A bioreactor deep learning control system based on multi-source heterogeneous data

Abstract

The application discloses a kind of biological reactor deep learning control systems based on multi-source heterogeneous data, including the multi-source heterogeneous data acquisition module, data preprocessing and fusion module, deep learning control model module, execution adjustment module, feedback correction module, monitoring interaction module, data storage module and safety protection module connected in turn.Effectively eliminate noise, missing, interference and data fragmentation problem, greatly improve data quality and feature expression ability by Kalman filtering, outlier correction, hierarchical fusion and multimodal Transformer architecture, through the construction multi-source heterogeneous data driven, deep learning and causal inference fusion biological reactor full closed loop intelligent control system, realize the deep fusion of multi-source heterogeneous and multi-modal data, the multi-type data such as physical, chemical, biological, equipment, environment and spectrum, image is uniformly represented.

Description

Technical Field

[0001] This invention relates to the field of bioreactor control system technology, specifically a deep learning control system for bioreactors based on multi-source heterogeneous data. Background Technology

[0002] As the core equipment in the biomanufacturing process, the operating status of bioreactors directly determines product yield, quality, and production efficiency, making them a key carrier for achieving efficient, stable, and scalable production in the biomanufacturing industry. Bioreaction processes are characterized by high nonlinearity, time-varying nature, multivariate coupling, and complex reaction mechanisms. Traditional control methods, relying mainly on a limited number of directly measurable process parameters, struggle to cope with the dynamic changes in cellular metabolic networks, resulting in low control precision, weak anti-interference capabilities, and an inability to meet the precision requirements of modern biomanufacturing.

[0003] With the development of biomanufacturing technology, various advanced sensors, detection equipment, and information systems are widely used in bioreactors, forming a multi-source heterogeneous process data system. This data covers different types, dimensions, and formats, specifically including physical parameters, chemical parameters, biological parameters, as well as equipment operating parameters and environmental parameters. Furthermore, the data exhibits multi-scale characteristics, integrating time spans from seconds to days, spatial scales from single cells to reactors, and various data formats such as scalars, spectra, and images.

[0004] Currently, existing bioreactor control systems have significant shortcomings in processing multi-source heterogeneous data: First, they lack effective data fusion mechanisms, with various types of data stored independently and in fragmented form, making it impossible to uncover the intrinsic relationships between data and thus hindering the realization of data value. Second, their data preprocessing capabilities are insufficient, failing to effectively eliminate sensor noise, bubble interference, data gaps, outliers, and other issues, affecting the accuracy of the control model. Third, control models are mostly based on traditional mechanistic modeling or simple machine learning methods, making it difficult to capture the complex nonlinear relationships of bioreaction processes, resulting in poor generalization ability and an inability to achieve accurate prediction and adaptive control of the reaction process. Furthermore, traditional PID control struggles to cope with the highly nonlinear and time-varying characteristics of cell culture, and large-scale bioreactors also face technical bottlenecks such as the difficulty in balancing mixing efficiency and aseptic assurance, and the challenge of process scale-up. Moreover, due to the singularity of data or the single control objective, a fully closed-loop deep learning control system driven by multi-source heterogeneous data has not been formed, failing to fundamentally solve the problem of accurate control of bioreaction processes. Therefore, we propose a deep learning control system for bioreactors based on multi-source heterogeneous data. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning control system for bioreactors based on multi-source heterogeneous data, so as to solve the problems that need to be solved in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning control system for a bioreactor based on multi-source heterogeneous data, comprising a multi-source heterogeneous data acquisition module, a data preprocessing and fusion module, a deep learning control model module, an execution adjustment module, a feedback correction module, a monitoring and interaction module, a data storage module, and a safety protection module connected in sequence.

[0007] The multi-source heterogeneous data acquisition module is used to collect physical, chemical, biological, equipment operating, and environmental parameters during the operation of the bioreactor.

[0008] The data preprocessing and fusion module includes a data preprocessing unit and a multi-source data fusion unit;

[0009] The data preprocessing unit uses Kalman filtering to eliminate noise and bubble interference, interpolation to fill in missing data, and normalization processing is performed after identifying and correcting outliers in the data.

[0010] The multi-source data fusion unit includes a data layer, a feature layer, and a decision layer. The feature layer fusion uses CNN to extract local features and LSTM to extract time series features. It combines an attention mechanism to allocate feature weights and uses a multimodal Transformer unit to achieve unified representation of time series data, image data, and spectral data. The decision layer fusion combines biological reaction mechanisms and key features, and standardizes and aligns data across reactors.

