A neural network-based selenium-rich yeast fermentation process parameter optimization method

Abstract

The application belongs to the technical field of industrial process automatic control, and particularly relates to a selenium-rich yeast fermentation process parameter optimization method based on a neural network, which comprises the following steps: collecting temperature, pH value and dissolved oxygen at each moment in the whole fermentation process to obtain a running parameter sequence; according to the running parameter sequence, calculating oxygen consumption rate and acid production rate at each moment, dividing the absolute values of the oxygen consumption rate and the acid production rate by the respective batch rolling maximum values to obtain oxygen consumption ratio and acid production ratio; multiplying the complement of the oxygen consumption ratio and the complement of the acid production ratio to obtain window period membership; for each historical batch, extracting a time period with a window period membership not lower than a preset threshold to obtain a window time period. The application breaks the limitation of static segmented control, realizes dynamic parameter optimization driven by physiological state, significantly improves the consistency of selenium organic rate between batches, and effectively reduces the product non-compliance rate.

Description

Technical Field

[0001] This invention belongs to the field of industrial process automatic control technology, specifically relating to a method for optimizing parameters in a selenium-enriched yeast fermentation process based on neural networks. Background Technology

[0002] With the continuous development of bioengineering technology, selenium-enriched yeast, as a safe and efficient organic selenium supplement, is increasingly widely used in food fortification and pharmaceuticals. The liquid fermentation production process of selenium-enriched yeast typically includes a lag phase, a logarithmic growth phase, a stationary phase, and apoptosis phase. The core process of yeast converting inorganic selenium into organic selenium is highly concentrated in the window phase transitioning from the logarithmic growth phase to the stationary phase. During this critical fermentation stage, environmental parameters such as temperature, pH, and dissolved oxygen have a strongly coupled physiological influence on selenium uptake and conversion. Therefore, accurately identifying this critical stage and controlling various fermentation parameters to improve the consistency of selenium organication rates between batches and reduce the proportion of substandard products has become an urgent technical need to be addressed in current industrial fermentation production.

[0003] To meet the demands of large-scale production, the most commonly used traditional technology in the industry is a static segmented control strategy. In existing fermentation production, operators typically rely on historical production experience to pre-set fixed time points in the control system, and then set fixed target values ​​for temperature, pH, and dissolved oxygen at different time intervals. For example, after a specific number of fermentation hours are reached, the system automatically switches the environmental parameter settings according to a preset schedule, thereby guiding the yeast to complete the growth and metabolic processes of the entire fermentation cycle. This control method, based on a fixed time axis and experienced thresholds, is currently the fundamental technical path for achieving industrial-scale fermentation of selenium-enriched yeast.

[0004] However, the aforementioned parameter control method based on static time nodes has significant technical flaws. In actual industrial fermentation, due to slight deviations in the inoculum quantity between different batches and differences in seed culture activity, the actual time when yeast enters the selenium accumulation window often experiences unpredictable stage drift. Traditional static control schemes cannot perceive the real-time physiological metabolic state of yeast in the fermenter and cannot coordinately adjust the three strongly coupled parameters based on the real dynamic changes in oxygen consumption and acid production rates. This rigid control method, which separates the real physiological stage from the parameter control logic, often leads to mismatches in environmental parameters during the most critical selenium accumulation period, resulting in significant fluctuations in selenium organication rate between batches and failing to achieve data-driven real-time closed-loop optimization. Summary of the Invention

[0005] This invention provides a neural network-based method for optimizing parameters in the fermentation process of selenium-enriched yeast, thereby solving the technical problem that the traditional static segmented control scheme cannot perceive the real-time physiological state of fermentation, leading to parameter mismatch in key fermentation stages and significant fluctuations in selenium organication rate between batches.

[0006] This invention provides a method for optimizing parameters in a selenium-enriched yeast fermentation process based on neural networks, comprising the following steps: Temperature, pH, and dissolved oxygen were collected at various points throughout the fermentation process to obtain an operational parameter sequence. Based on the operational parameter sequence, the oxygen consumption rate and acid production rate at each point were calculated. The absolute values ​​of the oxygen consumption rate and acid production rate were divided by their respective batch rolling maximum values ​​to obtain the oxygen consumption ratio and acid production ratio. The complement of the oxygen consumption ratio and the complement of the acid production ratio were multiplied to obtain the window period membership degree. For each historical batch, extract the time period with a membership degree no lower than a preset threshold to obtain the window period; calculate and normalize the time mean values ​​of temperature, pH value and dissolved oxygen within the window period to obtain the batch feature vector; add one to the window period mean value to obtain the batch loss weight; use the batch feature vector and the measured selenium organicization rate as training samples, and train the BP neural network by differential weighting with the batch loss weight to obtain the network weight. In the production batch, the membership degree of the window period is calculated in real time, and the dynamic search step size is determined according to the rule that is negatively correlated with the membership degree of the window period. Within the parameter constraint range, candidate parameter combinations are enumerated according to the dynamic search step size. After normalizing each candidate parameter combination, it is input into the BP neural network. The candidate parameter combination with the largest predicted selenium organicization rate is selected to obtain the optimal parameter group. The optimal parameter set is sent to the fermenter control system.

