Cefuroxime axetil preparation process parameter optimization method and system based on deep learning

Abstract

The application relates to a cefuroxime axetil preparation process parameter optimization method and system based on deep learning, and relates to the technical field of deep learning, and comprises the following steps: acquiring a cefuroxime acid residual concentration sequence and a by-product cumulative generation rate sequence in an esterification reaction; inputting the two sequences into a dynamic response prediction model to output a by-product growth inertia coefficient; constructing a residual concentration estimation attenuation path according to the latest residual concentration and a comprehensive attenuation rate; calculating a target reaction termination time based on the intersection of the path and a standard threshold value, and generating a catalyst or esterification reagent dynamic supplement correction amount in combination with the inertia coefficient; offset adjusting a basic supplement scheme according to the correction amount to obtain a corrected supplement dosage and a supplement time point; and controlling a supplement device to perform additional operations of the catalyst and the esterification reagent. The application solves the problem that, in a traditional cefuroxime axetil esterification reaction, the reaction completeness and by-product inhibition cannot be simultaneously considered due to the inability to predict the by-product growth inertia in real time and the static supplement strategy.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a method and system for optimizing the process parameters of cefuroxime axetil preparation based on deep learning. Background Technology

[0002] Cefuroxime axetil is an oral second-generation cephalosporin antibiotic. Its preparation process uses 7-aminocephalosporanic acid as the starting material, which undergoes hydrolysis, acylation, and carbamylation to obtain cefuroxime axetil, which is then esterified with 1-bromoethyl acetate under an alkaline catalyst. In actual production, due to fluctuations in reaction conditions, catalyst activity decay, or deviations in material ratios, the esterification reaction often becomes incomplete, resulting in cefuroxime axetil residues exceeding the process control standards. The process control standards typically require the cefuroxime axetil residue concentration to be reduced to below 1000-2000 mg / L. To address this issue, the traditional adjustment method involves directly adding potassium carbonate and 1-bromoethyl acetate during the reaction process and extending the reaction time until the residue concentration drops to the acceptable range.

[0003] However, since the single-needle analysis time of high-performance liquid chromatography (HPLC) is typically 12-25 minutes, it is difficult to achieve continuous online monitoring at the second level. Traditional catalyst or esterification reagent replenishment methods lack real-time perception and quantitative prediction of reaction dynamics. The replenishment dosage and timing often rely on fixed procedures or operator experience, and cannot be adaptively adjusted according to the current residual concentration and the actual generation trend of by-products. This can easily lead to a continued increase in reaction time and reduced production efficiency when the replenishment is insufficient, or an aggravation of side reactions and a significant increase in by-products and degradation impurities when the replenishment is excessive, while also causing raw material waste. Therefore, how to accurately determine the replenishment dosage and timing of catalyst or esterification reagent without significantly increasing the risk of by-products, so as to reduce the residual concentration to a qualified level in the shortest possible time, has become an urgent problem to be solved in esterification process control. Summary of the Invention

[0004] This application provides a method and system for optimizing process parameters of cefuroxime axetil preparation based on deep learning, which improves the technical problem in the existing technology that in the esterification process of cefuroxime axetil preparation, it is difficult to achieve a balance between reaction efficiency and impurity control due to the lack of dynamic response to the accumulation rate of by-products and the fixed addition parameters.

[0005] This application discloses the following technical solution: In a first aspect, this application provides a method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning, the method comprising: The real-time residual concentration sequence of cefuroxime axetine and the corresponding cumulative formation rate sequence of by-products were obtained from multiple consecutive sampling times in the current reaction batch in the esterification reactor. The real-time residual concentration sequence of cefuroxime axetine and the byproduct accumulation rate sequence are synchronously input into a pre-trained dynamic response prediction model, and the corresponding byproduct growth inertia coefficient is output. Based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime axetine, and combined with the comprehensive decay rate, a residual concentration prediction decay path is constructed. Based on the intersection of the estimated decay path of the residual concentration and the standard residual concentration threshold, the target reaction termination time is calculated, and combined with the byproduct growth inertia coefficient, a dynamic replenishment correction amount of catalyst or esterification reagent is generated. Based on the dynamic replenishment correction amount of the catalyst or esterification reagent, the preset basic replenishment scheme is offset and adjusted to obtain fine replenishment control parameters including the corrected replenishment dosage and the corrected replenishment timing. The addition device is controlled to perform the additional addition of catalyst and esterification reagent to the esterification reactor according to the fine addition control parameters.

[0006] Secondly, this application provides a deep learning-based system for optimizing the process parameters of cefuroxime axetil preparation, the system comprising: The sequence acquisition module is used to acquire the real-time residual concentration sequence of cefuroxime axetine and the corresponding cumulative generation rate sequence of by-products collected at multiple consecutive sampling times in the current reaction batch in the esterification reactor. The model prediction module is used to synchronously input the real-time residual concentration sequence of cefuroxime axetine and the by-product accumulation rate sequence into the pre-trained dynamic response prediction model and output the corresponding by-product growth inertia coefficient. The path construction module is used to construct a residual concentration prediction decay path based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime axetine and the comprehensive decay rate. The correction generation module is used to calculate the target reaction termination time based on the intersection of the estimated decay path of the residual concentration and the standard residual concentration threshold, and to generate a dynamic replenishment correction amount of catalyst or esterification reagent in combination with the byproduct growth inertia coefficient. The parameter adjustment module is used to offset the preset basic addition scheme according to the dynamic addition correction amount of the catalyst or esterification reagent, and obtain fine addition control parameters including the correction addition dose and the correction addition time. The replenishment control module is used to control the replenishment device to perform additional operations of catalyst and esterification reagent to the esterification reactor according to the fine replenishment control parameters.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: The technical solution of this application provides a deep learning-based method for optimizing the process parameters of cefuroxime axetil preparation. First, by monitoring the residual concentration of cefuroxime axetil and the corresponding cumulative generation rate of by-products at multiple consecutive sampling times during the esterification reaction using high-performance liquid chromatography, dual-channel time-series data is constructed, realizing continuous and multi-dimensional quantitative characterization of the reaction process. This solves the problems of traditional methods, such as information lag, single dimension, and inability to simultaneously obtain the dynamic changes of residual concentration and by-products.

[0008] Furthermore, by synchronously inputting dual-channel time-series data into a pre-trained deep learning model and outputting the byproduct growth inertia coefficient, a quantitative prediction of the byproduct growth trend after the addition of catalyst or esterification reagent is achieved, solving the problems of traditional methods lacking the ability to perceive the future inertial growth of byproducts and lacking data-driven basis for addition decisions.

[0009] Furthermore, by constructing a residual concentration prediction decay path based on the latest residual concentration value and combined with the comprehensive decay rate, adaptive prediction of the future decay trend of residual concentration is achieved, solving the problem that the traditional method using a fixed decay model cannot reflect the actual dynamics of the current batch and the path cannot be adjusted.

[0010] Furthermore, by calculating the target reaction termination time based on the intersection of the decay path and the standard residual concentration threshold, and combining the byproduct growth inertia coefficient to generate a dynamic correction amount for catalyst or esterification reagent replenishment, real-time quantitative correction of catalyst or esterification reagent replenishment amount is achieved, solving the problem that traditional replenishment schemes are fixed and cannot be differentiated according to the remaining reaction time and byproduct risk.

[0011] Furthermore, by adjusting the preset basic supplementation scheme based on the supplementation correction amount, fine control parameters including the corrected supplementation dose and the corrected supplementation time are obtained, realizing the synergistic optimization of the supplementation dose and the supplementation time. This solves the problem that traditional schemes only adjust the dose while ignoring the timing, or supplementation at fixed times, which leads to a mismatch with the reaction state.

[0012] Finally, by controlling the addition device to perform the additional operation of catalyst and esterification reagent according to the fine control parameters, closed-loop automated control from real-time perception and dynamic decision-making to precise execution is realized, which solves the problems of large deviation and poor consistency in traditional manual operation, making it difficult to ensure process stability.

[0013] In summary, the technical solution of this application realizes process monitoring of the esterification reaction state, prediction of by-product growth, construction of decay path, dynamic generation and fine control of supplementation amount during the preparation of cefuroxime axetil. Through continuous sampling dual-channel acquisition, deep learning model prediction, adaptive construction of exponential decay path, collaborative calculation of multi-factor correction amount and joint adjustment of dosage and time, it effectively improves the technical problems of fixed supplementation scheme, lack of quantitative prediction of by-product inertial growth, reliance on manual experience adjustment and inability to coordinate and optimize reaction efficiency and impurity control in the prior art, which leads to improper supplementation causing excessive by-products and waste of raw materials. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A schematic diagram of the overall flow of the deep learning-based method for optimizing the preparation process parameters of cefuroxime axetil provided in this application embodiment; Figure 2 A schematic diagram of the fine supplementation control parameter generation process of the deep learning-based cefuroxime axetil preparation process parameter optimization method provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the deep learning-based cefuroxime axetil preparation process parameter optimization system provided in the embodiments of this application.

[0016] In the attached diagram, Figure 3 The components represented by each label are explained as follows: Sequence acquisition module 11, Model prediction module 12, Path construction module 13, Correction generation module 14, Parameter adjustment module 15, and Supplement control module 16. Detailed Implementation

[0017] This application provides a method and system for optimizing the process parameters of cefuroxime axetil preparation based on deep learning, which is used to solve the technical problem in the prior art that the supplementation scheme cannot be dynamically adjusted according to the growth trend of by-products during the esterification process of cefuroxime axetil, causing a conflict between the reaction endpoint and the generation of impurities.

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that the reaction time, concentration values, and other parameters involved in the embodiments are all set based on the statistical range of the actual production process of cefuroxime axetil, and are only used for illustrative purposes and do not constitute a limitation on this application.

