A fermentation environment control method based on multi-modal data fusion

Abstract

The application discloses a kind of based on multi-modal data fusion fermentation environment control method, comprising: collection multi-modal fermentation data and executes pre-processing, forms multivariate time series dataset;Improved MTAD-GAT model is input, calculates prediction error reconstruction error and consistency score, generates dynamic reliable weight;Coupling relationship matrix is constructed across mode, coupling residual is calculated and iterative correction data, reconstructs state vector and uncertainty;Coupling stability index is calculated, and state anchoring is executed, obtains stable reference state parameter;Based on state vector and reference state, construct metabolic feasible region, form control constraint interval;Self-scaling robust controller is constructed, dynamic scaling constraint and rolling optimization calculation, output fermentation control instruction.The application realizes stable identification to fermentation process state and adaptive control to fermentation environment parameter by multi-modal data fusion, coupling relationship analysis and self-scaling robust predictive control.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for fermentation processes, and in particular to a fermentation environment control method based on multimodal data fusion. Background Technology

[0002] Fermentation processes are widely used in food, pharmaceuticals, biomanufacturing, and fine chemicals. Their production is typically influenced by a combination of factors, including temperature, pH, dissolved oxygen concentration, aeration rate, stirring speed, feed rate, and gas metabolism. To ensure the target microbial strain or system operates in a suitable metabolic environment, existing fermentation equipment is usually equipped with temperature sensors, pH sensors, dissolved oxygen sensors, exhaust gas detection devices, and feed, stirring, and aeration actuators. These parameters are adjusted via preset programs or controllers. With advancements in sensor and data acquisition technologies, fermentation processes can now acquire multi-source heterogeneous data, including images, gas emissions, and stirring status, providing a foundation for multimodal information-based fermentation state identification and process control.

[0003] In existing technologies, most methods for controlling the fermentation environment still rely on single sensor feedback or combined regulation of a small number of process variables. These typically employ fixed threshold judgments, empirical rule control, or conventional closed-loop control to regulate parameters such as aeration rate, stirring speed, feeding rate, and temperature. While some technologies introduce multivariate monitoring or anomaly detection methods, most only involve simple data splicing, weighted processing, or independent analysis, lacking a unified characterization of the coupling relationships between different modes. This makes it difficult to accurately identify the true fermentation state when sensor drift, signal noise enhancement, fermentation stage switching, or environmental disturbances occur. Furthermore, the evaluation of multimodal data quality in existing technologies usually relies on static rules or fixed weights, making it difficult to reflect the reliability of each modality's data in real time according to process changes, thus easily leading to state estimation bias.

[0004] Existing fermentation control methods generally lack the ability to address the interplay between state uncertainties and metabolic constraints in the control decision-making process. The adjustment ranges of control variables are often fixed or simply dynamically adjusted, failing to reflect the metabolic conflicts between oxygen supply, feeding, acid-base regulation, and heat changes during fermentation. When inconsistencies exist in multimodal information, existing control methods are prone to issues such as control lag, mismatched control amplitudes, and even local instability, thereby affecting fermentation efficiency and process stability.

[0005] Therefore, how to provide a fermentation environment control method based on multimodal data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a fermentation environment control method based on multimodal data fusion. This invention achieves dynamic reliability assessment of multimodal sensor data by introducing an improved MTAD-GAT model, and realizes stable identification of fermentation state by constructing a cross-modal coupling relationship matrix and coupling stability index. At the same time, it combines metabolic feasible domain construction and self-scaling robust predictive controller to realize adaptive adjustment of fermentation environment, thereby completing the intelligent control process of fermentation process state identification, control constraint generation and control command output. It has the advantages of high multimodal data utilization, strong state identification stability, high control response robustness and high fermentation environment adjustment accuracy.

[0007] A fermentation environment control method based on multimodal data fusion according to an embodiment of the present invention includes: Multiple sensors and information acquisition devices are arranged in the fermentation device to collect multimodal data during the fermentation process. The multimodal data is preprocessed to form a multivariate time series dataset. The multivariate time series dataset is input into the improved MTAD-GAT model to calculate the prediction error, reconstruction error and cross-modal consistency score of each variable, and to determine the dynamic reliability weight of each modality data. The cross-modal coupling relationship matrix of the fermentation process is constructed based on the dynamic reliability weight, and the coupling residuals between each mode are calculated. The multimodal data is iteratively corrected by the coupling residual constraints, the state vector of the fermentation process is reconstructed, and the corresponding state uncertainty parameters are determined. The cross-modal coupling stability index is calculated based on the cross-modal coupling relationship matrix and the trend of coupling residual change. The coupling stability interval in the fermentation process is identified based on the coupling stability index, and the state vector is anchored to obtain stable reference state parameters. The metabolic feasible region of the fermentation process is constructed based on the state vector, state uncertainty parameters, and stable reference state parameters. The boundary of the metabolic feasible region is then updated through self-evolution based on the historical trajectory of the fermentation state over time, forming a control variable constraint interval. A self-scaling robust predictive controller is constructed. The state vector, state uncertainty parameters, and control variable constraint intervals are input into the self-scaling robust predictive controller. Control commands are generated by dynamically scaling the control variable constraint intervals and performing rolling optimization calculations. The control commands are then executed by the actuators of the fermentation unit to adjust the fermentation environment parameters.

[0008] Optionally, the multimodal data includes temperature data, pH data, dissolved oxygen concentration data, gas emission data, stirring speed data, aeration rate data, feeding rate data, and fermentation broth image feature data.

[0009] Optionally, the preprocessing of multimodal data includes outlier detection, signal filtering, missing data imputation, feature standardization, and time synchronization of the acquired multimodal data.

