Automatic control method and system for fumaric acid production

Abstract

The invention discloses an automatic control method and system for fumaric acid production. The method comprises the following steps: collecting process variables in a fermentation process; constructing a soft measurement model fusing physical modeling and a residual neural network to obtain an acid production rate prediction value and prediction confidence; calculating a feed-forward control quantity, and adjusting the weight according to the prediction confidence to obtain a feed-forward control output; constructing a state vector, and establishing a Markov decision process; training a neural network control strategy by adopting a near-end strategy optimization algorithm, and outputting a proportionality coefficient, an integral coefficient and a differential coefficient; calculating feedback control quantity, fusing the feedback control quantity with feed-forward control output, and applying change rate and amplitude limitation to obtain final control output; and switching to a feedback priority control mode under a specific condition, and regularly updating a soft measurement model and a control strategy parameter. According to the invention, high-precision stable control of pH in the fumaric acid fermentation process is realized, the production efficiency and the process robustness are improved, and the method is suitable for automatic control of the large-scale fermentation process.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for fermentation processes, and in particular to an automated control method and system for fumaric acid production. Background Technology

[0002] With the increasing demand for bio-based chemical products, the automated control of fumaric acid's fermentation production process, as an important organic acid platform chemical, has attracted widespread attention. Existing fumaric acid fermentation process control methods mainly rely on online pH detection and fixed-parameter PID control to regulate alkali dosage. However, these methods generally suffer from the following problems in practical applications:

[0003] The collected process variables, such as pH value, dissolved oxygen concentration, and alkali addition acceleration rate, are greatly affected by environmental disturbances and instrument noise, resulting in insufficient data stability. This leads to limited accuracy in the estimation of acid production rate based on physical modeling, and the prediction results are highly volatile and unreliable. Traditional control methods mostly rely on single feedback regulation, lacking the ability to respond in advance to changes in the acid production rate. When the acid production rate rises rapidly or the buffer capacity changes drastically, alkali addition often overshoots or lags, affecting pH stability. Existing PID parameters are usually manually set or periodically adjusted, making it difficult to adapt to the nonlinear, time-varying, and uncertain conditions in the fermentation process. The control effect gradually declines over time. When the accuracy of the soft sensor model or controller decreases, the lack of confidence monitoring and adaptive update mechanisms increases the risk of model parameter drift and control instability, making it difficult to guarantee the stability and consistency of large-scale fermentation production.

[0004] Therefore, how to provide an automated control method and system for fumaric acid production is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automated control method and system for fumaric acid production. This invention fully utilizes a soft measurement method that combines physical modeling with residual neural networks and attention mechanisms to predict the rate of acid production per unit time. It employs a proximal policy optimization algorithm based on Markov decision processes to train a neural network control strategy to achieve adaptive adjustment of PID parameters. The invention details the weighted fusion of feedforward control and feedback control, as well as the confidence-driven control mode switching, and possesses the advantages of high prediction accuracy, fast control response, and strong stability of the fermentation process.

[0006] An automated control method for fumaric acid production according to an embodiment of the present invention includes the following steps:

[0007] Collect process variables during fermentation;

[0008] A soft measurement model integrating physical modeling and residual neural networks is constructed based on process variables. The input feature weights are allocated using an attention mechanism to obtain the predicted acid production rate per unit time and output the prediction confidence.

[0009] The feedforward control quantity is calculated based on the relationship between the predicted acid production rate and the neutralization reaction, and the weight of the feedforward control quantity is adjusted according to the prediction confidence level to obtain the feedforward control output.

[0010] A state vector is constructed based on process variables, predicted acid production rate, prediction confidence level, and historical control parameters, and a Markov decision process is established based on the state vector.

[0011] Within the Markov decision process, a neural network control strategy is trained using a proximal policy optimization algorithm, which receives the state vector and outputs the proportional coefficient, integral coefficient, and derivative coefficient.

[0012] The feedback control quantity is calculated based on the proportional coefficient, integral coefficient, derivative coefficient, and pH deviation. The feedback control quantity is then fused with the feedforward control output, and a rate of change limit and an amplitude limit are applied to obtain the final control output.

[0013] When the prediction confidence level is lower than the preset threshold or the process variable is abnormal, the system switches to feedback priority control mode and updates the parameters of the soft measurement model and neural network control strategy regularly.

[0014] Optionally, the process variables collected during the fermentation process specifically include:

[0015] Establish a unified time base and sampling period;

[0016] Collect pH value, dissolved oxygen concentration, stirring power, inlet air flow rate, alkali addition acceleration rate and exhaust gas composition according to the sampling cycle;

[0017] The collected data is timestamped and aligned to a time base to form synchronized process variables.

[0018] Optionally, the construction of a soft measurement model based on process variables, integrating physical modeling and residual neural networks, and the allocation of input feature weights using an attention mechanism to obtain a predicted acid production rate per unit time and output the prediction confidence level, specifically includes:

[0019] A soft measurement model integrating physical modeling and residual neural network is constructed. The soft measurement model takes process variables as input and outputs the predicted acid production rate and prediction confidence level per unit time.

[0020] Based on the principle of acid-base neutralization reaction and the mass conservation relationship of buffer solution, a physical modeling formula is established to estimate the acid production rate per unit time.

[0021] Extract the process variables and acid production rate estimates at the current time and the previous K time steps, and combine the process variables and corresponding acid production rate estimates at each time step to form a joint input, thus forming a joint input sequence containing multiple time steps;

[0022] The joint input sequence is fed into the time attention mechanism, the attention weight of each time step in the current prediction task is calculated, and the joint input sequence is weighted and aggregated according to the attention weight to generate the time-aware feature vector of the current time.