[0019] Example 1, as shown in the appendix Figure 1 As shown, this application provides a method for optimizing the process parameters of cefuroxime axetil preparation based on deep learning. The method includes the following steps: S100: Obtain the real-time residual concentration sequence of cefuroxime acid and the corresponding cumulative generation rate sequence of by-products collected at multiple consecutive sampling times in the current reaction batch in the esterification reactor; In this embodiment, during the esterification reaction of cefuroxime axetil, the residual concentration of cefuroxime acid gradually decreases as the reaction progresses, while byproducts continuously accumulate. These two factors are key indicators for judging the reaction progress and the severity of side reactions. The main byproduct is the delta-3 isomer, with a final process concentration typically between 2000 and 4000 mg / L. The final process control target for the residual concentration of cefuroxime axetil is 1000 to 2000 mg / L. Together, these two factors constitute the core parameters for reaction monitoring.

[0020] However, existing technologies typically only detect the residual concentration of cefuroxime axetine at a single moment through offline sampling, which cannot obtain concentration sequence data arranged in chronological order, nor can they simultaneously monitor the cumulative generation rate of byproducts. This results in a lag in understanding the reaction process and a single dimension of information, making it difficult to comprehensively and in real time assess the reaction status.

[0021] Step S100 in the method provided in this application embodiment includes: A sample of the reaction liquid from the esterification reactor in the current batch is collected. The peak area of ​​cefuroxime acid and the cumulative area of ​​adjacent impurity peaks at each sampling time are sequentially obtained using a high-performance liquid chromatography (HPLC) analyzer. The peak area of ​​cefuroxime acid is substituted into a pre-calibrated concentration-area standard curve to obtain the real-time residual concentration of cefuroxime acid at each sampling time. The residual concentrations at all sampling times are arranged in chronological order to form a real-time residual concentration sequence of cefuroxime acid. Based on the cumulative area of ​​adjacent impurity peaks and the peak area of ​​cefuroxime acid at each sampling time, the relative abundance of impurities is calculated. Based on the difference in relative abundance of impurities between two adjacent sampling times and the corresponding time interval, the cumulative formation rate of by-products within each sampling interval is calculated. The rates corresponding to all sampling intervals are arranged in chronological order to form a cumulative formation rate sequence of by-products. The real-time residual concentration sequence of cefuroxime acid and the cumulative formation rate sequence of by-products are aligned according to the sampling time to form a dual-channel input data stream. Detailed explanation follows: In this embodiment, the built-in ultraviolet detector of the high-performance liquid chromatography (HPLC) analyzer is used to monitor the absorbance of cefuroxime axetine and its byproduct (delta-3 isomer) at a specific wavelength. Sampling is performed periodically during the reaction. It should be noted that the sampling interval can be set according to the actual chromatographic analysis cycle; in this application, it is preferably 12-20 minutes to match the time window required for single-needle HPLC analysis. The entire esterification reaction of cefuroxime axetine typically lasts 80-300 minutes. Within this timescale, a sampling interval of 12-20 minutes can obtain a sufficient number of data points for subsequent model prediction and pathway construction.

[0022] It should be noted that the byproduct in this application is the delta-3 isomer generated by the esterification reaction during the preparation of cefuroxime axetil. This isomer is the product of the isomerization reaction of cefuroxime axetil under alkaline conditions. Its generation rate is closely related to the activity of potassium carbonate in the reaction system and the reaction temperature, and is a key indicator for measuring the intensity of the side reaction.

[0023] Furthermore, the cumulative area of ​​adjacent impurity peaks refers to the total peak area of ​​the delta-3 isomer impurity peak and its related degradation product peaks that are adjacent to the retention time of the cefuroxime acid main peak in the chromatogram. This area value is used to characterize the cumulative formation level of byproducts. The relative abundance of impurities is the ratio of the cumulative area of ​​adjacent impurity peaks to the chromatographic peak area of ​​cefuroxime acid. This ratio can eliminate the interference of absolute area changes caused by injection volume fluctuations or instrument response drift, making the byproduct levels comparable between different sampling times. The byproduct cumulative formation rate is defined as the difference in the relative abundance of impurities between two adjacent sampling times divided by the corresponding time interval. This rate reflects the net amount of byproducts formed per unit time and is one of the core input features of the dynamic response prediction model.

[0024] A concentration-area standard curve is a linear regression curve obtained by preparing a series of cefuroxime acid standard solutions of known concentrations, measuring their chromatographic peak areas, and then converting the measured peak areas into concentration values. A matrix effect correction factor is the ratio of the measured recovered chromatographic peak area to the theoretical peak area calculated from the initial standard curve after adding a known concentration of cefuroxime acid standard to a blank reaction matrix. It is used to correct for the influence of the reaction matrix on the chromatographic response.

[0025] Specifically, in order to obtain real-time component information of the reaction solution, samples need to be extracted periodically and chromatographically separated using a high-performance liquid chromatography (HPLC) analyzer. This allows for the acquisition of the peak area of ​​cefuroxime acid and the cumulative area of ​​adjacent impurity peaks at each sampling time. For example, when the sampling interval is 15 minutes, the retention time of cefuroxime acid in the chromatogram acquired at the 15th minute is 2.9 minutes, the peak area is 12500 mAU·s, and the cumulative area of ​​adjacent impurity peaks is 3200 mAU·s.

[0026] Furthermore, in order to convert the chromatographic peak area into physical concentration, the peak area of ​​cefuroxime acid needs to be substituted into a pre-calibrated concentration-area standard curve to obtain the real-time residual concentration of cefuroxime acid at that moment. For example, according to the standard curve y=2.5x+30 (y is the peak area unit mAU·s, x is the concentration unit mg / L), substituting the peak area of ​​12500mAU·s yields a concentration of 4988mg / L, and the concentrations at all times are arranged in chronological order to form a concentration sequence.

[0027] Furthermore, in order to characterize the relative level of byproducts, the relative abundance of impurities needs to be calculated based on the cumulative area of ​​adjacent impurity peak groups and the peak area of ​​cefuroxime acid at each sampling time. This can eliminate the absolute area deviation caused by fluctuations in the injection volume. For example, the relative abundance at the 15th minute is 3200 / 12500 = 0.256.

[0028] Furthermore, to obtain the dynamic generation rate of byproducts, the cumulative generation rate of byproducts needs to be calculated based on the difference in relative abundance between two adjacent sampling times and the time interval. This quantifies the generation speed of byproducts. For example, if the relative abundance is 0.256 at the 15th minute and 0.269 at the 30th minute, with an interval of 15 minutes (900 seconds), then the cumulative generation rate of byproducts = (0.269 - 0.256) / 900 ≈ 0.0000144 s. -1 .

[0029] Finally, to ensure that subsequent models can utilize both concentration and rate information simultaneously, the two sequences need to be aligned according to sampling time points to form a dual-channel input data stream. This guarantees that each time point contains both concentration and rate features. For example, at minutes 15, 30, and 45, the concentration sequence is [4988, 4770, 4550] mg / L, and the rate sequence is [0.0000144, 0.0000152, to be supplemented] s. -1 .

[0030] Step S100 in the method provided in this application embodiment further includes: Multiple cefuroxime acid standard solutions of different concentrations were prepared and injected into the high-performance liquid chromatograph (HPLC) analyzer, and the peak area of ​​the standard corresponding to each concentration was recorded. A linear regression was performed with the concentration of the cefuroxime acid standard solution as the x-axis and the corresponding peak area as the y-axis to obtain an initial concentration-area linear relationship. A blank reaction matrix sample with the same composition as the current reaction batch but without cefuroxime acid was collected. A cefuroxime acid standard of known concentration was added to the blank reaction matrix sample, and the recovered peak area was measured. Substituting the cefuroxime axetine standard into the initial concentration-area linear relationship, the theoretical peak area is calculated; based on the recovered chromatographic peak area and the theoretical peak area, a matrix effect correction factor is calculated; the slope in the initial concentration-area linear relationship is obtained, and combined with the matrix effect correction factor, the corrected slope is calculated; based on the corrected slope and the intercept in the initial concentration-area linear relationship, the concentration-area standard curve is generated, wherein the intercept is the ordinate value corresponding to the initial concentration-area linear relationship when the concentration is zero on the x-axis. Detailed explanation follows: In this embodiment, the blank reaction matrix sample refers to a reaction solution with the exact same composition as the raw materials in the current reaction batch but without the addition of cefuroxime acid. It contains solvents, catalysts, 7-ACA derivatives, 1-bromoethyl acetate, etc., and is used to simulate the real sample matrix environment. The recovered chromatographic peak area refers to the peak area of ​​cefuroxime acid measured by chromatographic analysis after adding a known concentration of cefuroxime acid standard to the blank reaction matrix. The theoretical peak area refers to the peak area calculated by directly substituting the added known concentration of standard into the initial concentration-area linear relationship; this area does not consider matrix interference. The corrected slope refers to the slope obtained by multiplying the slope of the initial linear relationship by a matrix effect correction factor, making the standard curve applicable to the actual reaction matrix.

[0031] First, to establish an initial concentration-area standard curve, at least five different concentrations of cefuroxime acid standard solutions (e.g., 100, 500, 1000, 1500, 2000, 3000 mg / L) need to be prepared. These solutions are injected into a high-performance liquid chromatograph (HPLC) analyzer, and the peak areas are recorded. A linear regression is then performed with concentration on the x-axis and peak area on the y-axis to obtain the initial linear relationship. ,For example =2.5, =30.

[0032] Furthermore, to evaluate the effect of the reaction matrix on the chromatographic response, a blank reaction matrix sample needs to be collected and cefuroxime acid standard at a known concentration added. The recovery peak area is then measured; for example, the recovery peak area is measured after adding 100 mg / L of the standard. =260mAU·s.

[0033] Furthermore, in order to obtain the matrix effect correction factor, the known concentration of the added material needs to be substituted into the initial linear relationship to calculate the theoretical peak area. =2.5×100+30=280mAU·s, then calculate the correction factor. =260 / 280≈0.9286.

[0034] Finally, to generate a concentration-area standard curve suitable for the actual reaction solution, the initial slope needs to be corrected using a correction factor to obtain k= ×f = 2.5 × 0.9286 = 2.3215, the intercept remains unchanged. =30, thus obtaining the corrected standard curve y=2.3215x+30, which is used for subsequent inversion calculation of residual concentration.

