Method for optimizing control of gluconic acid derivative production process based on multi-data fusion

Abstract

This invention relates to the field of production process optimization control, and particularly to a method for optimizing the production process of gluconic acid derivatives based on multi-data fusion. The method includes: preprocessing multi-source data acquired from the reactor in a semi-acid, semi-calcium process, as well as the alkali flow rate variable to be solved, to generate a reaction state vector; passing the reaction state vector through a neutralization reaction fitting model based on a gated cyclic unit, a temporal convolutional network, and a Transformer encoder architecture to generate a predicted reactor pH; solving the flow rate target equation for calcium acid based on the predicted reactor pH and the feed safety boundary to determine the calcium feed flow rate to be executed in the reactor; and periodically calculating the predicted reactor pH and alkali feed flow rate to optimize the reactor control. This invention balances multi-variable characteristics, prediction accuracy, and process uniformity in optimizing the production process of gluconic acid derivatives, thereby achieving high-quality and stable production of gluconic acid derivatives.

Description

Technical Field

[0001] This invention relates to the field of production process optimization and control, and in particular to a method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion. Background Technology

[0002] Gluconic acid derivatives, including gluconic acid, calcium gluconate, and gluconolactone, are important raw materials in the food, pharmaceutical, and chemical industries. In their production, the semi-acid, semi-calcium process uses pre-neutralized calcium carbonate solution to neutralize with calcium hydroxide alkaline solution in a reactor, forming a semi-acid, semi-calcium mixed system where gluconic acid and calcium gluconate coexist. This provides a crucial intermediate state for subsequent fractional production of different gluconic acid derivatives. Compared to traditional all-calcium and all-acid processes, the semi-acid, semi-calcium process significantly reduces the amount of calcium sulfate waste, lowers environmental treatment costs and raw material losses, and improves the conversion efficiency of gluconic acid, making it a key process promoted in the industry.

[0003] However, the difficulty in controlling the semi-acid and semi-calcium process lies in the fact that the reaction process has the characteristics of large lag and many interferences, and the progress of the acid-base neutralization reaction is strongly coupled with the feeding parameters of calcium carbonate solution and calcium hydroxide solution. As the core indicator for characterizing reaction equilibrium and the semi-acid and semi-calcium ratio, the stability of the acid-base value directly determines the quality consistency of the intermediate product.

[0004] Currently, the feeding control for the semi-acid, semi-calcium process only adopts traditional PID feedback control. This involves online detection of the real-time pH value of the reactor and adjustment of the calcium acid feeding parameters based on the target pH value. However, the traditional control method relies solely on single pH detection data and does not fully integrate multi-source sensor data from the reaction process. This single-data-driven control mode is prone to pH detection lag and insufficient prediction accuracy, which in turn leads to feeding adjustment lag, pH overshoot, oscillation, and other problems. This causes the semi-acid, semi-calcium ratio to easily deviate from the optimal target range. When the ratio exceeds the optimal target range, feeding too quickly will cause over-neutralization and insufficient free gluconic acid, affecting the efficiency of subsequent acidification and decalcification. Feeding too slowly will lead to insufficient neutralization, insufficient calcium gluconate formation, and reduced product yield.

[0005] Therefore, how to optimize and control the production process of gluconic acid derivatives while taking into account multivariate characteristics, prediction accuracy, and process uniformity, so as to achieve high-quality and stable production of gluconic acid derivatives, is a technical problem that needs to be solved. Summary of the Invention

[0006] To address this issue, the present invention provides a method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion, which overcomes the problem in the prior art that it is impossible to simultaneously consider multivariate characteristics, prediction accuracy, and process uniformity in optimizing and controlling the production process of gluconic acid derivatives, thus failing to achieve high-quality and stable production of gluconic acid derivatives.

[0007] To achieve the above objectives, this invention proposes an optimization control method for the production process of gluconic acid derivatives based on multi-data fusion, comprising: The current acid concentration data, alkali concentration data, temperature data, stirring speed data, alkali feed flow rate data, reactor liquid level data, conductivity data, component concentration data, and the alkali flow rate variable to be solved in the previous time step are preprocessed to generate a reaction state vector. The reaction state vector is passed through a neutralization reaction fitting model based on a gated recurrent unit, a temporal convolutional network, and a Transformer encoder architecture to generate a predicted reaction vessel pH. The target equation for calcium acid flow rate is solved based on the predicted reactor pH and feed safety boundary to determine the calcium solution feed flow rate to be implemented in the reactor. The neutralization reaction fitting model is executed periodically and the predicted acidity / alkalinity of the reactor and the alkali feed flow rate are calculated periodically to optimize the control of the reactor.

[0008] Furthermore, the process of determining the calcium solution feed flow rate for the reactor includes: Calculate the sum of the deviations between the predicted reactor pH and the set pH at all predicted time steps; calculate the first sum of the alkali flow rate variables to be solved at all predicted time steps; and calculate the second sum of the alkali flow rate changes to be solved at all predicted time steps. The sum of the deviation values, the first sum, and the second sum are semi-dynamically weighted and summed to generate the process value of the target equation for the flow rate of calcium acid. The objective equation for the flow rate of calcium acid is solved based on the process values ​​and the feed safety boundary to determine the calcium liquid feed flow rate to be applied to the reactor.