[0007] Its effect is as follows: by multiplying the complement of the oxygen consumption ratio with the complement of the acid production ratio to obtain the membership degree of the window period, and applying it to the weighting of model training and the control of the step size of online optimization, the physical change of metabolic rate is transformed into a precision adjustment signal, eliminating the stage drift interference caused by the inoculum deviation, achieving the matching of computing power with the key fermentation stage, and improving the effectiveness of parameter decision-making.

[0008] Furthermore, the calculation of oxygen consumption rate and acid production rate at each time point includes: dividing the difference between dissolved oxygen at the current time and dissolved oxygen at the previous sampling time by the sampling interval to obtain the oxygen consumption rate; and dividing the difference between pH value at the current time and pH value at the previous sampling time by the sampling interval to obtain the acid production rate.

[0009] Furthermore, the absolute values ​​of oxygen consumption rate and acid production rate are divided by their respective batch rolling maximum values ​​to obtain the oxygen consumption ratio and acid production ratio, including: dividing the absolute value of oxygen consumption rate by the larger of the maximum oxygen consumption value and the division-by-zero protection constant to obtain the oxygen consumption ratio; dividing the absolute value of acid production rate by the larger of the maximum acid production value and the division-by-zero protection constant to obtain the acid production ratio; wherein, the maximum oxygen consumption value is the rolling maximum absolute value of the oxygen consumption rate of the current batch up to the current time, the maximum acid production value is the rolling maximum absolute value of the acid production rate of the current batch up to the current time, and the value of the division-by-zero protection constant is less than the historical minimum recorded value of the absolute values ​​of oxygen consumption rate and acid production rate under normal operating conditions.

[0010] Its effect is that by dividing the absolute values ​​of oxygen consumption and acid production rates by the larger of the batch rolling maximum value and the division-by-zero protection constant, the dimensional differences between different physicochemical indicators are effectively eliminated, and a denominator fallback protection is provided for extreme abnormal operating conditions, ensuring the numerical stability of proportional calculations and enhancing the robustness of the system in harsh industrial control environments.

[0011] Furthermore, before calculating the oxygen consumption rate and acid production rate at each time point, the process includes: initializing the maximum oxygen consumption value of the current batch to the prior maximum oxygen consumption value obtained from historical batch statistics; initializing the maximum acid production value of the current batch to the prior maximum acid production value obtained from historical batch statistics; during fermentation, when the absolute value of the oxygen consumption rate at the current time point is greater than the maximum oxygen consumption value, updating the maximum oxygen consumption value to the absolute value of the oxygen consumption rate; and when the absolute value of the acid production rate at the current time point is greater than the maximum acid production value, updating the maximum acid production value to the absolute value of the acid production rate.

[0012] Furthermore, training the BP neural network to obtain network weights involves: inputting the batch feature vectors of each historical batch into the BP neural network to obtain the predicted selenium organication rate of each historical batch; multiplying the square of the difference between the predicted selenium organication rate and the measured selenium organication rate of each historical batch by the corresponding batch loss weight, summing the results, and then dividing by the sum of the batch loss weights of all historical batches to obtain the weighted loss value; and iteratively updating the parameters of the BP neural network through backpropagation with the goal of minimizing the weighted loss value to obtain the network weights.

[0013] Its effect is as follows: multiply the square of the prediction error by the batch loss weight, sum them, and then divide by the total weight to obtain the weighted loss value for iteration. This method, while maintaining the stability of the magnitude of the loss value, enables high-quality batches with significant physiological characteristics to generate a larger driving gradient, guiding the model to converge toward the regularity of the key fermentation period, and significantly improving the prediction accuracy of the key stage.

[0014] Furthermore, determining the dynamic search step size according to the rule of negative correlation with the window period membership degree includes: multiplying the temperature base step size by 1 and the difference between the product of the window period membership degree and the step size contraction coefficient to obtain the temperature step size; multiplying the pH base step size by 1 and the difference between the product of the window period membership degree and the step size contraction coefficient to obtain the pH step size; multiplying the dissolved oxygen base step size by 1 and the difference between the product of the window period membership degree and the step size contraction coefficient to obtain the dissolved oxygen step size; the temperature step size, pH step size, and dissolved oxygen step size together constitute the dynamic search step size, which is negatively correlated with the window period membership degree, that is, the larger the window period membership degree, the smaller the dynamic search step size.

[0015] Furthermore, the normalization of each candidate parameter combination before inputting it into the BP neural network includes: for each candidate parameter combination, the product of the window accumulation points and the mean of the accumulated parameters, plus the product of the remaining estimated points and the candidate parameter combination, is divided by the sum of the window accumulation points and the remaining estimated points to obtain the expected window mean; wherein, the window accumulation points are the cumulative number of sampling points that the current batch has entered the window period, the mean of the accumulated parameters is the average accumulated time of temperature, pH value and dissolved oxygen within the window period of the current batch, and the remaining estimated points are the difference between the average number of sampling points and the window accumulation points of the historical batch window periods; the expected window mean is normalized before being input into the BP neural network.

