A method and system for early warning of low-temperature crystallization process of peptide drugs based on digital twins

Abstract

This application belongs to the technical field of peptide drugs and discloses a method and system for early warning of low-temperature crystallization process of peptide drugs based on digital twins. The method includes: acquiring batch information of peptide drugs, as well as crystallization process data and online spectral data during the low-temperature crystallization process; inputting the batch information and crystallization process data into a digital twin model to predict the state parameters during the low-temperature crystallization process; extracting feature vectors related to the crystallization process from the online spectral data to obtain real-time spectral features; calculating the degree of inconsistency between key state parameters and real-time spectral features to obtain prediction deviation evaluation results; and issuing an early warning signal when the prediction deviation evaluation result is detected to be greater than a preset evaluation threshold or when the real-time spectral features are abnormal data. By issuing an early warning signal when the prediction deviation evaluation result is detected to be greater than a preset evaluation threshold or when the real-time spectral features are abnormal data, the early warning efficiency during the low-temperature crystallization process is improved.

Description

Technical Field

[0001] This application relates to the technical field of peptide drugs, and more specifically, to a method and system for early warning of low-temperature crystallization process of peptide drugs based on digital twins. Background Technology

[0002] Peptide drugs are an important component of modern medicine, and their production process requires extremely high purity and quality control. Low-temperature crystallization is a key purification step in the industrial production line of peptide drugs, aiming to precipitate peptides from solution in high-purity crystal form by controlling process parameters such as temperature, solvent composition, and stirring rate.

[0003] In the industrial-scale, multi-batch production environment of peptide drugs, even with seemingly identical process procedures, equipment conditions, and qualified raw materials and solvents, subtle differences that are difficult to completely eliminate still exist between different production batches. These batch-to-batch variations may originate from minor fluctuations in the purity or impurity profile of the active pharmaceutical ingredient (API) between batches, slight differences in trace moisture content or composition ratios between solvent batches, or even minor wear and tear or changes in the condition of equipment after long-term operation. These seemingly insignificant batch-specific factors can accumulate and have an unexpected impact on the fine crystallization behavior of peptides at low temperatures.

[0004] Existing early warning systems based on digital twins and machine learning face challenges in addressing batch-to-batch variability. On one hand, historical data used to train both the digital twin and machine learning models may not adequately cover subtle variation patterns unique to the current batch. On the other hand, machine learning models exhibit significantly reduced recognition capabilities when faced with these insufficiently learned batch-specific patterns. Due to the limitations of digital twin models in adapting to batch variations and the limitations of machine learning models in identifying insufficiently learned early-stage anomalies, early warning systems may fail to issue accurate warnings in the early stages, potentially leading to the scrapping of the entire production batch or requiring time-consuming and costly post-processing, significantly increasing production costs.

[0005] Therefore, in order to solve the technical problem that existing early warning methods for low-temperature crystallization of peptide drugs are insufficient in adaptability to batch differences and inability to recognize abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate early warnings, there is an urgent need for an early warning method and system for low-temperature crystallization of peptide drugs based on digital twins. Summary of the Invention

[0006] The purpose of this application is to provide a digital twin-based early warning method and system for the low-temperature crystallization process of peptide drugs. By calculating the degree of inconsistency between key state parameters and real-time spectral features, a prediction deviation assessment result is obtained. When the prediction deviation assessment result exceeds a preset assessment threshold or the real-time spectral features are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for the low-temperature crystallization process of peptide drugs have insufficient adaptability to batch differences and insufficient ability to identify abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate early warnings. By predicting the actual state through a digital twin model and reflecting it with online spectra, the method can adaptively capture the batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0007] In a first aspect, this application provides a method for early warning of the low-temperature crystallization process of peptide drugs based on digital twins, including:

[0008] Obtain specific batch information of the peptide drug to be tested, and acquire real-time crystallization process data and online spectral data of the peptide drug during the low-temperature crystallization process;

[0009] The specific batch information and the crystallization process data are input into a preset digital twin model to simulate the low-temperature crystallization process and predict the key state parameters of the test peptide drug at the next moment in the low-temperature crystallization process.

[0010] Feature vectors related to the crystallization process are extracted from the online spectral data to obtain real-time spectral features;

[0011] The degree of inconsistency between the key state parameters and the real-time spectral features is calculated to obtain the prediction deviation assessment result;

[0012] When the prediction deviation assessment result is detected to be greater than the preset assessment threshold, or when the real-time spectral features are abnormal data, an early warning signal is issued.

[0013] The digital twin-based early warning method for the low-temperature crystallization process of peptide drugs provided in this application can provide early warning of anomalies in the low-temperature crystallization process of peptide drugs. By calculating the degree of inconsistency between key state parameters and real-time spectral features, a prediction deviation assessment result is obtained. When the prediction deviation assessment result is greater than a preset assessment threshold or the real-time spectral features are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for the low-temperature crystallization process of peptide drugs have insufficient adaptability to batch differences and insufficient ability to identify abnormal patterns when using digital twin and machine learning methods, making it difficult to issue accurate early warnings. By predicting the actual state through a digital twin model and reflecting the online spectrum, the method can adaptively capture the batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0014] Optionally, feature vectors related to the crystallization process are extracted from the online spectral data to obtain real-time spectral features, including:

[0015] The online spectral data is preprocessed to obtain preprocessed Raman spectral data;

[0016] The characteristic peak area related to peptide concentration, the characteristic peak position related to solvent composition, the characteristic peak intensity ratio related to crystal form, and the specific band integral area related to impurities or intermediate products in the crystallization process are extracted from the preprocessed Raman spectral data to construct the corresponding feature vectors and obtain real-time spectral features.

[0017] The digital twin-based early warning method for low-temperature crystallization of peptide drugs provided in this application can provide early warning of anomalies in the low-temperature crystallization process of peptide drugs. It extracts characteristic peak areas related to peptide concentration, characteristic peak positions related to solvent composition, characteristic peak intensity ratios related to crystal form, and specific band integral areas related to impurities or intermediate products in the crystallization process from preprocessed Raman spectral data. By extracting these physicochemically significant features, high-dimensional spectral data can be transformed into low-dimensional and information-rich feature vectors. These feature vectors can be more reliably used for subsequent comparison with key state parameters predicted by the digital twin model, thereby improving the accuracy of prediction deviation assessment and ultimately enhancing the reliability of the early warning system.

[0018] Optionally, the online spectral data is preprocessed to obtain preprocessed Raman spectral data, including:

[0019] The online spectral data is filtered to obtain filtered Raman spectral data;

[0020] The filtered Raman spectral data is smoothed using a smoothing algorithm to obtain smoothed Raman spectral data.

[0021] The fluorescence background interference of the smoothed Raman spectral data was corrected by using the least squares method to obtain the preprocessed Raman spectral data.

[0022] Optionally, the degree of inconsistency between the key state parameters and the real-time spectral features is calculated to obtain the prediction deviation assessment result, including:

[0023] The key state parameters and the real-time spectral features are standardized respectively to obtain the standardized key state parameters and standardized real-time spectral features.

[0024] Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector.

[0025] The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into a preset deviation prediction model to calculate the prediction deviation evaluation result.

