A method and apparatus for optimizing biological reaction processes using deep learning

Abstract

This application discloses a method and apparatus for optimizing a bioreaction process using deep learning, relating to the field of cell bioengineering. The method includes acquiring the first bioreaction type, the first reaction stage, the first basic configuration parameters of the cell bioreactor, and the first monitoring parameters collected by multiple types of sensors in the current bioreaction process within a cell bioreactor; retrieving contextual data related to the first bioreaction type and the first reaction stage from a bioreaction knowledge vector base; constructing prompt words based on the contextual data, the first monitoring parameters, the first bioreaction type, the first reaction stage, the first basic configuration parameters, preset data analysis instructions, and historical question-and-answer data from monitoring and analysis; and inputting the prompt words into a trained large-scale cell bioreaction model, so that the large-scale cell bioreaction model outputs first control parameters for a programmable logic controller to predict the operating state of the cell bioreactor and correct for faults.

Description

Technical Field

[0001] This application relates to the field of cell bioengineering technology, and in particular to a method and apparatus for optimizing biological reaction processes using deep learning. Background Technology

[0002] A cell bioreactor is a container used for the controlled cultivation of cells and other biological organisms to produce or develop targeted products or substances, or to conduct specific reactions. It is a key piece of equipment in biotechnology development and is the most important and widely used equipment in scientific research and industrial fermentation. Fermentation culture can be considered the heart of the entire fermentation industry; most new biotechnological achievements require cell bioreactors to be transformed into the desired products before production. Cell bioreactors should provide a suitable reaction environment for living cells or microorganisms to achieve cell proliferation or product formation. The structure, operation methods, operating conditions, and production control of cell bioreactors are closely related to product quality, conversion rate, and energy consumption. Establishing new models and optimizing systems for the production control of cell bioreactors is crucial for intelligent and efficient production, effectively improving production efficiency, reducing enterprise costs, and increasing economic benefits.

[0003] In related technologies, the control of cell bioreactors mainly relies on human experience to set the operating parameters. Operators manually adjust the parameters in the PLC (Programmable Logic Controller) based on their understanding of the bioreactor process and past experimental data to control various operating states of the cell bioreactor, such as temperature, pressure, and stirring speed. When facing changes in the production line environment, professional technicians usually need to conduct comprehensive testing and analysis of the equipment and then manually adjust the PLC parameters to adapt to the changes. This process consumes a lot of manpower and time. Furthermore, if personnel change, even with previous operating manuals, it can lead to product instability to some extent. Summary of the Invention

[0004] The purpose of this application is to provide an optimized control method, apparatus, equipment, medium, and product for cell bioreactors.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] In a first aspect, this application provides an optimized control method for a cell bioreactor, comprising:

[0007] The first bioreaction type, the first reaction stage, the first basic configuration parameters of the cell bioreactor, and the first monitoring parameters collected by multiple types of sensors are obtained in the current bioreaction process within the cell bioreactor.

[0008] Retrieve contextual data related to the first biological reaction type and the first reaction stage from a pre-generated biological reaction knowledge vector library;

[0009] Based on the context data, the first monitoring parameters, the first biological reaction type, the first reaction stage, the first basic configuration parameters, the preset data analysis instructions, and the historical question and answer data of monitoring and analysis, prompt words are constructed;

[0010] The prompt word is input into the trained large-scale cell bioreactor model, so that the large-scale cell bioreactor model outputs a first control parameter, which is then used by the programmable logic controller to predict the operating state of the cell bioreactor and correct faults based on the first control instruction corresponding to the first control parameter.

[0011] Secondly, this application provides an optimized control device for a cell bioreactor, comprising:

[0012] The first acquisition module is used to acquire the first bioreaction type, the first reaction stage, the first basic configuration parameters of the cell bioreactor, and the first monitoring parameters collected by multiple types of sensors in the current bioreaction process within the cell bioreactor.

[0013] The retrieval module is used to retrieve contextual data related to the first biological reaction type and the first reaction stage from a pre-generated biological reaction knowledge vector base;

[0014] The construction module is used to construct prompt words based on the context data, the first monitoring parameters, the first biological reaction type, the first reaction stage, the first basic configuration parameters, preset data analysis instructions, and historical question and answer data.

[0015] The output module is used to input the prompt words into the trained large-scale cell bioreactor model, so that the large-scale cell bioreactor model outputs first control parameters, which are then used by the programmable logic controller to predict the operating state of the cell bioreactor and correct faults based on the first control instructions corresponding to the first control parameters.

[0016] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the optimized control method for the cell bioreactor described in any one of the above.

[0017] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the optimized control method for the cell bioreactor described above.

[0018] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the optimized control method for the cell bioreactor described above.

[0019] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0020] This application provides an optimized control method, device, equipment, medium, and product for a cell bioreactor. By acquiring multi-dimensional data of the current bioreaction process within the cell bioreactor, including bioreaction type, reaction stage, basic configuration parameters, and monitoring parameters, and retrieving relevant contextual data from a bioreaction knowledge vector base, it achieves effective integration and rapid retrieval of massive knowledge and real-time data, providing comprehensive data support for subsequent precise analysis and improving data processing efficiency. Based on multi-source data, prompt words are constructed and input into a trained large-scale cell bioreaction model. This fully utilizes the powerful learning and reasoning capabilities of the large model, combining historical question-and-answer data with preset analysis instructions to generate first control parameters that conform to actual operating conditions, improving the intelligence and accuracy of the bioreaction process. This effectively enhances the control precision of the cell bioreactor, optimizes the bioreaction process, and improves product quality and yield. The programmable logic controller (PLC) corrects the reactor state based on the control instructions corresponding to the control parameters output by the large model, forming a closed-loop optimization system of "data acquisition - analysis and decision-making - control execution". When changes occur in the bioreactor process or reactor state, the system can acquire new data in real time and repeat the above process, dynamically adjusting control parameters to ensure the reactor is always in optimal operating condition. This enhances the system's adaptability and robustness to complex operating conditions. The introduction of a bioreactor knowledge vector base enables structured storage and efficient reuse of knowledge in the bioreactor field, avoiding redundant research and development. Simultaneously, with the continuous accumulation of historical question-and-answer data from monitoring and analysis, the large-scale cell bioreactor model can be continuously optimized through incremental learning and other methods, further improving the effectiveness and sophistication of control strategies, reducing manpower and time costs, and providing strong support for the long-term stable operation and process optimization of cell bioreactors. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A schematic flowchart illustrating an optimized control method for a cell bioreactor provided in an embodiment of this application;

[0023] Figure 2 A schematic diagram of the tissue structure of a cell bioreactor training dataset provided in one embodiment of this application;

[0024] Figure 3 A schematic diagram of a large-scale model of cellular biological reactions provided in an embodiment of this application;

[0025] Figure 4 A schematic flowchart illustrating a large-scale model analysis of cellular biological responses provided in an embodiment of this application;

[0026] Figure 5 A schematic diagram of a control system module provided in an embodiment of this application;

[0027] Figure 6 A schematic diagram of a cell bioreactor culture parameter optimization control system based on a large model, provided as an embodiment of this application;

[0028] Figure 7 This is a schematic diagram of the functional modules of an optimized control device for a cell bioreactor provided in one embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] In one exemplary embodiment, an optimized control method for a cell bioreactor is provided. This method is executed by a computer device, specifically, it can be executed by a computer device such as a terminal or server alone, or it can be executed by both a terminal and a server. In this embodiment, such as... Figure 1 As shown, the optimized control method for a cell bioreactor includes steps 102 to 108. Wherein:

[0032] Step 102: Obtain the first bioreaction type, the first reaction stage, the first basic configuration parameters of the cell bioreactor, and the first monitoring parameters collected by multiple types of sensors in the current bioreaction process within the cell bioreactor;

[0033] The first biological reaction type is mainly classified according to the biological entities (such as cells, microorganisms, etc.) involved in the reaction and the reaction mechanism. Common types include fermentation reaction, enzyme-catalyzed reaction, cell culture reaction, and biotransformation reaction. For example, the cell culture reaction can include animal cell culture, plant cell culture, and microbial cell culture. The first reaction stage can include the delayed phase, exponential phase, stationary phase, and death phase. The first basic configuration parameters need to meet the requirements of cell growth and metabolism, product synthesis, and process control. The first basic configuration parameters can include the gas flow rate, number of impellers, impeller design, stirring rate, and power input of the cell bioreactor. The first monitoring parameters are parameters monitored in real time during the current biological reaction process, such as temperature, pressure, pH value, dissolved oxygen, biomass concentration and substrate concentration, volume, and foam.

