A machine learning-driven quality control method for pharmaceutical intermediates
Through convolutional neural networks and recurrent neural networks, the physical and chemical characteristics and time series data of pharmaceutical intermediates are analyzed, and the virtual samples are generated by combining the generation of adversarial networks, which solves the nonlinear and long-term dependence problems that are difficult to deal with by traditional statistical methods, and realizes the pharmaceutical intermediates Real-time monitoring and prediction of quality improves prediction accuracy and automation capabilities of production lines.
Patent Information
- Application Number
- CN202411186727.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-28
AI Technical Summary
Traditional statistical methods are difficult to capture the nonlinear relationship and long-term dependence in the production process of pharmaceutical intermediates, making it difficult to detect abnormal changes in the production process in a timely manner.
Convolutional neural network (CNN) is used to extract deep features related to the physical and chemical characteristics of pharmaceutical intermediates, and analyze data sequences over time through recurrent neural network (RNN) to construct a comprehensive training data set containing deep features and time series data. Generate virtual samples using Generative Adversarial Networks (GANs) to enhance the diversity and complexity of training data, establish quality detection models and integrate them into the production line.
Effectively capture dynamic trends and abnormal changes in the production process, improve the generalization ability and prediction accuracy of the quality detection model, realize real-time monitoring and prediction of product quality, and solve the limitations of manual monitoring.
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information processing technology, and in particular to a method for quality control of pharmaceutical intermediates driven by machine learning. Background Art
[0002] At present, pharmaceutical intermediates refer to compounds produced during the drug synthesis process, and they are an important part of the production process of APIs. These intermediates are usually not final products, but are used as starting materials or reactants for further chemical reactions to synthesize active pharmaceutical ingredients (APIs). The quality of pharmaceutical intermediates directly affects the safety and effectiveness of the final drug product.
[0003] The quality control of pharmaceutical intermediates is a key link in ensuring the quality of drugs. It involves a series of processes, including but not limited to raw material inspection, monitoring during the synthesis of intermediates, and finished product inspection. The goal of quality control is to ensure that the intermediates meet the specified quality standards, including purity, impurity levels, stability, etc. During the production process of pharmaceutical intermediates, a large number of time-varying data series such as temperature, pressure, and pH value will be generated. Traditional statistical methods have difficulty capturing the nonlinear relationships and long-term dependencies in these data series, making it difficult to detect abnormal changes in the production process in a timely manner.
[0004] Therefore, there is a need for a machine learning-driven quality control method for pharmaceutical intermediates that can solve the above problems. Summary of the invention
[0005] The present invention provides a method for quality control of pharmaceutical intermediates driven by machine learning. The present invention uses a convolutional neural network (CNN) to extract deep features related to the physical and chemical properties of pharmaceutical intermediates, enhancing the model's ability to understand complex image data. By analyzing the data sequence that changes over time through a recurrent neural network (RNN), dynamic trends and abnormal changes in the production process can be captured, solving the nonlinear and long-term dependency problems that are difficult to handle with traditional statistical methods.
[0006] The technical solution adopted by the present invention to solve the above technical problems is: a method for quality control of pharmaceutical intermediates based on machine learning, characterized in that it comprises the following steps:
[0007] Step S1: Collecting multimodal data of pharmaceutical intermediates;
[0008] Step S2: performing data cleaning and standardization on the multimodal data;
[0009] Step S3: extracting deep features of pharmaceutical intermediates from the multimodal data using a convolutional neural network;
[0010] Step S4: applying a recurrent neural network to analyze the data sequence that changes over time during the production process of the pharmaceutical intermediate;
[0011] Step S5: Based on the deep features obtained in step S3 and the data sequence obtained in step S4, a comprehensive training data set including deep features and time series data is constructed;
[0012] Step S6: Generate virtual samples using a generative adversarial network to enhance the comprehensive training data set in step S5;
[0013] Step S7: Based on the feature type of the deep features extracted in step S3 and the data sequence obtained in step S4, a quality detection model is established, and hyper parameters of the quality detection model are defined;
[0014] Step S8: training a quality detection model using the comprehensive training data set obtained in step S6;
[0015] Step S9: Integrate the quality inspection model into the production line of pharmaceutical intermediates.