[0026] The digital twin-based early warning method for the low-temperature crystallization process of peptide drugs provided in this application can provide early warning of abnormalities in the low-temperature crystallization process of peptide drugs. By considering the reliability of the data and weighting it, and using a prediction model for comprehensive evaluation, the accuracy and sensitivity of the assessment of the degree of inconsistency between key state parameters and real-time spectral characteristics are improved. This enables more effective identification of early or minor process abnormalities and supports subsequent early warning judgments.

[0027] Optionally, based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector, including:

[0028] Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined and denoted as the first weight and the second weight, respectively; wherein, the greater the uncertainty, the smaller the first weight, and the greater the extraction deviation, the smaller the second weight.

[0029] Based on the first weight and the second weight, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector.

[0030] Optionally, based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined, including:

[0031] The uncertainty of the key state parameters is determined by repeatedly inputting the specific batch information and the crystallization process data into a preset digital twin model to calculate the variance of the key state parameters.

[0032] The extraction deviation of the real-time spectral features is determined by calculating the actual difference of the real-time spectral features obtained by extracting feature vectors related to the crystallization process from the online spectral data multiple times.

[0033] Based on the uncertainty and the extraction bias, the weights of the corresponding standardized key state parameters and the weights of the corresponding standardized real-time spectral features are determined.

[0034] Optionally, the preset deviation prediction model is a model constructed using historical data to represent the nonlinear relationship between key state parameters and real-time spectral characteristics; the construction steps of the preset deviation prediction model include:

[0035] Retrieve historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results between the historical crystallization process data and the historical spectral data for multiple polypeptide drugs from the database.

[0036] Based on the historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results, corresponding historical key state parameters and historical spectral features are generated.

[0037] Based on the historical key state parameters and the historical spectral characteristics, a machine learning algorithm is used to construct the preset deviation prediction model.

[0038] Optionally, when the prediction deviation assessment result is detected to be greater than a preset assessment threshold, or when the real-time spectral features are abnormal data, an early warning signal is issued, including:

[0039] Based on historical spectral data in the database, pre-determine the conditions for identifying abnormal data;

[0040] Based on the abnormal data determination conditions, it is determined whether the real-time spectral features are abnormal data. At the same time, the relationship between the prediction deviation evaluation result and the preset evaluation threshold is compared.

[0041] When the prediction deviation assessment result is detected to be greater than the preset assessment threshold, or when the real-time spectral features are detected to be abnormal data, an early warning signal is issued.

[0042] Optionally, the abnormal data determination conditions include numerical anomaly conditions and pattern anomaly conditions; based on the abnormal data determination conditions, determining whether the real-time spectral features are abnormal data includes:

[0043] The spectral data of historical normal batches and the spectral data of historical abnormal patterns are extracted from the historical spectral data.

[0044] The Mahalanobis distance between the real-time spectral feature and the spectral data of the historical normal batch is calculated. The real-time spectral feature is determined to be abnormal data by whether the Mahalanobis distance is greater than a preset normal threshold. When the Mahalanobis distance is greater than the preset normal threshold, the real-time spectral feature is determined to meet the numerical abnormality condition and is determined to be abnormal data.

[0045] The Euclidean distance between the real-time spectral feature and the spectral data of the historical abnormal pattern is calculated. The real-time spectral feature is determined to be abnormal data by whether the Euclidean distance is less than a preset abnormal threshold. When the Euclidean distance is less than the preset abnormal threshold, the real-time spectral feature is determined to meet the abnormal conditions of the pattern and is therefore determined to be abnormal data.

[0046] Secondly, this application provides an early warning system for the low-temperature crystallization process of peptide drugs based on digital twins, comprising:

[0047] The acquisition module is used to acquire specific batch information of the peptide drug to be tested, and to acquire crystallization process data and online spectral data of the peptide drug in real time during the low-temperature crystallization process.

[0048] The prediction module is used to input the specific batch information and the crystallization process data into a preset digital twin model to simulate the low-temperature crystallization process and predict the key state parameters of the test peptide drug at the next moment in the low-temperature crystallization process.

[0049] The extraction module is used to extract feature vectors related to the crystallization process from the online spectral data to obtain real-time spectral features;

[0050] The extraction module is used to calculate the degree of inconsistency between the key state parameters and the real-time spectral features, and obtain the prediction deviation evaluation result;

[0051] The early warning module is used to issue an early warning signal when the prediction deviation evaluation result is detected to be greater than a preset evaluation threshold, or when the real-time spectral features are abnormal data.

[0052] This digital twin-based early warning system for the low-temperature crystallization process of peptide drugs calculates the degree of inconsistency between key state parameters and real-time spectral features to obtain a prediction deviation assessment result. When the prediction deviation assessment result exceeds a preset assessment threshold or the real-time spectral features are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for the low-temperature crystallization process of peptide drugs have insufficient adaptability to batch differences and insufficient ability to identify abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate warnings. By predicting the actual state through a digital twin model and reflecting the online spectrum, it can adaptively capture the batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0053] Beneficial Effects: The digital twin-based early warning method and system for low-temperature crystallization of peptide drugs provided in this application obtains the prediction deviation assessment result by calculating the degree of inconsistency between key state parameters and real-time spectral features. When the prediction deviation assessment result exceeds a preset assessment threshold or the real-time spectral features are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for low-temperature crystallization of peptide drugs have insufficient batch-to-batch adaptability and insufficient ability to identify abnormal patterns when using digital twin and machine learning methods, making it difficult to issue accurate early warnings. By predicting the actual state through a digital twin model and reflecting it with online spectra, the method can adaptively capture the batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs. Attached Figure Description

[0054] Figure 1 A flowchart of a method for early warning of low-temperature crystallization process of peptide drugs based on digital twins, provided in an embodiment of this application.

[0055] Figure 2 This is a schematic diagram of the structure of an early warning system for low-temperature crystallization of polypeptide drugs based on digital twins, provided in an embodiment of this application.

[0056] Labeling Explanation: 1. Acquisition Module; 2. Prediction Module; 3. Extraction Module; 4. Calculation Module; 5. Early Warning Module. Detailed Implementation

[0057] 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 a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0058] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0059] Please refer to Figure 1 , Figure 1 This application discloses a method for early warning of abnormalities in the low-temperature crystallization process of peptide drugs based on digital twins, which is used to provide early warning of abnormalities in the low-temperature crystallization process of peptide drugs. The method includes the following steps:

[0060] Step S101: Obtain batch information of the peptide drug, as well as crystallization process data and online spectral data of the peptide drug during low-temperature crystallization.

[0061] Step S102: Input batch information and crystallization process data into a preset digital twin model to simulate the low-temperature crystallization process and predict the state parameters of the peptide drug at the next moment in the low-temperature crystallization process.

[0062] Step S103: Extract feature vectors related to the crystallization process from online spectral data to obtain real-time spectral features;

[0063] Step S104: Calculate the degree of inconsistency between key state parameters and real-time spectral characteristics to obtain the prediction deviation assessment result;

[0064] Step S105: When the prediction deviation evaluation result is detected to be greater than the preset evaluation threshold, or when the real-time spectral characteristics are abnormal data, an early warning signal is issued.