[0034] It should be noted that data preprocessing can be performed on the first biological reaction type, the first reaction stage, the first basic configuration parameters of the cell bioreactor, and the first monitoring parameters collected by multiple types of sensors using at least one of the following methods: P / V (constant volumetric power input), vvm (constant gas velocity per unit volume), dynamic method Kla (volume oxygen transfer coefficient), DoE (factorial design, RSM (response surface methodology), mixing design, etc.), constant impeller tip velocity, mixing time, PID (Ziegler-Nichols method), IVCC (integral live cell density), metabolic model, and other relevant indicators. Among these, P / V represents the stirring power per unit volume of fermentation broth (kW / m³). 3 The stirring intensity reflects the effect of stirring intensity on oxygen transfer and mixing; vvm is the ratio of the volume of gas introduced into the fermentation broth per minute to the volume of the fermentation broth; the dynamic method Kla is a key parameter for measuring the ability of gas to transfer oxygen into the liquid (unit: h). -1The mass transfer efficiency of the equipment is reflected; DoE is used to evaluate the effects of multiple factors (such as temperature, pH, vvm, and stirring speed) and their interactions on the response value (such as product concentration and cell growth); constant impeller tip speed is used to control the shear force level; PID parameters include proportional (P), integral (I), and derivative (D). The Ziegler-Nichols method tunes the parameters using the critical proportional gain method: gradually increase the proportional gain until the system oscillates, record the critical gain and period, calculate the P, I, and D values ​​according to empirical formulas, and maintain the set value by adjusting the aeration rate or feed rate through PID; IVCC is used to monitor the live cell density in real time based on capacitance, laser diffraction, or near-infrared (NIR) spectroscopy, avoiding offline sampling lag; metabolic models are based on input-output data (such as substrate consumption and product generation) and fitted with kinetic equations (such as the Logistic equation describing cell / bacterial growth).

[0035] Step 104: Retrieve contextual data related to the first biological reaction type and the first reaction stage from the pre-generated biological reaction knowledge vector library;

[0036] The bioreaction knowledge vector library can be a knowledge storage and management system built based on vector representation technology (such as word vectors, knowledge graph embedding, etc.) to store bioreaction-related knowledge in a structured and digital manner, and to support efficient retrieval, reasoning, and application. The construction process of the bioreaction knowledge vector library is as follows: First, the knowledge data of the cell bioreactor is extracted and fragmented. Due to the special nature of the cell bioreactor, this embodiment of the application fragments the data according to each control adjustment of the cell bioreactor. Then, the data is vectorized. This embodiment of the application can use the T5 vectorization model to process the data. Next, the vector data is indexed and stored in the bioreaction knowledge vector library.

[0037] The context data can be data and operational examples similar to the current reaction type (i.e., the first biological reaction type) and the current reaction stage (i.e., the first reaction stage); such as Figure 4As shown, this application embodiment can use a hybrid semantic and keyword retrieval method to filter contextual data: First, BM25 semantic retrieval is performed using Elasticsearch to obtain the BM25 score of the retrieval results; keyword retrieval is performed using Milvus to obtain the distance score of the retrieval results. These two different retrieval methods mine relevant information in the knowledge vector base from different perspectives, and the relevant retrieval results can be merged and ranked: The above two retrieval results are deduplicated, and the two scores are combined using the Reciprocal Rank Fusion (RRF) method to rearrange the retrieval results and obtain the RRF score; Then, the retrieval results are rearranged and filtered: the retrieval results are further processed using the Final Ranking model to filter out the most relevant contextual data, providing accurate knowledge support for subsequent instruction construction.

[0038] Step 106: Based on the context data, the first monitoring parameter, the first biological reaction type, the first reaction stage, the first basic configuration parameter, the preset data analysis instructions, and the historical question and answer data of monitoring and analysis, construct prompt words;

[0039] The data analysis instructions can be pre-defined rules, algorithms, or operational procedures used to guide the analysis of biological response-related data. These instructions typically specify the analysis objectives, methods, and output requirements. Figure 4 As shown, the historical question-and-answer data for monitoring and analysis can be data pre-stored in a cache (Memory) to provide basic information for subsequent processing. The most relevant contextual data retrieved from the bioreaction knowledge vector library is integrated with the first monitoring data of the cell bioreactor, the basic description of the cell bioreactor and reaction (e.g., the first bioreaction type, the first reaction stage, and the first basic configuration parameters), the data analysis instructions of the cell bioreactor, and the historical question-and-answer data for monitoring and analysis of the cell bioreactor (e.g., the first three items) to construct a complete Prompt. The Prompt contains rich and targeted information, providing a foundation for large-scale model analysis of bioreaction process issues.

[0040] Step 108: Input the prompt word into the trained large-scale cell bioreactor model so that the large-scale cell bioreactor model outputs first control parameters, so that the programmable logic controller can predict the operating state of the cell bioreactor and correct faults based on the first control command corresponding to the first control parameters.

[0041] The operational status can include growth status, product status, and harvest status. The operational status prediction can be based on current bioreactor data and model analysis to predict future operational trends, key indicators, or bioreactor processes. This includes bioreactor process prediction, bio-growth status prediction, bio-product prediction, bio-harvest status prediction, equipment operating parameter prediction, and bio-status early warning. Fault correction can be implemented automatically or with manual assistance using model-generated control commands to restore normal operation in response to abnormal states or faults during reactor operation. This includes automatic adjustment of equipment (sensors, actuators, etc.) malfunctions and intervention for abnormal bioreactor processes. Control parameters can also be called control adjustment parameters, fermentation condition control parameters, or PLC parameters. Control commands are also called PLC commands. The large-scale cell bioreactor model is also called a large model or bioreactor large model. The large-scale cell bioreactor model can be a customized large-scale cell bioreactor model generated by applying LLM (Large Language Model) technology to the biomedical field. Figure 4 As shown, the constructed Prompt can be input into the trained large-scale cell bioreactor model. Based on its own training and learning, the large-scale cell bioreactor model performs in-depth analysis and reasoning on the input information. The large-scale cell bioreactor model analyzes the response, thereby analyzing the state of the cell bioreactor and providing the first control parameters. The large-scale cell bioreactor model can output in a structured JSON format according to the requirements of the first control command, so as to facilitate the parsing of subsequent commands.

[0042] Control command parsing is a crucial step in achieving precise and effective control of cell bioreactor operation. Control commands generated by large-scale cell bioreactor models are often based on the analysis of massive amounts of data and the reasoning of complex algorithms. However, these control commands must be accurately parsed and understood to be translated into practically operable steps, thereby enabling precise regulation of key parameters such as temperature, pressure, flow rate, and material addition in the cell bioreactor, ensuring that the bioreactor proceeds in the expected direction. Control command parsing can be divided into three parts: key information extraction, logic verification and integration, and command construction and execution.

[0043] Key Information Extraction: Rule-Based Matching: Instruction types can be identified based on predefined rules. For example, if the text contains words related to "temperature" accompanied by descriptions of actions like "increase" or "decrease," it can be identified as a temperature regulation instruction. Target parameters and values ​​can be matched using regular expressions; for example, for a temperature regulation instruction, a pattern like "[0-9] + degrees Celsius" can be matched to extract the target temperature value. Semantic Understanding Model Assistance: Semantic understanding models from natural language processing (such as BERT based on the Transformer architecture) can be used to perform semantic analysis on control instructions, further accurately extracting key information. Large-scale cellular biological response models can learn the semantic relationships between words in the control instruction text, improving the accuracy of key information extraction, especially when the control instructions are complex or ambiguous.

[0044] Logical Verification and Integration: Condition Reasonableness Check: Logical verification is performed on the condition constraints in the control instructions to check their reasonableness and for any conflicts. For example, it checks whether multiple conditions are mutually exclusive or whether the conditions contradict the actual operating state of the current cell bioreactor. Information Integration: The extracted key information, such as instruction type, target parameters, time information, and condition constraints, is integrated to form a complete instruction parsing result data structure usable by the execution module. For example, it can be stored in dictionary form: {"Instruction Type": "Temperature Adjustment Instruction", "Target Parameter": "Temperature", "Target Value": 38, "Execution Time": "Next 2 Hours", "Condition": "None (or Specific Condition Description)"}

[0045] Instruction construction and execution: such as Figure 4 As shown, structured data can be parsed and constructed into PLC instructions using a template-based approach. These instructions are then converted into control signals or instruction codes that can be recognized and executed by the actual control equipment of the cell bioreactor (such as temperature controllers, material conveying pump controllers, agitator controllers, flow controllers, and other functional controls). The PLC instructions are then sent to the PLC via the Modbus protocol to ensure that the equipment performs operations according to the parsed PLC instructions.

[0046] By implementing steps 102 to 108 above, multi-dimensional data on the current bioreaction process within the cell bioreactor is acquired, including bioreaction type, reaction stage, basic configuration parameters, and monitoring parameters. Relevant contextual data is retrieved from the bioreaction knowledge vector base, enabling effective integration and rapid retrieval of massive amounts of knowledge and real-time data. This provides comprehensive data support for subsequent precise analysis and improves data processing efficiency. Based on multi-source data, prompt words are constructed and input into the trained large-scale cell bioreaction model. This fully utilizes the powerful learning and reasoning capabilities of the large model, combining historical question-and-answer data with preset analysis instructions to generate first control parameters that conform to actual operating conditions. This improves the intelligence and precision of the bioreaction process, effectively enhancing the control accuracy of the cell bioreactor, optimizing the bioreaction process, and increasing product quality and yield. The programmable logic controller (PLC) corrects the reactor state based on the control instructions corresponding to the control parameters output by the large model, forming a closed-loop optimization system of "data acquisition - analysis and decision-making - control execution." When changes occur in the bioreactor process or reactor state, the system can acquire new data in real time and repeat the above process, dynamically adjusting control parameters to ensure the reactor is always in optimal operating condition. This enhances the system's adaptability and robustness to complex operating conditions. The introduction of a bioreactor knowledge vector base enables structured storage and efficient reuse of knowledge in the bioreactor field, avoiding redundant research and development. Simultaneously, with the continuous accumulation of historical question-and-answer data from monitoring and analysis, the large-scale cell bioreactor model can be continuously optimized through incremental learning and other methods, further improving the effectiveness and sophistication of control strategies, reducing manpower and time costs, and providing strong support for the long-term stable operation and process optimization of cell bioreactors.

