Pharmaceutical industry management method and system based on industrial big data

Abstract

The application discloses a pharmaceutical industry management method and system based on industrial big data, relates to the technical field of industrial pharmaceutical management, and carries out unified structured processing on process parameters and end quality indexes collected in each batch of production through a data regularization and quality index standardization module, thereby not only enhancing the comparability between multiple batches of data, but also providing consistent and stable original data support for subsequent fluctuation analysis and modeling, and effectively overcoming the technical obstacles that existing process data is difficult to be used for horizontal analysis due to different sampling frequencies, different cycle lengths and non-uniform index scales. Through quality fluctuation amplitude calculation and large fluctuation batch screening mechanism, the system can not only identify quality abnormal batches exceeding the early warning line, but more importantly, can detect slight but continuous deviation trends that are not identified in the traditional qualified judgment system, which has great significance for preventing quality risk advance and avoiding lag response.

Description

Technical Field

[0001] This invention relates to the field of industrial pharmaceutical management technology, specifically to a pharmaceutical industrial management method and system based on industrial big data. Background Technology

[0002] Industrial big data technology is being gradually introduced to improve traditional management models that rely primarily on manual recording and static processes. Particularly in pharmaceutical manufacturing, automated platforms such as MES and SCADA have become core tools for data acquisition and process monitoring. Batch-level production records and process tracking information provide the data foundation for achieving end-to-end quality management. Against this backdrop, identifying potential quality anomaly trends from these batch-level process and endpoint quality data has become a key issue for pharmaceutical companies in quality assurance and process optimization.

[0003] In the existing quality control system of the pharmaceutical industry, although batch data is recorded, its main purpose is still focused on conformity assessment and post-event accountability, lacking the ability to dynamically model the deep correlation between process behavior and quality fluctuations between batches. In particular, slight quality deviations occurring in different batches of the same product usually do not trigger systemic intervention because they have not crossed the warning line. However, such fluctuations are very likely to be early signals of "suboptimal process segments".

[0004] The aforementioned problems typically occur at a certain stage of multiple batches of operation, exhibiting similar but subtle abnormal process states, such as excessively high initial reaction temperature, slightly unstable stirring rate, or delayed pH adjustment response. These abnormalities do not immediately lead to non-compliance, but after accumulating over a period of time, they will manifest as a systematic shift or increased fluctuation in a certain indicator in the final product quality. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a pharmaceutical industry management method and system based on industrial big data, which solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a pharmaceutical industrial management system based on industrial big data, including a data regularization and quality index standardization module, a batch-to-batch quality index fluctuation identification module, a process parameter residual extraction module for fluctuating batches, a process parameter and quality fluctuation factor construction module, and a suboptimal segment identification and offset risk score calculation module. The data normalization and quality index standardization module collects the process parameters Gcs and the endpoint quality index Gzx of the same batch of drugs, and performs preprocessing to obtain the drug dataset YW; The batch-to-batch quality index fluctuation identification module extracts features from the drug dataset YW, extracts the endpoint quality index Gzx, calculates and obtains the quality fluctuation amplitude Zfu, and filters out the batch subset Bva with large fluctuations. The process parameter residual extraction module for fluctuating batches performs in-depth analysis of the process parameters Gcs in the large fluctuating batch subset Bva, and calculates and obtains the offset residual PR of each parameter. The process parameter and quality fluctuation factor construction module matches the correlation between offset residual PR and quality fluctuation amplitude Zfu, and calculates and obtains the standardized correlation factor H; The suboptimal segment identification and offset risk score calculation module quantifies the risk level of each parameter based on the standardized relevant factor H, obtains the comprehensive offset risk score DF, and provides feedback.

[0007] Preferably, the data regularization and quality index standardization module includes a raw data fitting and acquisition unit and a structure mapping-driven preprocessing unit; The raw data fitting and acquisition unit collects the process parameters Gcs and the endpoint quality index Gzx through sensors and acquisition settings, and fits them into the raw dataset CW; the process parameters Gcs include reaction temperature Trx, pH value pH, and stirring speed v; The reaction temperature Trx was obtained using an industrial thermocouple. pH value was obtained by online pH electrode sampling; The stirring speed v is acquired through the feedback interface of the servo motor encoder and the frequency converter. The endpoint quality indicator Gzx includes the content of active ingredient Ca, purity Pu, and residual moisture Sr; The content of active ingredient Ca was obtained by high-performance liquid chromatography (HPLC). Specifically, a finished product sample was collected after each batch of production was completed. After sample pretreatment, the sample was sent to the HPLC system. The content of the target ingredient was calculated based on the retention time and peak area. Purity Pu was obtained by gas chromatography; specifically, the purity was calculated by analyzing the proportion of impurity peak areas and then determining the proportion of the main component. Residual moisture (Sr) was collected using an infrared moisture analyzer. Specifically, the finished drug product was sampled, and trace moisture was determined using the coulometric method. The result was then converted to a percentage mass concentration.

[0008] Preferably, the structure mapping-driven preprocessing unit cleans and normalizes the original dataset CW to obtain the drug dataset YW; Cleaning includes multi-scale structural anomaly removal and dynamic structural offset correction; Multi-scale structural anomaly removal is performed by processing the original CW dataset using the sliding perturbation standard deviation method to eliminate anomalous perturbation points. The processing method is as follows: First, set a sliding time window of length w, and take w consecutive sampling points starting from time point t; collect the parameter value sequence within the time period, average the data of the w sampling points to obtain the baseline level within the sliding time window, compare the data of each sampling point with the baseline level to obtain the squared deviation; finally, calculate the average of all squared deviations and take the square root to obtain the degree of abnormal fluctuation. When the degree of abnormal fluctuation exceeds the preset fluctuation threshold, it is identified as an abnormal disturbance point and is removed. The structural dynamic offset correction uses a dynamic time normalization function to uniformly map the production time of different batches; The processing method is as follows: First, obtain the total runtime of the current batch; process each time t, divide the current time t by the total runtime to obtain a normalized time position λ, whose value ranges from 0 to 1; re-mark the parameter values ​​at the original time t as the time-normalized parameter values. The normalization formula is as follows: ; In the formula, YWo represents the o-th data in the drug dataset YW, CWo represents the o-th data in the original dataset CW, minCWo represents the valley value of the o-th data in the original dataset CW, and maxCWo represents the peak value of the o-th data in the original dataset CW.

