 nm.

[0098] R 1700 The spectral reflectance was measured at a wavelength of 1700 nm.

[0099] 0.09 and 0.11 are weighting coefficients extracted based on historical fermentation data using the PLS algorithm;

[0100] 1.2 represents the constant terms extracted by the algorithm.

[0101] After normalization, the final carbon source ratio was calculated to be 45.5% and the total nitrogen source ratio to be 12.8%, generating a quantitative table of raw material composition ratios.

[0102] S103: Based on the calculated raw material composition ratio table (carbon source 45.5%, total nitrogen source 12.8%) and initial fermenter parameters (temperature 34 ℃, pH 7.0, dissolved oxygen 25%), a multidimensional input matrix was constructed using principal component analysis for multivariate fusion analysis. Carbon source and total nitrogen with absolute loading values ​​greater than 0.4 were extracted as key feature combinations to generate a raw material-environment correlation feature vector.

[0103] S104: Based on the raw material-environment association feature vector, input the adaptive fuzzy logic controller. Activate the rules in the fuzzy rule base for high-density fermentation of E. coli: if the carbon source is medium-high and the nitrogen source is medium, then infer and output the parameters for the optimal start-up stage, namely, temperature setpoint 34.5 ℃, pH setpoint 6.95, and dissolved oxygen target value 28% (approximately 2.0 mg / L), generating the optimal start-up environment parameter set.

[0104] S105: Integrate the raw material component ratio quantification table with the raw material-environment association feature vector to form a key identifier that characterizes the unique attributes of this batch of raw materials and their impact on the early stage of E. coli fermentation, and finally generate the raw material component feature code [14.2%, 9.8%, 3.1%].

[0105] S205: Mid-stage of primary fermentation, real-time pH 6.92, predicted average pH for the next 30 minutes 6.88, deviation Δe c =-0.05. Based on this evaluation index and the adaptive PID control parameter set, the algorithm is executed to calculate the adjustment amount:

[0106] (Equation 1)

[0107] Where e(t) = Δe c =-0.05, integral term Take the cumulative deviation of the first 10 minutes, and the differential term. By taking the two most recent deviation changes and substituting the parameters into the formula, the system outputs the precise alkali adjustment amount (e.g., 1.5 L / min) and cooling water adjustment amount for the ammonia addition system, generating a pH stability control command.

[0108] S3: A random forest model was trained based on HPLC data (tyrosine 32 g / L, acetic acid 1.5 g / L). The optimal operating combination was determined by NSGA-II: temperature 34.0 ℃, pH 6.90.

[0109] S4: The reinforcement learning network outputs an increase of 1 in stirring speed (88 rpm) based on the DO observation vector, maintaining aeration. Energy consumption is reduced by 8.5%.

[0110] S502: The physical fermenter model updates the internal temperature distribution of the tank based on the heat conduction equation. The improved E. coli fermentation kinetics model updates the cell growth rate (using the Monod equation):

[0111] (Equation 3)

[0112] The maximum specific growth rate μ max =0.45 h -1 saturation constant K s=0.5 g / L, where S is the current carbon source substrate concentration. Simultaneously, based on the metabolic characteristics of tyrosine synthesis in *E. coli*, the product formation rate was updated (using the Luedeking-Piret equation):

[0113] (Equation 4)

[0114] The CFD model simulates the dissolved oxygen distribution at different depths from the liquid surface in the tank. The digital twin shows a uniform distribution of bacterial density within the tank. The yield curves of actual and estimated tyrosine values ​​are shown below. Figure 8 As shown, the final tyrosine yield for this batch reached 61.7 g / L.

[0115] Example 2: High-concentration raw material fluctuation adaptation mode

[0116] This embodiment dynamically adjusts the system in response to fluctuations in raw material composition (high carbon source concentration).

[0117] S1: The feed carbon source concentration was increased to 52.0% using the PLS formula, with a total nitrogen source of 11.5%. Spectral signatures detected a deviation in the carbon-to-nitrogen ratio. The system automatically adjusted the startup environment: the initial temperature was lowered to 33.8 ℃ to prevent excessively rapid initial growth.

[0118] S2: pH LSTM prediction during fermentation showed a significant increasing trend in acidic products. The adaptive control system increased the ammonia pre-feed rate in advance using a PID formula, stabilizing the pH at 6.90.

[0119] S3: The multi-objective decision module detected the risk of acetic acid accumulation and increased the weight of "minimizing byproduct accumulation" to 0.4. The solution yielded a command to reduce the flow acceleration rate.