[0065] This digital twin-based early warning method for the low-temperature crystallization process of peptide drugs calculates the degree of inconsistency between key state parameters and real-time spectral features to obtain a prediction deviation assessment result. When the prediction deviation assessment result exceeds a preset assessment threshold or the real-time spectral features are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for the low-temperature crystallization process of peptide drugs have insufficient adaptability to batch differences and insufficient ability to identify abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate warnings. By predicting the actual state through a digital twin model and reflecting it with online spectra, this method can adaptively capture batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0066] Specifically, in step S101, batch information of the peptide drug is acquired, as well as crystallization process data and online spectral data of the peptide drug during low-temperature crystallization. The batch information includes raw material batch information and solvent batch information, etc. The crystallization process data includes various data reflecting the macroscopic state of the crystallization process, such as temperature, pressure, stirring rate, and feed rate, which can be acquired in real time through various sensors. The online spectral data can be acquired using equipment such as an online Raman spectrometer. Online spectral data can reflect the vibration and rotation information of molecules and is closely related to the peptide, solvent, impurities, and crystal structure.

[0067] Specifically, in step S102, a preset digital twin model is used to simulate and predict the crystallization process. Batch information and crystallization process data are used as model inputs, enabling the model to reflect the specific state of the current batch and provide key state parameters for subsequent steps. These key state parameters are important indicators in the process, including peptide concentration and crystal size. The digital twin model refers to a digital model highly similar to the physical entity (i.e., a model constructed using digital twin technology) built in virtual space through data acquisition, transmission, processing, and analysis, modeling, and simulation. Specifically, it can include a peptide crystallization kinetic model under a specific solvent system. Digital twin models are existing technology and will not be described in detail here.

[0068] Specifically, in step S103, feature vectors related to the crystallization process are extracted from the online Raman spectroscopy data to obtain real-time spectral features, including:

[0069] The online Raman spectral data is preprocessed to obtain preprocessed Raman spectral data;

[0070] The characteristic peak areas related to peptide concentration, characteristic peak positions related to solvent composition, characteristic peak intensity ratios related to crystal form, and specific band integral areas related to impurities or intermediate products during crystallization are extracted from the preprocessed Raman spectral data to construct the corresponding feature vectors and obtain real-time spectral features.

[0071] Specifically, in step S103, the online Raman spectral data is preprocessed to obtain preprocessed Raman spectral data, including:

[0072] The online Raman spectral data is filtered to obtain filtered Raman spectral data.

[0073] The filtered Raman spectral data is smoothed using a smoothing algorithm to obtain smoothed Raman spectral data.

[0074] The least squares method was used to correct the fluorescence background interference in the smoothed Raman spectral data, resulting in preprocessed Raman spectral data.

[0075] In step S103, the raw online Raman spectral data may contain random noise and fluorescence background interference, which can mask the true Raman signal and lead to inaccurate feature vectors extracted from the data. To address this issue, the online Raman spectral data is first filtered (e.g., Fourier transform filtering or wavelet filtering) to remove high-frequency noise, resulting in data with reduced noise. Next, the filtered data is smoothed (e.g., using Savitzky-Golay smoothing or moving average smoothing algorithms) to further reduce data fluctuations and flatten the spectral curve, which is beneficial for subsequent peak identification and measurement. Finally, the least squares method is used to correct the fluorescence background in the smoothed data, separating the true Raman signal and eliminating the influence of fluorescence interference on peak intensity and area measurements. These preprocessing steps yield high-quality Raman spectral data. The filtering, smoothing, and least squares methods are existing techniques and will not be detailed here.

[0076] In step S103, the features extracted from the preprocessed Raman spectroscopy data include characteristic peak areas related to peptide concentration, characteristic peak positions related to solvent composition, characteristic peak intensity ratios related to crystal form, and specific band integral areas related to impurities or intermediate products during crystallization. These features represent key information during the crystallization process: changes in peptide content, changes in the solvent environment, the formation and transformation of crystal structures, and potential side reactions or impurity generation. By extracting these physicochemically significant features, high-dimensional spectral data can be transformed into low-dimensional and information-rich feature vectors.

[0077] Specifically, in step S104, the degree of inconsistency between key state parameters and real-time spectral characteristics is calculated to obtain the prediction deviation assessment result, including:

[0078] The key state parameters and real-time spectral features were standardized separately to obtain the standardized key state parameters and standardized real-time spectral features.

[0079] Based on the uncertainty of key state parameters and the extraction deviation of real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in the parameter weighted deviation vector and the spectral weighted deviation vector.

[0080] The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into the preset deviation prediction model to calculate the prediction deviation evaluation result.

[0081] In step S104, existing standardization methods, such as Z-score standardization, are used to standardize the key state parameters and real-time spectral features. Standardization eliminates differences in dimensions and numerical ranges between different parameters and features, making different types of data comparable. This standardization process is existing technology and will not be described in detail here.

[0082] Specifically, in step S104, based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector, including:

[0083] Based on the uncertainty of key state parameters and the extraction deviation of real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined and denoted as the first weight and the second weight, respectively. Among them, the greater the uncertainty, the smaller the first weight; the greater the extraction deviation, the smaller the second weight.

[0084] Based on the first and second weights, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in the parameter weighted deviation vector and the spectral weighted deviation vector.

[0085] Specifically, in step S104, based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined, including:

[0086] The uncertainty of key state parameters is determined by repeatedly inputting specific batch information and crystallization process data into a preset digital twin model to calculate the variance of key state parameters.

[0087] The extraction bias of real-time spectral features is determined by calculating the actual difference of real-time spectral features obtained by extracting feature vectors related to the crystallization process from online Raman spectral data multiple times.

[0088] Based on uncertainty and extraction bias, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined.

[0089] In step S104, to quantify the uncertainty of the model prediction, the simulation process (inputting specific batch information and real-time acquired crystallization process data into the digital twin model for simulation) is repeated multiple times. Each simulation may introduce minor perturbations or use different model parameters for sampling. The variance of the key state parameter results obtained from multiple simulations is calculated. The variance value reflects the stability of the model prediction, i.e., the uncertainty. Simultaneously, to quantify the feature extraction bias, feature extraction (i.e., the step of extracting feature vectors related to the crystallization process from online Raman spectroscopy data) is repeated multiple times on the same set of original spectral data. The average value of the extracted spectral features is calculated, thereby calculating the actual difference between the real-time spectral features and the average value. This actual difference reflects the stability of the feature extraction process, i.e., the extraction bias. Based on the calculated uncertainty (variance) and extraction bias (actual difference), the weights of the key state parameters and real-time spectral features in calculating the weighted bias are determined. Key state parameters with high uncertainty are assigned lower weights, and real-time spectral features with large extraction bias are assigned lower weights. Therefore, when calculating the weighted bias between the key state parameters and real-time spectral features, the influence of uncertain or unreliable data is reduced, making the calculation of the weighted bias more objective and accurate. This more reliable deviation assessment helps to more accurately identify anomalies in the low-temperature crystallization process, especially for minor batch-to-batch variations or early, rare abnormal patterns. More reliable deviation assessment can improve the sensitivity and accuracy of early warning.