[0047] In another exemplary embodiment of this application, the optimized control method for a cell bioreactor can be applied to an optimized control system for a cell bioreactor. The optimized control system can consist of a cell bioreactor sensing module, a process modeling module, and a control system module. The cell bioreactor sensing module primarily senses the state of the entire fermentation process of the reactor, including the sensing and calculation of constant or industrially variable factors during fermentation. The process modeling module primarily establishes a large model to correlate the sensed factors of the bioreaction process with the process control parameters, and uses reinforcement learning to fit the experience of experienced experts in adjusting the equipment's control parameters. The control system module primarily implements an intelligent control system for the fermentation process, and achieves encrypted inference and adaptive learning through model deployment. The optimized control method for a cell bioreactor also includes:

[0048] Step 1011: Obtain the second biological reaction type, second reaction stage, second basic configuration parameters of the cell bioreactor, and second monitoring parameters collected by the multi-type sensors for each of the multiple historical biological reaction processes within the cell bioreactor.

[0049] Environmental conditions are crucial to the microbial fermentation process. Accurate monitoring of the fermentation state and maintaining constant conditions are essential for preserving fermentation quality and increasing yield. This embodiment requires pre-determining the second bioreaction type, reaction process data (i.e., the second monitoring parameters), and the second basic configuration parameters of the cell bioreactor. Various types of sensors within the cell bioreactor can monitor historical bioreaction process parameters in real time. The cell bioreactor sensing module is primarily responsible for real-time acquisition of various parameters during the bioreaction process, preliminary processing and storage of the acquired data, and transmission of the data to the control system or host computer via a communication interface for system analysis and operator review.

[0050] The second monitoring data collected by the multi-type sensors can be multimodal data, including image data and text data. The second monitoring parameters can include temperature, pressure, air flow rate, pH value, dissolved oxygen, biomass concentration, substrate concentration (ammonia nitrogen concentration and concentrations of certain key enzymes and products), product concentration, growth rate, substrate consumption rate, product formation rate, stirring speed, liquid level, liquid viscosity, corrosion, redox potential, cell microscopic images, etc. Depending on the hardware of the cell bioreactor, the embodiments of this application include, but are not limited to, the above data.

[0051] The sensor selection in this application embodiment may include: Temperature sensor: High-precision thermocouples or resistance temperature detectors (RTDs) are selected to quickly and accurately measure temperature changes within the cell bioreactor, with a measurement range of 0-60℃ and an accuracy of ±0.1℃. pH sensor: Glass electrode pH sensors are used, offering good stability and accuracy, with a measurement range of 2-12 and an accuracy of ±0.05pH. Dissolved oxygen sensor: Polarographic or fluorescence-based dissolved oxygen sensors are selected to monitor dissolved oxygen concentration in the reactor in real time, with a measurement range of 0-100% saturation and an accuracy of ±2%. Biomass concentration sensor: Biomass sensors using the optical density method are used to monitor biomass growth online, with a measurement range depending on the specific sensor and an accuracy of ±5%. Substrate concentration sensor: For specific substrates, such as glucose and amino acids, near-infrared spectroscopy sensors, LC-MS, and ICP-MS are selected for real-time / offline monitoring, with a measurement range depending on the substrate type and concentration and an accuracy of ±5%. Product concentration sensor: Depending on the properties of the product, an online monitoring system such as a high-performance liquid chromatography (HPLC) analyzer can be selected to accurately measure the product concentration.

[0052] Ultrasonic or capacitive level sensors can accurately measure the liquid level in a cell bioreactor. ORP electrodes can be used to measure the redox potential within the cell bioreactor, reflecting the redox state and playing a crucial role in controlling biological reaction processes. Vibrating viscosity sensors can measure liquid viscosity. High-resolution microscopy can acquire microscopic images of cells, recording dynamic changes in cell morphology, density, and growth status.

[0053] Since the signals output by sensors are typically weak analog signals, they need to be amplified, filtered, and linearized by signal conditioning circuits for subsequent data acquisition and processing. The signal conditioning circuits employ high-precision operational amplifiers and filters to ensure signal accuracy and stability.

[0054] The hardware devices, such as sensors, signal conditioning circuits, data acquisition modules, and communication interfaces, are integrated to form a cell bioreactor sensing module. This module is installed in a suitable location within the cell bioreactor and then wired and connected.

[0055] All sensor output signals in this application embodiment can be transmitted to the lower-level computer, processed by the lower-level computer, transmitted to the industrial control computer via communication, and stored in the industrial control computer's database for subsequent training data and knowledge data construction.

[0056] Step 1012: Based on the second biological reaction type, second reaction stage, second basic configuration parameter and second monitoring parameter corresponding to the multiple historical biological reaction processes, generate a training dataset containing optimal reaction data, manual parameter tuning records and simulated comparative experimental data, wherein the manual parameter tuning records include control parameters corresponding to the second monitoring parameters;

[0057] The cell bioreactor process modeling module in this embodiment mainly performs fitting and learning on complete biological reaction process data, manually tuned parameter data, and logic to enable the reactor to automatically adjust parameters and ensure constant product generation. It mainly consists of three parts: modeling data construction, lightweight large model construction, and end-side inference deployment scheme.

[0058] Data construction for cell bioreactor process modeling: First, identify the scope of model learning, including the design features, reaction types, scale, and operating modes of the cell bioreactor.

[0059] Basic design features of cell bioreactors are provided as prerequisites for model learning, including gas velocity (apparent gas velocity), distributor rate and coverage rate, gas phase mass transfer coefficient, gas distributor design, type and location, number of impellers, impeller design, type and location, fluid forces and properties, probe location, stirring rate, power input, impeller tip linear velocity, number of baffles, reactor baffle configuration, reactor diameter, impeller diameter, gas holdup, bubble diameter, vortex size, bubble residence time, kinetic energy dissipation rate, carbon dioxide stripping time, mixing time, shear rate, and other cell bioreactor design features, such as top and bottom clearance design.

[0060] The main monitoring parameters learned in this application embodiment may include temperature, volume, pressure, air flow rate, pH value, dissolved oxygen, biomass concentration, substrate concentration, product concentration, growth rate, substrate consumption rate, and product formation rate. Among them, the substrate consumption rate and product formation rate are secondary parameters calculated from the process parameters. The secondary parameters include respiratory entropy, oxygen uptake rate, oxygen transfer coefficient, growth rate, and product yield coefficient, which are used as optimization indicators.

[0061] Secondly, the fermentation conditions that can be controlled in a cell bioreactor include temperature, pH, aeration rate, pressure, dissolved oxygen concentration, and feed control. The control principle lies in controlling the intermediate metabolic balance of cells and microorganisms, and guiding them in a direction conducive to product accumulation based on the laws of cell and microbial growth, metabolism, and biosynthesis.

[0062] Finally, the optimization indices for cell bioreactors include respiratory entropy, oxygen uptake rate, oxygen transfer coefficient, growth rate, and product yield coefficient. Among these, growth rate and product yield have the highest weights. By modeling the fermentation process, the relationship between the optimal fermentation reaction conditions and operating strategies is determined to achieve efficient operation of the cell bioreactor and high-yield, high-quality production of the product.

[0063] Training dataset construction: For example, step 1012, "generating a training dataset containing optimal reaction data, manual parameter tuning records, and simulated comparative experimental data based on the second biological reaction type, second reaction stage, second basic configuration parameter, and second monitoring parameter corresponding to the multiple historical biological reaction processes," can be replaced by the following steps 10121 to 10124:

[0064] Step 10121: Select the second biological reaction type, second reaction stage, second basic configuration parameters, and second monitoring parameters corresponding to the complete and optimal reaction process from multiple historical biological reaction processes actually produced by the cell bioreactor as the optimal reaction data;

[0065] Among them, data on the complete and optimal reaction process can be selected from historical data of actual production in cell bioreactors as the first part of the training data.

[0066] Step 10122: Record the status of the second monitoring parameter and the optimization index corresponding to each adjustment of the control parameter during each of the historical biological reaction processes, as well as the change status of the second monitoring parameter after adjustment, and mark the reason for each adjustment of the control parameter and the expected change. The recorded data corresponding to the multiple historical biological reaction processes are used as manual parameter adjustment records.

[0067] This can be achieved by establishing the status of monitoring and optimization parameters each time a technician adjusts the control fermentation conditions, as well as the changes in the monitoring parameters after the adjustment. At the same time, the reasons for each adjustment of the fermentation conditions by the technician and the expected changes can be manually labeled, and the manual parameter adjustment records can be used as the second part of the training data.