[0035] For example, taking the esterification reaction of a current batch as an example, at three sampling times (15 minutes, 30 minutes, and 45 minutes after the start of the reaction, the peak areas of cefuroxime acid obtained by high performance liquid chromatography were 12516 mAU·s, 11604 mAU·s, and 11102 mAU·s, respectively, and the cumulative areas of adjacent impurity peak groups were 3040 mAU·s, 2971 mAU·s, and 2994 mAU·s, respectively. Substituting into the calibrated concentration-area standard curve y=2.3215x+30, the concentrations obtained by inversion were 5378 mg / L, 4988 mg / L, and 4770 mg / L, respectively. The relative abundance of impurities at each time point was calculated as follows: 15 minutes: 3040 / 12516 ≈ 0.2429; 30 minutes: 2971 / 11604 ≈ 0.2560; 45 minutes: 2994 / 11102 ≈ 0.2697. The cumulative formation rate of byproducts in adjacent intervals was: 15-30 minutes: (0.2560-0.2429) / 900 ≈ 0.0000146 s.-1 Between the 30th and 45th minutes, (0.2697 - 0.2560) / 900 ≈ 0.0000152 seconds. -1 After aligning the concentration and rate at the three time points according to time, a dual-channel input data stream was obtained: the concentration at time point 15 was 5378 mg / L, and the rate was 0.0000146 s. -1 At 30 minutes, the corresponding concentration was 4988 mg / L and the rate was 0.0000152 s. -1 At the 45th minute, the concentration was 4770 mg / L, and the rate was yet to be determined, but a complete input sequence had been formed.

[0036] In summary, this step utilizes high-performance liquid chromatography (HPLC) to obtain the peak areas of cefuroxime acid and impurities at each sampling time. Combined with the matrix-effect corrected concentration-area standard curve and the calculation of impurity relative abundance differences, a dual-channel time series for concentration and byproduct rate is constructed. Compared with existing technologies, this step has the following advantages: First, the use of HPLC enables timed automatic monitoring of the reaction process, replacing intermittent offline sampling and obtaining dynamic information matching the reaction timescale. Second, the relative abundance of impurities eliminates the influence of fluctuations in injection volume and absolute peak area, making the byproduct rate calculation more robust. Third, the introduction of matrix effect correction avoids interference from reaction solution components on the chromatographic response, improves the accuracy of concentration inversion, and provides reliable dual-channel input data for subsequent models.

[0037] S200: The real-time residual concentration sequence of cefuroxime axetine and the byproduct accumulation rate sequence are synchronously input into a pre-trained dynamic response prediction model, and the corresponding byproduct growth inertia coefficient is output; In this embodiment, after obtaining the dual-channel time series of cefuroxime axetine residual concentration and by-product accumulation rate using a high-performance liquid chromatography (HPLC) analyzer, it is necessary to effectively extract quantitative information related to the by-product growth trend after supplementation. However, existing technologies typically rely solely on the current single residual concentration value or simple trend extrapolation for judgment, failing to comprehensively utilize the historical change patterns of the dual sequences to predict the inertial growth degree of by-products after supplementation. This lack of a unified, quantitative by-product growth trend indicator results in a lack of data-driven basis for subsequent supplementation decisions.

[0038] Step S200 in the method provided in this application embodiment includes: Timed monitoring data were collected from multiple historical reaction batches after the supplementary addition operation during the esterification reaction. The synchronized cefuroxime axetine residual concentration sequence and byproduct accumulation rate sequence were extracted to form a complete input sample. The byproduct growth inertia coefficient corresponding to each complete input sample was used as the output label, where the byproduct growth inertia coefficient is defined as the ratio of the byproduct accumulation rate at the time of supplementation completion to the byproduct accumulation rate at a future stable time. Simultaneously, the residual concentration values ​​and byproduct rate values ​​at several consecutive sampling times within the intermediate time period in the complete input sample were masked to generate a masked input sample. The masked input sample serves as the query input. The residual concentration sequence and byproduct rate sequence corresponding to the unmasked portion of the complete input sample are used as key references. The output label corresponding to the complete input sample is used as the supervision signal to construct a pre-training sample set. A network architecture for a dynamic response prediction network is constructed, comprising a temporal coding layer, a cross-correlation layer, and an inertia regression layer. The absolute error between the predicted byproduct growth inertia coefficient and the output label is used as the loss function. The network weights of the network architecture are updated using a backpropagation algorithm until the loss function converges. The trained network architecture is then used as the dynamic response prediction model. Detailed explanation follows: In this embodiment, the dynamic response prediction model refers to a time-series prediction network based on deep learning. The input is a dual-channel time-series sequence, and the output is a byproduct growth inertia coefficient. The network adopts an encoding-cross-correlation-regression architecture and is obtained through pre-training on historical data. The byproduct growth inertia coefficient is defined as the ratio of the cumulative byproduct generation rate at the time of replenishment completion to the cumulative byproduct generation rate at a future stable time. The future stable time is defined as the time when the relative change rate of the cumulative byproduct generation rate within three consecutive sampling intervals is less than a preset threshold of 1%. A coefficient greater than 1 indicates accelerated byproduct growth after replenishment, less than 1 indicates gradual decay, and equal to 1 indicates stability. The complete input sample refers to a dual sequence of residual concentration and byproduct rate extracted from historical reaction batches, containing a complete time range after the replenishment operation, with the sequence length covering from the completion of replenishment to the termination of the reaction. The masked input sample refers to randomly selecting a continuous intermediate interval from the complete input sample, replacing the concentration and rate values ​​at all times within that interval with specific masking markers, forcing the model to infer information from the masked portion using the unmasked portion.

[0039] First, in order to construct supervised training data, it is necessary to collect timed monitoring data after the replenishment operation in multiple historical reaction batches, extract the synchronized cefuroxime axetine concentration sequence and byproduct rate sequence to form a complete input sample, and use the byproduct growth inertia coefficient corresponding to each sample as the output label to obtain paired datasets for model training. For example, 5,000 samples were extracted from 50 historical batches, and each sample contained double sequence data within 240 minutes (i.e. 4 hours) after the replenishment was completed to cover the entire esterification reaction duration range of 80 to 300 minutes.

[0040] Furthermore, in order to enhance the model's ability to resist interference from incomplete latest data in online applications, the concentration and rate values ​​of several consecutive sampling times in the middle of the time period in the complete input sample need to be masked simultaneously to generate masked input samples. This can simulate the situation where some data is not obtained in the actual scenario, such as masking the data of three consecutive sampling times (corresponding to 45 minutes), forcing the model to learn to infer the missing segments using the data before and after.

[0041] Furthermore, in order to construct a pre-training task, a pre-training sample set needs to be built using occluded input samples as query inputs, unoccluded parts as key-value references, and the output labels of complete samples as supervision signals. This forces the model to learn and predict the inertia coefficient of byproducts while inferring occlusion information. For example, after the model inputs occluded samples, the temporal coding layer extracts the features of the unoccluded parts, the cross-association layer calculates the matching degree between the query and the key value, and finally the regression layer outputs the inertia coefficient.

[0042] Furthermore, in order to achieve the above mapping, a network architecture containing a temporal coding layer, a cross-correlation layer, and an inertial regression layer needs to be constructed. The temporal coding layer extracts local features at each time step, the cross-correlation layer calculates the response matching degree between the occluded and unoccluded positions, and the inertial regression layer regresses the inertial coefficient based on the matching degree features, thus obtaining a complete deep network structure.

[0043] Finally, in order to optimize the network parameters, the absolute error between the predicted byproduct growth inertia coefficient and the output label should be used as the loss function. The network weights should be updated through the backpropagation algorithm until the loss converges. This will result in a dynamic response prediction model that can accurately predict the byproduct growth inertia coefficient. For example, after 100 rounds of training, the average absolute error on the validation set is reduced to less than 0.05.

[0044] For example, the pre-training configuration of the dynamic response prediction model is as follows: A Transformer-based temporal coding layer is used, containing two multi-head self-attention layers with a hidden dimension of 128 and 4 attention heads; the cross-association layer uses a scaled dot product attention mechanism, with the key-value dimension aligned to the query dimension at 128; the inertial regression layer consists of two fully connected layers with 64 and 1 hidden units respectively, and the activation function is ReLU. Training samples are drawn from 50 historical response batches. For each batch, 16 consecutive sampling times are extracted after replenishment (calculated at 15-minute sampling intervals, totaling 240 minutes), generating a total of 10,000 complete input samples, of which 80% are used for training and 20% for validation. The loss function is the mean absolute error, the optimizer is Adam, the initial learning rate is set to 0.001, the weight decay is 1e-5, and the learning rate decays to 0.5 every 10 rounds. The batch size is set to 32, the maximum number of training rounds is 50, and an early stopping strategy is adopted: training stops if the validation loss does not decrease for 5 consecutive rounds. After training, the model's mean absolute error on the validation set was 0.048, indicating a good fit.

[0045] For example, taking the esterification reaction of a current batch as an example, the concentration sequence [5378, 4988, 4770] mg / L and the rate sequence [0.0000144, 0.0000152, to be supplemented but assumed to be known] at three time points (15 minutes, 30 minutes, and 45 minutes) are input into a trained dynamic response prediction model. The model's internal temporal coding layer extracts features for each time point, the cross-correlation layer compares the matching degree between the current batch and historical supplementation patterns, and the inertia regression layer outputs a byproduct growth inertia coefficient of 0.85. This coefficient is less than 1, indicating that under the current trends of residual concentration and byproduct rate changes, even with supplementation, the cumulative byproduct generation rate is not expected to increase drastically, but rather show a slow decay trend, providing a core quantitative basis for the generation of subsequent supplementation correction amounts.

[0046] In summary, this step directly outputs the byproduct growth inertia coefficient from a dual-channel time series using a pre-trained deep learning model. Compared with existing technologies, this step has the following advantages: First, it utilizes deep networks to extract complex time series patterns from historical data, solving the problem that traditional methods struggle to quantitatively predict byproduct growth trends from dual sequences; second, the masking pre-training strategy enhances the model's robustness to incomplete input data, making it more suitable for real-world online application scenarios; and third, the output inertia coefficient, as a unified quantitative indicator, provides a data-driven core parameter for subsequent supplementary decisions, avoiding the subjectivity of empirical judgments.

[0047] S300: Based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime acid, and combined with the comprehensive decay rate, construct a residual concentration prediction decay path.