[0009] Preferably, the batch-to-batch quality index fluctuation identification module includes an endpoint quality index extraction and construction unit and a batch fluctuation amplitude calculation unit; The endpoint quality index extraction construction unit performs feature extraction on the drug dataset YW, and extracts the endpoint quality index Gzx, including active ingredient content Ca, purity Pu and residual moisture Sr; The final quality index Gzx of each batch is constructed into a batch feature vector to obtain the quality index matrix Q; The quality indicator matrix Q is obtained using the following formula: ; In the formula, rows represent each batch, and columns represent each quality index q∈{Ca,Pu,Sr}.

[0010] Preferably, the batch fluctuation amplitude calculation unit constructs a fluctuation function for each quality index q∈{Ca,Pu,Sr} in each column of the quality index matrix Q, and calculates the quality fluctuation amplitude Zfu of each batch under each index through a multi-level fluctuation measurement function. For each quality index q, the deviation value of the i-th batch is defined as Δq(i); The formula is: ; In the formula, Qq(i) represents the actual value of quality index q in the i-th batch, and pQq represents the average value of quality index q in all batches; The quality fluctuation range Zfu is obtained using the following formula: ; In the formula, Zfu(i) represents the quality fluctuation range of the i-th batch, Ca(i) represents the content of active ingredients in the i-th batch, Pu(i) represents the purity of the i-th batch, and Sr(i) represents the residual moisture in the i-th batch. The obtained quality fluctuation amplitude Zfu is compared with the preset fluctuation screening threshold Tfu to obtain the large fluctuation batch subset Bva; the formula is as follows: Bva={i|Zfu(i)>Tfu}.

[0011] Preferably, the process parameter residual extraction module for fluctuating batches includes a window feature aggregation unit and a window feature aggregation unit; The window feature aggregation unit performs in-depth analysis of the process parameters Gcs in the large fluctuation batch subset Bva, and performs sliding feature extraction to extract the average reaction temperature PTrx, average pH value PpH, and average stirring speed Pv, which are then fitted into a process parameter feature set GS={PTrx, PpH, Pv}. The parameter offset residual calculation unit analyzes the process parameter feature set GS of each fluctuating batch and compares it with the preset reference batch process feature mean set Sref to calculate and obtain the parameter residual CR. CRj = |GSj(i) - Srefj|; In the formula, CRj represents the parameter residual of the j-th process parameter, GSj(i) represents the statistical characteristic value of the j-th process parameter in the i-th batch, and j∈{Trx, pH, v}; The acquired parameter residual CR is compared with the preset recognition threshold Tj to obtain the offset residual PR; The offset residual PR is obtained as follows: ; In the formula, PRj(i) represents the offset residual of the j-th process parameter in the i-th batch.

[0012] Preferably, the process parameter and quality fluctuation factor construction module includes a fluctuation value recombination and alignment unit and a causal sensitivity output unit; The fluctuation value recombination and alignment unit aligns the acquired offset residuals PR with the quality fluctuation amplitudes Zfu according to the batch number, forming a joint data structure and constructing a two-dimensional residual fluctuation joint dataset ELD={PRj(i),Zfu(i)}, i∈Bva, j∈Gcs; Each row represents the relationship between the structural offset of the j-th process parameter in the i-th batch and the overall quality fluctuation of the batch. The causal sensitivity output unit standardizes the data in the two-dimensional residual fluctuation joint dataset ELD and uses the Pearson standardized correlation coefficient formula to calculate the statistical correlation strength between each process parameter and quality fluctuation, and obtains the standardized correlation factor H. Statistical feature processing was performed on all offset residuals PR and quality fluctuation amplitude Zfu to obtain the mean offset residual μPR, the standard deviation of offset residual σPR, the mean quality fluctuation amplitude μZfu, and the standard deviation of quality fluctuation amplitude σZfu. The standardized correlation factor H is obtained using the following formula: ; In the formula, Hj represents the standardized correlation factor between the j-th process parameter and the quality fluctuation range, μPRj represents the mean of the offset residuals of the j-th process parameter, and σPRj represents the standard deviation of the offset residuals of the j-th process parameter. The relationship between process parameters and quality fluctuations is analyzed using a standardized correlation factor H, as follows: When 0 < standardized correlation factor H < 1, it indicates that the deviation trend of process parameters is strongly positively correlated with quality fluctuation. When -1 < standardized correlation factor H < 0, it indicates that the deviation of process parameters is negatively correlated with quality fluctuation; When the standardized correlation factor H=0, it indicates that there is no linear relationship between the two.

[0013] Preferably, the suboptimal segment identification and offset risk score calculation module includes a risk factor-driven offset intensity fusion unit and a risk segment dynamic judgment unit; The risk factor-driven offset intensity fusion unit uses the obtained standardized correlation factor H as a sensitive factor for the degree of influence of process parameters, performs risk amplification fusion calculation on the offset residual PR of each fluctuating batch, and obtains the comprehensive offset risk score DF. The overall offset risk score (DF) is obtained using the following formula: ; In the formula, DF(i) represents the comprehensive deviation risk score of the i-th batch, and m represents the number of process parameters involved in the evaluation.

[0014] Preferably, the risk segment dynamic judgment unit compares the comprehensive offset risk score DF with the risk threshold Tdf to determine whether the batch falls into a high-risk offset segment; The risk threshold Tdf is obtained using the following formula: Tdf = μDF + κ × σDF; In the formula, μDF represents the mean of the comprehensive deviation risk score, and σDF represents the standard deviation of the comprehensive deviation risk score; The method for identifying high-risk deviation zones is as follows: When the overall offset risk score DF ≤ risk threshold Tdf, it indicates that the batch is normal and in a safe area; When the comprehensive offset risk score DF is greater than the risk threshold Tdf, it indicates that the batch is abnormal, is in a high-risk offset zone, and is marked.