[0011] The deep learning control module includes a model training unit, a model inference unit, and integrates a soft measurement unit, a transfer learning unit, a green and low-carbon optimization unit, and a digital twin simulation unit.

[0012] The execution adjustment module includes a controller, an actuator, and a drive unit. The controller adopts a PLC and embedded chip combined architecture and is used to realize material stirring, temperature control, and pH adjustment. The drive unit is used to drive the actuator to work and adopts closed-loop feedback control.

[0013] The feedback correction module includes a deviation calculation unit, an adaptive correction unit, and a model update unit;

[0014] The deviation calculation unit is used to calculate the deviation between the real-time running parameters and the model prediction parameters. The adaptive correction unit uses the PI algorithm to fine-tune the control parameters and combines the adaptive weight adjustment strategy to adapt to the time-varying reaction. The model update unit uses incremental learning to update the model parameters. When the deviation exceeds the preset threshold, an abnormal warning is triggered.

[0015] The monitoring and interaction module includes a display unit, an operation unit, and an alarm unit;

[0016] The data storage module uses a distributed database, and the security protection module includes a hardware electromagnetic shielding module, a software data encryption module, and a hierarchical access control module.

[0017] Existing bioreactor control systems suffer from significant shortcomings in handling multi-source heterogeneous data: First, they lack effective data fusion mechanisms, resulting in fragmented and independent data storage that fails to uncover intrinsic correlations and undermines data value. Second, their data preprocessing capabilities are insufficient, failing to effectively eliminate sensor noise, bubble interference, missing data, and outliers, thus affecting the accuracy of control models. Third, control models are often based on traditional mechanistic modeling or simple machine learning methods, making it difficult to capture the complex nonlinear relationships in bioreactor processes. They also exhibit poor generalization ability, hindering accurate prediction and adaptive control of the reaction process. Furthermore, traditional PID control struggles to handle cell culture. The bioreactor is characterized by high nonlinearity and time-varying characteristics. Furthermore, large-scale bioreactors face technical bottlenecks such as the difficulty in balancing mixing efficiency and aseptic assurance, and challenges in process scale-up. Moreover, due to the singularity of data or a single control objective, a fully closed-loop deep learning control system driven by multi-source heterogeneous data has not been established, failing to fundamentally solve the problem of precise control of the bioreactor process. This invention constructs a fully closed-loop intelligent control system for bioreactors that integrates multi-source heterogeneous data driving, deep learning, and causal inference. Compared to traditional PID control, single-data-driven control, and conventional machine learning control schemes, this system achieves comprehensive improvements in data utilization efficiency, control accuracy, adaptability, process applicability, and production efficiency. This invention represents a breakthrough, possessing significant technological advantages and industrial application value. It achieves deep fusion of multi-source heterogeneous and multimodal data, unifying the representation of various data types including physical, chemical, biological, equipment, environmental, spectral, and image data. Through Kalman filtering, outlier correction, hierarchical fusion, and a multimodal Transformer architecture, it effectively eliminates noise, missing values, interference, and data fragmentation, significantly improving data quality and feature representation capabilities. Employing a multi-task deep learning architecture, it simultaneously achieves key parameter prediction, reaction state identification, and control parameter optimization. Combined with reinforcement learning and a green and low-carbon objective function, it can increase production while reducing energy consumption and carbon emissions. The entire system integrates a soft-sensor virtual sensing unit to achieve online estimation of difficult-to-measure indicators such as cell activity, specific growth rate, metabolic flux, and substrate conversion rate, reducing monitoring costs. Moreover, this invention introduces a causal inference decision-making mechanism, breaking through the limitations of traditional pure data association fitting. It constructs a true causal link between operating parameters and performance indicators through a causal discovery algorithm, eliminating spurious correlations and indirect correlation interference, making control decisions more interpretable and robust, avoiding misadjustment caused by data fluctuations, significantly improving long-term stability under complex reaction scenarios, and combining digital twins and CFD simulation to realize control strategy pre-simulation and risk avoidance in advance, further improving operational safety and batch consistency.

