An intelligent management and control platform for biopharmaceutical fermentation process based on industrial internet

Abstract

The application relates to the field of industrial internet, and discloses a bio-pharmaceutical fermentation process intelligent management and control platform based on industrial internet, which aims to solve the technical problems that the fermentation process state is unknown and the control decision lags behind the reaction dynamic change; the platform comprises a device layer, an edge layer and a platform layer; the device layer is used for fermentation data acquisition; the edge layer is used for data preprocessing and abnormality detection; and the platform layer is used for multi-source data fusion, fermentation process digital twinning and intelligent decision. The fermentation process digital twinning realizes fermentation state real-time simulation and trend prediction by error correction hybrid modeling of a mechanism model and a data-driven model; and the intelligent decision generates an optimal control strategy by using reinforcement learning based on the prediction result of the digital twinning. The application realizes a complete closed loop from state perception, trend prediction to intelligent decision and automatic execution, changes the traditional passive response type control into active forward-looking intelligent control, and improves the intelligent level of the fermentation process and product consistency.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and in particular to an intelligent control platform for biopharmaceutical fermentation processes based on industrial internet. Background Technology

[0002] Fermentation is a core link in the biopharmaceutical industry chain, and its control level directly determines the quality stability and production cost of the drug. Fermentation is essentially a complex biochemical reaction system involving the regulation of microbial metabolism, requiring coordinated control of multiple parameters such as temperature, pH, dissolved oxygen concentration, stirring speed, tank pressure, feed rate, and exhaust gas composition within strict process parameter boundaries. Microbial growth and metabolism exhibit highly nonlinear and time-varying characteristics, and the various process parameters are coupled and influence each other, forming a complex kinetic system.

[0003] In traditional fermentation process control models, engineers primarily rely on accumulated personal experience and trial-and-error methods for process regulation. Due to the complexity of microbial metabolic networks and the strong coupling between parameters, the intrinsic mechanisms of the process remain largely "black boxes," making it difficult for operators to accurately grasp the specific impact of parameter adjustments on the final product. More significantly, the industry commonly uses offline sampling and testing methods to obtain key quality indicators of the fermentation broth, such as potency, residual sugar content, amino acid concentration, and cell concentration. The parameters measured during fermentation are limited to certain physical and chemical parameters, while mature online measurement technologies are currently lacking for the more complex biological parameters—the most critical ones. The results of these key indicators typically take several hours or even longer to reach the production site. During this delay, substantial changes in the biochemical reactions within the fermenter occur, resulting in a reactive and reactive approach to process control decisions.

[0004] To address the challenges of online measurement of key biological parameters in fermentation processes and the significant lag in offline detection results, those skilled in the art have undertaken numerous explorations. Chinese invention patent CN116052786B discloses a soft measurement method and controller for key parameters in the fermentation process of marine alkaline protease. This method achieves real-time estimation of key parameters by acquiring process data online through measuring instruments, utilizes a bidirectional long short-term memory network to build a soft measurement model, combines an immediate learning strategy to establish a local model to replace the traditional global model, and employs a multi-head self-attention mechanism to capture the correlation between data. This method, to some extent, solves the problem of lag in manual offline sampling and detection, enabling prediction results to better reflect the temporal sequence of data. Chinese invention patent CN107589669B also proposes an intelligent measurement and control method and system for fermentation processes utilizing measurement lag information. By introducing a soft measurement strategy for the fermentation process based on a covariance fusion algorithm, it achieves the acquisition and utilization of measurement lag biomass information, and on this basis, optimizes and regulates the fermentation process.

[0005] However, the above-mentioned existing technical solutions still have the following shortcomings: Soft measurement methods mainly rely on pure data-driven models (such as BiLSTM, support vector regression, etc.). Although these models have a certain prediction accuracy, they lack a description of the intrinsic biochemical mechanism of the fermentation process, resulting in poor interpretability and a significant decrease in prediction reliability when the operating conditions deviate from the range of historical data. Existing solutions mostly focus on soft measurement estimation of a single key parameter (such as cell concentration and product concentration), failing to systematically integrate soft measurement prediction with digital twin simulation of the fermentation process and optimization of control strategies, making it difficult to achieve a complete closed loop from "state perception" to "optimization decision-making" and then to "automatic execution". The "black box" state of the fermentation process is not only reflected in the difficulty of measuring key parameters online, but also in the lack of forward-looking prediction capabilities for the future evolution trend of the fermentation process. Existing solutions can only estimate the state at the current moment and cannot predict key process turning points in advance (such as the cell growth entering the stable period, the product synthesis rate starting to decline, and the nutrients about to be exhausted). As a result, the control decision still lags behind the dynamic changes of the process. Due to the lack of a unified system architecture, the soft measurement module, simulation module and control module are isolated from each other, and data cannot be connected, making it difficult to form a complete data link from data acquisition, state estimation, trend prediction to the generation of optimal control strategy. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide an intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet, in order to solve the long-standing dilemma of traditional fermentation control being trapped in "estimating the state based on experience and then passively making corrections after waiting for several hours to measure the results." The root cause is that it is impossible to know the real process status inside the fermenter in real time (black box), which leads to the problem that every control decision lags behind the dynamic changes of the reaction itself.

[0007] This invention provides an intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet, comprising a device layer, an edge layer, and a platform layer, wherein: The equipment layer is located in the biopharmaceutical fermentation workshop and is used to realize the physical connection and data acquisition of fermenters, seed tanks, sensors and actuators, and transmit the acquired data to the edge layer through an industrial intelligent gateway; The edge layer is set up in the workshop and is used to receive raw data transmitted from the equipment layer, perform real-time data preprocessing, anomaly detection and emergency control, and upload the preprocessed data to the platform layer. The platform layer is located in the enterprise data center or cloud server and is used to receive pre-processed data transmitted from the edge layer, build the core functions of the industrial internet platform, and realize multi-source data fusion, digital twin of fermentation process, intelligent decision-making, intelligent control execution, blockchain traceability and predictive maintenance. The digital twin of the fermentation process includes: Mechanistic models were constructed based on microbial growth kinetics equations and metabolic network models. Train a data-driven model using historical batch fermentation data; The mechanistic model and the data-driven model are integrated using an error correction method to form a hybrid model; The prediction is based on the output of the mechanistic model and compensated by the prediction error correction amount output by the data-driven model. The algebraic sum of the two is used as the final prediction output. Based on real-time data collected during the current fermentation process, the data is input into the mixing model for continuous rolling prediction. The intelligent decision-making process includes: using a reinforcement learning algorithm to solve for the optimal control strategy based on the prediction results of the current fermentation state variables and the digital twin of the fermentation process.

[0008] The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet provided by this invention has the following beneficial effects: This invention establishes a complete data link from data acquisition in the fermentation workshop to intelligent decision-making in the cloud through a three-layer collaborative architecture of "end-edge-cloud". Based on this, it integrates the interpretability of the mechanistic model and the predictive accuracy of the data-driven model through error correction hybrid modeling, achieving forward-looking rolling prediction of the future evolution trend of the fermentation process. This makes the previously invisible state of the fermentation process perceptible and predictable. Furthermore, based on the prediction results of the hybrid model, a reinforcement learning algorithm is used to automatically solve for the optimal control strategy, transforming the traditional passive response regulation relying on human experience into a data-driven proactive and forward-looking intelligent control. This fundamentally solves the technical problem of regulatory decisions lagging behind dynamic changes in the fermentation process due to the difficulty in measuring key biological parameters online and the significant lag in offline detection results. Simultaneously, the intelligent control execution module automatically sends the AI ​​decision results to the field actuators, forming a complete closed-loop control link from state perception, trend prediction, intelligent decision-making to automatic execution. This achieves real-time optimization and precise control of the fermentation process, significantly improving the stability of product yield, reducing batch failure risk, and lowering energy costs. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, and these are all within the protection scope of the present invention.

[0010] Figure 1 This is a schematic diagram of the structure of an intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet; Figure 2 This is a flowchart of an intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. In the description of the present invention, it should be understood that the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, the element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Where there is no conflict, embodiments of the present invention and the various features thereof can be combined with each other, all of which are within the scope of protection of the present invention. Example 1

[0012] Please see Figure 1 This invention provides an intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet, comprising a device layer, an edge layer, and a platform layer, wherein: The equipment layer is located in the biopharmaceutical fermentation workshop and is used to realize the physical connection and data acquisition of fermenters, seed tanks, sensors and actuators, and transmit the acquired data to the edge layer through an industrial intelligent gateway; In the specific implementation at the device layer, the IoT gateway unit adopts an industrial-grade embedded hardware platform. This platform is equipped with a multi-protocol parallel parsing engine, capable of simultaneously supporting access via OPC UA, Modbus TCP, Profinet, and EtherNet / IP protocols. The hardware architecture of the IoT gateway unit is built on an x86 or ARM processor, equipped with at least four Ethernet interfaces and at least two serial communication interfaces to meet the access requirements of different types of equipment within the fermentation workshop, including the fermenter distributed control system, seed tank programmable logic controller (PLC), various sensors, and actuators. The sensor network unit includes temperature sensors, pH sensors, dissolved oxygen sensors, agitator speed sensors, tank pressure sensors, exhaust gas analyzers, and feed metering devices. These sensors are installed according to the process measurement point layout of the fermenter, and sensor signals are transmitted to the IoT gateway unit via shielded twisted-pair cables or industrial Ethernet. The actuator network unit includes a feed pump valve control module, a temperature control system adjustment module, an agitator motor speed control module, and a tank pressure control module. Each actuator receives control commands from the upper layer through the IoT gateway unit and executes corresponding actions. The raw data collected at the device layer is aggregated by the IoT gateway unit and then transmitted to the edge layer via industrial Ethernet or 5G communication links.

[0013] The edge layer is set up in the workshop and is used to receive raw data transmitted from the equipment layer, perform real-time data preprocessing, anomaly detection and emergency control, and upload the preprocessed data to the platform layer. In the specific implementation of the edge layer, the edge computing nodes employ industrial-grade embedded computing devices deployed within control cabinets in the fermentation workshop. These edge computing nodes are built on an octa-core ARM architecture CPU with a clock speed of at least 2 GHz, equipped with at least 8 gigabytes of RAM and at least 256 gigabytes of high-speed solid-state storage, and run a containerized edge computing operating system. After receiving raw data from the device layer, the edge computing nodes undergo filtering, noise reduction, format standardization, timestamp alignment, and missing value imputation by a data preprocessing module. The preprocessed data is then used by an anomaly detection module for real-time anomaly detection based on preset process parameter thresholds and a single-class classification model trained on historical data. When an anomaly is detected, the emergency control module automatically generates control commands and sends them to the actuators via an IoT gateway to implement emergency protection operations for the fermenter. Simultaneously, the preprocessed data is uploaded to the platform layer via the MQTT or AMQP protocol.

[0014] The platform layer is located in the enterprise data center or cloud server and is used to receive pre-processed data transmitted from the edge layer, build the core functions of the industrial internet platform, and realize multi-source data fusion, digital twin of fermentation process, intelligent decision-making, intelligent control execution, blockchain traceability and predictive maintenance. Specifically, the platform layer is located in an enterprise data center or cloud server, deployed using a containerized microservice architecture, and includes a big data storage cluster, core service modules of the industrial internet platform, and model training and inference modules. After receiving preprocessed data uploaded from the edge layer, the platform layer first uses a multi-source data acquisition and fusion module to aggregate real-time sensor data, offline detection data, environmental data, and equipment operation data into the big data storage cluster, forming a complete data view of the fermentation process. Based on this, the platform layer runs six core functions: digital twin of the fermentation process, intelligent decision-making, intelligent control execution, blockchain traceability, and predictive maintenance.