[0010] Optionally, determining the dynamic reliability weights of each modal data includes: The input of the multivariate time series dataset is organized by arranging the continuous observations of each modal variable within a preset time window in chronological order to form a set of time series data segments for modeling. An improved MTAD-GAT model is constructed based on a set of time-series data segments. The improved MTAD-GAT model includes a time-series coding layer, a coupled enhanced graph attention layer, and a dual-branch decoding layer. In the temporal coding layer, the temporal data segment corresponding to each modal variable is input into a shared temporal coding unit, and feature transformation and nonlinear mapping are performed on the observation values ​​at each time point to obtain a temporal coding feature sequence that corresponds one-to-one with each modal variable. In the coupled enhanced graph attention layer, static modal correlation is constructed based on the long-term statistical correlation of multivariate time series data, and dynamic modal correlation is obtained by combining the short-term collaborative change characteristics of each modal variable within the current time window. When calculating the graph attention weight, both static and dynamic modal correlation are introduced. The dynamic reliability weights of each modality obtained in the previous time window are used as gating factors to modulate the attention weights in order to generate the fused latent feature representation. In the dual-branch decoding layer, the latent feature representations are input into the prediction branch and the reconstruction branch respectively. The prediction branch outputs the predicted values ​​of each modal variable, and the reconstruction branch outputs the reconstructed values ​​of each modal variable within the current time window. Based on the deviation between the predicted values ​​and the actual observed values ​​of each modal variable, as well as the deviation between the reconstructed values ​​and the actual observed values, the prediction error, reconstruction error, and cross-modal consistency score of each modal variable are obtained. The dynamic reliability weight of each modal data is determined by the prediction error and the reconstruction error according to the preset mapping rules.

[0011] Optionally, the reconstructing of the fermentation process state vector and the determination of the corresponding state uncertainty parameters include: Based on the dynamic reliability weights of each modality data, the time series of each modality in the multivariate time series data are weighted to obtain the weighted time series data corresponding to each modality. Within a preset time window, based on the coordinated changes of weighted time series data, the statistical correlation characteristics between modes in the long-term historical data of the fermentation process, and the predefined causal constraints in the fermentation process, a basic coupling sub-matrix, a stage correction sub-matrix, and a reliability correction sub-matrix are constructed respectively. The long-term coupling relationship reflected by the basic coupling sub-matrix, the current fermentation stage characteristic coupling relationship reflected by the stage correction sub-matrix, and the dynamic reliability correction effect reflected by the reliability correction sub-matrix are superimposed to obtain the cross-modal coupling relationship matrix of the fermentation process. Based on the cross-modal coupling relationship matrix, the weighted time series data of each mode are coupled and estimated to obtain the coupling estimate value of each mode based on other modes within the current time window. The actual observation data of each mode is compared with the corresponding coupling estimate value to calculate the coupling residual sequence between modes. Iterative correction processing is performed on multimodal data under coupled residual constraints. In each iteration, the weighted time series data is updated using coupled residuals according to the preset convergence criteria, and data change amplitude constraints and smoothness constraints are applied at the same time. The iteration stops when the coupled residuals meet the preset threshold conditions, and the corrected multimodal time series data is obtained. The fermentation process state vectors corresponding to each time point are extracted from the corrected multimodal time series data. The coupled residual distribution characteristics at each time point are statistically analyzed to obtain the state uncertainty parameters corresponding to the state vectors.

[0012] Optionally, obtaining stable reference state parameters includes: Obtain the cross-modal coupling relationship matrix and the coupling residual sequence of each mode. Under a preset sampling period, use the coupling relationship strength and dynamic reliability weight in the cross-modal coupling relationship matrix as weights to perform weighted summation of the coupling residuals of each mode to obtain the coupling stability evaluation value sequence. Within a sliding time window, the window mean, window standard deviation, and number of times the coupling stability evaluation value sequence is calculated. The window mean represents the average level of coupling deviation, the window standard deviation represents the degree of fluctuation of coupling deviation, and the number of times the threshold is exceeded within the window represents the duration of coupling deviation exceeding the preset residual threshold. The cross-modal coupling stability index is calculated based on the window mean, window standard deviation, and the number of times the threshold is exceeded within the window. The cross-modal coupling stability index is obtained by weighting and summing the window mean, window standard deviation, and the number of times the threshold is exceeded within the window according to a preset weight. The preset weight is a constant parameter that satisfies the weight sum to 1. The calculated cross-modal coupling stability index is then formed into a cross-modal coupling stability index sequence. The cross-modal coupling stability index sequence is compared with the stability judgment threshold interval. The time period in which the cross-modal coupling stability index falls into the stability judgment threshold interval and continuously meets the preset duration is determined as the coupling stability interval. The fermentation process state vector at the corresponding time is extracted in the coupling stability interval to form a candidate anchoring state vector set. Clustering is performed on the candidate anchoring state vector set to obtain cluster center vectors, and time smoothing is performed on the cluster center vectors. The time-smoothed cluster center vectors are then determined as the stable reference state parameters of the coupled stable interval.

[0013] Optionally, forming the control variable constraint interval includes: Obtain the state vector, state uncertainty parameters, and stable reference state parameters. Compare the state vector with the stable reference state parameters dimension by dimension. Extract the deviation direction, deviation magnitude, and duration of each state component relative to the stable reference state parameters to form fermentation state deviation description information. Based on the description information of the deviation of fermentation state, the control relationship between each state component is identified as a metabolic conflict relationship, and a set of metabolic conflict constraint units is constructed based on the metabolic conflict relationship. Based on the set of metabolic conflict constraint units, corresponding local feasible boundary fragments are generated for aeration rate, stirring speed, feeding rate, temperature regulation power and pH regulator addition amount, respectively. The local feasible boundary fragments are then screened, spliced ​​and closed according to the compatibility relationship between each metabolic conflict constraint unit to form the metabolic feasible domain of the fermentation process corresponding to the current fermentation state. Boundary memory trajectories are established for each boundary segment of the metabolic feasible domain in the fermentation process. The boundary memory trajectories record the position changes, length changes and boundary switching order of each boundary segment within multiple consecutive time windows. When a change in state uncertainty parameter, a shift in state vector deviation direction or a reorganization of metabolic conflict relationship is detected, boundary migration update is triggered, so that each boundary segment in the current time window completes position migration, length contraction or length expansion based on the boundary memory trajectory of the previous time window, and obtains the self-evolutionary updated boundary of the metabolic feasible domain. Based on the metabolic feasible domain boundary updated through self-evolution, the constraint intervals of the control variables corresponding to ventilation rate, stirring speed, feeding rate, temperature regulation power, and pH regulator addition amount are determined.