[0018] As a further description of the above technical solution:

[0019] The multi-source heterogeneous data acquisition module includes multiple sets of sensors, a data acquisition card, a communication module, and a data caching unit. The sensors include physical sensors, chemical sensors, biological sensors, and equipment status sensors. Multiple sets of sensors enable multimodal data acquisition. The communication module supports multiple industrial communication protocols. The data caching unit adopts a distributed caching architecture for temporary storage and fast retrieval of raw data.

[0020] As a further description of the above technical solution:

[0021] The model training unit adopts a multi-task deep learning model, which includes three parallel task branches: key parameter prediction branch, reaction state recognition branch, and control parameter optimization branch.

[0022] The key parameter prediction branch is based on an LSTM network to predict cell density, product concentration, dissolved oxygen, and pH parameters within the next 1-60 minutes.

[0023] The reaction state recognition branch is based on CNN and attention mechanism and is used to identify the growth period, stationary period, death period, as well as the abnormal states of contamination, parameter drift, and equipment failure.

[0024] The control parameter optimization branch is based on a reinforcement learning algorithm to extend the green and low-carbon objective function.

[0025] The model training unit uses a hybrid dataset combined with the Adam algorithm and an early stopping strategy to optimize parameters for model training. Biological response mechanism constraints are introduced during model training.

[0026] By integrating soft measurement units to construct a deep learning virtual sensing model, real-time online estimation of difficult-to-measure indicators such as cell specific growth rate, cell activity, metabolic flux, substrate conversion rate, and product yield can be achieved.

[0027] The transfer learning unit adopts a domain adaptive algorithm, and the digital twin simulation unit establishes a 1:1 digital twin of the bioreactor, synchronizing the physical entity's operating status in real time. Combined with CFD flow field simulation, it enables control strategy pre-playing, simulation verification, and early risk avoidance.

[0028] As a further description of the above technical solution:

[0029] The feedback correction module also includes a newly added fault knowledge graph and causal inference unit, which automatically locates the root cause of anomalies based on real-time data and historical fault database, including sensor drift, insufficient oxygen supply, abnormal feeding, stirring failure, sealing failure, and risk of bacterial contamination, and outputs treatment suggestions.

[0030] As a further description of the above technical solution:

[0031] The display unit combines an industrial touchscreen with a host computer for real-time data display. The operation unit switches between manual and automatic modes. The alarm unit uses an audible and visual alarm.

[0032] As a further description of the above technical solution:

[0033] It also includes an edge-cloud collaboration module and an early warning module. The edge-cloud collaboration module includes an edge terminal and a cloud terminal. The edge terminal is used for real-time control, low-latency inference, local closed-loop regulation, and data caching and preprocessing. The cloud terminal is used for big data training and simultaneous scheduling of multiple reactor clusters. The early warning module collects parameter data and evaluates the data in real time based on a deep learning model to achieve early warning and proactive intervention.

[0034] As a further description of the above technical solution:

[0035] It also includes a causal inference decision module, which is integrated between the deep learning control model module and the feedback correction module. It is used to construct a causal relationship model between operating parameters and key performance indicators in the biological reaction process. The causal inference decision module automatically mines the real causal links between parameters from multi-source heterogeneous data through causal discovery algorithms, clarifies the strength and direction of the real influence of control parameters on the data, eliminates false associations, pseudo-correlation and indirect association interference, and forms causal confidence weights.

[0036] As a further description of the above technical solution:

[0037] A deep learning control method for bioreactors based on multi-source heterogeneous data includes the following steps:

[0038] Step S1: Synchronously collect raw operating data of the bioreactor through the multi-source heterogeneous data acquisition module, adaptively adjust the sampling frequency and store it in the data cache unit;

[0039] Step S2: The data preprocessing and multimodal fusion module performs noise removal, missing value imputation, outlier correction and normalization on the original running data, and obtains a unified feature vector through hierarchical fusion and multimodal Transformer strategy;

[0040] Step S3: Output key parameter prediction and reaction state identification results through the multi-task deep learning model, combine them with the causal inference unit to determine the real impact parameters, and output the optimized control parameters;

[0041] Step S4: Real-time synchronization of the physical reactor state using a 1:1 digital twin, combined with CFD flow field simulation to complete control strategy pre-playing and risk avoidance;

[0042] Step S5: Convert the optimized control parameters into control signals to drive the actuators to adjust the operating state of the bioreactor, and use multiple structures to ensure a sterile environment;

[0043] Step S6: Collect and execute the adjusted operating parameters, calculate the deviation from the model prediction value, fine-tune the control parameters and update the model using incremental learning, and automatically locate the root cause of the anomaly through the fault knowledge graph;

[0044] Step S7: Complete local real-time closed-loop control through the edge terminal, and realize model iteration, transfer learning and multi-reactor cluster scheduling through the cloud;

[0045] Step S8: Real-time display of system operating status, digital twin screen and low-carbon indicators; trigger audible and visual alarms and sterility risk warnings when abnormalities occur.