[0016] Its effect is as follows: by combining the accumulated points of the window, the remaining estimated points, and the candidate parameters to calculate the expected window mean and inputting it into the neural network, this method transforms the current transient control operation into a prediction of the average level of the entire window period, connecting local adjustment with overall quality, avoiding short-sighted parameter adjustment behavior, and ensuring the global optimality of online optimization decision-making.

[0017] Furthermore, the BP neural network was trained, resulting in a three-layer structure in the network weights, including: an input layer with 3 neurons, each corresponding to the normalized mean temperature, mean pH, and mean dissolved oxygen values ​​in the batch feature vector; a single hidden layer using the Sigmoid function as the activation function; and an output layer with 1 neuron, which outputs the predicted selenium organication rate.

[0018] Furthermore, training the BP neural network to obtain network weights also includes: dividing the historical batch dataset into a training set and a validation set according to a preset ratio; when the weighted loss value on the validation set no longer decreases within a preset number of consecutive rounds, terminating the training, and taking the network parameters at the moment when the weighted loss value is minimum as the network weights.

[0019] Furthermore, after the optimal parameter set is sent to the fermenter control system, the process also includes: after the fermentation of the current batch is completed, offline detection of the measured selenium organication rate of the current batch, calculation of the batch feature vector of the current batch, and appending the batch feature vector and the measured selenium organication rate to the historical batch dataset; when the number of appended batches reaches a preset number, the process of training the BP neural network by differential weighting with batch loss weights and obtaining the network weights is re-executed using the updated historical batch dataset, and the network weights are updated.

[0020] Its effects are as follows: by offline detection and actual measurement of selenium organication rate, after accumulating a preset number of batches, the model and benchmark parameters are triggered to undergo rolling retraining, giving the system the ability to learn and adapt to the drift of the production environment, effectively overcoming the model degradation caused by equipment aging or fluctuations in raw materials, and maintaining the long-term stability of fermentation control level and product quality.

[0021] The beneficial effects are as follows: This invention accurately solves the membership degree of the window period by calculating the oxygen consumption ratio and acid production ratio, and constructs batch loss weights accordingly to achieve differentiated weighted training of the BP neural network. In actual production, the parameter search step size is dynamically adjusted and optimization is carried out based on the real-time membership degree of the fermentation window period, which breaks through the technical limitations of traditional static segmented control and realizes dynamic regulation of fermentation parameters driven by the physiological metabolic state of yeast. This significantly improves the consistency of selenium organic content between batches of selenium-enriched yeast production and effectively reduces the product defect rate. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for optimizing parameters in the fermentation process of selenium-enriched yeast based on neural networks.

[0023] Figure 2 This is a comparison chart showing the training convergence process of the standard error and weighted loss value of a BP neural network.

[0024] Figure 3 This is a batch comparison chart of selenium organication rate in selenium-enriched yeast under traditional static segmented control and neural network dynamic optimization control. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] An embodiment of the neural network-based parameter optimization method for selenium-enriched yeast fermentation provided by this invention: like Figure 1As shown, the parameter optimization method for selenium-enriched yeast fermentation process based on neural networks includes the following steps: S101: Collect temperature, pH value, and dissolved oxygen at various times throughout the fermentation process to obtain the operating parameter sequence; based on the operating parameter sequence, calculate the oxygen consumption rate and acid production rate at each time point, divide the absolute values ​​of the oxygen consumption rate and acid production rate by their respective batch rolling maximum values ​​to obtain the oxygen consumption ratio and acid production ratio; multiply the complement of the oxygen consumption ratio and the complement of the acid production ratio to obtain the window period membership degree.

[0027] In one embodiment, this solution is preferably applied to an industrial selenium-enriched yeast liquid fermentation production scenario using sodium selenite as the selenium source and *Saccharomyces cerevisiae* as the strain. For example, in a fermentation tank with a volume of 50 liters to 5000 liters, a temperature sensor, pH electrode, and dissolved oxygen probe are already installed inside, and an industrial control bus is connected externally. The entire fermentation process of the selenium-enriched yeast typically lasts 16 to 24 hours. After fermentation begins, based on the industrial control bus, the values ​​of the above three sensors are continuously read at fixed sampling intervals, thereby collecting the temperature, pH value, and dissolved oxygen at various moments throughout the fermentation process to obtain the operating parameter sequence.

[0028] To avoid the failure of subsequent feature calculations due to abnormal fluctuations in sensors when data is scarce in the early stages of fermentation, a reliable comparison benchmark needs to be established before calculating the real-time rate of change. The maximum oxygen consumption of the current batch is initialized to the prior maximum oxygen consumption obtained from historical batch statistics; the maximum acid production of the current batch is also initialized to the prior maximum acid production obtained from historical batch statistics. During fermentation, when the absolute value of the current oxygen consumption rate is greater than the maximum oxygen consumption, the maximum oxygen consumption is updated to the absolute value of the oxygen consumption rate; when the absolute value of the current acid production rate is greater than the maximum acid production, the maximum acid production is updated to the absolute value of the acid production rate. It should be noted that the aforementioned prior maximum oxygen consumption and acid production values ​​are obtained by collecting data from at least 50 successful fermentation batches and extracting the maximum global absolute value as a benchmark, thereby ensuring that the physiological stage identification throughout the entire batch fermentation process has strong noise resistance from the outset.