[0023] The time-aware feature vector is input into the channel attention mechanism, and the attention weight is calculated based on the importance of each input feature dimension in the feature representation. The weighting operation is then performed on each channel to generate a structured feature vector.

[0024] The structured feature vector is input into the residual neural network, and the output includes the predicted acid production rate per unit time and the corresponding prediction confidence.

[0025] The soft measurement model is trained using a joint loss function based on uncertainty modeling. The joint loss function consists of a prediction error term and a confidence regularization term.

[0026] Optionally, the step of calculating the feedforward control quantity based on the predicted acid production rate and the relationship between the neutralization reaction, and adjusting the weight of the feedforward control quantity according to the prediction confidence level to obtain the feedforward control output, specifically includes:

[0027] Based on the difference between the predicted acid production rate per unit time and the preset target acid production rate, combined with the system volume and buffer capacity, the theoretical amount of alkali solution required for the neutralization reaction is calculated.

[0028] Based on the current alkali concentration, the theoretical alkali dosage is converted into the target alkali dosing flow rate;

[0029] The target alkali solution dosing flow rate is weighted by the predicted confidence level to generate a confidence-corrected feedforward control quantity, and the feedforward control output is obtained.

[0030] Optionally, the step of constructing a state vector based on process variables, predicted acid production rate, prediction confidence level, and historical control parameters, and establishing a Markov decision process based on the state vector, specifically includes:

[0031] Extract the process variables at the current moment, the predicted acid production rate per unit time, the prediction confidence level, and the proportional coefficient, integral coefficient, and differential coefficient of the previous moment, and concatenate them in a preset order to form a state vector;

[0032] The state space is defined as all combinations of state vectors within the allowed range of values.

[0033] The action space is defined by a set of triplets of continuous values ​​of the proportional coefficient, integral coefficient, and differential coefficient within their respective preset upper and lower limits, and the value boundaries of each coefficient are set.

[0034] A Markov decision process is established based on the state space and action space. The reward function is set as a weighted combination of the squared penalty term for pH deviation and the penalty term for changes in alkali addition flow rate, thus forming a correspondence between state, action and reward.

[0035] Optionally, within the Markov decision process, the neural network control strategy is trained using a proximal policy optimization algorithm, receiving the state vector and outputting proportional coefficients, integral coefficients, and derivative coefficients, specifically including:

[0036] In the Markov decision process, a proximal policy optimization algorithm is used to train the neural network control strategy. The initial values ​​of the shearing coefficient and the learning rate are set, and the shearing coefficient is set to a value that monotonically increases with the prediction confidence.

[0037] The state vector is input into the neural network control strategy, which outputs proportional coefficient, integral coefficient, and derivative coefficient, and projects the proportional coefficient, integral coefficient, and derivative coefficient into the action space boundary.

[0038] The data sequence is formed by collecting the state vector, proportional coefficient, integral coefficient, and differential coefficient, the reward calculated based on the reward function, and the state vector at the next time step in chronological order.

[0039] Calculate the cumulative discount reward based on the data sequence, set the discount factor to a fixed value, and apply the same discount to all time steps;

[0040] A policy update objective function based on the shearing target is constructed, and the prediction confidence is used as the sample weight to weight the policy update objective function;

[0041] Batch stochastic gradient updates are performed on the neural network control strategy. The learning rate is set to a value that decreases monotonically with the absolute value of pH deviation and the rate of pH change. Upper limit constraints are imposed on the single-step changes of the proportional coefficient, integral coefficient, and derivative coefficient.

[0042] The process of collecting, calculating, building, and updating data is repeated until the preset convergence condition is met.

[0043] Optionally, the step of calculating the feedback control quantity based on the proportional coefficient, integral coefficient, derivative coefficient, and pH deviation, and fusing the feedback control quantity with the feedforward control output, applying rate-of-change limits and amplitude limits to obtain the final control output, specifically includes:

[0044] The proportionality term is calculated based on the proportionality coefficient and pH deviation to obtain the proportionality term value;

[0045] The pH deviation is numerically integrated within a preset time window. The integral term is calculated based on the integration coefficient and the integration result. The upper and lower limits of the integral term and the anti-saturation backoff are applied to the integral term to obtain the integral term value.

[0046] After applying a first-order low-pass filter to the pH deviation, the rate of change of the deviation is calculated. The differential term is then calculated based on the differential coefficients and the rate of change of the deviation. High-frequency suppression and amplitude limiting are applied to the differential term to obtain its value.

[0047] The proportional term, integral term, and derivative term are added together to obtain the feedback control quantity.

[0048] The feedforward weights are calculated based on the prediction confidence level. The feedforward control output and the feedback control quantity are then weighted and fused according to the feedforward weights and complementary weights to obtain the fused control quantity.

[0049] Apply a rate-of-change limit to the fusion control quantity to limit the allowable variation within adjacent sampling periods;

[0050] An amplitude limit is applied to the fusion control quantity after the rate of change is limited, restricting it to the boundary between the minimum and maximum allowable flow rates of the equipment, thus obtaining the final control output.

[0051] Optionally, when the prediction confidence level is lower than a preset threshold or the process variable is abnormal, the system switches to a feedback-priority control mode and periodically updates the parameters of the soft sensor model and the neural network control strategy, specifically including:

[0052] The system compares the predicted confidence level with the preset threshold and detects whether the process variable exceeds the preset range or whether the pH change rate exceeds the preset upper limit. If any condition is met, it enters the feedback priority control mode.