[0046] Step S9: Repeat steps S1-S8 to form a fully closed-loop intelligent control until the biological reaction is completed.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) First, this invention constructs a fully closed-loop intelligent control system for bioreactors that integrates multi-source heterogeneous data, deep learning, and causal inference, thereby achieving deep integration of multi-source heterogeneous and multi-modal data. It unifies the representation of various types of data, including physical, chemical, biological, equipment, environmental, spectral, and image data. Through Kalman filtering, outlier correction, hierarchical fusion, and multi-modal Transformer architecture, it effectively eliminates noise, missing data, interference, and data fragmentation, significantly improving data quality and feature expression capabilities.

[0049] (2) Furthermore, the present invention adopts a multi-task deep learning architecture to simultaneously realize key parameter prediction, reaction state identification and control parameter optimization. Combined with reinforcement learning and green and low-carbon objective function, it can reduce energy consumption and carbon emissions while increasing production. Moreover, the entire system integrates a soft measurement virtual sensing unit to realize online estimation of difficult-to-measure indicators such as cell activity, specific growth rate, metabolic flux and substrate conversion rate, thereby reducing monitoring costs. Furthermore, the present invention introduces a causal inference decision-making mechanism to break through the limitations of traditional pure data association fitting. It constructs a real causal link between operating parameters and performance indicators through a causal discovery algorithm, eliminates pseudo-correlation and indirect correlation interference, and makes control decisions more interpretable and robust. It avoids misadjustment caused by data fluctuations, significantly improves long-term stability under complex reaction scenarios, and combines digital twins and CFD simulation to realize control strategy pre-playing and risk avoidance in advance, further improving operational safety and batch consistency.

[0050] (3) Finally, in terms of adaptability and generalization, the present invention has the ability to learn across batches, reactors and strains. The feedback correction module integrates fault knowledge graph and causal tracing, which can automatically locate the root causes of abnormalities such as sensor drift, insufficient oxygen supply, abnormal feeding, sealing failure and risk of contamination and provide disposal suggestions, effectively reducing the batch failure rate. In addition, the execution adjustment module adopts parameter smooth output and high-precision closed-loop control, which greatly improves the adjustment accuracy of temperature, pH, speed and ventilation. At the same time, it takes into account cell shear protection and mixing efficiency. With multiple aseptic protection structures, it significantly improves cell survival rate and culture stability. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the system principle of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] Please see Figure 1 The present invention provides a technical solution: a deep learning control system for a bioreactor based on multi-source heterogeneous data, comprising a multi-source heterogeneous data acquisition module, a data preprocessing and fusion module, a deep learning control model module, an execution adjustment module, a feedback correction module, a monitoring and interaction module, a data storage module, and a safety protection module connected in sequence.

[0055] The multi-source heterogeneous data acquisition module is used to collect physical, chemical, biological, equipment operating, and environmental parameters during the operation of the bioreactor.

[0056] The data preprocessing and fusion module includes a data preprocessing unit and a multi-source data fusion unit;

[0057] The data preprocessing unit uses Kalman filtering to eliminate noise and bubble interference, interpolation to fill in missing data, and normalization processing is performed after identifying and correcting outliers in the data.

[0058] The multi-source data fusion unit includes a data layer, a feature layer, and a decision layer. The feature layer fusion uses CNN to extract local features and LSTM to extract time series features. It combines an attention mechanism to allocate feature weights and uses a multimodal Transformer unit to achieve unified representation of time series data, image data, and spectral data. The decision layer fusion combines biological reaction mechanisms and key features, and standardizes and aligns data across reactors.

[0059] The deep learning control module includes a model training unit, a model inference unit, and integrates a soft measurement unit, a transfer learning unit, a green and low-carbon optimization unit, and a digital twin simulation unit.

[0060] The execution adjustment module includes a controller, an actuator, and a drive unit. The controller adopts a PLC and embedded chip combined architecture and is used to realize material stirring, temperature control, and pH adjustment. The drive unit is used to drive the actuator to work and adopts closed-loop feedback control.