[0016] Furthermore, the multimodal data in step S1 includes microscope images, chemical structure images, sensor data and chemical analysis data of the pharmaceutical intermediates.
[0017] Furthermore, the sensor data includes the temperature, pressure and pH value of the pharmaceutical intermediate, and the chemical analysis data includes the chromatographic data and spectral data of the pharmaceutical intermediate.
[0018] Furthermore, the step S2 of cleaning and standardizing the multimodal data includes:
[0019] Step S2-1: Check the collected multimodal data, identify and delete invalid data points or missing values;
[0020] Step S2-2: using a machine learning algorithm to detect outliers in the data and decide whether to keep or delete the outliers;
[0021] Step S2-3: For missing data points, interpolation is used to fill them in to ensure the integrity of the data set;
[0022] Step S2-4: Scale the data to a uniform scale using minimum-maximum scaling or z-score normalization.
[0023] Furthermore, the step S3 of extracting the deep features of the pharmaceutical intermediates from the multimodal data using a convolutional neural network includes:
[0024] Step S3-1: Select CNN architecture;
[0025] Step S3-2: data preprocessing;
[0026] Step S3-3: construct a CNN model;
[0027] Step S3-4: model initialization;
[0028] Step S3-5: training the CNN model;
[0029] Step S3-6: feature extraction;
[0030] Step S3-7: Feature selection and optimization.
[0031] Furthermore, the step S4 of applying a recurrent neural network to analyze the data sequence that changes over time during the production process of the pharmaceutical intermediates includes:
[0032] Step S4-1: Select RNN architecture;
[0033] Step S4-2: data preprocessing;
[0034] Step S4-3: constructing an RNN model;
[0035] Step S4-4: model initialization;
[0036] Step S4-5: training the RNN model;
[0037] Step S4-6: feature extraction;
[0038] Step S4-7: Perform trend analysis and anomaly detection on the extracted features.
[0039] Furthermore, in step S5, based on the deep features obtained in step S3 and the data sequence obtained in step S4, constructing a comprehensive training data set including deep features and time series data includes:
[0040] Step S5-1: merge the deep features extracted in step S3 and the time series features extracted in step S4 into one dataset;
[0041] Step S5-2: Pairing the depth features of the image data with the time series data in the corresponding time period according to the timestamp of the production process;
[0042] Step S5-3: Split the comprehensive dataset into a training set, a validation set, and a test set;
[0043] Step S5-4: Assign a label to each sample in the comprehensive dataset.
[0044] 1. This invention uses convolutional neural networks (CNN) to extract deep features related to the physical and chemical properties of pharmaceutical intermediates, enhancing the model's ability to understand complex image data. By analyzing time-varying data sequences through recurrent neural networks (RNN), dynamic trends and abnormal changes in the production process can be captured, solving the nonlinear and long-term dependency problems that are difficult to handle with traditional statistical methods.
[0045] 2. The present invention constructs a comprehensive training data set containing deep features and time series data, realizes the effective fusion of different modal data, and provides a more comprehensive source of information for the model. The reasonable division of the data set (training / validation / test set) ensures the training quality and generalization ability of the model, and generates virtual samples through generative adversarial networks (GANs), which increases the diversity and complexity of training data, improves the generalization ability and prediction accuracy of quality detection models, and realizes real-time monitoring and prediction of product quality by integrating quality detection models into pharmaceutical intermediate production lines, solving the limitations of manual monitoring. DETAILED DESCRIPTION
[0046] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0047] Example 1: The present invention provides a method for quality control of pharmaceutical intermediates based on machine learning, comprising the following steps:
[0048] Step S1: Collecting multimodal data of pharmaceutical intermediates;
[0049] Step S2: performing data cleaning and standardization on the multimodal data to ensure data quality;
[0050] Step S3: extracting deep features related to the physicochemical properties of the pharmaceutical intermediate from the multimodal data using a convolutional neural network;
[0051] Step S4: applying a recurrent neural network to analyze the time-varying data series during the production process of pharmaceutical intermediates to identify potential trends and anomalies;
[0052] Step S5: Based on the deep features obtained in step S3 and the data sequence obtained in step S4, a comprehensive training data set including deep features and time series data is constructed;
[0053] Step S6: Generate virtual samples using a generative adversarial network (GAN) to enhance the comprehensive training data set in step S5, thereby improving the generalization ability and prediction accuracy of the model.