[0053] In the feedback priority control mode, the feedforward weight is set to zero, the proportional coefficient, integral coefficient and derivative coefficient of the previous moment are kept unchanged, the feedback control quantity is calculated based on the current pH deviation, and the rate of change limit and amplitude limit are applied to the feedback control quantity to obtain the final control output.

[0054] In the feedback priority control mode, the state vector, final control output, pH value and alkali injection acceleration rate are collected within a preset time length. The soft sensor model is then calibrated, including bias correction and gain correction of the predicted acid production rate per unit time, and the soft sensor model parameters are updated.

[0055] In the feedback-priority control mode, a proximal policy optimization algorithm is used to perform incremental updates on the neural network control policy with a low learning rate.

[0056] When the prediction confidence is not lower than the preset recovery threshold within a continuous preset sampling period and the process variable is within a preset range, the feedback priority control mode is exited, the feedforward weight is set to the value calculated based on the prediction confidence, the feedforward control output and the feedback control quantity are weighted and fused according to the feedforward weight and the complementary weight, and the change rate limit and amplitude limit are applied to the weighted result to obtain the final control output.

[0057] An automated control system for fumaric acid production according to an embodiment of the present invention includes the following modules:

[0058] The process data acquisition module is used to acquire process variables during fermentation and transmit them to subsequent processing.

[0059] The acid production rate soft measurement module is used to obtain the predicted acid production rate per unit time based on physical modeling and residual neural network combined with time attention mechanism and output the prediction confidence.

[0060] The feedforward control calculation module is used to calculate the feedforward control output based on the relationship between the predicted acid production rate and the neutralization reaction, and to adjust the weight of the feedforward control output based on the prediction confidence.

[0061] The state vector construction module is used to concatenate process variables, predicted acid production rate, prediction confidence level and historical control parameters to form a state vector, and to establish a Markov decision process based on the state space and action space.

[0062] The control strategy training module is used to train the neural network control strategy using the proximal policy optimization algorithm in the Markov decision process, and inputs the state vector and outputs the proportional coefficient, integral coefficient and derivative coefficient.

[0063] The feedback control calculation module is used to calculate the feedback control quantity based on the proportional coefficient, integral coefficient and derivative coefficient, and then weight and fuse the feedforward control output with the feedback control quantity. After applying the rate of change limit and the amplitude limit, the final control output is obtained.

[0064] The model adaptation module is used to enter the feedback priority control mode when the prediction confidence is below the threshold or the process variable is abnormal. It performs parameter calibration on the soft measurement model and performs incremental updates on the neural network control strategy.

[0065] The beneficial effects of this invention are:

[0066] This invention constructs a soft sensor model by integrating physical modeling based on the acid-base neutralization reaction principle with a residual neural network incorporating a time attention mechanism. It combines this with a proximal policy optimization reinforcement learning control method based on Markov decision processes. Addressing the problems of unstable acid production rate estimation, feedforward lag, difficulty in adaptive PID parameters, and control instability under abnormal conditions in existing fumaric acid fermentation pH control, this invention employs a joint input sequence modeling of process variables and physical model estimates. By combining time attention and channel attention to achieve weighted aggregation of key moments and key features, it outputs a predicted acid production rate per unit time and its prediction confidence. Furthermore, it introduces a weighted aggregation mechanism into the feedforward control calculation. A confidence-weighted correction method is used to generate a stable feedforward control output. In the control strategy optimization stage, a Markov decision process is constructed based on the state vector, and a proximal strategy optimization algorithm is used to dynamically output the proportional, integral, and derivative coefficients. The stability of parameter updates is ensured by a confidence-weighted objective function and a rate-of-change constraint. In the control execution stage, the feedforward control output and feedback control quantity are weighted and fused, and rate-of-change and amplitude constraints are applied to obtain the final control output that satisfies the process constraints. When the prediction confidence is low or the process variables are abnormal, the system switches to a feedback-first mode and performs soft-sensor model calibration and incremental updates of the control strategy, achieving adaptive repair of the model and controller. This invention significantly improves the accuracy and stability of acid production rate prediction, achieves dynamic fusion of feedforward and feedback control, enhances the adaptability and convergence of PID parameter adjustment, effectively reduces the risk of control instability, and ensures the efficient, stable, and safe operation of the fumaric acid fermentation process. It has the value for widespread application in industrial-scale fermentation production. Attached Figure Description

[0067] 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:

[0068] Figure 1 This is a flowchart of an automated control method for fumaric acid production proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the soft sensor model in an automated control method for fumaric acid production proposed in this invention;

[0070] Figure 3 This is a schematic diagram of the structure of an automated control system for fumaric acid production proposed in this invention. Detailed Implementation

[0071] 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.

[0072] refer to Figure 1-2 An automated control method for fumaric acid production includes the following steps:

[0073] Collect process variables during fermentation;

[0074] A soft measurement model integrating physical modeling and residual neural networks is constructed based on process variables. The input feature weights are allocated using an attention mechanism to obtain the predicted acid production rate per unit time and output the prediction confidence.

[0075] The feedforward control quantity is calculated based on the relationship between the predicted acid production rate and the neutralization reaction, and the weight of the feedforward control quantity is adjusted according to the prediction confidence level to obtain the feedforward control output.

[0076] A state vector is constructed based on process variables, predicted acid production rate, prediction confidence level, and historical control parameters, and a Markov decision process is established based on the state vector.

[0077] Within the Markov decision process, a neural network control strategy is trained using a proximal policy optimization algorithm, which receives the state vector and outputs the proportional coefficient, integral coefficient, and derivative coefficient.