[0068] Step 10123: Multiple simulated biological reaction processes in the simulated environment of the cell bioreactor are used as simulated comparative experiments; during the simulated comparative experiments, the state data and final reaction result data of the cell bioreactor in different experimental groups are adjusted, and the adjustment of control parameters and the advantages or disadvantages compared with other experimental groups are recorded in each group of experimental data. The recorded data corresponding to the simulated comparative experiments are used as simulated comparative experimental data.

[0069] This involves establishing comparative experiments within a simulated cell bioreactor environment. During the simulation, the state data of the cell bioreactors for different experimental groups, as well as the final reaction results, are adjusted. For each set of experimental data, the adjustments to the control parameters and the advantages or disadvantages compared to other groups are manually labeled. This simulated comparative experimental data serves as the third part of the training data.

[0070] It should be noted that existing bioreaction process data and optimization strategies can also be obtained from scientific literature databases in relevant fields as the fourth part of the training data, and paired data of adjusting and controlling fermentation condition parameters and PLC instructions can be used as the fifth part of the training data.

[0071] Step 10124: Generate a training dataset based on the optimal response data, the manual parameter tuning records, and the simulated comparative experimental data.

[0072] The first part of the training data in this embodiment primarily allows the model to learn the optimal monitoring and control data states for this type of reaction. Simultaneously, the second and fourth parts of the training data enable the model to understand how to handle various abnormal indicators during the biological reaction process and the corresponding strategies and methods for adjusting control parameters. The third part of the training data uses a reinforcement learning reward model to learn from the experience of human technicians. The fifth part of the training data allows the model to construct control commands for the reactor through a thought process, thereby influencing the reaction process and forming a closed-loop optimization system.

[0073] Data preprocessing involves cleaning and normalizing the data, checking its integrity, and removing outliers and erroneous data. For example, measurements that significantly deviate from the normal range can be identified and processed using statistical methods (based on the mean and standard deviation). Then, data of different dimensions and ranges are normalized using Z-scores to make the data comparable.

[0074] The embodiments of this application aim to enable a large model to learn the potential logic and experience of the cell bioreactor in the reaction process by using the optimal reaction process data of the given cell bioreactor and the control and adjustment parameter data of the technician, and to provide interpretability of the adjustment operation.

[0075] Step 1013: Based on the training dataset, under biological prior constraints, the large model is trained through supervised learning and reinforcement learning to obtain the trained large model of cellular biological response. The large model of cellular biological response integrates multimodal features from the training dataset through a dynamic routing hybrid expert network and a bidirectional attention mechanism to establish the relationship between the second monitoring parameter and the control parameter.

[0076] Lightweight Large Model Construction: The model training parameters in this embodiment are designed as a combination of reaction type, monitoring parameters, and time series data for a cell bioreactor. The output includes state analysis of the cell bioreactor, fermentation condition adjustment parameters, and an explanation of the adjustment strategy; the specific presentation format is as follows: Figure 2 As shown, the amount of training data for the large model constructed in this application embodiment can be the amount of complete second monitoring data obtained in a reasonable number of iterations.

[0077] This application's embodiments modify the structure of a large model and fine-tune it to enable the model to perform state analysis and control parameter adjustment capabilities for cell bioreactors. Based on an interpretive prediction framework (prediction module + interpretation module), and utilizing reaction state analysis proxy and PPO (Proximal Policy Optimization) technology, the large model learns to control and adjust the cell bioreactor in an interpretable manner.

[0078] The interpretive prediction framework designed in this application provides a clear logical structure for the model's decision-making process, organically combining prediction and interpretation. The reaction state analysis agent, acting as a bridge between the large model and the cell bioreactor, is responsible for collecting the cell bioreactor's operational data in real time and converting it into an input format that the model can understand. Simultaneously, it enables the model's output control decision-making process.

[0079] Prediction Module: By learning from and analyzing a large amount of operational data from cell bioreactors, a mapping relationship is established between input features (such as various state parameters of the cell bioreactor, including temperature, pH, dissolved oxygen, substrate concentration, etc.) and output results (such as adjustment values ​​of control parameters, subsequent states of the biological reaction, etc.). Based on this data, the model can learn the operating patterns of the cell bioreactor under different states, thereby predicting future states or control parameters that need to be adjusted.

[0080] The explanation module is primarily responsible for interpreting and explaining the prediction results, providing the basis and process for the model's decision-making. It mines information such as the importance of features and data dependencies within the model, enabling the model to simulate the complex decision-making process and knowledge required by technicians. For example, by analyzing the model's weighting of various input features when predicting feed volume, it explains why the current substrate concentration is a key factor influencing feed volume decisions, and how other factors (such as temperature and pH) work together to affect the final decision.

[0081] Large-scale model structural design of cell bioreactors, such as Figure 3 As shown, the model design incorporates a lightweight approach based on the existing Transformer network structure. The base model uses DeepSeek 7B lightweight backbone network parameters as initialization parameters, while the multimodal module parameters are fine-tuned using random initialization and then trained through structured pruning. The model structure mainly consists of cell feature encoding, RMS Norm normalization, a feedforward neural network, a dynamic routing hybrid expert network, and an attention module.

[0082] Cellular Feature Encoding and Feature Space Alignment: Microscopic Image Sequence I, acquired using a high-resolution microscope, records dynamic changes in cell morphology, density, and growth status. Each image I... t The dimensions are H×W×C, where C is the number of channels in the image data. Sensor data S is acquired in real-time through a multi-sensor array and includes a series of physicochemical parameters, such as S0. t =[pH t DO t Temp t ,...] T , where DO t Temp represents dissolved oxygen concentration. tRepresents temperature, dimension d s Depending on the type of sensor, image and sensor data are interpolated to the same sampling frequency T to unify the time scale.

[0083] The cell feature encoding module is designed for visual feature extraction: an adaptive cell morphology perception encoder.

[0084] For microscopic images, we designed an adaptive cell morphology-aware encoder Φ morph The aim is to extract the spatiotemporal features of cell morphology and structure. The encoder combines multi-scale convolution and morphological attention mechanisms, and the specific process is shown in the following formula (1):

[0085]

[0086] Where, θ morph These are the learnable parameters of the encoder. The encoder first captures features at different spatial scales through multi-scale convolutional kernels, and then introduces a morphological attention mechanism (MorphAttn) to enhance attention to cellular structures. The formula for calculating morphological attention is shown in the following formula (2):

[0087]

[0088] Among them, Q morph K morph V morph These represent the query, key, and value matrices, respectively, generated from the input feature X through a linear transformation; γ is an adjustable scaling factor; d k is the dimension of the key vector, used for normalization. Biospecific mask M bio It is defined as shown in the following formula (3):

[0089]

[0090] Where, d bio (i, j) represents the biological distance between pixels i and j (based on the Euclidean distance of the cell edge or centroid), and λ and σ are hyperparameters that control the mask amplitude and decay rate, respectively. This design allows the model to focus on biologically significant regions (such as cell membranes or division regions) and reduce background noise interference.

[0091] Feature Space Alignment: This invention employs a multimodal fusion approach for input. Specifically, it uses a projection matrix and a neural network to align the image and text feature spaces. The specific formula is as follows:

[0092] Due to visual characteristics and sensor features Typically located in different feature spaces (with different dimensions and semantic distributions), we first map them to a shared d using linear projection. k - 1D space, as shown in formula (4) below:

[0093]

[0094] in: This is the query projection matrix, used to map visual features into query vectors; This is the key projection matrix, used to map sensor features to key vectors; The value projection matrix is ​​used to map sensor features into a value vector; d k To ensure a balance between computational efficiency and expressive power, the shared space dimension is set to 64.

[0095] Similarly, back projection (from sensor to vision) is defined as shown in the following formula (5):

[0096]

[0097] These projection matrices are learned through training to ensure that the two modalities have consistent semantic representations in the shared space.

[0098] After mapping to the shared latent space, a biodynamic-based prior matrix is ​​introduced. This is used to adjust attention weights to conform to the physical and chemical laws of biological reaction systems. For example, during fermentation, cell density (a visual feature) is typically negatively correlated with dissolved oxygen concentration (a sensor feature), while changes in pH may be related to cell metabolic rates. Prior matrix The process is divided into two phases. First, initial values ​​(static estimation) are constructed based on a biodynamic model. Then, real-time data is used to dynamically adjust the values, generating a prior matrix suitable for the current moment. Specifically, in the static estimation, the expected correlations between features are predefined according to the biodynamic model (Monod equation). For example, dissolved oxygen concentration (DO) and cell density (D). cell The relationship can be approximated by the following formula (6):

[0099]

[0100] Where α is the normalization coefficient, and i and j correspond to the indices of the feature dimensions. This reflects a negative correlation between dissolved oxygen concentration and the decrease in dissolved oxygen concentration with increasing cell density. This initial value... It is a static estimate that provides prior relationships between features based on a theoretical model, serving as the starting point for dynamic adjustment.