[0015] A pharmaceutical industry management method based on industrial big data includes the following steps: Step 1: The data preparation and quality index standardization module collects the process parameters Gcs and the endpoint quality index Gzx of the same batch of drugs, and performs preprocessing to obtain the drug dataset YW. Step 2: The batch-to-batch quality index fluctuation identification module extracts features from the drug dataset YW, extracts the endpoint quality index Gzx, calculates and obtains the quality fluctuation amplitude Zfu, and filters out the batch subset Bva with large fluctuations. Step 3: The process parameter residual extraction module for fluctuating batches performs in-depth analysis of the process parameters Gcs in the large fluctuating batch subset Bva, and calculates and obtains the offset residual PR of each parameter. Step 4: The process parameter and quality fluctuation factor construction module matches the correlation between the offset residual PR and the quality fluctuation amplitude Zfu, and calculates and obtains the standardized correlation factor H. Step 5: The suboptimal segment identification and offset risk score calculation module quantifies the risk level of each parameter based on the standardized relevant factor H, obtains the comprehensive offset risk score DF, and provides feedback.

[0016] This invention provides a pharmaceutical industry management method and system based on industrial big data, which has the following beneficial effects: (1) During system operation, the process parameters and endpoint quality indicators collected in each batch of production are uniformly structured through the data regularization and quality index standardization module. This not only enhances the comparability between multiple batches of data, but also provides strong consistency and high stability of original data support for subsequent fluctuation analysis and modeling. This effectively overcomes the technical obstacles that existing process data is difficult to use for horizontal analysis due to different sampling frequencies, different cycle lengths, and inconsistent index scales.

[0017] By calculating the magnitude of quality fluctuations and using a large-fluctuation batch screening mechanism, the system can not only identify batches with abnormal quality that exceed the warning line, but more importantly, it can detect "slight but persistent" deviation trends that are not identified in the traditional conformity assessment system. This is of great significance for preventing quality risks from shifting forward and avoiding delayed responses, and it solves the technical blind spot of traditional methods that only look at whether the product is conforming and do not look at trend deviations.

[0018] (2) By extracting the endpoint quality indicators for each batch, a standardized batch quality indicator matrix is ​​constructed, breaking the traditional batch-based, isolated information state in quality records and incorporating the quality performance of all batches into the same analytical framework. This approach enhances the insight into product fluctuation trends and helps achieve cross-batch quality consistency tracking and quality deviation distribution identification. By calculating the relative deviation degree of batches and the joint deviation trend of indicators, batches that, although not exceeding quality limits, have shown signs of systemic fluctuations can be effectively identified. These "potentially risky batches" are often overlooked in traditional judgment systems, but this method allows them to be identified in advance, enhancing the foresight and accuracy of process fluctuation management.

[0019] (3) By using the fluctuation value recombination and alignment unit, the process parameter offset residuals PR of each batch and their corresponding quality fluctuation amplitudes Zfu are mapped one-to-one under the batch number and structurally integrated to construct a unified two-dimensional residual-fluctuation dataset. This design breaks the limitation of the decentralized management of process control data and quality inspection data in the traditional system, and for the first time realizes the transformation from single variable trend analysis to the mining of three-dimensional structural relationships at the batch level, indicator level, and parameter level.

[0020] By employing standardized processing methods, and under the premise of eliminating dimensional interference and amplifying the essence of fluctuations, a statistical correlation model between process parameters and quality fluctuations was constructed using Pearson correlation analysis. Unlike conventional experience-based judgments or subjective screening, this method can output clear and rankable correlation factor values ​​H, providing enterprises with a clear priority for key process parameters, facilitating the focus on core intervention points, and improving the pertinence and effectiveness of process control.

[0021] (4) The risk threshold is dynamically calculated from the mean and standard deviation of the offset risk score, making the risk judgment statistically reasonable and dynamically adaptable, avoiding one-size-fits-all misjudgments or omissions, and more in line with the actual process operation background. By automatically identifying and marking batches with risk scores exceeding the threshold, the system realizes the transformation from passive response to proactive discovery. The system can not only output the risk level of the current batch, but also form the evolution trend of high-risk sections, providing substantial support for subsequent process adjustments, pre-control strategy deployment, and even pre-screening of production batches, thereby strengthening the risk identification capability of the entire production system. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a pharmaceutical industrial management system based on industrial big data according to the present invention. Figure 2 This is a schematic diagram illustrating the steps of a pharmaceutical industry management method based on industrial big data according to the present invention. Figure 3 This is a schematic diagram of the comprehensive offset risk score acquisition process of the present invention; Figure 4 This is a trend graph of the standardized relevant factors of this invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1

[0024] This invention provides a pharmaceutical industry management system based on industrial big data. Please refer to [link / reference]. Figures 1 to 4 It includes a data regularization and quality indicator standardization module, a batch-to-batch quality indicator fluctuation identification module, a process parameter residual extraction module for fluctuating batches, a process parameter and quality fluctuation factor construction module, and a suboptimal section identification and offset risk score calculation module. The data normalization and quality index standardization module collects the process parameters Gcs and the endpoint quality index Gzx of the same batch of drugs, and performs preprocessing to obtain the drug dataset YW; The batch-to-batch quality index fluctuation identification module extracts features from the drug dataset YW, extracts the endpoint quality index Gzx, calculates and obtains the quality fluctuation amplitude Zfu, and filters out the batch subset Bva with large fluctuations. The process parameter residual extraction module for fluctuating batches performs in-depth analysis of the process parameters Gcs in the large fluctuating batch subset Bva, and calculates and obtains the offset residual PR of each parameter. The process parameter and quality fluctuation factor construction module matches the correlation between offset residual PR and quality fluctuation amplitude Zfu, and calculates and obtains the standardized correlation factor H; The suboptimal segment identification and offset risk score calculation module quantifies the risk level of each parameter based on the standardized relevant factor H, obtains the comprehensive offset risk score DF, and provides feedback.

[0025] In this embodiment, the process parameters and endpoint quality indicators collected in each batch of production are uniformly structured through the data normalization and quality indicator standardization module. This not only enhances the comparability between multiple batches of data, but also provides strong and stable original data support for subsequent fluctuation analysis and modeling. It effectively overcomes the technical obstacles that existing process data is difficult to use for horizontal analysis due to different sampling frequencies, different cycle lengths, and inconsistent indicator scales.