[0120] S4: The reinforcement learning strategy network executes linkage control, increasing the aeration valve by 10% and fine-tuning the stirring speed.

[0121] S5: Through digital twin monitoring, the acetic acid concentration was maintained below 0.8 g / L, and the fermentation cycle was extended by 5 hours. The yield curves of actual and estimated tyrosine values ​​are shown below. Figure 9 As shown, the final yield reached 60.9 g / L.

[0122] Example 3: Low Energy Consumption and Energy Efficiency Priority Mode

[0123] This embodiment focuses on optimizing the overall energy consumption of the fermentation system while ensuring basic output.

[0124] S1: The optimal startup parameters are set to the normal median values: temperature 34.2 ℃, pH 7.0.

[0125] S2: The pH dynamic evaluation index is in the "low deviation" range, and the PID parameters Kp and Ti remain stable, reducing the energy consumption of frequent ammonia pumping.

[0126] S3: The multi-objective optimization function set increases the weight of "minimize energy consumption" from 0.1 to 0.3. NSGA-II seeks combinations of operations that keep dissolved oxygen levels below the lower limit (15-18%) but within the metabolically safe range.

[0127] S4: The deep reinforcement learning control strategy network optimizes the coordination between the stirring speed and the air intake valve, avoiding the high energy consumption range of the stirring motor, and reducing the cumulative energy consumption by 15.2%.

[0128] S5: Under twin monitoring, the yield curves of actual and estimated tyrosine measurements are shown below. Figure 10 As shown, the final yield reached 60.4 g / L, achieving synergistic optimization of yield and energy efficiency.

[0129] Example 4: Intelligent Repair Mode for Abnormal Disturbances

[0130] This embodiment simulates the repair control when abnormal deviations in parameters occur during fermentation.

[0131] S4: After 45 hours of fermentation, the DO suddenly dropped to 8%. The reinforcement learning module forcibly increased the rotation speed to the maximum of 115 rpm.

[0132] S5: Case-based reasoning calculations determined that the probability of localized blockage in the aeration system was 85%. The expert rule base recommended: perform backwashing and adjust the inlet air pressure. Dissolved oxygen levels returned to normal after the intervention.

[0133] Final result: Due to the system's powerful fault diagnosis and compensation capabilities, the batch was successfully salvaged, and the yield curves of actual and estimated tyrosine measurements were as follows: Figure 11 As shown, the yield reached 59.7 g / L.

[0134] Example 5: Metabolic Redirection Control Mode under Carbon-Nitrogen Ratio Imbalance

[0135] This embodiment simulates the process by which the system adaptively adjusts to ensure tyrosine production when the carbon-nitrogen ratio of the raw material batches is unbalanced.

[0136] S1: The PLS formula calculates a carbon source concentration of 46.5% and a total nitrogen source concentration reduced to 10.2%. The "high carbon, low nitrogen" characteristic is identified, and fuzzy logic inference indicates a decrease in temperature to 33.5 ℃, pH to 7.05, and a DO target of 25%.

[0137] S2: LSTM calculations revealed the risk of overflow metabolism. The PID algorithm adjusted the proportional coefficient Kp and the derivative time Td in real time, outputting a high-frequency micro-alkaline supplementation command, and the actual pH stabilized at 7.0.

[0138] S3: HPLC detection of acetic acid carries a risk of exceeding 2.5 g / L. In multi-objective optimization, the weight of "minimizing acetic acid accumulation" was increased to 0.5, and NSGA-II limited the carbon source flux acceleration rate.

[0139] S4: The control state observation vector shows a decrease in dissolved oxygen consumption rate. The strategy network lowers the intake valve to 35% and maintains stirring at 75 rpm to reduce energy consumption.

[0140] S5: The product formation constant was corrected according to the Luedeking-Piret equation. Twin monitoring showed that acetic acid accumulation was inhibited, and the tyrosine pathway was reactivated in the later stages. The yield curves of actual and estimated tyrosine values ​​are shown below. Figure 12 As shown, the final yield reached 60.8 g / L, overcoming the production reduction crisis.

[0141] The preferred embodiments of the present invention have been described in detail above; however, the present invention is not limited thereto. Within the scope of the inventive concept, various simple modifications can be made to the technical solutions of the present invention, including combinations of various technical features in any other suitable manner. These simple modifications and combinations should also be considered as the content disclosed in the present invention and are all within the protection scope of the present invention.