[0101] Considering the dynamic changes in biological reaction processes (such as logarithmic growth and stationary phases), static initial priors may not be fully adaptable to real-time reaction conditions. Therefore, we take... Based on this, combined with real-time sensor data S t-w：t (The sequence within the time window) and the visual features of the previous moment. Dynamic updates via a lightweight two-layer MLP The final prior matrix is ​​generated as shown in formula (7):

[0102]

[0103] in: The initial static prior matrix is ​​used as one of the inputs; S t-w：t Provides information on environmental changes over time; Introduce the visual context from the previous moment; θ bio It is a learnable parameter of the MLP; Sigmoid ensures that the output value is in the range [0, 1].

[0104] To avoid excessive interference from the prior matrix on the attention distribution, softmax normalization is applied as shown in Equation (8):

[0105]

[0106] Where τ is a temperature parameter that controls the smoothness of the distribution.

[0107] Based on projected features and biodynamic priors, attention weights from vision to sensor are calculated. As shown in formula (9) below:

[0108]

[0109] in: The original attention score represents the similarity between visual features and sensor features; is the scaling factor to avoid numerical instability caused by large dimensions; ⊙ represents element-wise multiplication, incorporating biodynamic priors into the attention distribution; Softmax represents row-wise normalization to ensure that the weight sum is 1.

[0110] Similarly, the reverse attention weights are as shown in formula (10):

[0111]

[0112] The aligned feature representation is generated by using the attention weights to weight the fused value vector, as shown in the following formula (11):

[0113]

[0114] in: For alignment sensor characterization guided by visual features; Aligned visual representations guided by sensor features.

[0115] This bidirectional alignment mechanism ensures that the two modalities are complementary before fusion. For example, visual features may highlight local changes in cell morphology, while sensor features provide global context of environmental conditions; feature alignment is achieved through the above method. After alignment, the two features are first superimposed along the channel dimension to obtain preliminary fused features. Where [·, ·] denotes concatenation along the channel axis, and α is an adjustable weight, typically set to 0.5 to balance the contributions of the two modalities. Subsequently, a simple linear transformation (e.g., 1×1 convolution) is used to reduce the dimensionality of the superimposed features, generating the final comprehensive features.

[0116] Dynamic Routing Hybrid Expert Network: A hybrid expert network consists of multiple "expert networks" and a gating network. Each expert network is an independent neural network, adept at handling specific types of data or tasks. The gating network assigns weights to each expert network based on the input data, and then the outputs of each expert network are weighted and summed to obtain the final output. This structure allows the model to dynamically select appropriate experts to process based on different inputs, thereby improving the model's expressive power and ability to handle complex tasks. The dynamic routing mechanism allows the model to dynamically determine which expert networks to route input data to during inference or training based on the real-time characteristics of the input data. This enables the model to utilize the capabilities of expert networks more efficiently and reduce unnecessary computation. The formula is designed as follows:

[0117] Let the input data be x, and the number of expert networks be K. The gating network calculates the routing weight w = [w1, w2, ..., w3] of each expert network using a function g(x). K ], where wi represents the weights of the input data assigned to the i-th expert network. Common gating networks are implemented using neural networks (such as multilayer perceptrons, MLPs), and finally the output is converted into a probability distribution through the softmax function to ensure that the sum of all weights is 1, as shown in the following formula (12):

[0118]

[0119] Where si is the score of the output of the gating network to the input x after linear transformation, which can be expressed as shown in the following formula (13):

[0120]

[0121] Here, vi is the parameter vector associated with the i-th expert network in the gated network, and MLP(x) is the feature extraction result of the multilayer perceptron on the input x.

[0122] In another exemplary embodiment of this application, step 1013, "based on the training dataset, under biological prior constraints, the large model is trained through supervised learning and reinforcement learning to obtain the trained large model of cellular biological response," can be replaced by the following steps 10131 to 10134:

[0123] Step 10131: Based on the optimal response data, perform initial fine-tuning training on the large model;

[0124] The fine-tuning method for the large-scale cell bioreactor model designed in this application is as follows:

[0125] First, the large-scale cell bioreactor model data is fine-tuned using training data constructed from analysis commands, reactor monitoring parameters, and reaction-related contextual information. The model's initialization parameters are based on the core Transformer Block parameters of the DeepSeek 7B model. The large-scale model is fine-tuned by freezing the core Transformer Block model parameters and training the Feed-Forward Network parameters, enabling it to model the cell bioreactor's data, process analysis, and parameter fine-tuning methods. The initial training of the large-scale model is completed using the first part of the training data constructed from the training data.

[0126] Step 10132: Based on the manually tuned data, the large model is fine-tuned and trained again through supervised learning;

[0127] Secondly, the large model is fine-tuned again using SFT (Supervised Fine-Tuning) to enable the model to learn the analytical logic for anomalies and adjustments in biological reactions. Using the second part of the training dataset constructed from the training dataset, in the parameter diagnosis of cell bioreactors, the model can learn the characteristics and corresponding treatment methods under different monitoring parameter states of the reactor through supervised training, thereby more accurately judging monitoring parameters and analyzing solutions.

[0128] Step 10133: Based on the simulated comparative experimental data, the large model is fine-tuned and trained using reinforcement learning.

[0129] Then, through reinforcement learning on a large-scale cell bioreactor model, the model can simulate a technician making optimal decisions in real time based on the state of the cell bioreactor, achieving automated control and optimized operation of the cell bioreactor. This is accomplished using the third part of the training set data.

[0130] The reinforcement learning framework is constructed as follows: First, define the State, which includes various monitoring parameters of the cell bioreactor, the current reaction stage, historical operation records, etc. This information comprehensively describes the current state of the cell bioreactor. Then, define the Action, which is an adjustment to the cell bioreactor parameters, such as adjusting temperature, flow rate, adding reactants, etc. Next, design the Reward Function, the core of reinforcement learning, which is used to evaluate the merits of each action. In the cell bioreactor, the Reward Function is designed based on factors such as reaction efficiency, product quality, energy consumption, cell morphology, density, and growth state. For example, if an action improves reaction efficiency, enhances product quality, and reduces energy consumption, a higher reward is given; conversely, a lower reward is given. The specific design of the reinforcement learning is as follows:

[0131] The state is a comprehensive description of the current state of the cell bioreactor, represented by the vector S in the following formula (14):

[0132] S = [s1,s2,…,sn]T(14);

[0133] Where si represents the i-th state characteristic, including temperature, pressure, air flow rate, pH, dissolved oxygen, biomass concentration, substrate concentration (ammonia nitrogen concentration and concentrations of certain key enzymes and products), product concentration, growth rate, substrate consumption rate, product formation rate, stirring speed, liquid level, liquid viscosity, corrosion, and redox potential of the cell bioreactor. Where n = 16 is the total number of state characteristics.

[0134] An action is an operation that the model can perform in a specific state, represented by vector A in the following formula (15):

[0135] A = [a1, a2, ..., am]T(15);

[0136] Where aj represents the j-th action variable, including adjusting temperature, pH, ventilation rate, pressure, dissolved oxygen concentration, and feed control. Where m = 6 is the total number of action variables.

[0137] The reward function is used to evaluate the merits of performing action A in state S, and it is a core part of reinforcement learning. It is expressed as R(S,A) in the following formula (16):

[0138] R(S,A)=w1R1(S,A)+w2R2(S,A) (16);

[0139] Where R1(S,A) is the growth rate reward function, R2(S,A) is the product yield reward function, and wi is the corresponding weight.

[0140] The growth of microorganisms in a cell bioreactor follows the logistic growth model, and its growth rate can be expressed as shown in the following formula (17):

[0141]

[0142] Where N is the number of microorganisms, t is time, r is the intrinsic growth rate, and K is the carrying capacity. The growth rate is calculated at discrete time step t as follows: Where Δt is the time interval.

[0143] The reward function R1(S,A) can be defined as: R1(S,A)=α*GRt, where α is a positive constant used to adjust the scale of the reward.

[0144] Assuming that at time step t, the actual output of the product is Pt, and the maximum possible output calculated based on substrate consumption and other theories is Pmax,t, then the product yield Yt = Pmax,tPt; the reward function R2(S,A) can be defined as: R2(S,A) = γ × Yt + δ × H(Yt-θ), where H(·) is a step function, H(Yt-θ) = 1 when Yt-θ > 0, otherwise H(Yt-θ) = 0, and δ is the intensity of the additional reward.

[0145] The Q-learning update formula can be expressed as shown in the following formula (18):

[0146]

[0147] Where: St and At are the state and action at time t, respectively. α is the learning rate, controlling the step size for each update. γ is the discount factor, ranging from [0,1], used to balance the importance of immediate and future rewards. R(St,At) is the immediate reward obtained by performing action At in state St. St+1 is the next state transitioned to after performing action At.

[0148] During training, the following greedy strategy, as shown in formula (19), is typically used to balance exploration and exploitation:

[0149]

[0150] Here, ∈ represents the exploration rate, which is typically set relatively high initially and gradually decreases as training progresses. Through learning, in the bio-fermentation process, the model can automatically adjust parameters such as stirring speed, aeration rate, and feed rate based on parameters like temperature, dissolved oxygen, and pH within the fermenter to improve fermentation efficiency and product quality. Furthermore, the model can continuously optimize itself based on long-term operational data, adapting to different reaction conditions and requirements.