[0054] Step S7: Based on the feature types of the deep features extracted in step S3 and the data sequence obtained in step S4, a quality detection model is established and the hyperparameters of the quality detection model are defined. Specifically, in this stage, the deep features extracted in step S3 and the time series data obtained in step S4 are combined. These feature types include: Deep features: Physical and chemical properties related features extracted from convolutional neural networks (CNNs), including molecular structure features and spectral features. Time series data: Information about the changes in process parameters over time extracted from recurrent neural networks (RNNs), including temperature, pressure, and component concentrations. In order to effectively utilize these two types of features, the following strategies are adopted: Feature fusion: Merge deep features and time series data into a unified representation, which is achieved by concatenation, that is, splicing the two features together to form a vector. Multimodal learning: Design a multimodal model that can process image / spectral data and time series data simultaneously (such as a combination of convolutional neural networks and recurrent neural networks, or using more advanced RNN variants such as long short-term memory networks). The choice of hyperparameters in this step is crucial to the performance of the model. Here are some key hyperparameters that need to be defined: Network architecture: The number of layers determines the complexity of the model. The number of nodes in each layer: affects the learning ability of the model. Activation function: such as ReLU, tanh, sigmoid, affects the nonlinear expression ability. Optimizer: Optimization algorithm (such as Adam, SGD) Learning rate: controls the speed of weight update. Momentum: accelerates the convergence process. Regularization: Dropout ratio: randomly discards a part of the neuron output to reduce overfitting. Weight decay: such as L1 or L2 regularization, used to penalize large weight values. Training parameters: Batch size: The number of samples used in each training iteration. Number of training rounds (epoch number): The number of times a training data set is fully traversed.
[0055] The quality detection model established in step S7 can improve the prediction accuracy and can improve the prediction accuracy of the model for the quality of pharmaceutical intermediates.
[0056] Step S8: using the comprehensive training data set obtained in step S6 to train a quality detection model, which may be a prediction model for specific physicochemical properties;
[0057] Step S9: Integrate the quality detection model into the production line of pharmaceutical intermediates to monitor and predict product quality in real time, and adjust process parameters to ensure product quality.
[0058] The multimodal data in step S1 includes microscope images, chemical structure images, sensor data and chemical analysis data of the pharmaceutical intermediate.
[0059] The sensor data include the temperature, pressure and pH value of the pharmaceutical intermediate, and the chemical analysis data include the chromatographic data and spectral data of the pharmaceutical intermediate.
[0060] The step S2 of cleaning and standardizing the multimodal data includes:
[0061] Step S2-1: Check the collected multimodal data, identify and delete invalid data points or missing values;
[0062] Ensure that the dataset does not contain invalid or incomplete data to improve the accuracy of subsequent analysis.
[0063] Step S2-2: Use a machine learning algorithm to detect outliers in the data and decide whether to keep or delete the outlier; outliers may cause deviations in subsequent analysis, so detecting and processing outliers is crucial to ensuring data quality. Specifically, a machine learning algorithm based on isolation forest, local anomaly factor, principal component analysis, support vector machine or autoencoder can be used to detect outliers in the data. Isolation forest is a tree-based unsupervised learning algorithm that is particularly suitable for anomaly detection tasks. It constructs a decision tree by randomly selecting features and randomly splitting feature values to isolate outliers. Outliers are often easier to isolate, so outliers can be identified by the average path length of the tree. It is suitable for high-dimensional data and does not require a large number of abnormal samples for training. Among them, the local anomaly factor is a density-based anomaly detection algorithm that identifies outliers by calculating the local density deviation of each data point relative to its neighbors. If the local density of a point is much lower than the density of its neighbors, the point is considered an anomaly, which is suitable for data sets with dense and sparse areas. Among them, principal component analysis is a dimensionality reduction technique that can be used for outlier detection. By projecting the data into a low-dimensional space, it is easier to identify data points that deviate from the normal pattern. It is suitable for high-dimensional data and can effectively reduce the data dimension and detect outliers. Among them, support vector machines can be used for anomaly detection, especially when using kernel techniques, which can map data to high-dimensional space to better separate normal data from abnormal data. It is suitable for small to medium-sized data sets, especially when the data is not linearly separable. Among them, the autoencoder is an unsupervised learning model that detects outliers by reconstructing the input data. If the reconstruction error of a data point is much larger than that of other data points, the data point is considered abnormal.