[0009] Furthermore, the process of generating the process value of the target equation for the flow rate of calcium acid includes: Calculate the first difference in the predicted reactor pH at adjacent time steps, calculate the second difference in the change in alkaline flow rate at adjacent time steps, and generate a buffer capacity index based on the ratio of the first difference and the second difference. The buffer capacity index is normalized within the process sensitivity range to generate a sensitivity coefficient. The sensitivity coefficient is passed through a first mapping function to generate a dynamic acid-base weight, wherein the first mapping function includes an acid-base adjustment factor and an acid-base reference value; The sensitivity coefficient is passed through a second mapping function to generate a dynamic weight for the alkali flow rate, wherein the second mapping function includes an alkali flow rate adjustment factor and an alkali reference value; Based on the dynamic weights of pH and alkali flow rate, the sum of the deviation values ​​and the second sum are dynamically weighted and summed, and then summed with the first sum based on fixed weights to generate the process value of the target equation for the flow rate of calcium acid.

[0010] Furthermore, the process of solving the objective equation for the flow rate of calcium acid to determine the calcium solution feed flow rate includes: The upper limit of the alkaline solution flow rate is determined by comparing the deviation between the predicted pH of the reactor and the set pH with the threshold. The current semi-calcium accumulation volume is generated based on the integral value of the product of the alkaline flow rate variable and the acid concentration data to be solved. The difference between the upper limit of the semi-calcium accumulation volume and the current semi-calcium accumulation volume is multiplied by the cumulative acid equivalent to generate the upper limit of the calcium equivalent increment. Based on the first product of the control step size and calcium solution concentration data, the upper limit of the calcium equivalent increment is divided by the first product to generate the maximum safe alkaline solution flow rate. The upper limit of the alkaline solution flow rate and the maximum safe alkaline solution flow rate are used as constraints, and the process values ​​are used to solve the objective equation of calcium acid flow rate through an optimization solver to determine the calcium solution feed flow rate. The feed safety boundary includes the upper limit of the alkali flow rate and the maximum safe alkali flow rate.

[0011] Furthermore, the process of generating a predicted pH value for the reaction vessel by fitting a neutralization reaction model includes: The reaction state vector is passed through a gated loop unit to generate reaction evolution features; The reaction state vector is passed through a temporal convolutional network and a Transformer encoder to generate globally coupled features; The concatenated vector of the reaction evolution features and global coupling features is passed through the output fully connected layer to generate the predicted acidity and alkalinity of the reactor. The neutralization reaction fitting model includes an output fully connected layer.

[0012] Furthermore, the process of generating reaction evolution features through gated cyclic units includes: The reaction state vector is passed through the first GRU layer to generate an initial hidden state sequence; The initial hidden state sequence is processed through drop-out and layer normalization operations to generate initial reaction evolution features; The initial reaction evolution features are passed through the second GRU layer to generate reaction evolution features; The gated loop unit includes a first GRU layer, a discard layer, a layer normalization operation, and a second GRU layer.

[0013] Furthermore, the process of generating globally coupled features through temporal convolutional networks and Transformer encoders includes: The reaction state vector is passed through an initial convolutional layer to generate initial reaction mapping features; The initial response mapping features are passed through causal dilated convolution residual blocks to generate response local features; The local reactive features are passed through a Transformer encoder to generate the global coupling features; The temporal convolutional network includes an initial convolutional layer and a causal dilated convolutional residual block.

[0014] Furthermore, the process of generating a predicted pH value for the reaction vessel through a neutralization reaction fitting model also includes: Acquire historical data from the reactor to construct a reaction dataset; After training the neutralization reaction fitting model using the reaction dataset through a loss function based on the mean squared error term and the acid-calcium neutralization guide term, the reaction state vector is used to generate a predicted acid-base balance of the reaction vessel through the neutralization reaction fitting model.

[0015] Furthermore, the preprocessing to generate the reaction state vector includes: The acid concentration data, alkali concentration data, temperature data, stirring speed data, alkali feed flow rate data, reactor liquid level data, conductivity data, component concentration data, and the alkali flow rate variable to be solved are standardized and preprocessed to generate a reaction state vector.

[0016] Furthermore, the process of optimizing the control of the reactor includes: After the set time for the calcium solution feed rate is executed in the reactor, the reaction state vector is reacquired. The reaction state vector is then used to calculate the predicted reactor pH and alkali feed rate through a neutralization reaction fitting model in order to optimize the control of the reactor.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: by introducing a buffer capacity index and a sensitivity coefficient, and by calculating the neutralization and buffering characteristics of the reaction system in real time, the present invention dynamically adjusts the weight coefficients of the pH deviation term and the flow rate change term in the target equation, thereby achieving semi-dynamic weighted optimization. When the reaction system is in the pH sensitive range, the weight of the pH deviation term is automatically increased to prioritize ensuring the pH control accuracy; when the reaction system is in the pH stable range, the weight of the flow rate change term is automatically increased to prioritize ensuring the stability of the feed flow rate. This allows the reactor to actively adapt to the operating conditions of raw material concentration fluctuations, enzyme activity changes, and reaction temperature disturbances, ensuring the process uniformity of the gluconic acid derivative production process.