[0151] Step 10134: Perform structured pruning, operator fusion, attention mechanism acceleration, and model quantization on the large model to obtain the trained large model of cellular biological response.

[0152] First, large models undergo structured pruning. Structured pruning is a model compression technique that reduces model size by removing unimportant structural units (such as entire channels, neurons, or layers) from neural networks, while preserving performance as much as possible. In the context of large-scale cell bioreactor models, structured pruning can help reduce computational cost and memory usage, improving inference speed. Removing less contributing units simplifies the model structure with minimal impact on prediction accuracy.

[0153] Secondly, operator fusion is applied to the model, merging multiple consecutive operators into a single operator to reduce memory access and computational overhead. Merging convolution and activation functions reduces the storage of intermediate results and the number of memory accesses, thereby improving computational efficiency. For large-scale cell bioreactor models, operator fusion can significantly accelerate model training and inference.

[0154] Then, Flash Attention is invoked for acceleration. Flash Attention is an efficient attention mechanism that significantly reduces the memory footprint and computation time of attention calculations by optimizing memory access patterns and computation processes.

[0155] Finally, the model is quantized to FP16 for processing. Quantizing the model to FP16 reduces memory usage and computational load, improves inference speed, and maintains model performance to some extent. Large cell bioreactor models can infer faster on the AGK orin device.

[0156] In another exemplary embodiment of this application, the optimized control method for the cell bioreactor further includes:

[0157] Step 10135: Divide the parameter matrix of the trained large-scale cell biological response model into blocks according to preset rules to obtain multiple data blocks;

[0158] Step 10136: Encrypt each data block using a preset encryption algorithm to obtain the encrypted large-scale model of cellular biological response;

[0159] Step 10137: Deploy the encrypted large-scale cellular biological response model to a local edge computing device.

[0160] After model training is complete, the parameter matrix of the trained large-scale cellular biological response model can be divided into blocks according to certain rules. Each data block is then encrypted using the AES-256 algorithm to generate an encrypted model file. A pair of RSA keys (public and private keys) can be generated. The private key is kept securely by the model owner, while the public key can be distributed to authorized devices or users who need to use the model. The symmetric encryption key (AES-256 key), after generation, is encrypted using the RSA public key and transmitted to the authorized device or user. When using the model, the authorized device or user uses the locally stored RSA private key to decrypt the model to obtain the symmetric encryption key, and then uses this key to decrypt the model file to obtain the original model parameters.

[0161] In another exemplary embodiment of this application, the sensor includes a first sensor and other sensors, the first monitoring parameter includes a first sub-monitoring parameter collected by the first sensor and a second sub-monitoring parameter collected by the other sensors, and the optimized control method of the cell bioreactor further includes:

[0162] Step 1031: Identify anomalies in the first monitoring parameter by at least one of multi-sensor data cross-validation and historical data trend analysis;

[0163] The system monitors the data transmitted by the sensors in real time. Under normal operating conditions, the sensor values ​​within the cell bioreactor will remain within a reasonable range, which can be determined through historical data analysis, theoretical requirements of the bioreactor, and practical production experience. The system continuously compares the sensor data with this reasonable range; if data is found to exceed this range, it will be immediately marked as abnormal data, triggering a fault-tolerant processing procedure.

[0164] Cellular bioreactors are typically equipped with multiple related sensors, such as temperature sensors, pressure sensors, and pH sensors. There is a certain correlation between the data from these sensors and specific sensor values ​​(such as dissolved oxygen levels). For example, under certain temperature and pressure conditions, there is a theoretical value for dissolved oxygen solubility within the cellular bioreactor. A large model collects data from these related sensors and uses a pre-trained correlation model to calculate the theoretical range of dissolved oxygen values ​​under the current conditions. If the dissolved oxygen sensor data exceeds this theoretical range, while the data from other related sensors are normal, then it is highly likely that the dissolved oxygen sensor is malfunctioning.

[0165] By comparing historical data with similar industrial control systems, including times when parameters such as bioreactor stage, temperature, pressure, and stirring speed are similar to the current operating conditions, the dissolved oxygen values ​​under these similar conditions are analyzed. If the current dissolved oxygen sensor data differs significantly from historical data under similar conditions, the analysis is performed using historical operational data of the cell bioreactor, including the trend of sensor values ​​over time. When sensor data anomalies occur, the large model analyzes the historical data trends before and after that point. If the current dissolved oxygen sensor data differs significantly from historical data under these similar conditions, the anomaly is likely due to sensor malfunction.

[0166] This application embodiment can use the results of cross-validation and / or analysis of historical data to make a final determination of whether the dissolved oxygen sensor is faulty. If both cross-validation and historical data analysis point to a high probability of sensor failure, then it can be determined that the dissolved oxygen sensor has a problem.

[0167] Step 1032: In the case of the abnormality indicating a failure of the first sensor, if there is a backup sensor for the first sensor, then switch to the backup sensor to collect the first sub-monitoring parameter; if there is no backup sensor for the first sensor, then estimate the first sub-monitoring parameter of the first sensor based on the second sub-monitoring parameter collected by the other sensors and the historical parameter collected by any sensor.

[0168] In system operation decisions, if redundant dissolved oxygen sensors exist and their data has been verified to be reliable, the large-scale model will automatically switch to using data from the backup sensors for subsequent control and monitoring. If no backup sensors are available, or if the backup sensors may be faulty, the large-scale model can estimate the current sensor value using its internal model based on data from relevant sensors and historical data. This estimate can temporarily replace the data from the faulty sensor to maintain the normal operation of the cell bioreactor.

[0169] Meanwhile, during sensor malfunctions, the large model will adjust the control strategy of the cell bioreactor appropriately based on the estimated dissolved oxygen value or other relevant parameters to ensure the stability and safety of the biological reaction.

[0170] In another exemplary embodiment of this application, the optimized control method for the cell bioreactor further includes:

[0171] Step 109: Generate a first rule vector based on the first monitoring parameter and the first control parameter;

[0172] Step 110: Add the first rule vector to the biological reaction knowledge vector library.

[0173] The construction process of the bioreactor knowledge vector base includes: First, extracting and fragmenting reactor knowledge data. Due to the special nature of cell bioreactors, this embodiment fragments the data according to each control adjustment of the reactor. Then, the data is vectorized. This embodiment uses the T5 vectorization model to process the data. Next, an index is created on the vector data and stored in the vector knowledge base. This embodiment designs a reasoning method based on the knowledge vector base and a large model, as follows: Figure 4 As shown.

[0174] Edge-side inference deployment scheme:

[0175] This application's embodiments employ a deployment scheme combining knowledge vector base retrieval with a large-scale cell bioreactor model. The knowledge vector base is deployed on an industrial control computer, while the large-scale model is deployed for inference on the AGK orin computing device of the lower-level machine. The specific architecture is as follows: Figure 5 The system architecture diagram is shown below.

[0176] In this embodiment, when the production line environment changes, such as equipment wear or the introduction of new material characteristics, the system collects new operational data and feeds this data back to the large model in the intelligent control module. By extracting the required knowledge structure from the knowledge vector base, the system can quickly add the new, adjusted knowledge to the knowledge vector base. When the new knowledge accumulates to a certain amount, the model can perform incremental learning based on this new data, fine-tuning its own parameters. This enables the large bioreactor model to output control commands that are more adapted to the new environment, controlling the execution state of the cell bioreactor. This achieves incremental learning and parameter fine-tuning of the model, adapting to changes in the production line environment (such as equipment wear or new material characteristics).

[0177] In this embodiment, technicians can interact with the system via voice. They can use voice to set system parameters, switch functions on and off, inquire about system status, results and product predictions, and answer questions related to the large-scale language model of cell bioreactors. The system's host computer has a built-in speech recognition module that can recognize and convert the technician's voice commands into text, and respond to questions and answers using a large-scale language model.

[0178] The overall design of the control system module in this embodiment is as follows: A distributed control system is adopted to automatically control various parameters in the cell bioreactor via a fieldbus. The entire control system module includes sensors, field instruments, lower-level microcontrollers or PLCs, a host computer, communication interfaces, actuators, etc.

[0179] The data acquisition and processing module in this embodiment uses an industrial-grade data acquisition card with multiple analog and digital input / output channels, capable of simultaneously acquiring data from various sensors. Its sampling frequency is no less than 10Hz. The data acquisition card converts the analog signals from the sensors into digital signals and transmits them to the central control unit. The central control unit can be a compute stick or compute card (such as the Jetson AGX Orin Developer kit 64G), possessing powerful data processing capabilities and stable operating performance, providing lightweight large-model analysis and processing capabilities. It runs the real-time operating system Linux-RT, ensuring rapid processing of sensor data and timely control of actuators. The knowledge vector database uses a high-performance industrial computer to ensure rapid execution of retrieval tasks and communicates with the compute stick or compute card device via a bus.

[0180] In this embodiment, the communication module uses Modbus-RTU fieldbus technology for internal communication between the sensors, actuators, and data acquisition module, which features strong anti-interference capability, high communication speed, and high reliability. For external communication, the control system communicates with the host computer or remote monitoring center via Ethernet. The TCP / IP protocol is used for remote data transmission, facilitating remote monitoring and management by operators.