[0048] In this embodiment, after obtaining the byproduct growth inertia coefficient, it is necessary to construct a predictive path that can quantitatively describe the future residual concentration decay trend over time based on the actual decay dynamics of the current reaction system. However, existing technologies typically rely solely on empirical judgment based on a single current concentration value or employ a fixed decay model, lacking a method for dynamic and adaptive path construction that combines real-time logarithmic decay rate, historical benchmarks, and inertia coefficients. This results in the inability to provide reliable quantitative basis for determining the subsequent reaction termination time.

[0049] Step S300 in the method provided in this application embodiment includes: The residual concentration values ​​of the latest three consecutive sampling times are extracted from the real-time residual concentration sequence of cefuroxime acid, and denoted as the first concentration value, the second concentration value, and the third concentration value, respectively. The timestamps of the corresponding three sampling times are obtained. The first logarithmic decay rate between the first and second concentration values, and the second logarithmic decay rate between the second and third concentration values ​​are calculated. Based on the first and second logarithmic decay rates, the effective decay rate of the current reaction system is calculated. A baseline decay rate constant obtained statistically from multiple historical reaction batches is obtained, wherein the baseline decay rate constant is the decay order of each historical batch after the replenishment operation. The arithmetic mean of the decay rates calculated from the first three sampling points is used. The byproduct growth inertia coefficient output by the dynamic response prediction model is used as a correction factor, combined with the baseline decay rate constant, to obtain the corrected decay rate constant. Based on the corrected decay rate constant and the effective decay rate, the comprehensive decay rate is calculated. Starting from the latest residual concentration value and using the comprehensive decay rate as the exponential decay parameter, an exponential decay curve of the residual concentration over time is constructed. The value of the exponential decay curve over the time interval is the estimated decay path of the residual concentration, and the time interval is determined by the maximum reaction time allowed by the reaction process. Detailed explanation follows: In this embodiment, the latest residual concentration value refers to the third concentration value among the last three consecutive sampling times in the real-time residual concentration sequence of cefuroxime axetine, i.e., the third concentration value, which represents the residual concentration at the current time. The logarithmic decay rate is the ratio of the natural logarithmic difference of the residual concentrations at adjacent times to the time interval, used to characterize the instantaneous rate of exponential decay. The first logarithmic decay rate is calculated from the first and second concentration values, and the second logarithmic decay rate is calculated from the second and third concentration values. The effective decay rate is the weighted average of the two latest logarithmic decay rates, with the second logarithmic decay rate closer to the current time given a greater weight. The baseline decay rate constant is the average of the initial decay rates of the decay stage obtained statistically from historical batches, reflecting the typical decay capacity under the process conditions. The corrected decay rate constant is the baseline decay rate constant multiplied by the byproduct growth inertia coefficient, adjusting the decay rate according to the expected byproduct growth; a coefficient greater than 1 accelerates the decay rate, and vice versa. The comprehensive decay rate is the weighted average of the effective decay rate and the corrected decay rate constant, with a preset fixed weight used to balance the current actual decay with historical experience.

[0050] First, to obtain the latest decay status of the current reaction, the concentration values ​​and corresponding timestamps of the three most recent consecutive sampling times need to be extracted from the residual concentration sequence. This yields raw data reflecting the recent decay trend. For example, the third concentration value of 4770 mg / L corresponds to the 45th minute (2700 s), the second concentration value of 4988 mg / L corresponds to the 30th minute (1800 s), and the first concentration value of 5378 mg / L corresponds to the 15th minute (900 s).

[0051] Furthermore, to quantify the instantaneous rate of concentration decay, the first logarithmic decay rate and the second logarithmic decay rate need to be calculated separately. For example: First logarithmic decay rate = 0.0000838s -1 Second logarithmic decay rate = 0.0000497s -1 .

[0052] Furthermore, to obtain an effective decay rate closer to the current state, a weighted average of the two logarithmic decay rates needs to be calculated. The first weight is set to 1, and the second weight to 2. The effective decay rate is then calculated as: (0.0000838 × 1 + 0.0000497 × 2) / (1 + 2) = 0.0000611 s. -1 .

[0053] Furthermore, in order to incorporate historical experience for correction, a baseline decay rate constant needs to be obtained. Let's assume the baseline value obtained from 50 historical batches is 0.00007s. -1The corrected decay rate constant was calculated using the byproduct growth inertia coefficient of 0.85 from the model output: 0.00007 × 0.85 = 0.0000595 s. -1 .

[0054] Furthermore, to integrate current realities and historical experience, the overall attenuation rate needs to be calculated. The effective attenuation rate weight is set at 0.4, and the corrected attenuation rate constant weight is set at 0.6. Therefore, the overall attenuation rate = 0.0000611 × 0.4 + 0.0000595 × 0.6 = 0.0000244 + 0.0000357 ≈ 0.0000601 s -1 .

[0055] Finally, to construct a complete decay curve, the third concentration value of 4770 mg / L should be used as the starting point, with a comprehensive decay rate of 0.0000601 s. -1 As the exponential decay parameter, an exponential decay curve was constructed within the maximum reaction time of 180 minutes (i.e., 10800 s). The residual concentration C(t) corresponding to any time t is 4770 × This curve represents the predicted decay path of the residual concentration.

[0056] For example, taking the esterification reaction of a current batch as an example, the concentration is 5378 mg / L at 15 minutes (900 s), 4988 mg / L at 30 minutes (1800 s), and 4770 mg / L at 45 minutes (2700 s). The first logarithmic decay rate is calculated to be 0.0000838 s. -1 The second logarithmic decay rate is 0.0000497s. -1 The effective decay rate is 0.0000611s. -1 The baseline decay rate constant is taken as the historical average of 0.00007 s. -1 The byproduct growth inertia coefficient is 0.85, and the corrected decay rate constant is 0.0000595 s. -1 The combined decay rate is taken as the weighted average of the two values, which is 0.0000601s. -1 Starting at the 45th minute (2700s), and setting the maximum reaction time to 180 minutes (10800s), the estimated decay path of the residual concentration is C(t) = 4770 × For example, at 120 minutes (7200 seconds), the estimated residual concentration = 4770 × The concentration was approximately 3639 mg / L, indicating that the residual concentration was expected to decrease to 3639 mg / L after 75 minutes.

[0057] In summary, this step extracts the residual concentrations from the three most recent consecutive time points, calculates the logarithmic decay rate, and weights it to obtain the effective decay rate. This is then corrected using historical benchmarks and byproduct inertia coefficients to construct a personalized exponential decay path. Compared to existing technologies, this step has the following advantages: First, it uses the concentration changes from the three most recent time points to calculate the logarithmic decay rate, achieving real-time and quantitative capture of the current decay trend; second, the weighted averaging method makes the data closer to the current time point contribute more, improving the sensitivity of the effective decay rate to the latest state; third, it integrates the actual effective decay rate with historical correction experience, ensuring that the decay path reflects both the individual characteristics of the current batch and incorporates statistical regularities, providing an accurate dynamic benchmark for calculating the termination time of the target reaction.

[0058] S400: Based on the intersection of the estimated decay path of the residual concentration and the standard residual concentration threshold, calculate the target reaction termination time, and combine it with the byproduct growth inertia coefficient to generate a dynamic replenishment correction amount of catalyst or esterification reagent.

[0059] In this embodiment, after obtaining the estimated decay path of the residual concentration and the inertia coefficient of by-product growth, it is necessary to determine an optimal amount of catalyst or esterification reagent to be added. However, in the prior art, the addition scheme is usually fixed or adjusted only based on experience, lacking a comprehensive quantitative consideration of the remaining reaction time, the difference between the current decay rate and the target decay rate, and the risk of inertial growth of by-products. This often leads to insufficient addition, which prolongs the reaction time, or excessive addition, which causes a surge in by-products.

[0060] It should be noted that in the esterification reaction described in this application, the catalyst and the esterification reagent are two essential materials that are added synergistically, and there is a predetermined ratio between them in the process. In this embodiment, the catalyst addition correction amount is used as the core control object, and the addition amount of the esterification reagent is adjusted synchronously with the catalyst correction amount according to the predetermined ratio. All subsequent fine addition control parameters generated based on the addition correction amount include instructions for the synchronous addition of both, and the synergistic addition is executed by the addition device.

[0061] Step S400 in the method provided in this application embodiment includes: The coordinates of the intersection point between the decay curve and the standard residual concentration threshold are read from the estimated decay path of the residual concentration, and the time corresponding to the intersection point is taken as the target reaction termination time. The current sampling time is obtained, and the remaining reaction time is calculated in combination with the target reaction termination time. Based on the corrected decay rate constant and the effective decay rate, the decay rate increase requirement factor is calculated, wherein the increase requirement factor is greater than one when the corrected decay rate constant is greater than the effective decay rate, and less than one otherwise. The supplementation intensity factor is determined according to the decay rate increase requirement factor and the by-product growth inertia coefficient, wherein the by-product growth inertia coefficient has a reverse adjustment effect on the supplementation intensity. Based on the unit supplementation dose corresponding to the baseline decay rate constant, combined with the supplementation intensity factor and the relative relationship between the remaining reaction time and the standard reaction time, a dynamic supplementation correction amount of catalyst or esterification reagent is generated. The unit supplementation dose corresponding to the baseline decay rate constant is the average catalyst supplementation amount required to increase the decay rate by one unit in historical batches, and the standard reaction time is the average time required for the residual concentration in historical batches to decay from the current average level to the standard residual concentration threshold. Detailed explanation is as follows: In this embodiment, the target reaction termination time refers to the moment corresponding to the intersection of the estimated residual concentration decay curve and the standard residual concentration threshold, used to determine the remaining reaction time to assess the urgency of supplementation. The decay rate increase requirement factor is defined as the ratio of the corrected decay rate constant to the effective decay rate; a value greater than 1 indicates that the current decay rate is lower than the target and needs to be increased. The supplementation intensity factor is defined as the increase requirement factor minus 1 divided by the byproduct growth inertia coefficient; a larger inertia coefficient results in a smaller supplementation intensity factor, reflecting a reverse adjustment effect. The unit supplementation dose refers to the average catalyst supplementation amount required to increase the decay rate by one unit in historical batches, serving as the basis for dimensional conversion. The standard reaction time refers to the average time required for the residual concentration in historical batches to decay from the current average level to the standard threshold, used to assess the relative urgency of the remaining time.