[0014] Optionally, the step of generating control commands by dynamically scaling the constraint interval of control variables and performing rolling optimization calculations, and having the fermentation device actuator execute the control commands to adjust the fermentation environment parameters, includes: Based on the state vector, state uncertainty parameters, and control variable constraint intervals, a self-scaling robust predictive controller is constructed. The self-scaling robust predictive controller includes a constraint self-scaling unit, a robust rolling optimization unit, and an execution sequence filtering unit. In the constrained self-scaling unit, the state uncertainty parameter, the deviation information between the state vector and the stable reference state parameter, and the boundary state of the metabolic feasible domain are read. According to the preset scaling rules, the boundary shrinkage, boundary preservation or boundary expansion processing is performed on each control variable constraint interval, so that the ventilation rate, stirring speed, feeding rate, temperature regulation power and pH regulator addition amount correspond to the updated control variable constraint interval respectively. In the robust rolling optimization unit, the updated control variable constraint interval is used as the control boundary. The joint constraint objectives are the convergence of the state vector to the stable reference state parameter, the continuous change of control variables, and the suppression of state uncertainty. Multiple candidate sequences of control variables are generated according to the preset prediction time domain. The state evolution results of each candidate sequence of control variables are evaluated cycle by cycle in the future continuous control cycle. Candidate sequences of control variables that exceed the updated control variable constraint interval or cause an increase in state uncertainty are eliminated. The candidate sequences of control variables that meet the constraint conditions are retained as executable control sequences. In the execution sequence screening unit, the executable control sequences are sorted, and the candidate control variable sequence with the largest reduction in state deviation, the smallest increase in state uncertainty, and the fewest number of control variable switching times is selected first. The first control variable of the best-ranked candidate control variable sequence is determined as the control instruction for the current control cycle. Control commands are sent to the actuators of the fermentation unit to drive the aeration actuator, stirring actuator, feeding actuator, temperature adjustment actuator, and pH adjustment actuator to perform corresponding adjustment actions. In the next control cycle, the updated state vector, state uncertainty parameters, and control variable constraint intervals are reread to repeatedly execute constraint self-scaling, robust rolling optimization, and execution sequence filtering.

[0015] The beneficial effects of this invention are: Compared with existing fermentation environment control technologies, this invention first achieves higher accuracy in state perception at the multimodal data processing level. By jointly acquiring multimodal data such as temperature, pH value, dissolved oxygen concentration, gas emission, stirring speed, aeration rate, feeding rate, and fermentation broth image features, and using an improved MTAD-GAT model to model the correlation between various modal data, this invention can identify abnormal changes and reliability of different modal data in real time during fermentation. This avoids interference caused by single sensor drift, local signal distortion, or modal conflicts in state judgment, and improves the utilization efficiency and effectiveness of multi-source monitoring data in fermentation control.

[0016] This invention achieves stable reconstruction and state anchoring of the actual operating state of the fermentation process by constructing a cross-modal coupling relationship matrix, calculating coupling residuals, and identifying coupling stability intervals. Compared to existing technologies that simply weight or independently analyze multimodal data, this invention maintains continuous identification of the fermentation process state even under conditions of fermentation stage switching, increased environmental disturbances, or inconsistent multimodal signals. It outputs stable reference state parameters as a control benchmark, thus effectively reducing control deviations caused by state identification fluctuations and improving the stability and consistency of fermentation process state judgment.

[0017] This invention constructs a metabolically feasible domain and combines it with a self-scaling robust predictive controller to perform synergistic optimization control of aeration rate, stirring speed, feeding rate, temperature regulation power, and pH regulator dosage. This allows the adjustment process of control variables to simultaneously consider state uncertainty, metabolic conflict relationships, and dynamic changes in control boundaries. When state deviations or increased uncertainty occur during fermentation, this invention automatically adjusts the constraint range of control variables and generates control commands adapted to the current fermentation state, reducing control lag, overshoot, and local instability, thereby improving the robustness, control accuracy, and overall process stability of fermentation environment regulation. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a fermentation environment control method based on multimodal data fusion proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved MTAD-GAT model, which is a fermentation environment control method based on multimodal data fusion proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A fermentation environment control method based on multimodal data fusion, comprising: Multiple sensors and information acquisition devices are arranged in the fermentation device to collect multimodal data during the fermentation process. The multimodal data is preprocessed to form a multivariate time series dataset. The multivariate time series dataset is input into the improved MTAD-GAT model to calculate the prediction error, reconstruction error and cross-modal consistency score of each variable, and to determine the dynamic reliability weight of each modality data. The cross-modal coupling relationship matrix of the fermentation process is constructed based on the dynamic reliability weight, and the coupling residuals between each mode are calculated. The multimodal data is iteratively corrected by the coupling residual constraints, the state vector of the fermentation process is reconstructed, and the corresponding state uncertainty parameters are determined. The cross-modal coupling stability index is calculated based on the cross-modal coupling relationship matrix and the trend of coupling residual change. The coupling stability interval in the fermentation process is identified based on the coupling stability index, and the state vector is anchored to obtain stable reference state parameters. The metabolic feasible region of the fermentation process is constructed based on the state vector, state uncertainty parameters, and stable reference state parameters. The boundary of the metabolic feasible region is then updated through self-evolution based on the historical trajectory of the fermentation state over time, forming a control variable constraint interval. A self-scaling robust predictive controller is constructed. The state vector, state uncertainty parameters, and control variable constraint intervals are input into the self-scaling robust predictive controller. Control commands are generated by dynamically scaling the control variable constraint intervals and performing rolling optimization calculations. The control commands are then executed by the actuators of the fermentation unit to adjust the fermentation environment parameters.

[0021] In this embodiment, the multimodal data includes temperature data, pH value data, dissolved oxygen concentration data, gas emission data, stirring speed data, aeration rate data, feeding rate data, and fermentation broth image feature data.

[0022] In this embodiment, the preprocessing of multimodal data includes outlier detection, signal filtering, missing data imputation, feature standardization, and time synchronization of the acquired multimodal data.