[0181] In this embodiment, a PLC controller can be used as the core of the cell bioreactor control system. This controller has advantages such as high performance, high reliability, and good scalability, and can meet the complex control requirements of the cell bioreactor. To achieve precise control of the cell bioreactor's rotation speed and aeration rate, a frequency converter, a mass flow meter, and a sensor module are integrated. The frequency converter is used to adjust the rotation speed of the stirring motor, achieving continuous adjustment of the rotation speed within a set range by changing the motor's input frequency. The mass flow meter is used to accurately measure the gas flow rate entering the cell bioreactor, ensuring the accuracy of the aeration rate. The sensor module collects various parameters inside the cell bioreactor in real time, such as temperature, pH value, and dissolved oxygen, and transmits this data to the PLC controller, providing real-time feedback information to the control system. Through the coordinated work of these hardware devices, precise adjustment of the cell bioreactor's rotation speed and aeration rate is achieved.

[0182] In the software portion of this application embodiment, the monitoring software employs a graphical interface design or DEMO monitoring to display the values ​​and change curves of various parameters (temperature, pressure, pH, dissolved oxygen, flow rate, etc.) of the cell bioreactor in real time. It provides intuitive operation buttons and menus for convenient parameter setting and control mode switching. The alarm software sets alarm thresholds; when parameters exceed normal ranges, it promptly issues audible and visual alarms and displays alarm information on the interface, including alarm parameters, alarm time, and alarm type (e.g., upper limit alarm, lower limit alarm). Operators can view historical alarm records and analyze the causes of alarms through the alarm interface.

[0183] The anti-interference design of this application includes setting software traps in the program to capture abnormal jumps during program execution, enabling the program to automatically resume normal operation. A watchdog timer is used to periodically check the program's running status. When the program crashes or malfunctions, the watchdog timer triggers a system reset, restarting the program or resuming seamless switching to traditional PID control.

[0184] Simultaneously, an anomaly handling method for cell bioreactors was incorporated into the large-scale model processing. When the system detects an anomaly in a cell bioreactor sensor, it triggers a system alarm through cross-validation. Simultaneously, a prompt command allows the large-scale model to automatically adapt to anomalies reported by specific sensors. In the absence of backup sensor data or when backup sensors are also faulty, the large-scale model can estimate the current dissolved oxygen level using relevant sensor data and historical data, leveraging its internal model. This estimate can temporarily replace the data from the faulty sensor to maintain the normal operation of the cell bioreactor. This process continues until the reaction concludes, after which technicians perform maintenance.

[0185] Adjusting the control strategy: During sensor malfunctions, the large model will adjust the control strategy of the cell bioreactor appropriately based on the estimated dissolved oxygen level or other relevant parameters to ensure the stability and safety of the biological reaction. For example, parameters such as aeration rate and stirring speed may be adjusted to maintain the normal operation of the biological reaction.

[0186] Meanwhile, if the system fails to detect an anomaly, it can still issue specific early warning instructions to the large model through manual intervention or directly intervene through the control system.

[0187] Compared to related technologies that rely on manual experience to set parameters, this application can automatically and accurately optimize cell bioreactor parameters, improving reaction efficiency and quality stability. In terms of responding to changes in the production line environment, the incremental learning and parameter fine-tuning mechanism can respond quickly without a lot of manual intervention, ensuring production continuity and product consistency. The locally encrypted deployment model avoids the data transmission delay and network instability problems of cloud deployment, while enhancing data security.

[0188] Compared to cell bioreactor modeling techniques in related technologies, the embodiments of this application introduce large models and knowledge vector bases, which can handle the complex nonlinear relationships in cell bioreactor reaction processes, and through reinforcement learning, make control analysis and decision-making more in line with the technician's optimal decisions, and better adapt to various processes.

[0189] This application embodiment utilizes the nonlinear modeling capability of large models and reinforcement learning to achieve process modeling and intelligent parameter optimization of cell bioreactors, enabling them to automatically sense internal states and output optimal control commands that are closer to those of professional operators.

[0190] The embodiments of this application are based on a knowledge vector database to achieve incremental learning and parameter fine-tuning of the model, enabling rapid adaptation to changes in the cell bioreactor production line environment;

[0191] This application demonstrates a lightweight and locally encrypted deployment of large-scale models, with edge computing and PLC working together to improve data processing efficiency and security.

[0192] The embodiments of this application include a fault-tolerance mechanism for a small number of sensor anomalies, which includes alarms for sensor anomalies and tolerance processing for sensor anomalies, thereby reducing the occurrence of errors in the cell bioreactor and improving the stability and yield of the cell bioreactor.

[0193] A successful bioreactor fermentation process depends on a specific environment for cell / microbial growth and product formation. Microorganisms are particularly sensitive to environmental conditions, and they possess multiple metabolic pathways; changes in environmental conditions can easily alter microbial metabolism. Therefore, correctly mastering and controlling fermentation reaction conditions is of great significance for improving fermentation yield.

[0194] like Figure 6As shown in the embodiments of this application, a method for optimizing and controlling cell bioreactor culture parameters based on a large model is proposed. This method utilizes the nonlinear modeling capabilities of the large model to achieve intelligent dynamic control of the cell bioreactor. First, multiple sensors are installed inside the cell bioreactor to collect real-time reactor detection data across various modes. This data is transmitted in real-time to the intelligent control system, which collects optimal state data and manual adjustment data for the entire cell bioreactor process, and manually annotates the reasons for each adjustment. Second, the large model is fine-tuned using the annotated training data to obtain a large bioreactor model. The powerful nonlinear modeling capabilities of the large model are used to model the relationship between sensor parameters and control parameters. Simultaneously, basic bioreactor knowledge and collected data are compiled into a vector knowledge base to provide contextual knowledge for subsequent large model analysis. Then, the intelligent control system senses the internal process state of the cell bioreactor. Through the edge large model and relevant knowledge retrieved from the knowledge vector base, a comprehensive analysis of the cell bioreactor's detection data is performed to determine whether it deviates from the optimal state of the specified fermentation process, and a control parameter adjustment and analysis process is provided. The PLC instructions for the cell bioreactor are generated by a large model, and the state of the cell bioreactor is corrected by the control system. Finally, the corrected state of the reactor is verified, and instructions are given for the optimal release time and alarms for abnormal phenomena.

[0195] This application's embodiments mainly consist of a cell bioreactor sensing module, a process modeling module, and a control system module. The cell bioreactor sensing module primarily senses the state of the entire fermentation process in the reactor, including the sensing and calculation of various constant or industrially variable factors during fermentation. The process modeling module mainly establishes a large model to map the sensing factors of the bioreaction process to the process control parameters, and uses reinforcement learning to fit the experience of experienced experts in adjusting the equipment's control parameters. The control system module mainly implements an intelligent control system for the fermentation process, and achieves encrypted inference and adaptive learning for model deployment.

[0196] Based on the same inventive concept, this application also provides an optimization control device for a cell bioreactor used to implement the above-described optimization control method for the cell bioreactor. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the optimization control device for cell bioreactors provided below can be found in the limitations of the optimization control method for cell bioreactors described above, and will not be repeated here.

[0197] In one exemplary embodiment, such as Figure 7 As shown, an optimized control device 200 for a cell bioreactor is provided, comprising:

[0198] The first acquisition module 201 is used to acquire the first bioreaction type, the first reaction stage, the first basic configuration parameters of the cell bioreactor, and the first monitoring parameters collected by multiple types of sensors in the current bioreaction process within the cell bioreactor.

[0199] The retrieval module 202 is used to retrieve contextual data related to the first biological reaction type and the first reaction stage from a pre-generated biological reaction knowledge vector library;

[0200] The construction module 203 is used to construct prompt words based on the context data, the first monitoring parameters, the first biological reaction type, the first reaction stage, the first basic configuration parameters, preset data analysis instructions, and historical question and answer data.

[0201] The output module 204 is used to input the prompt word into the trained large-scale cell bioreactor model so that the large-scale cell bioreactor model outputs first control parameters, which are then used by the programmable logic controller to predict the operating state of the cell bioreactor and correct faults based on the first control command corresponding to the first control parameters.

[0202] As an optional implementation, the device further includes: a second acquisition module, used to acquire, in multiple historical biological reaction processes within the cell bioreactor, the second biological reaction type, the second reaction stage, the second basic configuration parameters of the cell bioreactor, and the second monitoring parameters collected by the multi-type sensors; a first generation module, used to generate a training dataset containing optimal reaction data, manual parameter tuning records, and simulated comparative experimental data based on the second biological reaction type, second reaction stage, second basic configuration parameters, and second monitoring parameters corresponding to the multiple historical biological reaction processes, wherein the manual parameter tuning records include control parameters corresponding to the second monitoring parameters; and a training module, used to train a large model based on the training dataset under biological prior constraints through supervised learning and reinforcement learning to obtain the trained large model of cell biological reaction, wherein the large model of cell biological reaction integrates multimodal features in the training dataset through a dynamic routing hybrid expert network and a bidirectional attention mechanism to establish the relationship between the second monitoring parameters and the control parameters.