[0062] First, the target reaction termination time is obtained from the intersection of the predicted residual concentration decay curve and the standard residual concentration threshold. For example, curve C(t) = 4770 × The value intersects the threshold of 1500 mg / L at t≈11000s (approximately the 183rd minute), indicating that the target termination time is the 183rd minute.

[0063] Furthermore, at the current sampling time of 45 minutes (2700s), the remaining reaction time is calculated as 11000-2700=8300s (approximately 138 minutes).

[0064] Furthermore, based on the corrected decay rate constant of 0.0000595 and the effective decay rate of 0.0000611, the decay rate increase requirement factor is calculated as 0.0000595 / 0.0000611≈0.974. A value less than 1 indicates that the current decay rate is already faster than the target, and the increase needs to be reduced.

[0065] Furthermore, based on the demand multiple and the inertia coefficient of 0.85, the reinforcement coefficient is determined as (0.974-1)×(1 / 0.85)=-0.026×1.1765≈-0.0306.

[0066] Furthermore, based on the unit supplement dose of 10 mL, the supplement strength coefficient of -0.0306, and the remaining reaction time of 8300 s (approximately 138 minutes) and the standard reaction time of 9000 s (150 minutes), the dynamic supplement correction amount of catalyst or esterification reagent is calculated. The specific steps are detailed in the following content.

[0067] Step S400 in the method provided in this application embodiment further includes: Based on the unit supplement dose and the supplementation intensity coefficient, a base correction dose is calculated; a time urgency factor is determined according to the relative length of the remaining reaction time and the standard reaction time, wherein the shorter the remaining reaction time, the larger the time urgency factor; based on the base correction dose and the time urgency factor, a first intermediate correction amount is calculated; based on the relationship between the first intermediate correction amount and the unit supplement dose, an upper limit constraint is applied to the first intermediate correction amount to obtain a second intermediate correction amount; based on the degree to which the byproduct growth inertia coefficient deviates from one, a byproduct risk penalty factor is determined, wherein the greater the deviation, the larger the byproduct risk penalty factor; the second intermediate correction amount is adjusted for attenuation using the byproduct risk penalty factor to obtain the dynamic supplementation correction amount for the catalyst or esterification reagent. Detailed explanation follows: In this embodiment, the base correction dose refers to the product of the unit supplementary dose and the supplementary intensity factor, which is the initial correction amount without considering time urgency and risk penalty. The time urgency factor refers to the ratio of the standard reaction time to the remaining reaction time; the shorter the remaining time, the larger this factor, causing a corresponding increase in the supplementary dose to accelerate the reaction. The first intermediate correction amount refers to the value after multiplying the base correction dose by the time urgency factor. The second intermediate correction amount refers to the value after applying an upper limit constraint to the first intermediate correction amount to prevent the supplementary dose from exceeding the safe range. The byproduct risk penalty factor refers to the absolute difference between the byproduct growth inertia coefficient and 1, divided by the inertia coefficient; the larger the inertia coefficient deviates from 1, the larger this factor becomes, attenuating the supplementary dose to suppress byproduct risk.

[0068] First, based on the unit supplement dose of 10 mL and the supplementation strength coefficient of -0.0306, the base correction dose is calculated as 10 × (-0.0306) = -0.306 mL. Further, based on the remaining reaction time of 8300 s and the standard reaction time of 9000 s, the time urgency factor is calculated as 9000 / 8300 ≈ 1.084.

[0069] Furthermore, based on the baseline correction dose of -0.306 mL and the time urgency factor of 1.084, the first intermediate correction amount is calculated as -0.306 × 1.084 ≈ -0.332 mL. Further, an upper limit constraint is applied based on the relationship between the first intermediate correction amount and the unit supplementary dose, as detailed later. Further, based on the byproduct growth inertia coefficient of 0.85, the byproduct risk penalty factor is calculated as |0.85 - 1| / 0.85 = 0.15 / 0.85 = 0.1765. Finally, the second intermediate correction amount is adjusted for attenuation using the penalty factor, resulting in the dynamic supplementary correction amount for the catalyst or esterification reagent = second intermediate correction amount × (1 - 0.1765), thus completing the correction amount calculation.

[0070] Step S400 in the method provided in this application embodiment further includes: The ratio of the first intermediate correction amount to the unit supplementary dose is obtained and denoted as the correction multiple. When the correction multiple is less than or equal to a first multiple threshold, the first intermediate correction amount is used as the second intermediate correction amount. When the correction multiple is greater than the first multiple threshold and less than or equal to the second multiple threshold, the second intermediate correction amount is calculated based on the unit supplementary dose, combined with the first multiple threshold and the first intermediate correction amount. When the correction multiple is greater than the second multiple threshold, the unit supplementary dose is multiplied by the second multiple threshold to obtain the second intermediate correction amount. The first multiple threshold and the second multiple threshold are predetermined based on the distribution characteristics of supplementary correction amounts in historical reaction batches, and the first multiple threshold is less than the second multiple threshold. A detailed explanation follows: In this embodiment, the correction factor refers to the ratio of the first intermediate correction amount to the unit supplementary dose, used to determine the multiple of the correction amount relative to the baseline. The first and second correction factor thresholds are determined using the percentile method based on the distribution of supplementary correction factor ratios in historical batches, for example, using the 25th percentile as the first correction factor threshold and the 75th percentile as the second correction factor threshold. The superposition ratio coefficient is a constant greater than 0 and less than 1, used for partial superposition adjustment when the correction amount is between the first and second correction factor thresholds.

[0071] First, calculate the correction factor: First intermediate correction amount / Unit supplementary dose = -0.332 / 10 = -0.0332. Assume the first correction factor threshold is 0.5 and the second correction factor threshold is 1.5. Since the correction factor -0.0332 is less than the first correction factor threshold of 0.5, the second intermediate correction amount is directly taken as the first intermediate correction amount -0.0332 mL. Further, if the correction factor is between 0.5 and 1.5, a scaling factor will be used for calculation. For example, if the scaling factor is 0.6, then the second intermediate correction amount = (0.5 × 10 + (-0.0332 - 0.5 × 10) × 0.6). Finally, if the correction factor is greater than 1.5, multiply the unit supplementary dose by the second correction factor threshold of 1.5 to obtain the second intermediate correction amount = 15 mL.

[0072] For example, taking the esterification reaction of a current batch as an example, at the current sampling time of 45 minutes (2700s), the third concentration value is 4770mg / L, and the estimated decay path of the residual concentration is C(t) = 4770 × The standard residual concentration threshold is set at 1500 mg / L. At the intersection point, t ≈ 11000 s (minute 183), and the target reaction termination time is minute 183. The current sampling time is minute 45, with a remaining reaction time of 8300 s (approximately 138 minutes). The effective decay rate is 0.0000611 s. -1 The corrected decay rate constant is 0.0000595 s. -1 The attenuation rate increase factor is approximately 0.974 (0.0000595 / 0.0000611). The byproduct growth inertia coefficient is 0.85. The supplementation intensity factor is approximately -0.0306 (0.974-1)×(1 / 0.85). The unit supplementation dose is 10 mL (historical statistical value), and the standard reaction time is 9000 s (150 minutes). The calculated baseline correction dose is 10×(-0.0306)=-0.306 mL. The time urgency factor is approximately 1.084 (9000 / 8300). The first intermediate correction amount is approximately -0.306×1.084≈-0.332 mL. The correction factor is approximately -0.332 / 10=-0.0332. Assuming the first correction factor threshold is 0.5 and the second correction factor threshold is 1.5, the correction factor is less than 0.5, therefore the second intermediate correction amount is approximately -0.332 mL. Byproduct risk penalty factor = |0.85-1| / 0.85 = 0.1765, dynamic replenishment correction for catalyst or esterification reagent = -0.332×(1-0.1765) = -0.274mL. This negative value indicates that under the current reaction conditions, 0.274mL of catalyst should be reduced from the basic replenishment plan to avoid the risk of byproducts caused by excessive replenishment.

[0073] In summary, this step determines the remaining reaction time by the target reaction termination time, combines the decay rate to increase the demand factor, the byproduct growth inertia coefficient, the time urgency factor, and the byproduct risk penalty factor, and uses a multiplier threshold for upper limit constraint to generate a dynamic replenishment correction amount. Compared with existing technologies, this step has the following advantages: First, it introduces a comparison between the remaining reaction time and the standard reaction time to quantify the time urgency and avoid replenishment lag; second, the byproduct growth inertia coefficient has a reverse adjustment effect on the replenishment intensity, automatically reducing the replenishment amount when the byproduct inertia growth risk is high, thus suppressing impurity generation; third, the multiplier threshold constraint prevents the replenishment amount from abnormally deviating from the safe range, enhancing control robustness.

[0074] S500: Adjust the preset basic addition scheme according to the dynamic addition correction amount of the catalyst or esterification reagent to obtain fine addition control parameters including the correction addition dosage and the correction addition time.

[0075] In this embodiment, after calculating the dynamic replenishment correction amount of the catalyst or esterification reagent, it needs to be converted into a practically executable control parameter, i.e., simultaneously adjusting the replenishment dosage and timing. However, in the prior art, the replenishment scheme is usually completely fixed. Even if the replenishment dosage is adjusted based on experience, it is still executed at the original timing, resulting in a mismatch between the replenishment timing and the current reaction state, failing to fully realize the replenishment effect, and possibly even affecting reaction efficiency or increasing the risk of byproducts due to replenishment too early or too late. The process flow is as follows: Figure 2 As shown.

[0076] Step S500 in the method provided in this application embodiment includes: Obtain the baseline supplementation dose and baseline supplementation time in the baseline supplementation plan, wherein the baseline supplementation time is a fixed time offset calculated from the current sampling time; add the dynamic supplementation correction amount of the catalyst or esterification reagent to the baseline supplementation dose to obtain the corrected supplementation dose; calculate the supplementation time adjustment coefficient based on the ratio of the effective decay rate to the byproduct growth inertia coefficient, wherein the larger the byproduct growth inertia coefficient, the smaller the adjustment coefficient; calculate the corrected supplementation time based on the baseline supplementation time and the supplementation time adjustment coefficient; compare the corrected supplementation time with the minimum allowable supplementation interval; if the corrected supplementation time is less than the minimum allowable supplementation interval, then the minimum allowable supplementation interval is used as the final supplementation time; otherwise, the corrected supplementation time remains unchanged; combine the corrected supplementation dose and the final supplementation time as the fine supplementation control parameter.