[0078] The feedback control quantity is calculated based on the proportional coefficient, integral coefficient, derivative coefficient, and pH deviation. The feedback control quantity is then fused with the feedforward control output, and a rate of change limit and an amplitude limit are applied to obtain the final control output.

[0079] When the prediction confidence level is lower than the preset threshold or the process variable is abnormal, the system switches to feedback priority control mode and updates the parameters of the soft measurement model and neural network control strategy regularly.

[0080] In this embodiment, the process variables collected during the fermentation process specifically include:

[0081] Establish a unified time base and sampling period;

[0082] Collect pH value, dissolved oxygen concentration, stirring power, inlet air flow rate, alkali addition acceleration rate and exhaust gas composition according to the sampling cycle;

[0083] The collected data is timestamped and aligned to a time base to form synchronized process variables.

[0084] This implementation method synchronously collects data on pH value, dissolved oxygen concentration, stirring power, inlet air flow rate, alkali addition acceleration rate, and exhaust gas components during the fermentation process by setting a unified time base and sampling period, and aligns the data with timestamps to ensure consistency and comparability of process variables at the same time scale. This approach significantly reduces errors caused by data delay and offset, improves the accuracy of correlation analysis between multi-source variables, provides high-quality, time-consistent input data for subsequent soft sensor modeling and control strategy training, and enhances the real-time performance and reliability of the control method.

[0085] In this embodiment, the construction of a soft measurement model based on process variables, integrating physical modeling and residual neural networks, and the allocation of input feature weights using an attention mechanism to obtain a predicted acid production rate per unit time and output the prediction confidence level, specifically includes:

[0086] A soft measurement model integrating physical modeling and residual neural network is constructed. The soft measurement model takes process variables as input and outputs the predicted acid production rate and prediction confidence level per unit time.

[0087] Based on the principles of acid-base neutralization and the conservation of mass in buffer solutions, a physical modeling formula is established to estimate the rate of acid production per unit time. The calculation formula is:

[0088]

[0089] Among them, Q t C is the flow rate of the alkali solution. b Where β is the concentration of the alkali solution, V is the volume of the fermenter, and β is the concentration of the alkali solution. t Given the current buffer capacity, ΔpH t Δt represents the change in pH value per unit time, where Δt is the time interval.

[0090] Extract the process variables and acid production rate estimates at the current time and the previous K time steps, and combine the process variables and corresponding acid production rate estimates at each time step to form a joint input, thus forming a joint input sequence containing multiple time steps;

[0091] The joint input sequence is fed into the time attention mechanism to calculate the attention weight of each time step in the current prediction task. The joint input sequence is then weighted and aggregated based on the attention weight to generate the time-aware feature vector of the current moment, thereby enhancing the model's ability to represent changes at key moments.

[0092] The time-aware feature vector is input into the channel attention mechanism, and the attention weight is calculated based on the importance of each input feature dimension in the feature representation. The weighting operation is then performed on each channel to generate a structured feature vector.

[0093] The structured feature vector is input into the residual neural network, and the output includes the predicted acid production rate per unit time and the corresponding prediction confidence.

[0094] The soft measurement model is trained using a joint loss function based on uncertainty modeling. It consists of a prediction error term and a confidence regularization term, and its expression is:

[0095]

[0096] Where rt This represents the actual acid production rate. α represents the predicted rate of acid production. t This represents the prediction confidence level of the model output.

[0097] This implementation constructs a soft-sensor model that integrates physical modeling and residual neural networks. By combining time-attention and channel-attention mechanisms, it performs feature weighting and structured representation of multi-time- and multi-dimensional process variables, obtaining the predicted acid production rate per unit time and its prediction confidence level. This approach, while ensuring physical mechanism constraints, fully utilizes data-driven deep feature learning capabilities, improving the modeling accuracy of nonlinear dynamic processes and significantly enhancing the reliability and robustness of the prediction results, providing highly reliable input for subsequent control calculations.

[0098] In this embodiment, the step of calculating the feedforward control quantity based on the predicted acid production rate and the relationship between the neutralization reaction, and adjusting the weight of the feedforward control quantity according to the prediction confidence level to obtain the feedforward control output, specifically includes:

[0099] Based on the difference between the predicted acid production rate per unit time and the preset target acid production rate, combined with the system volume and buffer capacity, the theoretical amount of alkali solution required for the neutralization reaction is calculated.

[0100] Based on the current alkali concentration, the theoretical alkali dosage is converted into the target alkali dosing flow rate;

[0101] The target alkali solution dosing flow rate is weighted by the predicted confidence level to generate a confidence-corrected feedforward control quantity, and the feedforward control output is obtained.

[0102] This implementation calculates the required alkali dosage based on the predicted acid production rate and the relationship between the neutralization reaction, and then weights and corrects the target flow rate by incorporating the prediction confidence level to obtain the feedforward control output. This method enables early response to changes in the acid production rate, reduces the impact of pH fluctuations on system stability, and automatically reduces the feedforward effect when prediction uncertainty is high, thereby improving the adaptability and robustness of the control strategy and effectively ensuring the stable operation of the fermentation process and the efficiency of alkali utilization.

[0103] In this embodiment, the construction of a state vector based on process variables, predicted acid production rate, prediction confidence level, and historical control parameters, and the establishment of a Markov decision process based on the state vector, specifically includes:

[0104] Extract the process variables at the current moment, the predicted acid production rate per unit time, the prediction confidence level, and the proportional coefficient, integral coefficient, and differential coefficient of the previous moment, and concatenate them in a preset order to form a state vector;

[0105] The state space is defined as all combinations of state vectors within the allowed range of values.