[0090] In step S104, the standardized key state parameters and standardized real-time spectral features are multiplied by their corresponding weights to calculate the parameter weighted deviation vector and the spectral weighted deviation vector. Each component of the weighted deviation vector represents the difference between the predicted value and the actual value of the corresponding parameter or feature.

[0091] The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into a pre-defined deviation prediction model, which outputs a prediction deviation evaluation result. This evaluation result reflects the degree of deviation between the current process state and the normal state. The deviation prediction model can be a machine learning model, such as a support vector machine, random forest, or neural network algorithm. This deviation prediction model is used to accurately evaluate the deviation between the key state parameters predicted by the digital twin model and the spectral features acquired in real time.

[0092] The preset deviation prediction model is a model constructed using historical data to represent the nonlinear relationship between key state parameters and real-time spectral characteristics; the construction steps of the preset deviation prediction model include:

[0093] Retrieve historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results between historical crystallization process data and historical spectral data for multiple peptide drugs from the database;

[0094] Based on historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results, corresponding historical key state parameters and historical spectral characteristics are generated.

[0095] Based on historical key state parameters and historical spectral characteristics, a pre-defined deviation prediction model is constructed using machine learning algorithms.

[0096] Specifically, historical data is collected from a database, including information on different batches (normal batches) of peptide drugs, crystallization process data, online Raman spectroscopy data, and known nonlinear relationships between these historical data. This historical data provides the foundational information needed to train the model, covering different batches and process conditions. Based on this historical data, corresponding historical critical state parameters and historical spectral features are generated. This step transforms the raw historical data into a form directly related to the deviation to be evaluated, enabling the model to learn the correspondence between critical state parameters and spectral features. Using machine learning algorithms, a pre-defined deviation prediction model is constructed based on the generated historical critical state parameters and historical spectral features. Machine learning algorithms can learn and establish a complex nonlinear mapping relationship between critical state parameters and real-time spectral features from historical data. By learning this nonlinear relationship, the model can more accurately understand the correlation between critical state parameters and real-time spectral features, thus, in practical applications, when standardized critical state parameters and standardized real-time spectral features are input, it can output a prediction deviation evaluation result that reflects the degree of inconsistency between the two.

[0097] Specifically, in step S105, when the prediction deviation evaluation result is detected to be greater than a preset evaluation threshold, or when the real-time spectral characteristics are abnormal data, an early warning signal is issued, including:

[0098] Based on historical spectral data in the database, pre-determine the conditions for identifying abnormal data;

[0099] Based on the abnormal data determination conditions, it is determined whether the real-time spectral features are abnormal data. At the same time, the prediction deviation assessment results are compared with the preset assessment threshold.

[0100] An early warning signal is issued when the prediction deviation assessment result is greater than the preset assessment threshold, or when the real-time spectral features are abnormal data.

[0101] In step S105, historical spectral data accumulated in the database is used to analyze the normal batch data and known abnormal pattern data contained therein, and specific conditions for identifying abnormal data are pre-set. These conditions can be based on the numerical range of spectral characteristics, the appearance or disappearance of specific peaks, peak area ratios, etc. This step provides a basis for subsequent real-time judgment.

[0102] The conditions for identifying abnormal data include numerical anomalies and pattern anomalies.

[0103] In step S105, based on the abnormal data determination conditions, it is determined whether the real-time spectral features are abnormal data, including:

[0104] Spectral data of historical normal batches and spectral data of historical abnormal patterns are extracted from historical spectral data;

[0105] The Mahalanobis distance between real-time spectral features and historical normal batches of spectral data is calculated. The Mahalanobis distance is used to determine whether the corresponding real-time spectral features are abnormal data by checking whether the Mahalanobis distance is greater than a preset normal threshold. When the Mahalanobis distance is greater than the preset normal threshold, the corresponding real-time spectral features are determined to meet the numerical anomaly conditions and are therefore determined to be abnormal data.

[0106] The Euclidean distance between real-time spectral features and spectral data of historical abnormal patterns is calculated. The Euclidean distance is used to determine whether the corresponding real-time spectral feature is abnormal data by checking whether the Euclidean distance is less than a preset abnormal threshold. When the Euclidean distance is less than the preset abnormal threshold, the corresponding real-time spectral feature is determined to meet the pattern abnormality condition and is therefore determined to be abnormal data.

[0107] In step S105, when the real-time spectral features are extracted, the Mahalanobis distance between these features and the mean of historical normal batch spectral data is calculated and compared with a preset normal threshold. If the Mahalanobis distance exceeds the normal threshold, it indicates that the real-time spectral features deviate from the normal distribution numerically, meeting the conditions for numerical anomalies, and a warning signal is issued. The preset normal threshold can be set according to actual needs.

[0108] Simultaneously, the Euclidean distance between the real-time spectral features and each anomalous pattern in the historical anomalous pattern library is calculated and compared with a preset anomalous threshold. If the Euclidean distance with any historical anomalous pattern is less than the anomalous threshold, it indicates that the real-time spectral features are similar to a known anomalous pattern, meeting the pattern anomalous condition, and an early warning signal is issued. Through this dual judgment mechanism, even if the numerical deviation of the real-time spectral features is not significant, it can still be identified when its pattern is similar to a historical anomalous pattern. This effectively addresses the inherent batch-to-batch variability in the low-temperature crystallization process of peptide drugs and the scarcity of early and rare anomalous patterns in historical data, achieving rapid and accurate early warning of early and rare anomalous patterns before significant deviations in the process state.

[0109] In step S105, it is determined whether the real-time spectral features are abnormal data, and the prediction deviation evaluation result is compared with the preset evaluation threshold. When the prediction deviation evaluation result is detected to be greater than the preset evaluation threshold, or when the real-time spectral features are detected to be abnormal data, an early warning signal is issued.

[0110] By introducing abnormal data judgment conditions based on historical data and comparing the magnitude of prediction deviation assessment results with preset assessment thresholds, a dual independent dimension of anomaly detection is added. This not only detects anomalies based on the deviation between the digital twin model's predicted state and the actual spectral characteristics, but also directly identifies abnormal patterns in the spectral data itself, even if these abnormal patterns have not yet led to significant prediction deviations. This dual judgment mechanism improves the sensitivity and reliability of early warnings, enabling more comprehensive coverage of potential anomalies and addressing the problems caused by relying solely on deviation judgments or lacking clear methods for judging abnormal data.

[0111] For example, when pre-determining the conditions for identifying abnormal data, historical normal batch spectral data can be analyzed to establish normal fluctuation ranges for key spectral features (such as the intensity, area, and position of specific peaks). Simultaneously, historical abnormal batch spectral data can be analyzed to identify spectral patterns associated with the abnormality, such as the location and intensity threshold of specific impurity peaks. After acquiring online Raman spectral data of the low-temperature crystallization process of peptide drugs in real time, real-time spectral features can be extracted, such as calculating the area and position of key feature peaks. To determine whether real-time spectral features are abnormal, the extracted features can be compared with the pre-determined normal fluctuation range; if they exceed the range, they are judged as numerically abnormal. Simultaneously, it is checked whether features match historical abnormal patterns; for example, if a signal matching a historical abnormal pattern is detected in a specific band (using Mahalanobis and Euclidean distances), it is judged as pattern abnormal. If the real-time spectral features meet either the numerical abnormality condition or the pattern abnormality condition, they are determined to be abnormal data. At the same time, the system calculates the current prediction deviation assessment result and compares it with a preset assessment threshold. For example, the preset assessment threshold is set to 0.8. If the calculated prediction deviation assessment result is 0.9, which is greater than 0.8, the early warning condition is met. Alternatively, if the real-time spectral characteristics are judged to be abnormal data, the early warning condition is met even if the prediction deviation assessment result is less than or equal to 0.8. When either condition is met, the system immediately issues an early warning signal, such as through an audible and visual alarm or by sending a notification to the operator, indicating that there may be an anomaly in the low-temperature crystallization process, requiring further inspection or intervention.