Claims

1. An optimized control method for tyrosine production fermentation by *Escherichia coli*, comprising the steps of feeding a mixture of carbon source and composite nitrogen source into a fermenter for fermentation, and controlling the pH and dissolved oxygen during the fermentation process, characterized in that... Includes the following steps: S1: Based on the mixed raw materials of carbon source and composite nitrogen source in the feed, the raw material composition data is obtained in real time by online spectral analysis technology. The initial environmental parameters are associated by multivariate fusion analysis method to generate raw material component feature codes and optimal start-up environmental parameter set. The fermenter is then controlled to adjust to the corresponding initial state according to the optimal start-up environmental parameter set. S2: Based on the raw material component feature codes and real-time fermentation time series data, a dynamic prediction model based on time series data is used to calculate the pH change trend. Combined with the parameter adaptive control algorithm, a dynamic evaluation of the main fermentation state and pH stabilization control command are generated, and the ammonia water feeding system and temperature control system are controlled to execute the pH stabilization control command. S3: Based on the raw material component feature codes and dynamic evaluation of the main fermentation state, and by integrating the detection data of key products and by-products, a metabolite-operating parameter prediction model is constructed. The comprehensive optimization target of fermentation quality and the set value of operating parameters are generated by solving the Pareto front-based multi-objective optimization algorithm, and the set value of operating parameters is sent to the fermentation process control system. S4: Based on the comprehensive optimization target of fermentation quality and the set values ​​of operating parameters, a deep reinforcement learning control strategy model is adopted to optimize energy consumption output under the premise of meeting the dissolved oxygen (DO) control requirements of the fermenter, generate low-energy DO optimization control instructions, and control the agitator and air inlet valve of the fermenter to execute the low-energy DO optimization control instructions.

2. The optimized control method for tyrosine production fermentation by *Escherichia coli* according to claim 1, characterized in that, S1 includes the following steps: S101: Based on the mixed raw materials of carbon source and composite nitrogen source in the feed, online near-infrared spectroscopy scanning technology is used, and the raw material raw spectral data is generated by spectral acquisition in a specific wavelength range; S102: Based on the original spectral data of the raw materials, a multivariate regression analysis algorithm is used, combined with a pre-established quantitative analysis model, to calculate the content and ratio of key components (carbon source, total nitrogen, precursor amino acids) in the raw materials, and generate a quantitative table of raw material component ratios. S103: Based on the raw material component ratio quantification table and the initially set fermenter environmental parameters, a dimensionality reduction feature extraction method is used to perform multivariate fusion analysis, identify and extract key feature combinations that affect the fermentation start-up efficiency of Escherichia coli, and generate a raw material-environment association feature vector. S104: Based on the raw material-environment correlation feature vector, an adaptive fuzzy logic control algorithm is adopted. Through a preset fuzzy rule base, the optimal start-up stage temperature setpoint, pH setpoint, dissolved oxygen target value and its adjustment rate strategy for the current batch of raw materials are inferred and output to generate the optimal start-up environment parameter set. S105: Integrate the raw material component ratio quantification table with the raw material-environment association feature vector to form a key identifier that characterizes the unique properties of the raw material and its impact on the early stage of E. coli fermentation, and generate the raw material component feature code.

3. The optimized control method for tyrosine production fermentation by *Escherichia coli* according to claim 1, characterized in that, S2 includes the following steps: S201: Based on the feature codes of raw material components and the real-time collected time-series data sequence of the main fermentation process, a dynamic prediction model is constructed and run using a time-series deep neural network to generate an instance of the main fermentation process prediction model; S202: Based on the main fermentation process prediction model instance, input current and historical time series data, calculate the pH change trend and prediction value within a future control cycle, and generate short-term dynamic pH prediction value. S203: Based on the current actual pH measurement value, pH set value and the short-term dynamic prediction value of pH, calculate the current pH deviation, the predicted pH deviation and its rate of change, and generate a dynamic evaluation index of pH deviation. S204: Based on the pH deviation dynamic evaluation index and combined with the real-time acquired current fermentation stage identifier, a fuzzy inference system or an online optimization method based on gradient descent is used to adjust the proportional coefficient, integral time, and derivative time of the adaptive proportional-integral-derivative (PID) control algorithm in real time to generate an adaptive PID control parameter set. S205: Based on the dynamic evaluation index of pH deviation and the set of adaptive PID control parameters, execute the parameter adaptive PID control algorithm to calculate and output the precise adjustment amount of the heating / cooling system and the ammonia water flow system, and generate pH stabilization control command; S206: Integrate the aforementioned dynamic evaluation index of pH deviation, current key parameter values, and fermentation stage information to form a comprehensive judgment on the metabolic activity of E. coli and environmental stability in the fermenter, and generate a dynamic snapshot of the main fermentation state.