[0026] By calculating the magnitude of quality fluctuations and using a large-fluctuation batch screening mechanism, the system can not only identify batches with abnormal quality that exceed the warning line, but more importantly, it can detect "slight but persistent" deviation trends that are not identified in the traditional conformity assessment system. This is of great significance for preventing quality risks from shifting forward and avoiding delayed responses, and it solves the technical blind spot of traditional methods that only look at whether the product is conforming and do not look at trend deviations.

[0027] By analyzing the process parameters of identified fluctuating batches, the system constructs a quantitative expression of structural offset residuals, which can accurately identify the "suboptimal control state" generated at different stages and in different parameter dimensions. This structured offset information provides an operational basis for subsequent parameter-quality correlation analysis, making up for the limitation of traditional statistical process control methods in capturing time-sensitive disturbances.

[0028] Through the relevant factor construction module, the system extracts the actual impact strength of each key process parameter on quality fluctuations based on historical operational data, forming a data-driven process-quality coupling map. This allows parameter prioritization to be ranked according to their actual contribution to fluctuations, rather than relying on empirical judgment, providing a quantitative basis for process optimization decisions.

[0029] The system further generates a comprehensive offset risk score for each batch based on the calculated parameter sensitivity factors and residual information, and determines whether it is in a high-risk zone through a dynamically acquired threshold model. This non-static, real-time adaptive risk identification mechanism can effectively identify potentially unstable batches in a marginal state, enabling early identification and early intervention, and improving the foresight and proactivity of quality control. Example 2

[0030] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 and Figure 3 Specifically: the data regularization and quality index standardization module includes a raw data fitting and acquisition unit and a structure mapping-driven preprocessing unit; The raw data fitting and acquisition unit collects the process parameters Gcs and the endpoint quality index Gzx through sensors and acquisition settings, and fits them into the raw dataset CW; the process parameters Gcs include reaction temperature Trx, pH value pH, and stirring speed v; The reaction temperature Trx was obtained using an industrial thermocouple. pH value was obtained by online pH electrode sampling; The stirring speed v is acquired through the feedback interface of the servo motor encoder and the frequency converter. The endpoint quality indicator Gzx includes the content of active ingredient Ca, purity Pu, and residual moisture Sr; The content of active ingredient Ca was obtained by high-performance liquid chromatography (HPLC). Purity of Pu was obtained using gas chromatography. Residual moisture (Sr) was collected using an infrared moisture analyzer.

[0031] The structure mapping-driven preprocessing unit cleans and normalizes the original dataset CW to obtain the drug dataset YW; Cleaning includes multi-scale structural anomaly removal and dynamic structural offset correction; Multi-scale structural anomaly removal is performed by processing the original CW dataset using the sliding perturbation standard deviation method to eliminate anomalous perturbation points. The processing method is as follows: First, set a sliding time window of length w, and take w consecutive sampling points starting from time point t; collect the parameter value sequence within the time period, average the data of the w sampling points to obtain the baseline level within the sliding time window, compare the data of each sampling point with the baseline level to obtain the squared deviation; finally, calculate the average of all squared deviations and take the square root to obtain the degree of abnormal fluctuation. When the degree of abnormal fluctuation exceeds the preset fluctuation threshold, it is identified as an abnormal disturbance point and is removed. The structural dynamic offset correction uses a dynamic time normalization function to uniformly map the production time of different batches; The processing method is as follows: First, obtain the total runtime of the current batch; process each time t, divide the current time t by the total runtime to obtain a normalized time position λ, whose value ranges from 0 to 1; re-mark the parameter values ​​at the original time t as the time-normalized parameter values. The normalization formula is as follows: ; In the formula, YWo represents the o-th data in the drug dataset YW, CWo represents the o-th data in the original dataset CW, minCWo represents the valley value of the o-th data in the original dataset CW, and maxCWo represents the peak value of the o-th data in the original dataset CW.

[0032] In this embodiment, the process parameter Gcs and the endpoint quality index Gzx are acquired synchronously through the raw data fitting acquisition unit, realizing the data structure unification from the equipment acquisition layer to the experimental detection layer. This solves the problem of the disconnect between control system data and quality experimental data in the traditional acquisition process, ensuring that each batch of data is connected throughout the entire process from raw material processing to final inspection, and providing a complete and continuous data foundation for subsequent modeling and analysis.

[0033] In the structure mapping-driven preprocessing unit, the introduced sliding disturbance standard deviation method can identify abnormal disturbance points in the original process parameter curves in real time. This local fluctuation analysis method based on time windows does not rely on external rules for judgment, but identifies atypical disturbances or mechanical transients by statistically analyzing the intensity of the fluctuations themselves, and then eliminates them.

[0034] By employing a dynamic structural offset correction mechanism and introducing time normalization mapping, the timeline misalignment caused by differences in actual process runtime across batches is uniformly addressed. The original time points are converted into relative time proportions, effectively eliminating alignment barriers caused by variations in start and end times and the number of sampling points between different batches. This allows subsequent analytical models such as horizontal comparisons, batch clustering, and trend fitting to operate under the same coordinate semantics, significantly improving the modeling logic and reliability. Through normalization, parameters such as temperature, pH, and rotation speed are uniformly mapped to the same numerical range, avoiding data structure skew caused by different units and numerical scales. This provides a mathematical basis for equal comparison of features across dimensions in subsequent analyses, overcoming the inherent defect in traditional data mining where dimensional imbalance affects the accuracy of correlation calculations. Example 3

[0035] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 3 Specifically: the batch-to-batch quality index fluctuation identification module includes an endpoint quality index extraction and construction unit and a batch fluctuation amplitude calculation unit; The endpoint quality index extraction construction unit performs feature extraction on the drug dataset YW, and extracts the endpoint quality index Gzx, including active ingredient content Ca, purity Pu and residual moisture Sr; The final quality index Gzx of each batch is constructed into a batch feature vector to obtain the quality index matrix Q; The quality indicator matrix Q is obtained using the following formula: ; In the formula, rows represent each batch, and columns represent each quality index q∈{Ca,Pu,Sr}.