[0061] The feedback correction module includes a deviation calculation unit, an adaptive correction unit, and a model update unit;

[0062] The deviation calculation unit is used to calculate the deviation between the real-time running parameters and the model prediction parameters. The adaptive correction unit uses the PI algorithm to fine-tune the control parameters and combines the adaptive weight adjustment strategy to adapt to the time-varying reaction. The model update unit uses incremental learning to update the model parameters. When the deviation exceeds the preset threshold, an abnormal warning is triggered.

[0063] The monitoring and interaction module includes a display unit, an operation unit, and an alarm unit;

[0064] The data storage module uses a distributed database, and the security protection module includes a hardware electromagnetic shielding module, a software data encryption module, and a hierarchical access control module.

[0065] The system employs a multi-module collaborative closed-loop control architecture to comprehensively improve the operational stability, control precision, and intelligence level of the bioreactor. The multi-source heterogeneous data acquisition module can simultaneously acquire physical, chemical, biological, equipment, and environmental parameters, enabling comprehensive perception of the reaction process and providing a complete data foundation for precise control. The data preprocessing and fusion module utilizes Kalman filtering, interpolation, anomaly correction, and normalization to effectively remove noise, fill in missing data, and standardize data formats. Furthermore, through three-level fusion (data layer, feature layer, and decision layer), combined with CNN, LSTM, attention mechanisms, and multimodal Transformers, it achieves unified representation of time-series, image, and spectral data, improving feature effectiveness and cross-reactor data compatibility. The deep learning control module enables multi-task collaborative inference, enhancing model generalization ability and control rationality. Finally, the execution and regulation module uses a PLC and embedded chip architecture with closed-loop feedback control to control stirring, temperature, and pH. The system features high-precision and stable adjustment of parameters, while the feedback correction module can calculate deviations in real time, adaptively fine-tune parameters, and continuously update the model through incremental learning, enhancing the system's adaptability and anti-interference capabilities. The monitoring and interaction module enables status visualization, convenient human-machine operation, and timely alarms for anomalies, improving operational efficiency. The data storage and security protection module employs distributed storage and a combination of hardware and software protection to ensure data security, reliable storage, and stable system operation, achieving intelligent, standardized, and highly reliable closed-loop control of the bioreactor as a whole.

[0066] Example 2:

[0067] Please see Figure 1 The multi-source heterogeneous data acquisition module includes multiple sets of sensors, a data acquisition card, a communication module, and a data caching unit. The sensors include physical sensors, chemical sensors, biological sensors, and equipment status sensors. Multiple sets of sensors enable multimodal data acquisition. The communication module supports multiple industrial communication protocols. The data caching unit adopts a distributed caching architecture for temporary storage and fast retrieval of raw data.

[0068] The setup of multiple sensors allows for real-time monitoring and acquisition of multiple data points. Furthermore, the multi-source heterogeneous data acquisition module supports manual import of on-site experimental data, historical batch data, and cross-reactor process data. It also features an adaptive sampling frequency function, automatically adjusting the acquisition frequency according to the reaction status, with high-frequency acquisition during the logarithmic growth phase and low-frequency acquisition during the steady-state phase, reducing data redundancy, computing power consumption, and sensor wear.

[0069] The model training unit employs a multi-task deep learning model, which includes three parallel task branches: a key parameter prediction branch, a reaction state recognition branch, and a control parameter optimization branch. The key parameter prediction branch, based on an LSTM network, predicts cell density, product concentration, dissolved oxygen, and pH within the next 1-60 minutes. The parameter and reaction state recognition branch, based on CNN and attention mechanism, is used to identify the growth phase, stationary phase, decline phase, as well as contamination, parameter drift, and abnormal equipment malfunction states. The control parameter optimization branch, based on reinforcement learning algorithm, aims to maximize yield, optimize quality, and minimize energy consumption, extending the green and low-carbon objective function to achieve synergistic optimization of high yield, low consumption, and low carbon. The model training unit uses a hybrid dataset combined with the Adam algorithm and early shutdown strategy to optimize parameters for model training. During model training, biological reaction mechanism constraints are introduced. The integrated soft measurement unit constructs a deep learning virtual sensing model to achieve real-time online estimation of difficult-to-measure indicators such as cell specific growth rate, cell activity, metabolic flux, substrate conversion rate, and product yield, reducing dependence on imported sensors. The transfer learning unit uses a domain adaptive algorithm to achieve rapid transfer of process models across batches, reactors, and strains, enabling rapid modeling with small samples and significantly shortening the development cycle of new processes. The digital twin simulation unit establishes a 1:1 digital twin of the bioreactor, synchronizing the physical entity's operating status in real time. Combined with CFD flow field simulation, it enables control strategy pre-playing, simulation verification, and early risk avoidance.