[0064] Step S2-3: For missing data points, interpolation is used to fill them in to ensure the integrity of the data set; data interpolation can reduce the information loss caused by missing data and improve the overall quality of the data set.
[0065] Step S2-4: Scale the data to a uniform scale range, using minimum and maximum scaling or z-score standardization. Standardization can eliminate the impact of data dimensions and enable different features to be compared on the same scale.
[0066] The step S3 of extracting the deep features of the pharmaceutical intermediates from the multimodal data using a convolutional neural network includes:
[0067] Step S3-1: Select CNN architecture: According to the requirements of the physicochemical properties of the pharmaceutical intermediates, select a suitable convolutional neural network (CNN) architecture, such as VGG, ResNet, Inception, etc. Selecting a suitable CNN architecture is the basis for extracting effective features, and can perform efficient feature extraction for specific physicochemical properties.
[0068] Step S3-2: Data preprocessing: Preprocess the image data, including resizing, color space conversion, normalization, etc., to ensure that the data is suitable for input into CNN. Data preprocessing ensures that the format and size of the input data are consistent, which is conducive to improving the training efficiency of CNN.
[0069] Step S3-3: Construct a CNN model; construct the model according to the selected CNN architecture, and define the number of layers, convolution kernel size, pooling layer, fully connected layer, etc. Model construction is a key step in realizing CNN feature extraction, and needs to be carefully designed to meet specific physical and chemical property extraction requirements.
[0070] Step S3-4: Model initialization: Initialize the parameters of the CNN model, which can be done by random initialization or pre-trained model weights. Good initialization helps accelerate model convergence, and pre-trained models can also improve the generalization ability of the model.
[0071] Step S3-5: training the CNN model; using the image data preprocessed in step S3-2 to train the CNN model, the model parameters are updated through the back propagation algorithm. The training process is a key step for the model to learn how to extract features from the input data, and the model parameters are optimized through multiple iterations.
[0072] Step S3-6: Feature extraction: After training is completed, the middle layer of the CNN model is used to extract deep features related to the physicochemical properties of pharmaceutical intermediates. The extracted deep features are the basis for subsequent analysis and prediction, and can provide rich physicochemical information. Specifically, deep features refer to high-level abstract features automatically learned from raw data through deep learning models (such as convolutional neural networks, CNN). These features are usually able to capture complex patterns and structures in the input data, and are particularly effective for image and signal data. In the quality control of pharmaceutical intermediates, deep features can be extracted from multimodal data such as microscope images and chemical structure images to reflect the physicochemical properties of pharmaceutical intermediates.
[0073] The following is an example of the deep feature extraction process (microscope image feature extraction):
[0074] Data preparation: Collect a large amount of microscopic image data of pharmaceutical intermediates. These images should cover samples of different batches and quality levels. Preprocess the images, including resizing, normalization, and other operations. Model selection: Choose a suitable CNN architecture, such as VGG16 or ResNet50. In this example, we choose ResNet50, which can effectively extract deep features and has strong feature expression capabilities.
[0075] Model construction and initialization: Build a ResNet50 model, define the input layer, multiple convolutional layers, pooling layers, fully connected layers, etc. Use pre-trained weights to initialize the model to speed up training and improve model performance.
[0076] Model training: Use the preprocessed microscope image data to train the model, and adjust the model parameters through the back-propagation algorithm to minimize the loss function. Hyperparameters such as learning rate and batch size may need to be adjusted during the training process.