[0015] The digital twin of the fermentation process includes: constructing a mechanistic model based on microbial growth kinetic equations and metabolic network models; training a data-driven model using historical batch fermentation data; integrating the mechanistic model and the data-driven model using an error correction method to form a hybrid model; using the output of the mechanistic model as the base prediction value and the prediction error correction amount output by the data-driven model as compensation, with the algebraic sum of the two as the final prediction output; and inputting real-time collected data of the current fermentation process into the hybrid model for continuous rolling prediction. In this embodiment, the specific implementation of the digital twin of the fermentation process is as follows: First, the mechanism model construction unit establishes a mathematical model describing microbial growth, substrate consumption, and product synthesis based on the microbial growth kinetic equation and metabolic network model. The cell growth model adopts a specific growth rate model based on the Monod equation, and the input variables of the model include substrate concentration, temperature, pH value, and dissolved oxygen concentration; the product synthesis model adopts a product synthesis kinetic model based on the Luedeking-Piret equation. The model parameters are identified using nonlinear least squares method based on experimental data from historical batches. The identification algorithm adopts the Levenberg-Marquardt algorithm, and the identification results are stored in the model parameter database after cross-validation. Second, the data-driven model training unit trains a bidirectional long short-term memory network (Bi-LSTM) model using historical batch fermentation data (including complete process parameter records of successful batches and abnormal operating condition records of failed batches, with a data time span covering no less than three years and a batch number of no less than 500 batches). The Bi-LSTM network consists of three bidirectional recurrent hidden layers, each containing 128 hidden nodes. The model's input features include time series data of process parameters over the past twelve hours (temperature, pH, dissolved oxygen concentration, stirring speed, tank pressure, feed rate, etc.) and current offline detection index values ​​(potency, residual sugar concentration, amino acid concentration, and cell concentration). The model is trained using the Adam optimization algorithm with a learning rate of 0.001, a batch size of 32, and 500 training epochs, employing an early stopping strategy to prevent overfitting.

[0016] Then, the hybrid modeling unit integrates the mechanistic model and the data-driven model using an error correction method. The specific implementation process of the error correction method is as follows: The mechanistic model output serves as the base prediction, while the data-driven model output's prediction error correction is used as compensation. The algebraic sum of these two is the final prediction output. The error correction model also employs a bidirectional long short-term memory network architecture. Input features include the current mechanistic model prediction, the current process parameter vector, and the current offline detection index value. The hybrid model's output includes the predicted concentration of the fermentation endpoint product and the predicted trajectory of key metabolites. The real-time simulation unit performs a rolling prediction every five minutes based on real-time data collected during the fermentation process. It inputs the current process parameters into the hybrid model for continuous rolling prediction over a time domain of four to eight hours, outputting the changing trends of key indicators such as product concentration, substrate residue, dissolved oxygen changes, and pH changes. This transforms the previously "black box" fermentation process into a perceptible and predictable transparent process.

[0017] The intelligent decision-making process includes: using a reinforcement learning algorithm to solve for the optimal control strategy based on the prediction results of the current fermentation state variables and the digital twin of the fermentation process.

[0018] In this embodiment, the intelligent decision-making is implemented as follows: The optimal control strategy generation unit in the intelligent decision engine module solves for the optimal control strategy based on the prediction results of the current fermentation state variables and the digital twin model using the Deep Deterministic Policy Gradient (DDPG) algorithm. The state space of the reinforcement learning algorithm includes the current process parameter vector (temperature, pH value, dissolved oxygen concentration, stirring speed, tank pressure, substrate concentration, product concentration, etc., a total of fifteen dimensions) and the environmental parameter vector (ambient temperature, ambient humidity, and clean zone pressure difference, etc.). The action space includes the feed rate adjustment, temperature setpoint adjustment, pH adjuster addition rate adjustment, and dissolved oxygen control strategy switching command (a total of four dimensions). The adjustment range of each action is constrained according to the process safety limits. The reward function is designed as a multi-objective weighted form, comprehensively considering maximizing product yield (weight coefficient 0.5), minimizing energy consumption (weight coefficient 0.2), and minimizing process deviation (weight coefficient 0.3). The algorithm network comprises two neural network structures: a policy network and a value network. The number of nodes in the input layer of the policy network matches the dimension of the state space, and the number of nodes in the output layer matches the dimension of the action space. The hyperbolic tangent function is used for the output layer activation function to ensure that the output value is within a reasonable range. The value network uses an experience replay mechanism to store historical interaction data, with an experience replay buffer capacity of no less than 100,000 transformation samples. The DDPG algorithm is pre-trained in a digital twin virtual environment. Multiple simulation batches are run in parallel using the high-speed simulation capabilities of the digital twin model to explore policies. After pre-training, the policy network parameters are transferred to a real system for fine-tuning. The optimal control policy output by the algorithm includes the optimal setpoints for each control variable and the execution time window.

[0019] Furthermore, the intelligent control execution module distributes the optimal control strategy generated by intelligent decision-making to the distributed control system or programmable logic controller via the OPC UA or Modbus TCP standardized interface, forming a complete closed-loop control link from perception to analysis, from analysis to decision-making, and from decision-making to execution. Simultaneously, the blockchain traceability module writes key process parameters and operation records during fermentation into the consortium blockchain network, ensuring data authenticity and integrity. The predictive maintenance module constructs a fault prediction model based on data such as the operating current of the fermenter drive motor, bearing vibration acceleration, gearbox temperature, and sealing temperature, enabling early warning of equipment anomalies and intelligent recommendations for maintenance. These functional modules work collaboratively to transform traditional passive, experience-based control into data-driven, proactive, and forward-looking intelligent control.

[0020] Furthermore, to improve the prediction accuracy of the digital twin model of the fermentation process under different fermentation stages and operating conditions, the hybrid model in this embodiment adopts a dynamic error compensation mechanism based on fermentation state perception of the mechanistic model. Specifically, a mechanistic prediction model is first constructed using a microbial growth kinetics model, a product synthesis kinetics model, and a metabolic network model. By inputting process variables such as temperature, pH, dissolved oxygen concentration, stirring speed, feeding rate, and substrate concentration during the current fermentation process, the model outputs the cell concentration, product concentration, substrate consumption, and the changing trends of key metabolites at the corresponding time points, serving as the basic state prediction results of the fermentation process. Since the actual fermentation process is affected by factors such as batch differences in microbial strains, environmental disturbances, changes in equipment status, and dynamic changes in microbial metabolic pathways, the mechanistic model has certain prediction biases. Therefore, a data-driven model is further used to learn the error of the mechanistic model's output results, so that the data-driven model does not directly replace the mechanistic model, but is used to learn dynamic change characteristics that the mechanistic model cannot describe.

[0021] Furthermore, the dynamic error compensation mechanism includes an error calculation unit, a state evaluation unit, and an adaptive weight adjustment unit. Specifically, the error calculation unit acquires in real-time the difference between the actual detection results and the mechanistic model prediction results in historical fermentation batches, constructs a mechanistic model prediction error sequence, and associates the error sequence with the process state parameters at corresponding times. It then uses a bidirectional long short-term memory network to learn the error change pattern and obtain the error correction amount for future times. The state evaluation unit evaluates the reliability of the current mechanistic model based on the current fermentation stage, the rate of change of process parameters, and the trend of model prediction residual changes. The adaptive weight adjustment unit dynamically adjusts the fusion ratio between the mechanistic model output and the data-driven model compensation results based on the reliability, enabling the hybrid model to automatically adjust its model dependence according to changes in the fermentation process state.

[0022] Specifically, during the fermentation process, the system first determines whether the fermentation stage is in the lag phase, exponential growth phase, stationary phase, or decline phase through the fermentation stage identification module, and generates stage state parameters based on the microbial metabolic characteristics corresponding to different stages. When the microbial growth pattern is relatively stable, the fusion weight of the mechanistic model output is increased to ensure that the digital twin model maintains a strong mechanistic interpretation capability. When the deviation between the actual collected data and the mechanistic model prediction results is detected to be increasing continuously, or when abnormal changes are detected in parameters such as dissolved oxygen change rate, pH change rate, and feeding response characteristics, the weight of the mechanistic model is reduced and the error compensation weight of the data-driven model output is increased, so that the hybrid model can quickly adapt to the nonlinear dynamic change process.

[0023] Furthermore, the adaptive weight adjustment process determines the fusion parameters for the current moment based on the model prediction residuals, the magnitude of changes in state variables, and the historical prediction accuracy. The system first calculates the trend of prediction residual changes over multiple consecutive prediction periods. When the prediction residuals remain within a preset range, the mechanistic model is deemed to have high reliability. When the prediction residuals show a continuous increasing trend, it is determined that the current fermentation state may contain dynamic factors not covered by the mechanistic model, thus enhancing the compensation effect of the data-driven model. Through this dynamic error compensation method, the digital twin model can continuously correct model biases using real-time data while maintaining the constraints of microbial kinetic mechanisms, improving the model's adaptability to predicting fermentation states under different batches, environmental conditions, and abnormal operating conditions.

[0024] Furthermore, the hybrid digital twin model output, based on the aforementioned error dynamic compensation mechanism, includes not only the predicted current fermentation state but also the prediction deviation trend and model reliability evaluation results. These prediction deviation trend and reliability evaluation results are simultaneously input into the intelligent decision-making module for updating the state space of the reinforcement learning control strategy. By introducing the model's own uncertainty information into the control decision-making process, the reinforcement learning algorithm can select a more conservative or proactive control strategy based on the predictive reliability of the digital twin model, thereby reducing the control risk caused by model prediction deviations and achieving adaptive prediction and intelligent optimization control for biopharmaceutical fermentation processes.

[0025] Furthermore, the device layer includes: The Internet of Things (IoT) gateway unit is used to connect the distributed control system of the fermenter, the PLC controller, sensors and actuators of the seed tank through standard industrial communication protocols to realize real-time data acquisition and control command issuance. Specifically, the IoT gateway unit adopts an industrial-grade embedded hardware platform equipped with a multi-protocol parallel parsing engine, capable of simultaneously supporting OPC UA, Modbus TCP, Profinet, and EtherNet / IP protocols. The hardware architecture of the IoT gateway unit is built on an x86 or ARM processor, equipped with at least four Ethernet interfaces and at least two serial communication interfaces to meet the access requirements of various types of devices in the fermentation workshop, including the fermenter distributed control system (DCS), seed tank programmable logic controller (PLC), various sensors, and actuators. The IoT gateway unit integrates a protocol conversion engine with a plug-in architecture. Each communication protocol corresponds to an independent protocol parsing plug-in, and the plug-ins are independent of each other and can be dynamically loaded. In actual operation, the IoT gateway unit collects data packets from each device in parallel with a millisecond-level sampling period. The protocol conversion engine extracts and standardizes the data carried by different protocols according to predefined data mapping rules. The processed data is uniformly converted to JSON or Protobuf format and appended with a unified millisecond-level timestamp tag. The IoT gateway unit maintains a bidirectional communication channel with the edge layer. While uploading collected data to the edge layer, it receives control commands from the edge layer or platform layer, converts the control commands into a protocol format supported by the target actuator, and then sends them out.