[0070] The system employs a multi-task deep learning model, with three parallel branches working collaboratively: key parameter prediction, reaction state identification, and control parameter optimization. This enables precise prediction, intelligent identification, and autonomous regulation of biological reaction processes simultaneously. The key parameter prediction branch, based on an improved LSTM network, can accurately predict cell density, product concentration, dissolved oxygen, and pH 1-60 minutes in advance. Key indicators provide a basis for feedforward control; the reaction state recognition branch, combined with CNN and attention mechanism, can accurately identify the growth stage and abnormal states such as contamination, parameter drift, and equipment failure, improving process safety and stability; the control parameter optimization branch is based on reinforcement learning algorithm, and the model is trained with mixed datasets, combined with Adam optimization algorithm and early stop strategy, balancing training efficiency and generalization ability. The introduction of biological reaction mechanism constraints further improves the rationality and credibility of the model. Moreover, the integrated soft measurement virtual sensing unit can estimate indicators that are difficult to detect by traditional methods, such as cell specific growth rate, cell activity, metabolic flux, substrate conversion rate, and product yield, online; the transfer learning unit supports rapid reuse of the model across batches, reactors, and strains through domain adaptive algorithm, and can complete modeling under small sample conditions, significantly shortening the process development and scale-up cycle; the digital twin simulation unit constructs a 1:1 virtual reactor, synchronizes the physical state in real time and combines it with CFD flow field simulation, which can pre-simulate control strategies, verify the rationality of the scheme, and avoid operational risks, further improving control reliability and batch consistency.

[0071] The feedback correction module also includes a newly added fault knowledge graph and causal inference unit, which automatically locates the root cause of anomalies based on real-time data and historical fault database, including sensor drift, insufficient oxygen supply, abnormal feeding, stirring failure, sealing failure, and risk of bacterial contamination, and outputs treatment suggestions.

[0072] The display unit combines an industrial touchscreen with a host computer for real-time data display. The operation unit switches between manual and automatic modes. The alarm unit uses an audible and visual alarm.

[0073] It also includes an edge-cloud collaboration module and an early warning module. The edge-cloud collaboration module includes an edge terminal and a cloud terminal. The edge terminal is used for real-time control, low-latency inference, local closed-loop regulation, and data caching and preprocessing. The cloud terminal is used for big data training and simultaneous scheduling of multiple reactor clusters. The early warning module collects parameter data and evaluates the data in real time based on a deep learning model to achieve early warning and proactive intervention.

[0074] The inclusion of an edge-cloud collaboration module and an early warning module enhances the system's real-time control capabilities, scalability, and security. The edge module enables local real-time control, low-latency inference, rapid closed-loop adjustment, and data preprocessing and caching, ensuring the response speed and operational stability of the bioreactor control. The cloud module facilitates big data model training, process iteration optimization, and unified scheduling of multiple reactor clusters, effectively improving the system's computing power utilization and scalable production management capabilities. The early warning module collects multi-dimensional operating parameters and uses deep learning models for real-time status assessment, enabling early identification of potential risks and proactive intervention, effectively reducing the failure rate.

[0075] It also includes a causal inference decision module, which is integrated between the deep learning control model module and the feedback correction module. This module is used to construct a causal relationship model between operating parameters and key performance indicators in the biological reaction process. Unlike traditional data-driven methods that rely solely on correlation fitting, the causal inference decision module automatically mines the real causal links between parameters from multi-source heterogeneous data through causal discovery algorithms. It clarifies the strength and direction of the real impact of control parameters on the data, eliminates false associations, pseudo-correlations, and indirect association interference, and forms causal confidence weights.

[0076] In the control decision-making stage, the system makes adaptive parameter decisions based on causal weights, precisely adjusting only key parameters with significant causal effects and keeping parameters without direct causal relationship stable. This avoids erroneous adjustments caused by data fluctuations, noise interference, or indirect correlations, and improves the interpretability, robustness, and long-term stability of the control strategy.