[0077] Feature extraction: After training is complete, the penultimate layer of the model (usually the last feature map output layer before the fully connected layer) is used as the feature extraction layer. The features extracted from this layer are deep features, which can be used to characterize the physicochemical properties of pharmaceutical intermediates. Suppose we have collected 1,000 microscope images of pharmaceutical intermediates, and the image size is uniformly adjusted to 224x224 pixels. These images cover different batches of intermediate samples, including qualified and unqualified products.
[0078] Model construction: Select ResNet50 as the base model. Use ImageNet pre-trained weights for initialization. Define the input layer size as (224, 224, 3), which represents the image size and the number of color channels.
[0079] Training process: The dataset was divided into training set (70%), validation set (15%) and test set (15%). The Adam optimizer was used with an initial learning rate of 0.001 and a batch size of 32. The model was trained for 50 epochs and the learning rate was adjusted every 5 epochs.
[0080] Feature extraction: After training is complete, features are extracted from the penultimate layer of the model (e.g., the layer after the global average pooling layer of ResNet50).
[0081] Specific deep feature examples
[0082] Suppose we extract a set of 2048-dimensional feature vectors from the trained ResNet50 model above. These feature vectors contain information about the physicochemical properties such as particle morphology, crystallinity, and particle size distribution in the microscope images of pharmaceutical intermediates.
[0083] For example, one dimension of the feature vector may reflect the sharpness of the edge of the pharmaceutical intermediate particles, and another dimension may reflect the shape diversity of the particles. These feature vectors can be further used to build classification or regression models to predict the quality of pharmaceutical intermediates.
[0084] Step S3-7: Feature selection and optimization. By analyzing the extracted features, the most representative and discriminating features are selected, and the feature set is optimized. Feature selection and optimization can reduce redundant features and improve the efficiency and prediction performance of the model.
[0085] The step S4 in which a recurrent neural network is used to analyze the data sequence that changes over time during the production process of the pharmaceutical intermediates includes:
[0086] Step S4-1: Select RNN architecture: According to the characteristics of the data sequence in the production process of pharmaceutical intermediates, select a suitable recurrent neural network (RNN) architecture, such as LSTM (Long Short-Term Memory Network) or GRU (Gated Recurrent Unit). Selecting a suitable RNN architecture is the key to analyzing time series data, which can capture long-term dependencies in the data.
[0087] Step S4-2: Data preprocessing: preprocess the time series data, including data standardization, missing value processing, and outlier detection. Data preprocessing ensures the quality of the data and is conducive to improving the training effect of the RNN model.
[0088] Step S4-3: Build the RNN model; build the model according to the selected RNN architecture, define the number of layers, unit type (LSTM or GRU), and hidden layer size of the model. Building the model is the basis for time series analysis and needs to be carefully designed to meet specific data analysis needs.
[0089] Step S4-4: Model initialization: Initialize the parameters of the RNN model, which can be initialized randomly or with the weights of a pre-trained model. Good initialization helps accelerate model convergence, and a pre-trained model can also improve the generalization ability of the model.
[0090] Step S4-5: Train the RNN model; use the data sequence preprocessed in step S4-2 to train the RNN model, and update the model parameters through the back propagation algorithm. The training process is a key step for the model to learn how to extract features from time series data, and the model parameters are optimized through multiple iterations.
[0091] Step S4-6: Feature extraction: After training is completed, the hidden layer state of the RNN model is used as the feature representation of the data sequence that changes over time during the production process of pharmaceutical intermediates. The extracted features can capture the trends and periodic changes in the time series, providing valuable information for subsequent analysis.
[0092] Step S4-7: Perform trend analysis and anomaly detection on the extracted features. By analyzing the extracted features, potential trends and anomalies in the data sequence are identified. Trend analysis and anomaly detection can help identify abnormal situations in the production process and help take timely measures.
[0093] In step S5, based on the deep features obtained in step S3 and the data sequence obtained in step S4, a comprehensive training data set including deep features and time series data is constructed, including:
[0094] Step S5-1: Merge the deep features extracted in step S3 and the time series features extracted in step S4 into one data set; ensure that each sample in the data set contains the corresponding deep features and time series data to facilitate subsequent training.
[0095] Step S5-2: Pair the deep features of the image data with the time series data in the corresponding time period according to the timestamp of the production process; ensure that each deep feature is synchronized with the corresponding time series data in time, which helps the model learn the association between different modal data.