[0026] The sensor network unit includes a temperature sensor, a pH sensor, a dissolved oxygen sensor, a stirring speed sensor, a tank pressure sensor, a tail gas analyzer, and a feeding metering device, which are used to collect physical and chemical parameters during the fermentation process. Specifically, the sensor network units are installed and deployed according to the process measurement point layout of the fermenter. The temperature sensor uses a corrosion-resistant and high-temperature resistant Pt100 platinum resistance sensor, with a measurement range of 0°C to 150°C and an accuracy class of at least A. The sensor signal is transmitted to the IoT gateway unit via a four-wire connection to eliminate measurement errors caused by line resistance. The pH sensor uses a high-temperature resistant glass electrode pH sensor, with a measurement range of 0 to 14°C and an accuracy of at least ±0.02 pH units. The sensor is equipped with an automatic temperature compensation function, achieving a compensation accuracy of ±0.01 pH units within the 0°C to 120°C range. The dissolved oxygen sensor uses an optical polarographic dissolved oxygen sensor, with a measurement range of 0 to 100% air saturation and a response time of no more than 30 seconds. The sensors are equipped with in-situ cleaning devices to prevent fermentation broth from contaminating the optical diaphragm; the stirring speed sensor uses a non-contact photoelectric encoder or magnetoelectric speed sensor, with a measurement range of 0 to 1500 revolutions per minute, and the output signal is in pulse frequency form; the tank pressure sensor uses a diaphragm pressure sensor, with a measurement range of 0 to 500 kPa and an accuracy of not less than ±0.25%FS; the exhaust gas analyzer uses a combination of an infrared absorption carbon dioxide analyzer and a paramagnetic oxygen analyzer, capable of measuring the oxygen and carbon dioxide concentrations in the exhaust gas in real time, with a measurement accuracy of not less than ±1%FS; the feed metering device uses a mass flow meter or Coriolis mass flow meter, with a measurement accuracy of not less than ±0.5%FS. The signal cabling of the sensor network unit uses shielded twisted-pair cable, and the cable laying follows the industrial fieldbus cabling specifications. The sensor power supply uses an independent uninterruptible power supply to ensure that data acquisition can still be maintained for at least 30 minutes in the event of a main power failure.

[0027] The actuator network unit, including the feed pump valve control module, temperature control system adjustment module, stirring motor speed control module, and tank pressure control module, is used to receive and execute control commands from the edge layer or platform layer.

[0028] Specifically, the actuator network unit includes a feed pump valve control module, a temperature control system regulation module, a stirring motor speed regulation module, and a tank pressure control module. Each control module receives control commands from the edge layer or platform layer through an IoT gateway unit. The control commands are formatted using the standard process control object file format. The feed pump valve control module is equipped with an electric regulating valve or a pneumatic diaphragm valve. The valve positioning accuracy is no less than ±1%FS. The valve feedback signal is output to the IoT gateway unit via a position transmitter for closed-loop verification. The temperature control system regulation module includes a heating unit and a cooling unit. The heating unit uses jacketed steam heating or electric heating, and the cooling unit uses cooling water circulation or chilled brine circulation. After receiving the temperature setpoint command, the control module calculates the control output using a PID algorithm. The output signal controls the opening degree of the steam regulating valve or cooling water regulating valve. The temperature control accuracy is maintained under steady-state conditions. The temperature control accuracy reaches ±0.5 degrees Celsius. The stirring motor speed control module uses a frequency converter to drive the stirring motor. The frequency converter receives the speed setpoint signal from the IoT gateway unit and outputs a zero-to-ten-volt analog voltage or a four-to-twenty-milliampere analog current to the motor driver. The motor speed control accuracy is no less than ±1%FS. The tank pressure control module is equipped with a pressure transmitter and a pressure regulating valve. The pressure transmitter monitors the tank pressure in real time and feeds the signal back to the PID controller. The controller automatically adjusts the opening of the exhaust valve or pressure-replenishing valve according to the deviation between the set value and the measured value. The tank pressure control accuracy reaches ±2 kPa under steady-state conditions. Each actuator module is equipped with a status feedback mechanism, which transmits the actual execution status (such as the actual valve opening, the actual motor speed, the actual temperature value, etc.) back to the IoT gateway unit in real time, forming a complete closed loop of control command issuance and execution status feedback.

[0029] Furthermore, the IoT gateway unit supports parallel parsing of multiple industrial communication protocols, including OPC UA protocol, Modbus TCP protocol, Profinet protocol and EtherNet / IP protocol; it has a built-in protocol conversion engine that can uniformly convert data carried by different protocols into a standardized data format and add a unified timestamp; a lightweight data caching module is deployed on the edge side to automatically cache unuploaded data when the network is interrupted and automatically re-upload it after the network is restored.

[0030] In this embodiment, the IoT gateway unit has a built-in multi-protocol parallel parsing engine. This engine is based on a plug-in architecture design, with each communication protocol corresponding to an independent protocol parsing plug-in. Specifically, the OPC UA protocol plugin is developed based on the UA SDK provided by the OPC Foundation. It supports automatically locating the OPC UA server in the workshop through the discovery service, establishing a secure session channel, and subscribing to node data. The subscription period can be configured from 100 milliseconds to 1000 milliseconds. The Modbus TCP protocol plugin acts as a Modbus master station, communicating with each slave device through a TCP / IP connection. It supports common operations such as function codes 0x03 (read holding register), 0x04 (read input register), and 0x06 (write single register). The polling period is configured according to the real-time requirements of the device. The Profinet protocol plugin is based on the real-time Ethernet communication stack and supports both IRT (Isochronous Real-Time) and RT (Real-Time) communication modes. It establishes a connection with the PLC's IO controller and periodically exchanges process data. The EtherNet / IP protocol plugin is based on the CIP (Common Industrial Protocol) stack and communicates with devices through both implicit messages (I / O data) and explicit messages (explicit messages). It supports browsing and accessing CIP objects. Each protocol plugin operates independently and does not interfere with the others. It can process communication tasks from different devices in parallel at the same time. The gateway automatically calls the corresponding protocol plugin for data parsing based on the protocol type configured when the device is connected.

[0031] Furthermore, after receiving the raw data parsed by each protocol plugin, the protocol conversion engine initiates a format standardization process. Internally, this engine maintains a device data mapping table, which defines the metadata for each device data point in an object-oriented manner, including device ID, data point name, data type, original data format, physical quantity unit, conversion rules, and scaling factor. Based on this mapping table, the protocol conversion engine uniformly converts data carried by different protocols into a standardized data format, which is encapsulated using JSON or Protobuf. Simultaneously with format conversion, the protocol conversion engine periodically synchronizes time with the platform layer's clock source through a built-in IEEE 1588 Precision Time Protocol (PTP) or Network Time Protocol (NTP) client, attaching a unified millisecond-level timestamp tag to each frame of converted data. After data format standardization and timestamp alignment are completed, the standardized data is written to the gateway's local data buffer according to the device ID and data point ID organization method, and then uploaded to the edge layer via a separate sending thread.

[0032] The lightweight edge-side data caching module is deployed on the local storage medium of the IoT gateway. This caching module employs a ring buffer structure design with a total buffer capacity of 100 gigabytes, using industrial-grade eMMC or solid-state drives as storage media. When network communication is normal, the caching module forwards data to the edge layer in real-time using a first-in, first-out (FIFO) approach, with data only briefly residing in the buffer (typically no more than one second). When the gateway detects a network connection interruption with the edge layer (detected by periodically sending heartbeat packets; three consecutive heartbeat timeouts indicate a network interruption), the caching module automatically switches to data caching mode, writing the data to be uploaded sequentially into the ring buffer. Each data record contains complete standardized data content, a timestamp, and a cache tag. The caching module has a built-in data compression engine that performs lightweight LZ4 compression on the data written to the buffer to extend the cache duration. When the network is restored, the caching module automatically resumes the transmission process from the point of interruption, using a breakpoint resumption mechanism. It reads the compressed data blocks accumulated during the interruption from the buffer, decompresses them, and uploads them to the edge layer sequentially. During retransmission, the edge layer automatically identifies and removes duplicate data based on timestamps, ensuring no data duplication or loss. During network outages, if the circular buffer reaches its capacity limit, the caching module automatically overwrites the oldest data and records the overwrite log for later traceability. This lightweight caching mechanism ensures that fermentation process data is not lost during industrial network fluctuations or brief interruptions, guaranteeing the integrity and continuity of the data link.

[0033] Furthermore, the edge layer includes: Edge computing nodes employ industrial-grade embedded computing devices, equipped with multi-core central processing units and high-speed solid-state storage media, and run an edge computing operating system to perform data preprocessing tasks; Specifically, the edge computing nodes utilize industrial-grade embedded computing devices built on an ARM architecture eight-core CPU with a clock speed of at least 2 GHz, equipped with at least 8 gigabytes of RAM and at least 256 gigabytes of high-speed solid-state storage. Deployed within control cabinets in the fermentation workshop, these nodes possess an IP65 protection rating and can operate stably in industrial environments with temperatures ranging from 0 to 55 degrees Celsius and relative humidity from 10% to 90%. The edge computing nodes run a containerized edge computing operating system, specifically a customized Linux system built on YoctoProject, pre-installed with the Docker container engine and the Kubernetes lightweight orchestration framework (K3s), using contnerd as the runtime environment. This containerized deployment allows for flexible deployment of containers for data preprocessing, anomaly detection, and emergency control functions. Each container is resource-isolated and independent, facilitating independent upgrades and fault isolation of functional modules. The edge computing nodes have a built-in time-series database engine, using either InfluxDB or TimescaleDB architecture, for fast local data writing and querying. Local data retention is at least 24 hours, and the write performance of the time-series database is no less than 100,000 points per second, with a query response time of no more than 10 milliseconds. The edge computing nodes also feature power management capabilities. By monitoring the fermenter's operating status in real time, they automatically switch to a low-power mode when the fermenter is idle or in a closed-loop state. In low-power mode, unnecessary peripheral interfaces are disabled, and the CPU frequency is reduced to the minimum operating frequency to minimize energy consumption.

[0034] The data preprocessing module is used to filter and denoise the raw sensor data, standardize the format, align the timestamps, and impute missing values. Specifically, the data preprocessing module, as a core container service on the edge computing node, performs preprocessing on the raw sensor data uploaded by the IoT gateway. The filtering and denoising process employs a combination of sliding window mid-range filtering and Kalman filtering: the sliding window mid-range filtering is configured with a window width of five to eleven sampling points to remove impulse noise and outliers from the sensor data; Kalman filtering is used to preserve the true dynamic characteristics of the signal while removing noise. The process noise covariance matrix and measurement noise covariance matrix of the filtering algorithm are configured independently according to different sensor types. The process noise covariance of temperature and pH sensors is configured with smaller values ​​to maintain steady-state accuracy, while the configuration of dissolved oxygen and tank pressure sensors allows for a larger dynamic tracking bandwidth. Format standardization converts data from different sensors into a unified double-precision floating-point format, and physical quantity units are uniformly converted to SI units (temperature to Kelvin, pressure to Pascal, concentration to moles per cubic meter, etc.). Timestamp alignment is based on the unified timestamps of the IoT gateway units, compensating for and correcting inconsistencies in data arrival times caused by network transmission delays. Specifically, linear interpolation is used to resample the data from each sensor to the same time grid, ensuring that all sensor data at the same moment are aligned on the time axis. Missing value imputation uses cubic spline interpolation to fill in data with no more than five consecutive missing sampling points. The interpolation function uses ten valid data points before and after the missing point to construct a cubic spline curve for interpolation. For data segments with more than five consecutive missing sampling points, they are marked as invalid data and passed to the platform layer with a missing value label for further processing by the platform layer's big data analysis module.

[0035] The anomaly detection module is used to detect anomalies in sensor data in real time based on preset process parameter thresholds and a single-class classification model trained on historical data, and to trigger emergency control logic when an anomaly is detected. Specifically, the anomaly detection module is deployed in a containerized manner, receiving standardized time-series data output from the data preprocessing module in real time. It employs two parallel mechanisms: detection using preset process parameter thresholds and detection using a single-class classification model trained on historical data. Anomaly detection based on process parameter thresholds utilizes a multi-level alarm mechanism: the first level is the early warning threshold, where an early warning signal is generated and logged when a parameter exceeds the early warning threshold, but no emergency control is triggered; the second level is the alarm threshold, where an alarm signal is generated and emergency control logic is triggered when a parameter exceeds the alarm threshold; the third level is the danger threshold, where the highest priority interrupt-type emergency control is immediately triggered when a parameter exceeds the danger threshold. Threshold parameters are dynamically configured according to different fermentation varieties and fermentation stages. Edge computing nodes maintain a process threshold configuration table, where each row corresponds to the three-level thresholds for each parameter of a fermentation variety at a specific fermentation stage (lag phase, logarithmic growth phase, stationary phase, and decay phase). When a fermentation batch enters a new stage, the system automatically switches to the corresponding threshold configuration. The single-class classification model trained on historical data employs either the Isolation Forest algorithm or the Autoencoder algorithm. Model training is completed offline in the platform-level model training and inference module. After training, the model file is distributed to edge computing nodes for loading and execution. Model training uses only historical data under normal operating conditions. During inference, anomaly scores are calculated for input samples. When the anomaly score exceeds a set threshold, an abnormal operating condition is identified. The input features of the single-class classification model include the statistical characteristics of the sensor data time series (mean, variance, rate of change, peak value, etc.) within the most recent hour, the current process parameter vector, and the fermentation batch status information (current fermentation stage, batch runtime, etc.). When either of the two detection mechanisms triggers an anomaly determination, the anomaly detection module outputs the anomaly detection result and triggers emergency control logic.