[0018] In particular, this invention accurately predicts the future evolution trend of the reaction system through a two-layer GRU branch, and efficiently extracts the global dynamic features of the reaction process through TCN and Transformer branches, which can accurately discover the correlation of hidden variables and realize the production process of gluconic acid derivatives while taking into account the multivariate features and the accuracy of pH prediction. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the optimized control method for the production process of gluconic acid derivatives based on multi-data fusion, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the calcium solution feed flow rate solution in the gluconic acid derivative production process optimization control method based on multi-data fusion, according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the neutralization reaction fitting model of the gluconic acid derivative production process optimization control method based on multi-data fusion, according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the gated loop unit of the gluconic acid derivative production process optimization control method based on multi-data fusion, according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0023] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] like Figures 1 to 4 As shown, the present invention provides a method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion, which overcomes the problem that the existing technology cannot simultaneously take into account the characteristics of multiple variables, prediction accuracy and process uniformity in optimizing and controlling the production process of gluconic acid derivatives, thus failing to achieve high-quality and stable production of gluconic acid derivatives.

[0025] like Figure 1 As shown, this embodiment proposes an optimized control method for the production process of gluconic acid derivatives based on multi-data fusion, including: The current acid concentration data, alkali concentration data, temperature data, stirring speed data, alkali feed flow rate data, reactor liquid level data, conductivity data, component concentration data, and the alkali flow rate variable to be solved in the previous time step are preprocessed to generate a reaction state vector. The reaction state vector is passed through a neutralization reaction fitting model based on a gated recurrent unit, a temporal convolutional network, and a Transformer encoder architecture to generate a predicted reaction vessel pH. The target equation for calcium acid flow rate is solved based on the predicted reactor pH and feed safety boundary to determine the calcium solution feed flow rate to be implemented in the reactor. The neutralization reaction fitting model is executed periodically and the predicted acidity / alkalinity of the reactor and the alkali feed flow rate are calculated periodically to optimize the control of the reactor.

[0026] Specifically, in the pre-neutralization stage, after glucose is oxidized to gluconic acid by enzymatic or fermentation methods, a calcium carbonate suspension is first added to the reactor for pre-neutralization. This stage synchronously matches the glucose oxidation reaction process, promptly neutralizing the generated gluconic acid, which is the calcium solution. The oxidation reaction formula for this pre-neutralization stage is: , It is evident that the reaction in this pre-neutralization stage is mild, and the generated carbon dioxide gas can play an auxiliary role in stirring and improving mass transfer efficiency, thus eliminating the need for precise control.

[0027] After the glucose in the pre-neutralization stage of the reactor is completely converted into gluconic acid and the oxidation reaction reaches its endpoint, the final neutralization stage begins. The temperature is raised to inactivate the enzymes and terminate fermentation. Then, refined calcium hydroxide emulsion (i.e., alkali solution) is added to neutralize the remaining gluconic acid. The pH of the system is precisely controlled to the optimal range for the process, i.e., pH 6.0 to pH 6.3. The reaction formula for this final neutralization stage is: , Therefore, the final neutralization stage requires a slow and uniform gradient to the acid solution. Add alkaline solution To produce calcium solution By stabilizing the pH of the system to the optimal range of the process, the higher the control accuracy of the alkali feed flow rate, the better the consistency of the product quality, and thus the ability to meet the strict purity requirements of food and pharmaceutical grades.

[0028] Therefore, this embodiment achieves complete MPC (Model Predictive Control) control of the alkali feed flow rate by periodically executing the neutralization reaction fitting model and periodically calculating the predicted reactor pH and alkali feed flow rate.

[0029] like Figure 2 As shown, the process of determining the calcium solution feed flow rate for the reactor further includes: Calculate the sum of the deviations between the predicted reactor pH and the set pH at all predicted time steps; calculate the first sum of the alkali flow rate variables to be solved at all predicted time steps; and calculate the second sum of the alkali flow rate changes to be solved at all predicted time steps. The sum of the deviation values, the first sum, and the second sum are semi-dynamically weighted and summed to generate the process value of the target equation for the flow rate of calcium acid. The objective equation for the flow rate of calcium acid is solved based on the process values ​​and the feed safety boundary to determine the calcium liquid feed flow rate to be applied to the reactor.

[0030] Specifically, the target equation for the flow rate of calcium acid is: In the formula, J represents the optimization objective, N represents the total number of prediction time steps, and k represents the current k-th prediction time step. Let represent the dynamic weight of pH, the fixed weight, and the dynamic weight of alkali flow rate at the current k-th prediction time step, respectively. These represent the predicted pH and set pH of the reactor at the current prediction time step k, respectively. and represent the solution flow rate variable and change in solution flow rate at the current k-th prediction time step, respectively. The pH is set to 5.8, corresponding to a weakly acidic buffer system in the final neutralization stage.