[0036] The batch fluctuation amplitude calculation unit constructs a fluctuation function for each quality index q∈{Ca,Pu,Sr} in the quality index matrix Q, and calculates the quality fluctuation amplitude Zfu of each batch under each index through a multi-level fluctuation measurement function. For each quality index q, the deviation value of the i-th batch is defined as Δq(i); The formula is: ; In the formula, Qq(i) represents the actual value of quality index q in the i-th batch, and pQq represents the average value of quality index q in all batches; The quality fluctuation range Zfu is obtained using the following formula: ; In the formula, Zfu(i) represents the quality fluctuation range of the i-th batch, Ca(i) represents the content of active ingredients in the i-th batch, Pu(i) represents the purity of the i-th batch, and Sr(i) represents the residual moisture in the i-th batch. The obtained quality fluctuation amplitude Zfu is compared with the preset fluctuation screening threshold Tfu to obtain the large fluctuation batch subset Bva; the formula is as follows: Bva={i|Zfu(i)>Tfu}.

[0037] The process parameter residual extraction module for fluctuating batches includes a window feature aggregation unit and a window feature aggregation unit; The window feature aggregation unit performs in-depth analysis of the process parameters Gcs in the large fluctuation batch subset Bva, and performs sliding feature extraction to extract the average reaction temperature PTrx, average pH value PpH, and average stirring speed Pv, which are then fitted into a process parameter feature set GS={PTrx, PpH, Pv}. The parameter offset residual calculation unit analyzes the process parameter feature set GS of each fluctuating batch and compares it with the preset reference batch process feature mean set Sref to calculate and obtain the parameter residual CR. CRj = |GSj(i) - Srefj|; In the formula, CRj represents the parameter residual of the j-th process parameter, GSj(i) represents the statistical characteristic value of the j-th process parameter in the i-th batch, and j∈{Trx, pH, v}; The acquired parameter residual CR is compared with the preset recognition threshold Tj to obtain the offset residual PR; The offset residual PR is obtained as follows: ; In the formula, PRj(i) represents the offset residual of the j-th process parameter in the i-th batch.

[0038] In this embodiment, by extracting the endpoint quality indicators for each batch, a standardized batch quality indicator matrix is ​​constructed. This breaks away from the traditional batch-based, isolated information approach in quality records, allowing the quality performance of all batches to be incorporated into the same analytical framework. This approach enhances the insight into product fluctuation trends and facilitates cross-batch quality consistency tracking and quality deviation distribution identification.

[0039] To address the insufficient sensitivity of traditional statistical methods in handling minor fluctuations and boundary batches, this embodiment proposes a multi-level fluctuation measurement method. By calculating the relative deviation of batches and the joint offset trend of indicators, it effectively identifies batches that, although not exceeding quality limits, show signs of systemic fluctuations. These "potentially risky batches" are often overlooked in traditional judgment systems, but this method allows them to be identified in advance, enhancing the foresight and accuracy of process fluctuation management.

[0040] This embodiment further extracts the window features of the main process parameters of the identified fluctuating batches during actual operation and constructs an offset residual model between the batch and the reference parameter set to accurately reveal the key process control behavior changes that cause quality fluctuations.

[0041] By setting an offset recognition threshold, this embodiment can not only identify numerical deviations but also eliminate invalid interference based on a substantial comparison of the degree of deviation, retaining structural offsets with operational significance. This approach avoids false alarms for natural fluctuations or non-critical disturbances, making the system's judgment of risk parameters more realistic, while also improving the operability of prioritizing and refining process parameters.

[0042] This embodiment realizes the construction of a logical chain from the identification of endpoint quality deviation to the quantitative analysis of parameter residuals. It forms a mechanism within the system that drives process reverse engineering based on quality data and guides risk identification based on process characteristics. This significantly improves the structured response capability of the pharmaceutical process to fluctuations and anomalies, and provides high-dimensional data support for continuous process improvement, batch variation control and process capability verification. Example 4

[0043] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 3 and Figure 4 Specifically: the process parameter and quality fluctuation factor construction module includes a fluctuation value recombination and alignment unit and a causal sensitivity output unit; The fluctuation value recombination and alignment unit aligns the acquired offset residuals PR with the quality fluctuation amplitudes Zfu according to the batch number, forming a joint data structure and constructing a two-dimensional residual fluctuation joint dataset ELD={PRj(i),Zfu(i)}, i∈Bva, j∈Gcs; Each row represents the relationship between the structural offset of the j-th process parameter in the i-th batch and the overall quality fluctuation of the batch. The causal sensitivity output unit standardizes the data in the two-dimensional residual fluctuation joint dataset ELD and uses the Pearson standardized correlation coefficient formula to calculate the statistical correlation strength between each process parameter and quality fluctuation, and obtains the standardized correlation factor H. Statistical feature processing was performed on all offset residuals PR and quality fluctuation amplitude Zfu to obtain the mean offset residual μPR, the standard deviation of offset residual σPR, the mean quality fluctuation amplitude μZfu, and the standard deviation of quality fluctuation amplitude σZfu. The standardized correlation factor H is obtained using the following formula: ; In the formula, Hj represents the standardized correlation factor between the j-th process parameter and the quality fluctuation range, μPRj represents the mean of the offset residuals of the j-th process parameter, and σPRj represents the standard deviation of the offset residuals of the j-th process parameter. The relationship between process parameters and quality fluctuations is analyzed using a standardized correlation factor H, as follows: When 0 < standardized correlation factor H < 1, it indicates that the deviation trend of process parameters is strongly positively correlated with quality fluctuation. When -1 < standardized correlation factor H < 0, it indicates that the deviation of process parameters is negatively correlated with quality fluctuation; When the standardized correlation factor H=0, it indicates that there is no linear relationship between the two.

[0044] In this embodiment, through the fluctuation value recombination and alignment unit, the process parameter offset residuals PR of each batch and their corresponding quality fluctuation amplitudes Zfu are mapped one-to-one under the batch number and structurally fused to construct a unified two-dimensional residual-fluctuation dataset. This design breaks through the limitations of the decentralized management of process control data and quality inspection data in traditional systems, and for the first time realizes the transformation from single variable trend analysis to the mining of three-dimensional structural relationships at the batch, indicator, and parameter levels.