[0036] The emergency control module is used to automatically generate control commands and send them to the actuators via the IoT gateway when the anomaly detection module determines that emergency intervention is required, thereby realizing emergency protection operations for the fermenter.

[0037] Specifically, the emergency control module is activated immediately when the anomaly detection module determines that emergency intervention is required. This module maintains an emergency control strategy library, which stores preset protective action sequences for different anomaly types (temperature exceeding limits, pH exceeding limits, dissolved oxygen anomaly, tank pressure exceeding limits, feed anomaly, etc.) in the form of rule tables. Upon receiving an anomaly detection result, the emergency control module matches the corresponding protective action sequence according to the anomaly type. The action sequence may include combinations of the following operations: stopping feed, switching to maximum cooling power, initiating emergency venting, switching to standby stirring speed, etc. The emergency control module sends the generated control commands to the actuators via the MQTT or OPC UA protocol through the IoT gateway. The time delay from anomaly detection to control command issuance does not exceed 500 milliseconds. The emergency control module has a built-in safety interlock function; any emergency control action must undergo a safety status check before execution. The safety status check includes whether the fermenter pressure is within the safe range (not exceeding 80% of the design pressure), whether the stirring system is in an operable state, and whether the liquid level is above the safe level. If the safety status check fails, the emergency control module automatically escalates to an emergency shutdown operation—cutting off all heating sources, shutting down the feed pump, opening the emergency exhaust valve, and notifying on-site personnel for manual intervention. Each time the emergency control module executes an emergency control action, it sends an emergency event notification to the platform layer. The notification includes the anomaly type, anomaly parameter values, trigger time, the sequence of executed emergency control actions, and feedback on the action execution results. Upon receiving the emergency event notification, the platform layer automatically records it in the immutable ledger of the blockchain traceability module for subsequent auditing and analysis.

[0038] Furthermore, the platform layer also includes a multi-source data acquisition and fusion module, comprising: The real-time parameter acquisition unit is used to receive and analyze parameters such as temperature, pH value, dissolved oxygen concentration, stirring speed, tank pressure, exhaust gas oxygen concentration, exhaust gas carbon dioxide concentration, and cumulative feed amount from the sensor network unit in real time through the acquisition link composed of IoT gateway and edge computing node. Specifically, the data transmission in the acquisition link adopts either the MQTT or AMQP protocol. The MQTT protocol is configured with QoS=1 to ensure at least one data delivery, while the AMQP protocol is configured in automatic acknowledgment mode to balance reliability and transmission efficiency. The real-time parameter acquisition unit internally runs a data parsing engine, which parses the received raw JSON or Protobuf messages according to a pre-configured data template. The data template is defined in YAML format, with each data field corresponding to a parsing rule, including the field path, data type, valid value range, and unit information. The parsing engine extracts parameters from the messages, such as temperature, pH value, dissolved oxygen concentration, stirring speed, tank pressure, exhaust gas oxygen concentration, exhaust gas carbon dioxide concentration, and cumulative feed amount. After format and value range validation, these parameters are appended with the current system timestamp and written to the real-time data table of the big data storage cluster. The real-time parameter acquisition unit also maintains connection status monitoring with edge computing nodes. When a data interruption is detected exceeding a set threshold (e.g., no new data received for thirty seconds), a data anomaly alarm is automatically generated, notifying maintenance personnel to check the integrity of the acquisition link.

[0039] The offline data entry unit provides a standardized data entry interface, allowing quality control laboratory personnel to enter potency test data, residual sugar concentration test data, amino acid concentration test data, and bacterial cell concentration test data into the system. The offline data entry timestamp is automatically associated with the sampling timestamp. Specifically, the offline data entry unit provides a web-based data entry interface. This interface uses a responsive design developed with Vue.js or React frameworks, supporting access on various terminals such as desktop computers, tablets, and smartphones. The core of the data entry interface is a set of standardized data entry forms. The form fields correspond one-to-one with the testing items in the quality control laboratory, including potency (potency unit / mL), residual sugar concentration (g / L), amino acid concentration (mg / L), and cell concentration (OD value or g / L). Each field is configured with value range validation rules and logical relationship validation rules. The value range validation rules set upper and lower limits based on the reasonable range of historical data, while the logical relationship validation rules check for inconsistencies between different testing items in the same batch (e.g., a warning is triggered if the residual sugar concentration and cell concentration do not exhibit the expected fluctuation relationship). Operators need to fill in two time fields in the entry interface: sampling time (i.e., the actual time when the quality control personnel remove the sample from the fermenter) and testing completion time (i.e., the time when the laboratory issues the test results). The system automatically records the submission time of the data entry operation as the entry timestamp and uses the sampling time as the core timestamp attribute of the offline data, thus automatically associating the entry timestamp and sampling timestamp of the offline data and ensuring that the offline data can be accurately aligned to the timeline of the fermentation process. The offline data entry unit also supports batch data import, which supports CSV and Excel files. The import template predefines the mapping relationship between column names and data fields. During the import process, the same validation rules as for single entries are automatically executed. Records that fail validation generate detailed error reports for users to correct and re-import.

[0040] The environmental data access unit is used to establish a data connection with the enterprise's cleanroom building automation system through a standard API interface, and to obtain real-time data on the cleanroom's ambient temperature, relative humidity, and room pressure difference. Specifically, API authentication uses the OAuth 2.0 protocol's client credentials model, with the client ID and client key pre-configured during system deployment. The environmental data access unit actively calls the building automation system's API to retrieve data at a configurable polling interval (default 60 seconds). The API request URL format is "https: / / [BMS_HOST] / api / v1 / environment / current", and the request parameters include a list of room numbers for which data needs to be retrieved. The returned data is in JSON format, containing the ambient temperature (degrees Celsius), relative humidity (percentage RH), and room pressure differential (Pascals) for each room in the clean area. The environmental data access unit performs format conversion and timestamp calibration on the returned data before writing it to the environmental data table in the big data storage cluster. To ensure the continuity of data acquisition, the environmental data access unit has a retry mechanism and a circuit breaker mechanism: when an API request fails, it automatically performs exponential backoff retries (initial retry interval of one second, maximum retry interval of sixty seconds, and maximum number of retries of three); when the failure exceeds ten times consecutively, the circuit breaker is triggered, the API call is stopped and an alarm log is recorded, and the connection is re-established after manual recovery or automatic detection of API recovery by the system.

[0041] The equipment operation data acquisition unit is used to acquire the operating current of the fermenter drive motor, valve opening status, sterilization records, and equipment start-up and stop logs via OPC UA protocol or Modbus protocol; Specifically, the equipment operation data acquisition unit collects operational status data of the fermenter's supporting equipment via the OPC UA or Modbus protocol. The OPC UA protocol connection adopts a client-server architecture, with the platform-layer equipment operation data acquisition unit acting as an OPC UA client, actively connecting to the OPC UA server deployed on-site in the workshop. Based on the OPC UA subscription / monitoring item mechanism, it obtains data change notifications in real time, with a collection frequency configured to once per second. For equipment that does not support the OPC UA protocol, the Modbus TCP protocol is used for communication. The equipment operation data acquisition unit acts as the Modbus master station, sending read requests to each slave station device sequentially according to a preset polling table, with a polling period configured to two seconds. The collected equipment operation data includes the operating current (amperes) of the fermenter drive motor, the opening status (percentage or on / off status) of each key valve, sterilization operation records, and equipment start-up and stop logs (operation time, operation type, operator identification). The collected data, after protocol parsing and timestamp standardization, is written to the equipment operation data table in the big data storage cluster for subsequent predictive maintenance and equipment health status assessment.

[0042] The data governance unit is used to establish unified data coding standards and data cleaning rules, and to perform format standardization, unit unification, timestamp calibration and outlier marking on data from different data sources. Specifically, the data coding standard adopts a data coding system based on the ISO 80000 international standard and the Chinese national standard, assigning a globally unique coding identifier to each process parameter, equipment parameter, and quality parameter. The coding format is a four-segment structure of "domain code - equipment / system code - parameter code - version number". For example, "FMT-001-TEMP-001" represents the temperature parameter of tank number one in the fermentation system. The data cleaning rule processing flow is as follows: First, a format standardization check is performed to verify the field types, precision, and format of the original data. Data that does not meet the format requirements is recorded as having a format error and enters the abnormal data queue, where manual review determines whether to correct it before storing it in the database. Next, unit conversion is performed, converting all data to standard units according to the baseline units defined in the coding standard, such as converting temperature to degrees Celsius, pressure to Pascals, and concentration to grams per liter. Then, timestamp calibration is performed to check the continuity and rationality of the data timestamps. Data with abrupt timestamp changes or obvious anomalies is marked, and a traceability process is initiated to contact the data source system to confirm the correct time information. Finally, outlier marking is performed. Values ​​exceeding a predefined reasonable range are automatically marked as "suspicious data" and accompanied by an anomaly code. Suspicious data remains in the database but does not participate in the automated digital twin and decision-making processes until it is manually reviewed and confirmed to remove the suspicious status. All data cleaning operations during the data governance unit's operation generate operation logs, recording the original data values, cleaning rules, cleaned values, and operation times, ensuring the data governance process is traceable and auditable.

[0043] Furthermore, the mechanistic model is based on a cell growth model constructed using the Monod equation and a product synthesis model constructed using the Luedeking-Piret equation. The model parameters are identified using nonlinear least squares method based on experimental data from historical batches.

[0044] The cell growth model employed a specific growth rate model based on the Monod equation. The Monod equation describes the functional relationship between the specific growth rate of microorganisms and the concentration of the limiting substrate, and its expression is as follows: ,in Specific growth rate (unit: ), Maximum specific growth rate (unit: S represents the limiting substrate concentration (unit: g / L). This represents the saturation constant (unit: g / L). In actual fermentation processes, microbial growth is also affected by environmental factors such as temperature, pH, and dissolved oxygen concentration. Therefore, an environmental factor correction term is introduced based on the basic Monod equation. Specifically, the temperature correction uses the temperature coefficient in the form of the Arrhenius equation. The correction factor is Where T is the current temperature, This is a reference temperature (usually the optimal growth temperature for this strain). The temperature coefficient is determined through fitting experimental data, typically between 1.03 and 1.08; pH correction uses a quadratic parabolic pH influence function, with the pH influence factor reaching its maximum value of 1.0 at the optimal pH, decreasing parabolically as it deviates from the optimal pH; dissolved oxygen correction uses a Monod-form dissolved oxygen influence term, expressed as follows: ,in This represents the current dissolved oxygen concentration. This is the dissolved oxygen half-saturation constant. Combining the above corrections, the complete cell growth rate equation is: ,in This refers to the bacterial cell concentration (unit: g / L). , , These are correction factors for temperature, pH, and dissolved oxygen, respectively. The product synthesis model employs a product synthesis kinetic model based on the Luedeking-Piret equation.