[0077] In this embodiment, the basic replenishment scheme refers to fixed replenishment parameters pre-set based on process experience, including the basic replenishment dosage and the basic replenishment timing. The corrected replenishment dosage is the algebraic sum of the basic replenishment dosage and the dynamic replenishment correction amount of the catalyst or esterification reagent; a negative correction amount reduces the replenishment amount, while a positive value increases it. The replenishment timing adjustment coefficient is the ratio of the effective decay rate to the product of the corrected decay rate constant and the byproduct growth inertia coefficient, used to advance or postpone the replenishment execution timing based on the current decay state and byproduct risk. The minimum permissible replenishment interval refers to the shortest time to ensure sufficient mixing of the replenished material and reliable mechanical response; it is typically the larger of the mixing time and the mechanical delay, such as 300 seconds (5 minutes). The final replenishment timing refers to the actual replenishment time after adjustment and constraint by the minimum interval.

[0078] First, in order to obtain the baseline parameters, it is necessary to obtain the baseline supplement dose and baseline supplement time point in the baseline supplementation plan. For example, the baseline supplement dose is set to 10 mL, and the baseline supplementation time point is set to 60 minutes (3600 seconds) after the 45th minute (2700 seconds) from the current sampling time, that is, the 105th minute (6300 seconds).

[0079] Furthermore, in order to obtain the corrected dosage, the dynamic supplementation correction amount of the catalyst or esterification reagent needs to be added to the basic supplementation dosage. For example, if the correction amount is -0.274 mL, then the corrected supplementation dosage = 10 + (-0.274) = 9.726 mL.

[0080] Furthermore, in order to dynamically adjust the replenishment time, it is necessary to consider the effective decay rate of 0.0000611s. -1 The corrected decay rate constant is 0.0000595 s. -1 The adjustment factor for the supplementary time point is calculated using the by-product growth inertia coefficient of 0.85: 0.0000611 / (0.0000595×0.85) = 0.0000611 / 0.0000506≈1.207. The basic supplementary time point of 3600 seconds (60 minutes) is then multiplied by the adjustment factor to obtain the corrected supplementary time point: 3600×1.207≈4345 seconds (approximately 72.4 minutes). That is, the supplementary time point will be calculated from the 45th minute of the current time, 72.4 minutes later, corresponding to the 117.4th minute (7044 seconds).

[0081] Furthermore, to ensure operational safety, it is necessary to compare the corrected supplementary time point of 4345 seconds with the minimum allowable supplementary interval of 300 seconds (5 minutes). Since 4345 seconds is greater than 300 seconds, the final supplementary time point remains unchanged at 4345 seconds.

[0082] Finally, by combining the corrected supplemental dose of 9.726 mL with the final supplemental time of 4345 seconds, the fine supplemental control parameters were obtained as (9.726 mL, 4345 s).

[0083] For example, taking the esterification reaction of a current batch as an example, at the current sampling time of 45 minutes (2700 seconds), the dynamic replenishment correction amount of catalyst or esterification reagent is -0.274 mL. The basic replenishment plan is set to a basic replenishment dose of 10 mL, and the basic replenishment time is 60 minutes (3600 seconds) from the current time, i.e., the 105th minute. The corrected replenishment dose = 10 + (-0.274) = 9.726 mL. The replenishment time adjustment coefficient = effective decay rate 0.0000611 / (corrected decay rate constant 0.0000595 × inertia coefficient 0.85) = 0.0000611 / 0.0000506 ≈ 1.207. The corrected replenishment time = 3600 × 1.207 ≈ 4345 seconds (approximately 72.4 minutes). The minimum allowable supplementation interval is set to 300 seconds (5 minutes). Since 4345 is greater than 300, the final supplementation time is 4345 seconds. The fine supplementation control parameters are (9.726 mL, 4345 s), which means that the supplementation will be performed 72.4 minutes after the 45th minute from the current time, with a supplementation dose of 9.726 mL.

[0084] In summary, this step, by superimposing the dynamic replenishment correction amount of catalyst or esterification reagent with the basic replenishment scheme, not only adjusts the replenishment dosage but also dynamically optimizes the replenishment timing using a replenishment timing adjustment coefficient, and introduces a minimum permissible replenishment interval to prevent operational risks. Compared with existing technologies, this step has the following advantages: First, it achieves dual dynamic adjustment of replenishment dosage and timing, making the replenishment operation more closely aligned with the current reaction state and compensating for the defects of fixed-time replenishment being out of sync with the reaction; second, the replenishment timing adjustment coefficient integrates the current decay rate and the inertia of byproduct growth, and can automatically advance or delay the replenishment timing based on the magnitude of the inertia coefficient, further suppressing byproduct risks; third, the minimum permissible replenishment interval ensures the safety and reliability of the operation.

[0085] S600: Control the addition device to perform the additional addition of catalyst and esterification reagent to the esterification reactor according to the fine addition control parameters.

[0086] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: This application proposes a deep learning-based method for optimizing the process parameters of cefuroxime axetil preparation. First, the esterification reaction is monitored at preset sampling intervals (preferably 15-20 minutes). Real-time residual concentration sequences of cefuroxime axetil and corresponding byproduct accumulation rate sequences are obtained at multiple time-sequential sampling points. Then, a pre-trained deep learning model outputs the byproduct growth inertia coefficient from the dual sequences, and a residual concentration prediction decay path is constructed based on the latest residual concentration and the overall decay rate. Next, the target reaction termination time is calculated based on the intersection of the decay path and the standard residual concentration threshold, and a dynamic correction amount for catalyst or esterification reagent replenishment is generated using the inertia coefficient. Once the correction amount is determined, the basic replenishment scheme is adjusted to obtain fine-grained replenishment control parameters including the corrected replenishment dosage and timing. Finally, the replenishment device is controlled to perform the additional catalyst and esterification reagent replenishment operation. This method achieves timed sampling, dynamic prediction, and precise replenishment optimization of the cefuroxime axetil esterification process through dynamic response prediction and adaptive decay path construction.

[0087] Example 2, as shown in the appendix Figure 3 As shown, based on the inventive concept of the deep learning-based method for optimizing the preparation process parameters of cefuroxime axetil provided in Example 1, this application also provides a deep learning-based system for optimizing the preparation process parameters of cefuroxime axetil, specifically including: Sequence acquisition module 11 is used to acquire the real-time residual concentration sequence of cefuroxime axetine and the corresponding cumulative generation rate sequence of by-products collected at multiple consecutive sampling times in the current reaction batch in the esterification reactor. Model prediction module 12 is used to synchronously input the real-time residual concentration sequence of cefuroxime axetine and the by-product accumulation rate sequence into the pre-trained dynamic response prediction model and output the corresponding by-product growth inertia coefficient. Path construction module 13 is used to construct a residual concentration prediction decay path based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime axetine and the comprehensive decay rate. The correction generation module 14 is used to calculate the target reaction termination time based on the intersection of the estimated decay path of the residual concentration and the standard residual concentration threshold, and to generate a dynamic replenishment correction amount of catalyst or esterification reagent in combination with the byproduct growth inertia coefficient. The parameter adjustment module 15 is used to adjust the preset basic addition scheme according to the dynamic addition correction amount of the catalyst or esterification reagent, so as to obtain fine addition control parameters including the correction addition dose and the correction addition time. The replenishment control module 16 is used to control the replenishment device to perform additional operations of catalyst and esterification reagent to the esterification reactor according to the fine replenishment control parameters.

[0088] In one embodiment, module 11 is further configured to: acquire the real-time residual concentration sequence of cefuroxime acid and the corresponding cumulative generation rate sequence of by-products collected at multiple consecutive sampling times in the current reaction batch in the esterification reactor, including: collecting reaction liquid samples in the esterification reactor in the current reaction batch, and sequentially acquiring the chromatographic peak area of ​​cefuroxime acid and the cumulative area of ​​adjacent impurity peak groups at each sampling time using a high-performance liquid chromatography analyzer; substituting the chromatographic peak area of ​​cefuroxime acid into a pre-calibrated concentration-area standard curve to invert and obtain the real-time residual concentration of cefuroxime acid at each sampling time, and converting all sampling times into the real-time residual concentration of cefuroxime acid at each sampling time. The residual concentrations at each sampling time are arranged in chronological order to form a real-time residual concentration sequence of cefuroxime acid. Based on the cumulative area of ​​adjacent impurity peaks and the chromatographic peak area of ​​cefuroxime acid at each sampling time, the relative abundance of impurities is calculated. Based on the difference in relative abundance of impurities between two adjacent sampling times and the corresponding time interval, the cumulative generation rate of by-products in each sampling interval is calculated, and the rates corresponding to all sampling intervals are arranged in chronological order to form a cumulative generation rate sequence of by-products. The real-time residual concentration sequence of cefuroxime acid and the cumulative generation rate sequence of by-products are aligned according to the sampling time to form a dual-channel input data stream.

[0089] Furthermore, module 11 is also used for: preparing multiple cefuroxime acid standard solutions of different concentrations, injecting them into the high-performance liquid chromatograph, and recording the peak area of ​​the standard corresponding to each concentration; performing linear regression fitting with the concentration of the cefuroxime acid standard solution as the abscissa and the corresponding peak area of ​​the standard as the ordinate to obtain an initial concentration-area linear relationship; collecting a blank reaction matrix sample with the same composition as the current reaction batch but without cefuroxime acid, adding a known concentration of cefuroxime acid standard to the blank reaction matrix sample, and measuring the recovered peak area after addition; and so on. Substituting the known concentration of cefuroxime acid standard into the initial concentration-area linear relationship, the theoretical peak area is calculated; based on the recovered chromatographic peak area and the theoretical peak area, a matrix effect correction factor is calculated; the slope in the initial concentration-area linear relationship is obtained, and combined with the matrix effect correction factor, the corrected slope is calculated; based on the corrected slope and the intercept in the initial concentration-area linear relationship, the concentration-area standard curve is generated, wherein the intercept is the ordinate value corresponding to the initial concentration-area linear relationship when the concentration is zero on the abscissa.