[0029] After establishing the baseline for rolling updates, to quantify the real-time metabolic activity of yeast, it is necessary to calculate the current oxygen consumption and pH changes. The oxygen consumption rate is obtained by dividing the difference between the current dissolved oxygen level and the previous sampling time by the sampling interval; the acid production rate is obtained by dividing the difference between the current pH level and the previous sampling time by the sampling interval. Understandably, the standard first-order backward difference formula in this field is directly used for calculation, which accurately reflects the current transient metabolic characteristics without the need for complex smoothing filters. For example, the sampling interval is set to 0.5 minutes to achieve the best balance between capturing effective dynamic physiological changes and avoiding high-frequency signal noise.

[0030] Since oxygen consumption rate and acid production rate belong to completely different physical dimensions and have huge differences in magnitude, they cannot be directly calculated together. In order to eliminate the difference in dimensions and prevent calculation collapse under extreme abnormal conditions, the absolute value of oxygen consumption rate is divided by the larger of the maximum oxygen consumption rate and the division-by-zero protection constant to obtain the oxygen consumption ratio; the absolute value of acid production rate is divided by the larger of the maximum acid production rate and the division-by-zero protection constant to obtain the acid production ratio. Among them, the maximum oxygen consumption rate is the rolling maximum absolute value of oxygen consumption rate of the current batch up to the current time, the maximum acid production rate is the rolling maximum absolute value of acid production rate of the current batch up to the current time, and the value of the division-by-zero protection constant is less than the historical minimum recorded value of the absolute value of oxygen consumption rate and acid production rate under normal operating conditions.

[0031] The ratio of oxygen consumption to acid production satisfies the following relationship:

[0032]

[0033] In the formula, This represents the oxygen consumption ratio at the current moment; This indicates the acid production ratio at the current moment; This indicates the oxygen consumption rate at the current moment; This indicates the acid production rate at the current moment; This indicates the maximum oxygen consumption for the current batch; This indicates the maximum acid production value for the current batch; This represents the division-to-zero protection constant, preferably with a specific value of one ten-thousandth.

[0034] Understandably, introducing a very small division-to-zero protection constant does not affect the proportional calculation results under normal operating conditions. Its physical significance lies only in serving as a fallback denominator in abnormal scenarios where extremely high ventilation rates cause almost no change in dissolved oxygen, ensuring the numerical stability of the program. The oxygen consumption ratio and acid production ratio calculated in this way are dimensionless data. The closer their values ​​are to one, the closer the current metabolic rate is to the highest level of this batch, indicating that the yeast is in a high-speed logarithmic growth phase.

[0035] To accurately determine whether yeast has entered the most critical transitional stage where organic selenium can be accumulated, it is necessary to find a reliable indicator that both oxygen consumption and acid production slow down. The complement of the oxygen consumption ratio and the complement of the acid production ratio are multiplied to obtain the membership degree of the window period.

[0036] The membership degree during the window period satisfies the following relation:

[0037] In the formula, This represents the membership degree of the current window period; This represents the oxygen consumption ratio at the current moment; This represents the acid production ratio at the current moment.

[0038] It should be noted that the complement here is 1 minus the corresponding ratio. The above formula cleverly utilizes the algebraic product rule to achieve a strict judgment logic for the simultaneous occurrence of two fuzzy events. When both the oxygen consumption ratio and the acid production ratio approach zero, their complements both approach 1, and the membership degree of the window period obtained by multiplying them approaches one, indicating that the yeast has deeply entered the selenium accumulation stage; if only a single indicator decreases, its product result will still be suppressed to a lower level by the high value of the other term.

[0039] By calculating in real time using pure sensor data and eliminating differences in physical dimensions, the physiological stage transition state of the current fermentation can be accurately and smoothly quantified without relying on any fixed time point. This completely eliminates the stage drift interference caused by inoculum deviation and seed liquid vigor differences in engineering practice, providing an extremely reliable state benchmark for subsequent fully automated parameter optimization.

[0040] S102, for each historical batch, extract the time period with a membership degree not lower than a preset threshold to obtain the window period; calculate and normalize the time mean values ​​of temperature, pH value and dissolved oxygen within the window period to obtain the batch feature vector; add one to the window period membership mean value within the window period to obtain the batch loss weight; use the batch feature vector and the measured selenium organication rate as training samples, and train the BP neural network by differential weighting with the batch loss weight to obtain the network weight.

[0041] In one embodiment, S101 completes the precise quantification of the real-time physiological state throughout the fermentation process. To further explore the potential patterns between this physiological state and the final fermentation quality, this step will extract structured features from historical fermentation data during the offline training phase. Specifically, data from at least 50 historical fermentation batches are collected. For each historical batch, its corresponding window period membership sequence is traversed in chronological order. For each historical batch, the time period with a window period membership degree not lower than a preset threshold is extracted to obtain the window time period. For example, the preset threshold is preferably set to 0.5, that is, data segments that are highly likely to belong to the selenium accumulation physiological stage are extracted.