[0031] Therefore, the sum of the deviations forces the solver to find an alkaline flow rate sequence that makes the predicted pH of the reactor as close as possible to the target trajectory within the next N steps. The dynamic weight of pH is not fixed and can be adjusted in real time according to the buffer capacity. The penalty is automatically reduced in the high-gain sensitive area to prevent excessive alkaline flow rate output to correct small pH deviations, thus avoiding overneutralization at the source.

[0032] The first sum penalizes the square of the alkali flow rate itself, which is directly related to material costs. At the same time, it suppresses the solver from outputting a flow rate that far exceeds the actual need. In the scenario where only alkali is introduced during the pre-neutralization stage, the first sum constrains the pace of total alkali input, avoiding insufficient local reaction caused by a large injection at one time.

[0033] The second summation is a penalty for changes in the alkali flow rate. When pH is adjusted solely by alkali, frequent speed changes in the mechanical pump cause flow fluctuations. These fluctuations, transmitted to a highly sensitive neutralization system, directly translate into sawtooth-shaped pH fluctuations and lead to inconsistent instantaneous calcium ion concentrations, resulting in encapsulation and uneven local precipitation. This requirement mandates that changes in the alkali flow rate must be smooth and gradual, protecting the uniformity of the product stoichiometry within the reactor.

[0034] Furthermore, the process of generating the process value of the target equation for the flow rate of calcium acid includes: Calculate the first difference in the predicted reactor pH at adjacent time steps, calculate the second difference in the change in alkaline flow rate at adjacent time steps, and generate a buffer capacity index based on the ratio of the first difference and the second difference. The buffer capacity index is normalized within the process sensitivity range to generate a sensitivity coefficient. The sensitivity coefficient is passed through a first mapping function to generate a dynamic acid-base weight, wherein the first mapping function includes an acid-base adjustment factor and an acid-base reference value; The sensitivity coefficient is passed through a second mapping function to generate a dynamic weight for the alkali flow rate, wherein the second mapping function includes an alkali flow rate adjustment factor and an alkali reference value; Based on the dynamic weights of pH and alkali flow rate, the sum of the deviation values ​​and the second sum are dynamically weighted and summed, and then summed with the first sum based on fixed weights to generate the process value of the target equation for the flow rate of calcium acid.

[0035] Specifically, the process of calculating the buffer capacity index is as follows: Within the sliding window, for each set of adjacent time steps, the second difference of the unsolved alkaline flow variable and the first difference of the predicted reactor pH are used to fit a straight line passing through the origin from all the first and second differences, i.e. Where ΔpH represents the first difference, Δu represents the second difference, and k represents the fitting slope. Therefore, by calculating the median of the slope for all data points, or by iteratively removing outliers, it is possible to effectively resist instantaneous outliers caused by uneven stirring and bubble interference. The fitting slope representing the current local static gain is output. The absolute value of the fitting slope is taken to obtain the instantaneous buffer capacity index. The larger the buffer capacity index value, the more severe the pH fluctuation caused by the unit change in alkali flow rate, that is, the more sensitive the system is currently in the high gain region.

[0036] Specifically, the process of calculating the sensitivity coefficient is as follows: after subtracting the lower boundary from the buffer capacity index, divide it by the difference between the upper and lower boundaries to obtain the sensitivity coefficient. Then, the current real-time estimated buffer capacity index is linearly normalized within the process sensitivity range to obtain a sensitivity coefficient in the range of [0,1]. If the buffer capacity index is less than the lower boundary, the sensitivity coefficient is forced to be 0; if the buffer capacity index is greater than the lower boundary, the sensitivity coefficient is forced to be 0.

[0037] The process sensitivity range includes a lower boundary of 4.5 and an upper boundary of 5.8. It can be understood that when the pH is less than 4.5, it is a flat zone of the final neutralization stage reaction, at which point the reaction is sluggish and the alkali flow rate can be adjusted relatively aggressively. When the pH is greater than 4.5, it is a sensitive zone of the final neutralization stage reaction, at which point flow rate changes should be extremely suppressed.

[0038] Specifically, the process of calculating the dynamic weight of pH through the first mapping function is as follows: subtract the product of the sensitivity coefficient and the pH adjustment factor, and then multiply by the pH reference value to obtain the dynamic weight of pH. The alkali flow rate adjustment factor is 0.6, and the pH reference value is 100. Therefore, the larger the buffer capacity index, the more linearly the dynamic weight of pH decreases. This means that the flow rate calcium acid objective equation actively relaxes the strict requirements for tracking small pH deviations in the high-sensitivity zone, allowing for short-term static errors. This avoids drastic changes in alkali flow rate and over-neutralization oscillations caused by perfect correction. The alkali flow rate adjustment factor of 0.6 ensures that the weight is not less than half of the reference value. The pH reference value of 100 constitutes a significant driving term in the overall objective function, forcing the solver of the flow rate calcium acid objective equation to prioritize reducing pH deviations and avoid the flow rate calcium acid objective equation being too weak to overcome model errors and disturbances.