[0045] Specifically, the Luedeking-Piret equation describes the relationship between the product synthesis rate and the cell growth rate, and its expression is as follows: ,in Product concentration (unit: g / L). The product synthesis rate, This represents the bacterial growth rate. The product synthesis coefficient associated with growth (product formation is coupled with cell growth). This represents the product synthesis coefficient that is not related to growth. For fermentation processes such as recombinant protein fermentation, where the primary goal is the accumulation of metabolites, and The value of reflects the contribution weight of product synthesis at different fermentation stages. The substrate consumption model uses a substrate consumption equation based on the yield coefficient: ,in Substrate concentration (unit: g / L). The yield coefficient of the bacterial cells to the substrate (unit: g bacterial cells / g substrate). The yield coefficient of the product to the substrate (unit: g product / g substrate). To maintain the metabolic coefficient (unit: g substrate / g cell·h), the above mechanistic model includes the following parameters to be identified: maximum specific growth rate. saturation constant Temperature coefficient Coefficients of each order of the pH influence function, dissolved oxygen half-saturation constant Synthesis coefficient of growth-related products Non-growth-related product synthesis coefficient Bacterial cell yield coefficient Product yield coefficient and maintain metabolic coefficient .

[0046] Model parameter identification was performed using nonlinear least squares method based on historical batch experimental data. The dataset for parameter identification was derived from the company's historical batch database, selecting no fewer than fifty successful fermentation batches. Each batch of data included complete time series of process parameters (temperature, pH, dissolved oxygen, stirring speed, etc.) and corresponding offline detection data (cell concentration, substrate concentration, product concentration, etc.), covering various operating conditions under typical operating conditions for this strain. The objective function for parameter identification was defined as the weighted sum of squared residuals between the model predictions and experimentally measured values: ,in Let be the parameter vector to be identified. For the model at time The predicted value, The corresponding measured value, For the first The weighting coefficients at each time point. The calculation of the weighted sum of squared residuals considers three output variables simultaneously: cell concentration, substrate concentration, and product concentration. The weight of each output variable is configured according to the reciprocal of its measurement accuracy. The identification algorithm adopts the Levenberg-Marquardt algorithm combined with a global search particle swarm optimization algorithm to overcome the deficiency of the traditional LM algorithm in sensitivity to initial values: First, the particle swarm optimization algorithm is used to perform a coarse search in the global parameter space, with the number of particles configured to be one hundred and the number of iterations configured to be five hundred, to obtain a globally approximate optimal solution; then, the particle swarm optimization result is used as the initial value to input the Levenberg-Marquardt algorithm for local fine optimization. The initial value of the damping factor of the LM algorithm is set to 0.01. During the iteration process, the damping factor is dynamically adjusted according to the change of residuals. When the iteration step size is less than a preset threshold ( The iteration terminates when the maximum number of iterations (two hundred) is reached, and the final identified parameter values ​​are output. After parameter identification is completed, the identification results are cross-validated using ten batches of independent experimental data that did not participate in the identification, and the coefficient of determination R between the model predictions and the measured values ​​is calculated. 2 And the root mean square error (RMSE), when the R of each key output variable2 When all values ​​are greater than 0.85 and the RMSE is within an acceptable range, the identification results are considered valid and stored in the model parameter database. During subsequent operation, as new batches of data accumulate, the system periodically (e.g., quarterly) automatically triggers the model parameter re-identification process, using the accumulated new data to update and correct the parameters, ensuring that the mechanistic model always reflects the actual metabolic characteristics of the strain under current production conditions.

[0047] Furthermore, the data-driven model adopts a bidirectional long short-term memory network structure (Bi-LSTM). The network contains three bidirectional recurrent hidden layers, each containing 128 hidden nodes. The input features of the model include the time series of process parameters over the past twelve hours and the offline detection index value at the current moment.

[0048] In this embodiment, the Bi-LSTM network simultaneously sets forward and backward LSTM layers in the time dimension, enabling the network to utilize contextual information from both past and future moments for current-moment prediction. Compared to unidirectional LSTM, which can only utilize historical information, the bidirectional structure can extract richer temporal features in scenarios with strong temporal correlation, such as fermentation. The network contains three bidirectional recurrent hidden layers, each containing 128 hidden nodes. The three-layer structure balances the model's expressive power and training complexity: the first Bi-LSTM layer is responsible for extracting local temporal patterns of the original time-series signal, such as short-term trends and fluctuations in parameters; the second Bi-LSTM layer extracts mid-term evolution patterns based on local patterns, such as the delay time and recovery curve shape of dissolved oxygen response after feeding; the third Bi-LSTM layer further abstracts these into global fermentation stage features, such as the inflection point features of the transition from the logarithmic growth phase to the stationary phase. The configuration of 128 hidden nodes per layer is determined based on the time-series complexity of the fermentation process. The input sequence length is twelve hours, and the input feature dimension at each time step is twelve-dimensional. The 128 hidden nodes provide sufficient feature transformation space for this input dimension while avoiding the risk of overfitting due to excessive parameters. Dropout regularization is used between layers, with a dropout rate of 0.2, which randomly discards 20% of the neuron connections during training to prevent overfitting. The network's output layer is a fully connected layer, with the output dimension matching the number of prediction targets. The prediction targets include three key indicators: product concentration, substrate concentration, and cell concentration at a time point after the current time step.

[0049] Specifically, the process parameter time series over the past twelve hours, sampled at one-minute intervals, comprises 720 time steps. Each time step's process parameter vector includes eight real-time acquired process parameters: temperature, pH, dissolved oxygen concentration, stirring speed, tank pressure, feed rate, tail gas oxygen concentration, and tail gas carbon dioxide concentration, totaling 720 × 8 = 5760 time-series feature values. The current offline detection index values ​​include four indicators: potency (potency unit / mL), residual sugar concentration (g / L), amino acid concentration (mg / L), and bacterial cell concentration (OD value or g / L). In actual data organization, because the sampling frequency of offline detection indicators is much lower than that of real-time process parameters, there may be no corresponding offline detection value at the current moment. In this case, the system uses the most recent valid offline detection value as the offline detection input for the current moment, and adds a "time interval since the last offline sampling" feature (unit: minutes) to the input features to help the model understand the timeliness of the offline data. All input features are Z-score standardized before being fed into the network. This involves subtracting the historical mean and then dividing by the standard deviation to bring all features to the same order of magnitude, thereby accelerating the convergence process of network training.

[0050] The network training uses the Adam optimization algorithm with a learning rate of 0.001 and an exponential decay rate estimated by the first moment. The exponential decay rate is estimated by setting the second moment to 0.9. The numerical stability constant is configured to be 0.999. Configured as The batch size is configured to 32 sample sequences, each containing 720 consecutive time steps of input data. The training epochs are configured to 500, employing an early stopping strategy to prevent overfitting—the early stopping metric is the loss function value on the validation set; training is terminated early when the validation set loss does not decrease for 20 consecutive epochs, and the model is restored to the checkpoint where the validation set loss is minimized. The validation set accounts for 20% of the training process, using random stratification to ensure that both the training and validation sets contain samples from different fermentation batches and product levels. The loss function uses a combination of mean squared error and mean absolute error. ,in Configured to 0.7, The configuration is set to 0.3, and the combined loss balances the sensitivity of MSE to large errors and the robustness of MAE to outliers. After model training, performance is evaluated on an independent test set, with evaluation metrics including the coefficient of determination R0. 2 The root mean square error (RMSE) and mean absolute percentage error (MAPE) are calculated. When the R-squared value of each predicted target on the test set... 2When all values ​​are greater than 0.90 and all MAPE values ​​are less than 15%, the model is considered to have met the deployment criteria. The trained Bi-LSTM network model is exported in ONNX format and deployed to the model inference service at the platform layer for use by the real-time simulation unit. The single forward propagation latency of the model inference does not exceed fifty milliseconds, which can meet the real-time requirement of performing rolling predictions every five minutes.

[0051] Furthermore, the error correction method adopts an adaptive weighted fusion strategy: at each moment of model prediction, the system calculates the credibility index of the mechanism model at the current moment. When the credibility of the mechanism model is high, the correction weight of the data-driven model is reduced; when the credibility of the mechanism model is low, the correction weight of the data-driven model is increased.

[0052] Specifically, the error correction method employs an adaptive weighted fusion strategy. The core idea behind this strategy is that while the mechanistic model possesses high reliability and interpretability within the operating conditions covered by parameter identification, its prediction accuracy significantly decreases when the fermentation process deviates from typical operating conditions. In contrast, the data-driven model, although lacking physical interpretability, can provide relatively stable predictions based on historical data patterns under various operating conditions. Based on this understanding, the adaptive weighted fusion strategy dynamically evaluates the credibility of the mechanistic model at each moment of model prediction and adaptively adjusts the fusion weights of the data-driven model's correction based on the level of credibility, thereby achieving a complementary advantage between the two models.

[0053] Furthermore, at each moment of the model's prediction, the system calculates the reliability index of the current moment's mechanistic model. The calculation of the reliability index involves two dimensions: the first dimension is whether the input parameters of the mechanistic model are within the valid range of the model's parameter identification, specifically achieved by calculating the Mahalanobis distance between the current operating point and the parameter space of the parameter identification training data; let the input parameter vector at the current moment be... (Including substrate concentration, temperature, pH, dissolved oxygen concentration, etc.), the mean parameters of historical training data are... If the covariance matrix is ​​Σ, then the Mahalanobis distance is... ; The smaller the value, the closer the current operating condition is to the identified coverage area, and the higher the reliability of the mechanistic model. The second dimension is the similarity between the current operating condition and historical successful batches. This is quantified by calculating the average Euclidean distance between the current process parameter vector and the process parameter vectors of all successful batches at the same fermentation stage in the historical batch database. The smaller the distance, the higher the frequency of the current operating condition in historical successful batches, and the more historically validated the predictive reliability of the mechanistic model under that condition. The two dimensions are weighted and summed to obtain the comprehensive credibility score. (Range 0 to 1), the closer the score is to 1, the higher the credibility of the mechanism model at the current moment.

[0054] In obtaining the credibility score of the mechanism model Then, the system dynamically determines the fusion weights for the data-driven model correction based on the score. The principle for determining the fusion weights is: when the mechanistic model has high credibility ( When the reliability is close to 1), it indicates that the mechanistic model can provide a relatively reliable prediction under the current working conditions. At this time, the fusion weight of the data-driven model correction is reduced, and the final prediction is mainly based on the basic prediction value of the mechanistic model. When the reliability of the mechanistic model is low ( When the value is close to 0, it indicates that the predictive reliability of the mechanistic model under the current operating conditions is insufficient. At this point, the fusion weight of the data-driven model correction increases, and the final prediction relies more on the compensation of the mechanistic model error by the data-driven model. Specifically, the fusion weight of the data-driven model correction... Through a monotonically decreasing function Calculation, i.e. The reliability score of the mechanism model is inversely proportional to the reliability score. Based on this, to increase the smoothness and stability of the fusion strategy, a saturation interval is introduced: when... When >0.85, Set to the minimum value of 0.15, that is, retain at least 15% of the data-driven correction weights to prevent systematic bias that may result from relying entirely on the mechanistic model; when When <0.15, The maximum value is set to 0.85, meaning that at least 15% of the basic prediction weights of the mechanistic model are retained to maintain the physical rationality constraints of the model output. The introduction of the saturation interval allows the fusion strategy to maintain bidirectional redundancy in the boundary regions where the mechanistic model is extremely reliable or extremely unreliable, avoiding prediction abrupt changes caused by extreme weight allocation.

[0055] The final fusion prediction output is calculated using the following formula:

[0056] in For the final fusion prediction output, These are the basic predictions for the mechanistic model. This is the amount of correction for the prediction error output by the data-driven model. For adaptive fusion weights.