[0023] In this embodiment, determining the dynamic reliability weights of each modal data includes: The multivariate time series dataset is organized by arranging the continuous observations of each modal variable within a preset time window in chronological order to form a set of time series data segments for modeling, where the preset time window is 20 min. An improved MTAD-GAT model is constructed based on a set of time-series data segments. The improved MTAD-GAT model includes a time-series coding layer, a coupled enhanced graph attention layer, and a dual-branch decoding layer. In the temporal coding layer, the temporal data segment corresponding to each modal variable is input into a shared temporal coding unit. Feature transformation and nonlinear mapping are performed on the observations at each time step to obtain a temporal coding feature sequence that corresponds one-to-one with each modal variable. Specifically, the feature transformation and nonlinear mapping for the observations at each time step are as follows: Linear projection processing is performed on the observations of each modal variable within the current time window to map the original observations to a unified feature dimension space; The feature sequence after linear projection is input into a nonlinear activation function for nonlinear mapping to enhance the feature representation capability. The feature sequence after nonlinear mapping is encoded by time position, so that the observations at each time point retain the time order information in the feature space; The feature sequence after linear projection, nonlinear mapping and temporal position encoding is output as a temporal encoded feature sequence. In the coupled and enhanced graph attention layer, static modal correlations are constructed based on the long-term statistical correlations of multivariate time-series data, and dynamic modal correlations are obtained by updating the short-term collaborative change features of each modality variable within the current time window. When calculating the graph attention weights, both static and dynamic modal correlations are introduced. The dynamic reliability weights of each modality obtained from the previous time window are used as gating factors to modulate the attention weights, thereby generating a fused latent feature representation, where: Static modal association relationships are constructed based on the long-term statistical correlation of multivariate time series data, specifically as follows: Statistical analysis was performed on historical multivariate time series data to calculate the overall trend and correlation of each modal variable over multiple fermentation cycles. A modal correlation matrix was established based on the long-term stable correlation between different modal variables, and the modal correlation matrix was used as a static modal correlation. The dynamic modal correlation is obtained by combining the short-term collaborative change characteristics of each modal variable within the current time window, specifically as follows: Within the current time window, the observed sequences of each modal variable are analyzed for trend changes. The direction and magnitude of change of each modal variable within the time window are calculated. The modal correlation strength is updated based on the synchronous changes between different modal variables, forming a dynamic modal correlation relationship that reflects the current fermentation state change characteristics. In the dual-branch decoding layer, latent feature representations are input into the prediction branch and the reconstruction branch, respectively. The prediction branch outputs the predicted values ​​of each modal variable, and the reconstruction branch outputs the reconstructed values ​​of each modal variable within the current time window. Based on the deviations between the predicted and actual observations of each modal variable, as well as the deviations between the reconstructed and actual observations, the prediction error, reconstruction error, and cross-modal consistency score of each modal variable are obtained. According to a preset mapping rule, the dynamic reliability weights of each modal data are determined from the prediction error and the reconstruction error, where: The predicted values ​​are obtained as follows: The latent feature representation is input into the temporal prediction unit in the prediction branch. By performing feature propagation and linear mapping on the latent feature representation in the time dimension, the prediction results of each modal variable at the next time step are generated, and the predicted values ​​of each modal variable at the prediction time step are output according to the correspondence of modal variables. The reconstructed values ​​are obtained as follows: The latent feature representation is input into the feature reconstruction unit in the reconstruction branch. The latent feature representation is then subjected to inverse feature mapping to remap the features in the latent feature space back to the original observation space, thus obtaining the reconstruction values ​​of each modal variable within the current time window. The dynamic reliability weights of each modality data are determined according to the preset mapping rules based on the prediction error and reconstruction error, specifically as follows: The prediction error and reconstruction error of each modal variable are calculated separately, and the prediction error and reconstruction error are normalized so that the errors of different modal variables are within a uniform numerical range. The normalized prediction error and reconstruction error are weighted and fused according to preset weight coefficients to obtain the comprehensive error evaluation value of each modal variable; The dynamic reliability weights of each modal data are determined by performing reverse mapping processing based on the comprehensive error evaluation value and the inverse proportional relationship between the comprehensive error evaluation value and the dynamic reliability weight. The closer the comprehensive error evaluation value is to zero, the closer the corresponding dynamic reliability weight is to the upper limit of the weight; the closer the comprehensive error evaluation value is to the preset error threshold, the closer the corresponding dynamic reliability weight is to the lower limit of the weight.

[0024] In this embodiment, reconstructing the fermentation process state vector and determining the corresponding state uncertainty parameters includes: Based on the dynamic reliability weights of each modality data, the time series of each modality in the multivariate time series data are weighted to obtain the weighted time series data corresponding to each modality. Within a preset time window, based on the coordinated changes in weighted time-series data, the statistical correlation characteristics between modes in long-term historical data of the fermentation process, and the predefined causal constraints in the fermentation process, a basic coupling sub-matrix, a stage correction sub-matrix, and a reliability correction sub-matrix are constructed respectively. The long-term coupling relationships reflected by the basic coupling sub-matrix, the current fermentation stage characteristic coupling relationships reflected by the stage correction sub-matrix, and the dynamic reliability correction effect reflected by the reliability correction sub-matrix are superimposed to obtain the cross-modal coupling relationship matrix of the fermentation process. The predefined causal constraints are as follows: Based on the pre-established set of causal relationships between process variables according to the fermentation process mechanism, including the oxygen supply relationship between aeration rate and dissolved oxygen concentration, the mass transfer relationship between stirring speed and dissolved oxygen mass transfer efficiency, the nutrient supply relationship between feeding rate and substrate concentration and metabolic rate, the regulatory relationship between temperature and microbial metabolic rate, and the regulatory relationship between pH regulator addition amount and fermentation broth pH, ​​the causal relationship is used as the structural constraint for constructing a cross-modal coupling relationship matrix. Based on the cross-modal coupling relationship matrix, coupling estimation is performed on the weighted time series data of each mode to obtain the coupling estimation value of each mode based on other modes within the current time window. The actual observed data of each mode is compared with the corresponding coupling estimation value to calculate the coupling residual sequence between modes. Specifically, the coupling estimation for the weighted time series data of each mode is as follows: Using the coupling coefficient of the corresponding mode in the cross-modal coupling relationship matrix as the weight, the weighted time series data of other modes besides the current mode in the current time window are weighted and combined to obtain the coupling estimation value of the mode based on the information of other modes at the current time. The operation process is repeated for each time point in the entire time window to obtain the coupling estimation sequence of the mode in the time window. Iterative correction processing is performed on multimodal data under coupled residual constraints. In each iteration, the weighted time series data is updated using coupled residuals according to a preset convergence criterion, while data variation amplitude constraints and smoothness constraints are applied simultaneously. Iteration stops when the coupled residuals meet a preset threshold condition, resulting in corrected multimodal time series data. The preset convergence criterion is as follows: During two consecutive iterations, the coupled residual sequences corresponding to the same modal variable are compared. When the overall change between the coupled residual sequence of the current iteration and the coupled residual sequence of the previous iteration is lower than the preset residual change threshold, the iteration process is determined to have reached convergence. When the coupled residual sequences corresponding to all modal variables meet the residual change threshold condition, the iteration update process is stopped. Data change magnitude constraint refers to limiting the change between the updated value of each modal variable in the current time window and the corresponding value in the previous iteration when updating weighted time series data in each iteration. When the change exceeds the preset change threshold, the updated value is adjusted to the boundary range corresponding to the preset change threshold in order to avoid sudden changes in data during the iteration process. Smoothing constraint refers to constraining the trend of change between the updated values ​​of the same modal variable at adjacent time points when iterating and updating weighted time series data, so that the updated time series data maintains a continuous change relationship. When the change between adjacent time points exceeds the preset smoothing threshold, the data at adjacent time points are weighted and averaged to make the updated time series data meet the continuous change condition. The fermentation process state vectors corresponding to each time point are extracted from the corrected multimodal time series data. The coupled residual distribution characteristics at each time point are statistically analyzed to obtain the state uncertainty parameters corresponding to the state vectors.