[0090] In one embodiment, module 12 is further configured to: collect timed monitoring data from multiple historical reaction batches after the addition operation during the esterification reaction; extract the synchronized cefuroxime axetine residual concentration sequence and by-product cumulative generation rate sequence to form a complete input sample; use the by-product growth inertia coefficient corresponding to each complete input sample as an output label, wherein the by-product growth inertia coefficient is defined as the ratio of the by-product cumulative generation rate at the time of addition completion to the by-product cumulative generation rate at a future stable time; and simultaneously apply masking tags to the residual concentration values ​​and by-product rate values ​​at several consecutive sampling times in the middle time period of the complete input sample to generate... The input samples are masked; using the masked input samples as query input, and the residual concentration sequence and byproduct rate sequence corresponding to the unmasked portion of the complete input samples as key references, the output labels corresponding to the complete input samples are used as supervision signals to construct a pre-training sample set; a network architecture for a dynamic response prediction network is constructed, wherein the network architecture includes a temporal coding layer, a cross-correlation layer, and an inertial regression layer; using the absolute error between the predicted byproduct growth inertia coefficient and the output label as the loss function, the network weights of the network architecture are updated through a backpropagation algorithm until the loss function converges, and the trained network architecture is used as the dynamic response prediction model.

[0091] In one embodiment, module 13 is further configured to: extract the residual concentration values ​​of the latest three consecutive sampling times from the real-time residual concentration sequence of cefuroxime acid, denoted as the first concentration value, the second concentration value, and the third concentration value, respectively, and obtain the timestamps of the corresponding three sampling times; calculate the first logarithmic decay rate between the first concentration value and the second concentration value, and the second logarithmic decay rate between the second concentration value and the third concentration value; calculate the effective decay rate of the current reaction system based on the first logarithmic decay rate and the second logarithmic decay rate; and obtain the baseline decay rate constant statistically obtained from multiple historical reaction batches, wherein the baseline decay rate constant is the average residual concentration value of each historical batch at the time of sampling. The arithmetic mean of the decay rates calculated from the first three sampling points in the decay phase after the supplementary operation is obtained; the byproduct growth inertia coefficient output by the dynamic response prediction model is used as a correction factor, and combined with the baseline decay rate constant, the corrected decay rate constant is obtained; based on the corrected decay rate constant and the effective decay rate, the comprehensive decay rate is calculated; starting from the latest residual concentration value, and using the comprehensive decay rate as the exponential decay parameter, an exponential decay curve of the residual concentration over time is constructed, wherein the value of the exponential decay curve in the time interval is the estimated decay path of the residual concentration, and the time interval is determined by the maximum reaction time allowed by the reaction process.

[0092] In one embodiment, module 14 is further configured to: read the intersection coordinates of the decay curve and the standard residual concentration threshold from the estimated decay path of the residual concentration, and take the time corresponding to the intersection as the target reaction termination time; obtain the current sampling time, and calculate the remaining reaction time in combination with the target reaction termination time; calculate the decay rate increase requirement multiple based on the corrected decay rate constant and the effective decay rate, wherein the increase requirement multiple is greater than one when the corrected decay rate constant is greater than the effective decay rate, and less than one otherwise; determine the supplementation intensity coefficient according to the decay rate increase requirement multiple and the by-product growth inertia coefficient, wherein the by-product growth inertia coefficient has a reverse adjustment effect on the supplementation intensity; generate a dynamic supplementation correction amount of catalyst or esterification reagent based on the unit supplementation dose corresponding to the benchmark decay rate constant, combined with the supplementation intensity coefficient and the relative relationship between the remaining reaction time and the standard reaction time; the unit supplementation dose corresponding to the benchmark decay rate constant is the average catalyst supplementation amount required to increase the decay rate by one unit in historical batches, and the standard reaction time is the average time required for the residual concentration in historical batches to decay from the current average level to the standard residual concentration threshold.

[0093] Furthermore, module 14 is also used to: calculate a base correction dose based on the unit supplement dose and the supplementation intensity coefficient; determine a time urgency factor based on the relative length of the remaining reaction time and the standard reaction time, wherein the shorter the remaining reaction time, the larger the time urgency factor; calculate a first intermediate correction amount based on the base correction dose and the time urgency factor; impose an upper limit constraint on the first intermediate correction amount based on the relationship between the first intermediate correction amount and the unit supplement dose to obtain a second intermediate correction amount; determine a byproduct risk penalty factor based on the degree of deviation of the byproduct growth inertia coefficient from one, wherein the greater the deviation, the larger the byproduct risk penalty factor; and adjust the second intermediate correction amount by attenuation using the byproduct risk penalty factor to obtain a dynamic supplementation correction amount for the catalyst or esterification reagent.

[0094] Furthermore, module 14 is also configured to: obtain the ratio of the first intermediate correction amount to the unit supplementary dose, denoted as the correction multiple; when the correction multiple is less than or equal to a first multiple threshold, use the first intermediate correction amount as the second intermediate correction amount; when the correction multiple is greater than the first multiple threshold and less than or equal to the second multiple threshold, calculate the second intermediate correction amount based on the unit supplementary dose, combined with the first multiple threshold and the first intermediate correction amount; when the correction multiple is greater than the second multiple threshold, multiply the unit supplementary dose by the second multiple threshold to obtain the second intermediate correction amount; wherein, the first multiple threshold and the second multiple threshold are predetermined according to the distribution characteristics of supplementary correction amounts in historical reaction batches, and the first multiple threshold is less than the second multiple threshold.

[0095] In one embodiment, module 15 is further configured to: obtain the basic supplementation dose and basic supplementation time point in the basic supplementation plan, wherein the basic supplementation time point is a fixed time offset calculated from the current sampling time; add the dynamic supplementation correction amount of the catalyst or esterification reagent to the basic supplementation dose to obtain the corrected supplementation dose; calculate the supplementation time point adjustment coefficient according to the ratio of the effective decay rate to the byproduct growth inertia coefficient, wherein the larger the byproduct growth inertia coefficient, the smaller the adjustment coefficient; calculate the corrected supplementation time point based on the basic supplementation time point and the supplementation time point adjustment coefficient; compare the corrected supplementation time point with the minimum allowable supplementation interval, and if the corrected supplementation time point is less than the minimum allowable supplementation interval, then use the minimum allowable supplementation interval as the final supplementation time point, otherwise keep the corrected supplementation time point unchanged; combine the corrected supplementation dose and the final supplementation time point as the fine supplementation control parameter.

[0096] The deep learning-based cefuroxime axetil preparation process parameter optimization system provided in this application can achieve intelligent optimization and control of the esterification reaction process in scenarios such as cefuroxime axetil API production and quality control, pharmaceutical process parameter optimization and batch consistency improvement. This includes online monitoring data acquisition of the esterification reaction, prediction of byproduct growth inertia coefficients and construction of residual concentration decay paths, generation of supplementary correction amounts, and output of fine supplementary control parameters. It can be integrated into distributed control systems of pharmaceutical production lines or online monitoring platforms for laboratory reactors, effectively improving the prediction accuracy of cefuroxime axetil residual concentration decay trends and the quantitative assessment capability of byproduct growth risks. This reduces the risk of byproduct exceedances and raw material waste due to improper supplementation. Simultaneously, it provides structured process data including residual concentration sequences, byproduct accumulation rates, supplementary correction amounts, and reaction termination times for process optimization and batch consistency traceability. For specific optimization processes and control details of this system, please refer to Example 1.

[0097] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

Claims

1. A method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning, characterized in that, The method includes: The real-time residual concentration sequence of cefuroxime axetine and the corresponding cumulative formation rate sequence of by-products were obtained from multiple consecutive sampling times in the current batch of the esterification reactor. The real-time residual concentration sequence of cefuroxime axetine and the byproduct accumulation rate sequence are synchronously input into a pre-trained dynamic response prediction model, and the corresponding byproduct growth inertia coefficient is output. Based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime axetine, and combined with the comprehensive decay rate, a residual concentration prediction decay path is constructed. Based on the intersection of the estimated decay path of the residual concentration and the standard residual concentration threshold, the target reaction termination time is calculated, and combined with the byproduct growth inertia coefficient, a dynamic replenishment correction amount of catalyst or esterification reagent is generated. Based on the dynamic replenishment correction amount of the catalyst or esterification reagent, the preset basic replenishment scheme is offset and adjusted to obtain fine replenishment control parameters including the corrected replenishment dosage and the corrected replenishment timing. The addition device is controlled to perform the additional addition of catalyst and esterification reagent to the esterification reactor according to the fine addition control parameters.

2. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 1, characterized in that, Obtain the real-time residual concentration sequence of cefuroxime acid and the corresponding cumulative formation rate sequence of byproducts collected at multiple consecutive sampling times in the current reaction batch in the esterification reactor, including: Collect reaction liquid samples from the esterification reactor in the current batch of reaction, and obtain the peak area of ​​cefuroxime acid and the cumulative area of ​​adjacent impurity peaks at each sampling time using a high performance liquid chromatography analyzer. The peak area of ​​cefuroxime acid was substituted into the pre-calibrated concentration-area standard curve to obtain the real-time residual concentration of cefuroxime acid at each sampling time. The residual concentrations at all sampling times were arranged in chronological order to form a real-time residual concentration sequence of cefuroxime acid. The relative abundance of impurities is calculated based on the cumulative area of ​​adjacent impurity peak groups and the chromatographic peak area of ​​cefuroxime acid at each sampling time. Based on the difference in relative abundance of impurities between two adjacent sampling times and the corresponding time interval, the cumulative generation rate of by-products in each sampling interval is calculated, and the rates corresponding to all sampling intervals are arranged in chronological order to form a sequence of cumulative generation rates of by-products. The real-time residual concentration sequence of cefuroxime acid and the cumulative generation rate sequence of byproducts are aligned according to the sampling time to form a dual-channel input data stream.

3. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 2, characterized in that, The calibration steps for the concentration-area standard curve include: Multiple cefuroxime acid standard solutions of different concentrations were prepared and injected into the high performance liquid chromatograph, and the chromatographic peak area of ​​the standard corresponding to each concentration was recorded. Using the concentration of the cefuroxime acid standard solution as the abscissa and the corresponding chromatographic peak area of ​​the standard as the ordinate, a linear regression was performed to obtain the initial concentration-area linear relationship. Collect a blank reaction matrix sample with the same composition as the current reaction batch but without cefuroxime acid, add a cefuroxime acid standard of known concentration to the blank reaction matrix sample, and determine the recovery chromatographic peak area after addition; Substitute the known concentration of cefuroxime acid standard into the initial concentration-area linear relationship to calculate the theoretical peak area; Calculate the matrix effect correction factor based on the recovered chromatographic peak area and the theoretical peak area; Obtain the slope in the initial concentration-area linear relationship, and calculate the corrected slope by combining it with the matrix effect correction factor; Based on the corrected slope, and combined with the intercept in the initial concentration-area linear relationship, the concentration-area standard curve is generated, wherein the intercept is the ordinate value corresponding to the initial concentration-area linear relationship when the concentration is the zero point of the abscissa.

4. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 1, characterized in that, The pre-training steps of the dynamic response prediction model include: Timed monitoring data were collected from multiple historical reaction batches after the supplementary operation was performed during the esterification reaction. The cefuroxime axetine residual concentration sequence and by-product cumulative formation rate sequence were extracted after synchronization to form a complete input sample. The byproduct growth inertia coefficient corresponding to each complete input sample is used as the output label, wherein the byproduct growth inertia coefficient is defined as the ratio of the cumulative byproduct generation rate at the time of completion of supplementation to the cumulative byproduct generation rate at the future stable time. The residual concentration values ​​and byproduct rate values ​​at several consecutive sampling times in the middle time period of the complete input sample are simultaneously marked with a masking tag to generate a masked input sample; Using the masked input sample as the query input, the residual concentration sequence and byproduct rate sequence corresponding to the unmasked part of the complete input sample as key references, and the output label corresponding to the complete input sample as the supervision signal, a pre-training sample set is constructed. A network architecture for constructing a dynamic response prediction network is provided, wherein the network architecture includes a temporal coding layer, a cross-correlation layer, and an inertial regression layer; The absolute error between the predicted byproduct growth inertia coefficient and the output label is used as the loss function. The network weights of the network architecture are updated through the backpropagation algorithm until the loss function converges. The trained network architecture is then used as the dynamic response prediction model.

5. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 1, characterized in that, Based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime acid, and combined with the comprehensive decay rate, a residual concentration prediction decay path is constructed, including: Extract the residual concentration values ​​of the latest three consecutive sampling times from the real-time residual concentration sequence of cefuroxime axetine, and record them as the first concentration value, the second concentration value and the third concentration value respectively, and obtain the timestamps of the corresponding three sampling times; Calculate the first logarithmic decay rate between the first concentration value and the second concentration value, and the second logarithmic decay rate between the second concentration value and the third concentration value; Based on the first logarithmic decay rate and the second logarithmic decay rate, calculate the effective decay rate of the current reaction system; Obtain a baseline decay rate constant statistically derived from multiple historical reaction batches, wherein the baseline decay rate constant is the arithmetic mean of the decay rates calculated from the first three sampling points of each historical batch during the decay phase after the replenishment operation. The byproduct growth inertia coefficient output by the dynamic response prediction model is used as a correction factor, and combined with the baseline decay rate constant, the corrected decay rate constant is obtained. The overall attenuation rate is calculated based on the corrected attenuation rate constant and the effective attenuation rate. Starting from the latest residual concentration value, and using the comprehensive decay rate as the exponential decay parameter, an exponential decay curve of the residual concentration over time is constructed. The value of the exponential decay curve over the time interval is the estimated decay path of the residual concentration, and the time interval is determined by the maximum reaction time allowed by the reaction process.

6. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 1, characterized in that, Based on the intersection of the predicted decay path of the residual concentration and the standard residual concentration threshold, the target reaction termination time is calculated, and combined with the byproduct growth inertia coefficient, a dynamic replenishment correction amount for the catalyst or esterification reagent is generated, including: Read the coordinates of the intersection point between the decay curve and the standard residual concentration threshold from the estimated decay path of the residual concentration, and take the time corresponding to the intersection point as the target reaction termination time; Obtain the current sampling time and, in conjunction with the target reaction termination time, calculate the remaining reaction time; Based on the corrected decay rate constant and the effective decay rate, the decay rate improvement requirement multiple is calculated, wherein the improvement requirement multiple is greater than one when the corrected decay rate constant is greater than the effective decay rate, and less than one otherwise. The reinforcement coefficient is determined based on the attenuation rate increase demand multiple and the byproduct growth inertia coefficient, wherein the byproduct growth inertia coefficient has a reverse adjustment effect on the reinforcement. Based on the unit supplement dose corresponding to the baseline decay rate constant, combined with the supplement strength coefficient and the relative relationship between the remaining reaction time and the standard reaction time, a dynamic supplement correction amount for catalyst or esterification reagent is generated. The unit replenishment dose corresponding to the baseline decay rate constant is the average amount of catalyst replenishment required to increase the decay rate by one unit in historical batches, and the standard reaction time is the average time required for the residual concentration in historical batches to decay from the current average level to the standard residual concentration threshold.

7. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 6, characterized in that, Based on the unit supplement dose corresponding to the baseline decay rate constant, combined with the supplement strength coefficient and the relative relationship between the remaining reaction time and the standard reaction time, a dynamic supplement correction amount for the catalyst or esterification reagent is generated, including: The base correction dose is calculated based on the unit supplement dose and the supplementation strength coefficient; The time urgency factor is determined based on the relative length of the remaining reaction time and the standard reaction time, wherein the shorter the remaining reaction time, the greater the time urgency factor. Based on the baseline correction dose and the time urgency factor, calculate the first intermediate correction amount; Based on the relationship between the first intermediate correction amount and the unit supplementary dose, an upper limit constraint is applied to the first intermediate correction amount to obtain the second intermediate correction amount. The by-product risk penalty factor is determined based on the degree to which the by-product growth inertia coefficient deviates from one, wherein the greater the degree of deviation, the greater the by-product risk penalty factor. The second intermediate correction amount is attenuated and adjusted using the byproduct risk penalty factor to obtain the dynamic replenishment correction amount of the catalyst or esterification reagent.

8. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 7, characterized in that, Based on the relationship between the first intermediate correction amount and the unit supplementary dose, an upper limit constraint is applied to the first intermediate correction amount to obtain a second intermediate correction amount, including: Obtain the ratio of the first intermediate correction amount to the unit supplementary dose, and record it as the correction factor; When the correction factor is less than or equal to the first correction factor threshold, the first intermediate correction amount is used as the second intermediate correction amount; When the correction factor is greater than the first correction factor threshold and less than or equal to the second correction factor threshold, the second intermediate correction factor is calculated based on the unit supplementary dose, combined with the first correction factor threshold and the first intermediate correction factor. When the correction factor is greater than the second multiple threshold, the unit supplementary dose is multiplied by the second multiple threshold as the second intermediate correction amount; The first multiple threshold and the second multiple threshold are predetermined based on the distribution characteristics of the supplementary correction amount in historical reaction batches, and the first multiple threshold is less than the second multiple threshold.

9. The method for optimizing the preparation process parameters of cefuroxime axetil based on deep learning according to claim 1, characterized in that, Based on the dynamic replenishment correction amount of the catalyst or esterification reagent, the preset basic replenishment scheme is offset and adjusted to obtain fine replenishment control parameters including the corrected replenishment dosage and the corrected replenishment timing, including: Obtain the basic supplement dose and basic supplement time point in the basic supplement plan, wherein the basic supplement time point is a fixed time offset calculated from the current sampling time; The corrected dosage is obtained by adding the dynamic addition correction amount of the catalyst or esterification agent to the basic addition amount; The adjustment coefficient for the replenishment time is calculated based on the ratio of the effective decay rate to the byproduct growth inertia coefficient, wherein the larger the byproduct growth inertia coefficient, the smaller the adjustment coefficient. Based on the basic supplementary timing point and the supplementary timing point adjustment coefficient, the corrected supplementary timing point is calculated; Compare the corrected supplementation time point with the minimum allowable supplementation interval. If the corrected supplementation time point is less than the minimum allowable supplementation interval, then the minimum allowable supplementation interval is taken as the final supplementation time point; otherwise, the corrected supplementation time point remains unchanged. The corrected supplemental dose is combined with the final supplemental dose timing to form the fine supplemental dose control parameter.

10. A deep learning-based system for optimizing the process parameters of cefuroxime axetil preparation, characterized in that, The system is used to execute the deep learning-based method for optimizing the preparation process parameters of cefuroxime axetil according to any one of claims 1 to 9, the system comprising: The sequence acquisition module is used to acquire the real-time residual concentration sequence of cefuroxime axetine and the corresponding cumulative generation rate sequence of by-products collected at multiple consecutive sampling times in the current reaction batch in the esterification reactor. The model prediction module is used to synchronously input the real-time residual concentration sequence of cefuroxime axetine and the by-product accumulation rate sequence into the pre-trained dynamic response prediction model and output the corresponding by-product growth inertia coefficient. The path construction module is used to construct a residual concentration prediction decay path based on the latest residual concentration value in the real-time residual concentration sequence of cefuroxime axetine and the comprehensive decay rate. The correction generation module is used to calculate the target reaction termination time based on the intersection of the estimated decay path of the residual concentration and the standard residual concentration threshold, and to generate a dynamic replenishment correction amount of catalyst or esterification reagent in combination with the byproduct growth inertia coefficient. The parameter adjustment module is used to offset the preset basic addition scheme according to the dynamic addition correction amount of the catalyst or esterification reagent, and obtain fine addition control parameters including the correction addition dose and the correction addition time. The replenishment control module is used to control the replenishment device to perform additional operations of catalyst and esterification reagent to the esterification reactor according to the fine replenishment control parameters.

Images