[0057] Specifically, the mechanistic model provides a basic predictive framework with biochemical interpretability, while the data-driven model is only responsible for compensating for the predictive bias of the mechanistic model under the current operating conditions. The compensation magnitude is dynamically determined by the reliability of the mechanistic model. When the fermentation process is running smoothly and the operating conditions are within the effective range of the mechanistic model parameters, [the data-driven model will compensate]. Approaching 1, Approaching 0, the final output closely matches the prediction of a pure mechanistic model, demonstrating good interpretability. However, when the fermentation process encounters abnormal disturbances (such as batch differences in raw materials, equipment response deviations, sudden changes in ambient temperature, etc.) causing the operating conditions to deviate from the identification range, decline, The automatic weighting enhances the compensation effect of the data-driven model, enabling the final output to fully utilize the learning experience from similar operating conditions in historical data to compensate for the prediction bias of the mechanistic model. This adaptive weighted fusion strategy ensures that the hybrid model achieves the optimal balance between interpretability and prediction accuracy under any operating condition, fundamentally solving the shortcomings of traditional hybrid modeling methods that have fixed fusion weights and cannot adapt to the dynamic changes in the fermentation process.

[0058] Furthermore, the hybrid model has incremental learning and online update capabilities: after each batch of fermentation is completed, the system automatically includes the batch of data into the training dataset and uses the elastic weight consolidation method to incrementally train the data-driven model; the system automatically determines whether to trigger model retraining through the model performance drift detection mechanism, and seamlessly switches to the new model through the blue-green deployment strategy after the retraining is completed.

[0059] Specifically, after each batch of fermentation is completed, the system automatically triggers a model update process. First, the system incorporates the complete process data of that batch into the training dataset. This includes the real-time process parameter time series from inoculation to tank release, offline detection data at each time point, and the final batch output. During data incorporation, the system automatically performs a data quality assessment, calculating the data integrity rate for that batch. Only batch data with a integrity rate greater than 95% is included in the training set to ensure data quality for incremental training. After incorporation, the system uses a resilient weight consolidation method for incremental training of the data-driven model. The core mechanism of the resilient weight consolidation method is to protect old tasks from being overwritten by new tasks during incremental learning. Specifically, before incremental training, the resilient weight consolidation method first calculates the diagonal elements of the Fisher information matrix for each network parameter on the historical data prediction task, based on the old model. The diagonal elements of the Fisher information matrix reflect the importance of each parameter to changes in the model output. A larger Fisher information value indicates that even a small change in that parameter will have a significant impact on the model's prediction results; therefore, stronger constraints are needed in incremental training to protect the learned knowledge. During incremental training, the elastic weight consolidation method adds a regularization term to the standard regression loss function. This regularization term is the sum of the squares of the differences between the new and old values ​​of each parameter, multiplied by the corresponding Fisher information values, i.e.:

[0060] in For the regression loss on the new batch of data, The regularization intensity coefficient is . For the first Fisher information values ​​for each parameter These are the new values ​​of the parameters after incremental training. The values ​​are the old values ​​of the parameters before incremental training. Through this regularization constraint, the model applies a penalty proportional to the importance of changes in historical key parameters while fitting the features of the new batch of data, thus effectively preventing catastrophic forgetting.

[0061] Furthermore, the system automatically determines whether to trigger model retraining through a model performance drift detection mechanism. This mechanism monitors the model's prediction performance on the most recently completed batches, specifically monitoring the model's prediction error for offline detection metrics at each time point in that batch. The system maintains a sliding window, the size of which is the five most recent successful batches, and calculates the mean and standard deviation of the average prediction error for each batch within the window. When the prediction error of a newly completed batch exceeds the control limit of "historical mean + three standard deviations," the system determines that the model has experienced performance drift on that batch and triggers the retraining process. Performance drift detection also monitors the trend of error changes; when the prediction error of three consecutive batches shows a monotonically increasing trend, even if the errors of each batch have not yet exceeded the control limit, the system will determine it as potential drift and trigger retraining in advance. In addition, the system also sets up a timed triggering mechanism, automatically triggering retraining every twenty completed batches, regardless of whether performance drift is detected, to ensure that the model can periodically absorb the accumulated knowledge from all new batches. Once retraining is triggered, the system starts a training task in the background. The training uses all historical data to retrain the data-driven model instead of incremental training, in order to completely eliminate the errors and biases that may accumulate during incremental training.

[0062] After retraining, the system seamlessly switches to the new model using a blue-green deployment strategy. The blue-green deployment strategy is implemented as follows: the system maintains two independently running model service instances. The blue instance is the production model currently providing inference services, and the green instance is the new model that has completed retraining in the background. After deployment, the green instance first enters a warm-up phase, inputting the process parameter data from the most recent complete batch for inference, monitoring its resource consumption and output stability during inference. The warm-up phase continues until the simulation inference of the entire batch is completed without any abnormal alarms, after which it enters a ready-to-switch state. During the switch, the system simultaneously sends prediction requests for the new batch of data to both the blue and green instances, comparing the differences in their prediction outputs. Only when the prediction error of the green instance on the new batch of data is significantly lower than that of the blue instance, and the correlation between their outputs is greater than 0.95, does the system perform the formal switch. The formal switch employs a gradual traffic migration strategy, initially routing 10% of prediction requests to the green instance and 90% to the blue instance, monitoring the performance of the green instance under real production load. If no abnormalities are found, 20% of the traffic to the green instance is increased every ten minutes until all traffic is migrated to the green instance. The old model is retained for 24 hours as a rollback backup after all traffic is switched out. If a decrease in prediction quality, an increase in inference latency, or any abnormal alarm is detected within 24 hours after the new model goes live, the system will automatically switch all traffic back to the blue instance and record the switchover event log for technical personnel to analyze.

[0063] Furthermore, the hybrid model outputs a prediction confidence interval along with the predicted value: the uncertainty of the mechanistic model part is quantified through parameter sensitivity analysis and Monte Carlo simulation; the uncertainty of the data-driven model part is quantified through the Monte Carlo Dropout method; the two uncertainties are synthesized through variance decomposition to obtain the confidence interval of the final predicted value.

[0064] Specifically, the parameter sensitivity analysis employs the Sobol global sensitivity analysis method, the core idea of ​​which is to decompose the total variance of the model output into the contributions of each input parameter and their interactions. The mechanistic model's unidentified parameter (maximum specific growth rate) is used as an example. saturation constant Temperature coefficient Synthesis coefficient of growth-related products Non-growth-related product synthesis coefficient Ten parameters (including cell concentration, substrate concentration, and product concentration) were used as uncertain input variables, and the predicted values ​​of the model output (cell concentration, substrate concentration, and product concentration) were used as the output target. The first-order sensitivity index and the total sensitivity index of each parameter were calculated using the Sobol method. The first-order sensitivity index reflects the direct impact of changes in a single parameter on the variance of the model output, while the total sensitivity index reflects the total contribution of the parameter, both individually and in interaction with other parameters, to the variance of the output. Based on the sensitivity analysis results, the key parameters with the greatest impact on the model output (usually ten parameters) were identified. , and The overall sensitivity index is usually greater than 0.1, and these parameters are considered the main source of uncertainty propagation. Parameters with sensitivity indices below 0.01 are ignored to reduce computational complexity. After parameter screening, Monte Carlo simulations, based on the parameter covariance matrix obtained during parameter identification, establish probabilistic models of normal or log-normal distributions for each key parameter; for parameters with non-negative constraints (such as...), the overall sensitivity index is usually greater than 0.1. These parameters are considered the main source of uncertainty propagation, and parameters with sensitivity indices below 0.01 are ignored to reduce computational complexity. , (e.g., ) adopt a log-normal distribution; for parameters that can be positive or negative, a normal distribution is used. The mean of each distribution is taken as the optimal estimate of the parameter identification, and the standard deviation is taken as the standard error of the parameter identification. Then, a Monte Carlo simulation is run: ten thousand parameter combinations are generated by independently and randomly sampling from the probability distributions of each parameter. Each set of parameters is substituted into the mechanistic model for forward computation, yielding ten thousand sets of predicted output values. Statistical analysis is performed on these outputs, and the 2.5% quantile and 97.5% quantile are taken as the prediction uncertainty interval for the mechanistic model at a 95% confidence level.

[0065] The uncertainty of the data-driven model is quantified using the Monte Carlo Dropout method. The core principle of Monte Carlo Dropout is to keep the Dropout layer in the Bi-LSTM network active during inference (Dropout layers used to prevent overfitting during training are typically disabled during inference). Multiple forward passes are performed on the same input sample. Because the Dropout layer randomly discards a different subset of neurons each time, multiple forward passes will produce several slightly different output values. The statistical distribution of these output values ​​reflects the model's confidence in its predictions. Specifically, in each Dropout layer of the Bi-LSTM network, the Dropout ratio during inference is kept consistent with that during training, configured as 0.2, meaning that 20% of the neuron connections are randomly discarded during each forward pass. For the same input sample, fifty forward passes with Dropout are performed, resulting in fifty predicted values. The mean of these fifty predicted values ​​is calculated as the point estimate of the data-driven model, and its standard deviation is calculated as the uncertainty measure. The 2.5% quantile and 97.5% quantile of the fifty predicted values ​​are taken as the prediction uncertainty interval of the data-driven model at a 95% confidence level. The configuration of fifty forward propagations is based on a balance between model inference latency and uncertainty estimation accuracy; the total latency of fifty propagations is approximately 250 milliseconds, which ensures the stability of confidence interval estimation while meeting the real-time requirement of the real-time simulation unit to perform rolling predictions every five minutes.

[0066] The uncertainty of the mechanistic model and the data-driven model is synthesized using variance decomposition. The basic principle of variance decomposition is that the total prediction variance equals the sum of the prediction variances of the mechanistic model and the data-driven model, plus their covariance term. Since the modeling principles and error sources of the mechanistic model and the data-driven model are fundamentally different, the two types of errors are statistically approximately independent, and the covariance term can be ignored. Therefore, the total variance is approximately the algebraic sum of the two types of variances.

[0067] In the specific calculation, let the standard deviation of the predicted values ​​obtained from the Monte Carlo simulation of the mechanistic model be . The standard deviation of the predicted values ​​obtained by the data-driven model Monte Carlo Dropout is The standard deviation of the composite predicted value is then... The composite point estimate is the sum of the predicted value from the mechanistic model and the correction amount from the data-driven model, i.e., the point estimate output of the hybrid model. The final prediction confidence interval is the point estimate ± 1.96 × This constitutes a prediction confidence interval at a 95% confidence level. This confidence interval is output to the intelligent decision engine, which pre-defines a risk response strategy based on the confidence interval. When the predicted confidence interval width is less than 10% of the point estimate, it indicates high prediction reliability. The decision engine generates a control strategy in a conventional manner, with no limit on the action adjustment step size. When the predicted confidence interval width is between 10% and 30% of the point estimate, it indicates some uncertainty in the prediction. The decision engine applies a scaling factor of 0.7 to the action adjustment, making the control action more conservative. When the predicted confidence interval width exceeds 30% of the point estimate, it indicates insufficient prediction reliability. The decision engine pauses automatic decision output and triggers a manual verification process in the human-machine collaboration unit, where engineering technicians intervene to evaluate and manually set control parameters. Through the above uncertainty quantification and synthesis mechanism, the predicted value output by the hybrid model carries complete credibility information, providing risk perception capability for intelligent decision-making and significantly improving the system's safety and robustness under complex operating conditions.

[0068] Furthermore, the reinforcement learning algorithm adopts a deep deterministic policy gradient algorithm. The state space of the reinforcement learning algorithm includes the process parameter vector and the environmental parameter vector at the current moment. The action space includes the feed rate adjustment, temperature setpoint adjustment, pH adjuster addition rate adjustment, and dissolved oxygen control strategy switching instructions. The reward function is designed as a multi-objective optimization objective that comprehensively considers maximizing product yield, minimizing energy consumption, and minimizing process deviation.