[0106] The action space is defined by a set of triplets of continuous values ​​of the proportional coefficient, integral coefficient, and differential coefficient within their respective preset upper and lower limits, and the value boundaries of each coefficient are set.

[0107] A Markov decision process is established based on the state space and action space. The reward function is set as a weighted combination of the squared penalty term for pH deviation and the penalty term for changes in alkali addition flow rate, thus forming a correspondence between state, action and reward.

[0108] This implementation method establishes a Markov decision process based on state and action spaces, forming a mapping relationship between states, actions, and rewards. This approach can comprehensively reflect the dynamic characteristics of the fermentation process and the historical impact of control strategies, providing a unified modeling framework for subsequent control strategy training. It improves the learning efficiency and dynamic adaptability of control strategies in complex fermentation environments, enabling precise modeling and optimization of pH regulation.

[0109] In this embodiment, the step of training a neural network control strategy using a proximal policy optimization algorithm within the Markov decision process, receiving the state vector and outputting the proportional coefficient, integral coefficient, and derivative coefficient, specifically includes:

[0110] In the Markov decision process, a proximal policy optimization algorithm is used to train the neural network control strategy. The initial values ​​of the shearing coefficient and the learning rate are set, and the shearing coefficient is set to a value that monotonically increases with the prediction confidence.

[0111] The state vector is input into the neural network control strategy, which outputs proportional coefficient, integral coefficient, and derivative coefficient, and projects the proportional coefficient, integral coefficient, and derivative coefficient into the action space boundary.

[0112] The data sequence is formed by collecting the state vector, proportional coefficient, integral coefficient, and differential coefficient, the reward calculated based on the reward function, and the state vector at the next time step in chronological order.

[0113] Calculate the cumulative discount reward based on the data sequence, set the discount factor to a fixed value, and apply the same discount to all time steps;

[0114] A policy update objective function based on the shearing target is constructed, and the prediction confidence is used as the sample weight to weight the policy update objective function;

[0115] Batch stochastic gradient updates are performed on the neural network control strategy. The learning rate is set to a value that decreases monotonically with the absolute value of pH deviation and the rate of pH change. Upper limit constraints are imposed on the single-step changes of the proportional coefficient, integral coefficient, and derivative coefficient.

[0116] The process of collecting, calculating, building, and updating data is repeated until the preset convergence condition is met.

[0117] This implementation trains a neural network control strategy using a proximal policy optimization algorithm within a Markov decision process. The state vector is used as input, and the outputs are proportional, integral, and derivative coefficients. The training parameters are dynamically adjusted based on prediction confidence and pH deviation. This approach improves the convergence speed of strategy updates while ensuring training stability, achieving adaptive optimization of control parameters. This enhances the robustness and adjustment accuracy of the control strategy under different fermentation conditions, significantly improving the dynamic response performance and process stability of pH control.

[0118] In this embodiment, the step of calculating the feedback control quantity based on the proportional coefficient, integral coefficient, derivative coefficient, and pH deviation, and fusing the feedback control quantity with the feedforward control output, applying rate of change limits and amplitude limits to obtain the final control output, specifically includes:

[0119] The proportionality term is calculated based on the proportionality coefficient and pH deviation to obtain the proportionality term value;

[0120] The pH deviation is numerically integrated within a preset time window. The integral term is calculated based on the integration coefficient and the integration result. The upper and lower limits of the integral term and the anti-saturation backoff are applied to the integral term to obtain the integral term value.

[0121] After applying a first-order low-pass filter to the pH deviation, the rate of change of the deviation is calculated. The differential term is then calculated based on the differential coefficients and the rate of change of the deviation. High-frequency suppression and amplitude limiting are applied to the differential term to obtain its value.

[0122] The proportional term, integral term, and derivative term are added together to obtain the feedback control quantity.

[0123] The feedforward weights are calculated based on the prediction confidence level. The feedforward control output and the feedback control quantity are then weighted and fused according to the feedforward weights and complementary weights to obtain the fused control quantity.

[0124] Apply a rate-of-change limit to the fusion control quantity to limit the allowable variation within adjacent sampling periods;

[0125] An amplitude limit is applied to the fusion control quantity after the rate of change is limited, restricting it to the boundary between the minimum and maximum allowable flow rates of the equipment, thus obtaining the final control output.

[0126] This implementation calculates the feedback control quantity based on the proportional coefficient, integral coefficient, derivative coefficient, and pH deviation, and then weights and fuses it with the feedforward control output. Simultaneously, it applies rate-of-change and amplitude limits to obtain the final control output. This method fully leverages the complementary advantages of feedforward and feedback, ensuring control accuracy while avoiding abrupt changes in the control quantity. It improves the stability and smoothness of the pH adjustment process, thereby ensuring that the fermentation process remains in the optimal acid-base environment, promoting efficient fumaric acid production and reducing energy consumption.

[0127] In this embodiment, the step of switching to feedback-priority control mode when the prediction confidence is lower than a preset threshold or the process variable is abnormal, and periodically updating the parameters of the soft sensor model and the neural network control strategy, specifically includes:

[0128] The system compares the predicted confidence level with the preset threshold and detects whether the process variable exceeds the preset range or whether the pH change rate exceeds the preset upper limit. If any condition is met, it enters the feedback priority control mode.

[0129] In the feedback priority control mode, the feedforward weight is set to zero, the proportional coefficient, integral coefficient and derivative coefficient of the previous moment are kept unchanged, the feedback control quantity is calculated based on the current pH deviation, and the rate of change limit and amplitude limit are applied to the feedback control quantity to obtain the final control output.