[0112] As can be seen from the above, this early warning method for the low-temperature crystallization process of peptide drugs based on digital twins acquires batch information of the peptide drug, as well as crystallization process data and online spectral data of the peptide drug during the low-temperature crystallization process. The batch information and crystallization process data are input into a preset digital twin model to simulate the low-temperature crystallization process, predict the state parameters of the peptide drug at the next moment during the low-temperature crystallization process, extract feature vectors related to the crystallization process from the online spectral data to obtain real-time spectral features, calculate the degree of inconsistency between key state parameters and real-time spectral features, and obtain a prediction deviation assessment result. When the prediction deviation assessment result is detected to be greater than a preset assessment threshold, or the real-time spectral features are abnormal, the method will proceed accordingly. Based on the time, an early warning signal is issued; thus, by calculating the degree of inconsistency between key state parameters and real-time spectral characteristics, a prediction deviation assessment result is obtained, and when the prediction deviation assessment result is greater than the preset assessment threshold or the real-time spectral characteristics are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for low-temperature crystallization processes of peptide drugs have insufficient batch-to-batch adaptability and insufficient ability to identify abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate early warnings. By predicting the actual state through a digital twin model and reflecting it with online spectra, the method can adaptively capture batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0113] refer to Figure 2 This application provides a digital twin-based early warning system for the low-temperature crystallization process of peptide drugs, used to provide early warning of abnormalities in the low-temperature crystallization process of peptide drugs, including:

[0114] Module 1 is used to acquire batch information of peptide drugs, as well as crystallization process data and online spectral data of peptide drugs during low-temperature crystallization.

[0115] Prediction module 2 is used to input batch information and crystallization process data into a preset digital twin model to simulate the low-temperature crystallization process and predict the state parameters of the peptide drug at the next moment during the low-temperature crystallization process.

[0116] Extraction module 3 is used to extract feature vectors related to the crystallization process from online spectral data to obtain real-time spectral features;

[0117] Calculation module 4 is used to calculate the degree of inconsistency between key state parameters and real-time spectral characteristics, and obtain the prediction deviation assessment results;

[0118] The early warning module 5 is used to issue an early warning signal when the predicted deviation evaluation result is greater than the preset evaluation threshold, or when the real-time spectral characteristics are abnormal data.

[0119] This digital twin-based early warning system for the low-temperature crystallization process of peptide drugs calculates the degree of inconsistency between key state parameters and real-time spectral features to obtain a prediction deviation assessment result. When the prediction deviation assessment result exceeds a preset assessment threshold or the real-time spectral features are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for the low-temperature crystallization process of peptide drugs have insufficient adaptability to batch differences and insufficient ability to identify abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate warnings. By predicting the actual state through a digital twin model and reflecting the online spectrum, it can adaptively capture the batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0120] Specifically, during execution, module 1 acquires batch information of the peptide drug, as well as crystallization process data and online spectral data during the low-temperature crystallization process. The batch information includes raw material batch information and solvent batch information. The crystallization process data includes various data reflecting the macroscopic state of the crystallization process, such as temperature, pressure, stirring rate, and feed rate, which can be acquired in real time through various sensors. The online spectral data can be acquired using equipment such as an online Raman spectrometer. Online spectral data reflects the vibration and rotation information of molecules and is closely related to the peptide, solvent, impurities, and crystal structure.

[0121] Specifically, during execution, prediction module 2 utilizes a pre-set digital twin model to simulate and predict the crystallization process. Batch information and crystallization process data serve as model inputs, enabling the model to reflect the specific state of the current batch and provide key state parameters for subsequent steps. These key state parameters are important indicators in the process, including peptide concentration and crystal size. The digital twin model refers to a digital model highly similar to the physical entity (i.e., a model constructed using digital twin technology) built in virtual space through data acquisition, transmission, processing, and analysis, modeling, and simulation. Specifically, it can include peptide crystallization kinetics models under specific solvent systems. Digital twin models are existing technology and will not be detailed here.

[0122] Specifically, when extraction module 3 extracts feature vectors related to the crystallization process from online Raman spectral data to obtain real-time spectral features, it performs the following:

[0123] The online Raman spectral data is preprocessed to obtain preprocessed Raman spectral data;

[0124] The characteristic peak areas related to peptide concentration, characteristic peak positions related to solvent composition, characteristic peak intensity ratios related to crystal form, and specific band integral areas related to impurities or intermediate products during crystallization are extracted from the preprocessed Raman spectral data to construct the corresponding feature vectors and obtain real-time spectral features.

[0125] Specifically, when extraction module 3 preprocesses the online Raman spectral data to obtain preprocessed Raman spectral data, it executes the following:

[0126] The online Raman spectral data is filtered to obtain filtered Raman spectral data.

[0127] The filtered Raman spectral data is smoothed using a smoothing algorithm to obtain smoothed Raman spectral data.

[0128] The least squares method was used to correct the fluorescence background interference in the smoothed Raman spectral data, resulting in preprocessed Raman spectral data.

[0129] During the execution of extraction module 3, the raw online Raman spectral data may contain random noise and fluorescence background interference. These factors can mask the true Raman signal, leading to inaccurate feature vectors extracted from the data. To address this issue, the online Raman spectral data is first filtered (e.g., Fourier transform filtering or wavelet filtering) to remove high-frequency noise, resulting in data with reduced noise. Next, the filtered data is smoothed (e.g., using Savitzky-Golay smoothing or moving average smoothing algorithms) to further reduce data fluctuations and flatten the spectral curve, which is beneficial for subsequent peak identification and measurement. Finally, the least squares method is used to correct the fluorescence background in the smoothed data, separating the true Raman signal and eliminating the influence of fluorescence interference on peak intensity and area measurements. These preprocessing steps yield high-quality Raman spectral data. The filtering, smoothing, and least squares methods are existing technologies and will not be detailed here.

[0130] During execution, extraction module 3 extracts features from the preprocessed Raman spectral data, including characteristic peak areas related to peptide concentration, characteristic peak positions related to solvent composition, characteristic peak intensity ratios related to crystal form, and specific band integral areas related to impurities or intermediate products during crystallization. These features represent key information during the crystallization process: changes in peptide content, changes in the solvent environment, the formation and transformation of crystal structures, and potential side reactions or impurity generation. By extracting these physicochemically significant features, high-dimensional spectral data can be transformed into low-dimensional and information-rich feature vectors.