[0077] Example 3:

[0078] A deep learning control method for bioreactors based on multi-source heterogeneous data includes the following steps:

[0079] Step S1: Synchronously collect raw operating data of the bioreactor through the multi-source heterogeneous data acquisition module, adaptively adjust the sampling frequency and store it in the data cache unit;

[0080] Step S2: The data preprocessing and multimodal fusion module performs noise removal, missing value imputation, outlier correction and normalization on the original running data, and obtains a unified feature vector through hierarchical fusion and multimodal Transformer strategy;

[0081] Step S3: Output key parameter prediction and reaction state identification results through the multi-task deep learning model, combine them with the causal inference unit to determine the real impact parameters, and output the optimized control parameters;

[0082] Step S4: Real-time synchronization of the physical reactor state using a 1:1 digital twin, combined with CFD flow field simulation to complete control strategy pre-playing and risk avoidance;

[0083] Step S5: Convert the optimized control parameters into control signals to drive the actuators to adjust the operating state of the bioreactor, and use multiple structures to ensure a sterile environment;

[0084] Step S6: Collect and execute the adjusted operating parameters, calculate the deviation from the model prediction value, fine-tune the control parameters and update the model using incremental learning, and automatically locate the root cause of the anomaly through the fault knowledge graph;

[0085] Step S7: Complete local real-time closed-loop control through the edge terminal, and realize model iteration, transfer learning and multi-reactor cluster scheduling through the cloud;

[0086] Step S8: Real-time display of system operating status, digital twin screen and low-carbon indicators; trigger audible and visual alarms and sterility risk warnings when abnormalities occur.

[0087] Step S9: Repeat steps S1-S8 to form a fully closed-loop intelligent control until the biological reaction is completed.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning control system for a bioreactor based on multi-source heterogeneous data, characterized in that: It includes a multi-source heterogeneous data acquisition module, a data preprocessing and fusion module, a deep learning control model module, an execution adjustment module, a feedback correction module, a monitoring and interaction module, a data storage module, and a security protection module, which are connected in sequence. The multi-source heterogeneous data acquisition module is used to collect physical parameters, chemical parameters, biological parameters, equipment operating parameters, and environmental parameters during the operation of the bioreactor. The data preprocessing and fusion module includes a data preprocessing unit and a multi-source data fusion unit; The data preprocessing unit uses Kalman filtering to eliminate noise and bubble interference, interpolation to fill in missing data, and normalizes the data after identifying and correcting outliers. The multi-source data fusion unit includes a data layer, a feature layer, and a decision layer. The feature layer fusion uses CNN to extract local features and LSTM to extract time series features, combines an attention mechanism to allocate feature weights, and utilizes a multimodal Transformer unit to achieve unified representation of time series data, image data, and spectral data. The decision layer fusion combines biological reaction mechanisms and key features, and standardizes and aligns data across reactors. The deep learning control module includes a model training unit, a model inference unit, an integrated soft measurement unit, a transfer learning unit, a green and low-carbon optimization unit, and a digital twin simulation unit. The execution adjustment module includes a controller, an actuator, and a drive unit. The controller is used to realize material stirring, temperature control, and pH adjustment. The drive unit is used to drive the actuator to work, and the drive unit adopts closed-loop feedback control. The feedback correction module includes a deviation calculation unit, an adaptive correction unit, and a model update unit. The deviation calculation unit is used to calculate the deviation between the real-time running parameters and the model prediction parameters. The adaptive correction unit uses the PI algorithm to fine-tune the control parameters and combines the adaptive weight adjustment strategy to adapt to the time-varying reaction. The model update unit uses incremental learning to update the model parameters. When the deviation exceeds the preset threshold, an abnormal warning is triggered. The monitoring and interaction module includes a display unit, an operation unit, and an alarm unit; The data storage module uses a distributed database, and the security protection module includes a hardware electromagnetic shielding module, a software data encryption module, and a hierarchical access control module.

2. The deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 1, characterized in that: The multi-source heterogeneous data acquisition module includes multiple sets of sensors, a data acquisition card, a communication module, and a data caching unit. The sensors include physical sensors, chemical sensors, biological sensors, and equipment status sensors. The multiple sets of sensors realize multimodal data acquisition. The communication module supports multiple industrial communication protocols. The data caching unit adopts a distributed caching architecture for temporary storage and fast retrieval of raw data.