[0069] In this embodiment, the optimal control strategy generation unit in the intelligent decision engine module uses a deep deterministic policy gradient algorithm to solve for the optimal control strategy of the fermentation process. The deep deterministic policy gradient algorithm is chosen because the action space of the fermentation process control is continuous (feeding rate, temperature setpoint, etc., are continuously adjustable parameters), and the state space has a high dimension (containing fifteen state variables). As a deep reinforcement learning algorithm under the Actor-Critic framework, the deep deterministic policy gradient algorithm can effectively handle the mapping problem from a high-dimensional continuous state space to a continuous action space. Its deterministic policy gradient method is superior to the stochastic policy gradient method in sampling efficiency and is more suitable for application scenarios like fermentation where each interaction has a high cost. The algorithm network consists of two neural network structures: a policy network and a value network. The number of nodes in the input layer of the policy network matches the dimension of the state space, and the number of nodes in the output layer matches the dimension of the action space. The hidden layer uses a fully connected structure with 256 neurons, and the activation function is the ReLU function. The number of nodes in the input layer of the value network matches the sum of the dimensions of the state space and the action space. The output layer has a single neuron, and the output value is a scalar-form estimate of the long-term cumulative reward. Both the policy network and the value network use the Adam optimizer for parameter updates. The learning rate for the policy network is configured to be 0.0001, and the learning rate for the value network is configured to be 0.001. The target network uses soft update coefficients. The configuration is 0.001.

[0070] The state space of a reinforcement learning algorithm includes a process parameter vector and an environmental parameter vector at the current moment. The process parameter vector contains eleven dimensions: temperature (degrees Celsius), pH (dimensionless), dissolved oxygen concentration (% air saturation), stirring speed (rpm), tank pressure (kPa), substrate concentration (g / L), product concentration (g / L), cell concentration (g / L), cumulative feed volume (L), exhaust gas oxygen concentration (%), and exhaust gas carbon dioxide concentration (%). The environmental parameter vector contains four dimensions: cleanroom ambient temperature (degrees Celsius), ambient relative humidity (%RH), fermentation workshop room pressure difference (Pascals), and the current fermentation stage code (a value between 0 and 1 representing batch progress, 0 for inoculation start and 1 for tank discharge end). The state space has a total of fifteen dimensions. The observations of each dimension are normalized before being input into the policy network. The normalization method is Min-Max normalization, which linearly maps the values ​​of each dimension to the interval [-1,1]. The normalization parameters are determined based on the statistical distribution of historical batch data to ensure that parameters of different dimensions have comparable importance in the policy network.

[0071] Furthermore, the action space includes four continuous action variables, corresponding to the four main control inputs in the fermentation process: feed rate adjustment (unit: ml / min, adjustment range: ±20% of the current setpoint), temperature setpoint adjustment (unit: degrees Celsius, adjustment range: ±2℃ of the current setpoint), pH adjuster addition rate adjustment (unit: ml / min, adjustment range: ±30% of the current setpoint), and dissolved oxygen control strategy switching command (this action output is a continuous value, mapped to a soft switch of three discrete strategies: increasing stirring speed, increasing aeration rate, and increasing tank pressure; the action value changes continuously within the range [-1, 1], with negative values ​​biased towards the strategy of increasing stirring speed, positive values ​​biased towards the strategy of increasing aeration rate, and values ​​close to 0 biased towards the strategy of increasing tank pressure). After the action space output is processed by safety boundary trimming, it is ensured that the adjusted values ​​of each control variable are always within the safety limits specified in the process procedure. The adjustment range of each action is constrained according to the process safety limits, and the constraint rules are as follows: the feeding rate shall not exceed 120% of the maximum allowable feeding rate and shall not be lower than 20% of the maximum rate; the temperature setpoint shall not exceed the upper and lower limits of the allowable temperature for this fermentation stage; the pH adjuster addition rate shall not exceed 150% of the maximum addition rate, and the pH value change rate within fifteen seconds shall not exceed 0.02 pH units to prevent local over-acidity or over-alkaliness.

[0072] Furthermore, the reward function is designed as a multi-objective weighted form, comprehensively considering three optimization objectives: maximizing product yield, minimizing energy consumption, and minimizing process deviation. The specific expression of the reward function is: .

[0073] in, For the product yield bonus item, the ratio of the current product concentration to the target product concentration is normalized. When the product concentration reaches the target value, the reward saturation is 1.0 to avoid over-incentivizing the control strategy to pursue unrealistic high concentration targets. As an energy consumption penalty term, it is normalized by the ratio of the sum of the current stirring power and heating / cooling power to the historical average energy consumption. Taking a negative value achieves a penalty effect of minimizing energy consumption; As a process deviation penalty item, it is the weighted sum of the absolute values ​​of the deviations of each process parameter (temperature, pH, dissolved oxygen, tank pressure) from the target set value at the current time. The weights are configured according to the degree of influence of each parameter on product quality (temperature deviation weight is 1.0, pH deviation weight is 1.2, dissolved oxygen deviation weight is 0.8, and tank pressure deviation weight is 0.5). ,in This defines the allowable fluctuation range for each parameter, used to normalize the deviations of each parameter to a comparable order of magnitude. The weighting coefficients for the three sub-items. , , This is a configurable parameter; the default configuration is... , , These correspond to weighting coefficients of 0.5 for product output, 0.2 for energy consumption, and 0.3 for process deviation, respectively. These weighting coefficients can be adjusted by the company's technical personnel through the platform's parameter configuration interface, based on the current production priorities. When the company's capacity is tight, these coefficients can be increased. Prioritize maximizing output; when enterprises face energy cost pressures, they can increase... To strengthen energy conservation orientation; when product quality stability fluctuates, it can improve... To strictly control process deviations, after each fermentation batch is completed, the system calculates the cumulative sum of the actual reward values ​​for that batch and evaluates the optimization effect of the control strategy by comparing it with historical batches.

[0074] Furthermore, the policy network of the deep deterministic policy gradient algorithm is pre-trained in the virtual environment of the digital twin of the fermentation process. Multiple simulation batches are run in parallel using the high-speed simulation capability of the digital twin model to explore policies. After the pre-training is completed, the policy network parameters are transferred to the real system for fine-tuning training.

[0075] Specifically, the deep deterministic policy gradient algorithm employs a hierarchical reinforcement learning architecture. The motivation behind this architecture is that fermentation process control is essentially a hierarchical decision-making problem—the upper layer needs to make strategic decisions based on the stage characteristics of the fermentation process (such as when to switch from batch culture to fed-batch culture, and when to induce product expression), while the lower layer needs to execute specific continuous parameter adjustments under given stage objectives. Traditional single-layer reinforcement learning couples stage identification, policy selection, and continuous control into a single agent, resulting in excessively large state and action spaces, low training efficiency, and poor policy interpretability. The hierarchical reinforcement learning architecture decouples the decision problem according to time scale and control granularity. The upper-layer policy is responsible for identifying and switching fermentation stages at a slow time scale, while the lower-layer policy is responsible for continuous adjustment of process parameters at a fast time scale. The output of the upper-layer policy serves as a sub-objective of the lower-layer policy, constraining the range of values ​​for corresponding actions in the lower-layer policy. This decomposes the complex global optimization problem into two coupled but relatively independent training sub-problems.

[0076] The upper-level strategy is responsible for identifying and switching fermentation stages. The fermentation stages are divided into four phases based on biochemical characteristics: lag phase, logarithmic growth phase, stationary phase, and decay phase. The state space of the upper-level strategy includes the macroscopic state characteristics of the current batch: current cell concentration, substrate concentration, product concentration, the ratio between the three, the estimated specific growth rate, the duration of the current stage, and the statistical characteristics of all process parameters. The action space of the upper-level strategy is a discrete action space, containing four discrete actions: maintain the current stage, switch to the next stage, initiate a feed operation, and terminate a feed operation. The discrete actions are processed continuously using the Gumbel-Softmax technique, enabling the upper-level strategy to be trained using a continuous policy gradient method within the Actor-Critic framework of deep deterministic policy gradient algorithms. Specifically, the output layer of the policy network is replaced with a Gumbel-Softmax layer, which outputs continuous discrete action samples during training and directly outputs argmax discrete actions during inference. The training rewards of the upper-level strategy and the lower-level strategy share the same multi-objective reward function. That is, the reward signal of the upper-level strategy also comes from the comprehensive evaluation of product output, energy consumption, and process deviation. However, the time scale of reward accumulation is different—the rewards of the upper-level strategy are accumulated at intervals based on stage switching points to encourage it to make stage switching decisions at the right time. The output of the upper-level strategy is specifically a stage switching instruction, which is passed to the lower-level strategy as a sub-objective.

[0077] The lower-level strategy is responsible for continuously adjusting process parameters under given sub-objective constraints. The state space of the lower-level strategy contains all fifteen-dimensional process parameter vectors and environmental parameter vectors at the current moment, consistent with the state space defined in the above embodiments. The action space of the lower-level strategy is the same as the four-dimensional continuous action space defined in the above embodiments, including feed rate adjustment, temperature setpoint adjustment, pH adjuster addition rate adjustment, and dissolved oxygen control strategy switching instructions. During inference, the lower-level strategy obtains the sub-objectives output by the upper-level strategy and applies corresponding value range constraints to the action space based on the sub-objectives. The specific sub-target constraint mapping rules are as follows: When the sub-target is the "lag period", the feed rate adjustment is forcibly constrained to zero (feeding prohibited), the temperature setpoint adjustment is limited to the lower limit of the high-temperature range (promoting cell recovery), and the constraint on the pH adjuster addition rate adjustment is relaxed to a larger range; when the sub-target is the "logarithmic growth phase - feed mode", the constraint range of the feed rate adjustment is the positive range, and the dissolved oxygen control strategy switching command is biased towards increasing the stirring speed range; when the sub-target is the "stationary phase - product accumulation mode", the constraint range of the feed rate adjustment is to maintain or slowly increase, the dissolved oxygen control strategy switching command is biased towards increasing the aeration rate range, and the constraint range of the temperature setpoint adjustment is the lower limit of the low-temperature range; when the sub-target is the "decline period", the constraint range of the temperature setpoint adjustment is in the cooling direction, and the feed rate adjustment is forcibly constrained to be negative or zero. In actual execution, the lower-level strategy's output action is first clipped through the value range corresponding to the upper-level sub-target, and then clipped again through the process safety limits before being sent to the execution mechanism. The upper-layer strategy is scheduled to execute every 30 to 60 minutes, using an event-triggered approach. In addition to fixed-time triggers, a state-change-based trigger condition is configured: when the rate of change of the specific growth rate is negative three times consecutively and the cumulative decrease exceeds 20%, the system automatically triggers an additional decision from the upper-layer strategy. The lower-layer strategy is scheduled to execute every five minutes, consistent with the rolling prediction frequency of the real-time simulation unit, ensuring that the lower-layer strategy can respond promptly to dynamic changes in the fermentation process. Communication between the two strategies uses an asynchronous message queue model—the decision results of the upper-layer strategy are published through the platform-level message middleware, and the lower-layer strategy obtains the latest sub-targets through a subscription method. This decoupled operation of the two strategies ensures that the training of the upper-layer strategy is not affected by the real-time execution of the lower-layer strategy. The hierarchical reinforcement learning architecture significantly reduces the learning complexity of a single-layer agent and improves the interpretability and training efficiency of the strategy.