4. The optimized control method for tyrosine production fermentation by *Escherichia coli* according to claim 1, characterized in that, S3 includes the following steps: S301: Based on the feature codes of raw material components and dynamic evaluation of the main fermentation state, combined with the concentration data of key products (tyrosine) and by-products (acetic acid, etc.) obtained by periodic detection through high performance liquid chromatography (HPLC), and aligned by time delay compensation algorithm, a target metabolite-state association dataset is generated. S302: Based on the target metabolite-state association dataset, an ensemble learning regression algorithm is used to train and update the predictive association model between metabolite concentration and operating parameters, and generate a metabolite-operating parameter prediction model. S303: Define the optimization objective function, including maximizing the tyrosine ratio production rate, minimizing the accumulation of byproducts such as acetic acid, minimizing the remaining fermentation time, minimizing energy consumption, and generating a set of multi-objective optimization functions; S304: Based on the metabolite-operating parameter prediction model, the pre-built product synthesis rate model, the pre-built energy consumption model, the current state parameters, and the multi-objective optimization function set, a global heuristic optimization algorithm is used to perform online multi-objective optimization to find a combination of operating parameters that satisfies the constraints and approximates the Pareto front, thereby generating a multi-objective optimization solution set. S305: Based on the preset decision rules, select an optimal solution from the multi-objective optimization solution set, extract its corresponding operation parameter setting value, and generate the comprehensive optimization target and operation parameter setting value for fermentation quality.

5. The optimized control method for tyrosine production fermentation by *Escherichia coli* according to claim 1, characterized in that, S4 includes the following steps: S401: Based on the comprehensive optimization target of fermentation quality and the set value of operating parameters, combined with the real-time collected DO value, DO change rate, temperature, pH, current stirring speed, current aeration intensity, and cumulative energy consumption data, a DO control state observation vector is generated. S402: Construct a reinforcement learning environment state space, whose state includes the DO control state observation vector and the DO target range information in the operation parameter setting value; define the action space as the fermenter agitator speed adjustment gear and / or the air intake valve opening adjustment gear; design a reward function to give a positive reward for maintaining DO within the target range, a penalty for exceeding the range, and a negative reward for the action amplitude, and generate a reinforcement learning state-action-reward framework. S403: Load a pre-trained deep policy network model that can make decisions within the reinforcement learning state-action-reward framework and generate a deep reinforcement learning control policy network instance. S404: Based on the DO control state observation vector and the deep reinforcement learning control policy network instance, input the current state, obtain the optimal action output by the network, and generate recommended energy consumption optimization actions; S405: Based on the recommended energy consumption optimization actions and considering equipment safety constraints, generate the final executable stirrer speed control command and / or air intake valve opening control command, and generate low-energy DO optimization control command.

6. The optimized control method for tyrosine production fermentation by *Escherichia coli* according to claim 1, characterized in that, Following step S4, step S5, which involves full-process visual monitoring and anomaly-assisted decision-making, is also included. S501: Integrates raw material component feature codes, dynamic evaluation of main fermentation status, comprehensive optimization targets and operating parameter settings for fermentation quality, low-energy DO optimization control instructions and real-time sensor feedback data to construct the core input dataset of a high-fidelity digital twin and generate the core input dataset of the digital twin. S502: Based on the core input dataset of the digital twin, the physical fermenter model, the improved E. coli fermentation kinetics model, and the computational fluid dynamics model, update the internal state of the digital twin and generate an updated digital twin instance; S503: Using Kalman filtering or its improved algorithm, the data collected by the real-time sensor is fused and assimilated with the predicted value of the updated digital twin instance to ensure that the twin state is synchronized with the physical entity and generate a synchronized digital twin state. S504: Based on the synchronized digital twin state, render and display the temperature field distribution cloud map, dissolved oxygen concentration field distribution cloud map, simulated cell density distribution, real-time curves of key parameters, and the optimization target achievement indicator set in step S3 in real time through a three-dimensional visualization engine or mixed reality device interface, and generate a dynamic visualization map of the entire process. S505: When the system detects abnormal deviations in key parameters, it uses a case reasoning algorithm based on the historical case database to match similar historical cases and their handling solutions, and generates similar historical case matching results. S506: Based on the matching results of similar historical cases and the current abnormal data, the conditional probability of different potential fault causes is calculated using the Bayesian network algorithm to generate a fault probability diagnosis report. S507: Based on the fault probability diagnosis report and combined with the expert rule base, propose targeted operation intervention suggestions and generate intelligent auxiliary decision-making suggestions.