[0045] This embodiment employs a standardized processing approach. By eliminating dimensional interference and amplifying the inherent nature of fluctuations, it utilizes Pearson correlation analysis to construct a statistical correlation model between process parameters and quality fluctuations. Unlike conventional experience-based judgments or subjective screening, this method outputs clear and rankable correlation factor values ​​H, providing enterprises with explicit priorities for key process parameters. This facilitates focusing on core intervention points and enhances the targetedness and effectiveness of process control.

[0046] This embodiment not only outputs parameter-related factor values, but also establishes a process behavior interpretation logic based on positive and negative interval judgment through a causal sensitivity output unit. Specifically, the H value of different parameters falling into the positive interval, negative interval, or near-zero region represents positive correlation, negative correlation, or no correlation, respectively. Enterprises can use this to determine whether a parameter is "better the larger it is" or "more stable the lower it is." This provides a targeted suggestion basis for parameter limit adjustment and control strategy design in actual production, solving the core pain point of "unclear direction" in past process control.

[0047] Compared to traditional methods of tracing quality anomalies that rely on human experience, case review, and ontology identification, this module constructs a computable causal model between process and fluctuations through systematic data analysis. The strength, positive and negative relationships, and batch sensitivity of the influence of each parameter can be visualized and output, providing a clear path for tracing and controlling quality deviations and improving the explanatory power, reusability, and engineering capabilities of the tracing work. Example 5

[0048] This embodiment is an explanation based on Embodiment 4. Please refer to it. Figure 1 Specifically: the suboptimal segment identification and offset risk score calculation module includes a risk factor-driven offset intensity fusion unit and a risk segment dynamic judgment unit; The risk factor-driven offset intensity fusion unit uses the obtained standardized correlation factor H as a sensitive factor for the degree of influence of process parameters, performs risk amplification fusion calculation on the offset residual PR of each fluctuating batch, and obtains the comprehensive offset risk score DF. The overall offset risk score (DF) is obtained using the following formula: ; In the formula, DF(i) represents the comprehensive deviation risk score of the i-th batch, and m represents the number of process parameters involved in the evaluation.

[0049] The risk segment dynamic judgment unit compares the comprehensive offset risk score DF with the risk threshold Tdf to determine whether the batch falls into a high-risk offset segment. The risk threshold Tdf is obtained using the following formula: Tdf = μDF + κ × σDF; In the formula, μDF represents the mean of the comprehensive deviation risk score, and σDF represents the standard deviation of the comprehensive deviation risk score; The method for identifying high-risk deviation zones is as follows: When the overall offset risk score DF ≤ risk threshold Tdf, it indicates that the batch is normal and in a safe area; When the comprehensive offset risk score DF is greater than the risk threshold Tdf, it indicates that the batch is abnormal, is in a high-risk offset zone, and is marked.

[0050] In this embodiment, standardized relevant factors calculated by the preceding modules are integrated as risk sensitivity inputs, organically unifying information on whether fluctuations occur and whether the fluctuations are critical. This allows the deviation risk score to not only reflect the degree of deviation but also to demonstrate the substantial contribution of parameter deviations to quality risk, resulting in stronger interpretability and focus. By constructing a batch-level comprehensive deviation risk scoring system for all fluctuating batches, all batches to be judged can be systematically scored and ranked, achieving quantitative rating and trend archiving of quality fluctuations. This mechanism provides a unified, objective, and comparable criterion standard for subsequent batch audits, process reviews, and abnormal batch tracing, significantly reducing the risk of judgment errors relying on manual experience or static standards.

[0051] In this embodiment, the risk threshold is dynamically calculated from the mean and standard deviation of the offset risk scores, ensuring statistical rationality and dynamic adaptability in risk assessment. This avoids one-size-fits-all misjudgments or omissions and better aligns with the actual process operation context. By automatically identifying and marking batches with risk scores exceeding the threshold, a shift from passive response to proactive detection is achieved. The system not only outputs the risk level of the current batch but also generates an evolution trend of high-risk zones, providing substantial support for subsequent process adjustments, pre-control strategy deployment, and even pre-screening of production batches, thereby strengthening the risk identification capability of the entire production system.

[0052] Because this module is based on the fusion of parameter residuals, quality fluctuations, and statistical characteristics, it is not dependent on specific drug types or fixed processes, and has excellent model transferability and data adaptability. This mechanism is not only applicable to quality identification of the current batch, but can also be extended to fluctuation control and risk management in more process scenarios, making it a widely adaptable intelligent auxiliary management tool in the pharmaceutical industry. Example 6

[0053] A pharmaceutical industry management method based on industrial big data; please refer to [reference needed]. Figure 2 Specifically, it includes the following steps: Step 1: The data preparation and quality index standardization module collects the process parameters Gcs and the endpoint quality index Gzx of the same batch of drugs, and performs preprocessing to obtain the drug dataset YW. Step 2: The batch-to-batch quality index fluctuation identification module extracts features from the drug dataset YW, extracts the endpoint quality index Gzx, calculates and obtains the quality fluctuation amplitude Zfu, and filters out the batch subset Bva with large fluctuations. Step 3: The process parameter residual extraction module for fluctuating batches performs in-depth analysis of the process parameters Gcs in the large fluctuating batch subset Bva, and calculates and obtains the offset residual PR of each parameter. Step 4: The process parameter and quality fluctuation factor construction module matches the correlation between the offset residual PR and the quality fluctuation amplitude Zfu, and calculates and obtains the standardized correlation factor H. Step 5: The suboptimal segment identification and offset risk score calculation module quantifies the risk level of each parameter based on the standardized relevant factor H, obtains the comprehensive offset risk score DF, and provides feedback.

[0054] In this embodiment, by synchronously collecting and preprocessing process parameters and endpoint quality indicators in step one, this method achieves unified standardization of process data and result data in pharmaceutical production at the execution level. This integrates previously scattered and fragmented data into a drug dataset with a consistent structure and temporal semantics. This improvement allows quality analysis to move beyond being limited to a single batch or indicator, providing a holistic analytical foundation across batches and parameters, thus solving the problem of abundant but difficult-to-coordinate data in traditional management methods.

[0055] By introducing a batch-to-batch quality fluctuation identification mechanism in step two, this method can not only identify obviously abnormal batches, but also identify and screen batches that have not yet exceeded quality limits but have shown a fluctuating trend. This process-level design expands quality management from a static judgment of whether something is qualified to a dynamic assessment of whether it is stable, which helps to intervene before problems escalate and reduces the probability of subsequent rework, re-inspection, or systemic fluctuations.