[0203] As an optional implementation, the training module includes: a first training submodule for initial fine-tuning training of the large model based on the optimal response data; a second training submodule for further fine-tuning training of the large model using supervised learning based on the manually tuned parameter data; a third training submodule for final fine-tuning training of the large model using reinforcement learning based on the simulated comparative experimental data; and a processing submodule for performing structured pruning, operator fusion, attention mechanism acceleration, and model quantization on the large model to obtain the trained large model of cellular biological response.

[0204] As an optional implementation, the first generation module includes: a first generation submodule, used to select the second bioreaction type, second reaction stage, second basic configuration parameter, and second monitoring parameter corresponding to the complete and optimal reaction process from multiple historical bioreaction processes in the actual production of the cell bioreactor as optimal reaction data; a second generation submodule, used to record the status of the second monitoring parameter and optimization index corresponding to each adjustment of control parameters in each historical bioreaction process, as well as the change status of the second monitoring parameter after adjustment, and to mark the reason and expected change of each adjustment of control parameters, and to use the recorded data corresponding to the multiple historical bioreaction processes as manual parameter tuning records; a third generation submodule, used to use multiple simulated bioreaction processes in the simulated environment of the cell bioreactor as simulated comparative experiments; in the simulated comparative experiments, adjusting the status data of the cell bioreactor in different experimental groups and the final reaction result data, recording the adjustment of control parameters and the advantages or disadvantages compared with other experimental groups in each group of experimental data, and using the recorded data corresponding to the simulated comparative experiments as simulated comparative experimental data; a fourth generation submodule, used to generate a training dataset based on the optimal reaction data, the manual parameter tuning records, and the simulated comparative experimental data.

[0205] As an optional implementation, the device further includes: a block segmentation module, used to segment the parameter matrix of the trained large-scale cell biological response model into multiple data blocks according to a preset rule; an encryption module, used to encrypt each data block using a preset encryption algorithm to obtain an encrypted large-scale cell biological response model; and a deployment module, used to deploy the encrypted large-scale cell biological response model to a local edge computing device.

[0206] As an optional implementation, the sensor includes a first sensor and other sensors, and the first monitoring parameter includes a first sub-monitoring parameter collected by the first sensor and a second sub-monitoring parameter collected by the other sensors. The device further includes: an identification module, used to identify anomalies in the first monitoring parameter through at least one of multi-sensor data cross-validation and historical data trend analysis; and a switching module, used to switch to the backup sensor to collect the first sub-monitoring parameter if a backup sensor exists when the anomaly indicates a fault in the first sensor, and to estimate the first sub-monitoring parameter of the first sensor based on the second sub-monitoring parameter collected by the other sensors and historical parameters collected by any sensor if no backup sensor exists.

[0207] As an optional implementation, the device further includes: a second generation module, used to generate a first rule vector based on the first monitoring parameter and the first control parameter; and an adding module, used to add the first rule vector to the biological reaction knowledge vector base.

[0208] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0209] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0210] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

Claims

1. A method for optimal control of a cell bioreactor, characterized in that, The optimization control method of the cell bioreactor comprises: acquiring a first biological reaction type, a first reaction stage, first basic configuration parameters of the cell bioreactor, and first monitoring parameters collected by a multi-type sensor of a current biological reaction process in the cell bioreactor; retrieving context data related to the first biological reaction type and the first reaction stage in a pre-generated biological reaction knowledge vector library; constructing a prompt word based on the context data, the first monitoring parameters, the first biological reaction type, the first reaction stage, the first basic configuration parameters, a preset data analysis instruction, and monitoring analysis historical question and answer data; inputting the prompt word into a trained cell biological reaction large model to make the cell biological large model output first control parameters, so that a programmable logic controller predicts the running state of the cell bioreactor based on first control instructions corresponding to the first control parameters and corrects faults.

2. The method of optimizing control of a cell bioreactor according to claim 1, wherein, The optimization control method of the cell bioreactor further comprises: acquiring a second biological reaction type, a second reaction stage, second basic configuration parameters of the cell bioreactor, and second monitoring parameters collected by the multi-type sensor in each historical biological reaction process of a plurality of historical biological reaction processes in the cell bioreactor; generating a training data set containing optimal reaction data, artificial parameter adjustment records, and simulation comparison experiment data based on the second biological reaction type, the second reaction stage, the second basic configuration parameters, and the second monitoring parameters corresponding to the plurality of historical biological reaction processes, wherein the artificial parameter adjustment records include control parameters corresponding to the second monitoring parameters; training a large model based on the training data set under biological prior constraints through supervised learning and reinforcement learning to obtain the trained cell biological reaction large model, wherein the cell biological reaction large model fuses multi-modal features in the training data set through dynamic routing hybrid expert network and bidirectional attention mechanism to establish a relationship between the second monitoring parameters and the control parameters.

3. The method of optimizing control of a cell bioreactor according to claim 2, wherein, The training of the large model based on the training data set under biological prior constraints through supervised learning and reinforcement learning to obtain the trained cell biological reaction large model comprises: initially fine-tuning the large model based on the optimal reaction data; re-fine-tuning the large model through supervised learning based on the artificial parameter adjustment data; finally fine-tuning the large model through reinforcement learning based on the simulation comparison experiment data; performing structured pruning processing, operator fusion, attention mechanism acceleration, and model quantization on the large model to obtain the trained cell biological reaction large model.

4. The method of claim 2, wherein the optimization control of the cell bioreactor is performed by a computer program. The generation of the training data set containing optimal reaction data, artificial parameter adjustment records, and simulation comparison experiment data based on the second biological reaction type, the second reaction stage, the second basic configuration parameters, and the second monitoring parameters corresponding to the plurality of historical biological reaction processes comprises: screening a second bioreaction type, a second reaction stage, a second basic configuration parameter and a second monitoring parameter corresponding to a complete and optimal reaction process from a plurality of historical bioreaction processes of actual production of the cell bioreactor as optimal reaction data; recording the second monitoring parameter and the optimization index state corresponding to each adjustment of the control parameter in each of the historical bioreaction processes, the change state of the second monitoring parameter after the adjustment, and labeling the reason and expected change of each adjustment of the control parameter, and taking the recording data corresponding to the plurality of historical bioreaction processes as artificial parameter adjustment records; taking a plurality of simulated bioreaction processes of a simulated environment of the cell bioreactor as a simulated comparative experiment; in the simulated comparative experiment, adjusting the state data and the final reaction result data of the cell bioreactor in different experimental groups, recording the adjustment of the control parameter in each group of experimental data and the advantages or disadvantages compared with other experimental groups, and taking the recording data corresponding to the simulated comparative experiment as simulated comparative experiment data; generating a training data set based on the optimal reaction data, the artificial parameter adjustment records and the simulated comparative experiment data.

5. The method of claim 2, wherein the optimization control of the cell bioreactor is performed by a computer program. The optimization control method of the cell bioreactor further comprises: blocking the parameter matrix of the trained cell bioreaction large model according to a preset rule to obtain a plurality of data blocks; encrypting each of the data blocks using a preset encryption algorithm to obtain an encrypted cell bioreaction large model; deploying the encrypted cell bioreaction large model to a local edge computing device.

6. The method of claim 1, wherein the method is performed in a bioreactor. The sensor comprises a first sensor and other sensors, the first monitoring parameter comprises a first sub-monitoring parameter collected by the first sensor and a second sub-monitoring parameter collected by the other sensors, and the optimization control method of the cell bioreactor further comprises: performing abnormal identification on the first monitoring parameter through at least one of multi-sensor data cross-validation and historical data trend analysis; in the case that the abnormality indicates a failure of the first sensor, if the first sensor has a backup sensor, switching to the backup sensor to collect the first sub-monitoring parameter, and if the first sensor has no backup sensor, estimating the first sub-monitoring parameter of the first sensor based on the second sub-monitoring parameter collected by the other sensors and historical parameters collected by any sensor.

7. The method of claim 1, wherein the method is performed in a cell bioreactor. The optimization control method of the cell bioreactor further comprises: generating a first rule vector based on the first monitoring parameter and the first control parameter; adding the first rule vector to the bioreaction knowledge vector library.

8. An apparatus for optimized control of a cell bioreactor, characterized in that, The optimization control device of the cell bioreactor comprises: a first acquisition module configured to acquire a first bioreaction type, a first reaction stage, first basic configuration parameters of a cell bioreactor, and first monitoring parameters collected by a plurality of types of sensors in a current bioreaction process of the cell bioreactor; a retrieval module configured to retrieve context data related to the first bioreaction type and the first reaction stage from a pre-generated bioreaction knowledge vector library; The construction module is configured to construct a prompt word based on the context data, the first monitoring parameter, the first biological reaction type, the first reaction stage, the first basic configuration parameter, preset data analysis instructions, and monitoring analysis history question and answer data. An output module is configured to input the prompt word into a trained cell biological reaction large model, so that the cell biological large model outputs a first control parameter, and a programmable logic controller predicts an operation state of the cell biological reactor based on a first control instruction corresponding to the first control parameter and corrects a fault.

9. A computer device comprising: A memory and a processor to store a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the optimization control method of the cell biological reactor according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the optimization control method of the cell biological reactor according to any one of claims 1-7.

Images