[0042] After identifying the most critical time window, to eliminate the inconsistency in data dimensions caused by different batch fermentation durations, it is necessary to aggregate the time dimension. This involves calculating and normalizing the temporal mean values ​​of temperature, pH, and dissolved oxygen within the window to obtain the batch feature vector. Understandably, the well-known standard maximum-minimum normalization method is directly adopted here, mapping the calculated temporal mean values ​​to the interval between 0 and 1. This aggregation method explicitly defines the network input as the average parameter control level during the window period, which has a clear physical causal relationship with the output label, i.e., the final measured selenium organication rate of the batch.

[0043] Furthermore, since the physiological characteristics of different batches differ in degree when entering the window period, in order to quantify this difference in sample quality to guide subsequent model training, the batch loss weight is obtained by adding 1 to the mean membership degree of the window period within the window period.

[0044] Batch loss weights satisfy the following relationship:

[0045] In the formula, Indicates the first Batch loss weights for each historical batch; Indicates the first The average membership degree of each historical batch within the window period.

[0046] Understandably, this formula linearly maps the mean of continuous physiological characteristics to loss weights. The more significant the physiological characteristics of a batch, i.e., the more obvious the simultaneous slowdown in oxygen consumption and acid production, the closer the corresponding mean is to 1, and the closer the weight obtained for that batch is to 2; conversely, it is closer to 1. This endogenous weight calculation method, which does not require manually preset parameters, not only preserves the basic learning signals of all samples but also avoids gradient explosion caused by excessively large weights, thus possessing extremely high numerical stability.

[0047] To establish the mapping relationship between the window period parameter control level and the final fermentation quality, a neural network model with a specific structure needs to be constructed. The BP neural network is trained, resulting in a three-layer structure in the network weights: an input layer with three neurons, corresponding to the normalized mean temperature, mean pH, and mean dissolved oxygen values ​​in the batch feature vector; a single hidden layer using the Sigmoid function as the activation function; and an output layer with one neuron, outputting the predicted selenium organication rate.

[0048] Understandably, given that fermentation time has been implicitly eliminated by time averaging, the input layer only needs to retain the normalized values ​​of the three core physical quantities. The number of neurons in a single hidden layer is determined by an empirical formula, with seven being the preferred configuration. The output layer employs a linear identity mapping to accommodate regression prediction tasks with continuous values. This network structure, while ensuring sufficient nonlinear fitting capability, keeps the model complexity at an extremely low level, making it highly suitable for the computational resource constraints of industrial control environments.

[0049] After determining the topology of the neural network, to make the network pay more attention to high-quality batch data with significant physiological characteristics during training, thereby improving the fitting accuracy of key fermentation stages, the standard error function needs to be modified. The batch feature vectors of each historical batch are input into the BP neural network to obtain the predicted selenium organication rate for each historical batch. The squares of the differences between the predicted and measured selenium organication rates of each historical batch are multiplied by their respective batch loss weights, summed, and then divided by the sum of the batch loss weights of all historical batches to obtain the weighted loss value. With the goal of minimizing the weighted loss value, the parameters of the BP neural network are iteratively updated through backpropagation to obtain the network weights.

[0050] The weighted loss values ​​satisfy the following relationship:

[0051] In the formula, Indicates the weighted loss value; This indicates the total number of historical batches used in training; Indicates the first Batch loss weights for each historical batch; Indicates the first Predicted selenium organication rate for each historical batch; Indicates the first The measured selenium organication rate of each historical batch.

[0052] It should be noted that this scheme ensures that the magnitude of the overall loss value does not jump by introducing a weight term into the standard mean squared error function and replacing the denominator with a normalized sum of all weights. When backpropagating using the well-known stochastic gradient descent update rule, batches with more significant average metabolic states during the window period will generate larger driving gradients, forcing network weights to converge towards the underlying patterns of such high-quality data, greatly improving the model's predictive purity during the core fermentation period.

[0053] To prevent overfitting of the neural network to limited historical fermentation data during iterative updates and to ensure good generalization ability in actual online prediction, a scientific training termination mechanism must be introduced. The historical batch dataset is divided into training and validation sets according to a preset ratio. Training is terminated when the weighted loss value on the validation set no longer decreases within a preset number of consecutive rounds, and the network parameters at the moment of minimum weighted loss value are taken as the network weights.

[0054] The preferred preset ratio is 7:2:1, corresponding to the division of the training, validation, and test sets, respectively. The preferred number of consecutive preset rounds is 20 rounds. This early stopping mechanism effectively avoids the continuous consumption of ineffective computing power, ensuring that the final output network weights have optimal generalization performance.

[0055] This step overcomes the engineering challenge of traditional models weakening key features by precisely cropping and extracting features from historical sensor data and combining it with an innovative data-driven weighted learning mechanism. It outputs a set of highly customized network weights that are fully focused on the selenium organication window period, laying a solid model foundation for precise intervention in the subsequent production process.

[0056] S103: Calculate the membership degree of the window period in real time within the production batch, and determine the dynamic search step size according to the rule that is negatively correlated with the membership degree of the window period; within the parameter constraint range, enumerate candidate parameter combinations according to the dynamic search step size, normalize each candidate parameter combination and input it into the BP neural network, select the candidate parameter combination with the largest predicted selenium organicization rate, and obtain the optimal parameter set.