[0078] Furthermore, the platform layer also includes an intelligent control execution module, comprising: The closed-loop control unit is used to distribute the optimal control strategy generated by intelligent decision-making to the distributed control system or programmable logic controller through a standardized interface, forming a complete closed-loop control link from perception to analysis, from analysis to decision-making, and from decision-making to execution. Specifically, the closed-loop control unit distributes the optimal control strategy generated by the intelligent decision engine to the distributed control system or programmable logic controller (PLC) through a standardized interface, forming a complete closed-loop control chain from perception to analysis, from analysis to decision, and from decision to execution. The optimal control strategy output by the intelligent decision engine is in a structured data format, containing the setpoint of each control variable and the corresponding execution time window. The closed-loop control unit first performs protocol adaptation on the control strategy—automatically selecting the corresponding standardized interface protocol based on the type of control system attached to the target actuator. For distributed control systems, the OPC UA protocol is used for communication. The closed-loop control unit acts as an OPC UA client, writing the setpoint to the corresponding variable node of the distributed control system through the OPC UA write service, and simultaneously reading the actual value fed back by the distributed control system in real time through the OPC UA subscription mechanism. For PLCs, the Modbus TCP protocol is used for communication. The closed-loop control unit writes the setpoint to the corresponding holding register address according to the Modbus mapping table. After the control command is issued, the closed-loop control unit enters the execution effect monitoring state, monitoring the execution effect of the control command by comparing the deviation between the issued setpoint and the actual feedback value of the actuator in real time. When the deviation exceeds a preset threshold, the closed-loop control unit automatically triggers a control strategy adjustment process—based on the current deviation direction and magnitude, it generates a compensation adjustment amount and reissues it until the deviation converges to the allowable range. If the deviation still fails to converge after three consecutive adjustments, the closed-loop control unit generates an execution anomaly alarm and triggers the intervention process of the human-machine collaboration unit. The response time of the entire closed-loop control link, from the decision output to the actual response of the actuator, does not exceed two seconds, of which the protocol conversion and network transmission delay does not exceed five hundred milliseconds. The physical response time of the actuator depends on the specific equipment.

[0079] The human-machine collaboration unit is used to show the engineering and technical personnel the decision basis and expected effects before the execution of key control decisions. After the technical personnel confirm or modify the control strategy through the human-machine interaction terminal, the control command can be issued to the execution mechanism. Specifically, critical control decisions are defined as including the following types: initial switching decisions during the fermentation stage, significant cross-stage adjustments to process parameters, any decisions that may affect product quality, and decisions where the prediction confidence interval of the decision engine output exceeds 20% of the point estimate. For decisions deemed critical, the human-machine collaboration unit intervenes after the decision engine generates the control strategy but before the closed-loop control unit issues instructions, presenting the decision basis and expected effects to technical personnel through a human-machine interaction terminal.

[0080] The demonstration content specifically includes: the current state prediction curve of the digital twin model, the optimization objectives of the control strategy and the current achievement status of each objective, the name of the control variable to be adjusted, its current value, the suggested adjustment amount and direction, and the time curve of the expected effect. The interface also provides three operation options for the adjustment suggestions: "Confirm Execution"—directly adopt the suggestion and issue the control command to the actuator; "Modify and Execute"—technicians modify the set value of a control variable on the visual interface by dragging a slider or manually inputting (the modification range is constrained by process safety limits), and issue the command after confirmation; "Reject Execution"—temporarily refrain from issuing the control command, the system records the reason for rejection and maintains the current control state unchanged. All confirmation, modification, or rejection operations are recorded in an operation log, including the operator's identification, operation time, operation type, and parameter values ​​before and after the operation. These logs are automatically pushed to the blockchain traceability module for tamper-proof storage, meeting the compliance requirements for operational traceability in the pharmaceutical production process.

[0081] The multi-mode switching unit supports flexible switching between three control modes, including fully automatic mode, semi-automatic mode, and manual intervention mode.

[0082] Furthermore, in fully automatic mode, the intelligent decision engine generates and executes the control strategy entirely. Control commands are automatically issued to the actuators after passing through the key decision review process of the human-machine collaboration unit, eliminating the need for manual confirmation of non-critical decisions. Fully automatic mode is suitable for mature production stages where operation is stable and the strategy has been fully validated. In semi-automatic mode, the intelligent decision engine generates the control strategy, but all control commands must be displayed by the human-machine collaboration unit and confirmed by technicians before execution. Technicians can modify or reject any decision on the confirmation interface. Semi-automatic mode is suitable for the initial operation stage of new fermentation varieties or scenarios where the strategy has not yet been fully validated. In manual intervention mode, the intelligent decision engine continues to run in the background and generate suggested strategies, but all control commands are issued directly to the actuators by technicians who manually set the values ​​of each control parameter through the human-machine interaction terminal. The suggestions are for reference only. Manual intervention mode is suitable for emergency handling, equipment debugging, or scenarios where technicians wish to perform personalized and precise control on specific batches. Multi-mode switching is performed through a unified management interface with hierarchical access control: switching between fully automatic and semi-automatic modes requires confirmation from operators with engineer privileges, while enabling and disabling manual intervention modes requires confirmation from operators with senior engineer privileges. Mode switching operation records are automatically saved to the system log. Each record includes the mode before switching, the mode after switching, the operator performing the switch, the switching time, and the reason for the switch, ensuring complete audit traceability for any change in control permissions. Through flexible switching between these three control modes and hierarchical access control, the intelligent control execution module achieves a high degree of automation while retaining necessary human controllability and security fallback capabilities.

[0083] Furthermore, the platform layer also includes a blockchain traceability module, comprising: a key data on-chain unit for writing key process parameters and operation records during fermentation into the blockchain network; an immutable unit for ensuring that on-chain data is immutable once recorded using the distributed ledger technology and hash chain structure of the blockchain; and an audit traceability unit for providing a standardized data query interface, supporting drug regulatory authorities to query the complete process records of any batch through the interface. The blockchain traceability module adopts a consortium blockchain architecture, with on-chain nodes including enterprise core system nodes, regulatory department nodes, and third-party audit institution nodes; the consensus mechanism adopts a practical Byzantine fault-tolerant algorithm; and smart contracts are used to automatically execute data on-chain rules, including data format verification, timestamp verification, and on-chain permission confirmation.

[0084] Furthermore, the platform layer also includes a predictive maintenance module, comprising: an equipment health assessment unit, used to construct an equipment health status assessment model based on the operating current data of the fermenter drive motor, bearing vibration acceleration data, gearbox temperature data, and sealing component temperature data; the model output includes an equipment health score and a predicted remaining service life value; an early warning unit, used to automatically generate early warning information when the equipment health score is lower than a preset threshold; and a maintenance suggestion unit, used to recommend the optimal maintenance timing and maintenance plan based on the current health status of the equipment and historical maintenance records.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet, characterized in that, It includes the device layer, edge layer, and platform layer, among which: The equipment layer is located in the biopharmaceutical fermentation workshop and is used to realize the physical connection and data acquisition of fermenters, seed tanks, sensors and actuators, and transmit the acquired data to the edge layer through an industrial intelligent gateway; The edge layer is set up in the workshop and is used to receive raw data transmitted from the equipment layer, perform real-time data preprocessing, anomaly detection and emergency control, and upload the preprocessed data to the platform layer. The platform layer is located in the enterprise data center or cloud server and is used to receive pre-processed data transmitted from the edge layer, build the core functions of the industrial internet platform, and realize multi-source data fusion, digital twin of fermentation process, intelligent decision-making, intelligent control execution, blockchain traceability and predictive maintenance. The digital twin of the fermentation process includes: Mechanistic models were constructed based on microbial growth kinetics equations and metabolic network models. Train a data-driven model using historical batch fermentation data; The mechanistic model and the data-driven model are integrated using an error correction method to form a hybrid model; The prediction is based on the output of the mechanistic model and compensated by the prediction error correction amount output by the data-driven model. The algebraic sum of the two is used as the final prediction output. Based on real-time data collected during the current fermentation process, the data is input into the mixing model for continuous rolling prediction. The intelligent decision-making process includes: using a reinforcement learning algorithm to solve for the optimal control strategy based on the prediction results of the current fermentation state variables and the digital twin of the fermentation process.

2. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The platform layer also includes a multi-source data acquisition and fusion module, including: The real-time parameter acquisition unit is used to receive and parse parameters such as temperature, pH value, dissolved oxygen concentration, stirring speed, tank pressure, exhaust gas oxygen concentration, exhaust gas carbon dioxide concentration, and cumulative feed amount from the sensor network unit in real time through the acquisition link composed of IoT gateway and edge computing node. The offline data entry unit provides a standardized data entry interface, allowing quality control laboratory personnel to enter potency test data, residual sugar concentration test data, amino acid concentration test data, and bacterial cell concentration test data into the system. The offline data entry timestamp is automatically associated with the sampling timestamp. The environmental data access unit is used to establish a data connection with the enterprise's cleanroom building automation system through a standard API interface, and to obtain real-time data on the cleanroom's ambient temperature, relative humidity, and room pressure difference. The equipment operation data acquisition unit is used to acquire the operating current of the fermenter drive motor, valve opening status, sterilization records, and equipment start-up and stop logs via OPC UA protocol or Modbus protocol; The data governance unit is used to establish unified data coding standards and data cleaning rules, and to perform format standardization, unit unification, timestamp calibration, and outlier marking on data from different data sources.

3. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The mechanistic model is based on a cell growth model constructed using the Monod equation and a product synthesis model constructed using the Luedeking-Piret equation. The model parameters are identified using nonlinear least squares method based on experimental data from historical batches.

4. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The data-driven model adopts a bidirectional long short-term memory network structure. The network contains three bidirectional recurrent hidden layers, each containing 128 hidden nodes. The input features of the model include the time series of process parameters over the past twelve hours and the offline detection index value at the current moment.

5. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The error correction method adopts an adaptive weighted fusion strategy: at each moment of model prediction, the system calculates the credibility index of the mechanism model at the current moment. When the credibility of the mechanism model is high, the correction weight of the data-driven model is reduced. When the reliability of the mechanistic model is low, the correction weight of the data-driven model is increased.

6. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The hybrid model has incremental learning and online update capabilities: after each batch of fermentation is completed, the system automatically includes the batch of data into the training dataset and uses the elastic weight consolidation method to incrementally train the data-driven model; the system automatically determines whether to trigger model retraining through the model performance drift detection mechanism, and seamlessly switches to the new model through the blue-green deployment strategy after the retraining is completed.

7. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 6, characterized in that, The hybrid model outputs both the predicted value and the predicted confidence interval: the uncertainty of the mechanistic model part is quantified by parameter sensitivity analysis and Monte Carlo simulation; the uncertainty of the data-driven model part is quantified by the Monte Carlo Dropout method; the two uncertainties are synthesized by the variance decomposition method to obtain the confidence interval of the final predicted value.

8. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The reinforcement learning algorithm adopts a deep deterministic policy gradient algorithm. The state space of the reinforcement learning algorithm includes the process parameter vector and the environmental parameter vector at the current moment. The action space includes the feed rate adjustment, temperature setpoint adjustment, pH adjuster addition rate adjustment, and dissolved oxygen control strategy switching instructions. The reward function is designed as a multi-objective optimization objective that comprehensively considers maximizing product yield, minimizing energy consumption, and minimizing process deviation.

9. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 8, characterized in that, The policy network of the deep deterministic policy gradient algorithm is pre-trained in the virtual environment of the digital twin of the fermentation process. Multiple simulation batches are run in parallel using the high-speed simulation capability of the digital twin model to explore policies. After the pre-training is completed, the policy network parameters are transferred to the real system for fine-tuning training.

10. The intelligent control platform for biopharmaceutical fermentation processes based on the Industrial Internet as described in claim 1, characterized in that, The platform layer also includes an intelligent control execution module, comprising: The closed-loop control unit is used to distribute the optimal control strategy generated by intelligent decision-making to the distributed control system or programmable logic controller through a standardized interface, forming a complete closed-loop control link from perception to analysis, from analysis to decision-making, and from decision-making to execution. The human-machine collaboration unit is used to show the decision basis and expected effects to engineering and technical personnel before the execution of key control decisions. After the technical personnel confirm or modify the control strategy through the human-machine interaction terminal, the control command can be issued to the execution mechanism. The multi-mode switching unit supports flexible switching between three control modes, including fully automatic mode, semi-automatic mode, and manual intervention mode.

Images