[0130] In the feedback priority control mode, the state vector, final control output, pH value and alkali injection acceleration rate are collected within a preset time length. The soft sensor model is then calibrated, including bias correction and gain correction of the predicted acid production rate per unit time, and the soft sensor model parameters are updated.

[0131] In the feedback-priority control mode, a proximal policy optimization algorithm is used to perform incremental updates on the neural network control policy with a low learning rate.

[0132] When the prediction confidence is not lower than the preset recovery threshold within a continuous preset sampling period and the process variable is within a preset range, the feedback priority control mode is exited, the feedforward weight is set to the value calculated based on the prediction confidence, the feedforward control output and the feedback control quantity are weighted and fused according to the feedforward weight and the complementary weight, and the change rate limit and amplitude limit are applied to the weighted result to obtain the final control output.

[0133] This implementation switches to a feedback-priority control mode when the prediction confidence level falls below a preset threshold or when process variables are abnormal. It then performs parameter calibration on the soft sensor model and incremental updates to the neural network control strategy, ensuring stable output of control quantities even under abnormal conditions. This approach effectively avoids control failures caused by soft sensor biases or sudden operating conditions, enhancing the system's robustness and adaptability. It enables continuous and safe regulation of the fermentation process, thereby ensuring the stability of fumaric acid production and consistent product quality.

[0134] refer to Figure 3 An automated control system for fumaric acid production includes the following modules:

[0135] The process data acquisition module is used to acquire process variables during fermentation and transmit them to subsequent processing.

[0136] The acid production rate soft measurement module is used to obtain the predicted acid production rate per unit time based on physical modeling and residual neural network combined with time attention mechanism and output the prediction confidence.

[0137] The feedforward control calculation module is used to calculate the feedforward control output based on the relationship between the predicted acid production rate and the neutralization reaction, and to adjust the weight of the feedforward control output based on the prediction confidence.

[0138] The state vector construction module is used to concatenate process variables, predicted acid production rate, prediction confidence level and historical control parameters to form a state vector, and to establish a Markov decision process based on the state space and action space.

[0139] The control strategy training module is used to train the neural network control strategy using the proximal policy optimization algorithm in the Markov decision process, and inputs the state vector and outputs the proportional coefficient, integral coefficient and derivative coefficient.

[0140] The feedback control calculation module is used to calculate the feedback control quantity based on the proportional coefficient, integral coefficient and derivative coefficient, and then weight and fuse the feedforward control output with the feedback control quantity. After applying the rate of change limit and the amplitude limit, the final control output is obtained.

[0141] The model adaptation module is used to enter the feedback priority control mode when the prediction confidence is below the threshold or the process variable is abnormal. It performs parameter calibration on the soft measurement model and performs incremental updates on the neural network control strategy.

[0142] Example 1:

[0143] To verify the feasibility of this invention in practice, it was applied to the industrial production workshop of fumaric acid in a bio-fermentation company. The core production equipment in this workshop is a 20m³... 3 The stainless steel fermenter is equipped with a pH electrode, dissolved oxygen electrode, stirring system, air inlet pipeline, and automatic alkali dosing pump. During the production process, the fermentation strain rapidly metabolizes glucose to produce organic acids, causing the system pH to continuously decrease. Traditional methods often use a fixed-parameter PID controller for pH adjustment. When the acid production rate fluctuates significantly, the PID parameters cannot be adjusted in time, easily causing delayed or over-flush alkali dosing, resulting in excessive pH fluctuations, which in turn affects acid production efficiency and final yield.

[0144] First, data such as pH, dissolved oxygen concentration, stirring power, inlet air flow rate, alkali injection acceleration rate, and exhaust gas composition during the fermentation process are collected according to a sampling period and aligned to a time reference to form synchronized process variables. Then, a soft sensor model integrating physical modeling and residual neural networks is constructed based on these process variables. A time attention mechanism is used to enhance key moment features, and a channel attention mechanism is used to assign input feature weights, yielding a predicted acid production rate per unit time and its corresponding prediction confidence level. This predicted value is used for feedforward control calculations. The theoretical alkali injection flow rate is calculated using the acid-base neutralization formula, and then weighted and corrected using the prediction confidence level to form the feedforward control output.

[0145] Simultaneously, a state vector is constructed using process variables, predicted acid production rate, prediction confidence level, and historical control parameters. Within a Markov decision process framework, a neural network control strategy is trained using a proximal policy optimization algorithm, dynamically outputting proportional, integral, and derivative coefficients. Subsequently, the feedback control quantity is obtained according to PID calculation principles and fused with the feedforward control output according to confidence weights. Then, rate-of-change and amplitude constraints are applied to obtain the final control output. The final control output directly drives the alkali injection pump to regulate the alkali flow rate, achieving stable pH regulation of the fermentation broth.

[0146] To compare and verify the effectiveness, a parallel experiment was conducted between the method of this invention and the traditional PID control method. The experiment was designed to involve six consecutive batches of fermentation, each lasting 72 hours, with an initial liquid volume of 16 m³. 3 The target pH was set at 6.5 ± 0.05. The key performance indicators compared included: average pH fluctuation, total alkali consumption, fumaric acid yield, and average acid production rate per unit time. The experimental results are shown in Table 1.

[0147] Table 1 Comparison of Traditional PID and the Method of This Invention

[0148]

[0149] As can be seen from the data in Table 1, under traditional PID control, the pH fluctuation range is generally between ±0.18 and ±0.21, exceeding the set control range. The alkali consumption remains above 930L, the fumaric acid yield remains at 85% ±0.08, the alkali consumption is reduced by about 6%, and the average is controlled at around 872L. The fumaric acid yield is increased to 90.5% to 91.0%, and the average acid production rate reaches more than 3.15g / L·h.