[0057] In one embodiment, after obtaining the network weights during the offline training phase, online prospective optimization of parameters is required based on real-time physiological states in the actual fermentation production batch. As fermentation progresses, the window period membership degree is calculated in real time within the production batch, and the dynamic search step size is determined according to a rule negatively correlated with the window period membership degree.

[0058] Traditional fixed-step grid search employs the same search precision across all fermentation stages, while the growth phase is not a critical optimization period, leading to a significant waste of computational resources. Since the window period membership degree accurately characterizes the depth at which yeast enters the critical selenium accumulation stage, it can be used as an adjustment signal for search precision, automatically refining the search granularity at the most critical stages. Therefore, the temperature step size is obtained by multiplying the base temperature step size by 1 and the difference between the product of the window period membership degree and the step size contraction coefficient; the pH step size is obtained by multiplying the base pH step size by 1 and the difference between the product of the window period membership degree and the step size contraction coefficient; the dissolved oxygen step size is obtained by multiplying the base dissolved oxygen step size by 1 and the difference between the product of the window period membership degree and the step size contraction coefficient. The temperature step size, pH step size, and dissolved oxygen step size together constitute the dynamic search step size. The dynamic search step size is negatively correlated with the window period membership degree; that is, the larger the window period membership degree, the smaller the dynamic search step size.

[0059] Understandably, this dynamic step size calculation logic directly transforms physiological state indicators into precision control parameters for online inference.

[0060] The temperature step size, pH step size, and dissolved oxygen step size satisfy the following relationship:

[0061]

[0062]

[0063] In the formula, This indicates the temperature step size at the current moment; This indicates the pH step size at the current moment; This indicates the dissolved oxygen step size at the current moment; This represents the basic temperature step size, preferably 0.5 degrees Celsius. This represents the basic pH step size, which is preferably 0.2. This represents the basic step size for dissolved oxygen, preferably 0.05; This represents the step size contraction coefficient, which is preferably taken as 0.5; This represents the membership degree of the current window period.

[0064] It should be noted that when fermentation is in the normal growth stage, the membership degree of the window period approaches 0, and each search step size is maintained at a coarse basic accuracy, which greatly reduces the enumeration calculation amount in the normal stage; when the depth enters the critical selenium accumulation stage, the membership degree of the window period approaches 1, and each search step size is shrunk to the finest accuracy, thereby outputting extremely fine control instructions, realizing the precise matching between the computing power allocation of the industrial control computer and the actual process optimization needs.

[0065] After determining the current search accuracy, candidate parameter combinations are enumerated within the parameter constraints using a dynamic search step size. To ensure strict alignment between online inference prediction and offline model training in terms of data dimension and physical meaning—that is, to transform the transient setpoint at the current moment into its impact on the average level of the entire window period—each candidate parameter combination needs to be normalized before being input into the BP neural network. Specifically, for each candidate parameter combination, the product of the window accumulation points and the mean of accumulated parameters, plus the product of the remaining estimated points and the candidate parameter combination, is divided by the sum of the window accumulation points and the remaining estimated points to obtain the expected window mean. Here, the window accumulation points are the cumulative number of sampling points that have entered the window period in the current batch, the mean of accumulated parameters is the average accumulated time of temperature, pH value, and dissolved oxygen within the window period of the current batch, and the remaining estimated points are the difference between the average number of sampling points in the window period of historical batches and the window accumulation points. The expected window mean is then normalized before being input into the BP neural network.

[0066] Understandably, this forecast of the mean over the entire window period is a key bridge connecting transient control operations with overall fermentation quality.

[0067] The expected window mean satisfies the following relationship:

[0068] In the formula, This represents the expected window mean calculated at the current moment; This indicates the accumulated points in the window; This represents the average accumulated parameter value at the current moment. This represents the estimated number of remaining points derived from historical prior data. This indicates the currently enumerated combination of candidate parameters, including specific temperature, pH, and dissolved oxygen values.

[0069] It should be noted that the physical essence of the above formula lies in the following: assuming that the fermentation environment parameters are maintained at the level of the candidate parameter combination at the current moment and throughout the remaining window period, the overall average control state of this batch at the end of the window period will reach the expected window average. This expected average is then normalized using standard maximum and minimum values ​​and input into a BP neural network for forward propagation inference. The resulting network output is the measured selenium organication rate expected to be achievable during the remaining period under the control of this parameter combination.

[0070] After enumerating and predicting all legal values, the candidate parameter combination that predicts the highest selenium organication rate is selected to obtain the optimal parameter set. For example, considering the safety boundaries of engineering practice, the parameter constraints are preferably set as follows: temperature between 28°C and 32°C, pH between 4.5 and 5.5, and dissolved oxygen between 20% and 40%. Within this range, combined with a dynamically shrinking, extremely fine search granularity, the thousands of combinations generated, when performing simple matrix forward propagation calculations, have a total time consumption far less than the sampling interval of the control system, fully meeting the stringent constraints of real-time equipment response in industrial settings.

[0071] This step seamlessly transforms physiological state monitoring into an automatic scheduling basis for dynamic optimization computing power. It not only ensures the strict continuity of the physical meaning of the core data stream, the feature mean, during the feedforward inference stage, but also precisely locks the combination of fermentation parameters that can make the final product quality globally optimal in a massive control space with extremely low computational overhead.