[0131] Specifically, when calculating the degree of inconsistency between key state parameters and real-time spectral characteristics to obtain the prediction deviation assessment result, calculation module 4 executes:

[0132] The key state parameters and real-time spectral features were standardized separately to obtain the standardized key state parameters and standardized real-time spectral features.

[0133] Based on the uncertainty of key state parameters and the extraction deviation of real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in the parameter weighted deviation vector and the spectral weighted deviation vector.

[0134] The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into the preset deviation prediction model to calculate the prediction deviation evaluation result.

[0135] When calculation module 4 is executed, it uses existing standardization methods, such as Z-score standardization, to standardize key state parameters and real-time spectral features. Standardization eliminates differences in dimensions and numerical ranges between different parameters and features, making different types of data comparable. Standardization is an existing technology and will not be described in detail here.

[0136] Specifically, when calculation module 4 calculates the weighted deviation between the standardized key state parameters and the standardized real-time spectral features based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, and obtains the parameter weighted deviation vector and the spectral weighted deviation vector, the following is executed:

[0137] Based on the uncertainty of key state parameters and the extraction deviation of real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined and denoted as the first weight and the second weight, respectively. Among them, the greater the uncertainty, the smaller the first weight; the greater the extraction deviation, the smaller the second weight.

[0138] Based on the first and second weights, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in the parameter weighted deviation vector and the spectral weighted deviation vector.

[0139] Specifically, when calculation module 4 determines the weights of the corresponding standardized key state parameters and the weights of the corresponding standardized real-time spectral features based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, it executes the following:

[0140] The uncertainty of key state parameters is determined by repeatedly inputting specific batch information and crystallization process data into a preset digital twin model to calculate the variance of key state parameters.

[0141] The extraction bias of real-time spectral features is determined by calculating the actual difference of real-time spectral features obtained by extracting feature vectors related to the crystallization process from online Raman spectral data multiple times.

[0142] Based on uncertainty and extraction bias, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined.

[0143] During execution, calculation module 4 repeats the simulation process (inputting specific batch information and real-time acquired crystallization process data into the digital twin model for simulation) multiple times to quantify the uncertainty of model predictions. Each simulation may introduce minor perturbations or use different model parameters for sampling. The variance of the key state parameter results obtained from multiple simulations is calculated. The variance value reflects the stability of the model prediction, i.e., the uncertainty. Simultaneously, to quantify the feature extraction bias, feature extraction (i.e., the step of extracting feature vectors related to the crystallization process from online Raman spectroscopy data) is repeated multiple times on the same set of original spectral data. The average value of the extracted spectral features is calculated, thus obtaining the actual difference between the real-time spectral features and the average value. This actual difference reflects the stability of the feature extraction process, i.e., the extraction bias. Based on the calculated uncertainty (variance) and extraction bias (actual difference), the weights of key state parameters and real-time spectral features in calculating the weighted bias are determined. Key state parameters with high uncertainty are assigned lower weights, and real-time spectral features with large extraction bias are assigned lower weights. Therefore, the influence of uncertain or unreliable data is reduced when calculating the weighted bias between key state parameters and real-time spectral features, making the calculation of the weighted bias more objective and accurate. This more reliable deviation assessment helps to more accurately identify anomalies in the low-temperature crystallization process, especially for minor batch-to-batch variations or early, rare abnormal patterns. More reliable deviation assessment can improve the sensitivity and accuracy of early warning.

[0144] When the calculation module 4 is executed, it multiplies the standardized key state parameters and the standardized real-time spectral features by the corresponding weights to calculate the parameter weighted deviation vector and the spectral weighted deviation vector. Each component of the weighted deviation vector represents the difference between the predicted value and the actual value of the corresponding parameter or feature.

[0145] The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into a pre-defined deviation prediction model, which outputs a prediction deviation evaluation result. This evaluation result reflects the degree of deviation between the current process state and the normal state. The deviation prediction model can be a machine learning model, such as a support vector machine, random forest, or neural network algorithm. This deviation prediction model is used to accurately evaluate the deviation between the key state parameters predicted by the digital twin model and the spectral features acquired in real time.

[0146] The preset deviation prediction model is a model constructed using historical data to represent the nonlinear relationship between key state parameters and real-time spectral characteristics; the construction steps of the preset deviation prediction model include:

[0147] Retrieve historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results between historical crystallization process data and historical spectral data for multiple peptide drugs from the database;

[0148] Based on historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results, corresponding historical key state parameters and historical spectral characteristics are generated.

[0149] Based on historical key state parameters and historical spectral characteristics, a pre-defined deviation prediction model is constructed using machine learning algorithms.

[0150] Specifically, historical data is collected from a database, including information on different batches (normal batches) of peptide drugs, crystallization process data, online Raman spectroscopy data, and known nonlinear relationships between these historical data. This historical data provides the foundational information needed to train the model, covering different batches and process conditions. Based on this historical data, corresponding historical critical state parameters and historical spectral features are generated. This step transforms the raw historical data into a form directly related to the deviation to be evaluated, enabling the model to learn the correspondence between critical state parameters and spectral features. Using machine learning algorithms, a pre-defined deviation prediction model is constructed based on the generated historical critical state parameters and historical spectral features. Machine learning algorithms can learn and establish a complex nonlinear mapping relationship between critical state parameters and real-time spectral features from historical data. By learning this nonlinear relationship, the model can more accurately understand the correlation between critical state parameters and real-time spectral features, thus, in practical applications, when standardized critical state parameters and standardized real-time spectral features are input, it can output a prediction deviation evaluation result that reflects the degree of inconsistency between the two.

[0151] Specifically, when the early warning module 5 issues an early warning signal and detects that the prediction deviation assessment result is greater than the preset assessment threshold, or that the real-time spectral characteristics are abnormal data, it executes the following:

[0152] Based on historical spectral data in the database, pre-determine the conditions for identifying abnormal data;

[0153] Based on the abnormal data determination conditions, it is determined whether the real-time spectral features are abnormal data. At the same time, the prediction deviation assessment results are compared with the preset assessment threshold.

[0154] An early warning signal is issued when the prediction deviation assessment result is greater than the preset assessment threshold, or when the real-time spectral features are abnormal data.

[0155] When the early warning module 5 is executed, it utilizes historical spectral data accumulated in the database to analyze the normal batch data and known abnormal pattern data contained therein, and pre-sets specific conditions for identifying abnormal data. These conditions can be based on the numerical range of spectral characteristics, the appearance or disappearance of specific peaks, peak area ratios, etc. This step provides a basis for subsequent real-time judgment.

[0156] The conditions for identifying abnormal data include numerical anomalies and pattern anomalies.

[0157] When the early warning module 5 determines whether the real-time spectral characteristics are abnormal based on abnormal data determination conditions, it executes the following:

[0158] Spectral data of historical normal batches and spectral data of historical abnormal patterns are extracted from historical spectral data;

[0159] The Mahalanobis distance between real-time spectral features and historical normal batches of spectral data is calculated. The Mahalanobis distance is used to determine whether the corresponding real-time spectral features are abnormal data by checking whether the Mahalanobis distance is greater than a preset normal threshold. When the Mahalanobis distance is greater than the preset normal threshold, the corresponding real-time spectral features are determined to meet the numerical anomaly conditions and are therefore determined to be abnormal data.