[0150] The data further demonstrates that the method of this invention predicts changes in acid production rate in advance through a soft sensor model, and improves the reliability of feedforward control by combining it with a confidence correction mechanism. Simultaneously, the reinforcement learning strategy enables adaptive adjustment of PID parameters under dynamic operating conditions, making the feedback control more precise and stable. The combination of these two approaches effectively addresses the shortcomings of traditional PID control in fermentation processes characterized by strong nonlinearity and large fluctuations in operating conditions.

[0151] This embodiment demonstrates that the present invention can significantly reduce pH fluctuations, improve alkali utilization efficiency, stabilize the fermentation system, and increase fumaric acid yield and unit acid production rate on a real industrial scale, thereby achieving the goals of automated, intelligent, and energy-saving production control.

[0152] 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. An automatic control method for fumaric acid production, characterized by, The method comprises the following steps: collecting process variables in a fermentation process; constructing a soft-sensing model combining physical modeling and residual neural network based on the process variables, assigning input feature weights using an attention mechanism, obtaining a prediction value of acid production rate per unit time, and outputting a prediction confidence; calculating a feedforward control amount according to the prediction value of acid production rate and a neutralization reaction relationship, and adjusting the weight of the feedforward control amount according to the prediction confidence to obtain a feedforward control output; constructing a state vector based on the process variables, the prediction value of acid production rate, the prediction confidence, and historical control parameters, and establishing a Markov decision process based on the state vector; in the Markov decision process, training a neural network control strategy using a proximal policy optimization algorithm, receiving the state vector, and outputting a proportional coefficient, an integral coefficient, and a differential coefficient; calculating a feedback control amount based on the proportional coefficient, the integral coefficient, and the differential coefficient and a pH deviation, and fusing the feedback control amount and the feedforward control output, applying a rate limit and an amplitude limit to obtain a final control output; when the prediction confidence is lower than a preset threshold or the process variables are abnormal, switching to a feedback priority control mode, and periodically updating the parameters of the soft-sensing model and the neural network control strategy.

2. The automatic control method for fumaric acid production according to claim 1, characterized in that, The collecting process variables in a fermentation process specifically comprises: setting a unified time reference and a sampling period; collecting pH value, dissolved oxygen concentration, stirring power, gas flow rate, alkali dosage rate, and tail gas composition according to the sampling period; timestamping the collected data and aligning them to the time reference to form synchronized process variables.

3. The automatic control method for fumaric acid production according to claim 1, characterized in that, The constructing a soft-sensing model combining physical modeling and residual neural network based on the process variables, assigning input feature weights using an attention mechanism, obtaining a prediction value of acid production rate per unit time, and outputting a prediction confidence specifically comprises: constructing a soft-sensing model combining physical modeling and residual neural network, which takes the process variables as input and outputs a prediction value of acid production rate per unit time and a prediction confidence; establishing a physical modeling formula based on the principle of acid-base neutralization reaction and the mass conservation relationship of buffer solution to estimate the acid production rate per unit time; extracting the process variables and the acid production rate estimate at the current time and the previous K times, combining the process variables at each time with the corresponding acid production rate estimate to form a joint input, and constructing a joint input sequence containing multiple time steps; inputting the joint input sequence into a time attention mechanism to calculate the attention weight of each time step in the current prediction task, performing weighted aggregation on the joint input sequence according to the attention weight to generate a time-aware feature vector at the current time; inputting the time-aware feature vector into a channel attention mechanism to calculate the attention weight of each input feature dimension in the feature expression according to the importance, and performing weighted operation on each channel to generate a structured feature vector; inputting the structured feature vector into a residual neural network to output a prediction value of acid production rate per unit time and a corresponding prediction confidence; training the soft-sensing model using a joint loss function based on uncertainty modeling, which is composed of a prediction error term and a confidence regularization term.

4. The automatic control method for fumaric acid production according to claim 1, characterized by, The feedforward control output is obtained by calculating the feedforward control amount according to the relationship between the acid production rate prediction value and the neutralization reaction, and adjusting the weight of the feedforward control amount according to the prediction confidence, and specifically comprising: According to the difference between the acid production rate prediction value per unit time and the preset target acid production rate, combined with the system volume and the buffer capacity, the theoretical alkali consumption required for neutralization reaction is calculated; According to the current alkali concentration, the theoretical alkali consumption is converted into the target alkali addition flow rate; According to the prediction confidence, the target alkali addition flow rate is weighted to generate a confidence-corrected feedforward control amount, and a feedforward control output is obtained.

5. The automatic control method for fumaric acid production according to claim 1, characterized in that, The state vector is constructed based on the process variable, the acid production rate prediction value, the prediction confidence and the historical control parameter, and the Markov decision process is established based on the state vector, and specifically comprising: Extracting the process variable at the current time, the acid production rate prediction value per unit time, the prediction confidence, and the proportional coefficient, the integral coefficient and the differential coefficient at the last time, and splicing to form a state vector in a predetermined order; All combinations of the state vector within the allowed value range are defined as a state space; The continuous value triple of the proportional coefficient, the integral coefficient and the differential coefficient within the preset upper and lower limits is defined as an action space, and the value boundary of each coefficient is set; The Markov decision process is established based on the state space and the action space, the reward function is set as the weighted combination of the square penalty term of the pH deviation and the alkali addition flow rate change penalty term, and the corresponding relationship of state, action and reward is formed.