[0072] S104 sends the optimal parameter set to the fermenter control system.

[0073] In one embodiment, the real-time forward-looking optimization process described above has precisely identified the optimal parameter combination for the final fermentation quality at the current moment. Specifically, the optimal parameter set is sent to the fermenter control system. This optimal parameter set includes the target values ​​for temperature, pH, and dissolved oxygen at the current moment. These target values ​​are transmitted to the underlying programmable logic controller (PLC) or distributed control system via an industrial control bus. Upon receiving the target values, the underlying proportional-integral-derivative (PID) control loop adjusts the temperature by regulating the jacket cooling water flow rate; adjusts the pH by controlling the peristaltic pump replenishment of sodium hydroxide solution at a rate of two moles per liter; and controls the dissolved oxygen by adjusting the airflow rate of the air meter and the stirring speed. This architecture, which completely decouples the upper-level neural network-based optimization decision-making from the lower-level basic control, ensures that this solution requires no intervention in the lower-level control logic and is highly compatible with existing industrial fermentation production equipment.

[0074] Fermentation production is a long-cycle industrial process. As equipment ages, sensor sensitivity decreases, or there are slight fluctuations between batches of raw materials, fixed prediction models will inevitably experience a decline in generalization ability. To enable the model to continuously evolve with the accumulation of real production data, after the optimal parameter set is sent to the fermenter control system, the following steps are also included: after the fermentation of the current batch is completed, offline detection of the measured selenium organication rate of the current batch is performed, the batch feature vector of the current batch is calculated, and the batch feature vector and the measured selenium organication rate are added to the historical batch dataset; when the cumulative number of added batches reaches a preset number, the steps of training the BP neural network with differentiated weighting based on batch loss weights are re-executed using the updated historical batch dataset to obtain the network weights, and the network weights are updated.

[0075] It should be noted that the preset quantity is preferably ten batches. When the number of newly added fermentation batches reaches this threshold, an incremental retraining mechanism is automatically triggered. During this process, not only are the network weights updated based on the latest historical batch dataset, but the aforementioned baseline parameters such as the prior maximum values ​​of oxygen consumption and acid production are also updated simultaneously. This rolling optimization based on the latest real production closed-loop data is equivalent to giving the fermentation parameter optimization logic the ability to learn on its own and adapt to changes in the production environment, thereby preventing long-term changes in process conditions from causing model prediction distortion.

[0076] This step not only enables the precise and stable execution of real-time optimization commands at the industrial equipment level, but also completely establishes a closed-loop data flow throughout the entire lifecycle, from online control execution and offline testing to core model iteration. This ensures that the fermentation process parameter optimization method can always adapt to the ever-changing actual production environment, and has played a beneficial role in maintaining the long-term stable achievement of high product quality in engineering practice.

[0077] like Figure 2 As shown in the figure, the convergence process of the BP neural network trained with standard error and the weighted loss value of the present invention is compared. By introducing batch loss weights to differentiate the weights of high-quality historical batches, the model of the present invention can generate a larger driving gradient and converge to the underlying rules of the key fermentation period in fewer iterations, ultimately achieving a significantly lower prediction error than traditional standard training.

[0078] like Figure 3 As shown in the figure, the comparison between the measured selenium organication rate of traditional static segmented control and the dynamic optimization control of the present invention in 50 production batches is shown. As can be seen from the figure, the traditional method cannot detect physiological stage drift, resulting in drastic fluctuations in the selenium organication rate between batches; while after adopting the method of the present invention, the selenium organication rate is kept at a stable high level, which significantly improves the consistency between batches and effectively reduces the product non-compliance rate.

[0079] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing parameters in a selenium-enriched yeast fermentation process based on neural networks, characterized in that, Includes the following steps: Temperature, pH, and dissolved oxygen were collected at various points throughout the fermentation process to obtain an operational parameter sequence. Based on the operational parameter sequence, the oxygen consumption rate and acid production rate at each point were calculated. The absolute values ​​of the oxygen consumption rate and acid production rate were divided by their respective batch rolling maximum values ​​to obtain the oxygen consumption ratio and acid production ratio. The complement of the oxygen consumption ratio and the complement of the acid production ratio were multiplied to obtain the window period membership degree. For each historical batch, extract the time periods with a window period membership degree not lower than a preset threshold to obtain the window period; The time mean values ​​of temperature, pH, and dissolved oxygen within the window period are calculated and normalized to obtain the batch feature vector; the batch loss weight is obtained by adding one to the mean membership degree of the window period within the window period; the batch feature vector and the measured selenium organication rate are used as training samples, and the BP neural network is trained by differential weighting of the batch loss weight to obtain the network weight. In the production batch, the membership degree of the window period is calculated in real time, and the dynamic search step size is determined according to the rule that is negatively correlated with the membership degree of the window period. Within the parameter constraint range, candidate parameter combinations are enumerated according to the dynamic search step size. After normalizing each candidate parameter combination, it is input into the BP neural network. The candidate parameter combination with the largest predicted selenium organicization rate is selected to obtain the optimal parameter group. The optimal parameter set is sent to the fermenter control system.

2. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 1, characterized in that, The calculation of oxygen consumption rate and acid production rate at each time point includes: dividing the difference between dissolved oxygen at the current time and dissolved oxygen at the previous sampling time by the sampling interval to obtain the oxygen consumption rate; and dividing the difference between pH value at the current time and pH value at the previous sampling time by the sampling interval to obtain the acid production rate.

3. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 1, characterized in that, The oxygen consumption rate and acid production rate are obtained by dividing the absolute values ​​of their respective batch rolling maximum values ​​by the following methods: dividing the absolute value of the oxygen consumption rate by the larger of the maximum oxygen consumption rate and the zero-division protection constant; and dividing the absolute value of the acid production rate by the larger of the maximum acid production rate and the zero-division protection constant. The maximum oxygen consumption rate is the rolling maximum absolute value of the oxygen consumption rate for the current batch up to the current moment, and the maximum acid production rate is the rolling maximum absolute value of the acid production rate for the current batch up to the current moment. The zero-division protection constant is less than the historical minimum recorded value of the absolute values ​​of the oxygen consumption rate and acid production rate under normal operating conditions.

4. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 3, characterized in that, Before calculating the oxygen consumption rate and acid production rate at each time point, the following steps are also included: initializing the maximum oxygen consumption value of the current batch to the prior maximum oxygen consumption value obtained from historical batch statistics; initializing the maximum acid production value of the current batch to the prior maximum acid production value obtained from historical batch statistics; during fermentation, when the absolute value of the oxygen consumption rate at the current time point is greater than the maximum oxygen consumption value, updating the maximum oxygen consumption value to the absolute value of the oxygen consumption rate; when the absolute value of the acid production rate at the current time point is greater than the maximum acid production value, updating the maximum acid production value to the absolute value of the acid production rate.

5. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 1, characterized in that, Training the BP neural network to obtain network weights involves: inputting the batch feature vectors of each historical batch into the BP neural network to obtain the predicted selenium organication rate of each historical batch; multiplying the square of the difference between the predicted selenium organication rate and the measured selenium organication rate of each historical batch by the corresponding batch loss weight, summing the results, and then dividing by the sum of the batch loss weights of all historical batches to obtain the weighted loss value; and iteratively updating the parameters of the BP neural network through backpropagation with the goal of minimizing the weighted loss value to obtain the network weights.

6. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 1, characterized in that, The dynamic search step size is determined according to the rule that it is negatively correlated with the membership degree of the window period. This includes: multiplying the temperature base step size by 1 and the difference between the product of the membership degree of the window period and the step size contraction coefficient to obtain the temperature step size; multiplying the pH base step size by 1 and the difference between the product of the membership degree of the window period and the step size contraction coefficient to obtain the pH step size; and multiplying the dissolved oxygen base step size by 1 and the difference between the product of the membership degree of the window period and the step size contraction coefficient to obtain the dissolved oxygen step size. The temperature step size, pH step size, and dissolved oxygen step size together constitute the dynamic search step size. The dynamic search step size is negatively correlated with the membership degree of the window period, that is, the larger the membership degree of the window period, the smaller the dynamic search step size.

7. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 1, characterized in that, The normalization of each candidate parameter combination before inputting it into the BP neural network includes: for each candidate parameter combination, the product of the window accumulation points and the mean of the accumulated parameters, plus the product of the remaining estimated points and the candidate parameter combination, is divided by the sum of the window accumulation points and the remaining estimated points to obtain the expected window mean; where the window accumulation points are the cumulative number of sampling points in the current batch that have entered the window period, the mean of the accumulated parameters are the average accumulated time of temperature, pH value and dissolved oxygen in the current batch window period, and the remaining estimated points are the difference between the average number of sampling points in the historical batch window period and the window accumulation points; the expected window mean is normalized before being input into the BP neural network.

8. The method for optimizing parameters in the selenium-enriched yeast fermentation process based on neural networks according to claim 1, characterized in that, The BP neural network was trained, and the network weights were obtained as a three-layer structure, including: an input layer with 3 neurons, the 3 neurons corresponding to the normalized mean temperature, mean pH and mean dissolved oxygen in the batch feature vector respectively; a single hidden layer using the Sigmoid function as the activation function; and an output layer with 1 neuron, which outputs the predicted selenium organication rate.

9. The method for optimizing parameters in a selenium-enriched yeast fermentation process based on a neural network according to claim 5, characterized in that, Training a backpropagation (BP) neural network to obtain network weights also includes: dividing the historical batch dataset into a training set and a validation set according to a preset ratio; terminating training when the weighted loss value on the validation set no longer decreases within a preset number of consecutive rounds, and taking the network parameters at the moment when the weighted loss value is minimum as the network weights.

10. The method for optimizing parameters in a selenium-enriched yeast fermentation process based on a neural network according to claim 1, characterized in that, After the optimal parameter set is sent to the fermenter control system, the process also includes: after the fermentation of the current batch is completed, offline detection of the measured selenium organication rate of the current batch, calculation of the batch feature vector of the current batch, and appending the batch feature vector and the measured selenium organication rate to the historical batch dataset; when the number of appended batches reaches a preset number, the process of re-exercising the batch loss weight differential weighting, training the BP neural network, obtaining the network weights, and updating the network weights is performed using the updated historical batch dataset.

Images