[0039] Specifically, the process of calculating the dynamic weight of the alkali flow rate through the second mapping function is as follows: multiply the sensitivity coefficient by the alkali flow rate adjustment factor, add one, and then multiply by the alkali reference value to obtain the dynamic weight of the alkali flow rate. The alkali flow rate adjustment factor is 1.5, and the alkali reference value is 1.0. Therefore, the larger the buffer capacity index, the more linearly the dynamic weight of the alkali flow rate increases, the heavier the penalty for the change in alkali flow rate, and the more stable the flow rate curve output by the solver is forced. The alkali flow rate adjustment factor of 1.5 makes the penalty in the high-sensitivity zone reach twice the reference value. The alkali reference value is 1.0. In the reaction plateau zone, the change in the alkali flow rate to be solved in each step is usually between 0.1 and 0.5 L / min. Taking 0.5, its square is 0.25. Multiplying by the alkali reference value of 1.0 gives 0.25, which is on the same order of magnitude as the product of the sum of the deviation values ​​and the dynamic weight of pH, which is approximately 1.0. The two are in balance.

[0040] Specifically, the fixed weighting is 0.05, ensuring that when the deviation of the cumulative alkali equivalent from the theoretical endpoint is less than 5% and the pH has entered the target window, the solution's alkali flow rate variable linearly decreases to 0.02. The target window is set at pH ± 0.15. Therefore, conservative optimization can be achieved to reduce the solution's alkali flow rate variable without compromising quality and safety.

[0041] Furthermore, the process of solving the objective equation for the flow rate of calcium acid to determine the calcium solution feed flow rate includes: The upper limit of the alkaline solution flow rate is determined by comparing the deviation between the predicted pH of the reactor and the set pH with the threshold. The current semi-calcium accumulation volume is generated based on the integral value of the product of the alkaline flow rate variable and the acid concentration data to be solved. The difference between the upper limit of the semi-calcium accumulation volume and the current semi-calcium accumulation volume is multiplied by the cumulative acid equivalent to generate the upper limit of the calcium equivalent increment. Based on the first product of the control step size and calcium solution concentration data, the upper limit of the calcium equivalent increment is divided by the first product to generate the maximum safe alkaline solution flow rate. The upper limit of the alkaline solution flow rate and the maximum safe alkaline solution flow rate are used as constraints, and the process values ​​are used to solve the objective equation of calcium acid flow rate through an optimization solver to determine the calcium solution feed flow rate. The feed safety boundary includes the upper limit of the alkali flow rate and the maximum safe alkali flow rate.

[0042] Specifically, the process for determining the upper limit of the alkali flow rate is as follows: if the deviation value is greater than the large deviation threshold, the upper limit of the alkali flow rate is 0.5 for the fast speed setting; if the deviation value is less than the large deviation threshold but greater than the small deviation threshold, the upper limit of the alkali flow rate is 0.1 for the medium speed setting; if the deviation value is less than the small deviation threshold, the upper limit of the alkali flow rate is 0.02 for the low speed setting. The thresholds include a large deviation threshold of 0.3 and a small deviation threshold of 0.08.

[0043] Therefore, in the large deviation range, the alkali flow rate is balanced between speed and equipment not exceeding limits; in the medium deviation range, a safe step size is adapted and coordinated with the buffer capacity adaptive penalty; in the small deviation range, the endpoint fine-tuning accuracy is performed to approach the pump's minimum continuously controllable resolution and forced peristalsis is achieved.

[0044] Specifically, the maximum safe alkali flow rate is the maximum allowable calcium solution volumetric flow rate to maintain the molar ratio within limits. This allows the upper limit of the flow rate to automatically decrease as the molar ratio approaches the warning line, further forcing the alkali feed into a peristaltic state and fundamentally preventing alkali overload. The upper limit of the semi-calcium cumulative volume is 0.52, meaning a maximum calcium overload of 4%. At this point, the product still meets the requirements. The control step size is 10 seconds, which effectively captures the dynamic response of pH and satisfies Shannon's sampling theorem that the fastest reaction fluctuation period is no less than 30 seconds.

[0045] Specifically, the neutralization reaction fitting model can be simplified as predicting the acidity / alkalinity of the reactor as a function of the alkali flow rate variable. The neutralization reaction fitting model can obtain the predicted acidity / alkalinity of the reactor at multiple time steps. With minimizing the process value as the optimization objective, and the upper limit of the alkali flow rate and the maximum safe alkali flow rate as constraints, the flow rate calcium acid objective equation is solved by an optimization solver to determine the optimal value of the alkali flow rate variable, which is then used as the calcium solution feed flow rate.

[0046] Specifically, the optimization solver is CasaADi, which uses the IPOPT interior-point method to solve the above nonlinear programming problem. Key solver settings include: convergence tolerance of... The maximum number of iterations is 20 to 50, and convergence is accelerated by using real-time feasible initial values.

[0047] like Figure 3 As shown, the process of generating a predicted pH value for the reaction vessel through a neutralization reaction fitting model further includes: The reaction state vector is passed through a gated loop unit to generate reaction evolution features; The reaction state vector is passed through a temporal convolutional network and a Transformer encoder to generate globally coupled features; The concatenated vector of the reaction evolution features and global coupling features is passed through the output fully connected layer to generate the predicted acidity and alkalinity of the reactor. The neutralization reaction fitting model includes an output fully connected layer.

[0048] Specifically, the output fully connected layer is a linear fully connected layer without an activation function.