[0056] By conducting a targeted analysis of the process parameters of fluctuating batches in step three, this method establishes an analytical path that infers process behavior from quality results. This allows quality deviations to move beyond the result level and further pinpoint specific process control deviations. This in-depth analysis based on the screening of fluctuating batches avoids indiscriminate analysis of all batches and all parameters, improving the relevance of the analysis and its engineering operability.

[0057] By introducing standardized relevant factors in step four, this method achieves a quantitative expression of the impact of process parameters on the workflow. This means that which parameters are more likely to cause quality fluctuations no longer relies on human experience or subjective judgment, but is explained by the correlation strength extracted from historical data. This improvement enhances the objectivity and interpretability of the analysis results, providing a reliable basis for subsequent process optimization.

[0058] By calculating and dynamically assessing the comprehensive deviation risk score in step five, this method enables the automatic identification of batches in a marginal or potentially unstable state, shifting quality management from post-event traceability to process early warning and risk classification. This process helps managers proactively identify high-risk areas, optimize process control strategies, and prevent anomalies from spreading or becoming entrenched in subsequent batches.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. A pharmaceutical industrial management system based on industrial big data, characterized in that: It includes modules for data regularization and quality index standardization, batch-to-batch quality index fluctuation identification, process parameter residual extraction for fluctuating batches, process parameter and quality fluctuation factor construction, and suboptimal section identification and offset risk score calculation. The data normalization and quality index standardization module collects the process parameters Gcs and the endpoint quality index Gzx of the same batch of drugs, and performs preprocessing to obtain the drug dataset YW; The batch-to-batch quality index fluctuation identification module extracts features from the drug dataset YW, extracts the endpoint quality index Gzx, calculates and obtains the quality fluctuation amplitude Zfu, and filters out the batch subset Bva with large fluctuations. The process parameter residual extraction module for fluctuating batches performs in-depth analysis of the process parameters Gcs in the large fluctuating batch subset Bva, and calculates and obtains the offset residual PR of each parameter. The process parameter and quality fluctuation factor construction module matches the correlation between offset residual PR and quality fluctuation amplitude Zfu, and calculates and obtains the standardized correlation factor H; The suboptimal segment identification and offset risk score calculation module quantifies the risk level of each parameter based on the standardized relevant factor H, obtains the comprehensive offset risk score DF, and provides feedback.

2. The pharmaceutical industrial management system based on industrial big data according to claim 1, characterized in that: The data normalization and quality index standardization module includes a raw data fitting and acquisition unit and a structure mapping-driven preprocessing unit; The raw data fitting and acquisition unit collects the process parameter Gcs and the endpoint quality index Gzx through sensors and acquisition settings, and fits them into the raw dataset CW; The process parameters Gcs include reaction temperature Trx, pH value, and stirring speed v; The reaction temperature Trx was obtained using an industrial thermocouple. pH value was obtained by online pH electrode sampling; The stirring speed v is acquired through the feedback interface of the servo motor encoder and the frequency converter. The endpoint quality indicator Gzx includes the content of active ingredient Ca, purity Pu, and residual moisture Sr; The content of active ingredient Ca was obtained by high-performance liquid chromatography (HPLC). Purity of Pu was obtained using gas chromatography. Residual moisture (Sr) was collected using an infrared moisture analyzer.

3. The pharmaceutical industrial management system based on industrial big data according to claim 2, characterized in that: The structure mapping-driven preprocessing unit cleans and normalizes the original dataset CW to obtain the drug dataset YW; Cleaning includes multi-scale structural anomaly removal and dynamic structural offset correction; Multi-scale structural anomaly removal is performed by processing the original CW dataset using the sliding perturbation standard deviation method to eliminate anomalous perturbation points. The processing method is as follows: First, set a sliding time window of length w, and take w consecutive sampling points starting from time point t; collect the parameter value sequence within the time period, average the data of the w sampling points to obtain the baseline level within the sliding time window, compare the data of each sampling point with the baseline level to obtain the squared deviation; finally, calculate the average of all squared deviations and take the square root to obtain the degree of abnormal fluctuation. When the degree of abnormal fluctuation exceeds the preset fluctuation threshold, it is identified as an abnormal disturbance point and is removed. The structural dynamic offset correction uses a dynamic time normalization function to uniformly map the production time of different batches; The processing method is as follows: First, obtain the total runtime of the current batch; process each time t, divide the current time t by the total runtime to obtain a normalized time position λ, whose value ranges from 0 to 1; re-mark the parameter values ​​at the original time t as the time-normalized parameter values. The normalization formula is as follows: ； In the formula, YWo represents the o-th data in the drug dataset YW, CWo represents the o-th data in the original dataset CW, minCWo represents the valley value of the o-th data in the original dataset CW, and maxCWo represents the peak value of the o-th data in the original dataset CW.

4. The pharmaceutical industrial management system based on industrial big data according to claim 3, characterized in that: The batch-to-batch quality index fluctuation identification module includes an endpoint quality index extraction and construction unit and a batch fluctuation amplitude calculation unit. The endpoint quality index extraction construction unit performs feature extraction on the drug dataset YW, and extracts the endpoint quality index Gzx, including active ingredient content Ca, purity Pu and residual moisture Sr; The final quality index Gzx of each batch is constructed into a batch feature vector to obtain the quality index matrix Q; The quality indicator matrix Q is obtained using the following formula: ； In the formula, rows represent each batch, and columns represent each quality index q∈{Ca,Pu,Sr}.

5. A pharmaceutical industrial management system based on industrial big data according to claim 4, characterized in that: The batch fluctuation amplitude calculation unit constructs a fluctuation function for each quality index q∈{Ca,Pu,Sr} in the quality index matrix Q, and calculates the quality fluctuation amplitude Zfu of each batch under each index through a multi-level fluctuation measurement function. For each quality index q, the deviation value of the i-th batch is defined as Δq(i); The formula is: ; In the formula, Qq(i) represents the actual value of quality index q in the i-th batch, and pQq represents the average value of quality index q in all batches; The quality fluctuation range Zfu is obtained using the following formula: ； In the formula, Zfu(i) represents the quality fluctuation range of the i-th batch, Ca(i) represents the content of active ingredients in the i-th batch, Pu(i) represents the purity of the i-th batch, and Sr(i) represents the residual moisture in the i-th batch. The obtained quality fluctuation amplitude Zfu is compared with the preset fluctuation screening threshold Tfu to obtain the large fluctuation batch subset Bva; the formula is as follows: Bva={i|Zfu(i)>Tfu}.