[0096] Step S5-3: Split the comprehensive dataset into training set, validation set and test set to ensure the effectiveness of model training and the fairness of evaluation. Through data segmentation, the performance of the model on unseen data can be evaluated to ensure the generalization ability of the model.
[0097] Step S5-4: Assign a label to each sample in the comprehensive dataset. The label can be a quality indicator of the pharmaceutical intermediate (such as purity, yield, etc.). Labels are the key to supervised learning model training, ensuring that the model can learn the relationship between input data and quality indicators.
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for quality control of pharmaceutical intermediates based on machine learning, characterized in that: The following steps are involved: Step S1: Collecting multimodal data of pharmaceutical intermediates; The multimodal data includes image data, sensor data and chemical analysis data of the pharmaceutical intermediate; The image data of the pharmaceutical intermediate includes microscope image and chemical structure image data; The sensor data includes the temperature, pressure and pH value of the pharmaceutical intermediate, and the chemical analysis data includes the chromatographic data and spectral data of the pharmaceutical intermediate; Step S2: performing data cleaning and standardization on the multimodal data; Step S3: extracting deep features of the pharmaceutical intermediate from the image data of the pharmaceutical intermediate using a convolutional neural network; Step S4: applying a recurrent neural network to analyze the data sequence that changes over time during the production process of the pharmaceutical intermediate; Step S5: Based on the deep features obtained in step S3 and the data sequence obtained in step S4, a comprehensive training data set including deep features and time series data is constructed; Step S6: Generate virtual samples using a generative adversarial network to enhance the comprehensive training data set in step S5; Step S7: Based on the feature type of the deep features extracted in step S3 and the data sequence obtained in step S4, a quality detection model is established, and hyper parameters of the quality detection model are defined; Step S8: training a quality detection model using the comprehensive training data set obtained in step S6; Step S9: Integrate the quality inspection model into the production line of pharmaceutical intermediates.
2. A method for quality control of pharmaceutical intermediates based on machine learning drive according to claim 1, characterized in that: The step S2 of cleaning and standardizing the multimodal data includes: Step S2-1: Check the collected multimodal data, identify and delete invalid data points or missing values; Step S2-2: using a machine learning algorithm to detect outliers in the data and decide whether to keep or delete the outliers; Step S2-3: For missing data points, interpolation is used to fill them in to ensure the integrity of the data set; Step S2-4: Scale the data to a uniform scale using minimum-maximum scaling or z-score normalization.
3. The method for quality control of pharmaceutical intermediates based on machine learning drive according to claim 1, characterized in that: The step S3 of extracting the deep features of the pharmaceutical intermediates from the multimodal data using a convolutional neural network includes: Step S3-1: Select CNN architecture; Step S3-2: data preprocessing; Step S3-3: construct a CNN model; Step S3-4: model initialization; Step S3-5: training the CNN model; Step S3-6: feature extraction; Step S3-7: Feature selection and optimization.
4. The method for quality control of pharmaceutical intermediates based on machine learning drive according to claim 1, characterized in that: The step S4 in which a recurrent neural network is used to analyze the data sequence that changes over time during the production process of the pharmaceutical intermediates includes: Step S4-1: Select RNN architecture; Step S4-2: data preprocessing; Step S4-3: constructing an RNN model; Step S4-4: model initialization; Step S4-5: training the RNN model; Step S4-6: feature extraction; Step S4-7: Perform trend analysis and anomaly detection on the extracted features.
5. The method for quality control of pharmaceutical intermediates based on machine learning drive according to claim 1, characterized in that: In step S5, based on the deep features obtained in step S3 and the data sequence obtained in step S4, a comprehensive training data set including deep features and time series data is constructed, including: Step S5-1: merge the deep features extracted in step S3 and the time series features extracted in step S4 into one dataset; Step S5-2: Pairing the depth features of the image data with the time series data in the corresponding time period according to the timestamp of the production process; Step S5-3: Split the comprehensive dataset into a training set, a validation set, and a test set; Step S5-4: Assign a label to each sample in the comprehensive dataset.