7. An optimized control system for Escherichia coli fermentation for tyrosine production, comprising a fermenter, an ammonia feed system, a temperature control system, and a fermentation process control system, characterized in that, The optimized control system includes the following modules: The raw material intelligent start-up module is based on the mixed raw materials of carbon source and composite nitrogen source. It uses online spectral analysis technology to obtain raw material composition data and generates raw material component feature codes and optimal start-up environment parameter set through multivariate correlation analysis. The pH dynamic prediction and control module, based on the raw material component feature codes and real-time fermentation time data, uses a dynamic prediction model to calculate the pH change trend, combines a parameter adaptive control algorithm to generate dynamic evaluation of the main fermentation state and pH stability control instructions, and controls the ammonia water feeding system and temperature control system to execute them. The multi-objective decision-making module, based on the raw material component feature codes and the optimal start-up environment parameter set, combined with real-time fermentation data and product metabolite detection data, uses machine learning algorithms to construct a metabolite-operation parameter prediction model, and generates a comprehensive optimization objective for fermentation quality and set values ​​for operation parameters through multi-objective optimization. The low-energy execution module, based on the comprehensive optimization target of fermentation quality and the set values ​​of operating parameters, adopts a deep reinforcement learning control strategy model to optimize energy consumption output while meeting the DO control requirements of the fermenter, and generates and executes low-energy DO optimization control commands.

8. An optimized control system for Escherichia coli tyrosine production fermentation according to claim 7, characterized in that, The system also includes: The twin monitoring and feedback module, based on the data generated by the aforementioned modules and real-time sensor feedback data, constructs and updates a high-fidelity digital twin, realizing full-process visual monitoring, and generating intelligent auxiliary decision-making suggestions based on anomaly diagnosis algorithms when abnormal parameters are detected.

9. An optimized control system for Escherichia coli tyrosine production fermentation according to claim 7, characterized in that, The raw material intelligent start-up module includes the following sub-modules: The raw material characteristic analysis submodule, based on the mixed raw materials of carbon source and composite nitrogen source in the feed, uses online near-infrared spectroscopy scanning technology to collect raw spectral data, and calculates the content and ratio of carbon source, total nitrogen, and precursor amino acids through multivariate regression analysis algorithm or nonlinear classification algorithm to generate a raw material component ratio quantification table. The feature association submodule, based on the raw material component ratio quantification table and the initial fermenter environmental parameters, uses a dimensionality reduction feature extraction method to perform multivariate fusion analysis, extracts key feature combinations that affect the start-up efficiency of E. coli fermentation, and generates a raw material environment association feature vector. The startup strategy generation submodule, based on the raw material environment associated feature vector, adopts an adaptive fuzzy logic control algorithm to infer the optimal startup temperature setpoint, pH setpoint, dissolved oxygen target value and adjustment rate strategy through a preset fuzzy rule base, and generates raw material component feature codes and the optimal startup environment parameter set.

10. An optimized control system for Escherichia coli tyrosine production fermentation according to claim 7, characterized in that, The multi-objective decision-making module includes the following sub-modules: The metabolism modeling submodule, based on the feature codes of the raw material components, real-time fermentation time data and key product concentration data detected by HPLC, uses an ensemble learning regression algorithm to train and generate a metabolite-operating parameter prediction model. The objective definition submodule defines multi-objective optimization functions, including: maximizing tyrosine yield and synthesis rate, minimizing acetic acid accumulation, minimizing remaining fermentation time, and minimizing system stirring and temperature control energy consumption, generating a set of multi-objective optimization functions; The parameter optimization submodule, based on the metabolite-operation parameter prediction model, product synthesis model, energy consumption model, current state parameters, and the multi-objective optimization function set, uses a global heuristic optimization algorithm to perform online optimization and solve the problem, generating a multi-objective optimization solution set; based on preset decision rules, it selects the optimal combination of operation parameters to generate the comprehensive optimization target for fermentation quality and the set values ​​of operation parameters.

Images