3. The deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 2, characterized in that: The model training unit adopts a multi-task deep learning model, which includes three parallel task branches: a key parameter prediction branch, a reaction state recognition branch, and a control parameter optimization branch. The key parameter prediction branch is based on an LSTM network and predicts cell density, product concentration, dissolved oxygen, and pH parameters within the next 1-60 minutes. The reaction state recognition branch is based on CNN and attention mechanism and is used to identify the growth period, stationary period, death period, as well as the abnormal states of contamination, parameter drift, and equipment failure. The control parameter optimization branch extends the green and low-carbon objective function based on a reinforcement learning algorithm; The model training unit uses a hybrid dataset to combine the Adam algorithm with an early stopping strategy to optimize parameters for model training, and introduces biological response mechanism constraints during model training. The integrated soft measurement unit constructs a deep learning virtual sensing model to achieve real-time online estimation of difficult-to-measure indicators such as cell specific growth rate, cell activity, metabolic flux, substrate conversion rate, and product yield. The transfer learning unit adopts a domain adaptive algorithm, and the digital twin simulation unit establishes a 1:1 digital twin of the bioreactor, synchronizing the physical entity's operating status in real time. Combined with CFD flow field simulation, it enables control strategy pre-playing, simulation verification, and early risk avoidance.

4. The deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 3, characterized in that: The controller adopts a PLC and embedded chip combined architecture.

5. A deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 4, characterized in that: The feedback correction module also includes a new fault knowledge graph and causal inference unit. The feedback correction module automatically locates the root cause of the anomaly based on real-time data and historical fault database.

6. A deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 5, characterized in that: The display unit combines an industrial touch screen with a host computer for real-time data display. The operation unit achieves switching operations through a combination of manual and automatic switching modes. The alarm unit uses an audible and visual alarm.

7. A deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 6, characterized in that: It also includes an edge-cloud collaboration module and an early warning module. The edge-cloud collaboration module includes an edge terminal and a cloud terminal. The edge terminal is used for real-time control, low-latency inference, local closed-loop adjustment, and data caching and preprocessing. The cloud terminal is used for big data training and simultaneous scheduling of multiple reactor clusters. The early warning module collects parameter data and evaluates the data in real time based on a deep learning model to achieve early warning and proactive intervention.

8. A deep learning control system for a bioreactor based on multi-source heterogeneous data according to claim 7, characterized in that: It also includes a causal inference decision module, which is integrated between the deep learning control model module and the feedback correction module. It is used to construct a causal relationship model between operating parameters and key performance indicators in the biological reaction process. The causal inference decision module automatically mines the real causal links between parameters from multi-source heterogeneous data through causal discovery algorithms, clarifies the strength and direction of the real influence of control parameters on the data, eliminates false associations, pseudo-correlation and indirect association interference, and forms causal confidence weights.

9. A deep learning control method for a bioreactor based on multi-source heterogeneous data, as described in any one of claims 1-8, characterized in that: The control method includes the following steps: Step S1: Synchronously collect raw operating data of the bioreactor through the multi-source heterogeneous data acquisition module, adaptively adjust the sampling frequency and store it in the data cache unit; Step S2: The data preprocessing and multimodal fusion module performs noise removal, missing value imputation, outlier correction and normalization on the original running data, and obtains a unified feature vector through hierarchical fusion and multimodal Transformer strategy; Step S3: Output key parameter prediction and reaction state identification results through the multi-task deep learning model, combine them with the causal inference unit to determine the real impact parameters, and output the optimized control parameters; Step S4: Real-time synchronization of the physical reactor state using a 1:1 digital twin, combined with CFD flow field simulation to complete control strategy pre-playing and risk avoidance; Step S5: Convert the optimized control parameters into control signals to drive the actuators to adjust the operating state of the bioreactor, and use multiple structures to ensure a sterile environment; Step S6: Collect and execute the adjusted operating parameters, calculate the deviation from the model prediction value, fine-tune the control parameters and update the model using incremental learning, and automatically locate the root cause of the anomaly through the fault knowledge graph; Step S7: Complete local real-time closed-loop control through the edge terminal, and realize model iteration, transfer learning and multi-reactor cluster scheduling through the cloud; Step S8: Display the system's operating status, digital twin screen, and low-carbon indicators in real time, and trigger audible and visual alarms and aseptic risk warnings when abnormalities occur; Step S9: Repeat steps S1-S8 to form a fully closed-loop intelligent control until the biological reaction is completed.

Images