[0025] In this embodiment, obtaining stable reference state parameters includes: Obtain the cross-modal coupling relationship matrix and the coupling residual sequence of each mode. Under a preset sampling period, use the coupling relationship strength and dynamic reliability weight in the cross-modal coupling relationship matrix as weights to perform weighted summation of the coupling residuals of each mode to obtain the coupling stability evaluation value sequence. Within a sliding time window, the window mean, window standard deviation, and number of times the coupling stability evaluation value sequence exceeds a preset residual threshold are calculated. The window mean represents the average level of coupling deviation, the window standard deviation represents the degree of fluctuation in coupling deviation, and the number of times the threshold exceeds a preset residual threshold represents the duration of coupling deviation exceeding the preset residual threshold. The window mean is calculated as follows: Within the sliding time window, the evaluation values ​​corresponding to each time point in the coupling stability evaluation value sequence are summed, and the summation result is divided by the number of time points included in the sliding time window to obtain the average value of the coupling stability evaluation value within the sliding time window. The calculation of the window standard deviation is as follows: Within the sliding time window, the difference between the evaluation value at each time point in the coupling stability evaluation value sequence and the window mean is first calculated. Then, the difference is squared and averaged. Finally, the average value is squared to obtain the window standard deviation, which reflects the degree of fluctuation of the coupling stability evaluation value within the sliding time window. The calculation of the number of times the threshold is exceeded within the window is as follows: Within the sliding time window, the coupling stability evaluation value is compared with the preset residual threshold at each time step. When the coupling stability evaluation value is greater than the preset residual threshold, the time step is recorded as an over-threshold event. All over-threshold events within the sliding time window are accumulated and statistically analyzed to obtain the number of over-threshold events within the window. The cross-modal coupling stability index is calculated based on the window mean, window standard deviation, and the number of times the threshold is exceeded within the window. The cross-modal coupling stability index is obtained by weighting and summing the window mean, window standard deviation, and the number of times the threshold is exceeded within the window according to a preset weight. The preset weight is a constant parameter that satisfies the weight sum to 1. The calculated cross-modal coupling stability index is then formed into a cross-modal coupling stability index sequence. The cross-modal coupling stability index sequence is compared with the stability judgment threshold interval. The time period in which the cross-modal coupling stability index falls into the stability judgment threshold interval and continuously meets the preset duration is determined as the coupling stability interval. The fermentation process state vector at the corresponding time is extracted in the coupling stability interval to form a candidate anchoring state vector set. Clustering is performed on the candidate anchoring state vector set to obtain cluster center vectors, and time smoothing is performed on the cluster center vectors. The time-smoothed cluster center vectors are then determined as the stable reference state parameters of the coupled stable interval.

[0026] In this embodiment, forming the control variable constraint interval includes: Obtain the state vector, state uncertainty parameters, and stable reference state parameters. Compare the state vector with the stable reference state parameters dimension by dimension. Extract the deviation direction, deviation magnitude, and duration of each state component relative to the stable reference state parameters to form fermentation state deviation description information. Based on the description information of the deviation of fermentation state, the control relationship between the mutual constraints of each state component is identified as metabolic conflict relationship. Among them, the metabolic conflict relationship includes the conflict relationship between oxygen supply enhancement and foam growth, the conflict relationship between feed enhancement and acidification accumulation, the conflict relationship between temperature regulation and metabolic rate fluctuation, and the conflict relationship between agitation enhancement and local environmental disturbance. A set of metabolic conflict constraint units is constructed based on the metabolic conflict relationship. Based on the set of metabolic conflict constraint units, corresponding locally feasible boundary fragments are generated for aeration rate, stirring speed, feed rate, temperature regulation power, and pH regulator dosage. These locally feasible boundary fragments are then screened, spliced, and closed according to the compatibility relationships between the metabolic conflict constraint units to form the metabolically feasible domain of the fermentation process corresponding to the current fermentation state. Specifically, the generation of locally feasible boundary fragments involves: Based on the control variables and corresponding fermentation state components involved in each metabolic conflict constraint unit, the allowable adjustment direction of the control variables in the current fermentation stage is first determined. Based on the stable reference state parameters, and considering the deviation between the current state vector and the stable reference state parameters, an adjustable range is determined within the constraint interval of the corresponding control variables. Based on the restrictive relationship between control variables by the metabolic conflict constraint unit, the adjustable range is truncated so that the range does not enter the conflict region defined by the metabolic conflict constraint unit. The adjustable range of each control variable after truncation is represented as a local feasible boundary segment. Boundary memory trajectories are established for each boundary segment of the metabolic feasible domain in the fermentation process. The boundary memory trajectories record the position changes, length changes and boundary switching order of each boundary segment within multiple consecutive time windows. When a change in state uncertainty parameter, a shift in state vector deviation direction or a reorganization of metabolic conflict relationship is detected, boundary migration update is triggered, so that each boundary segment in the current time window completes position migration, length contraction or length expansion based on the boundary memory trajectory of the previous time window, and obtains the self-evolutionary updated boundary of the metabolic feasible domain. Based on the metabolic feasible domain boundary updated through self-evolution, the constraint intervals of the control variables corresponding to ventilation rate, stirring speed, feeding rate, temperature regulation power, and pH regulator addition amount are determined.