6. The automatic control method for fumaric acid production according to claim 1, characterized in that, In the Markov decision process, the neural network control strategy is trained by using the proximal policy optimization algorithm, the state vector is received and the proportional coefficient, the integral coefficient and the differential coefficient are outputted, and specifically comprising: In the Markov decision process, the neural network control strategy is trained by using the proximal policy optimization algorithm, the shear coefficient and the learning rate initial value are set, and the shear coefficient is set as a value monotonically increasing with the prediction confidence; The state vector is inputted into the neural network control strategy, and the proportional coefficient, the integral coefficient and the differential coefficient are outputted, and the proportional coefficient, the integral coefficient and the differential coefficient are projected into the action space boundary; The state vector, the proportional coefficient, the integral coefficient and the differential coefficient, the reward calculated according to the reward function, and the next time state vector are collected in time sequence to form a data sequence; The cumulative discounted reward is calculated according to the data sequence, the discount factor is set as a fixed value and the same discount is used for all time steps; A strategy update target function based on the shear target is constructed, and the prediction confidence is used as the sample weight to weight the strategy update target function; The neural network control strategy is updated by batch random gradient, the learning rate is set as a value monotonically decreasing with the absolute value of the pH deviation and the pH change rate, and an upper limit constraint is applied to the single-step change of the proportional coefficient, the integral coefficient and the differential coefficient; The collection, calculation, construction and update steps are repeatedly executed until the preset convergence condition is met.

7. The automatic control method for fumaric acid production according to claim 1, characterized by, The feedback control amount is calculated based on the proportional coefficient, the integral coefficient and the differential coefficient and the pH deviation, and the feedback control amount and the feedforward control output are fused, the change rate limit and the amplitude limit are applied, and the final control output is obtained, and specifically comprising: The proportional term value is obtained by calculating a proportional term based on a proportional coefficient and a pH deviation; The integral term value is obtained by calculating an integral term based on an integral coefficient and an integral result of the pH deviation in a preset time window, and performing upper limit and lower limit limiting and anti-saturation rollback on the integral term; The derivative term value is obtained by calculating a derivative term based on a derivative coefficient and a rate of change of the pH deviation after performing first-order low-pass filtering on the pH deviation; The feedback control quantity is obtained by adding the proportional term value, the integral term value and the derivative term value; The fusion control quantity is obtained by calculating a feedforward weight according to the prediction confidence, and weighting and fusing the feedforward control output and the feedback control quantity according to the feedforward weight and a complementary weight; The change rate limitation is applied to the fusion control quantity to limit the allowable change amplitude in adjacent sampling periods; The amplitude limitation is applied to the fusion control quantity after the change rate limitation to limit the minimum and maximum dosing flow rates within the boundaries allowed by the device, and the final control output is obtained.

8. The automatic control method for fumaric acid production according to claim 1, characterized by, When the prediction confidence is lower than a preset threshold or the process variable is abnormal, the feedback priority control mode is switched to, and the parameters of the soft measurement model and the neural network control strategy are updated regularly, specifically including: The prediction confidence is compared with the preset threshold, and whether the process variable is out of the preset range or the pH rate is out of the preset upper limit is detected, and when any condition is met, the feedback priority control mode is entered; In the feedback priority control mode, the feedforward weight is set to zero, the proportional coefficient, the integral coefficient and the derivative coefficient of the last moment are kept unchanged, the feedback control quantity is calculated based on the current pH deviation, and the change rate limitation and the amplitude limitation are applied to the feedback control quantity to obtain the final control output; In the feedback priority control mode, the state vector, the final control output, the pH value and the alkali dosing rate are collected within a preset time length, the parameters of the soft measurement model are calibrated, including bias correction and gain correction of the acid production rate prediction value per unit time, and the parameters of the soft measurement model are updated; In the feedback priority control mode, the proximal policy optimization algorithm is used to perform incremental update on the neural network control strategy with a low learning rate; When the prediction confidence is not lower than a preset recovery threshold for a continuous preset sampling period and the process variable is within the preset range, the feedback priority control mode is exited, the feedforward weight is set to a value calculated according to the prediction confidence, the feedforward control output and the feedback control quantity are weighted and fused according to the feedforward weight and the complementary weight, and the change rate limitation and the amplitude limitation are applied to the weighted result to obtain the final control output.

9. An automated control system for fumaric acid production, which executes an automated control method for fumaric acid production according to any one of claims 1 to 8, characterized by, The process data acquisition module is used to acquire process variables in the fermentation process and transmit them to subsequent processing; The acid production rate soft measurement module is used to obtain the acid production rate prediction value per unit time based on the combination of physical modeling and residual neural network with time attention mechanism, and output the prediction confidence; The feedforward control calculation module is used to calculate the feedforward control output according to the acid production rate prediction value and the neutralization reaction relationship, and adjust the weight of the feedforward control output based on the prediction confidence; ​ A state vector construction module is configured to splice the process variables, the acid production rate prediction value, the prediction confidence and the historical control parameters to form a state vector, and establish a Markov decision process based on a state space and an action space; A control strategy training module is configured to train a neural network control strategy in the Markov decision process by using a proximal policy optimization algorithm, input the state vector and output proportional, integral and derivative coefficients; A feedback control calculation module is configured to calculate a feedback control amount based on the proportional, integral and derivative coefficients, and to fuse the feedforward control output and the feedback control amount by weighting, and to obtain a final control output after applying a rate limit and an amplitude limit; A model adaptive module is configured to enter a feedback priority control mode when the prediction confidence is lower than a threshold or the process variables are abnormal, to perform parameter calibration on the soft measurement model, and to perform incremental update on the neural network control strategy.

Images