[0049] Therefore, by employing a dual-stream heterogeneous feature extraction and fusion architecture that integrates GRU and TCN-Transformer in a parallel ensemble model, the model can simultaneously understand the reaction process from both the perspectives of sequence evolution and complex multivariate dependencies, thereby achieving more accurate pH prediction. Specifically, the Gated Recurrent Unit (GRU) focuses on learning the temporal evolution of the reaction process, excelling at capturing short-term dynamics and sequential dependencies. The Temporal Convolutional Network (TCN) and Transformer simultaneously capture multi-scale local patterns and complex global multivariate coupling relationships, providing a powerful complement to the GRU.

[0050] like Figure 4 As shown, the process of generating reaction evolution features through the gated recurrent unit further includes: The reaction state vector is passed through the first GRU layer to generate an initial hidden state sequence; The initial hidden state sequence is processed through drop-out and layer normalization operations to generate initial reaction evolution features; The initial reaction evolution features are passed through the second GRU layer to generate reaction evolution features; The gated loop unit includes a first GRU layer, a discard layer, a layer normalization operation, and a second GRU layer.

[0051] Specifically, the first GRU layer has 128 memory units and outputs the hidden state sequence of all time steps. The discard rate of the discard layer is 0.2. The second GRU layer has 128 memory units and outputs the hidden state of the last time step, which is used as the reaction evolution feature.

[0052] Therefore, the first GRU layer re-expresses the reaction-state combination at each time step as a high-dimensional semantic vector, extracting rich context from each local time step. The dropout layer randomly discards some intermediate representations in the temporal dimension, forcing the network to be independent of noise at specific time steps and improving generalization ability. Layer normalization is used to stabilize training, making the deep second GRU layer easier to converge. The second GRU layer processes the internal state trajectory generated by the first layer frame by frame along the time axis, continuously updating, forgetting, and storing information through a gating mechanism, ultimately condensing the entire evolution process of the historical window into a single fixed-dimensional vector. The output vector captures the dynamic trend of the sequence from beginning to end.

[0053] Furthermore, the process of generating globally coupled features through temporal convolutional networks and Transformer encoders includes: The reaction state vector is passed through an initial convolutional layer to generate initial reaction mapping features; The initial response mapping features are passed through causal dilated convolution residual blocks to generate response local features; The local reactive features are passed through a Transformer encoder to generate the global coupling features; The temporal convolutional network includes an initial convolutional layer and a causal dilated convolutional residual block.

[0054] Specifically, the initial convolutional layer uses a 1×1 convolution to map the feature dimension to a higher dimension, providing a suitable dimension for subsequent multi-channel feature learning in the TCN. Causal dilated convolutional residual blocks are set with dilation rates of 1, 2, 4, and 8, respectively. Each causal dilated convolutional residual block sequentially sets up a causal dilated convolutional layer 1, weight normalization, GELU activation, Dropout, a causal dilated convolutional layer 2, weight normalization, GELU activation, Dropout, and a residual connection. Therefore, the local response features output by the causal dilated convolutional residual blocks can cover response dynamics at different scales, including rapid flow fluctuations and slow temperature drifts.

[0055] Specifically, the Transformer encoder sequentially sets up multi-head self-attention, residual connectivity and layer normalization 1, fully connected linear layer 1, GELU activation, fully connected linear layer 2, and layer normalization 2. Therefore, multi-head self-attention allows each time step to freely interact with all other time steps within the window, thereby fully capturing global long-range dependencies.

[0056] Specifically, the output fully connected layer is a linear fully connected layer without an activation function. The concatenated vector has been highly abstracted by GRU and TCN-Transformer, containing multi-scale dynamic trends, short-term inertia, and global coupling information. Through the affine transformation of the output fully connected layer, it is sufficient to map the comprehensive features of the concatenated vector to a specific pH value.

[0057] Furthermore, the process of generating a predicted pH value for the reaction vessel through a neutralization reaction fitting model also includes: Acquire historical data from the reactor to construct a reaction dataset; After training the neutralization reaction fitting model using the reaction dataset through a loss function based on the mean squared error term and the acid-calcium neutralization guide term, the reaction state vector is used to generate a predicted acid-base balance of the reaction vessel through the neutralization reaction fitting model.

[0058] Specifically, the mean squared error (MSE) term is the difference between the predicted pH of the reaction vessel and the historical actual pH in the reaction dataset. The calculation process for the calcium acid neutralization guide term is as follows: Based on acid volume data, acid concentration data, historical alkali flow rate data, alkali concentration data, and component concentration data, the increase in hydroxide ions is calculated. This increase in hydroxide ions is divided by the liquid volume in the reaction vessel to obtain the theoretical pH change. The difference between the predicted pH change in the reaction vessel and the theoretical pH change in the reaction dataset is averaged across all prediction time steps and used as the calcium acid neutralization guide term. The MSE term and the calcium acid neutralization guide term are weighted and summed with weights of 1.0 and 0.2 to obtain the loss function. Therefore, the loss function ensures the accuracy of pH prediction while ensuring that the model's prediction trajectory closely follows the stoichiometric laws of the half-acid, half-calcium reaction. The calcium acid neutralization guide term acts as a soft constraint, preventing the model output from violating the stoichiometric pH trend without excessively suppressing the real complex dynamics in the data.