[0160] The Euclidean distance between real-time spectral features and spectral data of historical abnormal patterns is calculated. The Euclidean distance is used to determine whether the corresponding real-time spectral feature is abnormal data by checking whether the Euclidean distance is less than a preset abnormal threshold. When the Euclidean distance is less than the preset abnormal threshold, the corresponding real-time spectral feature is determined to meet the pattern abnormality condition and is therefore determined to be abnormal data.

[0161] When the early warning module 5 is executed, upon extracting real-time spectral features, it calculates the Mahalanobis distance between these features and the mean of historical normal batch spectral data, and compares it with a preset normal threshold. If the Mahalanobis distance exceeds the normal threshold, it indicates that the real-time spectral features deviate numerically from the normal distribution, meeting the conditions for numerical anomalies, and an early warning signal is issued. The preset normal threshold can be set according to actual needs.

[0162] Simultaneously, the Euclidean distance between the real-time spectral features and each anomalous pattern in the historical anomalous pattern library is calculated and compared with a preset anomalous threshold. If the Euclidean distance with any historical anomalous pattern is less than the anomalous threshold, it indicates that the real-time spectral features are similar to a known anomalous pattern, meeting the pattern anomalous condition, and an early warning signal is issued. Through this dual judgment mechanism, even if the numerical deviation of the real-time spectral features is not significant, it can still be identified when its pattern is similar to a historical anomalous pattern. This effectively addresses the inherent batch-to-batch variability in the low-temperature crystallization process of peptide drugs and the scarcity of early and rare anomalous patterns in historical data, achieving rapid and accurate early warning of early and rare anomalous patterns before significant deviations in the process state.

[0163] In step S105, it is determined whether the real-time spectral features are abnormal data, and the prediction deviation evaluation result is compared with the preset evaluation threshold. When the prediction deviation evaluation result is detected to be greater than the preset evaluation threshold, or when the real-time spectral features are detected to be abnormal data, an early warning signal is issued.

[0164] By introducing abnormal data judgment conditions based on historical data and comparing the magnitude of prediction deviation assessment results with preset assessment thresholds, a dual independent dimension of anomaly detection is added. This not only detects anomalies based on the deviation between the digital twin model's predicted state and the actual spectral characteristics, but also directly identifies abnormal patterns in the spectral data itself, even if these abnormal patterns have not yet led to significant prediction deviations. This dual judgment mechanism improves the sensitivity and reliability of early warnings, enabling more comprehensive coverage of potential anomalies and addressing the problems caused by relying solely on deviation judgments or lacking clear methods for judging abnormal data.

[0165] For example, when pre-determining the conditions for identifying abnormal data, historical normal batch spectral data can be analyzed to establish normal fluctuation ranges for key spectral features (such as the intensity, area, and position of specific peaks). Simultaneously, historical abnormal batch spectral data can be analyzed to identify spectral patterns associated with the abnormality, such as the location and intensity threshold of specific impurity peaks. After acquiring online Raman spectral data of the low-temperature crystallization process of peptide drugs in real time, real-time spectral features can be extracted, such as calculating the area and position of key feature peaks. To determine whether real-time spectral features are abnormal, the extracted features can be compared with the pre-determined normal fluctuation range; if they exceed the range, they are judged as numerically abnormal. Simultaneously, it is checked whether features match historical abnormal patterns; for example, if a signal matching a historical abnormal pattern is detected in a specific band (using Mahalanobis and Euclidean distances), it is judged as pattern abnormal. If the real-time spectral features meet either the numerical abnormality condition or the pattern abnormality condition, they are determined to be abnormal data. At the same time, the system calculates the current prediction deviation assessment result and compares it with a preset assessment threshold. For example, the preset assessment threshold is set to 0.8. If the calculated prediction deviation assessment result is 0.9, which is greater than 0.8, the early warning condition is met. Alternatively, if the real-time spectral characteristics are judged to be abnormal data, the early warning condition is met even if the prediction deviation assessment result is less than or equal to 0.8. When either condition is met, the system immediately issues an early warning signal, such as through an audible and visual alarm or by sending a notification to the operator, indicating that there may be an anomaly in the low-temperature crystallization process, requiring further inspection or intervention.

[0166] As described above, this early warning system for the low-temperature crystallization process of peptide drugs based on digital twins acquires batch information of the peptide drugs, as well as crystallization process data and online spectral data during the low-temperature crystallization process. The batch information and crystallization process data are input into a preset digital twin model to simulate the low-temperature crystallization process, predict the state parameters of the peptide drugs at the next moment during the low-temperature crystallization process, extract feature vectors related to the crystallization process from the online spectral data to obtain real-time spectral features, calculate the degree of inconsistency between key state parameters and real-time spectral features, and obtain a prediction deviation assessment result. When the prediction deviation assessment result is detected to be greater than a preset assessment threshold, or the real-time spectral features are abnormal, the system will take action. Based on the time, an early warning signal is issued; thus, by calculating the degree of inconsistency between key state parameters and real-time spectral characteristics, a prediction deviation assessment result is obtained, and when the prediction deviation assessment result is greater than the preset assessment threshold or the real-time spectral characteristics are abnormal data, an early warning signal is issued. This solves the problem that existing early warning methods for low-temperature crystallization processes of peptide drugs have insufficient batch-to-batch adaptability and insufficient ability to identify abnormal patterns when using digital twins and machine learning methods, making it difficult to issue accurate early warnings. By predicting the actual state through a digital twin model and reflecting it with online spectra, the method can adaptively capture batch-to-batch variability and abnormal patterns in the low-temperature crystallization process of peptide drugs, thereby improving the early warning efficiency of the low-temperature crystallization process of peptide drugs.

[0167] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0168] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0169] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0170] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0171] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for early warning of low-temperature crystallization process of peptide drugs based on digital twins, used to provide early warning of abnormalities in the low-temperature crystallization process of peptide drugs, characterized in that, Including the following steps: Obtain specific batch information of the peptide drug to be tested, and acquire real-time crystallization process data and online spectral data of the peptide drug during the low-temperature crystallization process; The specific batch information and the crystallization process data are input into a preset digital twin model to simulate the low-temperature crystallization process and predict the key state parameters of the test peptide drug at the next moment in the low-temperature crystallization process. Feature vectors related to the crystallization process are extracted from the online spectral data to obtain real-time spectral features; The degree of inconsistency between the key state parameters and the real-time spectral features is calculated to obtain the prediction deviation assessment result; When the prediction deviation assessment result is detected to be greater than the preset assessment threshold, or when the real-time spectral features are abnormal data, an early warning signal is issued. Feature vectors related to the crystallization process are extracted from the online spectral data to obtain real-time spectral features, including: The online spectral data is preprocessed to obtain preprocessed Raman spectral data; The characteristic peak area related to peptide concentration, the characteristic peak position related to solvent composition, the characteristic peak intensity ratio related to crystal form, and the specific band integral area related to impurities or intermediate products in the crystallization process are extracted from the preprocessed Raman spectral data to construct the corresponding feature vector and obtain real-time spectral features. Calculate the degree of inconsistency between the key state parameters and the real-time spectral features to obtain the prediction deviation assessment result, including: The key state parameters and the real-time spectral features are standardized respectively to obtain the standardized key state parameters and standardized real-time spectral features. Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector. The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into a preset deviation prediction model to calculate the prediction deviation evaluation result. Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector, including: Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined and denoted as the first weight and the second weight, respectively; wherein, the greater the uncertainty, the smaller the first weight, and the greater the extraction deviation, the smaller the second weight. Based on the first weight and the second weight, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector. Based on the uncertainty of the key state parameters and the extraction bias of the real-time spectral features, the weights of the corresponding standardized key state parameters and the weights of the corresponding standardized real-time spectral features are determined, including: The uncertainty of the key state parameters is determined by repeatedly inputting the specific batch information and the crystallization process data into a preset digital twin model to calculate the variance of the key state parameters. The extraction deviation of the real-time spectral features is determined by calculating the actual difference of the real-time spectral features obtained by extracting feature vectors related to the crystallization process from the online spectral data multiple times. Based on the uncertainty and the extraction bias, the weights of the corresponding standardized key state parameters and the weights of the corresponding standardized real-time spectral features are determined.