6. A pharmaceutical industrial management system based on industrial big data according to claim 5, characterized in that: The process parameter residual extraction module for fluctuating batches includes a window feature aggregation unit and a window feature aggregation unit; The window feature aggregation unit performs in-depth analysis of the process parameters Gcs in the large fluctuation batch subset Bva, and performs sliding feature extraction to extract the average reaction temperature PTrx, average pH value PpH, and average stirring speed Pv, which are then fitted into a process parameter feature set GS={PTrx, PpH, Pv}. The parameter offset residual calculation unit analyzes the process parameter feature set GS of each fluctuating batch and compares it with the preset reference batch process feature mean set Sref to calculate and obtain the parameter residual CR. CRj = |GSj(i) - Srefj|; In the formula, CRj represents the parameter residual of the j-th process parameter, GSj(i) represents the statistical characteristic value of the j-th process parameter in the i-th batch, Srefj represents the mean value of the process characteristic of the j-th process parameter, and j∈{Trx, pH, v}; The acquired parameter residual CR is compared with the preset recognition threshold Tj to obtain the offset residual PR; The offset residual PR is obtained as follows: ； In the formula, PRj(i) represents the offset residual of the j-th process parameter in the i-th batch.

7. A pharmaceutical industrial management system based on industrial big data according to claim 6, characterized in that: The process parameter and quality fluctuation factor construction module includes a fluctuation value recombination and alignment unit and a causal sensitivity output unit; The fluctuation value recombination and alignment unit aligns the acquired offset residuals PR with the quality fluctuation amplitudes Zfu according to the batch number, forming a joint data structure and constructing a two-dimensional residual fluctuation joint dataset ELD={PRj(i),Zfu(i)}, i∈Bva, j∈Gcs; Each row represents the relationship between the structural offset of the j-th process parameter in the i-th batch and the overall quality fluctuation of the batch. The causal sensitivity output unit standardizes the data in the two-dimensional residual fluctuation joint dataset ELD and uses the Pearson standardized correlation coefficient formula to calculate the statistical correlation strength between each process parameter and quality fluctuation, and obtains the standardized correlation factor H. Statistical feature processing was performed on all offset residuals PR and quality fluctuation amplitude Zfu to obtain the mean offset residual μPR, the standard deviation of offset residual σPR, the mean quality fluctuation amplitude μZfu, and the standard deviation of quality fluctuation amplitude σZfu. The standardized correlation factor H is obtained using the following formula: ； In the formula, Hj represents the standardized correlation factor between the j-th process parameter and the quality fluctuation range, μPRj represents the mean of the offset residual of the j-th process parameter, σPRj represents the standard deviation of the offset residual of the j-th process parameter, and n represents the total number of batches. The relationship between process parameters and quality fluctuations is analyzed using a standardized correlation factor H, as follows: When 0 < standardized correlation factor H < 1, it indicates that the deviation trend of process parameters is strongly positively correlated with quality fluctuation. When -1 < standardized correlation factor H < 0, it indicates that the deviation of process parameters is negatively correlated with quality fluctuation; When the standardized correlation factor H=0, it indicates that there is no linear relationship between the two.

8. A pharmaceutical industrial management system based on industrial big data according to claim 7, characterized in that: The suboptimal segment identification and offset risk score calculation module includes a risk factor-driven offset intensity fusion unit and a risk segment dynamic judgment unit. The risk factor-driven offset intensity fusion unit uses the obtained standardized correlation factor H as a sensitive factor for the degree of influence of process parameters, performs risk amplification fusion calculation on the offset residual PR of each fluctuating batch, and obtains the comprehensive offset risk score DF. The overall offset risk score (DF) is obtained using the following formula: ； In the formula, DF(i) represents the comprehensive deviation risk score of the i-th batch, and m represents the number of process parameters involved in the evaluation.

9. A pharmaceutical industrial management system based on industrial big data according to claim 8, characterized in that: The risk segment dynamic judgment unit compares the comprehensive offset risk score DF with the risk threshold Tdf to determine whether the batch falls into a high-risk offset segment. The risk threshold Tdf is obtained using the following formula: Tdf = μDF + κ × σDF; In the formula, μDF represents the mean of the comprehensive deviation risk score, and σDF represents the standard deviation of the comprehensive deviation risk score; The method for identifying high-risk deviation zones is as follows: When the overall offset risk score DF ≤ risk threshold Tdf, it indicates that the batch is normal and in a safe area; When the comprehensive offset risk score DF is greater than the risk threshold Tdf, it indicates that the batch is abnormal, is in a high-risk offset zone, and is marked.

10. A pharmaceutical industry management method based on industrial big data, applied to the pharmaceutical industry management system based on industrial big data as described in any one of claims 1 to 9, characterized in that: Includes the following steps: Step 1: The data preparation and quality index standardization module collects the process parameters Gcs and the endpoint quality index Gzx of the same batch of drugs, and performs preprocessing to obtain the drug dataset YW. Step 2: The batch-to-batch quality index fluctuation identification module extracts features from the drug dataset YW, extracts the endpoint quality index Gzx, calculates and obtains the quality fluctuation amplitude Zfu, and filters out the batch subset Bva with large fluctuations. Step 3: The process parameter residual extraction module for fluctuating batches performs in-depth analysis of the process parameters Gcs in the large fluctuating batch subset Bva, and calculates and obtains the offset residual PR of each parameter. Step 4: The process parameter and quality fluctuation factor construction module matches the correlation between the offset residual PR and the quality fluctuation amplitude Zfu, and calculates and obtains the standardized correlation factor H. Step 5: The suboptimal segment identification and offset risk score calculation module quantifies the risk level of each parameter based on the standardized relevant factor H, obtains the comprehensive offset risk score DF, and provides feedback.

Images