[0059] Furthermore, the preprocessing to generate the reaction state vector includes: The acid concentration data, alkali concentration data, temperature data, stirring speed data, alkali feed flow rate data, reactor liquid level data, conductivity data, component concentration data, and the alkali flow rate variable to be solved are standardized and preprocessed to generate a reaction state vector.

[0060] Specifically, the standardization preprocessing is Z-score standardization.

[0061] Furthermore, the process of optimizing the control of the reactor includes: After the set time for the calcium solution feed rate is executed in the reactor, the reaction state vector is reacquired. The reaction state vector is then used to calculate the predicted reactor pH and alkali feed rate through a neutralization reaction fitting model in order to optimize the control of the reactor.

[0062] Specifically, the set duration is 10 seconds. The 10-second step size can fully capture the dynamic response of pH and satisfy the Shannon sampling theorem that the fastest fluctuation period is not less than 30 seconds.

[0063] Specifically, the neutralization reaction fitting model described in this embodiment is deployed on an edge computing node and communicates with the PLC of the reactor in real time via the OPCUA protocol. Ten batches of production were conducted using raw materials with a calcium carbonate suspension concentration of 25%±1%, a calcium hydroxide emulsion concentration of 15%±0.8%, and a glucose oxidase activity of 10000 U / g. The experimental group used the neutralization reaction fitting model described in this embodiment and periodically calculated the predicted reactor pH and alkali feed flow rate. The control group 1 used traditional single-loop PID control for pH: K. With p=2.5, Ki=0.1, and Kd=0.05, the alkaline solution flow rate was controlled using a fixed-ratio open-loop control. The pH fluctuation ranges during the pre-neutralization phase for the experimental group and control group 1 were ±0.12 and ±0.35, respectively. The standard deviations of pH during the pre-neutralization phase for the experimental group and control group 1 were 0.052 and 0.148, respectively. The final pH deviations during the final neutralization phase for the experimental group and control group 1 were ±0.06 and ±0.21, respectively. The pH attainment times during the final neutralization phase were 12 min and 20 min, respectively. The batch pH pass rates were 98% and 80%, respectively. It is evident that the control method in this embodiment solves the overshoot and oscillation problems of traditional control methods, significantly improving product quality consistency.

[0064] Specifically, the sampling frequency of the reaction dataset is 1Hz, and it is divided into training (70%, 168 batches), validation (15%, 36 batches), and test (15%, 36 batches) in chronological order to ensure the objectivity of the model's generalization ability verification. The neutralization reaction fitting model and the comparison model described in this embodiment are trained using the AdamW optimizer, with a learning rate of 1e-4, weight decay of 1e-5, batch size of 64, training epochs of 100, and the loss function. The MAE of the neutralization reaction fitting model, the pure GRU model, the pure TCN model, and the pure Transformer model are 0.124, 0.198, 0.231, and 0.247, respectively. It can be seen that under abnormal conditions, the prediction performance of the neutralization reaction fitting model remains stable and is far superior to other single models, proving that it has stronger generalization ability, accuracy, and robustness.

[0065] In this embodiment, by introducing a buffer capacity index and a sensitivity coefficient, and by calculating the neutralization and buffering characteristics of the reaction system in real time, the weighting coefficients of the pH deviation term and the flow rate change term in the objective equation are dynamically adjusted, achieving semi-dynamic weighted optimization. When the reaction system is in the pH sensitive range, the weight of the pH deviation term is automatically increased to prioritize pH control accuracy; when the reaction system is in the pH stable range, the weight of the flow rate change term is automatically increased to prioritize feed flow stability. This allows the reactor to actively adapt to fluctuations in raw material concentration, changes in enzyme activity, and reaction temperature disturbances, ensuring process uniformity in the production of gluconic acid derivatives. Through a two-layer GRU branch, the future evolution trend of the reaction system is accurately predicted. Through TCN and Transformer branches, the global dynamic features of the reaction process are efficiently extracted, accurately identifying hidden variable correlations. This achieves a balance between multivariate characteristics and pH prediction accuracy in the production process of gluconic acid derivatives.

[0066] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0067] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion, characterized in that, include: The current acid concentration data, alkali concentration data, temperature data, stirring speed data, alkali feed flow rate data, reactor liquid level data, conductivity data, component concentration data, and the alkali flow rate variable to be solved in the previous time step are preprocessed to generate a reaction state vector. The reaction state vector is passed through a neutralization reaction fitting model based on a gated recurrent unit, a temporal convolutional network, and a Transformer encoder architecture to generate a predicted reaction vessel pH. The target equation for calcium acid flow rate is solved based on the predicted reactor pH and feed safety boundary to determine the calcium solution feed flow rate to be implemented in the reactor. The neutralization reaction fitting model is executed periodically and the predicted acidity / alkalinity of the reactor and the alkali feed flow rate are calculated periodically to optimize the control of the reactor.

2. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 1, characterized in that, The process of determining the calcium solution feed flow rate for the reactor includes: Calculate the sum of the deviations between the predicted reactor pH and the set pH at all predicted time steps; calculate the first sum of the alkali flow rate variables to be solved at all predicted time steps; and calculate the second sum of the alkali flow rate changes to be solved at all predicted time steps. The sum of the deviation values, the first sum, and the second sum are semi-dynamically weighted and summed to generate the process value of the target equation for the flow rate of calcium acid. The objective equation for the flow rate of calcium acid is solved based on the process values ​​and the feed safety boundary to determine the calcium liquid feed flow rate to be applied to the reactor.

3. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 2, characterized in that, The process of generating the target value of the flow rate calcium acid equation includes: Calculate the first difference in the predicted reactor pH at adjacent time steps, calculate the second difference in the change in alkaline flow rate at adjacent time steps, and generate a buffer capacity index based on the ratio of the first difference and the second difference. The buffer capacity index is normalized within the process sensitivity range to generate a sensitivity coefficient. The sensitivity coefficient is passed through a first mapping function to generate a dynamic acid-base weight, wherein the first mapping function includes an acid-base adjustment factor and an acid-base reference value; The sensitivity coefficient is passed through a second mapping function to generate a dynamic weight for the alkali flow rate, wherein the second mapping function includes an alkali flow rate adjustment factor and an alkali reference value; Based on the dynamic weights of pH and alkali flow rate, the sum of the deviation values ​​and the second sum are dynamically weighted and summed, and then summed with the first sum based on fixed weights to generate the process value of the target equation for the flow rate of calcium acid.

4. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 2, characterized in that, The process of determining the calcium solution feed flow rate by solving the objective equation for calcium acid flow rate includes: The upper limit of the alkaline solution flow rate is determined by comparing the deviation between the predicted pH of the reactor and the set pH with the threshold. The current semi-calcium accumulation volume is generated based on the integral value of the product of the alkaline flow rate variable and the acid concentration data to be solved. The difference between the upper limit of the semi-calcium accumulation volume and the current semi-calcium accumulation volume is multiplied by the cumulative acid equivalent to generate the upper limit of the calcium equivalent increment. Based on the first product of the control step size and calcium solution concentration data, the upper limit of the calcium equivalent increment is divided by the first product to generate the maximum safe alkaline solution flow rate. The upper limit of the alkaline solution flow rate and the maximum safe alkaline solution flow rate are used as constraints, and the process values ​​are used to solve the objective equation of calcium acid flow rate through an optimization solver to determine the calcium solution feed flow rate. The feed safety boundary includes the upper limit of the alkali flow rate and the maximum safe alkali flow rate.

5. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 1, characterized in that, The process of generating a predicted pH level for the reaction vessel by fitting a neutralization reaction model includes: The reaction state vector is passed through a gated loop unit to generate reaction evolution features; The reaction state vector is passed through a temporal convolutional network and a Transformer encoder to generate globally coupled features; The concatenated vector of the reaction evolution features and global coupling features is passed through the output fully connected layer to generate the predicted acidity and alkalinity of the reactor. The neutralization reaction fitting model includes an output fully connected layer.

6. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 5, characterized in that, The process of generating reaction evolution features through gated recurrent units includes: The reaction state vector is passed through the first GRU layer to generate an initial hidden state sequence; The initial hidden state sequence is processed through drop-out and layer normalization operations to generate initial reaction evolution features; The initial reaction evolution features are passed through the second GRU layer to generate reaction evolution features; The gated loop unit includes a first GRU layer, a discard layer, a layer normalization operation, and a second GRU layer.

7. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 5, characterized in that, The process of generating globally coupled features using temporal convolutional networks and Transformer encoders includes: The reaction state vector is passed through an initial convolutional layer to generate initial reaction mapping features; The initial response mapping features are passed through causal dilated convolution residual blocks to generate response local features; The local reactive features are passed through a Transformer encoder to generate the global coupling features; The temporal convolutional network includes an initial convolutional layer and a causal dilated convolutional residual block.

8. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to claim 1, characterized in that, The process of generating a predicted pH level for the reaction vessel by fitting a neutralization reaction model also includes: Acquire historical data from the reactor to construct a reaction dataset; After training the neutralization reaction fitting model using the reaction dataset through a loss function based on the mean squared error term and the acid-calcium neutralization guide term, the reaction state vector is used to generate a predicted acid-base balance of the reaction vessel through the neutralization reaction fitting model.

9. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to any one of claims 1 to 8, characterized in that, The process of preprocessing to generate reaction state vectors includes: The acid concentration data, alkali concentration data, temperature data, stirring speed data, alkali feed flow rate data, reactor liquid level data, conductivity data, component concentration data, and the alkali flow rate variable to be solved are standardized and preprocessed to generate a reaction state vector.

10. The method for optimizing and controlling the production process of gluconic acid derivatives based on multi-data fusion according to any one of claims 1 to 8, characterized in that, The process of optimizing the control of the reactor includes: After the set time for the calcium solution feed rate is executed in the reactor, the reaction state vector is reacquired. The reaction state vector is then used to calculate the predicted reactor pH and alkali feed rate through a neutralization reaction fitting model in order to optimize the control of the reactor.

Images