2. The method for early warning of low-temperature crystallization process of polypeptide drugs based on digital twins according to claim 1, characterized in that, The online spectral data is preprocessed to obtain preprocessed Raman spectral data, including: The online spectral data is filtered to obtain filtered Raman spectral data; The filtered Raman spectral data is smoothed using a smoothing algorithm to obtain smoothed Raman spectral data. The fluorescence background interference of the smoothed Raman spectral data was corrected by using the least squares method to obtain the preprocessed Raman spectral data.

3. The method for early warning of low-temperature crystallization process of polypeptide drugs based on digital twins according to claim 1, characterized in that, The preset deviation prediction model is a model constructed using historical data to represent the nonlinear relationship between key state parameters and real-time spectral characteristics. The steps for constructing the preset deviation prediction model include: Retrieve historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results between the historical crystallization process data and the historical spectral data for multiple polypeptide drugs from the database. Based on the historical batch information, historical crystallization process data, historical spectral data, and historical nonlinear relationship results, corresponding historical key state parameters and historical spectral features are generated. Based on the historical key state parameters and the historical spectral characteristics, a machine learning algorithm is used to construct the preset deviation prediction model.

4. The method for early warning of low-temperature crystallization process of polypeptide drugs based on digital twins according to claim 1, characterized in that, When the prediction deviation assessment result is detected to be greater than a preset assessment threshold, or when the real-time spectral features are abnormal data, an early warning signal is issued, including: Based on historical spectral data in the database, pre-determine the conditions for identifying abnormal data; Based on the abnormal data determination conditions, it is determined whether the real-time spectral features are abnormal data. At the same time, the relationship between the prediction deviation evaluation result and the preset evaluation threshold is compared. When the prediction deviation assessment result is detected to be greater than the preset assessment threshold, or when the real-time spectral features are detected to be abnormal data, an early warning signal is issued.

5. The method for early warning of low-temperature crystallization process of polypeptide drugs based on digital twins according to claim 4, characterized in that, The conditions for determining abnormal data include numerical anomalies and pattern anomalies. Based on the aforementioned abnormal data determination conditions, determining whether the real-time spectral features are abnormal data includes: The spectral data of historical normal batches and the spectral data of historical abnormal patterns are extracted from the historical spectral data. The Mahalanobis distance between the real-time spectral feature and the spectral data of the historical normal batch is calculated. The real-time spectral feature is determined to be abnormal data by whether the Mahalanobis distance is greater than a preset normal threshold. When the Mahalanobis distance is greater than the preset normal threshold, the real-time spectral feature is determined to meet the numerical abnormality condition and is determined to be abnormal data. The Euclidean distance between the real-time spectral feature and the spectral data of the historical abnormal pattern is calculated. The real-time spectral feature is determined to be abnormal data by whether the Euclidean distance is less than a preset abnormal threshold. When the Euclidean distance is less than the preset abnormal threshold, the real-time spectral feature is determined to meet the abnormal conditions of the pattern and is therefore determined to be abnormal data.

6. A digital twin-based early warning system for the low-temperature crystallization process of peptide drugs, used to provide early warning of abnormalities in the low-temperature crystallization process of peptide drugs, characterized in that, include: The acquisition module is used to acquire specific batch information of the peptide drug to be tested, and to acquire crystallization process data and online spectral data of the peptide drug in real time during the low-temperature crystallization process. The prediction module is used to input the specific batch information and the crystallization process data into a preset digital twin model to simulate the low-temperature crystallization process and predict the key state parameters of the test peptide drug at the next moment in the low-temperature crystallization process. The extraction module is used to extract feature vectors related to the crystallization process from the online spectral data to obtain real-time spectral features; The calculation module is used to calculate the degree of inconsistency between the key state parameters and the real-time spectral features, and obtain the prediction deviation evaluation result; The early warning module is used to issue an early warning signal when the prediction deviation evaluation result is detected to be greater than a preset evaluation threshold, or when the real-time spectral features are abnormal data. The extraction module is used to extract feature vectors related to the crystallization process from the online spectral data to obtain real-time spectral features, including: The online spectral data is preprocessed to obtain preprocessed Raman spectral data; The characteristic peak area related to peptide concentration, the characteristic peak position related to solvent composition, the characteristic peak intensity ratio related to crystal form, and the specific band integral area related to impurities or intermediate products in the crystallization process are extracted from the preprocessed Raman spectral data to construct the corresponding feature vector and obtain real-time spectral features. The calculation module is used to calculate the degree of inconsistency between the key state parameters and the real-time spectral features, and to obtain the prediction deviation evaluation result, including: The key state parameters and the real-time spectral features are standardized respectively to obtain the standardized key state parameters and standardized real-time spectral features. Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector. The parameter-weighted deviation vector and the spectral-weighted deviation vector are input into a preset deviation prediction model to calculate the prediction deviation evaluation result. Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector, including: Based on the uncertainty of the key state parameters and the extraction deviation of the real-time spectral features, the weights of the corresponding standardized key state parameters and the corresponding standardized real-time spectral features are determined and denoted as the first weight and the second weight, respectively; wherein, the greater the uncertainty, the smaller the first weight, and the greater the extraction deviation, the smaller the second weight. Based on the first weight and the second weight, the weighted deviation between the standardized key state parameters and the standardized real-time spectral features is calculated, resulting in a parameter weighted deviation vector and a spectral weighted deviation vector. Based on the uncertainty of the key state parameters and the extraction bias of the real-time spectral features, the weights of the corresponding standardized key state parameters and the weights of the corresponding standardized real-time spectral features are determined, including: The uncertainty of the key state parameters is determined by repeatedly inputting the specific batch information and the crystallization process data into a preset digital twin model to calculate the variance of the key state parameters. The extraction deviation of the real-time spectral features is determined by calculating the actual difference of the real-time spectral features obtained by extracting feature vectors related to the crystallization process from the online spectral data multiple times. Based on the uncertainty and the extraction bias, the weights of the corresponding standardized key state parameters and the weights of the corresponding standardized real-time spectral features are determined.

Images