[0027] In this embodiment, the step of generating control commands by dynamically scaling the constraint interval of control variables and performing rolling optimization calculations, and then having the fermentation device actuator execute the control commands to adjust the fermentation environment parameters, includes: Based on the state vector, state uncertainty parameters, and control variable constraint intervals, a self-scaling robust predictive controller is constructed. The self-scaling robust predictive controller includes a constraint self-scaling unit, a robust rolling optimization unit, and an execution sequence filtering unit. In the constrained self-scaling unit, the state uncertainty parameters, the deviation information between the state vector and the stable reference state parameters, and the boundary state of the metabolic feasible region are read. According to a preset scaling rule, boundary contraction, boundary preservation, or boundary expansion processing is performed on each control variable constraint interval, so that the aeration rate, stirring speed, feeding rate, temperature regulation power, and pH adjuster addition amount correspond to the updated control variable constraint intervals, respectively. The preset scaling rule is as follows: The adjustment method of the control variable constraint interval is determined based on the deviation direction and magnitude between the state vector and the stable reference state parameter, as well as the state uncertainty parameter. When the deviation direction is consistent with the control variable adjustment direction and the state uncertainty parameter meets the stability condition, boundary expansion is performed. When the state uncertainty parameter is within the stable interval, the boundary remains unchanged. When the state uncertainty parameter exceeds the stability threshold or the deviation direction is opposite to the control variable adjustment direction, boundary contraction is performed. The boundary adjustment magnitude is determined by the remaining distance between the control variable constraint interval and the boundary of the metabolic feasible region. In the robust rolling optimization unit, the updated control variable constraint interval is used as the control boundary. The joint constraint objectives are convergence of the state vector to the stable reference state parameter, continuous change of control variables, and suppression of state uncertainty. Multiple candidate sequences of control variables are generated according to a preset prediction time domain. The state evolution results corresponding to each candidate sequence are evaluated cycle by cycle in future continuous control periods. Candidate sequences that exceed the updated control variable constraint interval or cause an increase in state uncertainty are eliminated. Candidate sequences that meet the constraints are retained as executable control sequences. The preset prediction time domain is based on the control period as the time step, predicting the state changes of the fermentation process within the next 15 control periods. The generation of the candidate control variable sequences is as follows: Within the current control cycle, the updated control variable constraint interval is read. Using the upper and lower boundaries of the control variable constraint interval as the range, the aeration rate, stirring speed, feeding rate, temperature regulation power, and pH regulator addition amount are discretized to obtain multiple candidate values ​​for each control variable within the current control cycle. Then, according to the preset prediction time domain, the candidate values ​​of each control variable within the continuous control cycle are combined and arranged to form multiple sets of control variable change trajectories. Each set of control variable change trajectories is then used as a control variable candidate sequence to generate a set of control variable candidate sequences for robust rolling optimization evaluation. In the execution sequence screening unit, the executable control sequences are sorted, and the candidate control variable sequence with the largest reduction in state deviation, the smallest increase in state uncertainty, and the fewest number of control variable switching times is selected first. The first control variable of the best-ranked candidate control variable sequence is determined as the control instruction for the current control cycle. Control commands are sent to the actuators of the fermentation unit to drive the aeration actuator, stirring actuator, feeding actuator, temperature adjustment actuator, and pH adjustment actuator to perform corresponding adjustment actions. In the next control cycle, the updated state vector, state uncertainty parameters, and control variable constraint intervals are reread to repeatedly execute constraint self-scaling, robust rolling optimization, and execution sequence filtering.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the antibiotic fermentation production workshop of a bio-fermentation production company. This workshop is equipped with 3000L industrial fermenters, primarily used for the production of streptomycin antibiotics. The fermentation cycle is typically 72 to 84 hours. In traditional production processes, the control system mainly relies on a few process variables such as temperature, pH, and dissolved oxygen concentration for control and adjustment, and manually intervenes in aeration rate, stirring speed, and feed rate using empirical rules. Because the fermentation process has distinct stages and metabolic changes, data from different sensors often exhibit asynchronous fluctuations, short-term drift, or local anomalies, leading to unstable fermentation state identification. For example, during the mid-fermentation feeding stage, an increase in the feed rate causes a short-term decrease in dissolved oxygen, while the carbon dioxide concentration in the exhaust gas increases simultaneously. If control is based solely on a single dissolved oxygen signal, it is easy to misjudge insufficient oxygen supply, leading to excessive increases in aeration rate or stirring speed, resulting in increased energy consumption and fluctuations in the fermentation environment. Therefore, in actual production processes, there is an urgent need for a fermentation environment control method that can integrate multimodal data and identify stable states.

[0029] The fermentation environment control method based on multimodal data fusion proposed in this invention is applied in the fermentation workshop. First, temperature sensors, pH sensors, dissolved oxygen sensors, exhaust gas analyzers, stirring speed detectors, and feed flow meters are installed on the fermentation unit. Simultaneously, an image acquisition device for the fermentation broth is configured to collect multimodal data in real time during the fermentation process. The collected data includes temperature data, pH data, dissolved oxygen concentration data, exhaust gas concentration data, stirring speed data, aeration rate data, and feed rate data. The system performs outlier detection, signal filtering, and time synchronization processing on the collected multimodal data to form a multivariate time-series dataset. Subsequently, this multivariate time-series data is input into an improved MTAD-GAT model. The model learns the correlations between the various modal data through its internal time-series coding layer, coupled enhanced graph attention layer, and dual-branch decoding layer, and calculates the prediction error and reconstruction error to obtain the dynamic reliability weights of each modality.

[0030] The system constructs a cross-modal coupling relationship matrix for the fermentation process based on dynamic reliability weights, and iteratively corrects the multimodal data by calculating coupling residuals to reconstruct the fermentation process state vector and calculate state uncertainty parameters. Subsequently, it calculates a cross-modal coupling stability index based on the trend of coupling residual changes. When the stability index meets the stability interval condition, it performs state anchoring processing on the state vector of the corresponding time period to obtain stable reference state parameters. Based on the state vector, state uncertainty parameters, and stable reference state parameters, it constructs the metabolic feasible region of the fermentation process, and performs self-evolutionary updates on the boundary of the metabolic feasible region based on the historical trajectory of fermentation state changes over time to obtain the constraint interval of control variables.

[0031] Finally, the system constructs a self-scaling robust predictive controller, which dynamically shrinks or expands the constraint interval of the control variables through a constraint self-scaling unit, and generates a candidate sequence of control variables through a robust rolling optimization unit. The optimal control command is determined by the execution sequence screening unit, thereby realizing the automatic adjustment of aeration rate, stirring speed, feeding rate, temperature regulation power, and pH adjustment agent addition.

[0032] During the application period, technicians conducted 12 batches of fermentation production experiments in the production workshop. Six batches were controlled using traditional methods, while the other six batches were controlled using the method of this invention.

[0033] Table 1. Key operating parameters and yield monitoring data for the fermentation process.

[0034] As shown in Table 1, the key parameters changed relatively steadily over time during fermentation, consistent with the general pattern of microbial fermentation. In the initial stage of fermentation (12h to 24h), the cells gradually adapted to the culture environment and began to grow rapidly. The temperature was maintained between 30.6℃ and 30.8℃, the pH value decreased from 6.72 to 6.55, and the dissolved oxygen concentration decreased from 38% to 34%. To ensure oxygen supply, the stirring speed was increased from 210 r / min to 225 r / min, and the product concentration increased from 0.45 g / L to 1.12 g / L, indicating that the cells had entered the product synthesis stage.

[0035] During the mid-fermentation stage (36-60 hours), the metabolic activity of the microorganisms increases, leading to increased oxygen consumption and a gradual decrease in dissolved oxygen concentration to between 30% and 28%. The control system correspondingly increases the stirring speed to 240-250 r / min to enhance gas-liquid mass transfer efficiency. During this stage, the pH slowly decreases to around 6.30, the temperature stabilizes at approximately 31.0℃, and the product concentration increases from 2.03 g / L to 3.48 g / L, indicating that fermentation has entered the main product accumulation stage.

[0036] During the later stages of fermentation (72 to 84 hours), the cell metabolism gradually stabilized, with dissolved oxygen concentration maintained at 27% to 28%. The stirring speed was slightly reduced to 240 r / min to decrease energy consumption. The temperature stabilized at 30.8℃ to 30.9℃, and the pH remained between 6.28 and 6.31, indicating relatively stable environmental conditions. At this point, the product concentration further increased to 4.18 g / L, indicating that the fermentation process maintained a certain capacity for product accumulation even in the later stages, and the overall fermentation process proceeded smoothly.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fermentation environment control method based on multimodal data fusion, characterized in that, include: Multiple sensors and information acquisition devices are arranged in the fermentation device to collect multimodal data during the fermentation process. The multimodal data is preprocessed to form a multivariate time series dataset. The multivariate time series dataset is input into the improved MTAD-GAT model to calculate the prediction error, reconstruction error and cross-modal consistency score of each variable, and to determine the dynamic reliability weight of each modality data. The cross-modal coupling relationship matrix of the fermentation process is constructed based on the dynamic reliability weight, and the coupling residuals between each mode are calculated. The multimodal data is iteratively corrected by the coupling residual constraints, the state vector of the fermentation process is reconstructed, and the corresponding state uncertainty parameters are determined. The cross-modal coupling stability index is calculated based on the cross-modal coupling relationship matrix and the trend of coupling residual change. The coupling stability interval in the fermentation process is identified based on the coupling stability index, and the state vector is anchored to obtain stable reference state parameters. The metabolic feasible region of the fermentation process is constructed based on the state vector, state uncertainty parameters, and stable reference state parameters. The boundary of the metabolic feasible region is then updated through self-evolution based on the historical trajectory of the fermentation state over time, forming a control variable constraint interval. A self-scaling robust predictive controller is constructed. The state vector, state uncertainty parameters, and control variable constraint intervals are input into the self-scaling robust predictive controller. Control commands are generated by dynamically scaling the control variable constraint intervals and performing rolling optimization calculations. The control commands are then executed by the actuators of the fermentation unit to adjust the fermentation environment parameters.

2. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The multimodal data includes temperature data, pH data, dissolved oxygen concentration data, gas emission data, stirring speed data, aeration rate data, feeding rate data, and fermentation broth image feature data.

3. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The preprocessing of multimodal data includes outlier detection, signal filtering, missing data imputation, feature standardization, and time synchronization.

4. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The determination of the dynamic reliability weights for each modality of data includes: The input of the multivariate time series dataset is organized by arranging the continuous observations of each modal variable within a preset time window in chronological order to form a set of time series data segments for modeling. An improved MTAD-GAT model is constructed based on a set of time-series data segments. The improved MTAD-GAT model includes a time-series coding layer, a coupled enhanced graph attention layer, and a dual-branch decoding layer. In the temporal coding layer, the temporal data segment corresponding to each modal variable is input into a shared temporal coding unit, and feature transformation and nonlinear mapping are performed on the observation values ​​at each time point to obtain a temporal coding feature sequence that corresponds one-to-one with each modal variable; In the coupled enhanced graph attention layer, static modal correlation is constructed based on the long-term statistical correlation of multivariate time series data, and dynamic modal correlation is obtained by combining the short-term collaborative change characteristics of each modal variable within the current time window. When calculating the graph attention weight, both static and dynamic modal correlation are introduced. The dynamic reliability weights of each modality obtained in the previous time window are used as gating factors to modulate the attention weights in order to generate the fused latent feature representation. In the dual-branch decoding layer, the latent feature representations are input into the prediction branch and the reconstruction branch respectively. The prediction branch outputs the predicted values ​​of each modal variable, and the reconstruction branch outputs the reconstructed values ​​of each modal variable within the current time window. Based on the deviation between the predicted values ​​and the actual observed values ​​of each modal variable, as well as the deviation between the reconstructed values ​​and the actual observed values, the prediction error, reconstruction error, and cross-modal consistency score of each modal variable are obtained. The dynamic reliability weight of each modal data is determined by the prediction error and the reconstruction error according to the preset mapping rules.

5. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The reconstructed fermentation process state vector, determining the corresponding state uncertainty parameters, includes: Based on the dynamic reliability weights of each modality data, the time series of each modality in the multivariate time series data are weighted to obtain the weighted time series data corresponding to each modality. Within a preset time window, based on the coordinated changes of weighted time series data, the statistical correlation characteristics between modes in the long-term historical data of the fermentation process, and the predefined causal constraints in the fermentation process, a basic coupling sub-matrix, a stage correction sub-matrix, and a reliability correction sub-matrix are constructed respectively. The long-term coupling relationship reflected by the basic coupling sub-matrix, the current fermentation stage characteristic coupling relationship reflected by the stage correction sub-matrix, and the dynamic reliability correction effect reflected by the reliability correction sub-matrix are superimposed to obtain the cross-modal coupling relationship matrix of the fermentation process. Based on the cross-modal coupling relationship matrix, the weighted time series data of each mode are coupled and estimated to obtain the coupling estimate value of each mode based on other modes within the current time window. The actual observation data of each mode is compared with the corresponding coupling estimate value to calculate the coupling residual sequence between modes. Iterative correction processing is performed on multimodal data under coupled residual constraints. In each iteration, the weighted time series data is updated using coupled residuals according to the preset convergence criteria, and data change amplitude constraints and smoothness constraints are applied at the same time. The iteration stops when the coupled residuals meet the preset threshold conditions, and the corrected multimodal time series data is obtained. The fermentation process state vectors corresponding to each time point are extracted from the corrected multimodal time series data. The coupled residual distribution characteristics at each time point are statistically analyzed to obtain the state uncertainty parameters corresponding to the state vectors.

6. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The process of obtaining stable reference state parameters includes: Obtain the cross-modal coupling relationship matrix and the coupling residual sequence of each mode. Under a preset sampling period, use the coupling relationship strength and dynamic reliability weight in the cross-modal coupling relationship matrix as weights to perform weighted summation of the coupling residuals of each mode to obtain the coupling stability evaluation value sequence. Within a sliding time window, the window mean, window standard deviation, and number of times the coupling stability evaluation value sequence is calculated. The window mean represents the average level of coupling deviation, the window standard deviation represents the degree of fluctuation of coupling deviation, and the number of times the threshold is exceeded within the window represents the duration of coupling deviation exceeding the preset residual threshold. The cross-modal coupling stability index is calculated based on the window mean, window standard deviation, and the number of times the threshold is exceeded within the window. The cross-modal coupling stability index is obtained by weighting and summing the window mean, window standard deviation, and the number of times the threshold is exceeded within the window according to a preset weight. The preset weight is a constant parameter that satisfies the weight sum to 1. The calculated cross-modal coupling stability index is then formed into a cross-modal coupling stability index sequence. The cross-modal coupling stability index sequence is compared with the stability judgment threshold interval. The time period in which the cross-modal coupling stability index falls into the stability judgment threshold interval and continuously meets the preset duration is determined as the coupling stability interval. The fermentation process state vector at the corresponding time is extracted in the coupling stability interval to form a candidate anchoring state vector set. Clustering is performed on the candidate anchoring state vector set to obtain cluster center vectors, and time smoothing is performed on the cluster center vectors. The time-smoothed cluster center vectors are then determined as the stable reference state parameters of the coupled stable interval.

7. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The formation of the control variable constraint interval includes: Obtain the state vector, state uncertainty parameters, and stable reference state parameters. Compare the state vector with the stable reference state parameters dimension by dimension. Extract the deviation direction, deviation magnitude, and duration of each state component relative to the stable reference state parameters to form fermentation state deviation description information. Based on the description information of the deviation of fermentation state, the control relationship between each state component is identified as a metabolic conflict relationship, and a set of metabolic conflict constraint units is constructed based on the metabolic conflict relationship. Based on the set of metabolic conflict constraint units, corresponding local feasible boundary fragments are generated for aeration rate, stirring speed, feeding rate, temperature regulation power and pH regulator addition amount, respectively. The local feasible boundary fragments are then screened, spliced ​​and closed according to the compatibility relationship between each metabolic conflict constraint unit to form the metabolic feasible domain of the fermentation process corresponding to the current fermentation state. Boundary memory trajectories are established for each boundary segment of the metabolic feasible domain in the fermentation process. The boundary memory trajectories record the position changes, length changes and boundary switching order of each boundary segment within multiple consecutive time windows. When a change in state uncertainty parameter, a shift in state vector deviation direction or a reorganization of metabolic conflict relationship is detected, boundary migration update is triggered, so that each boundary segment in the current time window completes position migration, length contraction or length expansion based on the boundary memory trajectory of the previous time window, and obtains the self-evolutionary updated boundary of the metabolic feasible domain. Based on the metabolic feasible domain boundary updated through self-evolution, the constraint intervals of the control variables corresponding to ventilation rate, stirring speed, feeding rate, temperature regulation power, and pH regulator addition amount are determined.

8. The fermentation environment control method based on multimodal data fusion according to claim 1, characterized in that, The process of generating control commands by dynamically scaling the constraint interval of control variables and performing rolling optimization calculations, and then having the fermentation device actuator execute the control commands to adjust the fermentation environment parameters, includes: Based on the state vector, state uncertainty parameters, and control variable constraint intervals, a self-scaling robust predictive controller is constructed. The self-scaling robust predictive controller includes a constraint self-scaling unit, a robust rolling optimization unit, and an execution sequence filtering unit. In the constrained self-scaling unit, the state uncertainty parameter, the deviation information between the state vector and the stable reference state parameter, and the boundary state of the metabolic feasible domain are read. According to the preset scaling rules, the boundary shrinkage, boundary preservation or boundary expansion processing is performed on each control variable constraint interval, so that the ventilation rate, stirring speed, feeding rate, temperature regulation power and pH regulator addition amount correspond to the updated control variable constraint interval respectively. In the robust rolling optimization unit, the updated control variable constraint interval is used as the control boundary. The joint constraint objectives are the convergence of the state vector to the stable reference state parameter, the continuous change of control variables, and the suppression of state uncertainty. Multiple candidate sequences of control variables are generated according to the preset prediction time domain. The state evolution results of each candidate sequence of control variables are evaluated cycle by cycle in the future continuous control cycle. Candidate sequences of control variables that exceed the updated control variable constraint interval or cause an increase in state uncertainty are eliminated. The candidate sequences of control variables that meet the constraint conditions are retained as executable control sequences. In the execution sequence screening unit, the executable control sequences are sorted, and the candidate control variable sequence with the largest reduction in state deviation, the smallest increase in state uncertainty, and the fewest number of control variable switching times is selected first. The first control variable of the best-ranked candidate control variable sequence is determined as the control instruction for the current control cycle. Control commands are sent to the actuators of the fermentation unit to drive the aeration actuator, stirring actuator, feeding actuator, temperature adjustment actuator, and pH adjustment actuator to perform corresponding adjustment actions. In the next control cycle, the updated state vector, state uncertainty parameters, and control variable constraint intervals are reread to repeatedly execute constraint self-scaling, robust rolling optimization, and execution sequence filtering.

Images